"RDNA 2" Instruction Set Architecture

Reference Guide

AMD

30-November-2020
"RDNA 2" Instruction Set Architecture

Specification Agreement

This Specification Agreement (this "Agreement") is a legal agreement between Advanced Micro Devices, Inc. ("AMD") and "You"
as the recipient of the attached AMD Specification (the "Specification"). If you are accessing the Specification as part of your
performance of work for another party, you acknowledge that you have authority to bind such party to the terms and
conditions of this Agreement. If you accessed the Specification by any means or otherwise use or provide Feedback (defined
below) on the Specification, You agree to the terms and conditions set forth in this Agreement. If You do not agree to the terms
and conditions set forth in this Agreement, you are not licensed to use the Specification; do not use, access or provide Feedback
about the Specification. In consideration of Your use or access of the Specification (in whole or in part), the receipt and
sufficiency of which are acknowledged, You agree as follows:

1. You may review the Specification only (a) as a reference to assist You in planning and designing Your product, service or
technology ("Product") to interface with an AMD product in compliance with the requirements as set forth in the
Specification and (b) to provide Feedback about the information disclosed in the Specification to AMD.

2. Except as expressly set forth in Paragraph 1, all rights in and to the Specification are retained by AMD. This Agreement
does not give You any rights under any AMD patents, copyrights, trademarks or other intellectual property rights. You
may not (i) duplicate any part of the Specification; (ii) remove this Agreement or any notices from the Specification, or (iii)
give any part of the Specification, or assign or otherwise provide Your rights under this Agreement, to anyone else.

3. The Specification may contain preliminary information, errors, or inaccuracies, or may not include certain necessary
information. Additionally, AMD reserves the right to discontinue or make changes to the Specification and its products at
any time without notice. The Specification is provided entirely "AS IS." AMD MAKES NO WARRANTY OF ANY KIND AND
DISCLAIMS ALL EXPRESS, IMPLIED AND STATUTORY WARRANTIES, INCLUDING BUT NOT LIMITED TO IMPLIED
WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE, NONINFRINGEMENT, TITLE OR THOSE
WARRANTIES ARISING AS A COURSE OF DEALING OR CUSTOM OF TRADE. AMD SHALL NOT BE LIABLE FOR DIRECT,
INDIRECT, CONSEQUENTIAL, SPECIAL, INCIDENTAL, PUNITIVE OR EXEMPLARY DAMAGES OF ANY KIND (INCLUDING
LOSS OF BUSINESS, LOSS OF INFORMATION OR DATA, LOST PROFITS, LOSS OF CAPITAL, LOSS OF GOODWILL)
REGARDLESS OF THE FORM OF ACTION WHETHER IN CONTRACT, TORT (INCLUDING NEGLIGENCE) AND STRICT
PRODUCT LIABILITY OR OTHERWISE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGES.

4. Furthermore, AMD’s products are not designed, intended, authorized or warranted for use as components in systems
intended for surgical implant into the body, or in other applications intended to support or sustain life, or in any other
application in which the failure of AMD’s product could create a situation where personal injury, death, or severe
property or environmental damage may occur.

5. You have no obligation to give AMD any suggestions, comments or feedback ("Feedback") relating to the Specification.
However, any Feedback You voluntarily provide may be used by AMD without restriction, fee or obligation of
confidentiality. Accordingly, if You do give AMD Feedback on any version of the Specification, You agree AMD may freely
use, reproduce, license, distribute, and otherwise commercialize Your Feedback in any product, as well as has the right to
sublicense third parties to do the same. Further, You will not give AMD any Feedback that You may have reason to believe
is (i) subject to any patent, copyright or other intellectual property claim or right of any third party; or (ii) subject to
license terms which seek to require any product or intellectual property incorporating or derived from Feedback or any
Product or other AMD intellectual property to be licensed to or otherwise provided to any third party.

6. You shall adhere to all applicable U.S., European, and other export laws, including but not limited to the U.S. Export
Administration Regulations ("EAR"), (15 C.F.R. Sections 730 through 774), and E.U. Council Regulation (EC) No 428/2009 of
5 May 2009. Further, pursuant to Section 740.6 of the EAR, You hereby certifies that, except pursuant to a license granted
by the United States Department of Commerce Bureau of Industry and Security or as otherwise permitted pursuant to a
License Exception under the U.S. Export Administration Regulations ("EAR"), You will not (1) export, re-export or release
to a national of a country in Country Groups D:1, E:1 or E:2 any restricted technology, software, or source code You receive
hereunder, or (2) export to Country Groups D:1, E:1 or E:2 the direct product of such technology or software, if such
foreign produced direct product is subject to national security controls as identified on the Commerce Control List
(currently found in Supplement 1 to Part 774 of EAR). For the most current Country Group listings, or for additional

ii of 283
"RDNA 2" Instruction Set Architecture

information about the EAR or Your obligations under those regulations, please refer to the U.S. Bureau of Industry and
Security’s website at http://www.bis.doc.gov/.

7. If You are a part of the U.S. Government, then the Specification is provided with "RESTRICTED RIGHTS" as set forth in
subparagraphs (c) (1) and (2) of the Commercial Computer Software-Restricted Rights clause at FAR 52.227-14 or
subparagraph (c) (1)(ii) of the Rights in Technical Data and Computer Software clause at DFARS 252.277-7013, as
applicable.

8. This Agreement is governed by the laws of the State of California without regard to its choice of law principles. Any
dispute involving it must be brought in a court having jurisdiction of such dispute in Santa Clara County, California, and
You waive any defenses and rights allowing the dispute to be litigated elsewhere. If any part of this agreement is
unenforceable, it will be considered modified to the extent necessary to make it enforceable, and the remainder shall
continue in effect. The failure of AMD to enforce any rights granted hereunder or to take action against You in the event
of any breach hereunder shall not be deemed a waiver by AMD as to subsequent enforcement of rights or subsequent
actions in the event of future breaches. This Agreement is the entire agreement between You and AMD concerning the
Specification; it may be changed only by a written document signed by both You and an authorized representative of
AMD.

DISCLAIMER

The information contained herein is for informational purposes only, and is subject to change without notice. This
document may contain technical inaccuracies, omissions and typographical errors, and AMD is under no obligation to
update or otherwise correct this information. Advanced Micro Devices, Inc. makes no representations or warranties
with respect to the accuracy or completeness of the contents of this document, and assumes no liability of any kind,
including the implied warranties of noninfringement, merchantability or fitness for particular purposes, with respect
to the operation or use of AMD hardware, software or other products described herein. No license, including implied
or arising by estoppel, to any intellectual property rights is granted by this document. Terms and limitations
applicable to the purchase or use of AMD’s products or technology are as set forth in a signed agreement between the
parties or in AMD’s Standard Terms and Conditions of Sale.

AMD, the AMD Arrow logo, and combinations thereof are trademarks of Advanced Micro Devices, Inc. OpenCL is a
trademark of Apple Inc. used by permission by Khronos Group, Inc. OpenGL® and the oval logo are trademarks or
registered trademarks of Hewlett Packard Enterprise in the United States and/or other countries worldwide. DirectX is
a registered trademark of Microsoft Corporation in the US and other jurisdictions. Other product names used in this
publication are for identification purposes only and may be trademarks of their respective companies.

© 2018-2020 Advanced Micro Devices, Inc. All rights reserved.

Advanced Micro Devices, Inc.

2485 Augustine Drive
Santa Clara, CA, 95054
www.amd.com

iii of 283
"RDNA 2" Instruction Set Architecture

Contents
Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1  

About This Document. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1  

Audience . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1  

Organization. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1  

Conventions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2  

Related Documents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2  

Feature Changes in RDNA2 Devices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2  

Instruction Changes. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3  

Additional Information . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3  

1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4  

1.1. Terminology. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5  

2. Program Organization. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7  

2.1. Wave32 and Wave64 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7  

2.2. Compute Shaders . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8  

2.3. Data Sharing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8  

2.3.1. Local Data Share (LDS) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9  

2.3.2. Global Data Share (GDS) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10  

2.4. Device Memory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10  

2.5. Shader Padding Requirement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11  

3. Kernel State . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12 

3.1. State Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12  

3.2. Program Counter (PC) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13  

3.3. EXECute Mask . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13  

3.4. Status registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14  

3.5. Mode register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15  

3.6. GPRs and LDS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16  

3.6.1. Out-of-Range behavior . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16  

3.6.2. SGPR Allocation and storage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17  

3.6.3. SGPR Alignment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18  

3.6.4. VGPR Allocation and Alignment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18  

3.6.5. Wave Shared VGPRs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18  

3.6.6. LDS Allocation and Clamping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18  

3.7. M# Memory Descriptor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18  

3.8. SCC: Scalar Condition code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19  

3.9. Vector Compares: VCC and VCCZ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19  

3.10. Trap and Exception registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20  

3.10.1. Trap Status register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21  

3.11. Memory Violations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22  

3.12. Initial Wave State . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22  

3.12.1. State Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23  

iv of 283
"RDNA 2" Instruction Set Architecture

3.12.2. SGPR Initialization. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   23

4. Program Flow Control. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   26
4.1. Program Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   26
4.1.1. Instruction Clauses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   26
4.2. Branching . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   27
4.2.1. Subvector Execution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   27
4.3. Workgroups. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   29
4.4. Data Dependency Resolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   29
4.5. Manually Inserted Wait States (NOPs) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   30
5. Scalar ALU Operations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   31
5.1. SALU Instruction Formats . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   31
5.2. Scalar ALU Operands . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   32
5.3. Scalar Condition Code (SCC) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   33
5.4. Integer Arithmetic Instructions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   34
5.5. Conditional Instructions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   34
5.6. Comparison Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   35
5.7. Bit-Wise Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   35
5.8. Access Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   37
6. Vector ALU Operations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   40
6.1. Microcode Encodings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   40
6.2. Operands . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   41
6.2.1. Instruction Inputs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   42
6.2.2. Instruction Outputs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   43
6.2.3. Wave64 use of SGPRs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   44
6.2.4. Wave64 Destination Restrictions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   44
6.2.5. OPSEL Field Restrictions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   44
6.2.6. Out-of-Range GPRs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   45
6.3. Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   45
6.4. Denormalized and Rounding Modes. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   47
6.5. ALU Clamp Bit Usage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   48
6.6. VGPR Indexing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   48
6.7. Packed Math . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   48
6.8. Sub-Dword Addressing (SDWA) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   49
6.9. Data Parallel Processing (DPP) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   49
6.10. PERMLANE Specific Rules. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   50
7. Scalar Memory Operations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   51
7.1. Microcode Encoding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   51
7.2. Operations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   51
7.2.1. S_LOAD_DWORD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   52
7.2.2. S_DCACHE_INV . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   53
7.2.3. S_MEMREALTIME . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   53
7.3. Dependency Checking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   53

v of 283
"RDNA 2" Instruction Set Architecture

7.4. Scalar Memory Clauses and Groups . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   53

7.5. Alignment and Bounds Checking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   54
8. Vector Memory Operations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   55
8.1. Vector Memory Buffer Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   55
8.1.1. Simplified Buffer Addressing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   56
8.1.2. Buffer Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   56
8.1.3. VGPR Usage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   58
8.1.4. Buffer Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   58
8.1.5. Buffer Addressing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   59
8.1.6. 16-bit Memory Operations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   64
8.1.7. Alignment. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   64
8.1.8. Buffer Resource. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   64
8.1.9. Memory Buffer Load to LDS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   65
8.1.10. GLC, DLC and SLC Bits Explained . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   66
8.2. Vector Memory (VM) Image Instructions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   67
8.2.1. Image Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   68
8.2.2. Image Non-Sequential Address (NSA) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   70
8.2.3. Image Opcodes with No Sampler . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   71
8.2.4. Image Opcodes with a Sampler . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   71
8.2.5. VGPR Usage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   73
8.2.6. Image Resource . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   75
8.2.7. Image Sampler . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   76
8.2.8. Data Formats. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   78
8.2.9. Vector Memory Instruction Data Dependencies . . . . . . . . . . . . . . . . . . . . . . . . . . .   79
8.2.10. Ray Tracing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   80
9. Flat Memory Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   83
9.1. Flat Memory Instruction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   83
9.2. Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   85
9.2.1. Ordering. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   86
9.2.2. Important Timing Consideration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   86
9.3. Addressing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   86
9.3.1. Legal Addressing Combinations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   86
9.4. Global . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   87
9.5. Scratch . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   87
9.6. Memory Error Checking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   88
9.7. Data. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   88
9.8. Scratch Space (Private) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   88
10. Data Share Operations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   90
10.1. Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   90
10.2. Dataflow in Memory Hierarchy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   91
10.3. LDS Modes and Allocation: CU vs. WGP Mode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   91
10.4. LDS Access. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   92

vi of 283
"RDNA 2" Instruction Set Architecture

10.4.1. LDS Direct Reads . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92  

10.4.2. LDS Parameter Reads . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93  

10.4.3. Data Share Indexed and Atomic Access . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94  

10.4.4. LDS Lane-permute Ops. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96  

10.5. Global Data Share. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97  

11. Exporting Pixel and Vertex Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98  

11.1. Microcode Encoding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98  

11.2. Operations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99  

11.2.1. Pixel Shader Exports . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99  

11.2.2. Vertex Shader Exports. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99  

11.3. Primitive Shader Exports. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100  

11.4. Dependency Checking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100  

12. Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101

 

12.1. SOP2 Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101  

12.2. SOPK Instructions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106  

12.3. SOP1 Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111  

12.4. SOPC Instructions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122  

12.5. SOPP Instructions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123  

12.5.1. Send Message. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129  

12.6. SMEM Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130  

12.7. VOP2 Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130  

12.7.1. VOP2 using VOP3 encoding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137  

12.8. VOP1 Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137  

12.8.1. VOP1 using VOP3 encoding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150  

12.9. VOPC Instructions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150  

12.9.1. VOPC using VOP3A encoding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162  

12.10. VOP3P Instructions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162  

12.11. VINTERP Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166  

12.11.1. VINTERP using VOP3 encoding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166  

12.12. VOP3A & VOP3B Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167  

12.13. LDS & GDS Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188  

12.13.1. DS_SWIZZLE_B32 Details . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207  

12.13.2. LDS Instruction Limitations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209  

12.14. MUBUF Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209  

12.15. MTBUF Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215  

12.16. MIMG Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 216  

12.17. EXPORT Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223  

12.18. FLAT, Scratch and Global Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 224  

12.18.1. Flat Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 224  

12.18.2. Scratch Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 228  

12.18.3. Global Instructions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229  

12.19. Instruction Limitations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 234  

vii of 283
"RDNA 2" Instruction Set Architecture

12.19.1. DPP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   234

12.19.2. SDWA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   235
13. Microcode Formats . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   236
13.1. Scalar ALU and Control Formats . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   237
13.1.1. SOP2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   238
13.1.2. SOPK. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   240
13.1.3. SOP1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   241
13.1.4. SOPC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   243
13.1.5. SOPP. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   245
13.2. Scalar Memory Format . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   246
13.2.1. SMEM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   246
13.3. Vector ALU Formats . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   247
13.3.1. VOP2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   247
13.3.2. VOP1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   249
13.3.3. VOPC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   252
13.3.4. VOP3A. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   257
13.3.5. VOP3B. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   260
13.3.6. VOP3P. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   263
13.3.7. SDWA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   265
13.3.8. SDWAB . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   266
13.3.9. DPP16 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   267
13.3.10. DPP8 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   268
13.4. Vector Parameter Interpolation Format. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   269
13.4.1. VINTRP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   269
13.5. LDS and GDS format . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   270
13.5.1. DS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   270
13.6. Vector Memory Buffer Formats . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   272
13.6.1. MTBUF . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   272
13.6.2. MUBUF . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   274
13.7. Vector Memory Image Format . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   276
13.7.1. MIMG. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   276
13.8. Flat Formats . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   279
13.8.1. FLAT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   279
13.8.2. GLOBAL . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   281
13.8.3. SCRATCH . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   282
13.9. Export Format . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   282
13.9.1. EXP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   282

viii of 283
"RDNA 2" Instruction Set Architecture

Preface

About This Document

This document describes the current environment, organization and program state of AMD
"RDNA" Generation devices. It details the instruction set and the microcode formats native to
this family of processors that are accessible to programmers and compilers.

The document specifies the instructions (include the format of each type of instruction) and the
relevant program state (including how the program state interacts with the instructions). Some
instruction fields are mutually dependent; not all possible settings for all fields are legal. This
document specifies the valid combinations.

The main purposes of this document are to:

1. Specify the language constructs and behavior, including the organization of each type of
instruction in both text syntax and binary format.
2. Provide a reference of instruction operation that compiler writers can use to maximize
performance of the processor.

Audience
This document is intended for programmers writing application and system software, including
operating systems, compilers, loaders, linkers, device drivers, and system utilities. It assumes
that programmers are writing compute-intensive parallel applications (streaming applications)
and assumes an understanding of requisite programming practices.

Organization
This document begins with an overview of the AMD RDNA processors' hardware and
programming environment (Chapter 1).
Chapter 2 describes the organization of RDNA programs.
Chapter 3 describes the program state that is maintained.
Chapter 4 describes the program flow.
Chapter 5 describes the scalar ALU operations.
Chapter 6 describes the vector ALU operations.
Chapter 7 describes the scalar memory operations.
Chapter 8 describes the vector memory operations.
Chapter 9 provides information about the flat memory instructions.
Chapter 10 describes the data share operations.
Chapter 11 describes exporting the parameters of pixel color and vertex shaders.
Chapter 12 describes instruction details, first by the microcode format to which they belong,

About This Document 1 of 283

"RDNA 2" Instruction Set Architecture

then in alphabetic order.

Finally, Chapter 13 provides a detailed specification of each microcode format.

Conventions
The following conventions are used in this document:

mono-spaced font A filename, file path or code.

* Any number of alphanumeric characters in the name of a code format,

parameter, or instruction.

<> Angle brackets denote streams.

[1,2) A range that includes the left-most value (in this case, 1), but excludes the right-
most value (in this case, 2).

[1,2] A range that includes both the left-most and right-most values.

{x | y} One of the multiple options listed. In this case, X or Y.

0.0 A single-precision (32-bit) floating-point value.

1011b A binary value, in this example a 4-bit value.

7:4 A bit range, from bit 7 to bit 4, inclusive. The high-order bit is shown first.

italicized word or phrase The first use of a term or concept basic to the understanding of stream
computing.

Related Documents
• Intermediate Language (IL) Reference Manual. Published by AMD.
• AMD Accelerated Parallel Processing OpenCL™ Programming Guide. Published by AMD.
• The OpenCL™ Specification. Published by Khronos Group. Aaftab Munshi, editor.
• OpenGL® Programming Guide, at http://www.glprogramming.com/red/
• Microsoft DirectX® Reference Website, at https://msdn.microsoft.com/en-us/library/
windows/desktop/ee663274(v=vs.85).aspx

Feature Changes in RDNA2 Devices

This section highlights some notable changes:

• Ray Tracing
• Dot product ALU operations added accelerate inferencing and deep-learning:
◦ V_DOT2_F32_F16 / V_DOT2C_F32_F16
◦ V_DOT2_I32_I16 / V_DOT2_U32_U16
◦ V_DOT4_I32_I8 / V_DOT4C_I32_I8

Conventions 2 of 283
"RDNA 2" Instruction Set Architecture

◦ V_DOT4_U32_U8
◦ V_DOT8_I32_I4
◦ V_DOT8_U32_U4
• Image Load MSAA
• Global memory loads with "Add-TID"
• Atomic clamped subtract buffer and global instructions
• VGPR & LDS allocation-unit size doubled
• S_MEMTIME replaced by "s_getreg_b32 Sn, SHADER_CYCLES"

Instruction Changes
Removed:

• V_MAC_LEGACY_F32 (replaced by V_FMAC_LEGACY_F32)

• V_MAD_LEGACY_F32 (replaced by V_FMA_LEGACY_F32)
• V_MAC_F32, V_MADMK_F32, V_MADAK_F32 (replaced by FMA equivalents)

Additional Information
For more information on AMD GPU architectures please visit https://GPUOpen.com

Additional Information 3 of 283

"RDNA 2" Instruction Set Architecture

Chapter 1. Introduction
The AMD RDNA processor implements a parallel micro-architecture that provides an excellent
platform not only for computer graphics applications but also for general-purpose data parallel
applications. Data-intensive applications that require high bandwidth or are computationally
intensive may be run on an AMD RDNA processor.

The figure below shows a block diagram of the AMD RDNA Generation series processors

Figure 1. AMD RDNA Generation Series Block Diagram

The RDNA device includes a data-parallel processor (DPP) array, a command processor, a
memory controller, and other logic (not shown). The RDNA command processor reads
commands that the host has written to memory-mapped RDNA registers in the system-memory
address space. The command processor sends hardware-generated interrupts to the host when
the command is completed. The RDNA memory controller has direct access to all RDNA device
memory and the host-specified areas of system memory. To satisfy read and write requests, the
memory controller performs the functions of a direct-memory access (DMA) controller, including
computing memory-address offsets based on the format of the requested data in memory. In the
RDNA environment, a complete application includes two parts:

• a program running on the host processor, and

• programs, called kernels, running on the RDNA processor.

The RDNA programs are controlled by host commands that

• set RDNA internal base-address and other configuration registers,

4 of 283
"RDNA 2" Instruction Set Architecture

• specify the data domain on which the RDNA GPU is to operate,

• invalidate and flush caches on the RDNA GPU, and
• cause the RDNA GPU to begin execution of a program.

The RDNA driver program runs on the host.

The DPP array is the heart of the RDNA processor. The array is organized as a set of
workgroup processor pipelines, each independent from the others, that operate in parallel on
streams of floating-point or integer data. The workgroup processor pipelines can process data
or, through the memory controller, transfer data to, or from, memory. Computation in a
workgroup processor pipeline can be made conditional. Outputs written to memory can also be
made conditional.

When it receives a request, the workgroup processor pipeline loads instructions and data from
memory, begins execution, and continues until the end of the kernel. As kernels are running, the
RDNA hardware is designed to automatically fetch instructions from memory into on-chip
caches; RDNA software plays no role in this. RDNA kernels can load data from off-chip memory
into on-chip general-purpose registers (GPRs) and caches.

The AMD RDNA devices can detect floating point exceptions and can generate interrupts. In
particular, they detect IEEE-754 floating-point exceptions in hardware; these can be recorded
for post-execution analysis. The software interrupts shown in the previous figure from the
command processor to the host represent hardware-generated interrupts for signaling
command-completion and related management functions.

The RDNA processor hides memory latency by keeping track of potentially hundreds of work-
items in various stages of execution, and by overlapping compute operations with memory-
access operations.

1.1. Terminology
Table 1. Basic Terms
Term Description

RDNA Processor The RDNA shader processor is a scalar and vector ALU designed to run complex
programs on behalf of a wavefront.

Dispatch A dispatch launches a 1D, 2D, or 3D grid of work to the RDNA processor array.

Workgroup A workgroup is a collection of wavefronts that have the ability to synchronize with each
other quickly; they also can share data through the Local Data Share.

Wavefront A collection of 32 or 64 work-items that execute in parallel on a single RDNA processor.

Work-item A single element of work: one element from the dispatch grid, or in graphics a pixel or
vertex.

Literal Constant A 32-bit integer or float constant that is placed in the instruction stream.

Scalar ALU (SALU) The scalar ALU operates on one value per wavefront and manages all control flow.

1.1. Terminology 5 of 283

"RDNA 2" Instruction Set Architecture

Term Description

Vector ALU (VALU) The vector ALU maintains Vector GPRs that are unique for each work item and execute
arithmetic operations uniquely on each work-item.

Workgroup Processor The basic unit of shader computation hardware, including scalar & vector ALU’s and
(WGP) memory, as well as LDS and scalar caches.

Compute Unit (CU) One half of a WGP. Contains 2 SIMD32’s which share one path to memory.

Microcode format The microcode format describes the bit patterns used to encode instructions. Each
instruction is either 32 or more bits, in units of 32-bits.

Instruction An instruction is the basic unit of the kernel. Instructions include: vector ALU, scalar
ALU, memory transfer, and control flow operations.

Quad A quad is a 2x2 group of screen-aligned pixels. This is relevant for sampling texture
maps.

Texture Sampler (S#) A texture sampler is a 128-bit entity that describes how the vector memory system
reads and samples (filters) a texture map.

Texture Resource (T#) A texture resource descriptor describes an image in memory: address, data format,
width, height, depth, etc.

Buffer Resource (V#) A buffer resource descriptor describes a buffer in memory: address, data format, stride,
etc.

UTC Universal (Address) Translation Cache : used for virtual memory translating logical to
physical addresses.

1.1. Terminology 6 of 283

"RDNA 2" Instruction Set Architecture

Chapter 2. Program Organization

RDNA kernels are programs executed by the RDNA processor. Conceptually, the kernel is
executed independently on every work-item, but in reality the RDNA processor groups 32 or 64
work-items into a wavefront, which executes the kernel on all 32 or 64 work-items in one pass.

The RDNA processor consists of:

• A scalar ALU, which operates on one value per wavefront (common to all work items).
• A vector ALU, which operates on unique values per work-item.
• Local data storage, which allows work-items within a workgroup to communicate and share
data.
• Scalar memory, which can transfer data between SGPRs and memory through a cache.
• Vector memory, which can transfer data between VGPRs and memory, including sampling
texture maps.

All kernel control flow is handled using scalar ALU instructions. This includes if/else, branches
and looping. Scalar ALU (SALU) and memory instructions work on an entire wavefront and
operate on up to two SGPRs, as well as literal constants.

Vector memory and ALU instructions operate on all work-items in the wavefront at one time. In
order to support branching and conditional execute, every wavefront has an EXECute mask that
determines which work-items are active at that moment, and which are dormant. Active work-
items execute the vector instruction, and dormant ones treat the instruction as a NOP. The
EXEC mask can be changed at any time by Scalar ALU instructions.

Vector ALU instructions can take up to three arguments, which can come from VGPRs, SGPRs,
or literal constants that are part of the instruction stream. They operate on all work-items
enabled by the EXEC mask. Vector compare and add with- carryout return a bit-per-work-item
mask back to the SGPRs to indicate, per work-item, which had a "true" result from the compare
or generated a carry-out.

Vector memory instructions transfer data between VGPRs and memory. Each work-item
supplies its own memory address and supplies or receives unique data. These instructions are
also subject to the EXEC mask.

2.1. Wave32 and Wave64

The shader hardware is designed to support both wavefronts of 32 workitems ("wave32") and
wavefronts of 64 workitems ("wave64"). Both wave sizes are supported for all operations, but
shader programs must be compiled for a particular wave size. The underlying hardware is
primarily natively wave32, and wave64 vector ALU and memory operations are executed by
issuing the instruction twice: once for the low 32 workitems, and then again for the high 32
workitems. Either half of the execution of a wave64 may be skipped if there is no work to do for

2.1. Wave32 and Wave64 7 of 283

"RDNA 2" Instruction Set Architecture

that half (i.e. EXEC == 0 for that half). Wave64 VALU instructions which return a scalar (SGPR
or VCC) value do not skip either pass. Wave64 Vector Memory instructions can skip either pass,
but do not skip both passes.

The upper half of EXEC and VCC are ignored for wave32 waves.

2.2. Compute Shaders

Compute kernels (shaders) are generic programs that can run on the RDNA processor, taking
data from memory, processing it, and writing results back to memory. Compute kernels are
created by a dispatch, which causes the RDNA processors to run the kernel over all of the work-
items in a 1D, 2D, or 3D grid of data. The RDNA processor walks through this grid and
generates wavefronts, which then run the compute kernel. Each work-item is initialized with its
unique address (index) within the grid. Based on this index, the work-item computes the
address of the data it is required to work on and what to do with the results.

2.3. Data Sharing

The AMD RDNA stream processors are designed to share data between different work-items.
Data sharing can boost performance. The figure below shows the memory hierarchy that is
available to each work-item.

2.2. Compute Shaders 8 of 283

"RDNA 2" Instruction Set Architecture

Figure 2. Shared Memory Hierarchy

2.3.1. Local Data Share (LDS)

Each workgroup processor (WGP) has a 128 kB memory space that enables low-latency
communication between work-items within a workgroup, or the work-items within a wavefront;
this is the local data share (LDS). This memory is configured with 64 banks, each with 512
entries of 4 bytes. The AMD RDNA processors use a 128 kB local data share (LDS) memory for
each WGP; this enables 128 kB of low-latency bandwidth to the processing elements. The
shared memory contains 64 integer atomic units to enable fast, unordered atomic operations.
This memory can be used as a software cache for predictable re-use of data, a data exchange
machine for the work-items of a workgroup, or as a cooperative way to enable efficient access
to off-chip memory. A single workgroup may allocate up to 64kB of LDS space.

LDS Allocation Modes

When a workgroup is dispatched or a graphics draw is launched, the waves can allocate LDS
space in one of two modes: CU or WGP mode. The shader can simultaneously execute some
waves in LDS mode and other waves in CU mode.

• CU mode: in this mode, the LDS is effectively split into a separate upper and lower LDS,

2.3. Data Sharing 9 of 283

"RDNA 2" Instruction Set Architecture

each serving two SIMD32’s.

Wave are allocated LDS space within the half of LDS which is associated with the SIMD the
wave is running on.
For workgroups, all waves are assigned to the pair of SIMD32’s. This mode may provide
faster operation since both halves run in parallel, but limits data sharing
(upper waves cannot read data in the lower half of LDS and vice versa).
When in CU mode, all waves in the workgroup are resident within the same CU.
• WGP mode: in this mode, the LDS is one large contiguous memory that all waves on the
WGP can access.
In WGP mode, waves of a workgroup may be distributed across both CU’s (all 4 SIMD32’s)
in the WGP.

2.3.2. Global Data Share (GDS)

The AMD RDNA devices use a 64 kB global data share (GDS) memory that can be used by
wavefronts of a kernel on all WGPs. This memory provides 128 bytes per cycle of memory
access to all the processing elements. The GDS is configured with 32 banks, each with 512
entries of 4 bytes each. It is designed to provide full access to any location for any processor.
The shared memory contains 32 integer atomic units to enable fast, unordered atomic
operations. This memory can be used as a software cache to store important control data for
compute kernels, reduction operations, or a small global shared surface. Data can be preloaded
from memory prior to kernel launch and written to memory after kernel completion. The GDS
block contains support logic for unordered append/consume and domain launch ordered
append/consume operations to buffers in memory. These dedicated circuits enable fast
compaction of data or the creation of complex data structures in memory.

2.4. Device Memory

The AMD RDNA devices offer several methods for access to off-chip memory from the
processing elements (PE) within each WGP. On the primary read path, the device consists of
multiple channels of L2 cache that provides data to Read-only L1 caches, and finally to L0
caches per WGP. Specific cache-less load instructions can force data to be retrieved from
device memory during an execution of a load clause. Load requests that overlap within the
clause are cached with respect to each other. The output cache is formed by two levels of
cache: the first for write-combining cache (collect scatter and store operations and combine
them to provide good access patterns to memory); the second is a read/write cache with atomic
units that lets each processing element complete unordered atomic accesses that return the
initial value. Each processing element provides the destination address on which the atomic
operation acts, the data to be used in the atomic operation, and a return address for the
read/write atomic unit to store the pre-op value in memory. Each store or atomic operation can
be set up to return an acknowledgment to the requesting PE upon write confirmation of the
return value (pre-atomic op value at destination) being stored to device memory.

This acknowledgment has two purposes:

2.4. Device Memory 10 of 283

"RDNA 2" Instruction Set Architecture

• enabling a PE to recover the pre-op value from an atomic operation by performing a cache-
less load from its return address after receipt of the write confirmation acknowledgment,
and
• enabling the system to maintain a relaxed consistency model.

Each scatter write from a given PE to a given memory channel maintains order. The
acknowledgment enables one processing element to implement a fence to maintain serial
consistency by ensuring all writes have been posted to memory prior to completing a
subsequent write. In this manner, the system can maintain a relaxed consistency model
between all parallel work-items operating on the system.

2.5. Shader Padding Requirement

Due to aggressive instruction prefetching used in some graphics devices, the user must pad all
shaders with 64 extra dwords (256 bytes) of data past the end of the shader. It is recommended
to use the S_CODE_END instruction as padding. This ensures that if the instruction prefetch
hardware goes beyond the end of the shader, it may not reach into uninitialized memory (or
unmapped memory pages).

The amount of shader padding required is related to how far the shader hardware may prefetch
ahead. The shader can be set to prefetch 1, 2 or 3 cachelines (64 bytes) ahead of the current
program counter. This is controlled via a wave-launch state register, or by the shader program
itself with S_INST_PREFETCH.

2.5. Shader Padding Requirement 11 of 283

"RDNA 2" Instruction Set Architecture

Chapter 3. Kernel State

This chapter describes the kernel states visible to the shader program.

3.1. State Overview

The table below shows the hardware states readable or writable by a shader program. All
registers below are unique to each wave except for TBA and TMA which are shared.

Table 2. Readable and Writable Hardware States

Abbrev. Name Size Description
(bits)

PC Program Counter 48 Points to the memory address of the next shader

instruction to execute.

V0-V255 VGPR 32 Vector general-purpose register.

S0-S105 SGPR 32 Scalar general-purpose register.

LDS Local Data Share 64kB Local data share is a scratch RAM with built-in
arithmetic capabilities that allow data to be shared
between threads in a workgroup.

EXEC Execute Mask 64 A bit mask with one bit per thread, which is applied to
vector instructions and controls that threads execute
and that ignore the instruction.

EXECZ EXEC is zero 1 A single bit flag indicating that the EXEC mask is all
zeros.

VCC Vector Condition Code 64 A bit mask with one bit per thread; it holds the result
of a vector compare operation.

VCCZ VCC is zero 1 A single bit-flag indicating that the VCC mask is all
zeros.

SCC Scalar Condition Code 1 Result from a scalar ALU comparison instruction.

FLAT_SCRATCH Flat scratch address 64 The base address of scratch memory.

STATUS Status 32 Read-only shader status bits.

MODE Mode 32 Writable shader mode bits.

M0 Memory Reg 32 A temporary register that has various uses, including

GPR indexing and bounds checking.

TRAPSTS Trap Status 32 Holds information about exceptions and pending

traps.

TBA Trap Base Address 64 Holds the pointer to the current trap handler program.

TMA Trap Memory Address 64 Temporary register for shader operations. For
example, can hold a pointer to memory used by the
trap handler.

3.1. State Overview 12 of 283

"RDNA 2" Instruction Set Architecture

Abbrev. Name Size Description

(bits)

TTMP0-TTMP15 Trap Temporary SGPRs 32 16 SGPRs available only to the Trap Handler for
temporary storage.

VMCNT Vector memory instruction 6 Counts the number of VMEM load instructions issued
count but not yet completed.

VSCNT Vector memory instruction 6 Counts the number of VMEM store instructions
count issued but not yet completed.

EXPCNT Export Count 3 Counts the number of Export and GDS instructions
issued but not yet completed. Also counts VMEM
writes that have not yet sent their write-data to the
last level cache.

LGKMCNT LDS, GDS, Constant and 4 Counts the number of LDS, GDS, constant-fetch
Message count (scalar memory read), and message instructions
issued but not yet completed.

3.2. Program Counter (PC)

The program counter (PC) is a byte address pointing to the next instruction to execute. When a
wavefront is created, the PC is initialized to the first instruction in the program.

The PC interacts with three instructions: S_GET_PC, S_SET_PC, S_SWAP_PC. These transfer
the PC to, and from, an even-aligned SGPR pair.

Branches jump to (PC_of_the_instruction_after_the_branch + offset). The shader program

cannot directly read from, or write to, the PC. Branches, GET_PC and SWAP_PC, are PC-
relative to the next instruction, not the current one. S_TRAP saves the PC of the S_TRAP
instruction itself.

3.3. EXECute Mask

The Execute mask (64-bit) determines which threads in the vector are executed:
1 = execute, 0 = do not execute.

EXEC can be read from, and written to, through scalar instructions; it also can be written as a
result of a vector-ALU compare (V_CMPX). This mask affects vector-ALU, vector-memory, LDS,
GDS, and export instructions. It does not affect scalar (ALU or memory) execution or branches.

A helper bit (EXECZ) can be used as a condition for branches to skip code when EXEC is zero.

Wave32: the upper 32-bit of EXEC are ignored, and EXECZ represents the status of only the
lower 32-bits of EXEC.

3.2. Program Counter (PC) 13 of 283

"RDNA 2" Instruction Set Architecture

This GPU can optimize instruction execution when EXEC = 0.

The shader hardware can skip vector ALU and memory instructions if EXEC is known to
be zero, but with some limitations:

• VALU instructions can be skipped, unless they write SGPRs (these are not skipped)
• Wave64 memory instructions: can skip one half but not the entire instruction
• Wave32 memory instructions: not skipped

Use CBRANCH to rapidly skip over code when it is likely that the EXEC mask is zero.

3.4. Status registers

Status register fields can be read, but not written to, by the shader. These bits are initialized at
wavefront-creation time. The table below lists and briefly describes the status register fields.

Table 3. Status Register Fields

Field Bit Description
Position

SCC 1 Scalar condition code. Used as a carry-out bit. For a comparison instruction,
this bit indicates failure or success. For logical operations, this is 1 if the
result was non-zero.

SPI_PRIO 2:1 Wavefront priority set by the shader processor interpolator (SPI) when the
wavefront is created. See the S_SETPRIO instruction (page 12-49) for
details. 0 is lowest, 3 is highest priority.

USER_PRIO 4:3 User settable wave-priority set by the shader program. See the
S_SETPRIO instruction (page 12-49) for details.

PRIV 5 Privileged mode. Can only be active when in the trap handler. Gives write
access to the TTMP, TMA, and TBA registers.

TRAP_EN 6 Indicates that a trap handler is present. When set to zero, traps are not
taken.

TTRACE_EN 7 Indicates whether thread trace is enabled for this wavefront. If zero, also
ignore any shader-generated (instruction) thread-trace data.

EXPORT_RDY 8 This status bit indicates if export buffer space has been allocated. The
shader stalls any export instruction until this bit becomes 1. It is set to 1
when export buffer space has been allocated. Before a Pixel or Vertex
shader can export, the hardware checks the state of this bit. If the bit is 1,
export can be issued. If the bit is zero, the wavefront sleeps until space
becomes available in the export buffer. Then, this bit is set to 1, and the
wavefront resumes.

EXECZ 9 Exec mask is zero.

VCCZ 10 Vector condition code is zero.

IN_WG 11 Wavefront is a member of a work-group of more than one wavefront.

3.4. Status registers 14 of 283

"RDNA 2" Instruction Set Architecture

Field Bit Description

Position

IN_BARRIER 12 Wavefront is waiting at a barrier.

HALT 13 Wavefront is halted or scheduled to halt. HALT can be set by the host
through wavefront-control messages, or by the shader. This bit is ignored
while in the trap handler (PRIV = 1); it also is ignored if a host-initiated trap
is received (request to enter the trap handler).

TRAP 14 Wavefront is flagged to enter the trap handler as soon as possible.

TTRACE_SIMD_EN 15 Enables/disables thread trace for this SIMD. This bit allows more than one
SIMD to be outputting USERDATA (shader initiated writes to the thread-
trace buffer). Note that wavefront data is only traced from one SIMD per
shader engine. Wavefront user data (instruction based) can still be output if
this bit is zero.

VALID 16 Wavefront is active (has been created and not yet ended).

ECC_ERR 17 An ECC error has occurred.

SKIP_EXPORT 18 For Vertex Shaders only. 1 = this shader is not allocated export buffer
space; all export instructions are ignored (treated as NOPs). Formerly
called VS_NO_ALLOC. Used for stream-out of multiple streams (multiple
passes over the same VS), and for DS running in the VS stage for
wavefronts that produced no primitives.

PERF_EN 19 Performance counters are enabled for this wavefront.

COND_DBG_USER 20 Conditional debug indicator for user mode

COND_DBG_SYS 21 Conditional debug indicator for system mode.

FATAL_HALT 23 Set if the wave experienced a fatal error.

MUST_EXPORT 27 This wavefront is required to perform an export with Done=1 before

terminating.

3.5. Mode register

Mode register fields can be read from, and written to, by the shader through scalar instructions.
The table below lists and briefly describes the mode register fields.

Table 4. Mode Register Fields

Field Bit Description
Position

FP_ROUND 3:0 [1:0] Single precision round mode. [3:2] Double/Half-precision round mode.
Round Modes: 0=nearest even, 1= +infinity, 2= -infinity, 3= toward zero.

FP_DENORM 7:4 [1:0] Single denormal mode. [3:2] Double/Half-precision denormal mode.
Denorm modes:
0 = flush input and output denorms.
1 = allow input denorms, flush output denorms.
2 = flush input denorms, allow output denorms.
3 = allow input and output denorms.

3.5. Mode register 15 of 283

"RDNA 2" Instruction Set Architecture

Field Bit Description

Position

DX10_CLAMP 8 Used by the vector ALU to force DX10-style treatment of NaNs: when set,
clamp NaN to zero; otherwise, pass NaN through.

IEEE 9 Floating point opcodes that support exception flag gathering quiet and
propagate signaling NaN inputs per IEEE 754-2008. Min_dx10 and max_dx10
become IEEE 754-2008 compliant due to signaling NaN propagation and
quieting.

LOD_CLAMPED 10 Sticky bit indicating that one or more texture accesses had their LOD
clamped.

DEBUG 11 Forces the wavefront to jump to the exception handler after each instruction is
executed (but not after ENDPGM). Only works if TRAP_EN = 1.

EXCP_EN 20:12 Enable mask for exceptions. Enabled means if the exception occurs and
TRAP_EN==1, a trap is taken.
[12] : invalid.
[13] : inputDenormal.
[14] : float_div0.
[15] : overflow.
[16] : underflow.
[17] : inexact.
[18] : int_div0.
[19] : address watch

FP16_OVFL 23 If set, an overflowed FP16 result is clamped to +/- MAX_FP16, regardless of

round mode, while still preserving true INF values.

DISABLE_PERF 27 1 = disable performance counting for this wave

3.6. GPRs and LDS

This section describes how GPR and LDS space is allocated to a wavefront, as well as how out-
of-range and misaligned accesses are handled.

3.6.1. Out-of-Range behavior

This section defines the behavior when a source or destination GPR or memory address is
outside the legal range for a wavefront.

Out-of-range can occur through GPR-indexing or bad programming. It is illegal to index from
one register type into another (for example: SGPRs into trap registers or inline constants). It is
also illegal to index within inline constants.

The following describe the out-of-range behavior for various storage types.

• SGPRs
◦ SGPRs cannot be "out of range".

3.6. GPRs and LDS 16 of 283

"RDNA 2" Instruction Set Architecture

However, it is illegal to index from one range to another, or for a 64-bit operand to
straddle two ranges.
The ranges are: [ SGPRs 0-105 and VCCH, VCCL], [ Trap Temps 0-15 ], [ all other
values ]
• VGPRs
◦ It is illegal to index from SGPRs into VGPRs, or vice versa.
◦ Out-of-range = (vgpr < 0 || (vgpr >= vgpr_size))
◦ If a source VGPR is out of range, VGPR0 is used.
◦ If a destination VGPR is out-of-range, the instruction is ignored and nothing is written
(treated as an NOP).
• LDS
◦ If the LDS-ADDRESS is out-of-range (addr < 0 or > (MIN(lds_size, m0)):
▪ Writes out-of-range are discarded; it is undefined if SIZE is not a multiple of write-
data-size.
▪ Reads return the value zero.
◦ If any source-VGPR is out-of-range, the VGPR0 value is used.
◦ If the dest-VGPR is out of range, nullify the instruction (issue with exec=0)
• Memory, LDS, and GDS: Reads and atomics with returns.
◦ If any source VGPR or SGPR is out-of-range, the data value is undefined.
◦ If any destination VGPR is out-of-range, the operation is nullified by issuing the
instruction as if the EXEC mask were cleared to 0.
▪ This out-of-range check must check all VGPRs that can be returned (for example:
VDST to VDST+3 for a BUFFER_LOAD_DWORDx4).
▪ This check must also include the extra PRT (partially resident texture) VGPR and
nullify the fetch if this VGPR is out-of-range, no matter whether the texture system
actually returns this value or not.
▪ Atomic operations with out-of-range destination VGPRs are nullified: issued, but
with exec mask of zero.

Instructions with multiple destinations (for example: V_ADDC): if any destination is out-of-range,
no results are written.

3.6.2. SGPR Allocation and storage

Every wavefront is allocated a fixed number of SGPRs:

• 106 normal SGPRs

• VCCh and VCCl (stored in SGPRs 106 and 107)
• 16 Trap-temporary SGPRs, meant for use by the trap handler

3.6. GPRs and LDS 17 of 283

"RDNA 2" Instruction Set Architecture

3.6.3. SGPR Alignment

Even-aligned SGPRs are required in the following cases.

• When 64-bit data is used. This is required for moves to/from 64-bit registers, including the
PC.
• When scalar memory reads that the address-base comes from an SGPR-pair (either in
SGPR).

Quad-alignment is required for the data-GPR when a scalar memory read returns four or more
Dwords. When a 64-bit quantity is stored in SGPRs, the LSBs are in SGPR[n], and the MSBs
are in SGPR[n+1].

3.6.4. VGPR Allocation and Alignment

VGPRs are allocated in groups of 8 Dwords for wave64, and 16 Dwords for wave32. Operations
using pairs of VGPRs (for example: double-floats) have no alignment restrictions. Physically,
allocations of VGPRs can wrap around the VGPR memory pool.

3.6.5. Wave Shared VGPRs

Wave64’s can be allocated wave-private and wave-shared VGPRs. Private GPRs are the
normal ones where each lane has a unique value. Shared VGPRS are shared between the high
and low halves of a wave64. This can be useful to reduce overall VGPR usage when combined
with subvector execution. Shared VGPRs are allocated in blocks of 16 Dwords.

Shared VGPRs logically occupy the VGPR addresses immediately following the private VGPRs.
E.g. if a wave has 8 private VGPRs, they are V0-V7 and shared VGPRs start at V8. If there are
16 shared VGPRs, they are accessed as V8-23.

Shared VGPRs cannot be used for: Exports or GDS.

3.6.6. LDS Allocation and Clamping

LDS is allocated per work-group or per-wavefront when work-groups are not in use. LDS space
is allocated to a work-group or wavefront in contiguous blocks of 256 Dwords on 256-Dword
alignment. LDS allocations do not wrap around the LDS storage. All accesses to LDS are
restricted to the space allocated to that wavefront/work-group.

3.7. M# Memory Descriptor

There is one 32-bit M# (M0) register per wavefront, which can be used for:

3.7. M# Memory Descriptor 18 of 283

"RDNA 2" Instruction Set Architecture

• Local Data Share (LDS)

◦ Interpolation: holds { 1’b0, new_prim_mask[15:1], parameter_offset[15:0] } // in bytes
◦ LDS direct-read offset and data type: { 13’b0, DataType[2:0], LDS_address[15:0] } //
addr in bytes
◦ LDS “add_TID” read/write: { 16’h0, lds_offset[15:0] } // offset in bytes
• Global Data Share (GDS)
◦ { base[15:0] , size[15:0] } // base and size are in bytes
• Indirect GPR addressing for both vector and scalar instructions. M0 is an unsigned index.
• Send-message value. EMIT/CUT use M0 and EXEC as the send-message data.
• Index value used by S_MOVREL and V_MOVREL

3.8. SCC: Scalar Condition code

Most scalar ALU instructions set the Scalar Condition Code (SCC) bit, indicating the result of the
operation.

Compare operations: 1 = true

Arithmetic operations: 1 = carry out
Bit/logical operations: 1 = result was not zero
Move: does not alter SCC

The SCC can be used as the carry-in for extended-precision integer arithmetic, as well as the
selector for conditional moves and branches.

3.9. Vector Compares: VCC and VCCZ

Vector ALU comparisons set the Vector Condition Code (VCC) register (1=pass, 0=fail). Also,
vector compares have the option of setting EXEC to the VCC value.

There is also a VCC summary bit (vccz) that is set to 1 when the VCC result is zero. This is
useful for early-exit branch tests. VCC is also set for selected integer ALU operations (carry-
out).

Vector compares have the option of writing the result to VCC (32-bit instruction encoding) or to
any SGPR (64-bit instruction encoding). VCCZ is updated every time VCC is updated: vector
compares and scalar writes to VCC.

The EXEC mask determines which threads execute an instruction. The VCC indicates which
executing threads passed the conditional test, or which threads generated a carry-out from an
integer add or subtract.

3.8. SCC: Scalar Condition code 19 of 283

"RDNA 2" Instruction Set Architecture

V_CMP_* ⇒ VCC[n] = EXEC[n] & (test passed for thread[n])

VCC is fully written; there are no partial mask updates.

VCC physically resides in the SGPR register file, so when an instruction


sources VCC, that counts against the limit on the total number of SGPRs that
can be sourced for a given instruction. VCC physically resides in the highest
two user SGPRs.

When used by a wave32, the upper 32 bits of VCC are unused and only the lower 32 bits of
VCC contribute to the value of VCCZ.

3.10. Trap and Exception registers

Each type of exception can be enabled or disabled independently by setting, or clearing, bits in
the TRAPSTS register’s EXCP_EN field. This section describes the registers which control and
report kernel exceptions.

All Trap temporary SGPRs (TTMP*) are privileged for writes - they can be written only when in
the trap handler (status.priv = 1). When not privileged, writes to these are ignored. TMA and
TBA are read-only; they can be accessed through S_GETREG_B32.

When a trap is taken (either user initiated, exception or host initiated), the shader hardware is
designed to generate an S_TRAP instruction. This loads trap information into a pair of SGPRS:

{TTMP1, TTMP0} = {1'h0, pc_rewind[5:0], HT[0],trapID[7:0], PC[47:0]}.

HT is set to one for host initiated traps, and zero for user traps (s_trap) or exceptions. TRAP_ID
is zero for exceptions, or the user/host trapID for those traps. When the trap handler is entered,
the PC of the faulting instruction is: (PC - PC_rewind*4).

STATUS . TRAP_EN - This bit indicates to the shader whether or not a trap handler is present.
When one is not present, traps are not taken, no matter whether they’re floating point, user-, or
host-initiated traps. When the trap handler is present, the wavefront uses an extra 16 SGPRs for
trap processing. If trap_en == 0, all traps and exceptions are ignored, and s_trap is converted
by hardware to NOP.

MODE . EXCP_EN[8:0] - Floating point exception enables. Defines which exceptions and
events cause a trap.

3.10. Trap and Exception registers 20 of 283

"RDNA 2" Instruction Set Architecture

Bit Exception

0 Invalid

1 Input Denormal

2 Divide by zero

3 Overflow

4 Underflow

5 Inexact

6 Integer divide by zero

7 Address Watch - the cache has witnessed a thread access to an

'address of interest'

3.10.1. Trap Status register

The trap status register records previously seen traps or exceptions. It can be read and written
by the kernel.

Table 5. Exception Field Bits

Field Bits Description

EXCP 8:0 Status bits of which exceptions have occurred. These bits are sticky and
accumulate results until the shader program clears them. These bits are
accumulated regardless of the setting of EXCP_EN. These can be read or
written without shader privilege.
Bit Exception
0 invalid
1 Input Denormal
2 Divide by zero
3 overflow
4 underflow
5 inexact
6 integer divide by zero
7 address watch 0
8 memory violation

SAVECTX 10 A bit set by the host command indicating that this wave must jump to its trap
handler and save its context. This bit must be cleared by the trap handler using
S_SETREG. Note - a shader can set this bit to 1 to cause a save-context trap,
and due to hardware latency the shader may execute up to 2 additional
instructions before taking the trap.

ILLEGAL_INST 11 An illegal instruction has been detected.

ADDR_WATCH1-3 14:12 Indicates that address watch 1, 2, or 3 has been hit. Bit 12 is address watch 1;
bit 13 is 2; bit 14 is 3.

BUFFER_OOB 15 A buffer instruction has addresses data which is out of range.

XNACK_ERROR 28 A memory address translation error has occurred.

3.10. Trap and Exception registers 21 of 283

"RDNA 2" Instruction Set Architecture

Field Bits Description

DP_RATE 31:29 Determines how the shader interprets the TRAP_STS.cycle. Different Vector
Shader Processors (VSP) process instructions at different rates.

3.11. Memory Violations

A Memory Violation is reported from:

• LDS alignment error.

• Memory read/write/atomic alignment error.
• Flat access where the address is invalid (does not fall in any aperture).
• Write to a read-only surface.
• GDS alignment or address range error.
• GWS operation aborted (semaphore or barrier not executed).

Memory violations are not reported for instruction or scalar-data accesses.

Memory Buffer to LDS does NOT return a memory violation if the LDS address is out of range,
but masks off EXEC bits of threads that would go out of range.

When a memory access is in violation, the appropriate memory (LDS or cache) returns
MEM_VIOL to the wave. This is stored in the wave’s TRAPSTS.mem_viol bit. This bit is sticky,
so once set to 1, it remains at 1 until the user clears it.

Memory violations are fatal: if a trap handler is present and the wave is not already in the trap
handler, the wave jumps to the trap handler; otherwise it signals an interrupt and halt.

Memory violations are not precise. The violation is reported when the LDS or cache processes
the address; during this time, the wave may have processed many more instructions. When a
mem_viol is reported, the Program Counter saved is that of the next instruction to execute; it
has no relationship the faulting instruction.

3.12. Initial Wave State

When a wave is launched, some of the state data is pre-initialized. This section describes what
state is initialized per shader stage. Note that as usual in this spec, the shader stages refer to
hardware shader stages and these often are not identical to software shader stages.

State initialization is controlled by state registers which are defined in other documentation.

3.11. Memory Violations 22 of 283

"RDNA 2" Instruction Set Architecture

3.12.1. State Registers

Table 6. State Register Initialization
Program Counter (PC) program start – from SPI_SHADER_PGM_LO/HI

Execute mask (EXEC) workitem valid mask. Indicates which workitems are valid for this wavefront.
Wave32 uses only bits 31-0. The combined ES+GS, HS+LS loads a dummy
non-zero value into EXEC, and the shader must calculate the real value from
initialized SGPRs.

Trap Status (TRAPSTS) 0

MODE.round/denorm Round and denormal modes are initialized from:

SPI_SHADER_PGM_RSRC1_*.{float,round}_mode

MODE.debug/dx_clamp Similar for "debug" and "dx10_clamp".

EXCP_EN Initalized from SPI_SHADER_PGM_RSRC2_*.excp_en

3.12.2. SGPR Initialization

SGPRs are initialized based on various SPI_PGM_RSRC* register setting. It is important to
know that only the enabled values are loaded, and they are packed into consecutive SGPRs.

Pixel Shader (PS)

SGPR Order Description Enable

First 0..32 of User data registers SPI_SHADER_PGM_RSRC2_PS.user_sgpr

then {bc_optimize, prim_mask[14:0] , lds_offset[15:0]} N/A

then {ps_wave_id[9:0], ps_wave_index[2:0]} SPI_SHADER_PGM_RSRC2_PS.wave_cnt_en

then POPS collision wave ID SPI_SHADER_PGM_RSRC2_PS.load_collision

{DidOverlap, 2’b0, Packer ID[0], 4’b0, Newest _waveID
Overlapped WaveID[11:0], WaveID[11:0]}

then {16’b0, Intra-Wave Quad Overlap[15:0]} SPI_SHADER_PGM_RSRC2_PS.load_intrawav

e_collisions

then Provoking Vertex SPI_SHADER_PGM_RSRC1_PS.load_provokin

g_vtx

then Scratch offset, in bytes SPI_SHADER_PGM_RSRC2_PS.scratch_en

Vertex Shader (VS)

SGPR Order Description Enable

First 0.. 32 of User data registers SPI_SHADER_PGM_RSRC2_VS.user_sgpr

3.12. Initial Wave State 23 of 283

"RDNA 2" Instruction Set Architecture

SGPR Order Description Enable

then {stream_id[1:0], is_offchip, SPI_SHADER_PGM_RSRC2_VS.so_en ||

streamout_vtx_count[6:0] , SPI_SHADER_PGM_RSRC2_VS.oc_lds_en
streamout_enable[15:0]}

then streamout_write_index[31:0] SPI_SHADER_PGM_RSRC2_VS.so_en

then streamout_offset0[31:0] SPI_SHADER_PGM_RSRC2_VS.so_base0_en

then streamout_offset1[31:0] SPI_SHADER_PGM_RSRC2_VS.so_base1_en

then streamout_offset2[31:0] SPI_SHADER_PGM_RSRC2_VS.so_base2_en

then streamout_offset3[31:0] SPI_SHADER_PGM_RSRC2_VS.so_base3_en

then offchip_lds_pointer[31:0] in bytes SPI_SHADER_PGM_RSRC2_VS.oc_lds_en

then waveID (dispatch draw term) SPI_SHADER_PGM_RSRC2_VS.dispatch_draw

_en

then pc_base SPI_SHADER_PGM_RSRC2_VS.pc_base_en

then Scratch offset (in bytes) SPI_SHADER_PGM_RSRC2_VS.scratch_en

Geometry Shader (ES+GS)

SGPR Order Description Enable

0 GS User Data Address Low ([31:0]) automatically enabled

comes from:
SPI_SHADER_USER_DATA_LO_GS

1 GS User Data Address High ([47:32]) automatically enabled

comes from:
SPI_SHADER_USER_DATA_HI_GS

2 GS2VS Ring buffer offset[31:0] (byte) OR VGT_SHADER_STAGES.primgen_en

Control SB offset | Ordered Wave ID
{wave_crawler_inc[2:0], 16’h0,
ordered_wave_id[12:0]}

3 { TGsize[3:0], WaveInSubgroup[3:0], automatically enabled

GSWaveID[7:0], GSPrimCount[7:0],
ESVertCount[7:0]}

4 Off-chip LDS base [31:0] SPI_SHADER_PGM_RSRC2_GS.oc_lds_en

5 Shared Scratch Offset SPI_SHADER_PGM_RSRC2_GS.scratch_en

6 GS Shader address low comes from: automatically enabled

SPI_SHADER_PGM_LO_GS

7 GS Shader address high comes from: automatically enabled

SPI_SHADER_PGM_HI_GS

then 0.. 32 of User data registers of GS shader SPI_SHADER_PGM_RSRC2_GS.user_sgpr

3.12. Initial Wave State 24 of 283

"RDNA 2" Instruction Set Architecture

Front End Shader (LS+HS)

SGPR Order Description Enable

0 HS User Data Address Low ([31:0]) SPI_SHADER_USER_DATA_LO_HS

1 HS User Data Address High ([47:32]) SPI_SHADER_USER_DATA_HI_HS

2 Off-chip LDS base [31:0] automatically enabled

3 {first_wave, lshs_TGsize[6:0], automatically enabled

lshs_PatchCount[7:0], HS_vertCount[7:0],
LS_vertCount[7:0]}

4 TF buffer base [15:0] automatically enabled

5 Shared Scratch Offset SPI_SHADER_PGM_RSRC2_HS.scratch_en

6 HS Shader address low SPI_SHADER_PGM_LO_HS

7 HS Shader address high SPI_SHADER_PGM_HI_HS

then 0.. 32 of User data registers of HS shader SPI_SHADER_PGM_RSRC2_HS.user_sgpr

Compute Shader (CS)

SGPR Order Description Enable

First 0.. 16 of User data registers COMPUTE_PGM_RSRC2.user_sgpr

then threadgroup_id0[31:0] COMPUTE_PGM_RSRC2.tgid_x_en

then threadgroup_id1[31:0] COMPUTE_PGM_RSRC2.tgid_y_en

then threadgroup_id2[31:0] COMPUTE_PGM_RSRC2.tgid_z_en

then {first_wave, 6’h00, wave_id_in_group[4:0], 2’h0, COMPUTE_PGM_RSRC2.tg_size_en

ordered_append_term[11:0],
threadgroup_size_in_waves[5:0]}

then Scratch offset (in bytes) COMPUTE_PGM_RSRC2.scratch_en

Compute shaders have up to 3 VGPRs initialized as well:

• VGPR0 = thread ID in group, X dimension

• VGPR1 = thread ID in group, Y dimension
• VGPR2 = thread ID in group, Z dimension

3.12. Initial Wave State 25 of 283

"RDNA 2" Instruction Set Architecture

Chapter 4. Program Flow Control

All program flow control is programmed using scalar ALU instructions. This includes loops,
branches, subroutine calls, and traps. The program uses SGPRs to store branch conditions and
loop counters. Constants can be fetched from the scalar constant cache directly into SGPRs.

4.1. Program Control

The instructions in the table below control the priority and termination of a shader program, as
well as provide support for trap handlers.

Table 7. Control Instructions

Instructions Description

S_ENDPGM Terminates the wavefront. It can appear anywhere in the kernel and can appear multiple
times.

S_ENDPGM_SAVED Terminates the wavefront due to context save. It can appear anywhere in the kernel and can
appear multiple times.

S_NOP Does nothing; it can be repeated in hardware up to eight times.

S_TRAP Jumps to the trap handler.

S_RFE Returns from the trap handler

S_SETPRIO Modifies the priority of this wavefront: 0=lowest, 3 = highest.

S_SLEEP Causes the wavefront to sleep for 64 - 960 clock cycles.

S_SENDMSG Sends a message (typically an interrupt) to the host CPU.

S_CLAUSE Define a clause of instructions which are executed together.

S_VERSION Does nothing (treated as S_NOP), but can be used as a code comment to indicate the
hardware version the shader is compiled for (using the SIMM16 field).

S_CODE_END Treated as an illegal instruction. Used to pad past the end of shaders.

4.1.1. Instruction Clauses

An instruction clause is a group of instructions of the same type which are to be executed in
an uninterruped sequence. Normally the shader hardware may interleave instructions from
different waves in order to maintain performance, but a clause can be used to override that
behavior and force the hardware to service only one wave for a given instruction type for the
duration of the clause.

Clauses are defined by the S_CLAUSE instructions, which specifies the number of instructions
that make up the clause. The clause-type is implicitly defined by the type of instruction
immediately following the clause. Clause types are:

4.1. Program Control 26 of 283

"RDNA 2" Instruction Set Architecture

• VALU
• SMEM
• LDS
• FLAT
• Texture, buffer, global and scratch

Clauses must contain only one instruction type.

4.2. Branching
Branching is done using one of the following scalar ALU instructions.

Table 8. Branch Instructions

Instructions Description

S_BRANCH Unconditional branch.

S_CBRANCH_<test> Conditional branch. Branch only if <test> is true. Tests are VCCZ, VCCNZ,
EXECZ, EXECNZ, SCCZ, and SCCNZ.

S_CBRANCH_CDBGSYS Conditional branch, taken if the COND_DBG_SYS status bit is set.

S_CBRANCH_CDBGUSER Conditional branch, taken if the COND_DBG_USER status bit is set.

S_CBRANCH_CDBGSYS_AND_ Conditional branch, taken only if both COND_DBG_SYS and

USER COND_DBG_USER are set.

S_SETPC Directly set the PC from an SGPR pair.

S_SWAPPC Swap the current PC with an address in an SGPR pair.

S_GETPC Retrieve the current PC value (does not cause a branch).

S_CALL_B64 Jump to a subroutine, and save return address. SGPR_pair = PC+4; PC =

PC+4+SIMM16*4.

S_SUBVECTOR_LOOP_BEGIN Starts a subvector execution loop. The SIMM16 field is the branch offset to the
instruction after S_SUBVECTOR_LOOP_END, and the SGPR is used for
temporary EXEC storage.

S_SUBVECTOR_LOOP_END Marks the end of the subvector execution loop. The SIMM16 field points back to
the instruction after S_SUBVECTOR_LOOP_BEGIN, and the SGPR is used for
temporary EXEC storage.

For conditional branches, the branch condition can be determined by either scalar or vector
operations. A scalar compare operation sets the Scalar Condition Code (SCC), which then can
be used as a conditional branch condition. Vector compare operations set the VCC mask, and
VCCZ or VCCNZ then can be used to determine branching.

4.2.1. Subvector Execution

“Subvector execution” is an alternate method of handling wave64 instruction execution. The

4.2. Branching 27 of 283

"RDNA 2" Instruction Set Architecture

normal method is to issue each half of a wave64 as two wave32 instructions, then move on to
the next instruction. This alternative method is to issue a group of instructions, all for the first 32
workitems and then come back and execute the same instructions but for the second 32
workitems. This has two potential advantages:

• Memory operations are for smaller units of work and may cache better
◦ example: reading multiple entries from a strided buffer
• Wave-temporary VGPRs are available:
◦ In Wave64 each wave may declare N normal VGPRs (the wave gets 64 * N dwords,
with N per work-item), and M temp VGPRs which may only be used in this mode. The
temp VGPRs are physically adjacent to the normal ones, but logically are from just after
the private VGPRs. These can be used on each pass of the subvector execution.

This mode is explicitly declared in shader code as a loop:

  S_subvector_loop_begin S0, skip_all

loop_start:
  < vec32 code >
  S_subvector_loop_end S0, loop_start
skip_all:

Table 9. SubVector Execution Order

Shader Normal Execution Sequence Subvector Loop Execution Sequence
Program

inst0 inst0 - low inst0 - low

inst1 inst0 - high inst1 - low
inst2 inst1 - low inst2 - low
inst3 inst1 - high inst3 - low
inst2 - low inst0 - high
inst2 - high inst1 - high
inst3 - low inst2 - high
inst3 – high inst3 – high

Subvector execution is simply a loop construct where half of the EXEC mask is zero for each
pass over the body of the code. All wave64 rules still apply. The loop executes zero, one or two
times, depending on the initial state of the EXEC mask. During each pass of the loop, one half
of EXEC is forced to zero (after being saved in an SGPR). The EXEC mask is restored at the
end of the loop.

If EXECHI = 0: the body is executed only once: EXECLO is stored in S0 and restored at the
end, but it was zero anyway. If EXEC_LO was zero at the start, the same thing happens. If both
halves of EXEC are non-zero, do the low pass first (storing EXECHI in S0), then restore
EXECHI and save off EXECLO and do it again. Restore EXECLO at the end of the second
pass. The “pass #” is encoded by observing which half of EXEC is zero.

Subvector looping imposes a rule that the “body code” cannot let the working half of the exec
mask go to zero. If it might go to zero, it must be saved at the start of the loop and be restored

4.2. Branching 28 of 283

"RDNA 2" Instruction Set Architecture

before the end since the S_SUBVECTOR_LOOP_* instructions determine which pass they’re in
by looking at which half of EXEC is zero.

4.3. Workgroups
Work-groups are collections of wavefronts running on the same workgroup processor which can
synchronize and share data. Up to 1024 work-items (16 wave64’s or 32 wave32’s) can be
combined into a work-group. When multiple wavefronts are in a workgroup, the S_BARRIER
instruction can be used to force each wavefront to wait until all other wavefronts reach the same
instruction; then, all wavefronts continue. Any wavefront may terminate early using S_ENDPGM,
and the barrier is considered satisfied when the remaining live waves reach their barrier
instruction.

4.4. Data Dependency Resolution

Shader hardware can resolve most data dependencies, but a few cases must be explicitly
handled by the shader program. In these cases, the program must insert S_WAITCNT
instructions to ensure that previous operations have completed before continuing.

The shader has four counters that track the progress of issued instructions. S_WAITCNT waits
for the values of these counters to be at, or below, specified values before continuing.

These allow the shader writer to schedule long-latency instructions, execute unrelated work,
and specify when results of long-latency operations are needed.

Instructions of a given type return in order, but instructions of different types can complete out-
of-order. For example, both GDS and LDS instructions use LGKM_cnt, but they can return out-
of-order. VMEM loads update VM_CNT in the order the instructions were issued, so waiting on
VM_CNT to be less-than a particular value ensures all previous loads have completed. It is
possible for data to be written to VGPRs out-of-order. Stores from a wave are not kept in order
with stores from that wave.

VM_CNT
Vector memory count (reads, atomic with return). Determines when memory reads have
finished.

• Incremented every time a vector-memory read or atomic-with-return (MIMG, MUBUF,

MTBUF, or FLAT/Scratch/Global format) instruction is issued.
• Decremented for reads when all of the data for that instruction and those before it have
completed.

VS_CNT
Vector memory store count (writes, atomic without return). Determines when memory writes
have completed.

4.3. Workgroups 29 of 283

"RDNA 2" Instruction Set Architecture

• Incremented every time a vector-memory write or atomic-without-return (MIMG, MUBUF,

MTBUF, or Flat/Scratch/Global format) instruction is issued.
• Decremented for writes when the data has been written to the L2 cache.

LGKM_CNT
(LDS, GDS, (K)constant, (M)essage) Determines when one of these low-latency instructions
have completed.

• Incremented by 1 for every LDS or GDS instruction issued, as well as by Dword-count

for scalar-memory reads. For example.
• Decremented by 1 for LDS/GDS reads or atomic-with-return when the data has been
returned to VGPRs.
• Incremented by 1 for each S_SENDMSG issued. Decremented by 1 when message is
sent out.
• Decremented by 1 for LDS/GDS writes when the data has been written to LDS/GDS.
• Decremented by 1 for each Dword returned from the data-cache (SMEM).
Ordering:
◦ Instructions of different types are returned out-of-order.
◦ Instructions of the same type are returned in the order they were issued, except
scalar-memory-reads, which can return out-of-order (in which case only
S_WAITCNT 0 is the only legitimate value).

EXP_CNT
VGPR-export count. Determines when data has been read out of the VGPR and sent to
GDS, at which time it is safe to overwrite the contents of that VGPR.

• Incremented when an Export/GDS instruction is issued from the wavefront buffer.

• Decremented for exports/GDS when the last cycle of the export instruction is granted
and executed (VGPRs read out). Ordering
◦ Exports are kept in order only within each export type (color/null, position,
parameter cache).

4.5. Manually Inserted Wait States (NOPs)

Inserting S_NOP is not required to achieve correct operation.

4.5. Manually Inserted Wait States (NOPs) 30 of 283

"RDNA 2" Instruction Set Architecture

Chapter 5. Scalar ALU Operations

Scalar ALU (SALU) instructions operate on a single value per wavefront. These operations
consist of 32-bit integer arithmetic and 32- or 64-bit bit-wise operations. The SALU also can
perform operations directly on the Program Counter, allowing the program to create a call stack
in SGPRs. Many operations also set the Scalar Condition Code bit (SCC) to indicate the result
of a comparison, a carry-out, or whether the instruction result was zero.

5.1. SALU Instruction Formats

SALU instructions are encoded in one of five microcode formats, shown below:

Figure 3. Scalar ALU format with one source operand

Figure 4. Scalar ALU format with two source operands

Figure 5. Scalar ALU format with one immediate value source operands

Figure 6. Scalar ALU format for compares, with two sources but no destinaton

Figure 7. Scalar ALU format for program flow operations

Each of these instruction formats uses some of these fields:

Field Description

OP Opcode: instruction to be executed.

SDST Destination SGPR.

SSRC0 First source operand.

SSRC1 Second source operand.

SIMM16 Signed immediate 16-bit integer constant.

The lists of similar instructions sometimes use a condensed form using curly braces { } to
express a list of possible names. For example, S_AND_{B32, B64} defines two legal
instructions: S_AND_B32 and S_AND_B64.

5.1. SALU Instruction Formats 31 of 283

"RDNA 2" Instruction Set Architecture

5.2. Scalar ALU Operands

Valid operands of SALU instructions are:

• SGPRs, including trap temporary SGPRs.

• Mode register.
• Status register (read-only).
• M0 register.
• TrapSts register.
• EXEC mask.
• VCC mask.
• SCC.
• Inline constants: integers from -16 to 64, and a some floating point values.
• VCCZ, EXECZ, and SCC.
• Hardware registers.
• 32-bit literal constant.

In the table below, 0-127 can be used as scalar sources or destinations; 128-255 can only be
used as sources.

Table 10. Scalar Operands

Code Meaning Description

Scalar 0 - 105 SGPR 0 to 105 Scalar GPRs

Dest
106 VCC_LO Holds the low Dword of the vector condition code
(7 bits)
107 VCC_HI Holds the high Dword of the vector condition code

108-123 TTMP0 to TTMP15 Trap temps (privileged)

124 M0 Misc register

125 NULL Reads return zero, writes are discarded.

126 EXEC_LO Execute mask, low Dword

127 EXEC_HI Execute mask, high Dword

128 0 zero

129-192 int 1 to 64 Positive integer values.

193-208 int -1 to -16 Negative integer values.

209-234 reserved Unused.

235 SHARED_BASE Memory Aperture definition. Values are affected by

system addressing mode: 32 or 64 bit.
236 SHARED_LIMIT

237 PRIVATE_BASE

238 PRIVATE_LIMIT

239 POPS_EXITING_WAVE_ID Primitive Ordered Pixel Shading wave ID.

5.2. Scalar ALU Operands 32 of 283

"RDNA 2" Instruction Set Architecture

Code Meaning Description

240 0.5 single or double floats

241 -0.5

242 1.0

243 -1.0

244 2.0

245 -2.0

246 4.0

247 -4.0

248 1.0 / (2 * PI)

249-250 reserved unused

251 VCCZ { zeros, VCCZ }

252 EXECZ { zeros, EXECZ }

253 SCC { zeros, SCC }

254 reserved unused

255 Literal constant 32-bit constant from instruction stream.

The SALU cannot use VGPRs or LDS. SALU instructions can use a 32-bit literal constant. This
constant is part of the instruction stream and is available to all SALU microcode formats except
SOPP and SOPK. Literal constants are used by setting the source instruction field to "literal"
(255), and then the following instruction dword is used as the source value.

If the destination SGPR is out-of-range, no SGPR is written with the result. However, SCC and
possibly EXEC (if saveexec) is still written.

If an instruction uses 64-bit data in SGPRs, the SGPR pair must be aligned to an even
boundary. For example, it is legal to use SGPRs 2 and 3 or 8 and 9 (but not 11 and 12) to
represent 64-bit data.

5.3. Scalar Condition Code (SCC)

The scalar condition code (SCC) is written as a result of executing most SALU instructions.

The SCC is set by many instructions:

• Compare operations: 1 = true.

• Arithmetic operations: 1 = carry out.
◦ SCC = overflow for signed add and subtract operations. For add, overflow = both
operands are of the same sign, and the MSB (sign bit) of the result is different than the
sign of the operands. For subtract (AB), overflow = A and B have opposite signs and

5.3. Scalar Condition Code (SCC) 33 of 283

"RDNA 2" Instruction Set Architecture

the resulting sign is not the same as the sign of A.

• Bit/logical operations: 1 = result was not zero.

5.4. Integer Arithmetic Instructions

This section describes the arithmetic operations supplied by the SALU. The table below shows
the scalar integer arithmetic instructions:

Table 11. Integer Arithmetic Instructions

Instruction Encoding Sets SCC? Operation

S_ADD_I32 SOP2 y D = S0 + S1, SCC = overflow.

S_ADD_U32 SOP2 y D = S0 + S1, SCC = carry out.

S_ADDC_U32 SOP2 y D = S0 + S1 + SCC = overflow.

S_SUB_I32 SOP2 y D = S0 - S1, SCC = overflow.

S_SUB_U32 SOP2 y D = S0 - S1, SCC = carry out.

S_SUBB_U32 SOP2 y D = S0 - S1 - SCC = carry out.

S_ABSDIFF_I32 SOP2 y D = abs (S0 - S1), SCC = result not zero.

S_MIN_I32 SOP2 y D = (S0 < S1) ? S0 : S1. SCC = 1 if S0 was min.

S_MIN_U32

S_MAX_I32 SOP2 y D = (S0 > S1) ? S0 : S1. SCC = 1 if S0 was max.

S_MAX_U32

S_MUL_I32 SOP2 n D = S0 * S1. Low 32 bits of result.

S_ADDK_I32 SOPK y D = D + simm16, SCC = overflow. Sign extended

version of simm16.

S_MULK_I32 SOPK n D = D * simm16. Return low 32bits. Sign extended

version of simm16.

S_ABS_I32 SOP1 y D.i = abs (S0.i). SCC=result not zero.

S_SEXT_I32_I8 SOP1 n D = { 24{S0[7]}, S0[7:0] }.

S_SEXT_I32_I16 SOP1 n D = { 16{S0[15]}, S0[15:0] }.

5.5. Conditional Instructions

Conditional instructions use the SCC flag to determine whether to perform the operation, or (for
CSELECT) which source operand to use.

Table 12. Conditional Instructions

Instruction Encoding Sets SCC? Operation

S_CSELECT_{B32, B64} SOP2 n D = SCC ? S0 : S1.

5.4. Integer Arithmetic Instructions 34 of 283

"RDNA 2" Instruction Set Architecture

Instruction Encoding Sets SCC? Operation

S_CMOVK_I32 SOPK n if (SCC) D = signext(simm16).

S_CMOV_{B32,B64} SOP1 n if (SCC) D = S0, else NOP.

5.6. Comparison Instructions

These instructions compare two values and set the SCC to 1 if the comparison yielded a TRUE
result.

Table 13. Conditional Instructions

Instruction Encoding Sets Operation
SCC?

S_CMP_EQ_U64, SOPC y Compare two 64-bit source values. SCC = S0 <cond>

S_CMP_LG_U64 S1.

S_CMP_{EQ,LG,GT,GE,LE,LT}_ SOPC y Compare two source values. SCC = S0 <cond> S1.

{I32,U32}

S_CMPK_{EQ,LG,GT,GE,LE,LT} SOPK y Compare Dest SGPR to a constant. SCC = DST

_{I32,U32} <cond> simm16. simm16 is zero-extended (U32) or
sign-extended (I32).

S_BITCMP0_{B32,B64} SOPC y Test for "is a bit zero". SCC = !S0[S1].

S_BITCMP1_{B32,B64} SOPC y Test for "is a bit one". SCC = S0[S1].

5.7. Bit-Wise Instructions

Bit-wise instructions operate on 32- or 64-bit data without interpreting it has having a type. For
bit-wise operations if noted in the table below, SCC is set if the result is nonzero.

Table 14. Bit-Wise Instructions

Instruction Encoding Sets Operation
SCC?

S_MOV_{B32,B64} SOP1 n D = S0

S_MOVK_I32 SOPK n D = signext(simm16)

{S_AND,S_OR,S_XOR}_{B32,B64} SOP2 y D = S0 & S1, S0 OR S1, S0 XOR S1

{S_ANDN2,S_ORN2}_{B32,B64} SOP2 y D = S0 & ~S1, S0 OR ~S1, S0 XOR ~S1,

{S_NAND,S_NOR,S_XNOR}_{B32,B64} SOP2 y D = ~(S0 & S1), ~(S0 OR S1), ~(S0 XOR S1)

S_LSHL_{B32,B64} SOP2 y D = S0 << S1[4:0], [5:0] for B64.

S_LSHR_{B32,B64} SOP2 y D = S0 >> S1[4:0], [5:0] for B64.

S_ASHR_{I32,I64} SOP2 y D = sext(S0 >> S1[4:0]) ([5:0] for I64).

5.6. Comparison Instructions 35 of 283

"RDNA 2" Instruction Set Architecture

Instruction Encoding Sets Operation

SCC?

S_BFM_{B32,B64} SOP2 n Bit field mask. D = ((1 << S0[4:0]) - 1) << S1[4:0].

S_BFE_U32, S_BFE_U64 SOP2 y Bit Field Extract, then sign-extend result for I32/64
S_BFE_I32, S_BFE_I64 instructions.
(signed/unsigned) S0 = data,
S1[5:0] = offset, S1[22:16]= width.

S_NOT_{B32,B64} SOP1 y D = ~S0.

S_WQM_{B32,B64} SOP1 y D = wholeQuadMode(S0). If any bit in a group of

four is set to 1, set the resulting group of four bits
all to 1.

S_QUADMASK_{B32,B64} SOP1 y D[0] = OR(S0[3:0]), D[1]=OR(S0[7:4]), etc.

S_BITREPLICATE_B64_B32 SOP1 n Replicate each bit in 32-bit S0 twice:

D = { … S0[1], S0[1], S0[0], S0[0] }.
Two of these instructions is the inverse of
S_QUADMASK.

S_BREV_{B32,B64} SOP1 n D = S0[0:31] are reverse bits.

S_BCNT0_I32_{B32,B64} SOP1 y D = CountZeroBits(S0).

S_BCNT1_I32_{B32,B64} SOP1 y D = CountOneBits(S0).

S_FF0_I32_{B32,B64} SOP1 n D = Bit position of first zero in S0 starting from

LSB. -1 if not found.

S_FF1_I32_{B32,B64} SOP1 n D = Bit position of first one in S0 starting from LSB.

-1 if not found.

S_FLBIT_I32_{B32,B64} SOP1 n Find last bit. D = the number of zeros before the
first one starting from the MSB. Returns -1 if none.

S_FLBIT_I32 SOP1 n Count how many bits in a row (from MSB to LSB)
S_FLBIT_I32_I64 are the same as the sign bit. Return -1 if the input
is zero or all 1’s (-1). 32-bit pseudo-code:
if (S0 == 0 || S0 == -1) D = -1
else
D=0
for (I = 31 .. 0)
if (S0[I] == S0[31])
D++
else break
This opcode behaves the same as V_FFBH_I32.

S_BITSET0_{B32,B64} SOP1 n D[S0[4:0], [5:0] for B64] = 0

S_BITSET1_{B32,B64} SOP1 n D[S0[4:0], [5:0] for B64] = 1

S_{and,or,xor,andn1,andn2,orn1,orn2,n SOP1 y Save the EXEC mask, then apply a bit-wise

and, nor,xnor}_SAVEEXEC_{B32,B64} operation to it.
D = EXEC
EXEC = S0 <op> EXEC
SCC = (exec != 0)

5.7. Bit-Wise Instructions 36 of 283

"RDNA 2" Instruction Set Architecture

Instruction Encoding Sets Operation

SCC?

S_{ANDN{1,2}_WREXEC_B{32,64} SOP1 y N1: EXEC, D = ~S0 & EXEC

N2: EXEC, D = S0 & ~EXEC
Both D and EXEC get the same result. SCC =
(result != 0).

S_MOVRELS_{B32,B64} SOP1 n Move a value into an SGPR relative to the value in

S_MOVRELD_{B32,B64} M0.
MOVERELS: D = SGPR[S0+M0]
MOVERELD: SGPR[D+M0] = S0
Index must be even for 64. M0 is an unsigned
index.

5.8. Access Instructions

These instructions access hardware internal registers.

Table 15. Hardware Internal Registers

Instruction Encoding Sets Operation
SCC?

S_GETREG_B32 SOPK* n Read a hardware register into the LSBs of D.

S_SETREG_B32 SOPK* n Write the LSBs of D into a hardware register. (Note that D is a
source SGPR.) Must add an S_NOP between two consecutive
S_SETREG to the same register.

S_SETREG_IMM32_B32 SOPK* n S_SETREG where 32-bit data comes from a literal constant (so
this is a 64-bit instruction format).

S_ROUND_MODE SOPP n Set the round mode from an immediate: simm16[3:0]

S_DENORM_MODE SOPP n Set the denorm mode from an immediate: simm16[3:0]

The hardware register is specified in the DEST field of the instruction, using the values in the
table above. Some bits of the DEST specify which register to read/write, but additional bits
specify which bits in the register to read/write:

SIMM16 = {size[4:0], offset[4:0], hwRegId[5:0]}; offset is 0..31, size is 1..32.

Table 16. Hardware Register Values

Code Register Description

0 reserved

1 MODE R/W.

2 STATUS Read only.

3 TRAPSTS R/W.

5.8. Access Instructions 37 of 283

"RDNA 2" Instruction Set Architecture

Code Register Description

5 GPR_ALLOC Read only. {sgpr_size, sgpr_base, vgpr_size, vgpr_base }.

6 LDS_ALLOC Read only. {lds_size, lds_base}.

7 IB_STS Read only. {valu_cnt, lgkm_cnt, exp_cnt, vm_cnt}.

8 - 14 reserved.

15 SH_MEM_BASES Bits [15:0] = Private Base; [31:16] = Shared Base.

16 TBA_LO Trap base address register [31:0].

17 TBA_HI Trap base address register [47:32].

18 TMA_LO Trap memory address register [31:0].

19 TMA_HI Trap memory address register [47:32].

20 FLAT_SCRATCH_L Flat Scratch memory address, bits [31:0]

O

21 FLAT_SCRATCH_ Flat Scratch memory address, bits [63:32]

HI

25 POPS_PACKER Bit [0] = POPS enabled for this wave; bits [2:1] = Pops Packer ID

29 SHADER_CYCLES Return the value of a 20-bit clock cycle counter. Used for measuring time-delta within a
wave, not between waves.

Table 17. IB_STS

Code Register Description

VM_CNT 23:22, Number of VMEM load instructions issued but not yet returned.
3:0

EXP_CNT 6:4 Number of Exports issued but have not yet read their data from VGPRs.

LGKM_CNT 11:8 LDS, GDS, Constant-memory and Message instructions issued-but-not-completed count.

VS_CNT 31:26 Number of VMEM store instructions issued but not yet returned.

Table 18. GPR_ALLOC

Code Register Description

VGPR_BASE 5:0 Physical address of first VGPR assigned to this wavefront, as [7:2]

VGPR_SIZE 13:8 Number of VGPRs assigned to this wavefront, as [7:2]+4. 0=4 VGPRs, 1=8 VGPRs, etc.

Table 19. LDS_ALLOC

Code Register Description

LDS_BASE 7:0 Physical address of first LDS location assigned to this wavefront, in
units of 64 Dwords.

LDS_SIZE 20:12 Amount of LDS space assigned to this wavefront, in units of 64

Dwords.

5.8. Access Instructions 38 of 283

"RDNA 2" Instruction Set Architecture

Code Register Description

VGPR_SHARED_SIZE 27:24 Number of shared VGPRs allocate to this wave, in units of 8 VGPRs.
(0=0vgprs, 1=8vgprs, …)

5.8. Access Instructions 39 of 283

"RDNA 2" Instruction Set Architecture

Chapter 6. Vector ALU Operations

Vector ALU instructions (VALU) perform an arithmetic or logical operation on data for each of 32
or 64 threads and write results back to VGPRs, SGPRs or the EXEC mask.

Parameter interpolation is a mixed VALU and LDS instruction, and is described in the Data
Share chapter.

6.1. Microcode Encodings

Most VALU instructions are available in two encodings: VOP3 which uses 64-bits of instruction,
and one of three 32-bit encodings that offer a restricted set of capabilities. A few instructions are
only available in the VOP3 encoding.

When an instruction is available in two microcode formats, it is up to the user to decide which to
use. It is recommended to use the 32-bit encoding whenever possible.

The microcode encodings are shown below:

VOP2 is for instructions with two inputs and a single vector destination. Instructions that have a
carry-out implicitly write the carry-out to the VCC register.

VOP1 is for instructions with no inputs or a single input and one destination.

VOPC is for comparison instructions.

VINTRP is for parameter interpolation instructions.

6.1. Microcode Encodings 40 of 283

"RDNA 2" Instruction Set Architecture

VOP3 is for instructions with up to three inputs, input modifiers (negate and absolute value), and
output modifiers. There are two forms of VOP3: one which uses a scalar destination field (used
only for div_scale, integer add and subtract); this is designated VOP3b. All other instructions
use the common form, designated VOP3a.

Any of the 32-bit microcode formats may use a 32-bit literal constant, as well VOP3. Note
however that VOP3 plus a literal makes a 96-bit instruction and excessive use of this
combination may reduce performance.

VOP3P is for instructions that use "packed math": These instructions perform an operation on a
pair of input values that are packed into the high and low 16-bits of each operand; the two 16-bit
results are written to a single VGPR as two packed values.

6.2. Operands
All VALU instructions take at least one input operand (except V_NOP and V_CLREXCP). The
data-size of the operands is explicitly defined in the name of the instruction. For example,
V_FMA_F32 operates on 32-bit floating point data.

Table 20. Instruction Operands

Value Name Description

0-105 SGPR 0 .. 105

106 VCC_LO vcc[31:0].

107 VCC_HI vcc[63:32].

108-123 TTMP0 to TTMP 15 Trap handler temps (privileged).

124 M0 M0 register

125 NULL Reads return zero, writes are discarded.

126 EXEC_LO exec[31:0].

127 EXEC_HI exec[63:32].

128 0 Zero

6.2. Operands 41 of 283

"RDNA 2" Instruction Set Architecture

Value Name Description

129-192 int 1.. 64 Integer inline constants.

193-208 int -1 .. -16

209-232 reserved Unused.

233 DPP8 DPP - 8 lane transfer. (only valid as source-0)

234 DPP8FI DPP - 8 lane transfer with fetch from invalid lanes. (only valid as source-
0)

235 SHARED_BASE Memory Aperture definition.

236 SHARED_LIMIT

237 PRIVATE_BASE

238 PRIVATE_LIMIT

239 POPS_EXITING_WAVE_ID Primitive Ordered Pixel Shading wave ID.

240 0.5 Single, double, or half-precision inline floats.

241 -0.5
1/(2*PI) is 0.15915494.
242 1.0 The exact value used is:
half: 0x3118
243 -1.0 single: 0x3e22f983
244 2.0 double: 0x3fc45f306dc9c882

245 -2.0

246 4.0

247 -4.0

248 1/(2*PI)

249 SDWA Sub Dword Address (only valid as Source-0)

250 DPP16 DPP over 16 lanes (only valid as Source-0)

251 VCCZ { zeros, VCCZ }

252 EXECZ { zeros, EXECZ }

253 SCC { zeros, SCC }

254 LDS direct Use LDS direct read to supply 32-bit value Vector-alu instructions only.

255 Literal constant 32-bit constant from instruction stream.

256-511 VGPR 0 .. 255

6.2.1. Instruction Inputs

VALU instructions can use any of the following sources for input, subject to restrictions listed
below:

• VGPRs

6.2. Operands 42 of 283

"RDNA 2" Instruction Set Architecture

• SGPRs
• Inline constants - constant selected by a specific VSRC value
• Literal constant - 32-bit value in the instruction stream.
• LDS direct data read
• M0
• EXEC mask

Limitations

• At most two scalar values can be read per instructions, but the values can be used for more
than one operand.
◦ Scalar values include: SGPRs, VCC, EXEC (used as data), and literal constants
◦ Some instructions implicitly read an SGPR (which includes VCC), and this implicit read
counts agains the total supported limit.
▪ These are: Add/sub with carry-in, FMAS and CNDMASK
◦ 64-bit shift instructions can use only a single scalar value, not two
• At most one literal constant can be used
• Inline constants are free, and do not count against these limits
• Only SRC0 can use LDS_DIRECT (see Chapter 10, "Data Share Operations")

Instructions using the VOP3 form and also using floating-point inputs have the option of
applying absolute value (ABS field) or negate (NEG field) to any of the input operands.

Literal Expansion to 64 bits

Literal constants are 32-bits, but they can be used as sources which normally require 64-bit
data. They are expanded to 64 bits following these rules:

• 64 bit float: the lower 32-bit are padded with zero.

• 64-bit unsigned integer: zero extended to 64 bits
• 64-bit signed integer: sign extended to 64 bits

6.2.2. Instruction Outputs

VALU instructions typically write their results to VGPRs specified in the VDST field of the
microcode word. A thread only writes a result if the associated bit in the EXEC mask is set to 1.

All V_CMPX instructions write the result of their comparison (one bit per thread) the EXEC
mask.

Instructions producing a carry-out (integer add and subtract) write their result to VCC when used
in the VOP2 form, and to an arbitrary SGPR-pair when used in the VOP3 form.

6.2. Operands 43 of 283

"RDNA 2" Instruction Set Architecture

When the VOP3 form is used, instructions with a floating-point result can apply an output
modifier (OMOD field) that multiplies the result by: 0.5, 1.0, 2.0 or 4.0. Optionally, the result can
be clamped (CLAMP field) to the range [0.0, +1.0].

Output modifiers apply only to floating point results and are ignored for integer or bit results.
Output modifiers are not compatible with output denormals: if output denormals are enabled,
then output modifiers are ignored. If output demormals are disabled, then the output modifier is
applied and denormals are flushed to zero. Output modifiers are not IEEE compatible: -0 is
flushed to +0. Output modifiers are ignored if the IEEE mode bit is set to 1.

In the table below, all codes can be used when the vector source is nine bits; codes 0 to 255
can be the scalar source if it is eight bits; codes 0 to 127 can be the scalar source if it is seven
bits; and codes 256 to 511 can be the vector source or destination.

6.2.3. Wave64 use of SGPRs

VALU instructions may use SGPRs as a uniform input, shared by all workitems. If the value is
used as simple data value, then the same SGPR is distributed to all 64 workitems. If, on the
other hand, the data value represents a mask (e.g. carry-in, mask for CNDMASK), then each
workitem receives a separate value, and two consecutive SGPRs are read.

6.2.4. Wave64 Destination Restrictions

When a VALU instruction is issued from a wave64, it is actually issued twice as two wave32
instructions. While in most cases the programmer need not be aware of these, it does impose a
prohibition on wave64 VALU instructions which both write and read the same SGPR value.
Doing this may lead to unpredictable results. Specifically, the first pass of a wave64 VALU
instruction may not overwrite a scalar value used by the second half.

6.2.5. OPSEL Field Restrictions

The OPSEL field (of VOP3) is usable only for a subset of VOP3 instructions, and not for VOP1,
VOP2 or VOPC instructions using the VOP3 encoding (these should use SDWA instead).

Table 21. Opcodes usable with OPSEL

 V_MAD_I16  V_INTERP_P2_F16  V_ADD_NC_U16

 V_FMA_F16  V_CVT_PKNORM_I16_F16  V_SUB_NC_U16

 V_ALIGNBIT_B32  V_CVT_PKNORM_U16_F16  V_MUL_LO_U16

 V_ALIGNBYTE_B32  V_MAD_U32_U16  V_LSHLREV_B16

 V_DIV_FIXUP_F16  V_MAD_I32_I16  V_LSHRREV_B16

 V_MIN3_{F16,I16,U16}  V_ASHRREV_I16

 V_MAX3_{F16,I16,U16}  V_MAX_U16

6.2. Operands 44 of 283

"RDNA 2" Instruction Set Architecture

 V_MED3_{F16,I16,U16}  V_MAX_I16

 V_PACK_F16  V_MIN_U16

 V_MIN_I16

6.2.6. Out-of-Range GPRs

When a source VGPR is out-of-range, the instruction uses as input the value from VGPR0.

When the destination GPR is out-of-range, the instruction executes but does not write the
results.

6.3. Instructions
The table below lists the complete VALU instruction set by microcode encoding, except for
VOP3P instructions which are listed in a later section.

Table 22. VALU Instruction Set

VOP3 VOP3 – 1-2 operand opcodes VOP2 VOP1

V_ADD_LSHL_U32  One Operand:  V_ADD_{F16, F32}  V_BFREV_B32

V_ADD3_U32  V_LDEXP_F32  V_ADD_CO_CI_U32  V_CEIL_{ F16,F32, F64}

V_ALIGNBIT_B32  V_LDEXP_F64  V_ADD_NC_U32  V_CLREXCP

V_ALIGNBYTE_B32  V_AND_B32  V_COS_ {F16,F32}

V_AND_OR_B32  Two Operands:  V_ASHRREV_B32  V_CVT_{I32,U32,F16,

F64}_F32

V_BFE_{U32 , I32 }  V_ADD_CO_U32  V_CNDMASK_B32  V_CVT_{I32,U32}_F64

V_BFI_B32  V_ADD_F64  V_CVT_PKRTZ_F16_F32  V_CVT_{U16, I16}_F16

V_CUBEID_F32  V_ADD_NC_{I32, U16, I16}  V_FMAAK_F16  V_CVT_F16_{U16, I16}

V_CUBEMA_F32  V_ASHRREV_{I16, I64}  V_FMAAK_F32  V_CVT_F32_{I32,U32,F16,

F64}

V_CUBESC_F32  V_BCNT_U32_B32  V_FMAC_F16  V_CVT_F32_UBYTE{0,1,2,3

}

V_CUBETC_F32  V_BFM_B32  V_FMAC_F32  V_CVT_F64_{I32,U32}

V_CVT_PK_U8_F32  V_CVT_PK_{I16, U16}_U32  V_FMAMK_F16  V_CVT_FLR_I32_F32

V_DIV_FIXUP_{  V_CVT_PKNORM_{I16, U16}_F16  V_FMAMK_F32  V_CVT_NORM_I16_F16

F16,F32,F64}

V_DIV_FMAS_{F32,F64}  V_CVT_PKNORM_{I16, U16}_F32  V_LDEXP_F16  V_CVT_NORM_U16_F16

V_DIV_SCALE_{F32,F64}  V_INTERP_MOV_F32  V_LSHLREV_B32  V_CVT_OFF_F32_I4

V_FMA_{ F16, F32, F64}  V_INTERP_P1_F32  V_LSHRREV_B32  V_CVT_RPI_I32_F32

V_LERP_U8  V_INTERP_P1LL_F16  V_MAC_{ F16,F32}  V_EXP_{ F16,F32}

V_LSHL_ADD_U32  V_INTERP_P1LV_F16  V_FFBH_{U32, I32}

6.3. Instructions 45 of 283

"RDNA 2" Instruction Set Architecture

VOP3 VOP3 – 1-2 operand opcodes VOP2 VOP1

V_LSHL_OR_B32  V_INTERP_P2_F16  V_FFBL_B32

V_MAD_{I16,U16}  V_INTERP_P2_F32  V_FLOOR_{ F16,F32, F64}

V_MAD_{U64_U32,  V_LSHLREV_{B16, B64}  V_MAX_{ F16,  V_FRACT_{ F16,F32,F64}

I64_I32} F32,I32,U32}

V_MAD_I32_I16  V_LSHRREV_{B16, B64}  V_MIN_{ F16,  V_FREXP_EXP_I16_F16

F32,I32,U32}

V_MAD_I32_I24  V_MAX_{U16, I16, F64}  V_MUL_{F16, F32}  V_FREXP_EXP_I32_F32

 V_MBCNT_HI_U32_B32  V_MUL_HI_I32_I24  V_FREXP_EXP_I32_F64

V_MAD_U32_U16  V_MBCNT_LO_U32_B32  V_MUL_HI_U32_U24  V_FREXP_MANT_{

F16,F32,F64}

V_MAD_U32_U24  V_MIN_ {U16, I16, F64}  V_MUL_I32_I24  V_LOG_ {F16,F32}

V_MAX3_{F16,I16,U16}  V_MUL_F64  V_MUL_LEGACY_F32  V_MOV_B32

V_MAX3_{F32,I32,U32}  V_MUL_HI_{I32,U32}  V_MUL_U32_U24  V_MOV_FED_B32

V_MED3_{F16,I16,U16}  V_MUL_LO_{U16, U32}  V_OR_B32  V_MOVREL{S,D,SD,

SD_2}_B32

V_MED3_{F32,I32,U32}  V_PACK_B32_F16  V_SUB_{F16, F32}  V_NOP

V_MIN3_{F16,I16,U16}  V_READLANE_B32  V_SUB_CO_CI_U32  V_NOT_B32

V_MIN3_{F32,I32,U32}  V_SUB_CO_U32  V_SUB_NC_U32  V_PIPEFLUSH

V_MQSAD_PK_U16_U8  V_SUB_NC_{I32, U16, I16}  V_SUBREV_{F16, F32}  V_RCP_{ F16,F32,F64}

V_MQSAD_PK_U32_U8  V_SUBREV_CO_U32  V_SUBREV_CO_CI_U32  V_RCP_IFLAG_F32

V_MSAD_U8  V_WRITELANE_B32  V_SUBREV_NC_U32  V_READFIRSTLANE_B32

V_MULLIT_F32  V_XNOR_B32  V_RNDNE_{ F16,F32, F64}

V_OR3_B32  V_XOR_B32  V_RSQ_{ F16,F32, F64}

V_PERM_B32  V_SAT_PK_U8_I16

V_PERMLANE16_B32  V_SIN_ {F16,F32}

V_PERMLANEX16_B32  V_SQRT_{ F16,F32,F64}

V_QSAD_PK_U16_U8  V_SWAP_B32

V_SAD_{U8, HI_U8, U16,  V_SWAPREL_B32

U32}

V_TRIG_PREOP_F64  V_TRUNC_{ F16,F32, F64}

V_XAD_U32

V_XOR3_B32

 VOP3P:  V_DOT2C_F32_F16

 V_DOT2_F32_F16  V_DOT4C_I32_I8

 V_DOT2_I32_I16

 V_DOT2_U32_U16

 V_DOT4_I32_I8

 V_DOT4_U32_U8

 V_DOT8_I32_I4

6.3. Instructions 46 of 283

"RDNA 2" Instruction Set Architecture

VOP3 VOP3 – 1-2 operand opcodes VOP2 VOP1

 V_DOT8_U32_U4

The next table lists the compare instructions.

Table 23. VALU Compare Instruction Set

Op Formats Functions Result

V_CMP I16, I32, I64, U16, F, LT, EQ, LE, GT, LG, GE, T Write VCC..
U32, U64
V_CMPX Write exec.

V_CMP F16, F32, F64 F, LT, EQ, LE, GT, LG, GE, T, Write VCC.
O, U, NGE, NLG, NGT, NLE, NEQ, NLT
V_CMPX (o = total order, u = unordered, Write exec.
N = NaN or normal compare)

V_CMP_CLASS F16, F32, F64 Test for one of: signaling-NaN, quiet-NaN, Write VCC.
positive or negative: infinity, normal,
V_CMPX_CLASS subnormal, zero. Write exec.

6.4. Denormalized and Rounding Modes

The shader program has explicit control over the rounding mode applied and the handling of
denormalized inputs and results. The MODE register is set using the S_SETREG instruction; it
has separate bits for controlling the behavior of single and double-precision floating-point
numbers.

Round and denormal modes can also be set using S_ROUND_MODE and
S_DENORM_MODE.

Table 24. Round and Denormal Modes

Field Bit Position Description

FP_ROUND 3:0 [1:0] Single-precision round mode.

[3:2] Double/Half-precision round mode.
Round Modes: 0=nearest even; 1= +infinity; 2= -infinity, 3= toward zero.

FP_DENORM 7:4 [5:4] Single-precision denormal mode.

[7:6] Double/Half-precision denormal mode.
Denormal modes:
0 = Flush input and output denorms.
1 = Allow input denorms, flush output denorms.
2 = Flush input denorms, allow output denorms.
3 = Allow input and output denorms.

6.4. Denormalized and Rounding Modes 47 of 283

"RDNA 2" Instruction Set Architecture

6.5. ALU Clamp Bit Usage

The clamp bit has multpiple uses. For V_CMP instructions, setting the clamp bit to 1 indicates
that the compare signals if a floating point exception occurs. For integer operations, it clamps
the result to the largest and smallest representable value. For floating point operations, it
clamps the result to the range: [0.0, 1.0].

6.6. VGPR Indexing

VGPR indexing allows a value stored in the M0 register to act as an index into the VGPRs for
either the source operand, destination or both for certain MOVE operations.

The table below describes the instructions which enable, disable and control VGPR indexing.

Table 25. VGPR Indexing Instructions

Instruction Encoding Operation

V_MOVRELD_B32 VOP1 Move with relative destination: VGPR[D+M0] = VGPR[S0].

V_MOVRELS_B32 VOP1 Move with relative source: VGPR[D] = VGPR[S0+M0].

V_MOVRELSD_B32 VOP1 Move with relative source and destination: VGPR[D+M0] =

VGPR[S0+M0].

V_MOVRELSD_2_B32 VOP1 Move with relative source and destination, each different:
VGPR[D+M0[25:16]] = VGPR[S0+M0[7:0]].

V_SWAPREL_B32 VOP1 Swap two VGPRs, each relative to a separate index: swap
VGPR[D+M0[25:16]] with VGPR[S0+M0[7:0]].

6.7. Packed Math

Packed math is a form of operation which accelerates arithmetic on two values packed into the
same VGPR. It performs operations on two 16-bit values within a DWORD as if they were
separate threads. For example, a packed add of V0=V1+V2 is really two separate adds: adding
the low 16 bits of each Dword and storing the result in the low 16 bits of V0, and adding the high
halves and storing the result in the high 16 bits of V0.

Packed math uses the instructions below and the microcode format "VOP3P". This format adds
op_sel and neg fields for both the low and high operands, and removes ABS and OMOD.

Table 26. Packed Math Opcodes:

V_PK_MAD_I16 V_PK_MUL_LO_U16 V_PK_ADD_I16 V_PK_SUB_I16

V_PK_LSHLREV_B16 V_PK_LSHRREV_B16 V_PK_ASHRREV_I16 V_PK_MAX_I16

V_PK_MIN_I16 V_PK_MAD_U16 V_PK_ADD_U16 V_PK_SUB_U16

6.5. ALU Clamp Bit Usage 48 of 283

"RDNA 2" Instruction Set Architecture

V_PK_MAX_U16 V_PK_MIN_U16 V_PK_FMA_F16 V_PK_ADD_F16

V_PK_MUL_F16 V_PK_MIN_F16 V_PK_MAX_F16 V_FMA_MIX_F32

V_FMA_MIX_* are not packed math, but perform a single MAD operation on
 a mixture of 16- and 32-bit inputs. They are listed here because they use the
VOP3P encoding.

6.8. Sub-Dword Addressing (SDWA)

Sub DWord Addressing allows a VOP1, VOP2 and VOPC instruction to reference 16 bits of data
in a 32-bit VGPR either the upper or lower half, or any of 4 bytes in the DWORD. The actual
SRC0 operand will be supplied by the SRC0 field of the SDWA word. Each operand can select
the high or low 16-bits or any byte as the destination portion of a VGPR. SDWA is indicated by
setting the SRC0 to the inline constant: SQ_SRC_SDWA. VOPC instructions use a slightly
different version of the SDWA instruction word which as “SD” and “SDST” fields, but not:
OMOD, CLMP, DST_U and DST_SEL.

6.9. Data Parallel Processing (DPP)

Data Parallel ALU operations allow VALU instruction to select operands from different lanes
(threads) rather than just using a thread’s own lane. DPP is compatible only with: VOP1 and
VOP2. There are no new instructions, but there are two new instruction formats in the form of an
extra DWORD of instruction: DPP8 or DPP16.

There are two forms of the DPP instruction word:

DPP8 allows arbitrary swizzling between groups of 8 lanes

DPP16 allows a set of predefined swizzles between groups of 16 lanes

A scan operation is one which computes a value per thread which is based on the values of the
previous threads and possibly itself. E.g. a running sum is the sum of the values from previous
threads in the vector. A reduction operation is essentially a scan which returns a single value
from the highest numbered active thread. These operations take the SP multiple instruction
cycles (at least 8 times what an ADD_F32 takes). Rather than make these a single macro in
SQ, the shader program will have unique instructions for each pass of the scan. This prevents
any instruction scheduling issues (any other waves may execute in between these individual
stage instruction) and allows more general flexibility.

Use of DPP is indicated by setting the SRC0 operand to a literal constant: DPP8 or DPP16.
Note that since SRC-0 is set to the literal value, the actual VGPR address for Source-0 comes

6.8. Sub-Dword Addressing (SDWA) 49 of 283

"RDNA 2" Instruction Set Architecture

from the literal constant (DPP). The scan operation requires the EXEC mask to be set to all 1’s
for proper operation. Unused threads (lanes) should be set to a value which will not change the
result prior to the scan. Readlane, readfirstlane and writelane cannot be used with DPP.

6.10. PERMLANE Specific Rules

V_PERMLANE may not occur immediately after a V_CMPX. To prevent this, any other VALU
opcode may be inserted (e.g. v_mov_b32 v0, v0).

6.10. PERMLANE Specific Rules 50 of 283

"RDNA 2" Instruction Set Architecture

Chapter 7. Scalar Memory Operations

Scalar Memory Read (SMEM) instructions allow a shader program to load data from memory
into SGPRs through the Scalar Data Cache. Instructions can read from 1 to 16 Dwords. Data is
read directly into SGPRs without any format conversion.

The scalar unit reads consecutive Dwords from memory to the SGPRs. This is intended
primarily for loading ALU constants and for indirect T#/S# lookup. No data formatting is
supported, nor is byte or short data.

7.1. Microcode Encoding

Scalar memory read instructions are encoded using the SMEM microcode format.

The fields are described in the table below:

Table 27. SMEM Encoding Field Descriptions

Field Size Description

OP 8 Opcode.

GLC 1 Globally Coherent.

Controls L1 cache policy: 0=hit_lru, 1=miss_evict.

DLC 1 Device Coherent. "1" indicates to bypass the GL1 cache.

SDATA 7 SGPRs to return read data to.

Reads of two Dwords must have an even SDST-sgpr.
Reads of four or more Dwords must have their DST-gpr aligned to a multiple of 4.
SDATA must be: SGPR or VCC. Not: exec or m0.

SBASE 6 SGPR-pair (SBASE has an implied LSB of zero) which provides a base address, or for BUFFER
instructions, a set of 4 SGPRs (4-sgpr aligned) which hold the resource constant. For BUFFER
instructions, the only resource fields used are: base, stride, num_records.

OFFSET 21 An immediate signed byte offset. Must be positive with s_buffer operations.

SOFFSET 7 The address of an SGPR which supplies an unsigned byte address offset. Set this to NULL to
disable.

7.2. Operations

7.1. Microcode Encoding 51 of 283

"RDNA 2" Instruction Set Architecture

7.2.1. S_LOAD_DWORD
These instructions load 1-16 Dwords from memory. The data in SGPRs is specified in SDATA,
and the address is composed of the SBASE, OFFSET, and SOFFSET fields.

Scalar Memory Addressing

S_LOAD :

ADDR = SGPR[base] + inst_offset + { M0 or SGPR[offset] or zero }

All components of the address (base, offset, inst_offset, M0) are in bytes, but the two LSBs are
ignored and treated as if they were zero.

It is illegal and undefined if the inst_offset is negative and the resulting

(inst_offset + (M0 or SGPR[offset])) is negative.

Scalar access to private (scratch) space must either use a buffer constant or manually convert
the address.

Reads using Buffer Constant

Buffer constant fields used: base_address, stride, num_records. Other fields are ignored.

Scalar memory read does not support "swizzled" buffers. Stride is used only for memory
address bounds checking, not for computing the address to access.

The SMEM supplies only a SBASE address (byte) and an offset (byte or Dword). Any "index *
stride" must be calculated manually in shader code and added to the offset prior to the SMEM.

The two LSBs of V#.base and of the final address are ignored to force Dword alignment.

"m_*" components come from the buffer constant (V#):

  offset = OFFSET + SOFFSET (M0, SGPR or zero)
  m_base = { SGPR[SBASE * 2 +1][15:0], SGPR[SBASE*2] }
  m_stride = SGPR[SBASE * 2 +1][31:16]
  m_num_records = SGPR[SBASE * 2 + 2]
  m_size = (m_stride == 0) ? 1 : m_num_records
  addr = (m_base + offset) & ~0x3
  SGPR[SDST] = read_Dword_from_dcache(addr, m_size)

  If more than 1 dword is being read, it is returned to SDST+1, SDST+2, etc,

  and the offset is incremented by 4 bytes per DWORD.

7.2. Operations 52 of 283

"RDNA 2" Instruction Set Architecture

7.2.2. S_DCACHE_INV
This instruction invalidates the entire scalar cache. It does not return anything to SDST.

7.2.3. S_MEMREALTIME
This instruction reads a 64-bit "real time-counter" and returns the value into a pair of SGPRS:
SDST and SDST+1. The time value is from a clock for which the frequency is constant (not
affected by power modes or core clock frequency changes).

7.3. Dependency Checking

Scalar memory reads can return data out-of-order from how they were issued; they can return
partial results at different times when the read crosses two cache lines. The shader program
uses the LGKM_CNT counter to determine when the data has been returned to the SDST
SGPRs. This is done as follows.

• LGKM_CNT is incremented by 1 for every fetch of a single Dword.

• LGKM_CNT is incremented by 2 for every fetch of two or more Dwords.
• LGKM_CNT is decremented by an equal amount when each instruction completes.

Because the instructions can return out-of-order, the only sensible way to use this counter is to
implement S_WAITCNT 0; this imposes a wait for all data to return from previous SMEMs
before continuing.

7.4. Scalar Memory Clauses and Groups

A clause is a sequence of instructions starting with S_CLAUSE and continuing for 2-63
instructions. Clauses lock the instruction arbiter onto this wave until the clause completes.

A group is a set of the same type of instruction that happen to occur in the code but are not
necessarily executed as a clause. A group ends when a non-SMEM instruction is encountered.
Scalar memory instructions are issued in groups. The hardware does not enforce that a single
wave will execute an entire group before issuing instructions from another wave.

Group restrictions:

1. INV must be in a group by itself

2. “TIME” instructions are considered as reads for group rules

Instruction ordering
The data cache is free to re-order instructions. The only assurance of ordering comes when the
shader executes an S_WAITCNT LGKMcnt==0. Cache invalidate instructions are not assured to

7.3. Dependency Checking 53 of 283

"RDNA 2" Instruction Set Architecture

have completed until the shader waits for LGKMcnt==0.

7.5. Alignment and Bounds Checking

SDST
The value of SDST must be even for fetches of two Dwords, or a multiple of four for larger
fetches. If this rule is not followed, invalid data can result. If SDST is out-of-range, the
instruction is not executed.

SBASE
The value of SBASE must be even for S_BUFFER_LOAD (specifying the address of an
SGPR which is a multiple of four). If SBASE is out-of-range, the value from SGPR0 is used.

OFFSET
The value of OFFSET has no alignment restrictions.

Memory Address : If the memory address is out-of-range (clamped), the operation is not
performed for any Dwords that are out-of-range.

7.5. Alignment and Bounds Checking 54 of 283

"RDNA 2" Instruction Set Architecture

Chapter 8. Vector Memory Operations

Vector Memory (VMEM) instructions read or write one piece of data separately for each work-
item in a wavefront into, or out of, VGPRs. This is in contrast to Scalar Memory instructions,
which move a single piece of data that is shared by all threads in the wavefront. All Vector
Memory (VM) operations are processed by the texture cache system.

Software initiates a load, store or atomic operation through the texture cache through one of
three types of VMEM instructions:

• MTBUF: Memory typed-buffer operations.

• MUBUF: Memory untyped-buffer operations.
• MIMG: Memory image operations.
• FLAT: Memory load/store/atomic on flat memory addresses (in subsequent chapter)
• GLOBAL: Memory load/store/atomic on simple address (in subsequent chapter)
• SCRATCH: Memory load/store/atomic to scratch memory (in subsequent chapter)

The instruction defines which VGPR(s) supply the addresses for the operation, which VGPRs
supply or receive data from the operation, and a series of SGPRs that contain the memory
buffer descriptor (V# or T#). Also, MIMG operations supply a texture sampler (S#) from a series
of four SGPRs; this sampler defines texel filtering operations to be performed on data read from
the image.

8.1. Vector Memory Buffer Instructions

Vector-memory (VM) operations transfer data between the VGPRs and buffer objects in memory
through the texture cache (TC). Vector means that one or more piece of data is transferred
uniquely for every thread in the wavefront, in contrast to scalar memory reads, which transfer
only one value that is shared by all threads in the wavefront.

Buffer reads have the option of returning data to VGPRs or directly into LDS.

Examples of buffer objects are vertex buffers, raw buffers, stream-out buffers, and structured
buffers.

Buffer objects support both homogeneous and heterogeneous data, but no filtering of read-data
(no samplers). Buffer instructions are divided into two groups:

• MUBUF: Untyped buffer objects.

◦ Data format is specified in the resource constant.
◦ Load, store, atomic operations, with or without data format conversion.
• MTBUF: Typed buffer objects.
◦ Data format is specified in the instruction.

8.1. Vector Memory Buffer Instructions 55 of 283

"RDNA 2" Instruction Set Architecture

◦ The only operations are Load and Store, both with data format conversion.

Atomic operations take data from VGPRs and combine them arithmetically with data already in
memory. Optionally, the value that was in memory before the operation took place can be
returned to the shader.

All VM operations use a buffer resource constant (V#) which is a 128-bit value in SGPRs. This
constant is sent to the texture cache when the instruction is executed. This constant defines the
address and characteristics of the buffer in memory. Typically, these constants are fetched from
memory using scalar memory reads prior to executing VM instructions, but these constants also
can be generated within the shader.

8.1.1. Simplified Buffer Addressing

The equation below shows how the hardware calculates the memory address for a buffer
access:

8.1.2. Buffer Instructions

Buffer instructions (MTBUF and MUBUF) allow the shader program to read from, and write to,
linear buffers in memory. These operations can operate on data as small as one byte, and up to
four Dwords per work-item. Atomic arithmetic operations are provided that can operate on the
data values in memory and, optionally, return the value that was in memory before the arithmetic
operation was performed.

The D16 instruction variants convert the results to packed 16-bit values. For example,
BUFFER_LOAD_FORMAT_D16_XYZW writes two VGPRs.

Table 28. Buffer Instructions

Instruction Description

MTBUF Instructions

TBUFFER_LOAD_FORMAT_{x,xy,xyz,xyzw} Read from, or write to, a typed buffer object. Also used for a
TBUFFER_STORE_FORMAT_{x,xy,xyz,xyzw} vertex fetch.
TBUFFER_LOAD_FORMAT_D16_{x,xy,xyz,xyzw}
TBUFFER_STORE_FORMAT_D16_{x,xy,xyz,xyzw}

MUBUF Instructions

8.1. Vector Memory Buffer Instructions 56 of 283

"RDNA 2" Instruction Set Architecture

Instruction Description

BUFFER_LOAD_FORMAT_{x,xy,xyz,xyzw} Read to, or write from, an untyped buffer object.

BUFFER_STORE_FORMAT_{x,xy,xyz,xyzw} <size> = byte, ubyte, short, ushort, Dword, Dwordx2, Dwordx3,
BUFFER_LOAD_FORMAT_D16_{x,xy,xyz,xyzw} Dwordx4
BUFFER_STORE_FORMAT_D16_{x,xy,xyz,xyzw} BUFFER_ATOMIC_<op>
BUFFER_LOAD_<size> BUFFER_ATOMIC_<op>_ x2
BUFFER_STORE_<size>

Table 29. Microcode Formats

Field Bit Size Description

OP 4 MTBUF: Opcode for Typed buffer instructions.

7 MUBUF: Opcode for Untyped buffer instructions.

VADDR 8 Address of VGPR to supply first component of address (offset or index). When both index and
offset are used, index is in the first VGPR, offset in the second.

VDATA 8 Address of VGPR to supply first component of write data or receive first component of read-
data.

SOFFSET 8 SGPR to supply unsigned byte offset. SGPR, M0, or inline constant.

SRSRC 5 Specifies which SGPR supplies V# (resource constant) in four consecutive SGPRs. This field
is missing the two LSBs of the SGPR address, since this address is be aligned to a multiple of
four SGPRs.

FORMAT 7 Data Format of data in memory buffer. See: Buffer Image format Table

OFFSET 12 Unsigned byte offset.

OFFEN 1 1 = Supply an offset from VGPR (VADDR). 0 = Do not (offset = 0).

IDXEN 1 1 = Supply an index from VGPR (VADDR). 0 = Do not (index = 0).

GLC 1 Globally Coherent. Controls how reads and writes are handled by the L0 texture cache.
READ
GLC = 0 Reads can hit on the L0 and persist across wavefronts
GLC = 1 Reads miss the L0 and force fetch to L2. No L0 persistence across waves.
WRITE
GLC = 0 Writes miss the L0, write through to L2, and persist in L0 across wavefronts.
GLC = 1 Writes miss the L0, write through to L2. No persistence across wavefronts.
ATOMIC
GLC = 0 Previous data value is not returned. No L0 persistence across wavefronts.
GLC = 1 Previous data value is returned. No L0 persistence across wavefronts.
Note: GLC means "return pre-op value" for atomics.

DLC 1 Device Level Coherent. When set, accesses are forced to miss in level 1

SLC 1 System Level Coherent. Used in conjunction with DLC to determine L2 cache policies.

8.1. Vector Memory Buffer Instructions 57 of 283

"RDNA 2" Instruction Set Architecture

Field Bit Size Description

TFE 1 Texel Fault Enable for PRT (partially resident textures). When set to 1 and fetch returns a
NACK, status is written to the VGPR at DST+1 (first VGPR after all fetch-dest VGPRs).

LDS 1 MUBUF-ONLY: 0 = Return read-data to VGPRs. 1 = Return read-data to LDS instead of

VGPRs.

8.1.3. VGPR Usage

VGPRs supply address and write-data; also, they can be the destination for return data (the
other option is LDS).

Address
Zero, one or two VGPRs are used, depending of the offset-enable (OFFEN) and index-
enable (IDXEN) in the instruction word, as shown in the table below:

Table 30. Address VGPRs

IDXEN OFFEN VGPRn VGPRn+1

0 0 nothing

0 1 uint offset

1 0 uint index

1 1 uint index uint offset

Write Data : N consecutive VGPRs, starting at VDATA. The data format specified in the
instruction word (FORMAT for MTBUF, or encoded in the opcode field for MUBUF) and D16
setting determines how many Dwords to write.

Read Data : Same as writes. Data is returned to consecutive VGPRs.

Read Data Format : Read data is 32 or 16 bits, based on the data format in the instruction or
resource and D16. Float or normalized data is returned as floats; integer formats are returned
as integers (signed or unsigned, same type as the memory storage format). Memory reads of
data in memory that is 32 or 64 bits do not undergo any format conversion unless they return as
16-bit due to D16 being set.

Atomics with Return : Data is read out of the VGPR(s) starting at VDATA to supply to the
atomic operation. If the atomic returns a value to VGPRs, that data is returned to those same
VGPRs starting at VDATA.

8.1.4. Buffer Data

The amount and type of data that is read or written is controlled by the following: the resource
format field, destination-component-selects (dst_sel), and the opcode. FORMAT can come from

8.1. Vector Memory Buffer Instructions 58 of 283

"RDNA 2" Instruction Set Architecture

the resource, instruction fields, or the opcode itself. Dst_sel comes from the resource, but is
ignored for many operations.

Table 31. Buffer Instructions

Instruction Data Format DST SEL

TBUFFER_LOAD_FORMAT_* instruction identity

TBUFFER_STORE_FORMAT_* instruction identity

BUFFER_LOAD_<type> derived identity

BUFFER_STORE_<type> derived identity

BUFFER_LOAD_FORMAT_* resource resource

BUFFER_STORE_FORMAT_* resource resource

BUFFER_ATOMIC_* derived identity

Instruction : The instruction’s format field is used instead of the resource’s fields.

Data format derived : The data format is derived from the opcode and ignores the resource
definition. For example, buffer_load_ubyte sets the data-format to 8 to uint.

The resource’s data format must not be INVALID; that format has specific
 meaning (unbound resource), and for that case the data format is not
replaced by the instruction’s implied data format.

DST_SEL identity : Depending on the number of components in the data-format, this is: X000,
XY00, XYZ0, or XYZW.

The MTBUF derives the data format from the instruction. The MUBUF
BUFFER_LOAD_FORMAT and BUFFER_STORE_FORMAT instructions use format from the
resource; other MUBUF instructions derive data-format from the instruction itself.

D16 Instructions : Load-format and store-format instructions also come in a "d16" variant. For
stores, each 32-bit VGPR holds two 16-bit data elements that are passed to the texture unit.
This texture unit converts them to the texture format before writing to memory. For loads, data
returned from the texture unit is converted to 16 bits, and a pair of data are stored in each 32-bit
VGPR (LSBs first, then MSBs). Control over int vs. float is controlled by FORMAT.

8.1.5. Buffer Addressing

A buffer is a data structure in memory that is addressed with an index and an offset. The index
points to a particular record of size stride bytes, and the offset is the byte-offset within the
record. The stride comes from the resource, the index from a VGPR (or zero), and the offset
from an SGPR or VGPR and also from the instruction itself.

8.1. Vector Memory Buffer Instructions 59 of 283

"RDNA 2" Instruction Set Architecture

Table 32. BUFFER Instruction Fields for Addressing

Field Size Description

inst_offset 12 Literal byte offset from the instruction.

inst_idxen 1 Boolean: get index from VGPR when true, or no index when false.

inst_offen 1 Boolean: get offset from VGPR when true, or no offset when false. Note that inst_offset is
present, regardless of this bit.

The "element size" for a buffer instruction is the amount of data the instruction transfers, or the
number of contiguous bytes of a record for a given index, and is fixed at 4 bytes.

Table 33. V# Buffer Resource Constant Fields for Addressing

Field Size Description

const_base 48 Base address, in bytes, of the buffer resource.

const_stride 14 Stride of the record in bytes (0 to 16,383 bytes, or 0 to

262,143 bytes).

const_num_records 32 Number of records in the buffer.

In units of:
Bytes if: const_stride == 0 || or const_swizzle_enable == false

Otherwise, in units of "stride".

const_add_tid_enable 1 Boolean. Add thread_ID within the wavefront to the index

when true.

const_swizzle_enable 1 Boolean. Indicates that the surface is swizzled when true.

const_index_stride 2 Used only when const_swizzle_en = true. Number of

contiguous indices for a single element (of element_size)
before switching to the next element. There are 8, 16, 32, or
64 indices.

Table 34. Address Components from GPRs

Field Size Description

SGPR_offset 32 An unsigned byte-offset to the address. Comes from an SGPR or M0.

VGPR_offset 32 An optional unsigned byte-offset. It is per-thread, and comes from a VGPR.

VGPR_index 32 An optional index value. It is per-thread and comes from a VGPR.

The final buffer memory address is composed of three parts:

8.1. Vector Memory Buffer Instructions 60 of 283

"RDNA 2" Instruction Set Architecture

• the base address from the buffer resource (V#),

• the offset from the SGPR, and
• a buffer-offset that is calculated differently, depending on whether the buffer is linearly
addressed (a simple Array-of-Structures calculation) or is swizzled.

Figure 8. Address Calculation for a Linear Buffer

Range Checking

Range checking determines if a given buffer memory address is in-range (valid) or out of range.
When an address is out of range, writes are ignored (dropped) and reads return zero. Range
checking is controlled by a 2-bit field in the buffer resource: OOB_SELECT (Out of Bounds
select).

Table 35. Buffer Out Of Bounds Selection

OOB Out of Bounds Check Description or use
SELECT

0 (index >= NumRecords) | (offset >= stride) structured buffers

1 (index >= NumRecords) Raw buffers

2 (NumRecords == 0) do not check bounds

8.1. Vector Memory Buffer Instructions 61 of 283

"RDNA 2" Instruction Set Architecture

OOB Out of Bounds Check Description or use

SELECT

3 Bounds check: Raw

In this mode,
“num_records” is
if (swizzle_en & const_stride != 0x0) reduced by
  OOB = (index >= NumRecords() || (offset+payload > stride)) “sgpr_offset”
else
  OOB = (offset+payload > NumRecords)

Where “payload” is the number of dwords the instruction transfers.

Notes:

1. Reads that go out-of-range return zero (except for components with V#.dst_sel = SEL_1
that return 1).
2. Writes that are out-of-range do not write anything.
3. Load/store-format-* instruction and atomics are range-checked "all or nothing" - either
entirely in or out.
4. Load/store-Dword-x{2,3,4} and range-check per component.

Swizzled Buffer Addressing

Swizzled addressing rearranges the data in the buffer which may improve performance for
arrays of structures. Swizzled addressing also requires Dword-aligned accesses. The buffer’s
STRIDE must be a multiple of element_size.

Index = (inst_idxen ? vgpr_index : 0) +

  (const_add_tid_enable ? thread_id[5:0] : 0)

Offset = (inst_offen ? vgpr_offset : 0) + inst_offset

index_msb = index / const_index_stride

index_lsb = index % const_index_stride
offset_msb = offset / element_size
offset_lsb = offset % element_size

buffer_offset = (index_msb * const_stride + offset_msb *

  element_size) * const_index_stride + index_lsb *
  element_size + offset_lsb

Final Address = const_base + sgpr_offset + buffer_offset

Remember that the "sgpr_offset" is not a part of the "offset" term in the above equations.

8.1. Vector Memory Buffer Instructions 62 of 283

"RDNA 2" Instruction Set Architecture

Figure 9. Example of Buffer Swizzling

Proposed Use Cases for Swizzled Addressing

Here are few proposed uses of swizzled addressing in common graphics buffers.

Table 36. Swizzled Buffer Use Cases

8.1. Vector Memory Buffer Instructions 63 of 283

"RDNA 2" Instruction Set Architecture

DX11 Raw Dx11 Structured Dx11 Structured Scratch Ring / Const

Uav OpenCL (literal offset) (gpr offset) stream-out Buffer
Buffer Object

inst_vgpr_offset_ T F T T T T
en

inst_vgpr_index_ F T T F F F
en

const_stride na <api> <api> scratchSize na na

const_add_tid_en F F F T T F
able

const_buffer_swiz F T T T F F
zle

8.1.6. 16-bit Memory Operations

The D16 buffer instructions allow a kernel to load or store just 16 bits per work item between
VGPRs and memory. There are two variants of these instructions:

• D16 loads data into or stores data from the lower 16 bits of a VGPR.
• D16_HI loads data into or stores data from the upper 16 bits of a VGPR.

For example, BUFFER_LOAD_UBYTE_D16 reads a byte per work-item from memory, converts
it to a 16-bit integer, then loads it into the lower 16 bits of the data VGPR.

8.1.7. Alignment
Formatted ops such as BUFFER_LOAD_FORMAT_* must always be aligned to element_size.

For Dword or larger reads or writes of non-formatted ops (such as BUFFER_LOAD_DWORD),

the two LSBs of the byte-address are ignored, thus forcing Dword alignment.

Memory alignment enforcement for non-formatted ops is controlled by a configuration register:

SH_MEM_CONFIG.alignment_mode.

• DWORD: Automatic alignment to multiple of the smaller of element size or a dword.

• DWORD_STRICT: Require alignment to multiple of the smaller of element size or a dword.
• STRICT: Require alignment to multiple of element size.
• UNALIGNED: No alignment requirements.

8.1.8. Buffer Resource

The buffer resource describes the location of a buffer in memory and the format of the data in
the buffer. It is specified in four consecutive SGPRs (four aligned SGPRs) and sent to the

8.1. Vector Memory Buffer Instructions 64 of 283

"RDNA 2" Instruction Set Architecture

texture cache with each buffer instruction.

The table below details the fields that make up the buffer resource descriptor.

Table 37. Buffer Resource Descriptor

Bits Size Name Description

47:0 48 Base address Byte address.

61:48 14 Stride Bytes 0 to 16383

62 1 Cache swizzle Buffer access. Optionally, swizzle texture cache TC L0 cache banks.

63 1 Swizzle enable Swizzle AOS according to stride, index_stride, and element_size,

else linear (stride * index + offset).

95:64 32 Num_records In units of stride if (stride \>=1), else in bytes.

98:96 3 Dst_sel_x Destination channel select:

0=0, 1=1, 4=R, 5=G, 6=B, 7=A
101:99 3 Dst_sel_y

104:102 3 Dst_sel_z

107:105 3 Dst_sel_w

114:108 7 Format Memory data type.

118:117 2 Index stride 0:8, 1:16, 2:32, or 3:64. Used for swizzled buffer addressing.

119 1 Add tid enable Add thread ID to the index for to calculate the address.

120 1 Resource Level Set to 1.

125:124 2 OOB_SELECT Out of bounds select.

127:126 2 Type Value == 0 for buffer. Overlaps upper two bits of four-bit TYPE field in
128-bit V# resource.

A resource set to all zeros acts as an unbound texture or buffer (return 0,0,0,0).

8.1.9. Memory Buffer Load to LDS

The MUBUF instruction format allows reading data from a memory buffer directly into LDS
without passing through VGPRs. This is supported for the following subset of MUBUF
instructions.

• BUFFER_LOAD_{ubyte, sbyte, ushort, sshort, dword, format_x}.

• It is illegal to set the instruction’s TFE bit for loads to LDS.

LDS_offset = 16-bit unsigned byte offset from M0[15:0].

Mem_offset = 32-bit unsigned byte offset from an SGPR (the SOFFSET SGPR).
idx_vgpr = index value from a VGPR (located at VADDR). (Zero if idxen=0.)
off_vgpr = offset value from a VGPR (located at VADDR or VADDR+1). (Zero if offen=0.)

8.1. Vector Memory Buffer Instructions 65 of 283

"RDNA 2" Instruction Set Architecture

The figure below shows the components of the LDS and memory address calculation:

TIDinWave is only added if the resource (V#) has the ADD_TID_ENABLE field set to 1, whereas
LDS adds it. The MEM_ADDR M# is in the VDATA field; it specifies M0.

Clamping Rules

Memory address clamping follows the same rules as any other buffer fetch. LDS address
clamping: the return data cannot be written outside the LDS space allocated to this wave.

• Set the active-mask to limit buffer reads to those threads that return data to a legal LDS
location.
• The LDSbase (alloc) is in units of 32 Dwords, as is LDSsize.
• M0[15:0] is in bytes.

8.1.10. GLC, DLC and SLC Bits Explained

GLC

The GLC bit means different things for loads, stores, and atomic ops.

GLC Meaning for Loads

• For GLC==0
◦ The load can read data from the GPU L0.
◦ Typically, all loads (except load-acquire) use GLC==0.
• For GLC==1
◦ The load intentionally misses the GPU L0 and reads from L2. If there was a line in the
GPU L0 that matched, it is invalidated; L2 is reread.
◦ NOTE: L2 is not re-read for every work-item in the same wave-front for a single load
instruction. For example: b=uav[N+tid] // assume this is a byte read w/ glc==1 and N is
aligned to 64B In the above op, the first Tid of the wavefront brings in the line from L2
or beyond, and all 63 of the other Tids read from same cache line in the L0.

8.1. Vector Memory Buffer Instructions 66 of 283

"RDNA 2" Instruction Set Architecture

GLC Meaning for Stores

For both GLC==0 and GLC==1, write data are combined across work-items of the wavefront
store clause which can contain multiple store ops; dirtied lines are written to the L2 cache
automatically and invalidated.

Atomics

• For GLC == 0 No return data (this is "write-only" atomic op).

• For GLC == 1 Returns previous value in memory (before the atomic operation).

DLC and SLC

The Device Level Coherent bit (DLC) and System Level Coherent (SLC) bits control the
behavior of the second and third level caches.

Table 38. Vector Load Operations

SLC DLC L2 Cache L1 Cache

0 0 LRU Hit LRU - reads can hit on previous data

0 1 LRU Miss Evict - reads miss

1 0 Stream Hit LRU

1 1 Hit No Allocate Miss Evict

Table 39. Vector Store & Atomic Operations

SLC DLC L2 Cache

0 0 LRU

0 1 Bypass

1 0 Stream - Hit leaves line in cache but do not reset age.

1 1 Hit No Allocate

For stores and atomics, the L1 cache is bypassed (but is coherent).

For stores the L0 cache is always Miss-Evict.

8.2. Vector Memory (VM) Image Instructions

Vector Memory (VM) operations transfer data between the VGPRs and memory through the
texture cache (TC). Vector means the transfer of one or more pieces of data uniquely for every
work-item in the wavefront. This is in contrast to scalar memory reads, which transfer only one
value that is shared by all work-items in the wavefront.

Examples of image objects are texture maps and typed surfaces.

8.2. Vector Memory (VM) Image Instructions 67 of 283

"RDNA 2" Instruction Set Architecture

Image objects are accessed using from one to four dimensional addresses; they are composed
of homogeneous data of one to four elements. These image objects are read from, or written to,
using IMAGE_* or SAMPLE_* instructions, all of which use the MIMG instruction format.
IMAGE_LOAD instructions read an element from the image buffer directly into VGPRS, and
SAMPLE instructions use sampler constants (S#) and apply filtering to the data after it is read.
IMAGE_ATOMIC instructions combine data from VGPRs with data already in memory, and
optionally return the value that was in memory before the operation.

All VM operations use an image resource constant (T#) that is a 128-bit or 256-bit value in
SGPRs. This constant is sent to the texture cache when the instruction is executed. This
constant defines the address, data format, and characteristics of the surface in memory. Some
image instructions also use a sampler constant that is a 128-bit constant in SGPRs. Typically,
these constants are fetched from memory using scalar memory reads prior to executing VM
instructions, but these constants can also be generated within the shader.

Texture fetch instructions have a data mask (DMASK) field. DMASK specifies how many data
components it receives. If DMASK is less than the number of components in the texture, the
texture unit only sends DMASK components, starting with R, then G, B, and A. if DMASK
specifies more than the texture format specifies, the shader receives data based on T#.dst_sel
for the missing components.

8.2.1. Image Instructions

This section describes the image instruction set, and the microcode fields available to those
instructions.

Table 40. Image Instructions

MIMG Description

SAMPLE_* Read and filter data from an image object.

GATHER_* Read up to four texels where each texel contains a single component of an image object
data format. It takes 4 instructions to read RGBA.

IMAGE_LOAD_<op> Read data from an image object using one of the following: image_load, image_load_mip,
image_load_{pck, pck_sgn, mip_pck, mip_pck_sgn}.

IMAGE_STORE Store data to an image object using one of the following:

image_store, image_store_mip, image_store_pck, image_store_mip_pck.

IMAGE_ATOMIC_<op Image atomic operation, which is one of the following: swap, cmpswap, add, sub, umin,
> smin, umax, smax, and, or, xor, inc, dec, fcmpswap, fmin, fmax.

GET_RESINFO Return resource information

8.2. Vector Memory (VM) Image Instructions 68 of 283

"RDNA 2" Instruction Set Architecture

Table 41. Instruction Fields

Field Bit Description
Size

OP 8 Opcode. Formed by joining the OPM and OP fields together.

NSA 2 Number of additional dwords of instruction: 0 - 3. 0 = instruction is 2 dwords in total; 3 =

instruction is 5 dwords in total. Values other than zero imply the "MIMG-NSA" usage of
addressing VGPRs.

VADDR 8 Address of VGPR to supply first component of address.

ADDR1 - 8 12 additional VGPR address fields, used by the MIMG-NSA format. (VADDR acts as
ADDR12 ADDR0).

VDATA 8 Address of VGPR to supply first component of write data or receive first component of
read-data.

SSAMP 5 SGPR to supply S# (sampler constant) in four consecutive SGPRs. Missing two LSBs of
SGPR-address since it is aligned to a multiple of four SGPRs.

SRSRC 5 SGPR to supply T# (resource constant) in four or eight consecutive SGPRs. Missing two
LSBs of SGPR-address since it is aligned to a multiple of four SGPRs.

UNRM 1 Force address to be un-normalized regardless of T#. Set to 1 for image loads, stores
and atomics.

R128 1 Texture buffer resource size: 0 = 256 bits, 1 = 128 bits.

DIM 3 Specifies the dimension of the surface:

0: 1D
1: 2D
2: 3D
3: Cube
4: 1D-array
5: 2D-array
6: 2D-msaa
7: 2D-msaa-array

DMASK 4 Data VGPR enable mask: one to four consecutive VGPRs. Reads: defines which
components are returned.
DMASK[0] = red, DMASK[1] = green, DMASK[2] = blue, DMASK[2] = alpha
For example: dst_sel=unity, DMASK=0110 writes green to VGPRn and blue to
VGPRn+1. D16 packs two components into one VGPR, so the example above would
return 1 VGPR with green in VGPRn[15:0] and blue into VGPRn[31:16].
Writes: defines which components are written with data from VGPRs (missing
components get 0). Enabled components come from consecutive VGPRs.
For example: DMASK=1001: Red is in VGPRn and alpha in VGPRn+1. For D16 writes,
two components' data are packed into one VGPR, so for this example Red ata comes
from VGPRn[15:0] and alpha data from VGPRn[31:16].

8.2. Vector Memory (VM) Image Instructions 69 of 283

"RDNA 2" Instruction Set Architecture

Field Bit Description

Size

GLC 1 Globally Coherent. Controls how reads and writes are handled by the L0 texture cache.
READ:
GLC = 0 Reads can hit on the L0 and persist across waves.
GLC = 1 Reads miss the L0 and force fetch to L2. No L0 persistence across waves.
WRITE:
GLC = 0 Writes miss the L0, write through to L2, and persist in L0 across wavefronts.
GLC = 1 Writes miss the L0, write through to L2. No persistence across wavefronts.
ATOMIC:
GLC = 0 Previous data value is not returned. No L0 persistence across wavefronts.
GLC = 1 Previous data value is returned. No L0 persistence across wavefronts.

DLC 1 Device Level Coherent. When set, accesses are forced to miss in level 1 texture cache.

SLC 1 System Level Coherent. Used in conjunction with DLC to determine L2 cache policies.

TFE 1 Texel Fault Enable for PRT (partially resident textures). When set to 1 and fetch returns
a NACK, status is written to the VGPR at DST+1 (first VGPR after all fetch-dest
VGPRs).

LWE 1 LOD Warning Enable. When set to 1, a texture fetch may return "LOD_CLAMPED = 1".

A16 1 When set, all address components are 16-bit UINT for image ops without sampler; 16-
bit float for image ops with sampler. Address components are packed two per VGPR
except for texel offset where one VGPR contains three 6bit uint offsets. PCF reference
(for _C instructions) ignores this field and is 32bit.

D16 1 VGPR-Data-16bit. On loads, convert data in memory to 16-bit format before storing it in
VGPRs (2 16bit values per VGPR). For stores, convert 16-bit data in VGPRs to memory
data format before writing to memory. Whether the data is treated as float or int is
decided by format. Allowed only with these opcodes:
IMAGE_SAMPLE*
IMAGE_GATHER4*
IMAGE_LOAD
IMAGE_LOAD_MIP
IMAGE_STORE
IMAGE_STORE_MIP

8.2.2. Image Non-Sequential Address (NSA)

To avoid having many move instructions to pack image address VGPRs together, MIMG
supports a “Non Sequential Address” version of the instruction where the VGPR of every
address component is uniquely defined. Data components are still packed. This new format
creates a larger instruction word, which can be up to 5 dwords long. The first address goes in
the VADDR field, and subsequent addresses go into ADDR1-12. The 3 dword form of the
instruction can supply up to 5 addresses; the 4 dword form 9 addresses and the 5 dword form
13 addresses which is the maximum any texture instruction can require.

When using 16-bit addresses, each VGPR holds a pair of addresses and these cannot be
located in different VGPRs.

8.2. Vector Memory (VM) Image Instructions 70 of 283

"RDNA 2" Instruction Set Architecture

8.2.3. Image Opcodes with No Sampler

For image opcodes with no sampler, all VGPR address values are taken as uint. For cubemaps,
face_id = slice * 6 + face.

The table below shows the contents of address VGPRs for the various image opcodes.

Table 42. Image Opcodes with No Sampler

Image Opcode Acnt dim VGPRn VGPRn+1 VGPRn+2 VGPRn+3

(Resource w/o
Sampler)

get_resinfo 0 Any mipid

load / store / atomics 0 1D x

1 1D Array x slice

1 2D x y

2 2D MSAA x y fragid

2 2D Array x y slice

3 2D Array MSAA x y slice fragid

2 3D x y z

2 Cube x y face_id

load_mip / store_mip 1 1D x mipid

2 1D Array x slice mipid

2 2D x y mipid

3 2D Array x y slice mipid

3 3D x y z mipid

3 Cube x y face_id mipid

8.2.4. Image Opcodes with a Sampler

For image opcodes with a sampler, all VGPR address values are taken as float. For cubemaps,
face_id = slice * 8 + face.

Certain sample and gather opcodes require additional values from VGPRs beyond what is
shown. These values are: offset, bias, z-compare, and gradients.

Table 43. Image Opcodes with Sampler

8.2. Vector Memory (VM) Image Instructions 71 of 283

"RDNA 2" Instruction Set Architecture

Image Opcode Acnt dim VGPRn VGPRn+1 VGPRn+2 VGPRn+3

(w/ Sampler)

sample 0 1D x

1 1D Array x slice

1 2D x y

2 2D interlaced x y field

2 2D Array x y slice

2 3D x y z

2 Cube x y face_id

sample_l 1 1D x lod

2 1D Array x slice lod

2 2D x y lod

3 2D interlaced x y field lod

3 2D Array x y slice lod

3 3D x y z lod

3 Cube x y face_id lod

sample_cl 1 1D x clamp

2 1D Array x slice clamp

2 2D x y clamp

3 2D interlaced x y field clamp

3 2D Array x y slice clamp

3 3D x y z clamp

3 Cube x y face_id clamp

gather4 1 2D x y

2 2D interlaced x y field

2 2D Array x y slice

2 Cube x y face_id

gather4_l 2 2D x y lod

3 2D interlaced x y field lod

3 2D Array x y slice lod

3 Cube x y face_id lod

gather4_cl 2 2D x y clamp

3 2D interlaced x y field clamp

3 2D Array x y slice clamp

3 Cube x y face_id clamp

8.2. Vector Memory (VM) Image Instructions 72 of 283

"RDNA 2" Instruction Set Architecture

1. Sample includes sample, sample_d, sample_b, sample_lz, sample_c, sample_c_d,

sample_c_b, sample_c_lz, and getlod.
2. Sample_l includes sample_l and sample_c_l.
3. Sample_cl includes sample_cl, sample_d_cl, sample_b_cl, sample_c_cl, sample_c_d_cl,
and sample_c_b_cl.
4. Gather4 includes gather4, gather4_lz, gather4_c, and gather4_c_lz.

The table below lists and briefly describes the legal suffixes for image instructions:

Table 44. Sample Instruction Suffix Key

Suffix Meaning Extra Description
Addresses

_L LOD - LOD is used instead of computed LOD.

_B LOD BIAS 1: lod bias Add this BIAS to the LOD computed.

_CL LOD CLAMP - Clamp the computed LOD to be no larger than this value.

_D Derivative 2,4 or 6: slopes Send dx/dv, dx/dy, etc. slopes to be used in LOD computation.

_LZ Level 0 - Force use of MIP level 0.

_C PCF 1: z-comp Percentage closer filtering.

_O Offset 1: offsets Send X, Y, Z integer offsets (packed into 1 Dword) to offset XYZ address.

8.2.5. VGPR Usage

Address: The address consists of up to four parts:

{ offset } { bias } { z-compare } { derivative } { body }

These are all packed into consecutive VGPRs.

• Offset: SAMPLE*O*, GATHER*O*

One Dword of offset_xyz. The offsets are six-bit signed integers: X=[5:0], Y=[13:8], and
Z=[21:16].
• Bias: SAMPLE*B*, GATHER*B*. One Dword float.
• Z-compare: SAMPLE*C*, GATHER*C*. One Dword.
• Derivatives (sample_d): 2, 4, or 6 Dwords, packed one Dword per derivative as:

Image Dim Vgpr N N+1 N+2 N+3 N+4 N+5

1D DX/DH DX/DV - - - -

2D DX/DH DY/DH DX/DV DY/DV - -

3D DX/DH DY/DH DZ/DH DX/DV DY/DV DZ/DV

8.2. Vector Memory (VM) Image Instructions 73 of 283

"RDNA 2" Instruction Set Architecture

• Body: One to four Dwords, as defined by the table: [Image Opcodes with Sampler] Address
components are X,Y,Z,W with X in VGPR_M, Y in VGPR_M+1, etc. The number of
components in "body" is the value of the ACNT field in the table, plus one.
• Data: Written from, or returned to, one to four consecutive VGPRs. The amount of data read
or written is determined by the DMASK field of the instruction.
• Reads: DMASK specifies which elements of the resource are returned to consecutive
VGPRs. The texture system reads data from memory and based on the data format
expands it to a canonical RGBA form, filling in zero or one for missing components. Then,
DMASK is applied, and only those components selected are returned to the shader.
• Writes: When writing an image object, it is only possible to write an entire element (all
components), not just individual components. The components come from consecutive
VGPRs, and the texture system fills in the value zero for any missing components of the
image’s data format; it ignores any values that are not part of the stored data format. For
example, if the DMASK=1001, the shader sends Red from VGPR_N, and Alpha from
VGPR_N+1, to the texture unit. If the image object is RGB, the texel is overwritten with Red
from the VGPR_N, Green and Blue set to zero, and Alpha from the shader ignored.
• Atomics: Image atomic operations are supported only on 32- and 64-bit-per pixel surfaces.
The surface data format is specified in the resource constant. Atomic operations treat the
element as a single component of 32- or 64-bits. For atomic operations, DMASK is set to
the number of VGPRs (Dwords) to send to the texture unit. DMASK legal values for atomic
image operations: no other values of DMASK are legal.
0x1 = 32-bit atomics except cmpswap.
0x3 = 32-bit atomic cmpswap.
0x3 = 64-bit atomics except cmpswap.
0xf = 64-bit atomic cmpswap.
• Atomics with Return: Data is read out of the VGPR(s), starting at VDATA, to supply to the
atomic operation. If the atomic returns a value to VGPRs, that data is returned to those
same VGPRs starting at VDATA.

D16 Instructions

Load-format and store-format instructions also come in a "d16" variant. For stores, each 32-bit
VGPR holds two 16-bit data elements that are passed to the texture unit. The texture unit
converts them to the texture format before writing to memory. For loads, data returned from the
texture unit is converted to 16 bits, and a pair of data are stored in each 32- bit VGPR (LSBs
first, then MSBs). The DMASK bit represents individual 16- bit elements; so, when
DMASK=0011 for an image-load, two 16-bit components are loaded into a single 32-bit VGPR.

A16 Instructions

The A16 instruction bit indicates that the address components are 16 bits instead of the usual
32 bits. Components are packed such that the first address component goes into the low 16 bits
([15:0]), and the next into the high 16 bits ([31:16]).

8.2. Vector Memory (VM) Image Instructions 74 of 283

"RDNA 2" Instruction Set Architecture

8.2.6. Image Resource

The image resource (also referred to as T#) defines the location of the image buffer in memory,
its dimensions, tiling, and data format. These resources are stored in four or eight consecutive
SGPRs and are read by MIMG instructions. All undefined or reserved bit must be set to zero
unless otherwise specified.

Table 45. Image Resource Definition

Bits Size Name Comments

128-bit Resource: 1D-tex, 2d-tex, 2d-msaa (multi-sample anti-aliasing)

39:0 40 base address 256-byte aligned (represents bits 47:8). Also used for fmask-ptr.

51:40 12 min lod 4.8 (four uint bits, eight fraction bits) format.

60:52 9 format Memory Data format

77:62 16 width width-1 of mip 0 in texels (2 MSBs must be set to zero)

93:78 16 height height-1 of mip 0 in texels (2 MSBs must be set to zero)

95 1 Resource level Set to 1.

98:96 3 dst_sel_x 0 = 0, 1 = 1, 4 = R, 5 = G, 6 = B, 7 = A.

101:99 3 dst_sel_y

104:102 3 dst_sel_z

107:105 3 dst_sel_w

111:108 4 base level largest mip level in the resource view. For MSAA, this should be set
to 0

115:112 4 last level smallest mip level in resource view. For MSAA, holds log2(number of
samples).

120:116 5 SW mode swizzling (tiling) mode

123:121 3 BC Swizzle Specifies channel ordering for border color data independent of the
T# dst_sel_*s. Internal xyzw channels get the following border color
channels as stored in memory. 0=xyzw, 1=xwyz, 2=wzyx, 3=wxyz,
4=zyxw, 5=yxwz

127:124 4 type 0 = buf, 8 = 1d, 9 = 2d, 10 = 3d, 11 = cube, 12 = 1d-array, 13 = 2d-

array, 14 = 2d-msaa, 15 = 2d-msaa-array. 1-7 are reserved.

256-bit Resource: 1d-array, 2d-array, 3d, cubemap, MSAA

140:128 13 depth Depth-1 of Mip0 for a 3D map; last array slice for a 2D-array or 1D-
array or cube-map.

141 1 Pitch[13] (pitch-1)[13] of mip0 for 1D, 2D and 2D-MSAA.

156:144 13 base array First slice in array of the resource view.

8.2. Vector Memory (VM) Image Instructions 75 of 283

"RDNA 2" Instruction Set Architecture

Bits Size Name Comments

163:160 4 array pitch For Arrays, array pitch for quilts, encoded as trunc(log2(array
pitch))+1. values 8..15 reserved
For 3D, bit 0 indicates SRV or UAV:
0: SRV (base_array ignored, depth w.r.t. base map)
1: UAV (base_array and depth are first and last layer in view, and
w.r.t. mip level specified)

167:164 14 max mip Resource MipLevels-1. Describes the resource, as opposed to base-
level and last-level which describes the resource-view. For MSAA,
holds the number of samples.

179:168 12 min lod warn feedback trigger for LOD

182:180 3 perf mod scales sampler’s perf Z, perf mip, aniso-bias, lod-bias-sec

183 1 corner samples Describes how texels were generated in the resource. 0=center
mod sampled, 1 = corner sampled.

201:200 2 LLC No-alloc 0: NoAlloc = PTE.NoAlloc

1: NoAlloc = Read ? PTE.NoAlloc : 1
2: NoAlloc = Read ? 1 : 0
3: NoAlloc = Read ? 1

202 1 Iterate 256 Indicates that compressed tiles in this surface have been flused out
to every 256B of the tile. Only applies to MSAA depth surfaces.

208:207 2 Max Maximum uncompressed block size used for compressed shader
Uncompressed writes
block size

210:209 2 Max Compressed Maximum compressed block size used for compressed shader writes
block size

211 1 Meta Pipe Aligned Maintains pipe alignment in metadata addressing (DCC and tiling)

212 1 Write compression Enable compressed writes from shader

enable

213 1 Compression enable delta color compression (DCC)

Enable

214 1 Alpha is on MSB Set to 1 if the surface’s component swap is not reversed (DCC)

215 1 Color Transform Auto=0, none=1 (DCC)

255:216 40 Meta Data Address Upper bits of meta-data address (DCC) [47:8]

8.2.7. Image Sampler

The sampler resource (also referred to as S#) defines what operations to perform on texture
map data read by sample instructions. These are primarily address clamping and filter options.
Sampler resources are defined in four consecutive SGPRs and are supplied to the texture
cache with every sample instruction.

Table 46. Image Sampler Definition

8.2. Vector Memory (VM) Image Instructions 76 of 283

"RDNA 2" Instruction Set Architecture

Bits Size Name Description

2:0 3 clamp x Clamp/wrap mode:

0: Wrap
1: Mirror
5:3 3 clamp y 2: ClampLastTexel
3: MirrorOnceLastTexel
4: ClampHalfBorder
8:6 3 clamp z 5: MirrorOnceHalfBorder
6: ClampBorder
7: MirrorOnceBorder

11:9 3 max aniso ratio 0 = 1:1

1 = 2:1
2 = 4:1
3 = 8:1
4 = 16:1

14:12 3 depth compare func 0: Never

1: Less
2: Equal
3: Less than or equal
4: Greater
5: Not equal
6: Greater than or equal
7: Always

15 1 force unnormalized Force address cords to be unorm: 0 = address coordinates are

normalized, in [0,1); 1 = address coordinates are
unnormalized:[0,dim).

18:16 3 aniso threshold threshold under which floor(aniso ratio) determines number of
samples and step size

19 1 mc coord trunc enables bilinear blend fraction truncation to 1 bit for motion
compensation

20 1 force degamma force format to srgb if data_format allows

26:21 6 aniso bias 6 bits, in u1.5 format.

27 1 trunc coord selects texel coordinate rounding or truncation.

28 1 disable cube wrap disables seamless DX10 cubemaps, allows cubemaps to clamp
according to clamp_x and clamp_y fields

30:29 2 filter_mode 0 = Blend (lerp); 1 = min, 2 = max.

31 1 skip degamma disabled degamma (sRGB→Linear) conversion.

43:32 12 min lod minimum LOD ins resource view space (0.0 = T#.base_level) u4.8.

55:44 12 max lod maximum LOD ins resource view space

59:56 4 perf_mip defines range of lod fractions that snap to nearest mip only when
mip_filter=Linear

63:60 4 perf z defines range of z fractions that snap to nearest z layer

z_filter=Linear

75:64 12 lod bias s6.8. This is bits[11:0] of the LOD bias.

8.2. Vector Memory (VM) Image Instructions 77 of 283

"RDNA 2" Instruction Set Architecture

Bits Size Name Description

77:76 2 lod bias high This is bits [13:12] of the LOD bias.

83:78 6 lod bias sec bias added to computed LOD, scaled by T#.perf_modulation. s2.4.

85:84 2 xy mag filter Magnification filter: 0=point, 1=bilinear, 2=aniso-point, 3=aniso-linear

87:86 2 xy min filter Minification filter: 0=point, 1=bilinear, 2=aniso-point, 3=aniso-linear

89:88 2 z filter Volume Filter: 0=none (use XY min/mag filter), 1=point, 2=linear

91:90 2 mip filter Mip level filter: 0=none (disable mipmapping,use base-leve),
1=point, 2=linear

93 1 Aniso_override Disable Aniso filtering if base_level = last_level

94 1 blend_zero_prt For PRT fetches, zero out texel if not resident

95 1 reserved reserved. set to zero.

107:96 12 border color ptr

127:126 2 border color type Opaque-black, transparent-black, white, use border color ptr.

8.2.8. Data Formats

The table below details all the data formats that can be used by image and buffer resources.

Table 47. Buffer and Image Data Formats

# Buffer and Image # Buffer and Image # Image Formats
Formats Formats

0 INVALID 43 11_11_10_FLOAT 128 8_SRGB

1 8_UNORM 44 10_10_10_2_UNORM 129 8_8_SRGB

2 8_SNORM 45 10_10_10_2_SNORM 130 8_8_8_8_SRGB

3 8_USCALED 48 10_10_10_2_UINT 131 6E4_FLOAT

4 8_SSCALED 49 10_10_10_2_SINT 132 5_9_9_9_FLOAT

5 8_UINT 50 2_10_10_10_UNORM 133 5_6_5_UNORM

6 8_SINT 51 2_10_10_10_SNORM 134 1_5_5_5_UNORM

7 16_UNORM 52 2_10_10_10_USCALED 135 5_5_5_1_UNORM

8 16_SNORM 53 2_10_10_10_SSCALED 136 4_4_4_4_UNORM

9 16_USCALED 54 2_10_10_10_UINT 137 4_4_UNORM

10 16_SSCALED 55 2_10_10_10_SINT 138 1_UNORM

11 16_UINT 56 8_8_8_8_UNORM 139 1_REVERSED_UNORM

12 16_SINT 57 8_8_8_8_SNORM 140 32_FLOAT_CLAMP

13 16_FLOAT 58 8_8_8_8_USCALED 141 8_24_UNORM

14 8_8_UNORM 59 8_8_8_8_SSCALED 142 8_24_UINT

8.2. Vector Memory (VM) Image Instructions 78 of 283

"RDNA 2" Instruction Set Architecture

# Buffer and Image # Buffer and Image # Image Formats

Formats Formats

15 8_8_SNORM 60 8_8_8_8_UINT 143 24_8_UNORM

16 8_8_USCALED 61 8_8_8_8_SINT 144 24_8_UINT

17 8_8_SSCALED 62 32_32_UINT 145 X24_8_32_UINT

18 8_8_UINT 63 32_32_SINT 146 X24_8_32_FLOAT

19 8_8_SINT 64 32_32_FLOAT 147 GB_GR_UNORM

20 32_UINT 65 16_16_16_16_UNORM 148 GB_GR_SNORM

21 32_SINT 66 16_16_16_16_SNORM 149 GB_GR_UINT

22 32_FLOAT 67 16_16_16_16_USCALE 150 GB_GR_SRGB

D

23 16_16_UNORM 68 16_16_16_16_SSCALE 151 BG_RG_UNORM

D

24 16_16_SNORM 69 16_16_16_16_UINT 152 BG_RG_SNORM

25 16_16_USCALED 70 16_16_16_16_SINT 153 BG_RG_UINT

26 16_16_SSCALED 71 16_16_16_16_FLOAT 154 BG_RG_SRGB

27 16_16_UINT 72 32_32_32_UINT

28 16_16_SINT 73 32_32_32_SINT Compressed Formats

29 16_16_FLOAT 74 32_32_32_FLOAT 169 BC1_UNORM

36 10_11_11_FLOAT 75 32_32_32_32_UINT 170 BC1_SRGB

76 32_32_32_32_SINT 171 BC2_UNORM

77 32_32_32_32_FLOAT 172 BC2_SRGB

173 BC3_UNORM

174 BC3_SRGB

175 BC4_UNORM

176 BC4_SNORM

177 BC5_UNORM

178 BC5_SNORM

179 BC6_UFLOAT

180 BC6_SFLOAT

181 BC7_UNORM

182 BC7_SRGB

8.2.9. Vector Memory Instruction Data Dependencies

When a VM instruction is issued, it schedules the reads of address and write-data from VGPRs
to be sent to the texture unit. Any ALU instruction which attempts to write this data before it has

8.2. Vector Memory (VM) Image Instructions 79 of 283

"RDNA 2" Instruction Set Architecture

been sent to the texture unit will be stalled.

The shader developer’s responsibility to avoid data hazards associated with VMEM instructions
include waiting for VMEM read instruction completion before reading data fetched from the TC
(VMCNT and VSCNT).

This is explained in the section: Data Dependency Resolution

8.2.10. Ray Tracing

Ray Tracing support includes the following instructions:

• IMAGE_BVH_INTERSECT_RAY
• IMAGE_BVH64_INTERSECT_RAY

These instructions receive ray data from the VGPRs and fetch BVH (Bounding Volume
Hierarchy) from memory.

• Box BVH nodes perform 4x Ray/Box intersection, sorts the 4 children based on intersection
distance and returns the child pointers and hit status.
• Triangle nodes perform 1 Ray/Triangle intersection test and returns the intersection point
and triangle ID.

The two instructions are identical, except that the “64” version supports a 64-bit address while
the normal version supports only a 32bit address. Both instructions can use the “A16” instruction
field to reduce some (but not all) of the address components to 16 bits (from 32). These
addresses are: ray_dir and ray_inv_dir.

Instruction definition and fields

image_bvh_intersect_ray vgpr_d[4], vgpr_a[11], sgpr_r[4]

image_bvh_intersect_ray vgpr_d[4], vgpr_a[8], sgpr_r[4] A16=1
image_bvh64_intersect_ray vgpr_d[4], vgpr_a[12], sgpr_r[4]
image_bvh64_intersect_ray vgpr_d[4], vgpr_a[9], sgpr_r[4] A16=1

Table 48. Ray Tracing VGPR Contents

VGPR BVH A16=0 BVH A16=1 BVH64 A16=0 BVH64 A16=1
_A

0 node_pointer (u32) node_pointer (u32) node_pointer [31:0] (u32) node_pointer [31:0] (u32)

1 ray_extent (f32) ray_extent (f32) node_pointer [63:32] (u32) node_pointer [63:32] (u32)

2 ray_origin.x (f32) ray_origin.x (f32) ray_extent (f32) ray_extent (f32)

3 ray_origin.y (f32) ray_origin.y (f32) ray_origin.x (f32) ray_origin.x (f32)

4 ray_origin.z (f32) ray_origin.z (f32) ray_origin.y (f32) ray_origin.y (f32)

8.2. Vector Memory (VM) Image Instructions 80 of 283

"RDNA 2" Instruction Set Architecture

VGPR BVH A16=0 BVH A16=1 BVH64 A16=0 BVH64 A16=1

_A

5 ray_dir.x (f32) [15:0] = ray_dir.x (f16) ray_origin.z (f32) ray_origin.z (f32)

[31:16] = ray_dir.y (f16)

6 ray_dir.y (f32) [15:0] = ray_dir.z (f16) ray_dir.x (f32) [15:0] = ray_dir.x (f16)
[31:16] = ray_inv_dir.x(f16) [31:16] = ray_dir.y (f16)

7 ray_dir.z (f32) [15:0] = ray_inv_dir.y (f16) ray_dir.y (f32) [15:0] = ray_dir.z (f16)
[31:16] = ray_inv_dir.z [31:16] = ray_inv_dir.x(f16)
(f16)

8 ray_inv_dir.x (f32) unused ray_dir.z (f32) [15:0] = ray_inv_dir.y (f16)

[31:16] = ray_inv_dir.z
(f16)

9 ray_inv_dir.y (f32) unused ray_inv_dir.x (f32) unused

10 ray_inv_dir.z (f32) unused ray_inv_dir.y (f32) unused

11 unused unused ray_inv_dir.z (f32) unused

Vgpr_d[4] are the destination VGPRs of the results of intersection testing. The values returned
here are different depending on the type of BVH node that was fetched. For box nodes the
results contain the 4 pointers of the children boxes in intersection time sorted order. For triangle
BVH nodes the results contain the intersection time and triangle ID of the triangle tested.

Sgpr_r[4] is the texture descriptor for the operation. The instruction is encoded with
use_128bit_resource=1.

Restrictions on image_bvh instructions

• DMASK must be set to 0xf (instruction returns all four DWORDs)

• D16 must be set to 0 (16 bit return data is not supported)
• R128 must be set to 1 (256 bit T#s are not supported)
• UNRM must be set to 1 (only unnormalized coordinates are supported)
• DIM must be set to 0 (BVH textures are 1D)
• LWE must be set to 0 (LOD warn is not supported)
• TFE must be set to 0 (no support for writing out the extra DWORD for the PRT hit status)
• SSAMP must be set to 0 (just a placeholder, since samplers are not used by the instruction)

The return order settings of the BVH ops are ignored instead they use the in-order read return
queue.

Texture Resource Definition

The T# used with these instructions is different from other image instructions.

Table 49. Ray Tracing Resource Contents

8.2. Vector Memory (VM) Image Instructions 81 of 283

"RDNA 2" Instruction Set Architecture

Field Bits Size Data

Base Address 39:0 40 Base address of the BVH texture 256 byte aligned

Reserved 54:40 15 Set to zero

Box growing amount 62:55 8 Number of ULPs to be added during ray-box test, encoded as
unsigned integer

Box sorting enable 63 1 Whether the ray-box test result need to be sorted

Size 105:64 42 Number of nodes minus 1 in the BVH texture used to enforce bounds
checking

Reserved 119:106 14 Set to zero

Triangle_return_mode 120 1 0: Return data for triangle tests are

{0: t_num, 1: t_denom, 2: triangle_id, 3: hit_status}
1: Return data for triangle tests are
{0: t_num, 1: t_denom, 2: I_num, 3: J_num}

Reserved 122:121 2 Set to zero

big_page 123 1 Describes resource page usage

0 : No page size override.
1 : Indicates when a whole resource is only using pages that are >=
64KB in size.

Type 127:124 4 Set to 0x8

Barycentrics
The ray-tracing hardware is designed to support computation of barycentric coordinates directly
in hardware. This uses the “triangle_return_mode” in the table in the previous section (T#
descriptor).

Table 50. Ray Tracing Return Mode

DWORD Return Mode =0 Return Mode = 1

Field Name Type Field Name Type

0 t_num float32 t_num float32

1 t_denom float32 t_denom float32

2 triangle_id uint32 I_num float32

3 hit_status uint32 (boolean value) J_num float32

8.2. Vector Memory (VM) Image Instructions 82 of 283

"RDNA 2" Instruction Set Architecture

Chapter 9. Flat Memory Instructions

Flat Memory instructions read, or write, one piece of data into, or out of, VGPRs; they do this
separately for each work-item in a wavefront. Unlike buffer or image instructions, Flat
instructions do not use a resource constant to define the base address of a surface. Instead,
Flat instructions use a single flat address from the VGPR; this addresses memory as a single
flat memory space. This memory space includes video memory, system memory, LDS memory,
and scratch (private) memory. It does not include GDS memory. Parts of the flat memory space
may not map to any real memory, and accessing these regions generates a memory-violation
error. The determination of the memory space to which an address maps is controlled by a set
of "memory aperture" base and size registers.

9.1. Flat Memory Instruction

Flat memory instructions let the kernel read or write data in memory, or perform atomic
operations on data already in memory. These operations occur through the texture L2 cache.
The instruction declares which VGPR holds the address (either 32- or 64-bit, depending on the
memory configuration), the VGPR which sends and the VGPR which receives data. Flat
instructions also use M0 as described in the table below:

Table 51. Flat, Global and Scratch Microcode Formats

Field Bit Size Description

OP 7 Opcode. Can be Flat, Scratch or Global instruction. See next table.

ADDR 8 VGPR which holds the address. For 64-bit addresses, ADDR has the LSBs, and ADDR+1 has
the MSBs.
As an offset a single VGPR has a 32 bit unsigned offset.
For FLAT_*: specifies an address.
For GLOBAL_* and SCRATCH_* when SADDR is NULL: specifies an address.
For GLOBAL_* and SCRATCH_* when SADDR is not NULL: specifies an offset.

DATA 8 VGPR which holds the first Dword of data. Instructions can use 0-4 Dwords.

VDST 8 VGPR destination for data returned to the kernel, either from LOADs or Atomics with GLC=1
(return pre-op value).

SLC 1 System Level Coherent. Used in conjunction with DLC to determine L2 cache policies.

DLC 1 Device Level Coherent. Controls GL1 cache bypass.

GLC 1 Global Level Coherent. For Atomics, GLC: 1 means return pre-op value, 0 means do not return
pre-op value.

SEG 2 Memory Segment: 0=FLAT, 1=SCRATCH, 2=GLOBAL, 3=reserved.

LDS 1 When set, data is moved from memory to LDS instead of to VGPRs. Available only for loads.
For Global and Scratch only; must be zero for Flat.

OFFSET 12 Address offset.

Scratch, Global: 12-bit signed byte offset.
Flat: must be positive.

9.1. Flat Memory Instruction 83 of 283

"RDNA 2" Instruction Set Architecture

Field Bit Size Description

SADDR 7 Scalar SGPR that provides an offset address. To disable use, set this field to NULL or 0x7f
(exec_hi).
Meaning of this field is different for Scratch and Global:
Flat: Unused.
Scratch: Use an SGPR (instead of VGPR) for the address.
Global: Use the SGPR to provide a base address; the VGPR provides a 32-bit offset per lane.

M0 16 Implied use of M0 for SCRATCH and GLOBAL only when LDS=1. Provides the LDS address-
offset.

Table 52. Flat, Global and Scratch Opcodes

Flat Opcodes Global Opcodes Scratch Opcodes

FLAT GLOBAL SCRATCH

FLAT_LOAD_UBYTE GLOBAL_LOAD_UBYTE SCRATCH_LOAD_UBYTE

FLAT_LOAD_UBYTE_D16 GLOBAL_LOAD_UBYTE_D16 SCRATCH_LOAD_UBYTE_D16

FLAT_LOAD_UBYTE_D16_HI GLOBAL_LOAD_UBYTE_D16_HI SCRATCH_LOAD_UBYTE_D16_HI

FLAT_LOAD_SBYTE GLOBAL_LOAD_SBYTE SCRATCH_LOAD_SBYTE

FLAT_LOAD_SBYTE_D16 GLOBAL_LOAD_SBYTE_D16 SCRATCH_LOAD_SBYTE_D16

FLAT_LOAD_SBYTE_D16_HI GLOBAL_LOAD_SBYTE_D16_HI SCRATCH_LOAD_SBYTE_D16_HI

FLAT_LOAD_USHORT GLOBAL_LOAD_USHORT SCRATCH_LOAD_USHORT

FLAT_LOAD_SSHORT GLOBAL_LOAD_SSHORT SCRATCH_LOAD_SSHORT

FLAT_LOAD_SHORT_D16 GLOBAL_LOAD_SHORT_D16 SCRATCH_LOAD_SHORT_D16

FLAT_LOAD_SHORT_D16_HI GLOBAL_LOAD_SHORT_D16_HI SCRATCH_LOAD_SHORT_D16_HI

FLAT_LOAD_DWORD GLOBAL_LOAD_DWORD SCRATCH_LOAD_DWORD

FLAT_LOAD_DWORDX2 GLOBAL_LOAD_DWORDX2 SCRATCH_LOAD_DWORDX2

FLAT_LOAD_DWORDX3 GLOBAL_LOAD_DWORDX3 SCRATCH_LOAD_DWORDX3

FLAT_LOAD_DWORDX4 GLOBAL_LOAD_DWORDX4 SCRATCH_LOAD_DWORDX4

FLAT_STORE_BYTE GLOBAL_STORE_BYTE SCRATCH_STORE_BYTE

FLAT_STORE_BYTE_D16_HI GLOBAL_STORE_BYTE_D16_HI SCRATCH_STORE_BYTE_D16_HI

FLAT_STORE_SHORT GLOBAL_STORE_SHORT SCRATCH_STORE_SHORT

FLAT_STORE_SHORT_D16_HI GLOBAL_STORE_SHORT_D16_HI SCRATCH_STORE_SHORT_D16_HI

FLAT_STORE_DWORD GLOBAL_STORE_DWORD SCRATCH_STORE_DWORD

FLAT_STORE_DWORDX2 GLOBAL_STORE_DWORDX2 SCRATCH_STORE_DWORDX2

FLAT_STORE_DWORDX3 GLOBAL_STORE_DWORDX3 SCRATCH_STORE_DWORDX3

FLAT_STORE_DWORDX4 GLOBAL_STORE_DWORDX4 SCRATCH_STORE_DWORDX4

9.1. Flat Memory Instruction 84 of 283

"RDNA 2" Instruction Set Architecture

Flat Opcodes Global Opcodes Scratch Opcodes

FLAT_ATOMIC_SWAP GLOBAL_ATOMIC_SWAP none

FLAT_ATOMIC_CMPSWAP GLOBAL_ATOMIC_CMPSWAP none

FLAT_ATOMIC_ADD GLOBAL_ATOMIC_ADD none

FLAT_ATOMIC_SUB GLOBAL_ATOMIC_SUB none

FLAT_ATOMIC_SMIN GLOBAL_ATOMIC_SMIN none

FLAT_ATOMIC_UMIN GLOBAL_ATOMIC_UMIN none

FLAT_ATOMIC_SMAX GLOBAL_ATOMIC_SMAX none

FLAT_ATOMIC_UMAX GLOBAL_ATOMIC_UMAX none

FLAT_ATOMIC_AND GLOBAL_ATOMIC_AND none

FLAT_ATOMIC_OR GLOBAL_ATOMIC_OR none

FLAT_ATOMIC_XOR GLOBAL_ATOMIC_XOR none

FLAT_ATOMIC_INC GLOBAL_ATOMIC_INC none

FLAT_ATOMIC_DEC GLOBAL_ATOMIC_DEC none

FLAT_ATOMIC_FMIN GLOBAL_ATOMIC_FMIN none

FLAT_ATOMIC_FMAX GLOBAL_ATOMIC_FMAX none

FLAT_ATOMIC_FCMPSWAP GLOBAL_ATOMIC_FCMPSWAP none

none GLOBAL_ATOMIC_CSUB none

none GLOBAL_LOAD_DWORD_ADDTID none

none GLOBAL_STORE_DWORD_ADDTID none

The atomic instructions above are also available in "_X2" versions (64-bit).

Note: GLOBAL_ATOMIC_CSUB requires that the user set GLC=1.

9.2. Instructions
The FLAT instruction set is nearly identical to the Buffer instruction set, but without the FORMAT
reads and writes. Unlike Buffer instructions, FLAT instructions cannot return data directly to
LDS, but only to VGPRs.

FLAT instructions do not use a resource constant (V#) or sampler (S#); however, they do require
a additional register (FLAT_SCRATCH) to hold scratch-space memory address information in
case any threads' address resolves to scratch space. See the scratch section for details.

Internally, FLAT instruction are executed as both an LDS and a Buffer instruction; so, they
increment both VM_CNT/VS_CNT and LGKM_CNT and are not considered done until both
have been decremented. There is no way beforehand to determine whether a FLAT instruction
uses only LDS or texture memory space.

9.2. Instructions 85 of 283

"RDNA 2" Instruction Set Architecture

9.2.1. Ordering
Flat instructions can complete out of order with each other. If one flat instruction finds all of its
data in Texture cache, and the next finds all of its data in LDS, the second instruction might
complete first. If the two fetches return data to the same VGPR, the result are unknown.

9.2.2. Important Timing Consideration

Since the data for a FLAT load can come from either LDS or the texture cache, and because
these units have different latencies, there is a potential race condition with respect to the
VM_CNT/VS_CNT and LGKM_CNT counters. Because of this, the only sensible S_WAITCNT
value to use after FLAT instructions is zero.

9.3. Addressing
FLAT instructions support both 64- and 32-bit addressing. The address size is set using a mode
register (PTR32), and a local copy of the value is stored per wave.

The addresses for the aperture check differ in 32- and 64-bit mode; however, this is not covered
here.

64-bit addresses are stored with the LSBs in the VGPR at ADDR, and the MSBs in the VGPR at
ADDR+1.

For scratch space, the texture unit takes the address from the VGPR and does the following.

Address = VGPR[addr] + TID_in_wave * Size

  - private aperture base (in SH_MEM_BASES)
  + offset (from flat_scratch)

9.3.1. Legal Addressing Combinations

Not every combination of addressing modes is legal for each type of instruction. The legal
combinations are:

• FLAT
a. VGPR (32 or 64 bit) supplies the complete address. SADDR must be NULL.
• Global
a. VGPR (32 or 64 bit) supplies the address. Indicated by: SADDR == NULL.
b. SGPR (64 bit) supplies an address, and a VGPR (32 bit) supplies an offset
• SCRATCH
a. VGPR (32 bit) supplies an offset. Indicated by SADDR==NULL.

9.3. Addressing 86 of 283

"RDNA 2" Instruction Set Architecture

b. SGPR (32 bit) supplies an offset. Indicated by SADDR!=NULL.

Every mode above can also add the "instruction immediate offset" to the address.

9.4. Global
Global instructions are similar to Flat instructions, but the programmer must ensure that no
threads access LDS space; thus, no LDS bandwidth is used by global instructions.

Global instructions offer two types of addressing:

• Memory_addr = VGPR-address + instruction offset.

• Memory_addr = SGPR-address + VGPR-offset + instruction offset.

The size of the address component is dependent on ADDRESS_MODE: 32-bits or 64-bit

pointers. The VGPR-offset is 32 bits.

These instructions also allow direct data movement to LDS from memory without going through
VGPRs.

Since these instructions do not access LDS, only VM_CNT/VS_CNT is used, not LGKM_CNT. If
a global instruction does attempt to access LDS, the instruction returns MEM_VIOL.

9.5. Scratch
Scratch instructions are similar to Flat, but the programmer must ensure that no threads access
LDS space, and the memory space is swizzled. Thus, no LDS bandwidth is used by scratch
instructions.

Scratch instructions also support multi-Dword access and mis-aligned access (although mis-
aligned is slower).

Scratch instructions use the following addressing:

• Memory_addr = flat_scratch.addr + swizzle(V/SGPR_offset + inst_offset, threadID)

The offset can come from either an SGPR or a VGPR, and is a 32- bit unsigned byte.

The size of the address component is dependent on the ADDRESS_MODE: 32-bits or 64-bit
pointers. The VGPR-offset is 32 bits.

These instructions also allow direct data movement to LDS from memory without going through
VGPRs.

Since these instructions do not access LDS, only VM_CNT/VS_CNT is used, not LGKM_CNT. It

9.4. Global 87 of 283

"RDNA 2" Instruction Set Architecture

is not possible for a Scratch instruction to access LDS; thus, no error or aperture checking is
done.

9.6. Memory Error Checking

Both the texture unit and LDS can report that an error occurred due to a bad address. This can
occur for the following reasons:

• invalid address (outside any aperture)

• write to read-only surface
• misaligned data
• out-of-range address:
◦ LDS access with an address outside the range: [ 0, MIN(M0, LDS_SIZE)-1 ]
◦ Scratch access with an address outside the range: [0, scratch-size -1 ]

The policy for threads with bad addresses is: writes outside this range do not write a value, and
reads return zero.

Addressing errors from either LDS or texture are returned on their respective "instruction done"
busses as MEM_VIOL. This sets the wave’s MEM_VIOL TrapStatus bit and causes an
exception (trap) if the corresponding EXCPEN bit is set.

9.7. Data
FLAT instructions can use zero to four consecutive Dwords of data in VGPRs and/or memory.
The DATA field determines which VGPR(s) supply source data (if any), and the VDST VGPRs
hold return data (if any). No data-format conversion is done.

“D16” instructions use only 16-bit of the VGPR instead of the full 32bits. “D16_HI” instructions
read or write only the high 16-bits, while “D16” use the low 16-bits. Scratch & Global D16 load
instructions with LDS=1 will write the entire 32-bits of LDS.

9.8. Scratch Space (Private)

Scratch (thread-private memory) is an area of memory defined by the aperture registers. When
an address falls in scratch space, additional address computation is automatically performed by
the hardware. The kernel must provide additional information for this computation to occur in the
form of the FLAT_SCRATCH register.

The wavefront must supply the scratch size and offset (for space allocated to this wave) with
every FLAT request. Prior to issuing any FLAT or Scratch instructions, the shader program must
initialize the FLAT_SCRATCH register with the base address of scratch space allocated this

9.6. Memory Error Checking 88 of 283

"RDNA 2" Instruction Set Architecture

wave.

FLAT_SCRATCH is a 64-bit, byte address. The shader composes the value by adding together
two separate values: the base address, which can be passed in via an initialized SGPR, or
perhaps through a constant buffer, and the per-wave allocation offset (also initialized in an
SGPR).

9.8. Scratch Space (Private) 89 of 283

"RDNA 2" Instruction Set Architecture

Chapter 10. Data Share Operations

Local data share (LDS) is a very low-latency, RAM scratchpad for temporary data with at least
one order of magnitude higher effective bandwidth than direct, uncached global memory. It
permits sharing of data between work-items in a work-group, as well as holding parameters for
pixel shader parameter interpolation. Unlike read-only caches, the LDS permits high-speed
write-to-read re-use of the memory space (gather/read/load and scatter/write/store operations).

10.1. Overview
The figure below shows the conceptual framework of the LDS is integration into the memory of
AMD GPUs using OpenCL.

Figure 10. High-Level Memory Configuration

Physically located on-chip, directly adjacent to the ALUs, the LDS is approximately one order of
magnitude faster than global memory (assuming no bank conflicts).

There are 128kB memory per workgroup processor split up into 64 banks of dword-wide RAMs.
These 64 banks are further sub-divided into two sets of 32-banks each where 32 of the banks
are affiliated with a pair of SIMD32’s, and the other 32 banks are affiliated with the other pair of
SIMD32’s within the WGP. Each bank is a 512x32 two-port RAM (1R/1W per clock cycle).
Dwords are placed in the banks serially, but all banks can execute a store or load
simultaneously. One work-group can request up to 64kB memory.

10.1. Overview 90 of 283

"RDNA 2" Instruction Set Architecture

The high bandwidth of the LDS memory is achieved not only through its proximity to the ALUs,
but also through simultaneous access to its memory banks. Thus, it is possible to concurrently
execute 32 write or read instructions, each nominally 32-bits; extended instructions,
read2/write2, can be 64-bits each. If, however, more than one access attempt is made to the
same bank at the same time, a bank conflict occurs. In this case, for indexed and atomic
operations, the hardware is designed to prevent the attempted concurrent accesses to the same
bank by turning them into serial accesses. This decreases the effective bandwidth of the LDS.
For increased throughput (optimal efficiency), therefore, it is important to avoid bank conflicts. A
knowledge of request scheduling and address mapping is key to achieving this.

10.2. Dataflow in Memory Hierarchy

The figure below is a conceptual diagram of the dataflow within the memory structure.

Data can be loaded into LDS either by transferring it from VGPRs to LDS using "DS"
instructions, or by loading in from memory. When loading from memory, the data may be loaded
into VGPRs first or for some types of loads it may be loaded directly into LDS from memory. To
store data from LDS to global memory, data is read from LDS and placed into the workitem’s
VGPRs, then written out to global memory. To make effective use of the LDS, a kernel must
perform many operations on what is transferred between global memory and LDS.

LDS atomics are performed in the LDS hardware. (Thus, although ALUs are not directly used for
these operations, latency is incurred by the LDS executing this function.)

10.3. LDS Modes and Allocation: CU vs. WGP Mode

Workgroups of waves are dispatched in one of two modes: CU or WGP. This mode controls
whether the waves of a workgroup are distributed across just two SIMD32’s (CU mode), or

10.2. Dataflow in Memory Hierarchy 91 of 283

"RDNA 2" Instruction Set Architecture

across all 4 SIMD32’s (WGP mode) within a WGP.

In CU mode, waves are allocated to two SIMD32’s which share a texture memory unit, and are
allocated LDS space which is all local (on the same side) as the SIMDs. This mode can provide
higher LDS memory bandwidth than WGP mode.

In WGP mode, the waves are distributed over all 4 SIMD32’s and LDS space maybe allocated
anywhere within the LDS memory. Waves may access data on the "near" or "far" side of LDS
equally, but performance may be lower in some cases. This mode provides more ALU and
texture memory bandwidth to a single workgroup (of at least 4 waves).

10.4. LDS Access

There are 3 forms of Local Data Share access:

• Direct Read – reads a single dword from LDS and broadcasts the data as input to a vector
ALU op.
• Indexed Read/write and Atomic ops – read/write address comes from a VGPR and data
to/from VGPR.
◦ LDS-ops require up to 3 inputs: 2data+1addr and immediate return VGPR.
• Parameter Interpolation – similar to direct read but with specific addressing.
◦ Reads up to 2 parameters (P0, P1-P0) or (P2-P0) from one attribute to be supplied to a
muladd.
◦ Also supplies individual parameter read for general interpolation (or select I,J=0.0)

The following subsections describe these methods.

10.4.1. LDS Direct Reads

Direct reads are only available in LDS, not in GDS.

LDS Direct reads occur in vector ALU (VALU) instructions and allow the LDS to supply a single
DWORD value which is broadcast to all threads in the wavefront and is used as the SRC0 input
to the ALU operations. A VALU instruction indicates that input is to be supplied by LDS by using
the LDS_DIRECT for the SRC0 field.

The LDS address and data-type of the data to be read from LDS comes from the M0 register:

10.4. LDS Access 92 of 283

"RDNA 2" Instruction Set Architecture

LDS_addr = M0[15:0] (byte address and must be Dword aligned)

DataType = M0[18:16]
  0 unsigned byte
  1 unsigned short
  2 Dword
  3 unused
  4 signed byte
  5 signed short

10.4.2. LDS Parameter Reads

Parameter reads are only available in LDS, not in GDS.

Pixel shaders use LDS to read vertex parameter values; the pixel shader then interpolates them
to find the per-pixel parameter values. LDS parameter reads occur when the following opcodes
are used.

• V_INTERP_P1_F32 D = P10 * S + P0 Parameter interpolation, first step.

• V_INTERP_P2_F32D = P20 * S + DParameter interpolation, second step.
• V_INTERP_MOV_F32D = {P10,P20,P0}[S]Parameter load.

The typical parameter interpolation operations involves reading three parameters: P0, P10, and
P20, and using the two barycentric coordinates, I and J, to determine the final per-pixel value:

Final value = P0 + P10 * I + P20 * J

Parameter interpolation instructions indicate the parameter attribute number (0 to 32) and the
component number (0=x, 1=y, 2=z and 3=w).

Table 53. Parameter Instruction Fields

Field Size Description

VDST 8 Destination VGPR. Also acts as source for v_interp_p2_f32.

OP 2 Opcode:
0: v_interp_p1_f32 VDST = P10 * VSRC + P0
1: v_interp_p2_f32 VDST = P20 * VSRC + VDST
2: v_interp_mov_f32 VDST = (P0, P10 or P20 selected by VSRC[1:0])
P0, P10 and P20 are parameter values read from LDS

ATTR 6 Attribute number: 0 to 32.

ATTRCHAN 2 0=X, 1=Y, 2=Z, 3=W

VSRC 8 Source VGPR supplies interpolation "I" or "J" value. For OP==v_interp_mov_f32: 0=P10,
1=P20, 2=P0.

( M0 ) 32 Use of the M0 register is automatic. M0 must contain: { 1’b0, new_prim_mask[15:1],

lds_param_offset[15:0] }

10.4. LDS Access 93 of 283

"RDNA 2" Instruction Set Architecture

Parameter interpolation and parameter move instructions must initialize the M0 register before
using it. The lds_param_offset[15:0] is an address offset from the beginning of LDS storage
allocated to this wavefront to where parameters begin in LDS memory for this wavefront.

The new_prim_mask is a 15-bit mask with one bit per quad; a one in this mask indicates that
this quad begins a new primitive, a zero indicates it uses the same primitive as the previous
quad. The mask is 15 bits, not 16, since the first quad in a wavefront begins a new primitive and
so it is not included in the mask.

Parameter Interpolation on 16-bit data

The above parameter interpolation opcodes use the VINTRP microcode format, but for
interpolation on 16-bit data, the VOP3 format is used. The opcodes supported are:

Opcode Operation Description

V_INTERP_P1LL_F16 d.f32 = lds.f16 * vgpr.f32 + lds.f16 attr_word selects LDS high or low 16bits. “LL” is
for “two LDS arguments.”

V_INTERP_P2_F16 d.f16 = lds.f16 * vgpr.f32 + vgpr.f32 Final computation. attr_word selects LDS high or
low 16bits. Result is written to the 16 LSB’s of the
dest-vgpr.

In the VOP3 encoding, the following fields are overloaded:

• SRC1 : this field holds the VINTERP VSRC value (I or J)

• SRC0 : this is treated as a set of bit-fields: {attr_word[1], attr_chan[2], attr[6]}
“attr_word” is a bit to select the low or high half of the LDS word. 1=high, 0=low.

10.4.3. Data Share Indexed and Atomic Access

Both LDS and GDS can perform indexed and atomic data share operations. For brevity, "LDS"
is used in the text below and, except where noted, also applies to GDS.

Indexed and atomic operations supply a unique address per work-item from the VGPRs to the
LDS, and supply or return unique data per work-item back to VGPRs. Due to the internal
banked structure of LDS, operations can complete in as little as one cycle (for wave32, or 2
cycles for wave64), or take as many 64 cycles, depending upon the number of bank conflicts
(addresses that map to the same memory bank).

Indexed operations are simple LDS load and store operations that read data from, and return
data to, VGPRs.

Atomic operations are arithmetic operations that combine data from VGPRs and data in LDS,
and write the result back to LDS. Atomic operations have the option of returning the LDS "pre-
op" value to VGPRs.

10.4. LDS Access 94 of 283

"RDNA 2" Instruction Set Architecture

The table below lists and briefly describes the LDS instruction fields.

Table 54. LDS Instruction Fields

Field Size Description

OP 7 LDS opcode.

GDS 1 0 = LDS, 1 = GDS.

OFFSET0 8 Immediate offset, in bytes. Instructions with one address combine the offset fields into a single 16-
bit unsigned offset: {offset1, offset0}. Instructions with two addresses (for example: READ2) use
OFFSET1 8 the offsets separately as two 8- bit unsigned offsets.

VDST 8 VGPR to which result is written: either from LDS-load or atomic return value.

ADDR 8 VGPR that supplies the byte address offset.

DATA0 8 VGPR that supplies first data source.

DATA1 8 VGPR that supplies second data source.

The M0 register is not used for most LDS-indexed operations: only the "ADD_TID" instructions
read M0 and for these it represents a byte address.

Table 55. LDS Indexed Load/Store

Load / Store Description

DS_READ_{B32,B64,B96,B128,U8,I8,U16,I16} Read one value per thread; sign extend to Dword, if signed.

DS_READ2_{B32,B64} Read two values at unique addresses.

DS_READ2ST64_{B32,B64} Read 2 values at unique addresses; offset *= 64.

DS_WRITE_{B32,B64,B96,B128,B8,B16} Write one value.

DS_WRITE2_{B32,B64} Write two values.

DS_WRITE2ST64_{B32,B64} Write two values, offset *= 64.

DS_WRXCHG2_RTN_{B32,B64} Exchange GPR with LDS-memory.

DS_WRXCHG2ST64_RTN_{B32,B64} Exchange GPR with LDS-memory; offset *= 64.

DS_PERMUTE_B32 Forward permute. Does not write any LDS memory.

LDS[dst] = src0
returnVal = LDS[thread_id]
where thread_id is 0..63.

DS_BPERMUTE_B32 Backward permute. Does not actually write any LDS memory.
LDS[thread_id] = src0
where thread_id is 0..63, and returnVal = LDS[dst].

Single Address Instructions

LDS_Addr = LDS_BASE + VGPR[ADDR] + {InstrOffset1,InstrOffset0}

Double Address Instructions

10.4. LDS Access 95 of 283

"RDNA 2" Instruction Set Architecture

LDS_Addr0 = LDS_BASE + VGPR[ADDR] + InstrOffset0*ADJ +

LDS_Addr1 = LDS_BASE + VGPR[ADDR] + InstrOffset1*ADJ
  Where ADJ = 4 for 8, 16 and 32-bit data types; and ADJ = 8 for 64-bit.

Note that LDS_ADDR1 is used only for READ2*, WRITE2*, and WREXCHG2*.

The address comes from VGPR, and both ADDR and InstrOffset are byte addresses.

At the time of wavefront creation, LDS_BASE is assigned to the physical LDS region owned by
this wavefront or work-group.

Specify only one address by setting both offsets to the same value. This causes only one read
or write to occur and uses only the first DATA0.

DS_{READ,WRITE}_ADD_TID Addressing

LDS_Addr = LDS_BASE + {Inst_offset1, Inst_offset0} + TID(0..63)*4 + M0

  Note: no part of the address comes from a VGPR. M0 must be dword-aligned.

The "ADD_TID" (add thread-id) is a separate form where the base address for the instruction is
common to all threads, but then each thread has a fixed offset added in based on its thread-ID
within the wave. This allows a convenient way to quickly transfer data between VGPRs and LDS
without having to use a VGPR to supply an address.

LDS Atomic Ops

DS_<atomicOp> OP, GDS=0, OFFSET0, OFFSET1, VDST, ADDR, Data0, Data1

Data size is encoded in atomicOp: byte, word, Dword, or double.

LDS_Addr0 = LDS_BASE + VGPR[ADDR] + {InstrOffset1,InstrOffset0}

ADDR is a Dword address. VGPRs 0,1 and dst are double-GPRs for doubles data.

VGPR data sources can only be VGPRs or constant values, not SGPRs.

10.4.4. LDS Lane-permute Ops

DS_PERMUTE instructions allow data to be swizzled arbitrarily across 32 lanes. Two versions
of the instruction are provided: a forward (scatter) and backward (gather).

Note that in wave64 mode the permute operates only across 32 lanes at a time of each half of a
wave64. In other words, it executes as if were two independent wave32’s. Each half-wave can

10.4. LDS Access 96 of 283

"RDNA 2" Instruction Set Architecture

use indices in the range 0-31 to reference lanes in that same half-wave.

These instructions use the LDS hardware but do not use any memory storage, and may be
used by waves which have not allocated any LDS space. The instructions supply a data value
from VGPRs and an index value per lane.

• ds_permute_b32 : Dst[index[0..31]] = src[0..31] Where [0..31] is the lane number

• ds_bpermute_b32 : Dst[0..31] = src[index[0..31]]

The EXEC mask is honored for both reading the source and writing the destination. Index
values out of range will wrap around (only index bits [6:2] are used, the other bits of the index
are ignored). Reading from disabled lanes returns zero.

In the instruction word: VDST is the dest VGPR, ADDR is the index VGPR, and DATA0 is the
source data VGPR. Note that index values are in bytes (so multiply by 4), and have the ‘offset0’
field added to them before use.

10.5. Global Data Share

Global data share is similar to LDS, but is a single memory accessible by all waves on the GPU.
Global Data share uses the same instruction format as local data share (indexed operations
only – no interpolation or direct reads). Instructions increment the LGKM_cnt for all reads, writes
and atomics, and decrement LGKM_cnt when the instruction completes.

M0 is used for:

• [15:0] holds SIZE, in bytes

• [31:16] holds BASE address in bytes

10.5. Global Data Share 97 of 283

"RDNA 2" Instruction Set Architecture

Chapter 11. Exporting Pixel and Vertex

Data
The export instruction copies pixel or vertex shader data from VGPRs into a dedicated output
buffer. The export instruction outputs the following types of data.

• Vertex Position
• Vertex Parameter
• Pixel color
• Pixel depth (Z)
• Primitive Data

11.1. Microcode Encoding

The export instruction uses the EXP microcode format.

Table 56. EXP Encoding Field Descriptions

Field Size Description

VM 1 Valid Mask. When set to 1, this indicates that the EXEC mask represents the
valid-mask for this wavefront. It can be sent multiple times per shader (the final
value is used), but must be sent at least once per pixel shader.

DONE 1 This is the final pixel shader or vertex-position export of the program. Used only
for pixel and position exports. Set to zero for parameters.

COMPR 1 Compressed data. When set, indicates that the data being exported is 16-bits
per component rather than the usual 32-bit.

TARGET 6 Indicates type of data exported.

  0..7 MRT 0..7

  8 Z
  9 Null (no data)
  12-16 Position 0..4
  20 Primitive data
  32-63 Param 0..31

11.1. Microcode Encoding 98 of 283

"RDNA 2" Instruction Set Architecture

Field Size Description

EN 4 COMPR==1: export half-Dword enable. Valid values are: 0x0,3,C,F.

[0] enables VSRC0 : R,G from one VGPGR
[2] enables VSRC1 : B,A from one VGPR
COMPR==0: [0-3] = enables for VSRC0..3.
EN can be zero (used when exporting only valid mask to NULL target).

VSRC3 8 VGPR from which to read data.

VSRC2 8 Pos & Param: vsrc0=X, 1=Y, 2=Z, 3=W
MRT: vsrc0=R, 1=G, 2=B, 3=A
VSRC1 8

VSRC0 8

11.2. Operations

11.2.1. Pixel Shader Exports

Export instructions copy color data to the MRTs. Data has up to four components (R, G, B, A).
Optionally, export instructions also output depth (Z) data.

Every pixel shader must have at least one export instruction. The last export instruction
executed must have the DONE bit set to one.

The EXEC mask is applied to all exports. Only pixels with the corresponding EXEC bit set to 1
export data to the output buffer. Results from multiple exports are accumulated in the output
buffer.

At least one export must have the VM bit set to 1. This export, in addition to copying data to the
color or depth output buffer, also informs the color buffer which pixels are valid and which have
been discarded. The value of the EXEC mask communicates the pixel valid mask. If multiple
exports are sent with VM set to 1, the mask from the final export is used. If the shader program
wants to only update the valid mask but not send any new data, the program can do an export
to the NULL target.

11.2.2. Vertex Shader Exports

The vertex shader uses export instructions to output vertex position data and vertex parameter
data to the output buffer. This data is passed on to subsequent pixel shaders.

Every vertex shader must output at least one position vector (x, y, z; w is optional) to the POS0
target. The last position export must have the DONE bit set to 1. A vertex shader can export
zero or more parameters. For optimized performance, it is recommended to output all position

11.2. Operations 99 of 283

"RDNA 2" Instruction Set Architecture

data as early as possible in the vertex shader.

11.3. Primitive Shader Exports

The primitive shader may export Position and Primitive data. Before exporting, the shader must
request that space be allocated in the output buffer using the ALLOC_REQ message.

11.4. Dependency Checking

Export instructions are executed by the hardware in two phases. First, the instruction is selected
to be executed, and EXPCNT is incremented by 1. At this time, the hardware requests the use
of internal busses needed to complete the instruction.

When access to the bus is granted, the EXEC mask is read and the VGPR data sent out. After
the last of the VGPR data is sent, the EXPCNT counter is decremented by 1.

Use S_WAITCNT on EXPCNT to prevent the shader program from overwriting EXEC or the
VGPRs holding the data to be exported before the export operation has completed.

Multiple export instructions can be outstanding at one time. Exports of the same type (for
example: position) are completed in order, but exports of different types can be completed out of
order.

If the STATUS register’s SKIP_EXPORT bit is set to one, the hardware treats all EXPORT
instructions as if they were NOPs.

11.3. Primitive Shader Exports 100 of 283

"RDNA 2" Instruction Set Architecture

Chapter 12. Instructions

This chapter lists, and provides descriptions for, all instructions in the RDNA Generation
environment. Instructions are grouped according to their format.

Instruction suffixes have the following definitions:

• B32 Bitfield (untyped data) 32-bit

• B64 Bitfield (untyped data) 64-bit
• F16 floating-point 16-bit
• F32 floating-point 32-bit (IEEE 754 single-precision float)
• F64 floating-point 64-bit (IEEE 754 double-precision float)
• I8 signed 8-bit integer
• I16 signed 16-bit integer
• I32 signed 32-bit integer
• I64 signed 64-bit integer
• U16 unsigned 16-bit integer
• U32 unsigned 32-bit integer
• U64 unsigned 64-bit integer

If an instruction has two suffixes (for example, _I32_F32), the first suffix indicates the destination
type, the second the source type.

The following abbreviations are used in instruction definitions:

• D = destination
• U = unsigned integer
• S = source
• SCC = scalar condition code
• I = signed integer
• B = bitfield

Note: .u or .i specifies to interpret the argument as an unsigned or signed integer.

Note: Rounding and Denormal modes apply to all floating-point operations unless otherwise
specified in the instruction description.

12.1. SOP2 Instructions

12.1. SOP2 Instructions 101 of 283

"RDNA 2" Instruction Set Architecture

Instructions in this format may use a 32-bit literal constant which occurs immediately after the
instruction.

Opcode Name Description

0 S_ADD_U32 Add two unsigned integers with carry-out.

  D.u32 = S0.u32 + S1.u32;

  SCC = S0.u32 + S1.u32 >= 0x100000000ULL ? 1 : 0.

1 S_SUB_U32 Subtract the second unsigned integer from the first with
carry-out.

  D.u = S0.u - S1.u;

  SCC = (S1.u > S0.u ? 1 : 0). // unsigned overflow or carry-out for
S_SUBB_U32.

2 S_ADD_I32 Add two signed integers with carry-out.

This opcode is not suitable for use with S_ADDC_U32 for

implementing 64-bit operations.

  D.i = S0.i + S1.i;

  SCC = (S0.u[31] == S1.u[31] && S0.u[31] != D.u[31]). // signed
overflow.

3 S_SUB_I32 Subtract the second signed integer from the first with
carry-out.

This opcode is not suitable for use with S_SUBB_U32 for

implementing 64-bit operations.

  D.i = S0.i - S1.i;

  SCC = (S0.u[31] != S1.u[31] && S0.u[31] != D.u[31]). // signed
overflow.

4 S_ADDC_U32 Add two unsigned integers with carry-in and carry-out.

  D.u32 = S0.u32 + S1.u32 + SCC;

  SCC = S0.u32 + S1.u32 + SCC >= 0x100000000ULL ? 1 : 0.

5 S_SUBB_U32 Subtract the second unsigned integer from the first with
carry-in and carry-out.

  D.u = S0.u - S1.u - SCC;

  SCC = (S1.u + SCC > S0.u ? 1 : 0). // unsigned overflow.

6 S_MIN_I32 Minimum of two signed integers.

  D.i = (S0.i < S1.i) ? S0.i : S1.i;

  SCC = (S0.i < S1.i).

7 S_MIN_U32 Minimum of two unsigned integers.

  D.u = (S0.u < S1.u) ? S0.u : S1.u;

  SCC = (S0.u < S1.u).

8 S_MAX_I32 Maximum of two signed integers.

  D.i = (S0.i > S1.i) ? S0.i : S1.i;

  SCC = (S0.i > S1.i).

12.1. SOP2 Instructions 102 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

9 S_MAX_U32 Maximum of two unsigned integers.

  D.u = (S0.u > S1.u) ? S0.u : S1.u;

  SCC = (S0.u > S1.u).

10 S_CSELECT_B32 Conditional select based on scalar condition code.

  D.u = SCC ? S0.u : S1.u.

11 S_CSELECT_B64 Conditional select base on scalar condition code.

  D.u64 = SCC ? S0.u64 : S1.u64.

14 S_AND_B32 Bitwise AND.

  D = S0 & S1;
  SCC = (D != 0).

15 S_AND_B64 Bitwise AND.

  D = S0 & S1;
  SCC = (D != 0).

16 S_OR_B32 Bitwise OR.

  D = S0 | S1;
  SCC = (D != 0).

17 S_OR_B64 Bitwise OR.

  D = S0 | S1;
  SCC = (D != 0).

18 S_XOR_B32 Bitwise XOR.

  D = S0 ^ S1;
  SCC = (D != 0).

19 S_XOR_B64 Bitwise XOR.

  D = S0 ^ S1;
  SCC = (D != 0).

20 S_ANDN2_B32 Bitwise ANDN2.

  D = S0 & ~S1;
  SCC = (D != 0).

21 S_ANDN2_B64 Bitwise ANDN2.

  D = S0 & ~S1;
  SCC = (D != 0).

22 S_ORN2_B32 Bitwise ORN2.

  D = S0 | ~S1;
  SCC = (D != 0).

23 S_ORN2_B64 Bitwise ORN2.

  D = S0 | ~S1;
  SCC = (D != 0).

12.1. SOP2 Instructions 103 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

24 S_NAND_B32 Bitwise NAND.

  D = ~(S0 & S1);

  SCC = (D != 0).

25 S_NAND_B64 Bitwise NAND.

  D = ~(S0 & S1);

  SCC = (D != 0).

26 S_NOR_B32 Bitwise NOR.

  D = ~(S0 | S1);
  SCC = (D != 0).

27 S_NOR_B64 Bitwise NOR.

  D = ~(S0 | S1);
  SCC = (D != 0).

28 S_XNOR_B32 Bitwise XNOR.

  D = ~(S0 ^ S1);
  SCC = (D != 0).

29 S_XNOR_B64 Bitwise XNOR.

  D = ~(S0 ^ S1);
  SCC = (D != 0).

30 S_LSHL_B32 Logical shift left.

  D.u = S0.u << S1.u[4:0];

  SCC = (D.u != 0).

31 S_LSHL_B64 Logical shift left.

  D.u64 = S0.u64 << S1.u[5:0];

  SCC = (D.u64 != 0).

32 S_LSHR_B32 Logical shift right.

  D.u = S0.u >> S1.u[4:0];

  SCC = (D.u != 0).

33 S_LSHR_B64 Logical shift right.

  D.u64 = S0.u64 >> S1.u[5:0];

  SCC = (D.u64 != 0).

34 S_ASHR_I32 Arithmetic shift right (preserve sign bit).

  D.i = signext(S0.i) >> S1.u[4:0];

  SCC = (D.i != 0).

35 S_ASHR_I64 Arithmetic shift right (preserve sign bit).

  D.i64 = signext(S0.i64) >> S1.u[5:0];

  SCC = (D.i64 != 0).

36 S_BFM_B32 Bitfield mask.

  D.u = ((1 << S0.u[4:0]) - 1) << S1.u[4:0].

12.1. SOP2 Instructions 104 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

37 S_BFM_B64 Bitfield mask.

  D.u64 = ((1ULL << S0.u[5:0]) - 1) << S1.u[5:0].

38 S_MUL_I32 Multiply two signed integers.

  D.i = S0.i * S1.i.

39 S_BFE_U32 Bit field extract. S0 is Data, S1[4:0] is field offset,

S1[22:16] is field width.

  D.u = (S0.u >> S1.u[4:0]) & ((1 << S1.u[22:16]) - 1);

  SCC = (D.u != 0).

40 S_BFE_I32 Bit field extract. S0 is Data, S1[4:0] is field offset,

S1[22:16] is field width.

  D.i = signext((S0.i >> S1.u[4:0]) & ((1 << S1.u[22:16]) - 1));

  SCC = (D.i != 0).

41 S_BFE_U64 Bit field extract. S0 is Data, S1[5:0] is field offset,

S1[22:16] is field width.

  D.u64 = (S0.u64 >> S1.u[5:0]) & ((1 << S1.u[22:16]) - 1);

  SCC = (D.u64 != 0).

42 S_BFE_I64 Bit field extract. S0 is Data, S1[5:0] is field offset,

S1[22:16] is field width.

  D.i64 = signext((S0.i64 >> S1.u[5:0]) & ((1 << S1.u[22:16]) - 1));

  SCC = (D.i64 != 0).

44 S_ABSDIFF_I32 Compute the absolute value of difference between two values.

  D.i = S0.i - S1.i;

  if(D.i < 0) then
  D.i = -D.i;
  endif;
  SCC = (D.i != 0).

Functional examples:

S_ABSDIFF_I32(0x00000002, 0x00000005) => 0x00000003

S_ABSDIFF_I32(0xffffffff, 0x00000000) => 0x00000001
S_ABSDIFF_I32(0x80000000, 0x00000000) => 0x80000000 // Note: result
is negative!
S_ABSDIFF_I32(0x80000000, 0x00000001) => 0x7fffffff
S_ABSDIFF_I32(0x80000000, 0xffffffff) => 0x7fffffff
S_ABSDIFF_I32(0x80000000, 0xfffffffe) => 0x7ffffffe

46 S_LSHL1_ADD_U32 Logical shift left by 1 bit and then add.

  D.u = (S0.u << N) + S1.u; // N is the shift value in the opcode

  SCC = (((S0.u << N) + S1.u) >= 0x100000000ULL ? 1 : 0). // unsigned
overflow.

47 S_LSHL2_ADD_U32 Logical shift left by 2 bits and then add.

  D.u = (S0.u << N) + S1.u; // N is the shift value in the opcode

  SCC = (((S0.u << N) + S1.u) >= 0x100000000ULL ? 1 : 0). // unsigned
overflow.

12.1. SOP2 Instructions 105 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

48 S_LSHL3_ADD_U32 Logical shift left by 3 bits and then add.

  D.u = (S0.u << N) + S1.u; // N is the shift value in the opcode

  SCC = (((S0.u << N) + S1.u) >= 0x100000000ULL ? 1 : 0). // unsigned
overflow.

49 S_LSHL4_ADD_U32 Logical shift left by 4 bits and then add.

  D.u = (S0.u << N) + S1.u; // N is the shift value in the opcode

  SCC = (((S0.u << N) + S1.u) >= 0x100000000ULL ? 1 : 0). // unsigned
overflow.

50 S_PACK_LL_B32_B16 Pack two short values into the destination.

  D.u[31:0] = { S1.u[15:0], S0.u[15:0] }.

51 S_PACK_LH_B32_B1 Pack two short values into the destination.

6
  D.u[31:0] = { S1.u[31:16], S0.u[15:0] }.

52 S_PACK_HH_B32_B1 Pack two short values into the destination.

6
  D.u[31:0] = { S1.u[31:16], S0.u[31:16] }.

53 S_MUL_HI_U32 Multiple two unsigned integers and store the high 32 bits.

  D.u = (S0.u * S1.u) >> 32.

54 S_MUL_HI_I32 Multiple two signed integers and store the high 32 bits.

  D.i = (S0.i * S1.i) >> 32.

12.2. SOPK Instructions

Instructions in this format may not use a 32-bit literal constant which occurs immediately after
the instruction.

Opcode Name Description

0 S_MOVK_I32 Sign extension from a 16-bit constant.

  D.i32 = signext(SIMM16[15:0]).

12.2. SOPK Instructions 106 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

1 S_VERSION Do nothing. Argument is ignored by hardware. This opcode is

not designed for inserting wait states as it is possible the
next instruction will issue in the same cycle. Do not use this
opcode to resolve wait state hazards, use S_NOP instead.

This opcode is used to specify the microcode version for tools

that interpret shader microcode; it may also be used to validate
microcode is running with the correct compatibility settings in
drivers and functional models that support multiple generations.
We strongly encourage this opcode be included at the top of
every shader block to simplify debug and catch configuration
errors.

This opcode must appear in the first 16 bytes of a block of

shader code in order to be recognized by external tools and
functional models. Avoid placing opcodes > 32 bits or encodings
that are not available in all versions of the microcode before
the S_VERSION opcode. If this opcode is absent then tools are
allowed to make a 'best guess' of the microcode version using
cues from the environment; the guess may be incorrect and lead
to an invalid decode. It is highly recommended that this be the
FIRST opcode of a shader block except for trap handlers, where
it should be the SECOND opcode (allowing the first opcode to be
a 32-bit branch to accommodate context switch).

SIMM16[7:0] specifies the microcode version.

SIMM16[15:8] must be set to zero.

2 S_CMOVK_I32 Conditional move with sign extension.

  if(SCC)
  D.i32 = signext(SIMM16[15:0]);
  endif.

3 S_CMPK_EQ_I32   SCC = (S0.i32 == signext(SIMM16[15:0])).

4 S_CMPK_LG_I32   SCC = (S0.i32 != signext(SIMM16[15:0])).

5 S_CMPK_GT_I32   SCC = (S0.i32 > signext(SIMM16[15:0])).

6 S_CMPK_GE_I32   SCC = (S0.i32 >= signext(SIMM16[15:0])).

7 S_CMPK_LT_I32   SCC = (S0.i32 < signext(SIMM16[15:0])).

8 S_CMPK_LE_I32   SCC = (S0.i32 <= signext(SIMM16[15:0])).

9 S_CMPK_EQ_U32   SCC = (S0.u32 == SIMM16[15:0]).

10 S_CMPK_LG_U32   SCC = (S0.u32 != SIMM16[15:0]).

11 S_CMPK_GT_U32   SCC = (S0.u32 > SIMM16[15:0]).

12 S_CMPK_GE_U32   SCC = (S0.u32 >= SIMM16[15:0]).

13 S_CMPK_LT_U32   SCC = (S0.u32 < SIMM16[15:0]).

14 S_CMPK_LE_U32   SCC = (S0.u32 <= SIMM16[15:0]).

12.2. SOPK Instructions 107 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

15 S_ADDK_I32 Add a 16-bit signed constant to the destination.

  int32 tmp = D.i32; // save value so we can check sign bits for
overflow later.
  D.i32 = D.i32 + signext(SIMM16[15:0]);
  SCC = (tmp[31] == SIMM16[15] && tmp[31] != D.i32[31]). // signed
overflow.

16 S_MULK_I32 Multiply a 16-bit signed constant with the destination.

  D.i32 = D.i32 * signext(SIMM16[15:0]).

18 S_GETREG_B32 Read some or all of a hardware register into the LSBs of D.

SIMM16 = {size[4:0], offset[4:0], hwRegId[5:0]}; offset is

0..31, size is 1..32.

  uint32 offset = SIMM16[10:6];

  uint32 size = SIMM16[15:11];
  uint32 id = SIMM16[5:0];
  D.u32 = hardware_reg[id][offset+size-1:offset].

19 S_SETREG_B32 Write some or all of the LSBs of S0 into a hardware register.

SIMM16 = {size[4:0], offset[4:0], hwRegId[5:0]}; offset is

0..31, size is 1..32.

  hardware-reg = S0.u.

21 S_SETREG_IMM32_B Write some or all of the LSBs of IMM32 into a hardware register;
32 this instruction requires a 32-bit literal constant.

SIMM16 = {size[4:0], offset[4:0], hwRegId[5:0]}; offset is

0..31, size is 1..32.

  hardware-reg = LITERAL.

22 S_CALL_B64 Implements a short call, where the return address (the next
instruction after the S_CALL_B64) is saved to D. Long calls
should consider S_SWAPPC_B64 instead.

  D.u64 = PC + 4;
  PC = PC + signext(SIMM16 * 4) + 4.

23 S_WAITCNT_VSCNT Wait for the counts of outstanding vector store events -- vector
memory stores and atomics that DO NOT return data -- to be at or
below the specified level. This counter is not used in
'all-in-order' mode.

Waits for the following condition to hold before continuing:

  vscnt <= S0.u[5:0] + S1.u[5:0].

  // Comparison is 6 bits, no clamping is applied for add overflow

To wait on a literal constant only, write 'null' for the GPR

argument.

See also S_WAITCNT.

12.2. SOPK Instructions 108 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

24 S_WAITCNT_VMCNT Wait for the counts of outstanding vector memory events --

everything except for memory stores and atomics-without-return
-- to be at or below the specified level. When in
'all-in-order' mode, wait for all vector memory events.

Waits for the following condition to hold before continuing:

  vmcnt <= S0.u[5:0] + S1.u[5:0].

  // Comparison is 6 bits, no clamping is applied for add overflow

To wait on a literal constant only, write 'null' for the GPR

argument or use S_WAITCNT.

See also S_WAITCNT.

25 S_WAITCNT_EXPCN Waits for the following condition to hold before continuing:

T
  expcnt <= S0.u[2:0] + S1.u[2:0].
  // Comparison is 3 bits, no clamping is applied for add overflow

To wait on a literal constant only, write 'null' for the GPR

argument or use S_WAITCNT.

See also S_WAITCNT.

26 S_WAITCNT_LGKMC Waits for the following condition to hold before continuing:

NT
  lgkmcnt <= S0.u[5:0] + S1.u[5:0].
  // Comparison is 6 bits, no clamping is applied for add overflow

To wait on a literal constant only, write 'null' for the GPR

argument or use S_WAITCNT.

See also S_WAITCNT.

12.2. SOPK Instructions 109 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

27 S_SUBVECTOR_LOO Begin execution of a subvector block of code.

P_BEGIN See also S_SUBVECTOR_LOOP_END.

  if(EXEC[63:0] == 0)
  // no passes, skip entire loop
  jump LABEL
  elif(EXEC_LO == 0)
  // execute high pass only
  D0 = EXEC_LO
  else
  // execute low pass first, either running both passes or
running low pass only
  D0 = EXEC_HI
  EXEC_HI = 0
  endif.

Example:
  s_subvector_loop_begin s0, SKIP_ALL
  LOOP_START:
  // instructions
  // ...
  LOOP_END:
  s_subvector_loop_end s0, LOOP_START
  SKIP_ALL:

This opcode is intended to be used in conjunction with

S_SUBVECTOR_LOOP_END but there is no dedicated subvector state
and internally it is equivalent to an S_CBRANCH with extra math.
This opcode has well-defined semantics in wave32 mode but the
author of this document is not aware of any practical wave32
programming scenario where it would make sense to use this
opcode.

28 S_SUBVECTOR_LOO End execution of a subvector block of code.

P_END See also S_SUBVECTOR_LOOP_START.

  if(EXEC_HI != 0)
  EXEC_LO = D0
  elif(S0 == 0)
  // done: executed low pass and skip high pass
  nop
  else
  // execute second pass of two-pass mode
  EXEC_HI = D0
  D0 = EXEC_LO
  EXEC_LO = 0
  jump LABEL
  endif.

This opcode is intended to be used in conjunction with

S_SUBVECTOR_LOOP_BEGIN but there is no dedicated subvector state
and internally it is equivalent to an S_CBRANCH with extra math.
This opcode has well-defined semantics in wave32 mode but the
author of this document is not aware of any practical wave32
programming scenario where it would make sense to use this
opcode.

12.2. SOPK Instructions 110 of 283

"RDNA 2" Instruction Set Architecture

12.3. SOP1 Instructions

Instructions in this format may use a 32-bit literal constant which occurs immediately after the
instruction.

Opcode Name Description

3 S_MOV_B32 Move data to an SGPR.

  D.u = S0.u.

4 S_MOV_B64 Move data to an SGPR.

  D.u64 = S0.u64.

5 S_CMOV_B32 Conditionally move data to an SGPR when scalar condition code is

true.

  if(SCC) then
  D.u = S0.u;
  endif.

6 S_CMOV_B64 Conditionally move data to an SGPR when scalar condition code is

true.

  if(SCC) then
  D.u64 = S0.u64;
  endif.

7 S_NOT_B32 Bitwise negation.

  D = ~S0;
  SCC = (D != 0).

8 S_NOT_B64 Bitwise negation.

  D = ~S0;
  SCC = (D != 0).

9 S_WQM_B32 Computes whole quad mode for an active/valid mask. If any pixel
in a quad is active, all pixels of the quad are marked active.

  for i in 0 ... opcode_size_in_bits - 1 do

  D[i] = (S0[(i & ~3):(i | 3)] != 0);
  endfor;
  SCC = (D != 0).

10 S_WQM_B64 Computes whole quad mode for an active/valid mask. If any pixel
in a quad is active, all pixels of the quad are marked active.

  for i in 0 ... opcode_size_in_bits - 1 do

  D[i] = (S0[(i & ~3):(i | 3)] != 0);
  endfor;
  SCC = (D != 0).

11 S_BREV_B32 Reverse bits.

  D.u[31:0] = S0.u[0:31].

12.3. SOP1 Instructions 111 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

12 S_BREV_B64 Reverse bits.

  D.u64[63:0] = S0.u64[0:63].

13 S_BCNT0_I32_B32 Count number of bits that are zero.

  D = 0;
  for i in 0 ... opcode_size_in_bits - 1 do
  D += (S0[i] == 0 ? 1 : 0)
  endfor;
  SCC = (D != 0).

Functional examples:

S_BCNT0_I32_B32(0x00000000) => 32
S_BCNT0_I32_B32(0xcccccccc) => 16
S_BCNT0_I32_B32(0xffffffff) => 0

14 S_BCNT0_I32_B64 Count number of bits that are zero.

  D = 0;
  for i in 0 ... opcode_size_in_bits - 1 do
  D += (S0[i] == 0 ? 1 : 0)
  endfor;
  SCC = (D != 0).

Functional examples:

S_BCNT0_I32_B32(0x00000000) => 32
S_BCNT0_I32_B32(0xcccccccc) => 16
S_BCNT0_I32_B32(0xffffffff) => 0

15 S_BCNT1_I32_B32 Count number of bits that are one.

  D = 0;
  for i in 0 ... opcode_size_in_bits - 1 do
  D += (S0[i] == 1 ? 1 : 0)
  endfor;
  SCC = (D != 0).

Functional examples:

S_BCNT1_I32_B32(0x00000000) => 0
S_BCNT1_I32_B32(0xcccccccc) => 16
S_BCNT1_I32_B32(0xffffffff) => 32

16 S_BCNT1_I32_B64 Count number of bits that are one.

  D = 0;
  for i in 0 ... opcode_size_in_bits - 1 do
  D += (S0[i] == 1 ? 1 : 0)
  endfor;
  SCC = (D != 0).

Functional examples:

S_BCNT1_I32_B32(0x00000000) => 0
S_BCNT1_I32_B32(0xcccccccc) => 16
S_BCNT1_I32_B32(0xffffffff) => 32

12.3. SOP1 Instructions 112 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

17 S_FF0_I32_B32 Returns the bit position of the first zero from the LSB (least
significant bit), or -1 if there are no zeros.

  D.i = -1; // Set if no zeros are found

  for i in 0 ... opcode_size_in_bits - 1 do // Search from LSB
  if S0[i] == 0 then
  D.i = i;
  break for;
  endif;
  endfor.

Functional examples:

S_FF0_I32_B32(0xaaaaaaaa) => 0
S_FF0_I32_B32(0x55555555) => 1
S_FF0_I32_B32(0x00000000) => 0
S_FF0_I32_B32(0xffffffff) => 0xffffffff
S_FF0_I32_B32(0xfffeffff) => 16

18 S_FF0_I32_B64 Returns the bit position of the first zero from the LSB (least
significant bit), or -1 if there are no zeros.

  D.i = -1; // Set if no zeros are found

  for i in 0 ... opcode_size_in_bits - 1 do // Search from LSB
  if S0[i] == 0 then
  D.i = i;
  break for;
  endif;
  endfor.

Functional examples:

S_FF0_I32_B32(0xaaaaaaaa) => 0
S_FF0_I32_B32(0x55555555) => 1
S_FF0_I32_B32(0x00000000) => 0
S_FF0_I32_B32(0xffffffff) => 0xffffffff
S_FF0_I32_B32(0xfffeffff) => 16

19 S_FF1_I32_B32 Returns the bit position of the first one from the LSB (least
significant bit), or -1 if there are no ones.

  D.i = -1; // Set if no ones are found

  for i in 0 ... opcode_size_in_bits - 1 do // Search from LSB
  if S0[i] == 1 then
  D.i = i;
  break for;
  endif;
  endfor.

Functional examples:

S_FF1_I32_B32(0xaaaaaaaa) => 1
S_FF1_I32_B32(0x55555555) => 0
S_FF1_I32_B32(0x00000000) => 0xffffffff
S_FF1_I32_B32(0xffffffff) => 0
S_FF1_I32_B32(0x00010000) => 16

12.3. SOP1 Instructions 113 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

20 S_FF1_I32_B64 Returns the bit position of the first one from the LSB (least
significant bit), or -1 if there are no ones.

  D.i = -1; // Set if no ones are found

  for i in 0 ... opcode_size_in_bits - 1 do // Search from LSB
  if S0[i] == 1 then
  D.i = i;
  break for;
  endif;
  endfor.

Functional examples:

S_FF1_I32_B32(0xaaaaaaaa) => 1
S_FF1_I32_B32(0x55555555) => 0
S_FF1_I32_B32(0x00000000) => 0xffffffff
S_FF1_I32_B32(0xffffffff) => 0
S_FF1_I32_B32(0x00010000) => 16

21 S_FLBIT_I32_B32 Counts how many zeros before the first one starting from the MSB
(most significant bit). Returns -1 if there are no ones.

  D.i = -1; // Set if no ones are found

  for i in 0 ... opcode_size_in_bits - 1 do
  // Note: search is from the MSB
  if S0[opcode_size_in_bits - 1 - i] == 1 then
  D.i = i;
  break for;
  endif;
  endfor.

Functional examples:

S_FLBIT_I32_B32(0x00000000) => 0xffffffff

S_FLBIT_I32_B32(0x0000cccc) => 16
S_FLBIT_I32_B32(0xffff3333) => 0
S_FLBIT_I32_B32(0x7fffffff) => 1
S_FLBIT_I32_B32(0x80000000) => 0
S_FLBIT_I32_B32(0xffffffff) => 0

12.3. SOP1 Instructions 114 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

22 S_FLBIT_I32_B64 Counts how many zeros before the first one starting from the MSB
(most significant bit). Returns -1 if there are no ones.

  D.i = -1; // Set if no ones are found

  for i in 0 ... opcode_size_in_bits - 1 do
  // Note: search is from the MSB
  if S0[opcode_size_in_bits - 1 - i] == 1 then
  D.i = i;
  break for;
  endif;
  endfor.

Functional examples:

S_FLBIT_I32_B32(0x00000000) => 0xffffffff

S_FLBIT_I32_B32(0x0000cccc) => 16
S_FLBIT_I32_B32(0xffff3333) => 0
S_FLBIT_I32_B32(0x7fffffff) => 1
S_FLBIT_I32_B32(0x80000000) => 0
S_FLBIT_I32_B32(0xffffffff) => 0

23 S_FLBIT_I32 Counts how many bits in a row (from MSB to LSB) are the same as
the sign bit. Returns -1 if all bits are the same.

  D.i = -1; // Set if all bits are the same

  for i in 1 ... opcode_size_in_bits - 1 do
  // Note: search is from the MSB
  if S0[opcode_size_in_bits - 1 - i] != S0[opcode_size_in_bits - 1]
then
  D.i = i;
  break for;
  endif;
  endfor.

Functional examples:

S_FLBIT_I32(0x00000000) => 0xffffffff

S_FLBIT_I32(0x0000cccc) => 16
S_FLBIT_I32(0xffff3333) => 16
S_FLBIT_I32(0x7fffffff) => 1
S_FLBIT_I32(0x80000000) => 1
S_FLBIT_I32(0xffffffff) => 0xffffffff

12.3. SOP1 Instructions 115 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

24 S_FLBIT_I32_I64 Counts how many bits in a row (from MSB to LSB) are the same as
the sign bit. Returns -1 if all bits are the same.

  D.i = -1; // Set if all bits are the same

  for i in 1 ... opcode_size_in_bits - 1 do
  // Note: search is from the MSB
  if S0[opcode_size_in_bits - 1 - i] != S0[opcode_size_in_bits - 1]
then
  D.i = i;
  break for;
  endif;
  endfor.

Functional examples:

S_FLBIT_I32(0x00000000) => 0xffffffff

S_FLBIT_I32(0x0000cccc) => 16
S_FLBIT_I32(0xffff3333) => 16
S_FLBIT_I32(0x7fffffff) => 1
S_FLBIT_I32(0x80000000) => 1
S_FLBIT_I32(0xffffffff) => 0xffffffff

25 S_SEXT_I32_I8 Sign extension of a signed byte.

  D.i = signext(S0.i[7:0]).

26 S_SEXT_I32_I16 Sign extension of a signed short.

  D.i = signext(S0.i[15:0]).

27 S_BITSET0_B32 Set a specific bit to zero.

  D.u[S0.u[4:0]] = 0.

28 S_BITSET0_B64 Set a specific bit to zero.

  D.u64[S0.u[5:0]] = 0.

29 S_BITSET1_B32 Set a specific bit to one.

  D.u[S0.u[4:0]] = 1.

30 S_BITSET1_B64 Set a specific bit to one.

  D.u64[S0.u[5:0]] = 1.

31 S_GETPC_B64 Save current program location. Destination receives the byte

address of the next instruction.

  D.u64 = PC + 4.

32 S_SETPC_B64 Jump to a new location. S0.u64 is a byte address of the

instruction to jump to.

  PC = S0.u64.

12.3. SOP1 Instructions 116 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

33 S_SWAPPC_B64 Save current program location and jump to a new location.

S0.u64 is a byte address of the instruction to jump to.
Destination receives the byte address of the instruction
immediately following the SWAPPC instruction.

  D.u64 = PC + 4;
  PC = S0.u64.

34 S_RFE_B64 Return from exception handler and continue. This instruction

may only be used within a trap handler.

  PRIV = 0;
  PC = S0.u64.

36 S_AND_SAVEEXEC_ Bitwise AND with EXEC mask. The original EXEC mask is saved to
B64 the destination SGPRs before the bitwise operation is
performed.

  D.u64 = EXEC;
  EXEC = S0.u64 & EXEC;
  SCC = (EXEC != 0).

37 S_OR_SAVEEXEC_B Bitwise OR with EXEC mask. The original EXEC mask is saved to
64 the destination SGPRs before the bitwise operation is
performed.

  D.u64 = EXEC;
  EXEC = S0.u64 | EXEC;
  SCC = (EXEC != 0).

38 S_XOR_SAVEEXEC_ Bitwise XOR with EXEC mask. The original EXEC mask is saved to
B64 the destination SGPRs before the bitwise operation is
performed.

  D.u64 = EXEC;
  EXEC = S0.u64 ^ EXEC;
  SCC = (EXEC != 0).

39 S_ANDN2_SAVEEXE Bitwise ANDN2 with EXEC mask. The original EXEC mask is saved
C_B64 to the destination SGPRs before the bitwise operation is
performed.

  D.u64 = EXEC;
  EXEC = S0.u64 & ~EXEC;
  SCC = (EXEC != 0).

40 S_ORN2_SAVEEXEC Bitwise ORN2 with EXEC mask. The original EXEC mask is saved to
_B64 the destination SGPRs before the bitwise operation is
performed.

  D.u64 = EXEC;
  EXEC = S0.u64 | ~EXEC;
  SCC = (EXEC != 0).

12.3. SOP1 Instructions 117 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

41 S_NAND_SAVEEXEC Bitwise NAND with EXEC mask. The original EXEC mask is saved to
_B64 the destination SGPRs before the bitwise operation is
performed.

  D.u64 = EXEC;
  EXEC = ~(S0.u64 & EXEC);
  SCC = (EXEC != 0).

42 S_NOR_SAVEEXEC_ Bitwise NOR with EXEC mask. The original EXEC mask is saved to
B64 the destination SGPRs before the bitwise operation is
performed.

  D.u64 = EXEC;
  EXEC = ~(S0.u64 | EXEC);
  SCC = (EXEC != 0).

43 S_XNOR_SAVEEXEC Bitwise XNOR with EXEC mask. The original EXEC mask is saved to
_B64 the destination SGPRs before the bitwise operation is
performed.

  D.u64 = EXEC;
  EXEC = ~(S0.u64 ^ EXEC);
  SCC = (EXEC != 0).

44 S_QUADMASK_B32 Reduce a pixel mask to a quad mask. To perform the inverse

operation see S_BITREPLICATE_B64_B32.

  D = 0;
  for i in 0 ... (opcode_size_in_bits / 4) - 1 do
  D[i] = (S0[i * 4 + 3:i * 4] != 0);
  endfor;
  SCC = (D != 0).

45 S_QUADMASK_B64 Reduce a pixel mask to a quad mask. To perform the inverse

operation see S_BITREPLICATE_B64_B32.

  D = 0;
  for i in 0 ... (opcode_size_in_bits / 4) - 1 do
  D[i] = (S0[i * 4 + 3:i * 4] != 0);
  endfor;
  SCC = (D != 0).

46 S_MOVRELS_B32 Move from a relative source address.

  SGPR[D.addr].u32 = SGPR[S0.addr+M0[31:0]].u32

Example: The following instruction sequence will perform a move

s5 <== s17:
  s_mov_b32 m0, 10
  s_movrels_b32 s5, s7

47 S_MOVRELS_B64 Move from a relative source address. The index in M0.u must be
even for this operation.

  SGPR[D.addr].u64 = SGPR[S0.addr+M0[31:0]].u64

12.3. SOP1 Instructions 118 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

48 S_MOVRELD_B32 Move to a relative destination address.

  SGPR[D.addr+M0[31:0]].u32 = SGPR[S0.addr].u32

Example: The following instruction sequence will perform a move

s15 <== s7:
  s_mov_b32 m0, 10
  s_movreld_b32 s5, s7

49 S_MOVRELD_B64 Move to a relative destination address. The index in M0.u must

be even for this operation.

  SGPR[D.addr+M0[31:0]].u64 = SGPR[S0.addr].u64

52 S_ABS_I32 Integer absolute value.

  D.i = (S.i < 0 ? -S.i : S.i);

  SCC = (D.i != 0).

Functional examples:

S_ABS_I32(0x00000001) => 0x00000001

S_ABS_I32(0x7fffffff) => 0x7fffffff
S_ABS_I32(0x80000000) => 0x80000000 // Note this is negative!
S_ABS_I32(0x80000001) => 0x7fffffff
S_ABS_I32(0x80000002) => 0x7ffffffe
S_ABS_I32(0xffffffff) => 0x00000001

55 S_ANDN1_SAVEEXE Bitwise ANDN1 with EXEC mask. The original EXEC mask is saved
C_B64 to the destination SGPRs before the bitwise operation is
performed.

  D.u64 = EXEC;
  EXEC = ~S0.u64 & EXEC;
  SCC = (EXEC != 0).

56 S_ORN1_SAVEEXEC Bitwise ORN1 with EXEC mask. The original EXEC mask is saved to
_B64 the destination SGPRs before the bitwise operation is
performed.

  D.u64 = EXEC;
  EXEC = ~S0.u64 | EXEC;
  SCC = (EXEC != 0).

57 S_ANDN1_WREXEC_ Bitwise ANDN1 with EXEC mask. Unlike the SAVEEXEC series of
B64 opcodes, the value written to destination SGPRs is the result of
the bitwise-op result. EXEC and the destination SGPRs will have
the same value at the end of this instruction. This instruction
is intended to accelerate waterfalling.

  EXEC = ~S0.u64 & EXEC;

  D.u64 = EXEC;
  SCC = (EXEC != 0).

12.3. SOP1 Instructions 119 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

58 S_ANDN2_WREXEC_ Bitwise ANDN2 with EXEC mask. Unlike the SAVEEXEC series of
B64 opcodes, the value written to destination SGPRs is the result of
the bitwise-op result. EXEC and the destination SGPRs will have
the same value at the end of this instruction. This instruction
is intended to accelerate waterfalling.

  EXEC = S0.u64 & ~EXEC;

  D.u64 = EXEC;
  SCC = (EXEC != 0).

In particular, the following sequence of waterfall code is

optimized by using a WREXEC instead of two separate scalar
ops:

  // V0 holds the index value per lane

  // save exec mask for restore at the end
  s_mov_b64 s2, exec
  // exec mask of remaining (unprocessed) threads
  s_mov_b64 s4, exec
  loop:
  // get the index value for the first active lane
  v_readfirstlane_b32 s0, v0
  // find all other lanes with same index value
  v_cmpx_eq s0, v0
  <OP> // do the operation using the current EXEC mask. S0 holds
the index.
  // mask out thread that was just executed
  // s_andn2_b64 s4, s4, exec
  // s_mov_b64 exec, s4
  s_andn2_wrexec_b64 s4, s4 // replaces above 2 ops
  // repeat until EXEC==0
  s_cbranch_scc1 loop
  s_mov_b64 exec, s2

59 S_BITREPLICATE_B6 Replicate the low 32 bits of S0 by 'doubling' each bit.

4_B32
  for i in 0 ... 31 do
  D.u64[i * 2 + 0] = S0.u32[i]
  D.u64[i * 2 + 1] = S0.u32[i]
  endfor.

This opcode can be used to convert a quad mask into a pixel

mask; given quad mask in s0, the following sequence will produce
a pixel mask in s2:
  s_bitreplicate_b64 s2, s0
  s_bitreplicate_b64 s2, s2

To perform the inverse operation see S_QUADMASK_B64.

60 S_AND_SAVEEXEC_ Bitwise AND with EXEC mask. The original EXEC mask is saved to
B32 the destination SGPRs before the bitwise operation is
performed.

  D.u32 = EXEC_LO;
  EXEC_LO = S0.u32 & EXEC_LO;
  SCC = (EXEC_LO != 0).

12.3. SOP1 Instructions 120 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

61 S_OR_SAVEEXEC_B Bitwise OR with EXEC mask. The original EXEC mask is saved to
32 the destination SGPRs before the bitwise operation is
performed.

  D.u32 = EXEC_LO;
  EXEC_LO = S0.u32 | EXEC_LO;
  SCC = (EXEC_LO != 0).

62 S_XOR_SAVEEXEC_ Bitwise XOR with EXEC mask. The original EXEC mask is saved to
B32 the destination SGPRs before the bitwise operation is
performed.

  D.u32 = EXEC_LO;
  EXEC_LO = S0.u32 ^ EXEC_LO;
  SCC = (EXEC_LO != 0).

63 S_ANDN2_SAVEEXE Bitwise ANDN2 with EXEC mask. The original EXEC mask is saved
C_B32 to the destination SGPRs before the bitwise operation is
performed.

  D.u32 = EXEC_LO;
  EXEC_LO = S0.u32 & ~EXEC_LO;
  SCC = (EXEC_LO != 0).

64 S_ORN2_SAVEEXEC Bitwise ORN2 with EXEC mask. The original EXEC mask is saved to
_B32 the destination SGPRs before the bitwise operation is
performed.

  D.u32 = EXEC_LO;
  EXEC_LO = S0.u32 | ~EXEC_LO;
  SCC = (EXEC_LO != 0).

65 S_NAND_SAVEEXEC Bitwise NAND with EXEC mask. The original EXEC mask is saved to
_B32 the destination SGPRs before the bitwise operation is
performed.

  D.u32 = EXEC_LO;
  EXEC_LO = ~(S0.u32 & EXEC_LO);
  SCC = (EXEC_LO != 0).

66 S_NOR_SAVEEXEC_ Bitwise NOR with EXEC mask. The original EXEC mask is saved to
B32 the destination SGPRs before the bitwise operation is
performed.

  D.u32 = EXEC_LO;
  EXEC_LO = ~(S0.u32 | EXEC_LO);
  SCC = (EXEC_LO != 0).

67 S_XNOR_SAVEEXEC Bitwise XNOR with EXEC mask. The original EXEC mask is saved to
_B32 the destination SGPRs before the bitwise operation is
performed.

  D.u32 = EXEC_LO;
  EXEC_LO = ~(S0.u32 ^ EXEC_LO);
  SCC = (EXEC_LO != 0).

12.3. SOP1 Instructions 121 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

68 S_ANDN1_SAVEEXE Bitwise ANDN1 with EXEC mask. The original EXEC mask is saved
C_B32 to the destination SGPRs before the bitwise operation is
performed.

  D.u32 = EXEC_LO;
  EXEC_LO = ~S0.u32 & EXEC_LO;
  SCC = (EXEC_LO != 0).

69 S_ORN1_SAVEEXEC Bitwise ORN1 with EXEC mask. The original EXEC mask is saved to
_B32 the destination SGPRs before the bitwise operation is
performed.

  D.u32 = EXEC_LO;
  EXEC_LO = ~S0.u32 | EXEC_LO;
  SCC = (EXEC_LO != 0).

70 S_ANDN1_WREXEC_ Bitwise ANDN1 with EXEC mask. Unlike the SAVEEXEC series of
B32 opcodes, the value written to destination SGPRs is the result of
the bitwise-op result. EXEC and the destination SGPRs will have
the same value at the end of this instruction. This instruction
is intended to accelerate waterfalling.

  EXEC_LO = ~S0.u32 & EXEC_LO;

  D.u32 = EXEC_LO;
  SCC = (EXEC_LO != 0).

71 S_ANDN2_WREXEC_ Bitwise ANDN2 with EXEC mask. Unlike the SAVEEXEC series of
B32 opcodes, the value written to destination SGPRs is the result of
the bitwise-op result. EXEC and the destination SGPRs will have
the same value at the end of this instruction. This instruction
is intended to accelerate waterfalling. See S_ANDN2_WREXEC_B64
for example code.

  EXEC_LO = S0.u32 & ~EXEC_LO;

  D.u32 = EXEC_LO;
  SCC = (EXEC_LO != 0).

73 S_MOVRELSD_2_B32 Move from a relative source address to a relative destination

address, with different offsets.

  SGPR[D.addr+M0[25:16]].u32 = SGPR[S0.addr+M0[9:0]].u32

Example: The following instruction sequence will perform a move

s25 <== s17:
  s_mov_b32 m0, ((20 << 16) | 10)
  s_movrelsd_2_b32 s5, s7

12.4. SOPC Instructions

Instructions in this format may use a 32-bit literal constant which occurs immediately after the
instruction.

12.4. SOPC Instructions 122 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

0 S_CMP_EQ_I32 Compare two integers for equality. Note that S_CMP_EQ_I32 and
S_CMP_EQ_U32 are identical opcodes, but both are provided for
symmetry.

  SCC = (S0 == S1).

1 S_CMP_LG_I32 Compare two integers for inequality. Note that S_CMP_LG_I32 and
S_CMP_LG_U32 are identical opcodes, but both are provided for
symmetry.

  SCC = (S0 != S1).

2 S_CMP_GT_I32   SCC = (S0.i > S1.i).

3 S_CMP_GE_I32   SCC = (S0.i >= S1.i).

4 S_CMP_LT_I32   SCC = (S0.i < S1.i).

5 S_CMP_LE_I32   SCC = (S0.i <= S1.i).

6 S_CMP_EQ_U32 Compare two integers for equality. Note that S_CMP_EQ_I32 and
S_CMP_EQ_U32 are identical opcodes, but both are provided for
symmetry.

  SCC = (S0 == S1).

7 S_CMP_LG_U32 Compare two integers for inequality. Note that S_CMP_LG_I32 and
S_CMP_LG_U32 are identical opcodes, but both are provided for
symmetry.

  SCC = (S0 != S1).

8 S_CMP_GT_U32   SCC = (S0.u > S1.u).

9 S_CMP_GE_U32   SCC = (S0.u >= S1.u).

10 S_CMP_LT_U32   SCC = (S0.u < S1.u).

11 S_CMP_LE_U32   SCC = (S0.u <= S1.u).

12 S_BITCMP0_B32   SCC = (S0.u[S1.u[4:0]] == 0).

13 S_BITCMP1_B32   SCC = (S0.u[S1.u[4:0]] == 1).

14 S_BITCMP0_B64   SCC = (S0.u64[S1.u[5:0]] == 0).

15 S_BITCMP1_B64   SCC = (S0.u64[S1.u[5:0]] == 1).

18 S_CMP_EQ_U64   SCC = (S0.i64 == S1.i64).

19 S_CMP_LG_U64   SCC = (S0.i64 != S1.i64).

12.5. SOPP Instructions

12.5. SOPP Instructions 123 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

0 S_NOP Do nothing. Repeat NOP 1..16 times based on SIMM16[3:0] -- 0x0 =

1 time, 0xf = 16 times.

Examples:
  s_nop 0 // Wait 1 cycle.
  s_nop 0xf // Wait 16 cycles.

1 S_ENDPGM End of program; terminate wavefront. The hardware implicitly

executes S_WAITCNT 0 and S_WAITCNT_VSCNT 0 before executing this
instruction. See S_ENDPGM_SAVED for the context-switch version
of this instruction and S_ENDPGM_ORDERED_PS_DONE for the POPS
critical region version of this instruction.

2 S_BRANCH Perform an unconditional short jump. For a long jump, use

S_SETPC_B64.

  PC = PC + signext(SIMM16 * 4) + 4. // short jump.

Examples:
  s_branch label // Set SIMM16 = +4 = 0x0004
  s_nop 0 // 4 bytes
label:
  s_nop 0 // 4 bytes
  s_branch label // Set SIMM16 = -8 = 0xfff8

3 S_WAKEUP Allow a wave to 'ping' all the other waves in its threadgroup to
force them to wake up early from an S_SLEEP instruction.
The ping is ignored if the waves are not sleeping. This allows
for efficient polling on a memory location. The waves which
are polling can sit in a long S_SLEEP between memory reads, but
the wave which writes the value can tell them all to wake up
early now that the data is available. This is useful for
fBarrier implementations (speedup). This method is also safe
from races because if any wave misses the ping, everything still
works fine (waves which missed it just complete their S_SLEEP).

If the wave executing S_WAKEUP is in a threadgroup (in_tg set),

then it will wake up all waves associated with the same
threadgroup ID. Otherwise, S_WAKEUP is treated as an S_NOP.

4 S_CBRANCH_SCC0 Perform a conditional short jump when SCC is zero.

  if(SCC == 0) then
  PC = PC + signext(SIMM16 * 4) + 4;
  endif.

5 S_CBRANCH_SCC1 Perform a conditional short jump when SCC is one.

  if(SCC == 1) then
  PC = PC + signext(SIMM16 * 4) + 4;
  endif.

6 S_CBRANCH_VCCZ Perform a conditional short jump when VCC is zero.

  if(VCC == 0) then
  PC = PC + signext(SIMM16 * 4) + 4;
  endif.

12.5. SOPP Instructions 124 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

7 S_CBRANCH_VCCNZ Perform a conditional short jump when VCC is nonzero.

  if(VCC != 0) then
  PC = PC + signext(SIMM16 * 4) + 4;
  endif.

8 S_CBRANCH_EXECZ Perform a conditional short jump when EXEC is zero.

  if(EXEC == 0) then
  PC = PC + signext(SIMM16 * 4) + 4;
  endif.

9 S_CBRANCH_EXECNZ Perform a conditional short jump when EXEC is nonzero.

  if(EXEC != 0) then
  PC = PC + signext(SIMM16 * 4) + 4;
  endif.

10 S_BARRIER Synchronize waves within a threadgroup. If not all waves of the

threadgroup have been created yet, waits for entire group before
proceeding. If some waves in the threadgroup have already
terminated, this waits on only the surviving waves. Barriers
are legal inside trap handlers.

11 S_SETKILL Set KILL bit to value of SIMM16[0]. Used primarily for

debugging kill wave host command behavior.

12 S_WAITCNT Wait for the counts of outstanding lds, vector-memory and

export/vmem-write-data to be at or below the specified levels.

Waits for all of the following conditions to hold before continuing:

  vmcnt <= {SIMM16[15:14], SIMM16[3:0]}

  expcnt <= SIMM16[6:4]
  lgkmcnt <= SIMM16[13:8]

NOTE: VMCNT only counts vector memory loads, image sample

instructions,
and vector memory atomics that return data. Contrast with the VSCNT
counter.

See also S_WAITCNT_VSCNT.

13 S_SETHALT S_SETHALT can set/clear the HALT or FATAL_HALT status bits. The
particular status bit is chosen by halt type control as
indicated in SIMM16[2]; 0 = HALT bit select; 1 = FATAL_HALT bit
select.

When halt type control is set to 0 = HALT bit select:

Set HALT bit to value of SIMM16[0]; 1 = halt, 0 = clear HALT bit.
The
halt flag is ignored while PRIV == 1 (inside trap handlers) but
the shader will halt after the handler returns if
HALT is still set at that time.

When halt type control is set to 1 = FATAL HALT bit select:

Set FATAL_HALT bit to value of SIMM16[0]; 1 = fatal_halt, 0 = clear
FATAL_HALT bit. Setting the fatal_halt flag halts the shader in
or outside of the trap handlers.

12.5. SOPP Instructions 125 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

14 S_SLEEP Cause a wave to sleep for (64*(SIMM16[6:0]-1) ..

64*SIMM16[6:0]) clocks. The exact amount of delay is
approximate. Compare with S_NOP. When SIMM16[6:0] is zero then
no sleep occurs.

Examples:
  s_sleep 0 // Wait for 0 clocks.
  s_sleep 1 // Wait for 1-64 clocks.
  s_sleep 2 // Wait for 65-128 clocks.

15 S_SETPRIO User settable wave priority is set to SIMM16[1:0]. 0 = lowest, 3

= highest. The overall wave priority is {SPIPrio[1:0] +
UserPrio[1:0], WaveAge[3:0]}.

16 S_SENDMSG Send a message upstream to VGT or the interrupt handler.

SIMM16[9:0] contains the message type.

17 S_SENDMSGHALT Send a message and then HALT the wavefront; see S_SENDMSG for
details.

18 S_TRAP Enter the trap handler. This instruction may be generated

internally as well in response to a host trap (HT = 1) or an
exception. TrapID 0 is reserved for hardware use and should not
be used in a shader-generated trap.

  TrapID = SIMM16[7:0];
  Wait for all instructions to complete;
  {TTMP1, TTMP0} = {1'h0, PCRewind[5:0], HT[0], TrapID[7:0],
PC[47:0]};
  PC = TBA; // trap base address
  PRIV = 1.

19 S_ICACHE_INV Invalidate entire L0 instruction cache.

The hardware invalidates the instruction buffer,

so no S_NOP instructions are required after S_ICACHE_INV.

20 S_INCPERFLEVEL Increment performance counter specified in SIMM16[3:0] by 1.

21 S_DECPERFLEVEL Decrement performance counter specified in SIMM16[3:0] by 1.

22 S_TTRACEDATA Send M0 as user data to the thread trace stream.

23 S_CBRANCH_CDBGSY Perform a conditional short jump when the system debug flag is
S set.

  if(conditional_debug_system != 0) then
  PC = PC + signext(SIMM16 * 4) + 4;
  endif.

24 S_CBRANCH_CDBGUS Perform a conditional short jump when the user debug flag is
ER set.

  if(conditional_debug_user != 0) then
  PC = PC + signext(SIMM16 * 4) + 4;
  endif.

12.5. SOPP Instructions 126 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

25 S_CBRANCH_CDBGSY Perform a conditional short jump when either the system or the
S_OR_USER user debug flag are set.

  if(conditional_debug_system || conditional_debug_user) then

  PC = PC + signext(SIMM16 * 4) + 4;
  endif.

26 S_CBRANCH_CDBGSY Perform a conditional short jump when both the system and the
S_AND_USER user debug flag are set.

  if(conditional_debug_system && conditional_debug_user) then

  PC = PC + signext(SIMM16 * 4) + 4;
  endif.

27 S_ENDPGM_SAVED End of program; signal that a wave has been saved by the
context-switch trap handler and terminate wavefront. The
hardware implicitly executes S_WAITCNT 0 and S_WAITCNT_VSCNT 0
before executing this instruction. See S_ENDPGM for additional
variants.

30 S_ENDPGM_ORDERED End of program; signal that a wave has exited its POPS critical
_PS_DONE section and terminate wavefront. The hardware implicitly
executes S_WAITCNT 0 and S_WAITCNT_VSCNT 0 before executing this
instruction. This instruction is an optimization that combines
S_SENDMSG(MSG_ORDERED_PS_DONE) and S_ENDPGM; there may be cases
where you still need to send the message separately, in which
case the shader must end with a regular S_ENDPGM instruction.
See S_ENDPGM for additional variants.

31 S_CODE_END Generate an illegal instruction interrupt.

This instruction should NEVER appear in typical shader code. It

is used to pad the end of a shader program to make it easier for
analysis programs to locate the end of a shader program buffer.
Use of this opcode in an embedded shader block may cause
analysis tools to fail.

To unambiguously mark the end of a shader buffer, this

instruction must be specified five times in a row (total of 20
bytes) and analysis tools must ensure the opcode occurs at least
five times to be certain they are at the end of the buffer. This
is because the bit pattern generated by this opcode could
incidentally appear in a valid instruction's second dword,
literal constant or as part of a multi-DWORD image
instruction.

In short: do not embed this opcode in the middle of a valid

shader program. DO use this opcode 5 times at the end of a
shader program to clearly mark the end of the program.

Example:
  ...
  s_endpgm // last real instruction in shader buffer
  s_code_end // 1
  s_code_end // 2
  s_code_end // 3
  s_code_end // 4
  s_code_end // done!

12.5. SOPP Instructions 127 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

32 S_INST_PREFETCH Change instruction prefetch mode.

SIMM16[1:0] specifies the prefetch mode to switch to.

Defined prefetch modes are:

  0: reserved
  1: SQ_PREFETCH_1_LINE -- prefetch 1 line
  2: SQ_PREFETCH_2_LINES -- prefetch 2 lines
  3: SQ_PREFETCH_3_LINES -- prefetch 3 lines

SIMM16[15:2] must be set to zero.

33 S_CLAUSE Mark the beginning of a clause. The next instruction determines

the clause type, which may be one of the following types.

  Texture, Buffer, Global, Scratch (clause may not mix atomics,

loads & stores)
Flat (loads, stores and atomics may not be combined in a clause)
  LDS
  SMEM
  VALU

Halting and killing a wave will break the clause.

The clause length is: (SIMM16[5:0] + 1), and clauses must be 2

instructions or longer and no more than 63 instructions.

SIMM16[11:8] determines the number of instructions per clause

break, in the range 0..15. If SIMM16[11:8] == 0 then there are
no clause breaks.

The following instruction types cannot appear in a clause:

  SALU
  Export
  Branch
  Message
  GDS

12.5. SOPP Instructions 128 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

35 S_WAITCNT_DEPCTR Bit mask of which dependency counters to wait to be zero,

intended for debug and bug-workarounds.

Waits for all of the following conditions to hold before continuing:

  va_vdst <= SIMM16[15:12] || SIMM16[15:12] == 0xf

  va_sdst <= SIMM16[11:9] || SIMM16[11:9] == 0x7
  va_ssrc == 0 || SIMM16[8] == 1
  hold_cnt == 0 || SIMM16[7] == 1
  vm_vsrc <= SIMM16[4:2] || SIMM16[4:2] == 0x7
  va_vcc == 0 || SIMM16[1] == 1
  sa_sdst == 0 || SIMM16[0] == 1

Some wait values are smaller than the counters: the max "wait"
value means "don't wait on this counter". For example, VM_VSRC
is 4 bits, but the wait field for VM_VSRC is only 3 bits. The
value 7 means don't wait on VM_VSRC, 6 means wait for VM_VSRC <=
6, etc. The wait value for VA_VCC is just 1 bit even though the
counter is 3 bits: 0 = wait for va_vcc==0, 1 = don't wait on
va_vcc.

36 S_ROUND_MODE Set floating point round mode using an immediate constant.

Avoids wait state penalty that would be imposed by S_SETREG.

37 S_DENORM_MODE Set floating point denormal mode using an immediate constant.

Avoids wait state penalty that would be imposed by S_SETREG.

40 S_TTRACEDATA_IMM Send SIMM16[7:0] as user data to the thread trace stream.

12.5.1. Send Message

The S_SENDMSG instruction encodes the message type in M0, and can also send data from
the SIMM16 field and in some cases from EXEC.

Message SIMM16[3:0] SIMM16[6:4] Payload

none 0 - illegal

GS 2 0=nop, 1=cut, GS output. M0[4:0]=gs-waveID, SIMM[9:8] = stream-id

2=emit,
GS-done 3 3=emit-cut

Save wave 4 - used in context switching

Stall Wave 5 - stop new wave generation

Gen

Halt Waves 6 - halt all running waves of this vmid

Ordered PS 7 - POPS ordered section done

Done

GS alloc req 9 - Request GS space in parameter cache. M0[9:0] = number of

vertices, M0[22:12] = number of primitives.

12.5. SOPP Instructions 129 of 283

"RDNA 2" Instruction Set Architecture

12.6. SMEM Instructions

Opcode Name Description

0 S_LOAD_DWORD Read 1 dword from scalar data cache.

If the offset is specified as an SGPR, the SGPR contains

an UNSIGNED BYTE offset (the 2 LSBs are ignored).

If the offset is specified as an immediate 21-bit

constant, the constant is a SIGNED BYTE offset.

1 S_LOAD_DWORDX2 Read 2 dwords from scalar data cache. See S_LOAD_DWORD

for details on the offset input.

2 S_LOAD_DWORDX4 Read 4 dwords from scalar data cache. See S_LOAD_DWORD

for details on the offset input.

3 S_LOAD_DWORDX8 Read 8 dwords from scalar data cache. See S_LOAD_DWORD

for details on the offset input.

4 S_LOAD_DWORDX16 Read 16 dwords from scalar data cache. See S_LOAD_DWORD

for details on the offset input.

8 S_BUFFER_LOAD_DWORD Read 1 dword from scalar data cache. See S_LOAD_DWORD

for details on the offset input.

9 S_BUFFER_LOAD_DWORDX Read 2 dwords from scalar data cache. See S_LOAD_DWORD

2 for details on the offset input.

10 S_BUFFER_LOAD_DWORDX Read 4 dwords from scalar data cache. See S_LOAD_DWORD

4 for details on the offset input.

11 S_BUFFER_LOAD_DWORDX Read 8 dwords from scalar data cache. See S_LOAD_DWORD

8 for details on the offset input.

12 S_BUFFER_LOAD_DWORDX Read 16 dwords from scalar data cache. See S_LOAD_DWORD

16 for details on the offset input.

31 S_GL1_INV Invalidate the GL1 cache only.

32 S_DCACHE_INV Invalidate the scalar data L0 cache.

36 S_MEMTIME Return current 64-bit timestamp.

37 S_MEMREALTIME Return current 64-bit RTC.

38 S_ATC_PROBE Probe or prefetch an address into the SQC data cache.

39 S_ATC_PROBE_BUFFER Probe or prefetch an address into the SQC data cache.

12.7. VOP2 Instructions

12.6. SMEM Instructions 130 of 283

"RDNA 2" Instruction Set Architecture

Instructions in this format may use a 32-bit literal constant, DPP or SDWA which occurs
immediately after the instruction.

Opcode Name Description

1 V_CNDMASK_B32 Conditional mask on each thread. In VOP3 the VCC source may be a
scalar GPR specified in S2.u.

Floating-point modifiers are valid for this instruction if S0.u

and S1.u are 32-bit floating point values. This instruction is
suitable for negating or taking the absolute value of a
floating-point value.

  D.u32 = VCC ? S1.u32 : S0.u32.

2 V_DOT2C_F32_F16 Dot product of packed FP16 values, accumulate with

destination.

  D.f32 =
  S0.f16[0] * S1.f16[0] +
  S0.f16[1] * S1.f16[1] + D.f32.

3 V_ADD_F32 Add two single-precision values. 0.5ULP precision, denormals

are supported.

  D.f32 = S0.f32 + S1.f32.

4 V_SUB_F32 Subtract the second single-precision input from the first

input.

  D.f32 = S0.f32 - S1.f32.

5 V_SUBREV_F32 Subtract the first single-precision input from the second

input.

  D.f32 = S1.f32 - S0.f32.

6 V_FMAC_LEGACY_F3 Multiply two single-precision values and accumulate the result

2 with the destination. Follows DX9 rules where 0.0 times
anything produces 0.0 (this is not IEEE compliant).

  D.f32 = S0.f32 * S1.f32 + S2.f32. // DX9 rules, 0.0 * x = 0.0

7 V_MUL_LEGACY_F32 Multiply two single-precision values. Follows DX9 rules where

0.0 times anything produces 0.0 (this is not IEEE compliant).

  D.f32 = S0.f32 * S1.f32. // DX9 rules, 0.0*x = 0.0

8 V_MUL_F32 Multiply two single-precision values. 0.5ULP precision,

denormals are supported.

  D.f32 = S0.f32 * S1.f32.

9 V_MUL_I32_I24 Multiply two signed 24-bit integers and store the result as a
signed 32-bit integer. This opcode is as efficient as basic
single-precision opcodes since it utilizes the single-precision
floating point multiplier. See also V_MUL_HI_I32_I24.

  D.i32 = S0.i24 * S1.i24.

12.7. VOP2 Instructions 131 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

10 V_MUL_HI_I32_I24 Multiply two signed 24-bit integers and store the high 32 bits
of the result as a signed 32-bit integer. See also
V_MUL_I32_I24.

  D.i32 = (S0.i24 * S1.i24)>>32;

11 V_MUL_U32_U24 Multiply two unsigned 24-bit integers and store the result as an
unsigned 32-bit integer. This opcode is as efficient as basic
single-precision opcodes since it utilizes the single-precision
floating point multiplier. See also V_MUL_HI_U32_U24.

  D.u32 = S0.u24 * S1.u24.

12 V_MUL_HI_U32_U24 Multiply two unsigned 24-bit integers and store the high 32 bits
of the result as an unsigned 32-bit integer. See also
V_MUL_U32_U24.

  D.u32 = (S0.u24 * S1.u24)>>32.

13 V_DOT4C_I32_I8 Dot product of packed byte values, accumulate with

destination.

  D.i32 =
  S0.i8[0] * S1.i8[0] +
  S0.i8[1] * S1.i8[1] +
  S0.i8[2] * S1.i8[2] +
  S0.i8[3] * S1.i8[3] + D.i32.

15 V_MIN_F32 Compute the minimum of two single-precision floats.

  D.f32 = min(S0.f32,S1.f32);
  if (IEEE_MODE && S0.f == sNaN)
  D.f = Quiet(S0.f);
  else if (IEEE_MODE && S1.f == sNaN)
  D.f = Quiet(S1.f);
  else if (S0.f == NaN)
  D.f = S1.f;
  else if (S1.f == NaN)
  D.f = S0.f;
  else if (S0.f == +0.0 && S1.f == -0.0)
  D.f = S1.f;
  else if (S0.f == -0.0 && S1.f == +0.0)
  D.f = S0.f;
  else
  // Note: there's no IEEE special case here like there is for
V_MAX_F32.
  D.f = (S0.f < S1.f ? S0.f : S1.f);
  endif.

12.7. VOP2 Instructions 132 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

16 V_MAX_F32 Compute the maximum of two single-precision floats.

  D.f32 = max(S0.f32,S1.f32);
  if (IEEE_MODE && S0.f == sNaN)
  D.f = Quiet(S0.f);
  else if (IEEE_MODE && S1.f == sNaN)
  D.f = Quiet(S1.f);
  else if (S0.f == NaN)
  D.f = S1.f;
  else if (S1.f == NaN)
  D.f = S0.f;
  else if (S0.f == +0.0 && S1.f == -0.0)
  D.f = S0.f;
  else if (S0.f == -0.0 && S1.f == +0.0)
  D.f = S1.f;
  else if (IEEE_MODE)
  D.f = (S0.f >= S1.f ? S0.f : S1.f);
  else
  D.f = (S0.f > S1.f ? S0.f : S1.f);
  endif.

17 V_MIN_I32 Compute the minimum of two signed integers.

D.i32 = (S0.i32 < S1.i32 ? S0.i32 : S1.i32).

18 V_MAX_I32 Compute the maximum of two signed integers.

  D.i32 = (S0.i32 >= S1.i32 ? S0.i32 : S1.i32).

19 V_MIN_U32 Compute the minimum of two unsigned integers.

  D.u32 = (S0.u32 < S1.u32 ? S0.u32 : S1.u32).

20 V_MAX_U32 Compute the maximum of two unsigned integers.

  D.u32 = (S0.u32 >= S1.u32 ? S0.u32 : S1.u32).

22 V_LSHRREV_B32 Logical shift right with shift count in the first operand.

  D.u32 = S1.u32 >> S0[4:0].

24 V_ASHRREV_I32 Arithmetic shift right (preserve sign bit) with shift count in
the first operand.

  D.i32 = S1.i32 >> S0[4:0].

26 V_LSHLREV_B32 Logical shift left with shift count in the first operand.

  D.u32 = S1.u32 << S0[4:0].

27 V_AND_B32 Bitwise AND. Input and output modifiers not supported.

  D.u32 = S0.u32 & S1.u32.

28 V_OR_B32 Bitwise OR. Input and output modifiers not supported.

  D.u32 = S0.u32 | S1.u32.

29 V_XOR_B32 Bitwise XOR. Input and output modifiers not supported.

  D.u32 = S0.u32 ^ S1.u32.

12.7. VOP2 Instructions 133 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

30 V_XNOR_B32 Bitwise XNOR. Input and output modifiers not supported.

  D.u32 = ~(S0.u32 ^ S1.u32).

37 V_ADD_NC_U32 Add two unsigned integers. No carry-in or carry-out.

  D.u32 = S0.u32 + S1.u32.

38 V_SUB_NC_U32 Subtract the second unsigned integer from the first unsigned
integer. No carry-in or carry-out.

  D.u32 = S0.u32 - S1.u32.

39 V_SUBREV_NC_U32 Subtract the first unsigned integer from the second unsigned
integer. No carry-in or carry-out.

  D.u32 = S1.u32 - S0.u32.

40 V_ADD_CO_CI_U32 Add two unsigned integers and a carry-in from VCC. Store the
result and also save the carry-out to VCC. In VOP3 the VCC
destination may be an arbitrary SGPR-pair, and the VCC source
comes from the SGPR-pair at S2.u.

  D.u32 = S0.u32 + S1.u32 + VCC;

  VCC = S0.u32 + S1.u32 + VCC >= 0x100000000ULL ? 1 : 0.

41 V_SUB_CO_CI_U32 Subtract the second unsigned integer from the first unsigned
integer and then subtract a carry-in from VCC. Store the result
and also save the carry-out to VCC. In VOP3 the VCC destination
may be an arbitrary SGPR-pair, and the VCC source comes from the
SGPR-pair at S2.u.

  D.u32 = S0.u32 - S1.u32 - VCC;

  VCC = S1.u32 + VCC > S0.u32 ? 1 : 0.

42 V_SUBREV_CO_CI_U3 Subtract the first unsigned integer from the second unsigned
2 integer and then subtract a carry-in from VCC. Store the result
and also save the carry-out to VCC. In VOP3 the VCC destination
may be an arbitrary SGPR-pair, and the VCC source comes from the
SGPR-pair at S2.u.

  D.u32 = S1.u32 - S0.u32 - VCC;

  VCC = S1.u32 + VCC > S0.u ? 1 : 0.

43 V_FMAC_F32 Fused multiply-add of single-precision floats, accumulate with

destination.

  D.f32 = S0.f32 * S1.f32 + D.f32. // Fused operation

44 V_FMAMK_F32 Multiply a single-precision float with a literal constant and

add a second single-precision float using fused multiply-add.
This opcode cannot use the VOP3 encoding and cannot use
input/output modifiers.

  D.f32 = S0.f32 * K.f32 + S1.f32. // K is a 32-bit literal constant.

45 V_FMAAK_F32 Multiply two single-precision floats and add a literal constant

using fused multiply-add. This opcode cannot use the VOP3
encoding and cannot use input/output modifiers.

  D.f32 = S0.f32 * S1.f32 + K.f32. // K is a 32-bit literal constant.

12.7. VOP2 Instructions 134 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

47 V_CVT_PKRTZ_F16_F Convert two single-precision floats into a packed FP16 result

32 and round to zero (ignore the current rounding mode).
This opcode is intended for use with 16-bit compressed exports.
See V_CVT_F16_F32 for a version that respects the current
rounding mode.

  D.f16_lo = f32_to_f16(S0.f32);
  D.f16_hi = f32_to_f16(S1.f32).
  // Round-toward-zero regardless of current round mode setting in
hardware.

50 V_ADD_F16 Add two FP16 values. 0.5ULP precision. Supports denormals,

round mode, exception flags and saturation.

  D.f16_lo = S0.f16_lo + S1.f16_lo.

51 V_SUB_F16 Subtract the second FP16 value from the first. 0.5ULP
precision, Supports denormals, round mode, exception flags and
saturation.

  D.f16_lo = S0.f16_lo - S1.f16_lo.

52 V_SUBREV_F16 Subtract the first FP16 value from the second. 0.5ULP
precision. Supports denormals, round mode, exception flags and
saturation.

  D.f16_lo = S1.f16_lo - S0.f16_lo.

53 V_MUL_F16 Multiply two FP16 values. 0.5ULP precision. Supports

denormals, round mode, exception flags and saturation.

  D.f16_lo = S0.f16_lo * S1.f16_lo.

54 V_FMAC_F16 Fused multiply-add of FP16 values, accumulate with destination.

0.5ULP precision. Supports denormals, round mode, exception
flags and saturation.

  D.f16_lo = S0.f16_lo * S1.f16_lo + D.f16_lo.

55 V_FMAMK_F16 Multiply a FP16 value with a literal constant and add a second
FP16 value using fused multiply-add. This opcode cannot use the
VOP3 encoding and cannot use input/output modifiers. Supports
round mode, exception flags, saturation.

  D.f16_lo = S0.f16_lo * K.f16_lo + S1.f16_lo.

  // K is a 32-bit literal constant stored in the following literal
DWORD.

56 V_FMAAK_F16 Multiply two FP16 values and add a literal constant using fused
multiply-add. This opcode cannot use the VOP3 encoding and
cannot use input/output modifiers. Supports round mode,
exception flags, saturation.

  D.f16_lo = S0.f16_lo * S1.f16_lo + K.f16_lo.

  // K is a 32-bit literal constant stored in the following literal
DWORD.

12.7. VOP2 Instructions 135 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

57 V_MAX_F16 Maximum of two FP16 values. IEEE compliant. Supports

denormals, round mode, exception flags, saturation.

  D.f16 = max(S0.f16,S1.f16);
  if (IEEE_MODE && S0.f16 == sNaN)
  D.f16 = Quiet(S0.f16);
  else if (IEEE_MODE && S1.f16 == sNaN)
  D.f16 = Quiet(S1.f16);
  else if (S0.f16 == NaN)
  D.f16 = S1.f16;
  else if (S1.f16 == NaN)
  D.f16 = S0.f16;
  else if (S0.f16 == +0.0 && S1.f16 == -0.0)
  D.f16 = S0.f16;
  else if (S0.f16 == -0.0 && S1.f16 == +0.0)
  D.f16 = S1.f16;
  else if (IEEE_MODE)
  D.f16 = (S0.f16 >= S1.f16 ? S0.f16 : S1.f16);
  else
  D.f16 = (S0.f16 > S1.f16 ? S0.f16 : S1.f16);
  endif.

58 V_MIN_F16 Minimum of two FP16 values. IEEE compliant. Supports

denormals, round mode, exception flags, saturation.

  D.f16 = min(S0.f16,S1.f16);
  if (IEEE_MODE && S0.f16 == sNaN)
  D.f16 = Quiet(S0.f16);
  else if (IEEE_MODE && S1.f16 == sNaN)
  D.f16 = Quiet(S1.f16);
  else if (S0.f16 == NaN)
  D.f16 = S1.f16;
  else if (S1.f16 == NaN)
  D.f16 = S0.f16;
  else if (S0.f16 == +0.0 && S1.f16 == -0.0)
  D.f16 = S1.f16;
  else if (S0.f16 == -0.0 && S1.f16 == +0.0)
  D.f16 = S0.f16;
  else
  // Note: there's no IEEE special case here like there is for
V_MAX_F16.
  D.f16 = (S0.f16 < S1.f16 ? S0.f16 : S1.f16);
  endif.

59 V_LDEXP_F16 Multiply an FP16 value by an integral power of 2, compare with

the ldexp() function in C. Note that the S1 has a format of f16
since floating point literal constants are interpreted as 16 bit
value for this opcode.

  D.f16 = S0.f16 * (2 ** S1.i16).

60 V_PK_FMAC_F16   Multiply packed FP16 values and accumulate with destination.

VOP2 version of V_PK_FMA_F16 with third source VGPR address is
the destination.

  D.f16_lo = S0.f16_lo * S1.f16_lo + D.f16_lo;

  D.f16_hi = S0.f16_hi * S1.f16_hi + D.f16_hi.

12.7. VOP2 Instructions 136 of 283

"RDNA 2" Instruction Set Architecture

12.7.1. VOP2 using VOP3 encoding

Instructions in this format may also be encoded as VOP3. VOP3 allows access to the extra
control bits (e.g. ABS, OMOD) at the expense of a larger instruction word. The VOP3 opcode is:
VOP2 opcode + 0x100.

12.8. VOP1 Instructions

Instructions in this format may use a 32-bit literal constant, DPP or SDWA which occurs
immediately after the instruction.

Opcode Name Description

0 V_NOP Do nothing, with style!

1 V_MOV_B32 Move data to a VGPR. Floating-point modifiers are valid for

this instruction if S0.u is a 32-bit floating point value. This
instruction is suitable for negating or taking the absolute
value of a floating-point value.

  D.u = S0.u.

Examples:
  v_mov_b32 v0, v1 // Move v1 to v0
  v_mov_b32 v0, -v1 // Set v1 to the negation of v0
  v_mov_b32 v0, abs(v1) // Set v1 to the absolute value of v0

2 V_READFIRSTLANE_B Copy one VGPR value to one SGPR. D = SGPR destination, S0 =

32 source data (VGPR# or M0 for lds direct access), Lane# =
FindFirst1fromLSB(exec) (Lane# = 0 if exec is zero). Ignores
exec mask for the access. Input and output modifiers not
supported; this is an untyped operation.

3 V_CVT_I32_F64 Convert from a double-precision float to a signed integer.

0.5ULP accuracy, out-of-range floating point values (including
infinity) saturate. NaN is converted to 0. Generation of the
INEXACT exception is controlled by the CLAMP bit. INEXACT
exceptions are enabled for this conversion iff CLAMP == 1.

  D.i = (int)S0.d.

12.8. VOP1 Instructions 137 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

4 V_CVT_F64_I32 Convert from a signed integer to a double-precision float, 0ULP

accuracy.

  D.d = (double)S0.i.

5 V_CVT_F32_I32 Convert from a signed integer to a single-precision float,

0.5ULP accuracy.

  D.f = (float)S0.i.

6 V_CVT_F32_U32 Convert from an unsigned integer to a single-precision float,

0.5ULP accuracy.

  D.f = (float)S0.u.

7 V_CVT_U32_F32 Convert from a single-precision float to an unsigned integer.

1ULP accuracy, out-of-range floating point values (including
infinity) saturate. NaN is converted to 0. Generation of the
INEXACT exception is controlled by the CLAMP bit. INEXACT
exceptions are enabled for this conversion iff CLAMP == 1.

  D.u = (unsigned)S0.f.

8 V_CVT_I32_F32 Convert from a single-precision float to a signed integer. 1ULP

accuracy, out-of-range floating point values (including
infinity) saturate. NaN is converted to 0. Generation of the
INEXACT exception is controlled by the CLAMP bit. INEXACT
exceptions are enabled for this conversion iff CLAMP == 1.

  D.i = (int)S0.f.

10 V_CVT_F16_F32 Convert from a single-precision float to an FP16 float. 0.5ULP

accuracy, supports input modifiers and creates FP16 denormals
when appropriate.

  D.f16 = flt32_to_flt16(S0.f).

11 V_CVT_F32_F16 Convert from an FP16 float to a single-precision float. 0ULP

accuracy, FP16 denormal inputs are accepted.

  D.f = flt16_to_flt32(S0.f16).

12 V_CVT_RPI_I32_F32 Convert from a single-precision float to a signed integer, round

to nearest integer.0.5ULP accuracy, denormals are supported.

  D.i = (int)floor(S0.f + 0.5).

13 V_CVT_FLR_I32_F32 Convert from a single-precision float to a signed integer, round

down. 1ULP accuracy, denormals are supported.

  D.i = (int)floor(S0.f).

12.8. VOP1 Instructions 138 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

14 V_CVT_OFF_F32_I4 4-bit signed int to 32-bit float. Used for interpolation in

shader.

S0___Result__
1000 -0.5000f
1001 -0.4375f
1010 -0.3750f
1011 -0.3125f
1100 -0.2500f
1101 -0.1875f
1110 -0.1250f
1111 -0.0625f
0000 +0.0000f
0001 +0.0625f
0010 +0.1250f
0011 +0.1875f
0100 +0.2500f
0101 +0.3125f
0110 +0.3750f
0111 +0.4375f

15 V_CVT_F32_F64 Convert from a double-precision float to a single-precision

float. 0.5ULP accuracy, denormals are supported.

  D.f = (float)S0.d.

16 V_CVT_F64_F32 Convert from a single-precision float to a double-precision

float. 0ULP accuracy, denormals are supported.

  D.d = (double)S0.f.

17 V_CVT_F32_UBYTE0 Convert an unsigned byte (byte 0) to a single-precision float.

  D.f = (float)(S0.u[7:0]).

18 V_CVT_F32_UBYTE1 Convert an unsigned byte (byte 1) to a single-precision float.

  D.f = (float)(S0.u[15:8]).

19 V_CVT_F32_UBYTE2 Convert an unsigned byte (byte 2) to a single-precision float.

  D.f = (float)(S0.u[23:16]).

20 V_CVT_F32_UBYTE3 Convert an unsigned byte (byte 3) to a single-precision float.

  D.f = (float)(S0.u[31:24]).

21 V_CVT_U32_F64 Convert from a double-precision float to an unsigned integer.

0.5ULP accuracy, out-of-range floating point values (including
infinity) saturate. NaN is converted to 0. Generation of the
INEXACT exception is controlled by the CLAMP bit. INEXACT
exceptions are enabled for this conversion iff CLAMP == 1.

  D.u = (unsigned)S0.d.

22 V_CVT_F64_U32 Convert from an unsigned integer to a double-precision float.

0ULP accuracy.

  D.d = (double)S0.u.

12.8. VOP1 Instructions 139 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

23 V_TRUNC_F64 Return integer part of S0.d, round-to-zero semantics.

  D.d = trunc(S0.d).

24 V_CEIL_F64 Round up to next whole integer.

  D.d = trunc(S0.d);
  if(S0.d > 0.0 && S0.d != D.d) then
  D.d += 1.0;
  endif.

25 V_RNDNE_F64 Round-to-nearest-even semantics.

  D.d = floor(S0.d + 0.5);

  if(floor(S0.d) is even && fract(S0.d) == 0.5) then
  D.d -= 1.0;
  endif.

26 V_FLOOR_F64 Round down to previous whole integer.

  D.d = trunc(S0.d);
  if(S0.d < 0.0 && S0.d != D.d) then
  D.d += -1.0;
  endif.

27 V_PIPEFLUSH Flush the VALU destination cache.

32 V_FRACT_F32 Return fractional portion of a number. 0.5ULP accuracy,

denormals are accepted.

  D.f = S0.f + -floor(S0.f).

NOTE: This complies with the DX specification of fract where the

function behaves like an extension of integer modulus; be aware
this may differ from how fract() is defined in other domains.
For example: fract(-1.2) = 0.8 in DX.

33 V_TRUNC_F32 Return integer part of S0.f, round-to-zero semantics.

  D.f = trunc(S0.f).

34 V_CEIL_F32 Round up to next whole integer.

  D.f = trunc(S0.f);
  if(S0.f > 0.0 && S0.f != D.f) then
  D.f += 1.0;
  endif.

35 V_RNDNE_F32 Round-to-nearest-even semantics.

  D.f = floor(S0.f + 0.5);

  if(floor(S0.f) is even && fract(S0.f) == 0.5) then
  D.f -= 1.0;
  endif.

36 V_FLOOR_F32 Round down to previous whole integer.

  D.f = trunc(S0.f);
  if(S0.f < 0.0 && S0.f != D.f) then
  D.f += -1.0;
  endif.

12.8. VOP1 Instructions 140 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

37 V_EXP_F32 Base 2 exponentiation. 1ULP accuracy, denormals are flushed.

  D.f = pow(2.0, S0.f).

Functional examples:

V_EXP_F32(0xff800000) => 0x00000000 // exp(-INF) = 0

V_EXP_F32(0x80000000) => 0x3f800000 // exp(-0.0) = 1
V_EXP_F32(0x7f800000) => 0x7f800000 // exp(+INF) = +INF

39 V_LOG_F32 Base 2 logarithm. 1ULP accuracy, denormals are flushed.

  D.f = log2(S0.f).

Functional examples:

V_LOG_F32(0xff800000) => 0xffc00000 // log(-INF) = NAN

V_LOG_F32(0xbf800000) => 0xffc00000 // log(-1.0) = NAN
V_LOG_F32(0x80000000) => 0xff800000 // log(-0.0) = -INF
V_LOG_F32(0x00000000) => 0xff800000 // log(+0.0) = -INF
V_LOG_F32(0x3f800000) => 0x00000000 // log(+1.0) = 0
V_LOG_F32(0x7f800000) => 0x7f800000 // log(+INF) = +INF

42 V_RCP_F32 Compute reciprocal with IEEE rules. 1ULP accuracy. Accuracy

converges to < 0.5ULP when using the Newton-Raphson method and 2
FMA operations. Denormals are flushed.

  D.f = 1.0 / S0.f.

Functional examples:

V_RCP_F32(0xff800000) => 0x80000000 // rcp(-INF) = -0

V_RCP_F32(0xc0000000) => 0xbf000000 // rcp(-2.0) = -0.5
V_RCP_F32(0x80000000) => 0xff800000 // rcp(-0.0) = -INF
V_RCP_F32(0x00000000) => 0x7f800000 // rcp(+0.0) = +INF
V_RCP_F32(0x7f800000) => 0x00000000 // rcp(+INF) = +0

43 V_RCP_IFLAG_F32 Compute reciprocal as part of integer divide. Can raise integer

DIV_BY_ZERO exception but cannot raise floating-point
exceptions. To be used in an integer reciprocal macro by the
compiler with one of the following sequences:

  D.f = 1.0 / S0.f.

Unsigned usage:
  CVT_F32_U32
  RCP_IFLAG_F32
  MUL_F32 (2**32 - 1)
  CVT_U32_F32

Signed usage:
  CVT_F32_I32
  RCP_IFLAG_F32
  MUL_F32 (2**31 - 1)
  CVT_I32_F32

12.8. VOP1 Instructions 141 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

46 V_RSQ_F32 Reciprocal square root with IEEE rules. 1ULP accuracy,

denormals are flushed.

  D.f = 1.0 / sqrt(S0.f).

Functional examples:

V_RSQ_F32(0xff800000) => 0xffc00000 // rsq(-INF) = NAN

V_RSQ_F32(0x80000000) => 0xff800000 // rsq(-0.0) = -INF
V_RSQ_F32(0x00000000) => 0x7f800000 // rsq(+0.0) = +INF
V_RSQ_F32(0x40800000) => 0x3f000000 // rsq(+4.0) = +0.5
V_RSQ_F32(0x7f800000) => 0x00000000 // rsq(+INF) = +0

47 V_RCP_F64 Reciprocal with IEEE rules. Precision is (2**29) ULP, and supports
denormals.

  D.d = 1.0 / S0.d.

49 V_RSQ_F64 Reciprocal square root with IEEE rules. Precision is (2**29) ULP, and
supports denormals.

  D.f16 = 1.0 / sqrt(S0.f16).

51 V_SQRT_F32 Square root. 1ULP accuracy, denormals are flushed.

  D.f = sqrt(S0.f).

Functional examples:

V_SQRT_F32(0xff800000) => 0xffc00000 // sqrt(-INF) = NAN

V_SQRT_F32(0x80000000) => 0x80000000 // sqrt(-0.0) = -0
V_SQRT_F32(0x00000000) => 0x00000000 // sqrt(+0.0) = +0
V_SQRT_F32(0x40800000) => 0x40000000 // sqrt(+4.0) = +2.0
V_SQRT_F32(0x7f800000) => 0x7f800000 // sqrt(+INF) = +INF

52 V_SQRT_F64 Square root. Precision is (2**29) ULP, and supports denormals.

  D.d = sqrt(S0.d).

53 V_SIN_F32 Trigonometric sine. Denormals are supported.

  D.f = sin(S0.f * 2 * PI).

Functional examples:

V_SIN_F32(0xff800000) => 0xffc00000 // sin(-INF) = NAN

V_SIN_F32(0xff7fffff) => 0x00000000 // -MaxFloat, finite
V_SIN_F32(0x80000000) => 0x80000000 // sin(-0.0) = -0
V_SIN_F32(0x3e800000) => 0x3f800000 // sin(0.25) = 1
V_SIN_F32(0x7f800000) => 0xffc00000 // sin(+INF) = NAN

12.8. VOP1 Instructions 142 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

54 V_COS_F32 Trigonometric cosine. Denormals are supported.

  D.f = cos(S0.f * 2 * PI).

Functional examples:

V_COS_F32(0xff800000) => 0xffc00000 // cos(-INF) = NAN

V_COS_F32(0xff7fffff) => 0x3f800000 // -MaxFloat, finite
V_COS_F32(0x80000000) => 0x3f800000 // cos(-0.0) = 1
V_COS_F32(0x3e800000) => 0x00000000 // cos(0.25) = 0
V_COS_F32(0x7f800000) => 0xffc00000 // cos(+INF) = NAN

55 V_NOT_B32 Bitwise negation. Input and output modifiers not supported.

  D.u = ~S0.u.

56 V_BFREV_B32 Bitfield reverse. Input and output modifiers not supported.

  D.u[31:0] = S0.u[0:31].

57 V_FFBH_U32 Counts how many zeros before the first one starting from the
MSB. Returns -1 if there are no ones.

  D.i = -1; // Set if no ones are found

  for i in 0 ... 31 do
  // Note: search is from the MSB
  if S0.u[31 - i] == 1 then
  D.i = i;
  break for;
  endif;
  endfor.

Functional examples:

V_FFBH_U32(0x00000000) => 0xffffffff

V_FFBH_U32(0x800000ff) => 0
V_FFBH_U32(0x100000ff) => 3
V_FFBH_U32(0x0000ffff) => 16
V_FFBH_U32(0x00000001) => 31

58 V_FFBL_B32 Returns the bit position of the first one from the LSB, or -1 if
there are no ones.

  D.i = -1; // Set if no ones are found

  for i in 0 ... 31 do // Search from LSB
  if S0.u[i] == 1 then
  D.i = i;
  break for;
  endif;
  endfor.

Functional examples:

V_FFBL_B32(0x00000000) => 0xffffffff

V_FFBL_B32(0xff000001) => 0
V_FFBL_B32(0xff000008) => 3
V_FFBL_B32(0xffff0000) => 16
V_FFBL_B32(0x80000000) => 31

12.8. VOP1 Instructions 143 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

59 V_FFBH_I32 Counts how many bits in a row (from MSB to LSB) are the same as
the sign bit. Returns -1 if all bits are the same.

  D.i = -1; // Set if all bits are the same

  for i in 1 ... 31 do
  // Note: search is from the MSB
  if S0.i[31 - i] != S0.i[31] then
  D.i = i;
  break for;
  endif;
  endfor.

Functional examples:

V_FFBH_I32(0x00000000) => 0xffffffff

V_FFBH_I32(0x40000000) => 1
V_FFBH_I32(0x80000000) => 1
V_FFBH_I32(0x0fffffff) => 4
V_FFBH_I32(0xffff0000) => 16
V_FFBH_I32(0xfffffffe) => 31
V_FFBH_I32(0xffffffff) => 0xffffffff

60 V_FREXP_EXP_I32_F6 Returns exponent of single precision float input, such that S0.d
4 = significand * (2 ** exponent). See also V_FREXP_MANT_F64,
which returns the significand. See the C library function
frexp() for more information.

  if(S0.f64 == +-INF || S0.f64 == NAN)

  D.i32 = 0;
  else
  D.i32 = S0.f64.exp - 1023 + 1;
  endif.

61 V_FREXP_MANT_F64 Returns binary significand of double precision float input, such

that S0.d = significand * (2 ** exponent). Result range is in
(-1.0,-0.5][0.5,1.0) in normal cases. See also
V_FREXP_EXP_I32_F64, which returns integer exponent. See the C
library function frexp() for more information.

  if(S0.d == +-INF || S0.d == NAN) then

  D.d = S0.d;
  else
  D.d = Mantissa(S0.d);
  endif.

62 V_FRACT_F64 Return fractional portion of a number. 0.5ULP accuracy,

denormals are accepted.

  D.d = S0.d + -floor(S0.d).

NOTE: This complies with the DX specification of fract where the

function behaves like an extension of integer modulus; be aware
this may differ from how fract() is defined in other domains.
For example: fract(-1.2) = 0.8 in DX.

12.8. VOP1 Instructions 144 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

63 V_FREXP_EXP_I32_F3 Returns exponent of single precision float input, such that S0.f
2 = significand * (2 ** exponent). See also V_FREXP_MANT_F32,
which returns the significand. See the C library function
frexp() for more information.

  if(S0.f == +-INF || S0.f == NAN) then

  D.i = 0;
  else
  D.i = TwosComplement(Exponent(S0.f) - 127 + 1);
  endif.

64 V_FREXP_MANT_F32 Returns binary significand of single precision float input, such

that S0.f = significand * (2 ** exponent). Result range is in
(-1.0,-0.5][0.5,1.0) in normal cases. See also
V_FREXP_EXP_I32_F32, which returns integer exponent. See the C
library function frexp() for more information.

  if(S0.f == +-INF || S0.f == NAN) then

  D.f = S0.f;
  else
  D.f = Mantissa(S0.f);
  endif.

65 V_CLREXCP Clear this wave's exception state in the SIMD (SP).

66 V_MOVRELD_B32 Move to a relative destination address.

  addr = VGPR address appearing in instruction DST field;

  addr += M0.u[31:0];
  VGPR[addr].u = S0.u.

Example: The following instruction sequence will perform a move

v15 <== v7:
  s_mov_b32 m0, 10
  v_movreld_b32 v5, v7

67 V_MOVRELS_B32 Move from a relative source address.

  addr = VGPR address appearing in instruction SRC0 field;

  addr += M0.u[31:0];
  D.u = VGPR[addr].u.

Example: The following instruction sequence will perform a move

v5 <== v17:
  s_mov_b32 m0, 10
  v_movrels_b32 v5, v7

12.8. VOP1 Instructions 145 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

68 V_MOVRELSD_B32 Move from a relative source address to a relative destination

address.

  addr_src = VGPR address appearing in instruction SRC0 field;

  addr_src += M0.u[31:0];
  addr_dst = VGPR address appearing in instruction DST field;
  addr_dst += M0.u[31:0];
  VGPR[addr_dst].u = VGPR[addr_src].u.

Example: The following instruction sequence will perform a move

v15 <== v17:
  s_mov_b32 m0, 10
  v_movrelsd_b32 v5, v7

72 V_MOVRELSD_2_B32 Move from a relative source address to a relative destination

address, with different relative offsets.

  addr_src = VGPR address appearing in instruction SRC0 field;

  addr_src += M0.u[9:0];
  addr_dst = VGPR address appearing in instruction DST field;
  addr_dst += M0.u[25:16];
  VGPR[addr_dst].u = VGPR[addr_src].u.

Example: The following instruction sequence will perform a move

v25 <== v17:
  s_mov_b32 m0, ((20 << 16) | 10)
  v_movrelsd_2_b32 v5, v7

80 V_CVT_F16_U16 Convert from an unsigned short to an FP16 float. 0.5ULP

accuracy, supports denormals, rounding, exception flags and
saturation.

  D.f16 = uint16_to_flt16(S.u16).

81 V_CVT_F16_I16 Convert from a signed short to an FP16 float. 0.5ULP accuracy,

supports denormals, rounding, exception flags and saturation.

  D.f16 = int16_to_flt16(S.i16).

82 V_CVT_U16_F16 Convert from an FP16 float to an unsigned short. 1ULP accuracy,

supports rounding, exception flags and saturation. FP16
denormals are accepted. Conversion is done with truncation.
Generation of the INEXACT exception is controlled by the CLAMP
bit. INEXACT exceptions are enabled for this conversion iff
CLAMP == 1.

  D.u16 = flt16_to_uint16(S.f16).

83 V_CVT_I16_F16 Convert from an FP16 float to a signed short. 1ULP accuracy,

supports rounding, exception flags and saturation. FP16
denormals are accepted. Conversion is done with truncation.
Generation of the INEXACT exception is controlled by the CLAMP
bit. INEXACT exceptions are enabled for this conversion iff
CLAMP == 1.

  D.i16 = flt16_to_int16(S.f16).

12.8. VOP1 Instructions 146 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

84 V_RCP_F16 Reciprocal with IEEE rules. 0.51ULP accuracy.

  D.f16 = 1.0 / S0.f16.

Functional examples:

V_RCP_F16(0xfc00) => 0x8000 // rcp(-INF) = -0

V_RCP_F16(0xc000) => 0xb800 // rcp(-2.0) = -0.5
V_RCP_F16(0x8000) => 0xfc00 // rcp(-0.0) = -INF
V_RCP_F16(0x0000) => 0x7c00 // rcp(+0.0) = +INF
V_RCP_F16(0x7c00) => 0x0000 // rcp(+INF) = +0

85 V_SQRT_F16 Square root. 0.51ULP accuracy, denormals are supported.

  D.f16 = sqrt(S0.f16).

Functional examples:

V_SQRT_F16(0xfc00) => 0xfe00 // sqrt(-INF) = NAN

V_SQRT_F16(0x8000) => 0x8000 // sqrt(-0.0) = -0
V_SQRT_F16(0x0000) => 0x0000 // sqrt(+0.0) = +0
V_SQRT_F16(0x4400) => 0x4000 // sqrt(+4.0) = +2.0
V_SQRT_F16(0x7c00) => 0x7c00 // sqrt(+INF) = +INF

86 V_RSQ_F16 Reciprocal square root with IEEE rules. 0.51ULP accuracy,

denormals are supported.

  D.f16 = 1.0 / sqrt(S0.f16).

Functional examples:

V_RSQ_F16(0xfc00) => 0xfe00 // rsq(-INF) = NAN

V_RSQ_F16(0x8000) => 0xfc00 // rsq(-0.0) = -INF
V_RSQ_F16(0x0000) => 0x7c00 // rsq(+0.0) = +INF
V_RSQ_F16(0x4400) => 0x3800 // rsq(+4.0) = +0.5
V_RSQ_F16(0x7c00) => 0x0000 // rsq(+INF) = +0

87 V_LOG_F16 Base 2 logarithm. 0.51ULP accuracy, denormals are supported.

  D.f16 = log2(S0.f).

Functional examples:

V_LOG_F16(0xfc00) => 0xfe00 // log(-INF) = NAN

V_LOG_F16(0xbc00) => 0xfe00 // log(-1.0) = NAN
V_LOG_F16(0x8000) => 0xfc00 // log(-0.0) = -INF
V_LOG_F16(0x0000) => 0xfc00 // log(+0.0) = -INF
V_LOG_F16(0x3c00) => 0x0000 // log(+1.0) = 0
V_LOG_F16(0x7c00) => 0x7c00 // log(+INF) = +INF

12.8. VOP1 Instructions 147 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

88 V_EXP_F16 Base 2 exponentiation. 0.51ULP accuracy, denormals are

supported.

  D.f16 = pow(2.0, S0.f16).

Functional examples:

V_EXP_F16(0xfc00) => 0x0000 // exp(-INF) = 0

V_EXP_F16(0x8000) => 0x3c00 // exp(-0.0) = 1
V_EXP_F16(0x7c00) => 0x7c00 // exp(+INF) = +INF

89 V_FREXP_MANT_F16 Returns binary significand of half precision float input, such

that S0.f16 = significand * (2 ** exponent). Result range is in
(-1.0,-0.5][0.5,1.0) in normal cases. See also
V_FREXP_EXP_I16_F16, which returns integer exponent. See the C
library function frexp() for more information.

  if(S0.f16 == +-INF || S0.f16 == NAN) then

  D.f16 = S0.f16;
  else
  D.f16 = Mantissa(S0.f16);
  endif.

90 V_FREXP_EXP_I16_F1 Returns exponent of half precision float input, such that S0.f16
6 = significand * (2 ** exponent). See also V_FREXP_MANT_F16,
which returns the significand. See the C library function
frexp() for more information.

  if(S0.f16 == +-INF || S0.f16 == NAN) then

  D.i = 0;
  else
  D.i = TwosComplement(Exponent(S0.f16) - 15 + 1);
  endif.

91 V_FLOOR_F16 Round down to previous whole integer.

  D.f16 = trunc(S0.f16);
  if(S0.f16 < 0.0f && S0.f16 != D.f16) then
  D.f16 -= 1.0;
  endif.

92 V_CEIL_F16 Round up to next whole integer.

  D.f16 = trunc(S0.f16);
  if(S0.f16 > 0.0f && S0.f16 != D.f16) then
  D.f16 += 1.0;
  endif.

93 V_TRUNC_F16 Return integer part of S0.f16, round-to-zero semantics.

  D.f16 = trunc(S0.f16).

94 V_RNDNE_F16 Round-to-nearest-even semantics.

  D.f16 = floor(S0.f16 + 0.5);

  if(floor(S0.f16) is even && fract(S0.f16) == 0.5) then
  D.f16 -= 1.0;
  endif.

12.8. VOP1 Instructions 148 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

95 V_FRACT_F16 Return fractional portion of a number. 0.5ULP accuracy,

denormals are accepted.

  D.f16 = S0.f16 + -floor(S0.f16).

NOTE: This complies with the DX specification of fract where the

function behaves like an extension of integer modulus; be aware
this may differ from how fract() is defined in other domains.
For example: fract(-1.2) = 0.8 in DX.

96 V_SIN_F16 Trigonometric sine. Denormals are supported.

  D.f16 = sin(S0.f16 * 2 * PI).

Functional examples:

V_SIN_F16(0xfc00) => 0xfe00 // sin(-INF) = NAN

V_SIN_F16(0xfbff) => 0x0000 // Most negative finite FP16
V_SIN_F16(0x8000) => 0x8000 // sin(-0.0) = -0
V_SIN_F16(0x3400) => 0x3c00 // sin(0.25) = 1
V_SIN_F16(0x7bff) => 0x0000 // Most positive finite FP16
V_SIN_F16(0x7c00) => 0xfe00 // sin(+INF) = NAN

97 V_COS_F16 Trigonometric cosine. Denormals are supported.

  D.f16 = cos(S0.f16 * 2 * PI).

Functional examples:

V_COS_F16(0xfc00) => 0xfe00 // cos(-INF) = NAN

V_COS_F16(0xfbff) => 0x3c00 // Most negative finite FP16
V_COS_F16(0x8000) => 0x3c00 // cos(-0.0) = 1
V_COS_F16(0x3400) => 0x0000 // cos(0.25) = 0
V_COS_F16(0x7bff) => 0x3c00 // Most positive finite FP16
V_COS_F16(0x7c00) => 0xfe00 // cos(+INF) = NAN

98 V_SAT_PK_U8_I16 Packed 8-bit saturating add.

  D.u32 = {16'b0, sat8(S.u[31:16]), sat8(S.u[15:0])}.

99 V_CVT_NORM_I16_F1 Convert from an FP16 float to a signed normalized short. 0.5ULP

6 accuracy, supports rounding, exception flags and saturation,
denormals are supported.

  D.i16 = flt16_to_snorm16(S.f16).

100 V_CVT_NORM_U16_F Convert from an FP16 float to an unsigned normalized short.

16 0.5ULP accuracy, supports rounding, exception flags and
saturation, denormals are supported.

  D.u16 = flt16_to_unorm16(S.f16).

101 V_SWAP_B32 Swap operands. Input and output modifiers not supported; this
is an untyped operation.

  tmp = D.u;
  D.u = S0.u;
  S0.u = tmp.

12.8. VOP1 Instructions 149 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

104 V_SWAPREL_B32 Swap operands. Input and output modifiers not supported; this
is an untyped operation. The two addresses are relatively
indexed using M0.

  addr_src = VGPR address appearing in instruction SRC0 field;

  addr_src += M0.u[9:0];
  addr_dst = VGPR address appearing in instruction DST field;
  addr_dst += M0.u[25:16];
  tmp = VGPR[addr_dst];
  VGPR[addr_dst] = VGPR[addr_src];
  VGPR[addr_src] = tmp.

Example: The following instruction sequence will swap v25 and

v17:
  s_mov_b32 m0, ((20 << 16) | 10)
  v_swaprel_b32 v5, v7

12.8.1. VOP1 using VOP3 encoding

Instructions in this format may also be encoded as VOP3. VOP3 allows access to the extra
control bits (e.g. ABS, OMOD) at the expense of a larger instruction word. The VOP3 opcode is:
VOP2 opcode + 0x180.

12.9. VOPC Instructions

The bitfield map for VOPC is:

  where:

  SRC0 = First operand for instruction.

  VSRC1 = Second operand for instruction.
  OP = Instruction opcode.
  All VOPC instructions can alternatively be encoded in the VOP3A format.

Compare instructions perform the same compare operation on each lane (workItem or thread)
using that lane’s private data, and producing a 1 bit result per lane into VCC or EXEC.

12.9. VOPC Instructions 150 of 283

"RDNA 2" Instruction Set Architecture

Instructions in this format may use a 32-bit literal constant or SDWA which occurs immediately
after the instruction.

Most compare instructions fall into one of two categories:

• Those which can use one of 16 compare operations (floating point types). "{COMPF}"
• Those which can use one of 8 compare operations (integer types). "{COMPI}"

The opcode number is such that for these the opcode number can be calculated from a base
opcode number for the data type, plus an offset for the specific compare operation.

Table 57. Sixteen Compare Operations

Compare Operation Opcode Offset Description

F 0 D.u = 0

LT 1 D.u = (S0 < S1)

EQ 2 D.u = (S0 == S1)

LE 3 D.u = (S0 <= S1)

GT 4 D.u = (S0 > S1)

LG 5 D.u = (S0 <> S1)

GE 6 D.u = (S0 >= S1)

O 7 D.u = (!isNaN(S0) && !isNaN(S1))

U 8 D.u = (!isNaN(S0) || !isNaN(S1))

NGE 9 D.u = !(S0 >= S1)

NLG 10 D.u = !(S0 <> S1)

NGT 11 D.u = !(S0 > S1)

NLE 12 D.u = !(S0 <= S1)

NEQ 13 D.u = !(S0 == S1)

NLT 14 D.u = !(S0 < S1)

TRU 15 D.u = 1

Table 58. Instructions with Sixteen Compare Operations

Instruction Description Hex Range

V_CMP_{COMPF}_F16 16-bit float compare. 0x20 to 0x2F

V_CMPX_{COMPF}_F16 16-bit float compare. Also writes EXEC. 0x30 to 0x3F

V_CMP_{COMPF}_F32 32-bit float compare. 0x40 to 0x4F

V_CMPX_{COMPF}_F32 32-bit float compare. Also writes EXEC. 0x50 to 0x5F

V_CMPS_{COMPF}_F64 64-bit float compare. 0x60 to 0x6F

V_CMPSX_{COMPF}_F64 64-bit float compare. Also writes EXEC. 0x70 to 0x7F

12.9. VOPC Instructions 151 of 283

"RDNA 2" Instruction Set Architecture

Table 59. Eight Compare Operations

Compare Operation Opcode Offset Description

F 0 D.u = 0

LT 1 D.u = (S0 < S1)

EQ 2 D.u = (S0 == S1)

LE 3 D.u = (S0 <= S1)

GT 4 D.u = (S0 > S1)

LG 5 D.u = (S0 <> S1)

GE 6 D.u = (S0 >= S1)

TRU 7 D.u = 1

Table 60. Instructions with Eight Compare Operations

Instruction Description Hex Range

V_CMP_{COMPI}_I16 16-bit signed integer compare. 0xA0 - 0xA7

V_CMP_{COMPI}_U16 16-bit signed integer compare. Also writes EXEC. 0xA8 - 0xAF

V_CMPX_{COMPI}_I16 16-bit unsigned integer compare. 0xB0 - 0xB7

V_CMPX_{COMPI}_U16 16-bit unsigned integer compare. Also writes EXEC. 0xB8 - 0xBF

V_CMP_{COMPI}_I32 32-bit signed integer compare. 0xC0 - 0xC7

V_CMP_{COMPI}_U32 32-bit signed integer compare. Also writes EXEC. 0xC8 - 0xCF

V_CMPX_{COMPI}_I32 32-bit unsigned integer compare. 0xD0 - 0xD7

V_CMPX_{COMPI}_U32 32-bit unsigned integer compare. Also writes EXEC. 0xD8 - 0xDF

V_CMP_{COMPI}_I64 64-bit signed integer compare. 0xE0 - 0xE7

V_CMP_{COMPI}_U64 64-bit signed integer compare. Also writes EXEC. 0xE8 - 0xEF

V_CMPX_{COMPI}_I64 64-bit unsigned integer compare. 0xF0 - 0xF7

V_CMPX_{COMPI}_U64 64-bit unsigned integer compare. Also writes EXEC. 0xF8 - 0xFF

Table 61. VOPC Compare Opcodes

Opcode Name Description

0 V_CMP_F_F32   D[threadId] = 0.
  // D = VCC in VOPC encoding.

1 V_CMP_LT_F32   D[threadId] = (S0 < S1).

  // D = VCC in VOPC encoding.

2 V_CMP_EQ_F32   D[threadId] = (S0 == S1).

  // D = VCC in VOPC encoding.

3 V_CMP_LE_F32   D[threadId] = (S0 <= S1).

  // D = VCC in VOPC encoding.

4 V_CMP_GT_F32   D[threadId] = (S0 > S1).

  // D = VCC in VOPC encoding.

12.9. VOPC Instructions 152 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

5 V_CMP_LG_F32   D[threadId] = (S0 <> S1).

  // D = VCC in VOPC encoding.

6 V_CMP_GE_F32   D[threadId] = (S0 >= S1).

  // D = VCC in VOPC encoding.

7 V_CMP_O_F32   D[threadId] = (!isNan(S0) && !isNan(S1)).

  // D = VCC in VOPC encoding.

8 V_CMP_U_F32   D[threadId] = (isNan(S0) || isNan(S1)).

  // D = VCC in VOPC encoding.

9 V_CMP_NGE_F32   D[threadId] = !(S0 >= S1)

  // With NAN inputs this is not the same operation as <.
  // D = VCC in VOPC encoding.

10 V_CMP_NLG_F32   D[threadId] = !(S0 <> S1)

  // With NAN inputs this is not the same operation as ==.
  // D = VCC in VOPC encoding.

11 V_CMP_NGT_F32   D[threadId] = !(S0 > S1)

  // With NAN inputs this is not the same operation as <=.
  // D = VCC in VOPC encoding.

12 V_CMP_NLE_F32   D[threadId] = !(S0 <= S1)

  // With NAN inputs this is not the same operation as >.
  // D = VCC in VOPC encoding.

13 V_CMP_NEQ_F32   D[threadId] = !(S0 == S1)

  // With NAN inputs this is not the same operation as !=.
  // D = VCC in VOPC encoding.

14 V_CMP_NLT_F32   D[threadId] = !(S0 < S1)

  // With NAN inputs this is not the same operation as >=.
  // D = VCC in VOPC encoding.

15 V_CMP_TRU_F32   D[threadId] = 1.
  // D = VCC in VOPC encoding.

16 V_CMPX_F_F32   EXEC[threadId] = 0.

17 V_CMPX_LT_F32   EXEC[threadId] = (S0 < S1).

18 V_CMPX_EQ_F32   EXEC[threadId] = (S0 == S1).

19 V_CMPX_LE_F32   EXEC[threadId] = (S0 <= S1).

20 V_CMPX_GT_F32   EXEC[threadId] = (S0 > S1).

21 V_CMPX_LG_F32   EXEC[threadId] = (S0 <> S1).

22 V_CMPX_GE_F32   EXEC[threadId] = (S0 >= S1).

23 V_CMPX_O_F32   EXEC[threadId] = (!isNan(S0) && !isNan(S1)).

24 V_CMPX_U_F32   EXEC[threadId] = (isNan(S0) || isNan(S1)).

25 V_CMPX_NGE_F32   EXEC[threadId] = !(S0 >= S1)

  // With NAN inputs this is not the same operation as <.

26 V_CMPX_NLG_F32   EXEC[threadId] = !(S0 <> S1)

  // With NAN inputs this is not the same operation as ==.

27 V_CMPX_NGT_F32   EXEC[threadId] = !(S0 > S1)

  // With NAN inputs this is not the same operation as <=.

12.9. VOPC Instructions 153 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

28 V_CMPX_NLE_F32   EXEC[threadId] = !(S0 <= S1)

  // With NAN inputs this is not the same operation as >.

29 V_CMPX_NEQ_F32   EXEC[threadId] = !(S0 == S1)

  // With NAN inputs this is not the same operation as !=.

30 V_CMPX_NLT_F32   EXEC[threadId] = !(S0 < S1)

  // With NAN inputs this is not the same operation as >=.

31 V_CMPX_TRU_F32   EXEC[threadId] = 1.

32 V_CMP_F_F64   D[threadId] = 0.
  // D = VCC in VOPC encoding.

33 V_CMP_LT_F64   D[threadId] = (S0 < S1).

  // D = VCC in VOPC encoding.

34 V_CMP_EQ_F64   D[threadId] = (S0 == S1).

  // D = VCC in VOPC encoding.

35 V_CMP_LE_F64   D[threadId] = (S0 <= S1).

  // D = VCC in VOPC encoding.

36 V_CMP_GT_F64   D[threadId] = (S0 > S1).

  // D = VCC in VOPC encoding.

37 V_CMP_LG_F64   D[threadId] = (S0 <> S1).

  // D = VCC in VOPC encoding.

38 V_CMP_GE_F64   D[threadId] = (S0 >= S1).

  // D = VCC in VOPC encoding.

39 V_CMP_O_F64   D[threadId] = (!isNan(S0) && !isNan(S1)).

  // D = VCC in VOPC encoding.

40 V_CMP_U_F64   D[threadId] = (isNan(S0) || isNan(S1)).

  // D = VCC in VOPC encoding.

41 V_CMP_NGE_F64   D[threadId] = !(S0 >= S1)

  // With NAN inputs this is not the same operation as <.
  // D = VCC in VOPC encoding.

42 V_CMP_NLG_F64   D[threadId] = !(S0 <> S1)

  // With NAN inputs this is not the same operation as ==.
  // D = VCC in VOPC encoding.

43 V_CMP_NGT_F64   D[threadId] = !(S0 > S1)

  // With NAN inputs this is not the same operation as <=.
  // D = VCC in VOPC encoding.

44 V_CMP_NLE_F64   D[threadId] = !(S0 <= S1)

  // With NAN inputs this is not the same operation as >.
  // D = VCC in VOPC encoding.

45 V_CMP_NEQ_F64   D[threadId] = !(S0 == S1)

  // With NAN inputs this is not the same operation as !=.
  // D = VCC in VOPC encoding.

46 V_CMP_NLT_F64   D[threadId] = !(S0 < S1)

  // With NAN inputs this is not the same operation as >=.
  // D = VCC in VOPC encoding.

47 V_CMP_TRU_F64   D[threadId] = 1.
  // D = VCC in VOPC encoding.

12.9. VOPC Instructions 154 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

48 V_CMPX_F_F64   EXEC[threadId] = 0.

49 V_CMPX_LT_F64   EXEC[threadId] = (S0 < S1).

50 V_CMPX_EQ_F64   EXEC[threadId] = (S0 == S1).

51 V_CMPX_LE_F64   EXEC[threadId] = (S0 <= S1).

52 V_CMPX_GT_F64   EXEC[threadId] = (S0 > S1).

53 V_CMPX_LG_F64   EXEC[threadId] = (S0 <> S1).

54 V_CMPX_GE_F64   EXEC[threadId] = (S0 >= S1).

55 V_CMPX_O_F64   EXEC[threadId] = (!isNan(S0) && !isNan(S1)).

56 V_CMPX_U_F64   EXEC[threadId] = (isNan(S0) || isNan(S1)).

57 V_CMPX_NGE_F64   EXEC[threadId] = !(S0 >= S1)

  // With NAN inputs this is not the same operation as <.

58 V_CMPX_NLG_F64   EXEC[threadId] = !(S0 <> S1)

  // With NAN inputs this is not the same operation as ==.

59 V_CMPX_NGT_F64   EXEC[threadId] = !(S0 > S1)

  // With NAN inputs this is not the same operation as <=.

60 V_CMPX_NLE_F64   EXEC[threadId] = !(S0 <= S1)

  // With NAN inputs this is not the same operation as >.

61 V_CMPX_NEQ_F64   EXEC[threadId] = !(S0 == S1)

  // With NAN inputs this is not the same operation as !=.

62 V_CMPX_NLT_F64   EXEC[threadId] = !(S0 < S1)

  // With NAN inputs this is not the same operation as >=.

63 V_CMPX_TRU_F64   EXEC[threadId] = 1.

128 V_CMP_F_I32   D[threadId] = 0.

  // D = VCC in VOPC encoding.

129 V_CMP_LT_I32   D[threadId] = (S0 < S1).

  // D = VCC in VOPC encoding.

130 V_CMP_EQ_I32   D[threadId] = (S0 == S1).

  // D = VCC in VOPC encoding.

131 V_CMP_LE_I32   D[threadId] = (S0 <= S1).

  // D = VCC in VOPC encoding.

132 V_CMP_GT_I32   D[threadId] = (S0 > S1).

  // D = VCC in VOPC encoding.

133 V_CMP_NE_I32   D[threadId] = (S0 <> S1).

  // D = VCC in VOPC encoding.

134 V_CMP_GE_I32   D[threadId] = (S0 >= S1).

  // D = VCC in VOPC encoding.

135 V_CMP_T_I32   D[threadId] = 1.

  // D = VCC in VOPC encoding.

12.9. VOPC Instructions 155 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

136 V_CMP_CLASS_F32 VCC = IEEE numeric class function specified in S1.u, performed on S0.f.

The function reports true if the floating point value is *any* of the
numeric types selected in S1.u according to the following list:

S1.u[0] -- value is a signaling NaN.

S1.u[1] -- value is a quiet NaN.
S1.u[2] -- value is negative infinity.
S1.u[3] -- value is a negative normal value.
S1.u[4] -- value is a negative denormal value.
S1.u[5] -- value is negative zero.
S1.u[6] -- value is positive zero.
S1.u[7] -- value is a positive denormal value.
S1.u[8] -- value is a positive normal value.
S1.u[9] -- value is positive infinity.

137 V_CMP_LT_I16   D[threadId] = (S0 < S1).

  // D = VCC in VOPC encoding.

138 V_CMP_EQ_I16   D[threadId] = (S0 == S1).

  // D = VCC in VOPC encoding.

139 V_CMP_LE_I16   D[threadId] = (S0 <= S1).

  // D = VCC in VOPC encoding.

140 V_CMP_GT_I16   D[threadId] = (S0 > S1).

  // D = VCC in VOPC encoding.

141 V_CMP_NE_I16   D[threadId] = (S0 <> S1).

  // D = VCC in VOPC encoding.

142 V_CMP_GE_I16   D[threadId] = (S0 >= S1).

  // D = VCC in VOPC encoding.

143 V_CMP_CLASS_F16 VCC = IEEE numeric class function specified in S1.u, performed on
S0.f16.

Note that the S1 has a format of f16 since floating point literal
constants are interpreted as 16 bit value for this opcode.

The function reports true if the floating point value is *any* of the
numeric types selected in S1.u according to the following list:

S1.u[0] -- value is a signaling NaN.

S1.u[1] -- value is a quiet NaN.
S1.u[2] -- value is negative infinity.
S1.u[3] -- value is a negative normal value.
S1.u[4] -- value is a negative denormal value.
S1.u[5] -- value is negative zero.
S1.u[6] -- value is positive zero.
S1.u[7] -- value is a positive denormal value.
S1.u[8] -- value is a positive normal value.
S1.u[9] -- value is positive infinity.

144 V_CMPX_F_I32   EXEC[threadId] = 0.

145 V_CMPX_LT_I32   EXEC[threadId] = (S0 < S1).

146 V_CMPX_EQ_I32   EXEC[threadId] = (S0 == S1).

147 V_CMPX_LE_I32   EXEC[threadId] = (S0 <= S1).

12.9. VOPC Instructions 156 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

148 V_CMPX_GT_I32   EXEC[threadId] = (S0 > S1).

149 V_CMPX_NE_I32   EXEC[threadId] = (S0 <> S1).

150 V_CMPX_GE_I32   EXEC[threadId] = (S0 >= S1).

151 V_CMPX_T_I32   EXEC[threadId] = 1.

152 V_CMPX_CLASS_F32 EXEC = IEEE numeric class function specified in S1.u, performed on
S0.f.

The function reports true if the floating point value is *any* of the
numeric types selected in S1.u according to the following list:

S1.u[0] -- value is a signaling NaN.

S1.u[1] -- value is a quiet NaN.
S1.u[2] -- value is negative infinity.
S1.u[3] -- value is a negative normal value.
S1.u[4] -- value is a negative denormal value.
S1.u[5] -- value is negative zero.
S1.u[6] -- value is positive zero.
S1.u[7] -- value is a positive denormal value.
S1.u[8] -- value is a positive normal value.
S1.u[9] -- value is positive infinity.

153 V_CMPX_LT_I16   EXEC[threadId] = (S0 < S1).

154 V_CMPX_EQ_I16   EXEC[threadId] = (S0 == S1).

155 V_CMPX_LE_I16   EXEC[threadId] = (S0 <= S1).

156 V_CMPX_GT_I16   EXEC[threadId] = (S0 > S1).

157 V_CMPX_NE_I16   EXEC[threadId] = (S0 <> S1).

158 V_CMPX_GE_I16   EXEC[threadId] = (S0 >= S1).

159 V_CMPX_CLASS_F16 EXEC = IEEE numeric class function specified in S1.u, performed on
S0.f16.

Note that the S1 has a format of f16 since floating point literal
constants are interpreted as 16 bit value for this opcode.

The function reports true if the floating point value is *any* of the
numeric types selected in S1.u according to the following list:

S1.u[0] -- value is a signaling NaN.

S1.u[1] -- value is a quiet NaN.
S1.u[2] -- value is negative infinity.
S1.u[3] -- value is a negative normal value.
S1.u[4] -- value is a negative denormal value.
S1.u[5] -- value is negative zero.
S1.u[6] -- value is positive zero.
S1.u[7] -- value is a positive denormal value.
S1.u[8] -- value is a positive normal value.
S1.u[9] -- value is positive infinity.

160 V_CMP_F_I64   D[threadId] = 0.

  // D = VCC in VOPC encoding.

12.9. VOPC Instructions 157 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

161 V_CMP_LT_I64   D[threadId] = (S0 < S1).

  // D = VCC in VOPC encoding.

162 V_CMP_EQ_I64   D[threadId] = (S0 == S1).

  // D = VCC in VOPC encoding.

163 V_CMP_LE_I64   D[threadId] = (S0 <= S1).

  // D = VCC in VOPC encoding.

164 V_CMP_GT_I64   D[threadId] = (S0 > S1).

  // D = VCC in VOPC encoding.

165 V_CMP_NE_I64   D[threadId] = (S0 <> S1).

  // D = VCC in VOPC encoding.

166 V_CMP_GE_I64   D[threadId] = (S0 >= S1).

  // D = VCC in VOPC encoding.

167 V_CMP_T_I64   D[threadId] = 1.

  // D = VCC in VOPC encoding.

168 V_CMP_CLASS_F64 VCC = IEEE numeric class function specified in S1.u, performed on S0.d.

The function reports true if the floating point value is *any* of the
numeric types selected in S1.u according to the following list:

S1.u[0] -- value is a signaling NaN.

S1.u[1] -- value is a quiet NaN.
S1.u[2] -- value is negative infinity.
S1.u[3] -- value is a negative normal value.
S1.u[4] -- value is a negative denormal value.
S1.u[5] -- value is negative zero.
S1.u[6] -- value is positive zero.
S1.u[7] -- value is a positive denormal value.
S1.u[8] -- value is a positive normal value.
S1.u[9] -- value is positive infinity.

169 V_CMP_LT_U16   D[threadId] = (S0 < S1).

  // D = VCC in VOPC encoding.

170 V_CMP_EQ_U16   D[threadId] = (S0 == S1).

  // D = VCC in VOPC encoding.

171 V_CMP_LE_U16   D[threadId] = (S0 <= S1).

  // D = VCC in VOPC encoding.

172 V_CMP_GT_U16   D[threadId] = (S0 > S1).

  // D = VCC in VOPC encoding.

173 V_CMP_NE_U16   D[threadId] = (S0 <> S1).

  // D = VCC in VOPC encoding.

174 V_CMP_GE_U16   D[threadId] = (S0 >= S1).

  // D = VCC in VOPC encoding.

176 V_CMPX_F_I64   EXEC[threadId] = 0.

177 V_CMPX_LT_I64   EXEC[threadId] = (S0 < S1).

178 V_CMPX_EQ_I64   EXEC[threadId] = (S0 == S1).

179 V_CMPX_LE_I64   EXEC[threadId] = (S0 <= S1).

180 V_CMPX_GT_I64   EXEC[threadId] = (S0 > S1).

12.9. VOPC Instructions 158 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

181 V_CMPX_NE_I64   EXEC[threadId] = (S0 <> S1).

182 V_CMPX_GE_I64   EXEC[threadId] = (S0 >= S1).

183 V_CMPX_T_I64   EXEC[threadId] = 1.

184 V_CMPX_CLASS_F64 EXEC = IEEE numeric class function specified in S1.u, performed on
S0.d.

The function reports true if the floating point value is *any* of the
numeric types selected in S1.u according to the following list:

S1.u[0] -- value is a signaling NaN.

S1.u[1] -- value is a quiet NaN.
S1.u[2] -- value is negative infinity.
S1.u[3] -- value is a negative normal value.
S1.u[4] -- value is a negative denormal value.
S1.u[5] -- value is negative zero.
S1.u[6] -- value is positive zero.
S1.u[7] -- value is a positive denormal value.
S1.u[8] -- value is a positive normal value.
S1.u[9] -- value is positive infinity.

185 V_CMPX_LT_U16   EXEC[threadId] = (S0 < S1).

186 V_CMPX_EQ_U16   EXEC[threadId] = (S0 == S1).

187 V_CMPX_LE_U16   EXEC[threadId] = (S0 <= S1).

188 V_CMPX_GT_U16   EXEC[threadId] = (S0 > S1).

189 V_CMPX_NE_U16   EXEC[threadId] = (S0 <> S1).

190 V_CMPX_GE_U16   EXEC[threadId] = (S0 >= S1).

192 V_CMP_F_U32   D[threadId] = 0.

  // D = VCC in VOPC encoding.

193 V_CMP_LT_U32   D[threadId] = (S0 < S1).

  // D = VCC in VOPC encoding.

194 V_CMP_EQ_U32   D[threadId] = (S0 == S1).

  // D = VCC in VOPC encoding.

195 V_CMP_LE_U32   D[threadId] = (S0 <= S1).

  // D = VCC in VOPC encoding.

196 V_CMP_GT_U32   D[threadId] = (S0 > S1).

  // D = VCC in VOPC encoding.

197 V_CMP_NE_U32   D[threadId] = (S0 <> S1).

  // D = VCC in VOPC encoding.

198 V_CMP_GE_U32   D[threadId] = (S0 >= S1).

  // D = VCC in VOPC encoding.

199 V_CMP_T_U32   D[threadId] = 1.

  // D = VCC in VOPC encoding.

200 V_CMP_F_F16   D[threadId] = 0.

  // D = VCC in VOPC encoding.

201 V_CMP_LT_F16   D[threadId] = (S0 < S1).

  // D = VCC in VOPC encoding.

12.9. VOPC Instructions 159 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

202 V_CMP_EQ_F16   D[threadId] = (S0 == S1).

  // D = VCC in VOPC encoding.

203 V_CMP_LE_F16   D[threadId] = (S0 <= S1).

  // D = VCC in VOPC encoding.

204 V_CMP_GT_F16   D[threadId] = (S0 > S1).

  // D = VCC in VOPC encoding.

205 V_CMP_LG_F16   D[threadId] = (S0 <> S1).

  // D = VCC in VOPC encoding.

206 V_CMP_GE_F16   D[threadId] = (S0 >= S1).

  // D = VCC in VOPC encoding.

207 V_CMP_O_F16   D[threadId] = (!isNan(S0) && !isNan(S1)).

  // D = VCC in VOPC encoding.

208 V_CMPX_F_U32   EXEC[threadId] = 0.

209 V_CMPX_LT_U32   EXEC[threadId] = (S0 < S1).

210 V_CMPX_EQ_U32   EXEC[threadId] = (S0 == S1).

211 V_CMPX_LE_U32   EXEC[threadId] = (S0 <= S1).

212 V_CMPX_GT_U32   EXEC[threadId] = (S0 > S1).

213 V_CMPX_NE_U32   EXEC[threadId] = (S0 <> S1).

214 V_CMPX_GE_U32   EXEC[threadId] = (S0 >= S1).

215 V_CMPX_T_U32   EXEC[threadId] = 1.

216 V_CMPX_F_F16   EXEC[threadId] = 0.

217 V_CMPX_LT_F16   EXEC[threadId] = (S0 < S1).

218 V_CMPX_EQ_F16   EXEC[threadId] = (S0 == S1).

219 V_CMPX_LE_F16   EXEC[threadId] = (S0 <= S1).

220 V_CMPX_GT_F16   EXEC[threadId] = (S0 > S1).

221 V_CMPX_LG_F16   EXEC[threadId] = (S0 <> S1).

222 V_CMPX_GE_F16   EXEC[threadId] = (S0 >= S1).

223 V_CMPX_O_F16   EXEC[threadId] = (!isNan(S0) && !isNan(S1)).

224 V_CMP_F_U64   D[threadId] = 0.

  // D = VCC in VOPC encoding.

225 V_CMP_LT_U64   D[threadId] = (S0 < S1).

  // D = VCC in VOPC encoding.

226 V_CMP_EQ_U64   D[threadId] = (S0 == S1).

  // D = VCC in VOPC encoding.

227 V_CMP_LE_U64   D[threadId] = (S0 <= S1).

  // D = VCC in VOPC encoding.

228 V_CMP_GT_U64   D[threadId] = (S0 > S1).

  // D = VCC in VOPC encoding.

229 V_CMP_NE_U64   D[threadId] = (S0 <> S1).

  // D = VCC in VOPC encoding.

12.9. VOPC Instructions 160 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

230 V_CMP_GE_U64   D[threadId] = (S0 >= S1).

  // D = VCC in VOPC encoding.

231 V_CMP_T_U64   D[threadId] = 1.

  // D = VCC in VOPC encoding.

232 V_CMP_U_F16   D[threadId] = (isNan(S0) || isNan(S1)).

  // D = VCC in VOPC encoding.

233 V_CMP_NGE_F16   D[threadId] = !(S0 >= S1)

  // With NAN inputs this is not the same operation as <.
  // D = VCC in VOPC encoding.

234 V_CMP_NLG_F16   D[threadId] = !(S0 <> S1)

  // With NAN inputs this is not the same operation as ==.
  // D = VCC in VOPC encoding.

235 V_CMP_NGT_F16   D[threadId] = !(S0 > S1)

  // With NAN inputs this is not the same operation as <=.
  // D = VCC in VOPC encoding.

236 V_CMP_NLE_F16   D[threadId] = !(S0 <= S1)

  // With NAN inputs this is not the same operation as >.
  // D = VCC in VOPC encoding.

237 V_CMP_NEQ_F16   D[threadId] = !(S0 == S1)

  // With NAN inputs this is not the same operation as !=.
  // D = VCC in VOPC encoding.

238 V_CMP_NLT_F16   D[threadId] = !(S0 < S1)

  // With NAN inputs this is not the same operation as >=.
  // D = VCC in VOPC encoding.

239 V_CMP_TRU_F16   D[threadId] = 1.

  // D = VCC in VOPC encoding.

240 V_CMPX_F_U64   EXEC[threadId] = 0.

241 V_CMPX_LT_U64   EXEC[threadId] = (S0 < S1).

242 V_CMPX_EQ_U64   EXEC[threadId] = (S0 == S1).

243 V_CMPX_LE_U64   EXEC[threadId] = (S0 <= S1).

244 V_CMPX_GT_U64   EXEC[threadId] = (S0 > S1).

245 V_CMPX_NE_U64   EXEC[threadId] = (S0 <> S1).

246 V_CMPX_GE_U64   EXEC[threadId] = (S0 >= S1).

247 V_CMPX_T_U64   EXEC[threadId] = 1.

248 V_CMPX_U_F16   EXEC[threadId] = (isNan(S0) || isNan(S1)).

249 V_CMPX_NGE_F16   EXEC[threadId] = !(S0 >= S1)

  // With NAN inputs this is not the same operation as <.

250 V_CMPX_NLG_F16   EXEC[threadId] = !(S0 <> S1)

  // With NAN inputs this is not the same operation as ==.

251 V_CMPX_NGT_F16   EXEC[threadId] = !(S0 > S1)

  // With NAN inputs this is not the same operation as <=.

252 V_CMPX_NLE_F16   EXEC[threadId] = !(S0 <= S1)

  // With NAN inputs this is not the same operation as >.

12.9. VOPC Instructions 161 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

253 V_CMPX_NEQ_F16   EXEC[threadId] = !(S0 == S1)

  // With NAN inputs this is not the same operation as !=.

254 V_CMPX_NLT_F16   EXEC[threadId] = !(S0 < S1)

  // With NAN inputs this is not the same operation as >=.

255 V_CMPX_TRU_F16   EXEC[threadId] = 1.

12.9.1. VOPC using VOP3A encoding

Instructions in this format may also be encoded as VOP3A. VOP3A allows access to the extra
control bits (e.g. ABS, OMOD) at the expense of a larger instruction word. The VOP3A opcode
is: VOP2 opcode + 0x000.

When the CLAMP microcode bit is set to 1, these compare instructions signal an exception
when either of the inputs is NaN. When CLAMP is set to zero, NaN does not signal an
exception. The second eight VOPC instructions have {OP8} embedded in them. This refers to
each of the compare operations listed below.

where:

  VDST = Destination for instruction in the VGPR.

  ABS = Floating-point absolute value.
  CLMP = Clamp output.
  OP = Instruction opcode.
  SRC0 = First operand for instruction.
  SRC1 = Second operand for instruction.
  SRC2 = Third operand for instruction. Unused in VOPC instructions.
  OMOD = Output modifier for instruction. Unused in VOPC instructions.
  NEG = Floating-point negation.

12.10. VOP3P Instructions

Opcode Name Description

0 V_PK_MAD_I16 Packed multiply-add on signed shorts.

  D.i[31:16] = S0.i[31:16] * S1.i[31:16] + S2.i[31:16];

  D.i[15:0] = S0.i[15:0] * S1.i[15:0] + S2.i[15:0].

12.10. VOP3P Instructions 162 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

1 V_PK_MUL_LO_U16 Packed multiply on unsigned shorts.

  D.u[31:16] = S0.u[31:16] * S1.u[31:16];

  D.u[15:0] = S0.u[15:0] * S1.u[15:0].

2 V_PK_ADD_I16 Packed addition on signed shorts.

  D.i[31:16] = S0.i[31:16] + S1.i[31:16];

  D.i[15:0] = S0.i[15:0] + S1.i[15:0].

3 V_PK_SUB_I16 Packed subtraction on signed shorts. The second operand is

subtracted from the first.

  D.i[31:16] = S0.i[31:16] - S1.i[31:16];

  D.i[15:0] = S0.i[15:0] - S1.i[15:0].

4 V_PK_LSHLREV_B16 Packed logical shift left. The shift count is in the first
operand.

  D.u[31:16] = S1.u[31:16] << S0.u[19:16];

  D.u[15:0] = S1.u[15:0] << S0.u[3:0].

5 V_PK_LSHRREV_B16 Packed logical shift right. The shift count is in the first
operand.

  D.u[31:16] = S1.u[31:16] >> S0.u[19:16];

  D.u[15:0] = S1.u[15:0] >> S0.u[3:0].

6 V_PK_ASHRREV_I16 Packed arithmetic shift right (preserve sign bit). The shift
count is in the first operand.

  D.i[31:16] = S1.i[31:16] >> S0.i[19:16];

  D.i[15:0] = S1.i[15:0] >> S0.i[3:0].

7 V_PK_MAX_I16 Packed maximum of signed shorts.

  D.i[31:16] = (S0.i[31:16] >= S1.i[31:16]) ? S0.i[31:16] :

S1.i[31:16];
  D.i[15:0] = (S0.i[15:0] >= S1.i[15:0]) ? S0.i[15:0] :
S1.i[15:0].

8 V_PK_MIN_I16 Packed minimum of signed shorts.

  D.i[31:16] = (S0.i[31:16] < S1.i[31:16]) ? S0.i[31:16] :

S1.i[31:16];
  D.i[15:0] = (S0.i[15:0] < S1.i[15:0]) ? S0.i[15:0] :
S1.i[15:0].

9 V_PK_MAD_U16 Packed multiply-add on unsigned shorts.

  D.u[31:16] = S0.u[31:16] * S1.u[31:16] + S2.u[31:16];

  D.u[15:0] = S0.u[15:0] * S1.u[15:0] + S2.u[15:0].

10 V_PK_ADD_U16 Packed addition on unsigned shorts.

  D.u[31:16] = S0.u[31:16] + S1.u[31:16];

  D.u[15:0] = S0.u[15:0] + S1.u[15:0].

12.10. VOP3P Instructions 163 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

11 V_PK_SUB_U16 Packed subtraction on unsigned shorts. The second operand is

subtracted from the first.

  D.u[31:16] = S0.u[31:16] - S1.u[31:16];

  D.u[15:0] = S0.u[15:0] - S1.u[15:0].

12 V_PK_MAX_U16 Packed maximum of unsigned shorts.

  D.u[31:16] = (S0.u[31:16] >= S1.u[31:16]) ? S0.u[31:16] :

S1.u[31:16];
  D.u[15:0] = (S0.u[15:0] >= S1.u[15:0]) ? S0.u[15:0] :
S1.u[15:0].

13 V_PK_MIN_U16 Packed minimum of unsigned shorts.

  D.u[31:16] = (S0.u[31:16] < S1.u[31:16]) ? S0.u[31:16] :

S1.u[31:16];
  D.u[15:0] = (S0.u[15:0] < S1.u[15:0]) ? S0.u[15:0] :
S1.u[15:0].

14 V_PK_FMA_F16 Packed fused-multiply-add of FP16 values.

  D.f[31:16] = S0.f[31:16] * S1.f[31:16] + S2.f[31:16];

  D.f[15:0] = S0.f[15:0] * S1.f[15:0] + S2.f[15:0].

15 V_PK_ADD_F16 Packed addition of FP16 values.

  D.f[31:16] = S0.f[31:16] + S1.f[31:16];

  D.f[15:0] = S0.f[15:0] + S1.f[15:0].

16 V_PK_MUL_F16 Packed multiply of FP16 values.

  D.f[31:16] = S0.f[31:16] * S1.f[31:16];

  D.f[15:0] = S0.f[15:0] * S1.f[15:0].

17 V_PK_MIN_F16 Packed minimum of FP16 values.

  D.f[31:16] = min(S0.f[31:16], S1.f[31:16]);

  D.f[15:0] = min(S0.f[15:0], S1.u[15:0]).

18 V_PK_MAX_F16 Packed maximum of FP16 values.

  D.f[31:16] = max(S0.f[31:16], S1.f[31:16]);

  D.f[15:0] = max(S0.f[15:0], S1.f[15:0]).

19 V_DOT2_F32_F16 Dot product of packed FP16 values.

  D.f32 =
  S0.f16[0] * S1.f16[0] +
  S0.f16[1] * S1.f16[1] + S2.f32.

20 V_DOT2_I32_I16 Dot product of signed shorts.

  D.i32 =
  S0.i16[0] * S1.i16[0] +
  S0.i16[1] * S1.i16[1] + S2.i32.

21 V_DOT2_U32_U16 Dot product of unsigned shorts.

  D.u32 =
  S0.u16[0] * S1.u16[0] +
  S0.u16[1] * S1.u16[1] + S2.u32.

12.10. VOP3P Instructions 164 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

22 V_DOT4_I32_I8 Dot product of signed bytes.

  D.i32 =
  S0.i8[0] * S1.i8[0] +
  S0.i8[1] * S1.i8[1] +
  S0.i8[2] * S1.i8[2] +
  S0.i8[3] * S1.i8[3] + S2.i32.

23 V_DOT4_U32_U8 Dot product of unsigned bytes.

  D.u32 =
  S0.u8[0] * S1.u8[0] +
  S0.u8[1] * S1.u8[1] +
  S0.u8[2] * S1.u8[2] +
  S0.u8[3] * S1.u8[3] + S2.u32.

24 V_DOT8_I32_I4 Dot product of signed nibbles.

  D.i32 =
  S0.i4[0] * S1.i4[0] +
  S0.i4[1] * S1.i4[1] +
  S0.i4[2] * S1.i4[2] +
  S0.i4[3] * S1.i4[3] +
  S0.i4[4] * S1.i4[4] +
  S0.i4[5] * S1.i4[5] +
  S0.i4[6] * S1.i4[6] +
  S0.i4[7] * S1.i4[7] + S2.i32.

25 V_DOT8_U32_U4 Dot product of unsigned nibbles.

  D.u32 =
  S0.u4[0] * S1.u4[0] +
  S0.u4[1] * S1.u4[1] +
  S0.u4[2] * S1.u4[2] +
  S0.u4[3] * S1.u4[3] +
  S0.u4[4] * S1.u4[4] +
  S0.u4[5] * S1.u4[5] +
  S0.u4[6] * S1.u4[6] +
  S0.u4[7] * S1.u4[7] + S2.u32.

32 V_FMA_MIX_F32 Fused-multiply-add of single-precision values with MIX encoding.

Size and location of S0, S1 and S2 controlled by OPSEL:
0=src[31:0], 1=src[31:0], 2=src[15:0], 3=src[31:16]. Also, for
MAD_MIX, the NEG_HI field acts instead as an absolute-value
modifier.

  D.f[31:0] = S0.f * S1.f + S2.f.

33 V_FMA_MIXLO_F16 Fused-multiply-add of FP16 values with MIX encoding, result

stored in low 16 bits of destination. Size and location of S0,
S1 and S2 controlled by OPSEL: 0=src[31:0], 1=src[31:0],
2=src[15:0], 3=src[31:16]. Also, for MAD_MIX, the NEG_HI field
acts instead as an absolute-value modifier.

  D.f[15:0] = S0.f * S1.f + S2.f.

12.10. VOP3P Instructions 165 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

34 V_FMA_MIXHI_F16 Fused-multiply-add of FP16 values with MIX encoding, result

stored in HIGH 16 bits of destination. Size and location of S0,
S1 and S2 controlled by OPSEL: 0=src[31:0], 1=src[31:0],
2=src[15:0], 3=src[31:16]. Also, for MAD_MIX, the NEG_HI field
acts instead as an absolute-value modifier.

  D.f[31:16] = S0.f * S1.f + S2.f.

12.11. VINTERP Instructions

Opcode Name Description

0 V_INTERP_P1_F32 Parameter interpolation, first pass.

  D.f32 = P10[S1.u32].f32 * S0.f32 + P0[S1.u3].f32.

CAUTION: when in HALF_LDS mode, D must not be the same GPR as S;

if D == S then data corruption will occur.

NOTE: In textual representations the I/J VGPR is the first

source and the attribute is the second source; however in the
VOP3 encoding the attribute is stored in the src0 field and the
VGPR is stored in the src1 field.

1 V_INTERP_P2_F32 Parameter interpolation, second pass.

  D.f = P20[S1.u] * S0.f + D.f.

NOTE: In textual representations the I/J VGPR is the first

source and the attribute is the second source; however in the
VOP3 encoding the attribute is stored in the src0 field and the
VGPR is stored in the src1 field.

2 V_INTERP_MOV_F32 Parameter load. Used for custom interpolation in the shader.

  D.f = {P10,P20,P0}[S1.u].

12.11.1. VINTERP using VOP3 encoding

Instructions in this format may also be encoded as VOP3A. VOP3A allows access to the extra
control bits (e.g. ABS, OMOD) at the expense of a larger instruction word. The VOP3A opcode
is: VOP2 opcode + 0x270.

12.11. VINTERP Instructions 166 of 283

"RDNA 2" Instruction Set Architecture

12.12. VOP3A & VOP3B Instructions

VOP3 instructions use one of two encodings:

VOP3B this encoding allows specifying a unique scalar destination, and is used only for:
V_ADD_CO_U32
V_SUB_CO_U32
V_SUBREV_CO_U32
V_ADDC_CO_U32
V_SUBB_CO_U32
V_SUBBREV_CO_U32
V_DIV_SCALE_F32
V_DIV_SCALE_F64
V_MAD_U64_U32
V_MAD_I64_I32

VOP3A all other VALU instructions use this encoding

Opcode Name Description

320 V_FMA_LEGACY_F32 Multiply and add single-precision values. Follows DX9 rules
where 0.0 times anything produces 0.0 (this is not IEEE
compliant).

  D.f = S0.f * S1.f + S2.f. // DX9 rules, 0.0 * x = 0.0

322 V_MAD_I32_I24 Multiply two signed 24-bit integers, add a signed 32-bit integer
and store the result as a signed 32-bit integer. This opcode is
as efficient as basic single-precision opcodes since it utilizes
the single-precision floating point multiplier.

  D.i = S0.i[23:0] * S1.i[23:0] + S2.i.

323 V_MAD_U32_U24 Multiply two unsigned 24-bit integers, add an unsigned 32-bit
integer and store the result as an unsigned 32-bit integer.
This opcode is as efficient as basic single-precision opcodes
since it utilizes the single-precision floating point
multiplier.

  D.u = S0.u[23:0] * S1.u[23:0] + S2.u.

12.12. VOP3A & VOP3B Instructions 167 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

324 V_CUBEID_F32 Cubemap Face ID determination. Result is a floating point face

ID.

  // Set D.f = cubemap face ID ({0.0, 1.0, ..., 5.0}).

  // XYZ coordinate is given in (S0.f, S1.f, S2.f).
  // S0.f = x
  // S1.f = y
  // S2.f = z
  if (abs(S2.f) >= abs(S0.f) && abs(S2.f) >= abs(S1.f))
  if (S2.f < 0)
  D.f = 5.0;
  else
  D.f = 4.0;
  endif;
  else if (abs(S1.f) >= abs(S0.f))
  if (S1.f < 0)
  D.f = 3.0;
  else
  D.f = 2.0;
  endif;
  else
  if (S0.f < 0)
  D.f = 1.0;
  else
  D.f = 0.0;
  endif;
  endif.

325 V_CUBESC_F32 Cubemap S coordinate.

  // D.f = cubemap S coordinate.

  // XYZ coordinate is given in (S0.f, S1.f, S2.f).
  // S0.f = x
  // S1.f = y
  // S2.f = z
  if (abs(S2.f) >= abs(S0.f) && abs(S2.f) >= abs(S1.f))
  if (S2.f < 0)
  D.f = -S0.f;
  else
  D.f = S0.f;
  endif;
  else if (abs(S1.f) >= abs(S0.f))
  D.f = S0.f;
  else
  if (S0.f < 0)
  D.f = S2.f;
  else
  D.f = -S2.f;
  endif;
  endif.

12.12. VOP3A & VOP3B Instructions 168 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

326 V_CUBETC_F32 Cubemap T coordinate.

  // D.f = cubemap T coordinate.

  // XYZ coordinate is given in (S0.f, S1.f, S2.f).
  // S0.f = x
  // S1.f = y
  // S2.f = z
  if (abs(S2.f) >= abs(S0.f) && abs(S2.f) >= bs(S1.f))
  D.f = -S1.f;
  else if (abs(S1.f) >= abs(S0.f))
  if (S1.f < 0)
  D.f = -S2.f;
  else
  D.f = S2.f;
  endif;
  else
  D.f = -S1.f;
  endif.

327 V_CUBEMA_F32 Determine cubemap major axis.

  // D.f = 2.0 * cubemap major axis.

  // XYZ coordinate is given in (S0.f, S1.f, S2.f).
  // S0.f = x
  // S1.f = y
  // S2.f = z
  if (abs(S2.f) >= abs(S0.f) && abs(S2.f) >= abs(S1.f))
  D.f = 2.0 * S2.f;
  else if (abs(S1.f) >= abs(S0.f))
  D.f = 2.0 * S1.f;
  else
  D.f = 2.0 * S0.f;
  endif.

328 V_BFE_U32 Bitfield extract with S0 = data, S1 = field_offset, S2 =

field_width.

  D.u = (S0.u >> S1.u[4:0]) & ((1 << S2.u[4:0]) - 1).

329 V_BFE_I32 Bitfield extract with S0 = data, S1 = field_offset, S2 =

field_width.

  D.i = (S0.i >> S1.u[4:0]) & ((1 << S2.u[4:0]) - 1).

330 V_BFI_B32 Bitfield insert.

  D.u = (S0.u & S1.u) | (~S0.u & S2.u).

331 V_FMA_F32 Fused single precision multiply add. 0.5ULP accuracy, denormals
are supported.

  D.f = S0.f * S1.f + S2.f.

332 V_FMA_F64 Fused double precision multiply add. 0.5ULP precision,

denormals are supported.

  D.d = S0.d * S1.d + S2.d.

12.12. VOP3A & VOP3B Instructions 169 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

333 V_LERP_U8 Unsigned 8-bit pixel average on packed unsigned bytes (linear
interpolation). S2 acts as a round mode; if set, 0.5 rounds up,
otherwise 0.5 truncates.

  D.u = ((S0.u[31:24] + S1.u[31:24] + S2.u[24]) >> 1) << 24

  D.u += ((S0.u[23:16] + S1.u[23:16] + S2.u[16]) >> 1) << 16;
  D.u += ((S0.u[15:8] + S1.u[15:8] + S2.u[8]) >> 1) << 8;
  D.u += ((S0.u[7:0] + S1.u[7:0] + S2.u[0]) >> 1).

334 V_ALIGNBIT_B32 Align a value to the specified bit position.

  D.u = ({S0,S1} >> S2.u[4:0]) & 0xffffffff.

335 V_ALIGNBYTE_B32 Align a value to the specified byte position.

  D.u = ({S0,S1} >> (8*S2.u[4:0])) & 0xffffffff.

336 V_MULLIT_F32 Multiply for lighting. Specific rules apply: 0.0 * x = 0.0;
Specific INF, NaN, overflow rules.

  D.f = S0.f * S1.f

337 V_MIN3_F32 Return minimum single-precision value of three inputs.

  D.f = V_MIN_F32(V_MIN_F32(S0.f, S1.f), S2.f).

338 V_MIN3_I32 Return minimum signed integer value of three inputs.

  D.i = V_MIN_I32(V_MIN_I32(S0.i, S1.i), S2.i).

339 V_MIN3_U32 Return minimum unsigned integer value of three inputs.

  D.u = V_MIN_U32(V_MIN_U32(S0.u, S1.u), S2.u).

340 V_MAX3_F32 Return maximum single precision value of three inputs.

  D.f = V_MAX_F32(V_MAX_F32(S0.f, S1.f), S2.f).

341 V_MAX3_I32 Return maximum signed integer value of three inputs.

  D.i = V_MAX_I32(V_MAX_I32(S0.i, S1.i), S2.i).

342 V_MAX3_U32 Return maximum unsigned integer value of three inputs.

  D.u = V_MAX_U32(V_MAX_U32(S0.u, S1.u), S2.u).

343 V_MED3_F32 Return median single precision value of three inputs.

  if (isNan(S0.f) || isNan(S1.f) || isNan(S2.f))

  D.f = V_MIN3_F32(S0.f, S1.f, S2.f);
  else if (V_MAX3_F32(S0.f, S1.f, S2.f) == S0.f)
  D.f = V_MAX_F32(S1.f, S2.f);
  else if (V_MAX3_F32(S0.f, S1.f, S2.f) == S1.f)
  D.f = V_MAX_F32(S0.f, S2.f);
  else
  D.f = V_MAX_F32(S0.f, S1.f);
  endif.

12.12. VOP3A & VOP3B Instructions 170 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

344 V_MED3_I32 Return median signed integer value of three inputs.

  if (V_MAX3_I32(S0.i, S1.i, S2.i) == S0.i)

  D.i = V_MAX_I32(S1.i, S2.i);
  else if (V_MAX3_I32(S0.i, S1.i, S2.i) == S1.i)
  D.i = V_MAX_I32(S0.i, S2.i);
  else
  D.i = V_MAX_I32(S0.i, S1.i);
  endif.

345 V_MED3_U32 Return median unsigned integer value of three inputs.

  if (V_MAX3_U32(S0.u, S1.u, S2.u) == S0.u)

  D.u = V_MAX_U32(S1.u, S2.u);
  else if (V_MAX3_U32(S0.u, S1.u, S2.u) == S1.u)
  D.u = V_MAX_U32(S0.u, S2.u);
  else
  D.u = V_MAX_U32(S0.u, S1.u);
  endif.

346 V_SAD_U8 Sum of absolute differences with accumulation, overflow into

upper bits is allowed.

  ABSDIFF(x, y) := (x > y ? x - y : y - x) // UNSIGNED comparison

  D.u = S2.u;
  D.u += ABSDIFF(S0.u[31:24], S1.u[31:24]);
  D.u += ABSDIFF(S0.u[23:16], S1.u[23:16]);
  D.u += ABSDIFF(S0.u[15:8], S1.u[15:8]);
  D.u += ABSDIFF(S0.u[7:0], S1.u[7:0]).

347 V_SAD_HI_U8 Sum of absolute differences with accumulation, accumulate into

the higher-order bits of S2.

  D.u = (V_SAD_U8(S0, S1, 0) << 16) + S2.u.

348 V_SAD_U16 Short SAD with accumulation.

  ABSDIFF(x, y) := (x > y ? x - y : y - x) // UNSIGNED comparison

  D.u = S2.u;
  D.u += ABSDIFF(S0.u[31:16], S1.u[31:16]);
  D.u += ABSDIFF(S0.u[15:0], S1.u[15:0]).

349 V_SAD_U32 Dword SAD with accumulation.

  ABSDIFF(x, y) := (x > y ? x - y : y - x) // UNSIGNED comparison

  D.u = ABSDIFF(S0.u, S1.u) + S2.u.

350 V_CVT_PK_U8_F32 Convert floating point value S0 to 8-bit unsigned integer and
pack the result into byte S1 of dword S2.

  D.u = (S2.u & ~(0xff << (8 * S1.u[1:0])));

  D.u = D.u | ((flt32_to_uint8(S0.f) & 0xff) << (8 * S1.u[1:0])).

12.12. VOP3A & VOP3B Instructions 171 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

351 V_DIV_FIXUP_F32 Single precision division fixup.

S0 = Quotient, S1 = Denominator, S2 = Numerator.

Given a numerator, denominator, and quotient from a divide, this

opcode will detect and apply specific case numerics, touching up
the quotient if necessary. This opcode also generates invalid,
denorm and divide by zero exceptions caused by the division.

  sign_out = sign(S1.f)^sign(S2.f);
  if (S2.f == NAN)
  D.f = Quiet(S2.f);
  else if (S1.f == NAN)
  D.f = Quiet(S1.f);
  else if (S1.f == S2.f == 0)
  // 0/0
  D.f = 0xffc0_0000;
  else if (abs(S1.f) == abs(S2.f) == +-INF)
  // inf/inf
  D.f = 0xffc0_0000;
  else if (S1.f == 0 || abs(S2.f) == +-INF)
  // x/0, or inf/y
  D.f = sign_out ? -INF : +INF;
  else if (abs(S1.f) == +-INF || S2.f == 0)
  // x/inf, 0/y
  D.f = sign_out ? -0 : 0;
  else if ((exponent(S2.f) - exponent(S1.f)) < -150)
  D.f = sign_out ? -underflow : underflow;
  else if (exponent(S1.f) == 255)
  D.f = sign_out ? -overflow : overflow;
  else
  D.f = sign_out ? -abs(S0.f) : abs(S0.f);
  endif.

12.12. VOP3A & VOP3B Instructions 172 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

352 V_DIV_FIXUP_F64 Double precision division fixup.

S0 = Quotient, S1 = Denominator, S2 = Numerator.

Given a numerator, denominator, and quotient from a divide, this

opcode will detect and apply specific case numerics, touching up
the quotient if necessary. This opcode also generates invalid,
denorm and divide by zero exceptions caused by the division.

  sign_out = sign(S1.d)^sign(S2.d);
  if (S2.d == NAN)
  D.d = Quiet(S2.d);
  else if (S1.d == NAN)
  D.d = Quiet(S1.d);
  else if (S1.d == S2.d == 0)
  // 0/0
  D.d = 0xfff8_0000_0000_0000;
  else if (abs(S1.d) == abs(S2.d) == +-INF)
  // inf/inf
  D.d = 0xfff8_0000_0000_0000;
  else if (S1.d == 0 || abs(S2.d) == +-INF)
  // x/0, or inf/y
  D.d = sign_out ? -INF : +INF;
  else if (abs(S1.d) == +-INF || S2.d == 0)
  // x/inf, 0/y
  D.d = sign_out ? -0 : 0;
  else if ((exponent(S2.d) - exponent(S1.d)) < -1075)
  D.d = sign_out ? -underflow : underflow;
  else if (exponent(S1.d) == 2047)
  D.d = sign_out ? -overflow : overflow;
  else
  D.d = sign_out ? -abs(S0.d) : abs(S0.d);
  endif.

356 V_ADD_F64 Add two double-precision values. 0.5ULP precision, denormals

are supported.

  D.d = S0.d + S1.d.

357 V_MUL_F64 Multiply two double-precision values. 0.5ULP precision,

denormals are supported.

  D.d = S0.d * S1.d.

12.12. VOP3A & VOP3B Instructions 173 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

358 V_MIN_F64 Compute the minimum of two double-precision floats.

  if (IEEE_MODE && S0.d == sNaN)

  D.d = Quiet(S0.d);
  else if (IEEE_MODE && S1.d == sNaN)
  D.d = Quiet(S1.d);
  else if (S0.d == NaN)
  D.d = S1.d;
  else if (S1.d == NaN)
  D.d = S0.d;
  else if (S0.d == +0.0 && S1.d == -0.0)
  D.d = S1.d;
  else if (S0.d == -0.0 && S1.d == +0.0)
  D.d = S0.d;
  else
  // Note: there's no IEEE special case here like there is for
V_MAX_F64.
  D.d = (S0.d < S1.d ? S0.d : S1.d);
  endif.

359 V_MAX_F64 Compute the maximum of two double-precision floats.

  if (IEEE_MODE && S0.d == sNaN)

  D.d = Quiet(S0.d);
  else if (IEEE_MODE && S1.d == sNaN)
  D.d = Quiet(S1.d);
  else if (S0.d == NaN)
  D.d = S1.d;
  else if (S1.d == NaN)
  D.d = S0.d;
  else if (S0.d == +0.0 && S1.d == -0.0)
  D.d = S0.d;
  else if (S0.d == -0.0 && S1.d == +0.0)
  D.d = S1.d;
  else if (IEEE_MODE)
  D.d = (S0.d >= S1.d ? S0.d : S1.d);
  else
  D.d = (S0.d > S1.d ? S0.d : S1.d);
  endif.

360 V_LDEXP_F64 Multiply a double-precision float by an integral power of 2,

compare with the ldexp() function in C.

  D.d = S0.d * (2 ** S1.i).

361 V_MUL_LO_U32 Multiply two unsigned integers. If you only need to multiply
integers with small magnitudes consider V_MUL_U32_U24, which is
a faster implementation.

  D.u = S0.u * S1.u.

362 V_MUL_HI_U32 Multiply two unsigned integers and store the high 32 bits of the
result. If you only need to multiply integers with small
magnitudes consider V_MUL_HI_U32_U24, which is a faster
implementation.

  D.u = (S0.u * S1.u) >> 32.

12.12. VOP3A & VOP3B Instructions 174 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

364 V_MUL_HI_I32 Multiply two signed integers and store the high 32 bits of the
result. If you only need to multiply integers with small
magnitudes consider V_MUL_HI_I32_I24, which is a faster
implementation.

  D.i = (S0.i * S1.i) >> 32.

365 V_DIV_SCALE_F32 Single precision division pre-scale. S0 = Input to scale

(either denominator or numerator), S1 = Denominator, S2 =
Numerator.

Given a numerator and denominator, this opcode will

appropriately scale inputs for division to avoid subnormal terms
during Newton-Raphson correction algorithm. S0 must be the same
value as either S1 or S2.

This opcode producses a VCC flag for post-scaling of the

quotient (using V_DIV_FMAS_F32).

  VCC = 0;
  if (S2.f == 0 || S1.f == 0)
  D.f = NAN
  else if (exponent(S2.f) - exponent(S1.f) >= 96)
  // N/D near MAX_FLOAT
  VCC = 1;
  if (S0.f == S1.f)
  // Only scale the denominator
  D.f = ldexp(S0.f, 64);
  end if
  else if (S1.f == DENORM)
  D.f = ldexp(S0.f, 64);
  else if (1 / S1.f == DENORM && S2.f / S1.f == DENORM)
  VCC = 1;
  if (S0.f == S1.f)
  // Only scale the denominator
  D.f = ldexp(S0.f, 64);
  end if
  else if (1 / S1.f == DENORM)
  D.f = ldexp(S0.f, -64);
  else if (S2.f / S1.f==DENORM)
  VCC = 1;
  if (S0.f == S2.f)
  // Only scale the numerator
  D.f = ldexp(S0.f, 64);
  end if
  else if (exponent(S2.f) <= 23)
  // Numerator is tiny
  D.f = ldexp(S0.f, 64);
  end if.

12.12. VOP3A & VOP3B Instructions 175 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

366 V_DIV_SCALE_F64 Double precision division pre-scale. S0 = Input to scale

(either denominator or numerator), S1 = Denominator, S2 =
Numerator.

Given a numerator and denominator, this opcode will

appropriately scale inputs for division to avoid subnormal terms
during Newton-Raphson correction algorithm. S0 must be the same
value as either S1 or S2.

This opcode producses a VCC flag for post-scaling of the

quotient (using V_DIV_FMAS_F64).

  VCC = 0;
  if (S2.d == 0 || S1.d == 0)
  D.d = NAN
  else if (exponent(S2.d) - exponent(S1.d) >= 768)
  // N/D near MAX_FLOAT
  VCC = 1;
  if (S0.d == S1.d)
  // Only scale the denominator
  D.d = ldexp(S0.d, 128);
  end if
  else if (S1.d == DENORM)
  D.d = ldexp(S0.d, 128);
  else if (1 / S1.d == DENORM && S2.d / S1.d == DENORM)
  VCC = 1;
  if (S0.d == S1.d)
  // Only scale the denominator
  D.d = ldexp(S0.d, 128);
  end if
  else if (1 / S1.d == DENORM)
  D.d = ldexp(S0.d, -128);
  else if (S2.d / S1.d==DENORM)
  VCC = 1;
  if (S0.d == S2.d)
  // Only scale the numerator
  D.d = ldexp(S0.d, 128);
  end if
  else if (exponent(S2.d) <= 53)
  // Numerator is tiny
  D.d = ldexp(S0.d, 128);
  end if.

367 V_DIV_FMAS_F32 Single precision FMA with fused scale.

This opcode performs a standard Fused Multiply-Add operation and

will conditionally scale the resulting exponent if VCC is set.

Input denormals are not flushed, but output flushing is

allowed.

  if (VCC[threadId])
  D.f = 2**32 * (S0.f * S1.f + S2.f);
  else
  D.f = S0.f * S1.f + S2.f;
  end if.

12.12. VOP3A & VOP3B Instructions 176 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

368 V_DIV_FMAS_F64 Double precision FMA with fused scale.

This opcode performs a standard Fused Multiply-Add operation and

will conditionally scale the resulting exponent if VCC is set.

Input denormals are not flushed, but output flushing is

allowed.

  if (VCC[threadId])
  D.d = 2**64 * (S0.d * S1.d + S2.d);
  else
  D.d = S0.d * S1.d + S2.d;
  end if.

369 V_MSAD_U8 Masked sum of absolute differences with accumulation, overflow

into upper bits is allowed. Components where the reference
value in S1 is zero are not included in the sum.

  ABSDIFF(x, y) := (x > y ? x - y : y - x) // UNSIGNED comparison

  D.u = S2.u;
  D.u += S1.u[31:24] == 0 ? 0 : ABSDIFF(S0.u[31:24], S1.u[31:24]);
  D.u += S1.u[23:16] == 0 ? 0 : ABSDIFF(S0.u[23:16], S1.u[23:16]);
  D.u += S1.u[15:8] == 0 ? 0 : ABSDIFF(S0.u[15:8], S1.u[15:8]);
  D.u += S1.u[7:0] == 0 ? 0 : ABSDIFF(S0.u[7:0], S1.u[7:0]).

370 V_QSAD_PK_U16_U8 Quad-byte SAD with 16-bit packed accumulation.

  D[63:48] = SAD_U8(S0[55:24], S1[31:0], S2[63:48]);

  D[47:32] = SAD_U8(S0[47:16], S1[31:0], S2[47:32]);
  D[31:16] = SAD_U8(S0[39:8], S1[31:0], S2[31:16]);
  D[15:0] = SAD_U8(S0[31:0], S1[31:0], S2[15:0]).

371 V_MQSAD_PK_U16_ Quad-byte masked SAD with 16-bit packed accumulation.

U8
  D[63:48] = MSAD_U8(S0[55:24], S1[31:0], S2[63:48]);
  D[47:32] = MSAD_U8(S0[47:16], S1[31:0], S2[47:32]);
  D[31:16] = MSAD_U8(S0[39:8], S1[31:0], S2[31:16]);
  D[15:0] = MSAD_U8(S0[31:0], S1[31:0], S2[15:0]).

372 V_TRIG_PREOP_F64 Look Up 2/PI (S0.d) with segment select S1.u[4:0]. This
operation returns an aligned, double precision segment of 2/PI
needed to do range reduction on S0.d (double-precision value).
Multiple segments can be specified through S1.u[4:0]. Rounding
is round-to-zero. Large inputs (exp > 1968) are scaled to avoid
loss of precision through denormalization.

  shift = S1.u * 53;

  if exponent(S0.d) > 1077 then
  shift += exponent(S0.d) - 1077;
  endif
  result = (double) ((2/PI[1200:0] << shift) & 0x1fffff_ffffffff);
  scale = (-53 - shift);
  if exponent(S0.d) >= 1968 then
  scale += 128;
  endif
  D.d = ldexp(result, scale).

12.12. VOP3A & VOP3B Instructions 177 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

373 V_MQSAD_U32_U8 Quad-byte masked SAD with 32-bit packed accumulation.

  D[127:96] = MSAD_U8(S0[55:24], S1[31:0], S2[127:96]);

  D[95:64] = MSAD_U8(S0[47:16], S1[31:0], S2[95:64]);
  D[63:32] = MSAD_U8(S0[39:8], S1[31:0], S2[63:32]);
  D[31:0] = MSAD_U8(S0[31:0], S1[31:0], S2[31:0]).

374 V_MAD_U64_U32 Multiply and add unsigned integers and produce a 64-bit
result.

  {vcc_out,D.u64} = S0.u32 * S1.u32 + S2.u64.

375 V_MAD_I64_I32 Multiply and add signed integers and produce a 64-bit result.

  {vcc_out,D.i64} = S0.i32 * S1.i32 + S2.i64.

376 V_XOR3_B32 Bitwise XOR of three inputs. Input and output modifiers not
supported.

  D.u32 = S0.u32 ^ S1.u32 ^ S2.u32.

767 V_LSHLREV_B64 Logical shift left, count is in the first operand. Only one
scalar broadcast constant is allowed.

  D.u64 = S1.u64 << S0.u[5:0].

768 V_LSHRREV_B64 Logical shift right, count is in the first operand. Only one
scalar broadcast constant is allowed.

  D.u64 = S1.u64 >> S0.u[5:0].

769 V_ASHRREV_I64 Arithmetic shift right (preserve sign bit), count is in the
first operand. Only one scalar broadcast constant is allowed.

  D.u64 = signext(S1.u64) >> S0.u[5:0].

771 V_ADD_NC_U16 Add two unsigned shorts. Supports saturation (unsigned 16-bit
integer domain). No carry-in or carry-out.

  D.u16 = S0.u16 + S1.u16.

772 V_SUB_NC_U16 Subtract the second unsigned short from the first. Supports
saturation (unsigned 16-bit integer domain). No carry-in or
carry-out.

  D.u16 = S0.u16 - S1.u16.

773 V_MUL_LO_U16 Multiply two unsigned shorts. Supports saturation (unsigned

16-bit integer domain).

  D.u16 = S0.u16 * S1.u16.

775 V_LSHRREV_B16 Logical shift right, count is in the first operand.

  D.u[15:0] = S1.u[15:0] >> S0.u[3:0].

776 V_ASHRREV_I16 Arithmetic shift right (preserve sign bit), count is in the
first operand.

  D.i[15:0] = signext(S1.i[15:0]) >> S0.i[3:0].

12.12. VOP3A & VOP3B Instructions 178 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

777 V_MAX_U16 Maximum of two unsigned shorts.

  D.u16 = (S0.u16 >= S1.u16 ? S0.u16 : S1.u16).

778 V_MAX_I16 Maximum of two signed shorts.

  D.i16 = (S0.i16 >= S1.i16 ? S0.i16 : S1.i16).

779 V_MIN_U16 Minimum of two unsigned shorts.

  D.u16 = (S0.u16 < S1.u16 ? S0.u16 : S1.u16).

780 V_MIN_I16 Minimum of two signed shorts.

  D.i16 = (S0.i16 < S1.i16 ? S0.i16 : S1.i16).

781 V_ADD_NC_I16 Add two signed shorts. Supports saturation (signed 16-bit
integer domain). No carry-in or carry-out.

  D.i16 = S0.i16 + S1.i16.

782 V_SUB_NC_I16 Subtract the second signed short from the first. Supports
saturation (unsigned 16-bit integer domain). No carry-in or
carry-out.

  D.i16 = S0.i16 - S1.i16.

783 V_ADD_CO_U32 Add two unsigned integers with carry-out. In VOP3 the VCC
destination may be an arbitrary SGPR-pair.

  D.u32 = S0.u32 + S1.u32;

  VCC = S0.u + S1.u >= 0x100000000ULL ? 1 : 0.

784 V_SUB_CO_U32 Subtract the second unsigned integer from the first with
carry-out. In VOP3 the VCC destination may be an arbitrary
SGPR-pair.

  D.u = S0.u - S1.u;

  VCC[threadId] = (S1.u > S0.u ? 1 : 0).
  // VCC is an UNSIGNED overflow/carry-out for V_SUB_CO_CI_U32.

785 V_PACK_B32_F16 Pack two FP16 values together.

  D[31:16].f16 = S1.f16;
  D[15:0].f16 = S0.f16.

786 V_CVT_PKNORM_I16 Convert two FP16 values into packed signed normalized shorts.
_F16
  D = {(snorm)S1.f16, (snorm)S0.f16}.

787 V_CVT_PKNORM_U1 Convert two FP16 values into packed unsigned normalized
6_F16 shorts.

  D = {(unorm)S1.f16, (unorm)S0.f16}.

788 V_LSHLREV_B16 Logical shift left, count is in the first operand.

  D.u[15:0] = S1.u[15:0] << S0.u[3:0].

12.12. VOP3A & VOP3B Instructions 179 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

793 V_SUBREV_CO_U32 Subtract the first unsigned integer from the second with
carry-out. In VOP3 the VCC destination may be an arbitrary
SGPR-pair.

  D.u = S1.u - S0.u;

  VCC[threadId] = (S0.u > S1.u ? 1 : 0).
  // VCC is an UNSIGNED overflow/carry-out for V_SUB_CO_CI_U32.

832 V_MAD_U16 Multiply and add unsigned shorts. Supports saturation (unsigned
16-bit integer domain).

If op_sel[3] is 0: Result is written to 16 LSBs of destination

VGPR and hi 16 bits are preserved.

If op_sel[3] is 1: Result is written to 16 MSBs of destination

VGPR and lo 16 bits are preserved.

  D.u16 = S0.u16 * S1.u16 + S2.u16.

834 V_INTERP_P1LL_F16 FP16 parameter interpolation. `LL' stands for `two LDS
arguments'. attr_word selects the high or low half 16 bits of
each LDS dword accessed. This opcode is available for 32-bank
LDS only.

NOTE: In textual representations the I/J VGPR is the first

source and the attribute is the second source; however in the
VOP3 encoding the attribute is stored in the src0 field and the
VGPR is stored in the src1 field.

  D.f32 = P10.f16 * S0.f32 + P0.f16.

835 V_INTERP_P1LV_F16 FP16 parameter interpolation. `LV' stands for `One LDS and one
VGPR argument'. S2 holds two parameters, attr_word selects the
high or low word of the VGPR for this calculation, as well as
the high or low half of the LDS data. Meant for use with
16-bank LDS.

NOTE: In textual representations the I/J VGPR is the first

source and the attribute is the second source; however in the
VOP3 encoding the attribute is stored in the src0 field and the
VGPR is stored in the src1 field.

  D.f32 = P10.f16 * S0.f32 + (S2.u32 >> (attr_word * 16)).f16.

12.12. VOP3A & VOP3B Instructions 180 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

836 V_PERM_B32 Byte permute.

  D.u[31:24] = byte_permute({S0.u, S1.u}, S2.u[31:24]);

  D.u[23:16] = byte_permute({S0.u, S1.u}, S2.u[23:16]);
  D.u[15:8] = byte_permute({S0.u, S1.u}, S2.u[15:8]);
  D.u[7:0] = byte_permute({S0.u, S1.u}, S2.u[7:0]);

  byte permute(byte in[8], byte sel) {

  if(sel>=13) then return 0xff;
  elsif(sel==12) then return 0x00;
  elsif(sel==11) then return in[7][7] * 0xff;
  elsif(sel==10) then return in[5][7] * 0xff;
  elsif(sel==9) then return in[3][7] * 0xff;
  elsif(sel==8) then return in[1][7] * 0xff;
  else return in[sel];
  }

837 V_XAD_U32 Bitwise XOR and then add. No carryin/carryout and no

saturation. This opcode exists to accelerate the SHA256 hash
algorithm.

  D.u32 = (S0.u32 ^ S1.u32) + S2.u32.

838 V_LSHL_ADD_U32 Logical shift left and then add.

  D.u = (S0.u << S1.u[4:0]) + S2.u.

839 V_ADD_LSHL_U32 Add and then logical shift left the result.

  D.u = (S0.u + S1.u) << S2.u[4:0].

843 V_FMA_F16 Fused half precision multiply add of FP16 values. 0.5ULP
accuracy, denormals are supported.

If op_sel[3] is 0 Result is written to 16 LSBs of destination

VGPR and hi 16 bits are preserved.

If op_sel[3] is 1 Result is written to 16 MSBs of destination

VGPR and lo 16 bits are preserved.

  D.f16 = S0.f16 * S1.f16 + S2.f16.

849 V_MIN3_F16 Return minimum FP16 value of three inputs.

  D.f16 = V_MIN_F16(V_MIN_F16(S0.f16, S1.f16), S2.f16).

850 V_MIN3_I16 Return minimum signed short value of three inputs.

  D.i16 = V_MIN_I16(V_MIN_I16(S0.i16, S1.i16), S2.i16).

851 V_MIN3_U16 Return minimum unsigned short value of three inputs.

  D.u16 = V_MIN_U16(V_MIN_U16(S0.u16, S1.u16), S2.u16).

852 V_MAX3_F16 Return maximum FP16 value of three inputs.

  D.f16 = V_MAX_F16(V_MAX_F16(S0.f16, S1.f16), S2.f16).

853 V_MAX3_I16 Return maximum signed short value of three inputs.

  D.i16 = V_MAX_I16(V_MAX_I16(S0.i16, S1.i16), S2.i16).

12.12. VOP3A & VOP3B Instructions 181 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

854 V_MAX3_U16 Return maximum unsigned short value of three inputs.

  D.u16 = V_MAX_U16(V_MAX_U16(S0.u16, S1.u16), S2.u16).

855 V_MED3_F16 Return median FP16 value of three inputs.

  if (isNan(S0.f16) || isNan(S1.f16) || isNan(S2.f16))

  D.f16 = V_MIN3_F16(S0.f16, S1.f16, S2.f16);
  else if (V_MAX3_F16(S0.f16, S1.f16, S2.f16) == S0.f16)
  D.f16 = V_MAX_F16(S1.f16, S2.f16);
  else if (V_MAX3_F16(S0.f16, S1.f16, S2.f16) == S1.f16)
  D.f16 = V_MAX_F16(S0.f16, S2.f16);
  else
  D.f16 = V_MAX_F16(S0.f16, S1.f16);
  endif.

856 V_MED3_I16 Return median signed short value of three inputs.

  if (V_MAX3_I16(S0.i16, S1.i16, S2.i16) == S0.i16)

  D.i16 = V_MAX_I16(S1.i16, S2.i16);
  else if (V_MAX3_I16(S0.i16, S1.i16, S2.i16) == S1.i16)
  D.i16 = V_MAX_I16(S0.i16, S2.i16);
  else
  D.i16 = V_MAX_I16(S0.i16, S1.i16);
  endif.

857 V_MED3_U16 Return median unsigned short value of three inputs.

  if (V_MAX3_U16(S0.u16, S1.u16, S2.u16) == S0.u16)

  D.u16 = V_MAX_U16(S1.u16, S2.u16);
  else if (V_MAX3_U16(S0.u16, S1.u16, S2.u16) == S1.u16)
  D.u16 = V_MAX_U16(S0.u16, S2.u16);
  else
  D.u16 = V_MAX_U16(S0.u16, S1.u16);
  endif.

858 V_INTERP_P2_F16 FP16 parameter interpolation. Final computation. attr_word

selects LDS high or low 16bits. Used for both 16- and 32-bank
LDS.

NOTE: In textual representations the I/J VGPR is the first

source and the attribute is the second source; however in the
VOP3 encoding the attribute is stored in the src0 field and the
VGPR is stored in the src1 field.

If op_sel[3] is 0 Result is written to 16 LSBs of destination

VGPR and hi 16 bits are preserved.

If op_sel[3] is 1 Result is written to 16 MSBs of destination

VGPR and lo 16 bits are preserved.

  D.f16 = P20.f16 * S0.f32 + S2.f32.

12.12. VOP3A & VOP3B Instructions 182 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

862 V_MAD_I16 Multiply and add signed short values. Supports saturation
(signed 16-bit integer domain).

If op_sel[3] is 0 Result is written to 16 LSBs of destination

VGPR and hi 16 bits are preserved.

If op_sel[3] is 1 Result is written to 16 MSBs of destination

VGPR and lo 16 bits are preserved.

  D.i16 = S0.i16 * S1.i16 + S2.i16.

863 V_DIV_FIXUP_F16 Half precision division fixup.

S0 = Quotient, S1 = Denominator, S2 = Numerator.

Given a numerator, denominator, and quotient from a divide, this

opcode will detect and apply specific case numerics, touching up
the quotient if necessary. This opcode also generates invalid,
denorm and divide by zero exceptions caused by the division.

If op_sel[3] is 0 Result is written to 16 LSBs of destination

VGPR and hi 16 bits are preserved.

If op_sel[3] is 1 Result is written to 16 MSBs of destination

VGPR and lo 16 bits are preserved.

  sign_out = sign(S1.f16)^sign(S2.f16);
  if (S2.f16 == NAN)
  D.f16 = Quiet(S2.f16);
  else if (S1.f16 == NAN)
  D.f16 = Quiet(S1.f16);
  else if (S1.f16 == S2.f16 == 0)
  // 0/0
  D.f16 = 0xfe00;
  else if (abs(S1.f16) == abs(S2.f16) == +-INF)
  // inf/inf
  D.f16 = 0xfe00;
  else if (S1.f16 ==0 || abs(S2.f16) == +-INF)
  // x/0, or inf/y
  D.f16 = sign_out ? -INF : +INF;
  else if (abs(S1.f16) == +-INF || S2.f16 == 0)
  // x/inf, 0/y
  D.f16 = sign_out ? -0 : 0;
  else
  D.f16 = sign_out ? -abs(S0.f16) : abs(S0.f16);
  end if.

864 V_READLANE_B32 Copy one VGPR value to one SGPR. D = SGPR-dest, S0 = Source
Data (VGPR# or M0(lds-direct)), S1 = Lane Select (SGPR or M0).
Lane is S1 % (32 if wave32, 64 if wave64). Ignores exec mask.

Input and output modifiers not supported; this is an untyped

operation.

  if(wave32)
  SMEM[D_ADDR] = VMEM[S0_ADDR][S1[4:0]]; // For wave32
  else
  SMEM[D_ADDR] = VMEM[S0_ADDR][S1[5:0]]; // For wave64
  endif.

12.12. VOP3A & VOP3B Instructions 183 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

865 V_WRITELANE_B32 Write value into one VGPR in one lane. D = VGPR-dest, S0 =
Source Data (sgpr, m0, exec or constants), S1 = Lane Select
(SGPR or M0). Lane is S1 % (32 if wave32, 64 if wave64).
Ignores exec mask.

Input and output modifiers not supported; this is an untyped

operation.

  if(wave32)
  VMEM[D_ADDR][S1[4:0]] = SMEM[S0_ADDR]; // For wave32
  else
  VMEM[D_ADDR][S1[5:0]] = SMEM[S0_ADDR]; // For wave64
  endif.

866 V_LDEXP_F32 Multiply a single-precision float by an integral power of 2,

compare with the ldexp() function in C.

  D.f = S0.f * (2 ** S1.i).

867 V_BFM_B32 Bitfield modify. S0 is the bitfield width and S1 is the bitfield
offset.

  D.u32 = ((1<<S0[4:0])-1) << S1[4:0].

868 V_BCNT_U32_B32 Bit count.

  D.u = S1.u;
  for i in 0 .. 31 do
  D.u += S0.u[i]; // count i'th bit
  endfor.

869 V_MBCNT_LO_U32_B Masked bit count, ThreadPosition is the position of this thread
32 in the wavefront (in 0..63). See also V_MBCNT_HI_U32_B32.

  ThreadMask = (1LL << ThreadPosition) - 1;

  MaskedValue = (S0.u & ThreadMask[31:0]);
  D.u = S1.u;
  for i in 0 ... 31 do
  D.u += (MaskedValue[i] == 1 ? 1 : 0);
  endfor.

870 V_MBCNT_HI_U32_B Masked bit count, ThreadPosition is the position of this thread
32 in the wavefront (in 0..63). See also V_MBCNT_LO_U32_B32. Note
that in Wave32 mode ThreadMask[63:32] == 0 and this instruction
simply performs a move from S1 to D.

  ThreadMask = (1LL << ThreadPosition) - 1;

  MaskedValue = (S0.u & ThreadMask[63:32]);
  D.u = S1.u;
  for i in 0 ... 31 do
  D.u += (MaskedValue[i] == 1 ? 1 : 0);
  endfor.

Example to compute each thread's position in 0..63:

  v_mbcnt_lo_u32_b32 v0, -1, 0
  v_mbcnt_hi_u32_b32 v0, -1, v0
  // v0 now contains ThreadPosition

12.12. VOP3A & VOP3B Instructions 184 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

872 V_CVT_PKNORM_I16 Convert two single-precision floats into a packed signed

_F32 normalized value.

  D.i16_lo = (snorm)S0.f32;
  D.i16_hi = (snorm)S1.f32.

873 V_CVT_PKNORM_U1 Convert two single-precision floats into a packed unsigned

6_F32 normalized value.

  D.u16_lo = (unorm)S0.f32;
  D.u16_hi = (unorm)S1.f32.

874 V_CVT_PK_U16_U32 Convert two unsigned integers into a packed unsigned short.

  D.u16_lo = u32_to_u16(S0.u32);
  D.u16_hi = u32_to_u16(S1.u32).

875 V_CVT_PK_I16_I32 Convert two signed integers into a packed signed short.

  D.i16_lo = i32_to_i16(S0.i32);
  D.i16_hi = i32_to_i16(S1.i32).

877 V_ADD3_U32 Add three unsigned integers.

  D.u = S0.u + S1.u + S2.u.

879 V_LSHL_OR_B32 Logical shift left and then bitwise OR.

  D.u = (S0.u << S1.u[4:0]) | S2.u.

881 V_AND_OR_B32 Bitwise AND and then bitwise OR.

  D.u = (S0.u & S1.u) | S2.u.

882 V_OR3_B32 Bitwise OR of three inputs.

  D.u = S0.u | S1.u | S2.u.

883 V_MAD_U32_U16 Multiply and add unsigned values.

  D.u32 = S0.u16 * S1.u16 + S2.u32.

885 V_MAD_I32_I16 Multiply and add signed values.

  D.i32 = S0.i16 * S1.i16 + S2.i32.

886 V_SUB_NC_I32 Subtract the second signed integer from the first. No carry-in
or carry-out. Supports saturation (signed 32-bit integer
domain).

  D.i = S0.i - S1.i.

12.12. VOP3A & VOP3B Instructions 185 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

887 V_PERMLANE16_B32 Perform arbitrary gather-style operation within a row (16

contiguous lanes).

The first source must be a VGPR and the second and third
sources must be scalar values; the second and third source
are combined into a single 64-bit value representing lane
selects used to swizzle within each row.

OP_SEL is not used in its typical manner for this instruction.

For this instruction OP_SEL[0] is overloaded to represent the
DPP 'FI' (Fetch Inactive) bit and OP_SEL[1] is overloaded to
represent the DPP 'BOUND_CTRL' bit. The remainin OP_SEL bits are
reserved for this instruction.

ABS, NEG and OMOD modifiers should all be zeroed for this
instruction.

Compare with V_PERMLANEX16_B32.

  lanesel = { S2.u, S1.u }; // Concatenate lane select bits

  for row in 0 ... 3 do // interval is 0 ... 1 for wave32 mode
  // Implement arbitrary swizzle within each row
  for i in 0 ... 15 do
  D.lane[row * 16 + i] = S0.lane[row * 16 + lanesel[i * 4 + 3:i
* 4]];
  endfor;
  endfor.

Example implementing a rotation within each row:

  v_mov_b32 s0, 0x87654321;

  v_mov_b32 s1, 0x0fedcba9;
  v_permlane16_b32 v1, v0, s0, s1;
  // ROW 0:
  // v1.lane[0] <-- v0.lane[1]
  // v1.lane[1] <-- v0.lane[2]
  // ...
  // v1.lane[14] <-- v0.lane[15]
  // v1.lane[15] <-- v0.lane[0]
  //
  // ROW 1:
  // v1.lane[16] <-- v0.lane[17]
  // v1.lane[17] <-- v0.lane[18]
  // ...
  // v1.lane[30] <-- v0.lane[31]
  // v1.lane[31] <-- v0.lane[16]

12.12. VOP3A & VOP3B Instructions 186 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

888 V_PERMLANEX16_B3 Perform arbitrary gather-style operation across two rows (each
2 row is 16 contiguous lanes).

The first source must be a VGPR and the second and third
sources must be scalar values; the second and third source
are combined into a single 64-bit value representing lane
selects used to swizzle within each row.

OP_SEL is not used in its typical manner for this instruction.

For this instruction OP_SEL[0] is overloaded to represent the
DPP 'FI' (Fetch Inactive) bit and OP_SEL[1] is overloaded to
represent the DPP 'BOUND_CTRL' bit. The remainin OP_SEL bits are
reserved for this instruction.

ABS, NEG and OMOD modifiers should all be zeroed for this
instruction.

Compare with V_PERMLANE16_B32.

  lanesel = { S2.u, S1.u }; // Concatenate lane select bits

  for row in 0 ... 3 do // interval is 0 ... 1 for wave32 mode
  // Implement arbitrary swizzle across two rows
  altrow = {row[1], ~row[0]}; // 1<->0, 3<->2
  for i in 0 ... 15 do
  D.lane[row * 16 + i] = S0.lane[altrow * 16 + lanesel[i * 4 +
3:i * 4]];
  endfor;
  endfor.

Example implementing a rotation across an entire wave32

wavefront:

  // Note for this to work, source and destination VGPRs must be

different.
  // For this rotation, lane 15 gets data from lane 16, lane 31 gets
data from lane 0.
  // These are the only two lanes that need to use v_permlanex16_b32.

v_mov_b32 exec_lo, 0x7fff7fff; // Lanes getting data from their own row
v_mov_b32 s0, 0x87654321;
v_mov_b32 s1, 0x0fedcba9;
v_permlane16_b32 v1, v0, s0, s1 fi; // FI bit needed for lanes 14 and
30

v_mov_b32 exec_lo, 0x80008000; // Lanes getting data from the other row
v_permlanex16_b32 v1, v0, s0, s1 fi; // FI bit needed for lanes 15 and
31

895 V_ADD_NC_I32 Add two signed integers. No carry-in or carry-out. Supports

saturation (signed 32-bit integer domain). No carry-in or
carry-out.

  D.i = S0.i + S1.i.

12.12. VOP3A & VOP3B Instructions 187 of 283

"RDNA 2" Instruction Set Architecture

12.13. LDS & GDS Instructions

This suite of instructions operates on data stored within the data share memory. The instructions
transfer data between VGPRs and data share memory.
The bitfield map for the LDS/GDS is:

where:
OFFSET0 = Unsigned byte offset added to the address from the ADDR VGPR.
OFFSET1 = Unsigned byte offset added to the address from the ADDR VGPR.
GDS = Set if GDS, cleared if LDS.
OP = DS instruction opcode
ADDR = Source LDS address VGPR 0 - 255.
DATA0 = Source data0 VGPR 0 - 255.
DATA1 = Source data1 VGPR 0 - 255.
VDST = Destination VGPR 0- 255.


All instructions with RTN in the name return the value that was in memory
before the operation was performed.

Opcode Name Description

0 DS_ADD_U32   // 32bit
  tmp = MEM[ADDR];
  MEM[ADDR] += DATA;
  RETURN_DATA = tmp.

1 DS_SUB_U32   // 32bit
  tmp = MEM[ADDR];
  MEM[ADDR] -= DATA;
  RETURN_DATA = tmp.

2 DS_RSUB_U32 Subtraction with reversed operands.

  // 32bit
  addr = VGPR[ADDR]+{INST1,INST0};
  tmp = DS[addr].u32;
  DS[addr].u32 = VGPR[DATA0].u32-DS[addr].u32;
  VGPR[VDST].u32 = tmp.

3 DS_INC_U32   // 32bit
  tmp = MEM[ADDR];
  MEM[ADDR] = (tmp >= DATA) ? 0 : tmp + 1; // unsigned compare
  RETURN_DATA = tmp.

4 DS_DEC_U32   // 32bit
  tmp = MEM[ADDR];
  MEM[ADDR] = (tmp == 0 || tmp > DATA) ? DATA : tmp - 1; //
unsigned compare
  RETURN_DATA = tmp.

12.13. LDS & GDS Instructions 188 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

5 DS_MIN_I32   // 32bit
  tmp = MEM[ADDR];
  MEM[ADDR] = (DATA < tmp) ? DATA : tmp; // signed compare
  RETURN_DATA = tmp.

6 DS_MAX_I32   // 32bit
  tmp = MEM[ADDR];
  MEM[ADDR] = (DATA > tmp) ? DATA : tmp; // signed compare
  RETURN_DATA = tmp.

7 DS_MIN_U32   // 32bit
  tmp = MEM[ADDR];
  MEM[ADDR] = (DATA < tmp) ? DATA : tmp; // unsigned compare
  RETURN_DATA = tmp.

8 DS_MAX_U32   // 32bit
  tmp = MEM[ADDR];
  MEM[ADDR] = (DATA > tmp) ? DATA : tmp; // unsigned compare
  RETURN_DATA = tmp.

9 DS_AND_B32   // 32bit
  tmp = MEM[ADDR];
  MEM[ADDR] &= DATA;
  RETURN_DATA = tmp.

10 DS_OR_B32   // 32bit
  tmp = MEM[ADDR];
  MEM[ADDR] |= DATA;
  RETURN_DATA = tmp.

11 DS_XOR_B32   // 32bit
  tmp = MEM[ADDR];
  MEM[ADDR] ^= DATA;
  RETURN_DATA = tmp.

12 DS_MSKOR_B32 Masked dword OR, D0 contains the mask and D1 contains the new
value.

  // 32bit
  tmp = MEM[ADDR];
  MEM[ADDR] = (MEM[ADDR] & ~DATA) | DATA2;
  RETURN_DATA = tmp.

13 DS_WRITE_B32 Write dword.

  // 32bit
  MEM[ADDR] = DATA.

14 DS_WRITE2_B32 Write 2 dwords.

  // 32bit
  MEM[ADDR + OFFSET0 * 4] = DATA;
  MEM[ADDR + OFFSET1 * 4] = DATA2.

15 DS_WRITE2ST64_B32 Write 2 dwords with larger stride.

  // 32bit
  MEM[ADDR + OFFSET0 * 4 * 64] = DATA;
  MEM[ADDR + OFFSET1 * 4 * 64] = DATA2.

12.13. LDS & GDS Instructions 189 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

16 DS_CMPST_B32 Compare and store. Caution, the order of src and cmp are the
*opposite* of the BUFFER_ATOMIC_CMPSWAP opcode.

  // 32bit
  tmp = MEM[ADDR];
  src = DATA2;
  cmp = DATA;
  MEM[ADDR] = (tmp == cmp) ? src : tmp;
  RETURN_DATA[0] = tmp.

17 DS_CMPST_F32 Floating point compare and store that handles NaN/INF/denormal

values. Caution, the order of src and cmp are the *opposite* of
the BUFFER_ATOMIC_FCMPSWAP opcode.

  // 32bit
  tmp = MEM[ADDR];
  src = DATA2;
  cmp = DATA;
  MEM[ADDR] = (tmp == cmp) ? src : tmp;
  RETURN_DATA[0] = tmp.

18 DS_MIN_F32 Floating point minimum that handles NaN/INF/denormal values.

Note that this opcode is slightly more general-purpose than
BUFFER_ATOMIC_FMIN.

  // 32bit
  tmp = MEM[ADDR];
  src = DATA;
  cmp = DATA2;
  MEM[ADDR] = (cmp < tmp) ? src : tmp.

19 DS_MAX_F32 Floating point maximum that handles NaN/INF/denormal values.

Note that this opcode is slightly more general-purpose than
BUFFER_ATOMIC_FMAX.

  // 32bit
  tmp = MEM[ADDR];
  src = DATA;
  cmp = DATA2;
  MEM[ADDR] = (tmp > cmp) ? src : tmp.

20 DS_NOP Do nothing.

21 DS_ADD_F32 Floating point add that handles NaN/INF/denormal values.

  float tmp = MEM[ADDR].f32;

  MEM[ADDR].f32 += DATA0.f32;
  VDST.f32 = tmp;

12.13. LDS & GDS Instructions 190 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

24 DS_GWS_SEMA_RELEA GDS Only: The GWS resource (rid) indicated will process this
SE_ALL opcode by updating the counter and labeling the specified
resource as a semaphore.

  // Determine the GWS resource to work on

  rid[5:0] = gds_base[5:0] + offset0[5:0];

  // Incr the state counter of the resource

  state.counter[rid] = state.wave_in_queue;
  state.type = SEMAPHORE;
  return rd_done; //release calling wave

This action will release ALL queued waves; it Will have no

effect if no waves are present.

25 DS_GWS_INIT GDS Only: Initialize a barrier or semaphore resource.

  // Determine the GWS resource to work on

  rid[5:0] = gds_base[5:0] + offset0[5:0];

  // Get the value to use in init

  index = find_first_valid(vector mask)
  value = DATA[thread: index]

  // Set the state of the resource

  state.counter[rid] = lsb(value); //limit #waves
  state.flag[rid] = 0;
  return rd_done; //release calling wave

26 DS_GWS_SEMA_V GDS Only: The GWS resource indicated will process this opcode by
updating the counter and labeling the resource as a
semaphore.

  //Determine the GWS resource to work on

  rid[5:0] = gds_base[5:0] + offset0[5:0];

  //Incr the state counter of the resource

  state.counter[rid] += 1;
  state.type = SEMAPHORE;
  return rd_done; //release calling wave

This action will release one waved if any are queued in this
resource.

12.13. LDS & GDS Instructions 191 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

27 DS_GWS_SEMA_BR GDS Only: The GWS resource indicated will process this opcode by
updating the counter by the bulk release delivered count and
labeling the resource as a semaphore.

  //Determine the GWS resource to work on

  rid[5:0] = gds_base[5:0] + offset0[5:0];
  index = find first valid (vector mask)
  count = DATA[thread: index];

  //Add count to the resource state counter

  state.counter[rid] += count;
  state.type = SEMAPHORE;
  return rd_done; //release calling wave

This action will release count number of waves, promptly if

queued, or as they arrive from the noted resource.

28 DS_GWS_SEMA_P GDS Only: The GWS resource indicated will process this opcode by
queueing it until counter enables a release and then
decrementing the counter of the resource as a semaphore.

  //Determine the GWS resource to work on

  rid[5:0] = gds_base[5:0] + offset0[5:0];
  state.type = SEMAPHORE;
  ENQUEUE until(state[rid].counter > 0)
  state[rid].counter -= 1;
  return rd_done;

12.13. LDS & GDS Instructions 192 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

29 DS_GWS_BARRIER GDS Only: The GWS resource indicated will process this opcode by
queueing it until barrier is satisfied. The number of waves
needed is passed in as DATA of first valid thread.

  //Determine the GWS resource to work on

  rid[5:0] = gds_base[5:0] + OFFSET0[5:0];
  index = find first valid (vector mask);
  value = DATA[thread: index];

  // Input Decision Machine

  state.type[rid] = BARRIER;
  if(state[rid].counter <= 0) then
  thread[rid].flag = state[rid].flag;
  ENQUEUE;
  state[rid].flag = !state.flag;
  state[rid].counter = value;
  return rd_done;
  else
  state[rid].counter -= 1;
  thread.flag = state[rid].flag;
  ENQUEUE;
  endif.

Since the waves deliver the count for the next barrier, this
function can have a different size barrier for each
occurrence.

  // Release Machine
  if(state.type == BARRIER) then
  if(state.flag != thread.flag) then
  return rd_done;
  endif;
  endif.

30 DS_WRITE_B8 Byte write.

  MEM[ADDR] = DATA[7:0].

31 DS_WRITE_B16 Short write.

  MEM[ADDR] = DATA[15:0].

32 DS_ADD_RTN_U32   // 32bit
  tmp = MEM[ADDR];
  MEM[ADDR] += DATA;
  RETURN_DATA = tmp.

33 DS_SUB_RTN_U32   // 32bit
  tmp = MEM[ADDR];
  MEM[ADDR] -= DATA;
  RETURN_DATA = tmp.

34 DS_RSUB_RTN_U32 Subtraction with reversed operands.

  // 32bit
  addr = VGPR[ADDR]+{INST1,INST0};
  tmp = DS[addr].u32;
  DS[addr].u32 = VGPR[DATA0].u32-DS[addr].u32;
  VGPR[VDST].u32 = tmp.

12.13. LDS & GDS Instructions 193 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

35 DS_INC_RTN_U32   // 32bit
  tmp = MEM[ADDR];
  MEM[ADDR] = (tmp >= DATA) ? 0 : tmp + 1; // unsigned compare
  RETURN_DATA = tmp.

36 DS_DEC_RTN_U32   // 32bit
  tmp = MEM[ADDR];
  MEM[ADDR] = (tmp == 0 || tmp > DATA) ? DATA : tmp - 1; //
unsigned compare
  RETURN_DATA = tmp.

37 DS_MIN_RTN_I32   // 32bit
  tmp = MEM[ADDR];
  MEM[ADDR] = (DATA < tmp) ? DATA : tmp; // signed compare
  RETURN_DATA = tmp.

38 DS_MAX_RTN_I32   // 32bit
  tmp = MEM[ADDR];
  MEM[ADDR] = (DATA > tmp) ? DATA : tmp; // signed compare
  RETURN_DATA = tmp.

39 DS_MIN_RTN_U32   // 32bit
  tmp = MEM[ADDR];
  MEM[ADDR] = (DATA < tmp) ? DATA : tmp; // unsigned compare
  RETURN_DATA = tmp.

40 DS_MAX_RTN_U32   // 32bit
  tmp = MEM[ADDR];
  MEM[ADDR] = (DATA > tmp) ? DATA : tmp; // unsigned compare
  RETURN_DATA = tmp.

41 DS_AND_RTN_B32   // 32bit
  tmp = MEM[ADDR];
  MEM[ADDR] &= DATA;
  RETURN_DATA = tmp.

42 DS_OR_RTN_B32   // 32bit
  tmp = MEM[ADDR];
  MEM[ADDR] |= DATA;
  RETURN_DATA = tmp.

43 DS_XOR_RTN_B32   // 32bit
  tmp = MEM[ADDR];
  MEM[ADDR] ^= DATA;
  RETURN_DATA = tmp.

44 DS_MSKOR_RTN_B32 Masked dword OR, D0 contains the mask and D1 contains the new
value.

  // 32bit
  tmp = MEM[ADDR];
  MEM[ADDR] = (MEM[ADDR] & ~DATA) | DATA2;
  RETURN_DATA = tmp.

45 DS_WRXCHG_RTN_B32 Write-exchange operation.

  tmp = MEM[ADDR];
  MEM[ADDR] = DATA;
  RETURN_DATA = tmp.

12.13. LDS & GDS Instructions 194 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

46 DS_WRXCHG2_RTN_B3 Write-exchange 2 separate dwords.

2

47 DS_WRXCHG2ST64_RT Write-exchange 2 separate dwords with a stride of 64 dwords.

N_B32

48 DS_CMPST_RTN_B32 Compare and store. Caution, the order of src and cmp are the
*opposite* of the BUFFER_ATOMIC_CMPSWAP opcode.

  // 32bit
  tmp = MEM[ADDR];
  src = DATA2;
  cmp = DATA;
  MEM[ADDR] = (tmp == cmp) ? src : tmp;
  RETURN_DATA[0] = tmp.

49 DS_CMPST_RTN_F32 Floating point compare and store that handles NaN/INF/denormal

values. Caution, the order of src and cmp are the *opposite* of
the BUFFER_ATOMIC_FCMPSWAP opcode.

  // 32bit
  tmp = MEM[ADDR];
  src = DATA2;
  cmp = DATA;
  MEM[ADDR] = (tmp == cmp) ? src : tmp;
  RETURN_DATA[0] = tmp.

50 DS_MIN_RTN_F32 Floating point minimum that handles NaN/INF/denormal values.

Note that this opcode is slightly more general-purpose than
BUFFER_ATOMIC_FMIN.

  // 32bit
  tmp = MEM[ADDR];
  src = DATA;
  cmp = DATA2;
  MEM[ADDR] = (cmp < tmp) ? src : tmp.

51 DS_MAX_RTN_F32 Floating point maximum that handles NaN/INF/denormal values.

Note that this opcode is slightly more general-purpose than
BUFFER_ATOMIC_FMAX.

  // 32bit
  tmp = MEM[ADDR];
  src = DATA;
  cmp = DATA2;
  MEM[ADDR] = (tmp > cmp) ? src : tmp.

52 DS_WRAP_RTN_B32   tmp = MEM[ADDR];

  MEM[ADDR] = (tmp >= DATA) ? tmp - DATA : tmp + DATA2;
  RETURN_DATA = tmp.

53 DS_SWIZZLE_B32 Dword swizzle, no data is written to LDS memory. See next section
for details.

54 DS_READ_B32 Dword read.

  RETURN_DATA = MEM[ADDR].

12.13. LDS & GDS Instructions 195 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

55 DS_READ2_B32 Read 2 dwords.

  RETURN_DATA[0] = MEM[ADDR + OFFSET0 * 4];

  RETURN_DATA[1] = MEM[ADDR + OFFSET1 * 4].

56 DS_READ2ST64_B32 Read 2 dwords with a larger stride.

  RETURN_DATA[0] = MEM[ADDR + OFFSET0 * 4 * 64];

  RETURN_DATA[1] = MEM[ADDR + OFFSET1 * 4 * 64].

57 DS_READ_I8 Signed byte read.

  RETURN_DATA = signext(MEM[ADDR][7:0]).

58 DS_READ_U8 Unsigned byte read.

  RETURN_DATA = {24'h0,MEM[ADDR][7:0]}.

59 DS_READ_I16 Signed short read.

  RETURN_DATA = signext(MEM[ADDR][15:0]).

60 DS_READ_U16 Unsigned short read.

  RETURN_DATA = {16'h0,MEM[ADDR][15:0]}.

61 DS_CONSUME LDS & GDS. Subtract (count_bits(exec_mask)) from the value

stored in DS memory at (M0.base + instr_offset). Return the
pre-operation value to VGPRs.

The DS will subtract count_bits(vector valid mask) from the

value stored at address M0.base + instruction based offset and
return the pre-op value to all valid lanes. This op can be used
in both the LDS and GDS. In the LDS this address will be an
offset to HWBASE and clamped by M0.size, but in the GDS the
M0.base constant will have the physical GDS address and the
compiler must force offset to zero. In GDS it is for the
traditional append buffer operations. In LDS it is for local
thread group appends and can be used to regroup divergent
threads. The use of the M0 register enables the compiler to do
indexing of UAV append/consume counters.

For GDS (system wide) consume, the compiler must use a

zero for {offset1,offset0}, for LDS the compiler will use
{offset1,offset0} to provide the relative address to the append
counter in the LDS for runtime index offset or index.

Inside DS --- Do one atomic add for first valid lane and
broadcast result to all valid lanes. Offset = 0ffset1:offset0;
Interpreted as byte offset --- For 10xx LDS designs only aligned
atomics are supported, so 2 lsbs of offset must be set to zero.

  addr = M0.base + offset; // offset by LDS HWBASE, limit to M.size

  rtnval = LDS(addr);
  LDS(addr) = LDS(addr) - countbits(valid mask);
  GPR[VDST] = rtnval; // return to all valid threads

12.13. LDS & GDS Instructions 196 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

62 DS_APPEND LDS & GDS. Add (count_bits(exec_mask)) to the value stored in DS

memory at (M0.base + instr_offset). Return the pre-operation
value to VGPRs.

The DS will add count_bits(vector valid mask) from the value

stored at address M0.base + instruction based offset and return
the pre-op value to all valid lanes. This op can be used in
both the LDS and GDS. In the LDS this address will be an offset
to HWBASE and clamped by M0.size, but in the GDS the M0.base
constant will have the physical GDS address and the compiler
must set offset to zero. In GDS it is for the
traditional append buffer operations. In LDS it is for local
thread group appends and can be used to regroup divergent
threads. The use of the M0 register enables the compiler to do
indexing of UAV append/consume counters.

For GDS (system wide) consume, the compiler must use a

zero for {offset1,offset0}, for LDS the compiler will use
{offset1,offset0} to provide the relative address to the append
counter in the LDS for runtime index offset or index.

Inside DS --- Do one atomic add for first valid lane and
broadcast result to all valid lanes. Offset = 0ffset1:offset0;
Interpreted as byte offset --- For 10xx LDS designs only aligned
atomics will be supported, so 2 lsbs of offset must be set to zero.

  addr = M0.base + offset; // offset by LDS HWBASE, limit to M.size

  rtnval = LDS(addr);
  LDS(addr) = LDS(addr) + countbits(valid mask);
  GPR[VDST] = rtnval; // return to all valid threads

12.13. LDS & GDS Instructions 197 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

63 DS_ORDERED_COUNT GDS-only. Add (count_bits(exec_mask)) to one of 4 dedicated

ordered-count counters (aka 'packers'). Additional bits of
instr.offset field are overloaded to hold packer-id, 'last'.

GDS Only: Intercepted by GDS and processed by ordered append

module. The ordered append module will queue request until this
request wave is the oldest in the queue at which time the oldest
wave request will be dispatched to the DS with an atomic add for
execution and broadcast back to ALL lanes of a wave. This is an
ordered count operation and can only be called once per issue
with the release flag set. If the release flag is not set, the
wave will have full control over the order count module until it
sends a request with the release flag set.

Unlike append/consume this op needs to be sent even if there are

no valid lanes when it is issued. The GDS will add zero and
advance the tracking walker that needs to match up with the
dispatch counter.

The shader will send the following data to identify which

wave to return the result to:

The shader will send the following pipeline_ID to the ordered count
unit
to be used to select the correct pipeline's tracking data.
Additionally pixel waves will use 4 counters depending on
the packer sourcing the pixel waves and generating the launch order.

Pipeline_id = ring_id + !pixel wave type;

  0 = ring0 pixel wave
  1 = ring0 CS
  2 = ring1 CS
  3 = ring2 CS
Physical_wave_id = {se_id, sh_id, wave_buf_id}

GDS_size from the M0.size register contains the

pkr_id (set at wave creation time) and logical_wave_id for pixel
waves and
launch order logical wave_id for compute shaders.

The pixel shader uses four counters for each instance,

so the pkr_id will need to be added to the gds_base to act on
the correct counter.

  index = find first valid (vector mask)

  count = src0[index][31:0];
  Pkr_id = gds_size[1:0];
  gds_atomic_address[15:2] = gds_base[15:2] will contain the dword
address in the ds for the count accumulation counter.
  ds_address[15:2] = gds_base[15:2] + offset0[7:2] +
(pipeline_id == 0)?Pkr_id:0
  //2 new control signals
  Wave_release = Offset1[0];
  Wave_done = offset1[1];
  Pixel_wave = offset1[2];

If this control is not set, hold the crawler until wave does an
additional access with the wave_release the wave. This feature
allows one wavefront to issue serial access to the any of the
12.13. LDS & GDS Instructions 198 of 283
"RDNA 2" Instruction Set Architecture

Opcode Name Description

64 DS_ADD_U64   // 64bit
  tmp = MEM[ADDR];
  MEM[ADDR] += DATA[0:1];
  RETURN_DATA[0:1] = tmp.

65 DS_SUB_U64   // 64bit
  tmp = MEM[ADDR];
  MEM[ADDR] -= DATA[0:1];
  RETURN_DATA[0:1] = tmp.

66 DS_RSUB_U64 Subtraction with reversed operands.

  // 64bit
  tmp = MEM[ADDR];
  MEM[ADDR] = DATA - MEM[ADDR];
  RETURN_DATA = tmp.

67 DS_INC_U64   // 64bit
  tmp = MEM[ADDR];
  MEM[ADDR] = (tmp >= DATA[0:1]) ? 0 : tmp + 1; // unsigned compare
  RETURN_DATA[0:1] = tmp.

68 DS_DEC_U64   // 64bit
  tmp = MEM[ADDR];
  MEM[ADDR] = (tmp == 0 || tmp > DATA[0:1]) ? DATA[0:1] : tmp - 1;
// unsigned compare
  RETURN_DATA[0:1] = tmp.

69 DS_MIN_I64   // 64bit
  tmp = MEM[ADDR];
  MEM[ADDR] = (DATA[0:1] < tmp) ? DATA[0:1] : tmp; // signed
compare
  RETURN_DATA[0:1] = tmp.

70 DS_MAX_I64   // 64bit
  tmp = MEM[ADDR];
  MEM[ADDR] = (DATA[0:1] > tmp) ? DATA[0:1] : tmp; // signed
compare
  RETURN_DATA[0:1] = tmp.

71 DS_MIN_U64   // 64bit
  tmp = MEM[ADDR];
  MEM[ADDR] = (DATA[0:1] < tmp) ? DATA[0:1] : tmp; // unsigned
compare
  RETURN_DATA[0:1] = tmp.

72 DS_MAX_U64   // 64bit
  tmp = MEM[ADDR];
  MEM[ADDR] = (DATA[0:1] > tmp) ? DATA[0:1] : tmp; // unsigned
compare
  RETURN_DATA[0:1] = tmp.

73 DS_AND_B64   // 64bit
  tmp = MEM[ADDR];
  MEM[ADDR] &= DATA[0:1];
  RETURN_DATA[0:1] = tmp.

12.13. LDS & GDS Instructions 199 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

74 DS_OR_B64   // 64bit
  tmp = MEM[ADDR];
  MEM[ADDR] |= DATA[0:1];
  RETURN_DATA[0:1] = tmp.

75 DS_XOR_B64   // 64bit
  tmp = MEM[ADDR];
  MEM[ADDR] ^= DATA[0:1];
  RETURN_DATA[0:1] = tmp.

76 DS_MSKOR_B64 Masked dword OR, D0 contains the mask and D1 contains the new
value.

  // 64bit
  tmp = MEM[ADDR];
  MEM[ADDR] = (MEM[ADDR] & ~DATA) | DATA2;
  RETURN_DATA = tmp.

77 DS_WRITE_B64 Write qword.

  // 64bit
  MEM[ADDR] = DATA.

78 DS_WRITE2_B64 Write 2 qwords.

  // 64bit
  MEM[ADDR + OFFSET0 * 8] = DATA;
  MEM[ADDR + OFFSET1 * 8] = DATA2.

79 DS_WRITE2ST64_B64 Write 2 qwords with a larger stride.

  // 64bit
  MEM[ADDR + OFFSET0 * 8 * 64] = DATA;
  MEM[ADDR + OFFSET1 * 8 * 64] = DATA2.

80 DS_CMPST_B64 Compare and store. Caution, the order of src and cmp are the
*opposite* of the BUFFER_ATOMIC_CMPSWAP_X2 opcode.

  // 64bit
  tmp = MEM[ADDR];
  src = DATA2;
  cmp = DATA;
  MEM[ADDR] = (tmp == cmp) ? src : tmp;
  RETURN_DATA[0] = tmp.

81 DS_CMPST_F64 Floating point compare and store that handles NaN/INF/denormal

values. Caution, the order of src and cmp are the *opposite* of
the BUFFER_ATOMIC_FCMPSWAP_X2 opcode.

  // 64bit
  tmp = MEM[ADDR];
  src = DATA2;
  cmp = DATA;
  MEM[ADDR] = (tmp == cmp) ? src : tmp;
  RETURN_DATA[0] = tmp.

12.13. LDS & GDS Instructions 200 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

82 DS_MIN_F64 Floating point minimum that handles NaN/INF/denormal values.

Note that this opcode is slightly more general-purpose than
BUFFER_ATOMIC_FMIN_X2.

  // 64bit
  tmp = MEM[ADDR];
  src = DATA;
  cmp = DATA2;
  MEM[ADDR] = (cmp < tmp) ? src : tmp.

83 DS_MAX_F64 Floating point maximum that handles NaN/INF/denormal values.

Note that this opcode is slightly more general-purpose than
BUFFER_ATOMIC_FMAX_X2.

  // 64bit
  tmp = MEM[ADDR];
  src = DATA;
  cmp = DATA2;
  MEM[ADDR] = (tmp > cmp) ? src : tmp.

85 DS_ADD_RTN_F32 Floating point add that handles NaN/INF/denormal values.

  float tmp = MEM[ADDR].f32;

  MEM[ADDR].f32 += DATA0.f32;
  VDST.f32 = tmp;

96 DS_ADD_RTN_U64   // 64bit
  tmp = MEM[ADDR];
  MEM[ADDR] += DATA[0:1];
  RETURN_DATA[0:1] = tmp.

97 DS_SUB_RTN_U64   // 64bit
  tmp = MEM[ADDR];
  MEM[ADDR] -= DATA[0:1];
  RETURN_DATA[0:1] = tmp.

98 DS_RSUB_RTN_U64 Subtraction with reversed operands.

  // 64bit
  tmp = MEM[ADDR];
  MEM[ADDR] = DATA - MEM[ADDR];
  RETURN_DATA = tmp.

99 DS_INC_RTN_U64   // 64bit
  tmp = MEM[ADDR];
  MEM[ADDR] = (tmp >= DATA[0:1]) ? 0 : tmp + 1; // unsigned compare
  RETURN_DATA[0:1] = tmp.

100 DS_DEC_RTN_U64   // 64bit

  tmp = MEM[ADDR];
  MEM[ADDR] = (tmp == 0 || tmp > DATA[0:1]) ? DATA[0:1] : tmp - 1;
// unsigned compare
  RETURN_DATA[0:1] = tmp.

101 DS_MIN_RTN_I64   // 64bit

  tmp = MEM[ADDR];
  MEM[ADDR] = (DATA[0:1] < tmp) ? DATA[0:1] : tmp; // signed
compare
  RETURN_DATA[0:1] = tmp.

12.13. LDS & GDS Instructions 201 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

102 DS_MAX_RTN_I64   // 64bit

  tmp = MEM[ADDR];
  MEM[ADDR] = (DATA[0:1] > tmp) ? DATA[0:1] : tmp; // signed
compare
  RETURN_DATA[0:1] = tmp.

103 DS_MIN_RTN_U64   // 64bit

  tmp = MEM[ADDR];
  MEM[ADDR] = (DATA[0:1] < tmp) ? DATA[0:1] : tmp; // unsigned
compare
  RETURN_DATA[0:1] = tmp.

104 DS_MAX_RTN_U64   // 64bit

  tmp = MEM[ADDR];
  MEM[ADDR] = (DATA[0:1] > tmp) ? DATA[0:1] : tmp; // unsigned
compare
  RETURN_DATA[0:1] = tmp.

105 DS_AND_RTN_B64   // 64bit

  tmp = MEM[ADDR];
  MEM[ADDR] &= DATA[0:1];
  RETURN_DATA[0:1] = tmp.

106 DS_OR_RTN_B64   // 64bit

  tmp = MEM[ADDR];
  MEM[ADDR] |= DATA[0:1];
  RETURN_DATA[0:1] = tmp.

107 DS_XOR_RTN_B64   // 64bit

  tmp = MEM[ADDR];
  MEM[ADDR] ^= DATA[0:1];
  RETURN_DATA[0:1] = tmp.

108 DS_MSKOR_RTN_B64 Masked dword OR, D0 contains the mask and D1 contains the new
value.

  // 64bit
  tmp = MEM[ADDR];
  MEM[ADDR] = (MEM[ADDR] & ~DATA) | DATA2;
  RETURN_DATA = tmp.

109 DS_WRXCHG_RTN_B64 Write-exchange operation.

  tmp = MEM[ADDR];
  MEM[ADDR] = DATA;
  RETURN_DATA = tmp.

110 DS_WRXCHG2_RTN_B6 Write-exchange 2 separate qwords.

4

111 DS_WRXCHG2ST64_RT Write-exchange 2 qwords with a stride of 64 qwords.

N_B64

12.13. LDS & GDS Instructions 202 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

112 DS_CMPST_RTN_B64 Compare and store. Caution, the order of src and cmp are the
*opposite* of the BUFFER_ATOMIC_CMPSWAP_X2 opcode.

  // 64bit
  tmp = MEM[ADDR];
  src = DATA2;
  cmp = DATA;
  MEM[ADDR] = (tmp == cmp) ? src : tmp;
  RETURN_DATA[0] = tmp.

113 DS_CMPST_RTN_F64 Floating point compare and store that handles NaN/INF/denormal
values. Caution, the order of src and cmp are the *opposite* of
the BUFFER_ATOMIC_FCMPSWAP_X2 opcode.

  // 64bit
  tmp = MEM[ADDR];
  src = DATA2;
  cmp = DATA;
  MEM[ADDR] = (tmp == cmp) ? src : tmp;
  RETURN_DATA[0] = tmp.

114 DS_MIN_RTN_F64 Floating point minimum that handles NaN/INF/denormal values.

Note that this opcode is slightly more general-purpose than
BUFFER_ATOMIC_FMIN_X2.

  // 64bit
  tmp = MEM[ADDR];
  src = DATA;
  cmp = DATA2;
  MEM[ADDR] = (cmp < tmp) ? src : tmp.

115 DS_MAX_RTN_F64 Floating point maximum that handles NaN/INF/denormal values.

Note that this opcode is slightly more general-purpose than
BUFFER_ATOMIC_FMAX_X2.

  // 64bit
  tmp = MEM[ADDR];
  src = DATA;
  cmp = DATA2;
  MEM[ADDR] = (tmp > cmp) ? src : tmp.

118 DS_READ_B64 Read 1 qword.

  RETURN_DATA = MEM[ADDR].

119 DS_READ2_B64 Read 2 qwords.

  RETURN_DATA[0] = MEM[ADDR + OFFSET0 * 8];

  RETURN_DATA[1] = MEM[ADDR + OFFSET1 * 8].

120 DS_READ2ST64_B64 Read 2 qwords with a larger stride.

  RETURN_DATA[0] = MEM[ADDR + OFFSET0 * 8 * 64];

  RETURN_DATA[1] = MEM[ADDR + OFFSET1 * 8 * 64].

126 DS_CONDXCHG32_RTN Conditional write exchange.

_B64

12.13. LDS & GDS Instructions 203 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

160 DS_WRITE_B8_D16_HI Byte write in to high word.

  MEM[ADDR] = DATA[23:16].

161 DS_WRITE_B16_D16_HI Short write in to high word.

  MEM[ADDR] = DATA[31:16].

162 DS_READ_U8_D16 Unsigned byte read with masked return to lower word.

  RETURN_DATA[15:0] = {8'h0,MEM[ADDR][7:0]}.

163 DS_READ_U8_D16_HI Unsigned byte read with masked return to upper word.

  RETURN_DATA[31:16] = {8'h0,MEM[ADDR][7:0]}.

164 DS_READ_I8_D16 Signed byte read with masked return to lower word.

  RETURN_DATA[15:0] = signext(MEM[ADDR][7:0]).

165 DS_READ_I8_D16_HI Signed byte read with masked return to upper word.

  RETURN_DATA[31:16] = signext(MEM[ADDR][7:0]).

166 DS_READ_U16_D16 Unsigned short read with masked return to lower word.

  RETURN_DATA[15:0] = MEM[ADDR][15:0].

167 DS_READ_U16_D16_HI Unsigned short read with masked return to upper word.

  RETURN_DATA[31:0] = MEM[ADDR][15:0].

176 DS_WRITE_ADDTID_B32 Write dword with thread ID offset.

  LDS_GS[LDS_BASE + {OFFSET1,OFFSET0} + M0[15:0] + TID*4].u32 =

VGPR[DATA0].u32

177 DS_READ_ADDTID_B32 Dword read with thread ID offset.

  VGPR[VDST].u32 = LDS_GS[LDS_BASE + {OFFSET1,OFFSET0} + M0[15:0] +

TID*4].u32

12.13. LDS & GDS Instructions 204 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

178 DS_PERMUTE_B32   // VGPR[index][thread_id] is the VGPR RAM

  // VDST, ADDR and DATA0 are from the microcode DS encoding
  tmp[0..63] = 0
  for i in 0..63 do
  // If a source thread is disabled, it will not propagate data.
  next if !EXEC[i]
  // ADDR needs to be divided by 4.
  // High-order bits are ignored.
  dst_lane = floor((VGPR[ADDR][i] + OFFSET) / 4) mod 64
  tmp[dst_lane] = VGPR[DATA0][i]
  endfor
  // Copy data into destination VGPRs. If multiple sources
  // select the same destination thread, the highest-numbered
  // source thread wins.
  for i in 0..63 do
  next if !EXEC[i]
  VGPR[VDST][i] = tmp[i]
  endfor

Forward permute. This does not access LDS memory and may be
called even if no LDS memory is allocated to the wave. It uses
LDS hardware to implement an arbitrary swizzle across threads in
a wavefront.

Note the address passed in is the thread ID multiplied by 4.

If multiple sources map to the same destination lane, the final

value is not predictable but will be the value from one of the
writers.

See also DS_BPERMUTE_B32.

Examples (simplified 4-thread wavefronts):

  VGPR[SRC0] = { A, B, C, D }
  VGPR[ADDR] = { 0, 0, 12, 4 }
  EXEC = 0xF, OFFSET = 0
  VGPR[VDST] := { B, D, 0, C }

  VGPR[SRC0] = { A, B, C, D }
  VGPR[ADDR] = { 0, 0, 12, 4 }
  EXEC = 0xA, OFFSET = 0
  VGPR[VDST] := { -, D, -, 0 }

12.13. LDS & GDS Instructions 205 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

179 DS_BPERMUTE_B32   // VGPR[index][thread_id] is the VGPR RAM

  // VDST, ADDR and DATA0 are from the microcode DS encoding
  tmp[0..63] = 0
  for i in 0..63 do
  // ADDR needs to be divided by 4.
  // High-order bits are ignored.
  src_lane = floor((VGPR[ADDR][i] + OFFSET) / 4) mod 64
  // EXEC is applied to the source VGPR reads.
  next if !EXEC[src_lane]
  tmp[i] = VGPR[DATA0][src_lane]
  endfor
  // Copy data into destination VGPRs. Some source
  // data may be broadcast to multiple lanes.
  for i in 0..63 do
  next if !EXEC[i]
  VGPR[VDST][i] = tmp[i]
  endfor

Backward permute. This does not access LDS memory and may be
called even if no LDS memory is allocated to the wave. It uses
LDS hardware to implement an arbitrary swizzle across threads in
a wavefront.

Note the address passed in is the thread ID multiplied by 4.

Note that EXEC mask is applied to both VGPR read and write. If
src_lane selects a disabled thread, zero will be returned.

See also DS_PERMUTE_B32.

Examples (simplified 4-thread wavefronts):

  VGPR[SRC0] = { A, B, C, D }
  VGPR[ADDR] = { 0, 0, 12, 4 }
  EXEC = 0xF, OFFSET = 0
  VGPR[VDST] := { A, A, D, B }

  VGPR[SRC0] = { A, B, C, D }
  VGPR[ADDR] = { 0, 0, 12, 4 }
  EXEC = 0xA, OFFSET = 0
  VGPR[VDST] := { -, 0, -, B }

222 DS_WRITE_B96 Tri-dword write.

  {MEM[ADDR + 8], MEM[ADDR + 4], MEM[ADDR]} = DATA[95:0].

223 DS_WRITE_B128 Quad-dword write.

  {MEM[ADDR + 12], MEM[ADDR + 8], MEM[ADDR + 4], MEM[ADDR]} =

DATA[127:0].

254 DS_READ_B96 Tri-dword read.

255 DS_READ_B128 Quad-dword read.

12.13. LDS & GDS Instructions 206 of 283

"RDNA 2" Instruction Set Architecture

12.13.1. DS_SWIZZLE_B32 Details

Dword swizzle, no data is written to LDS memory.

Swizzles input thread data based on offset mask and returns; note does not read or write the DS
memory banks.

Note that reading from an invalid thread results in 0x0.

This opcode supports two specific modes, FFT and rotate, plus two basic modes which swizzle
in groups of 4 or 32 consecutive threads.

The FFT mode (offset >= 0xe000) swizzles the input based on offset[4:0] to support FFT
calculation. Example swizzles using input {1, 2, … 20} are:

Offset[4:0]: Swizzle 0x00:

{1,11,9,19,5,15,d,1d,3,13,b,1b,7,17,f,1f,2,12,a,1a,6,16,e,1e,4,14,c,1c,8,18,10,20} 0x10:
{1,9,5,d,3,b,7,f,2,a,6,e,4,c,8,10,11,19,15,1d,13,1b,17,1f,12,1a,16,1e,14,1c,18,20} 0x1f: No
swizzle

The rotate mode (offset >= 0xc000 and offset < 0xe000) rotates the input either left (offset[10]
== 0) or right (offset[10] == 1) a number of threads equal to offset[9:5]. The rotate mode also
uses a mask value which can alter the rotate result. For example, mask == 1 will swap the odd
threads across every other even thread (rotate left), or even threads across every other odd
thread (rotate right).

Offset[9:5]: Swizzle 0x01, mask=0, rotate left:

{2,3,4,5,6,7,8,9,a,b,c,d,e,f,10,11,12,13,14,15,16,17,18,19,1a,1b,1c,1d,1e,1f,20,1} 0x01,
mask=0, rotate right:
{20,1,2,3,4,5,6,7,8,9,a,b,c,d,e,f,10,11,12,13,14,15,16,17,18,19,1a,1b,1c,1d,1e,1f} 0x01,
mask=1, rotate left:
{2,1,4,7,6,5,8,b,a,9,c,f,e,d,10,13,12,11,14,17,16,15,18,1b,1a,19,1c,1f,1e,1d,20,3} 0x01,
mask=1, rotate right:
{1e,1,4,3,2,5,8,7,6,9,c,b,a,d,10,f,e,11,14,13,12,15,18,17,16,19,1c,1b,1a,1d,20,1f}

If offset < 0xc000, one of the basic swizzle modes is used based on offset[15]. If offset[15] == 1,
groups of 4 consecutive threads are swizzled together. If offset[15] == 0, all 32 threads are
swizzled together.

The first basic swizzle mode (when offset[15] == 1) allows full data sharing between a group of 4
consecutive threads. Any thread within the group of 4 can get data from any other thread within
the group of 4, specified by the corresponding offset bits --- [1:0] for the first thread, [3:2] for the
second thread, [5:4] for the third thread, [7:6] for the fourth thread. Note that the offset bits apply
to all groups of 4 within a wavefront; thus if offset[1:0] == 1, then thread0 will grab thread1,
thread4 will grab thread5, etc.

The second basic swizzle mode (when offset[15] == 0) allows limited data sharing between 32

12.13. LDS & GDS Instructions 207 of 283

"RDNA 2" Instruction Set Architecture

consecutive threads. In this case, the offset is used to specify a 5-bit xor-mask, 5-bit or-mask,
and 5-bit and-mask used to generate a thread mapping. Note that the offset bits apply to each
group of 32 within a wavefront. The details of the thread mapping are listed below. Some
example usages:

SWAPX16 : xor_mask = 0x10, or_mask = 0x00, and_mask = 0x1f

SWAPX8 : xor_mask = 0x08, or_mask = 0x00, and_mask = 0x1f

SWAPX4 : xor_mask = 0x04, or_mask = 0x00, and_mask = 0x1f

SWAPX2 : xor_mask = 0x02, or_mask = 0x00, and_mask = 0x1f

SWAPX1 : xor_mask = 0x01, or_mask = 0x00, and_mask = 0x1f

REVERSEX32 : xor_mask = 0x1f, or_mask = 0x00, and_mask = 0x1f

REVERSEX16 : xor_mask = 0x0f, or_mask = 0x00, and_mask = 0x1f

REVERSEX8 : xor_mask = 0x07, or_mask = 0x00, and_mask = 0x1f

REVERSEX4 : xor_mask = 0x03, or_mask = 0x00, and_mask = 0x1f

REVERSEX2 : xor_mask = 0x01 or_mask = 0x00, and_mask = 0x1f

BCASTX32: xor_mask = 0x00, or_mask = thread, and_mask = 0x00

BCASTX16: xor_mask = 0x00, or_mask = thread, and_mask = 0x10

BCASTX8: xor_mask = 0x00, or_mask = thread, and_mask = 0x18

BCASTX4: xor_mask = 0x00, or_mask = thread, and_mask = 0x1c

BCASTX2: xor_mask = 0x00, or_mask = thread, and_mask = 0x1e

Pseudocode follows:

offset = offset1:offset0;

12.13. LDS & GDS Instructions 208 of 283

"RDNA 2" Instruction Set Architecture

if (offset >= 0xe000) {

  // FFT decomposition
  mask = offset[4:0];
  for (i = 0; i < 64; i++) {
  j = reverse_bits(i & 0x1f);
  j = (j >> count_ones(mask));
  j \|= (i & mask);
  j \|= i & 0x20;
  thread_out[i] = thread_valid[j] ? thread_in[j] : 0;
  }

} else if (offset >= 0xc000) { // rotate rotate = offset[9:5]; mask = offset[4:0]; if (offset[10]) { rotate
= -rotate; } for (i = 0; i < 64; i++) { j = (i & mask) \| ((i + rotate) & ~mask); j \|= i & 0x20;
thread_out[i] = thread_valid[j] ? thread_in[j] : 0; }

} else if (offset[15]) { // full data sharing within 4 consecutive threads for (i = 0; i < 64; i+=4) {
thread_out[i+0] = thread_valid[i+offset[1:0]]?thread_in[i+offset[1:0]]:0; thread_out[i+1] =
thread_valid[i+offset[3:2]]?thread_in[i+offset[3:2]]:0; thread_out[i+2] =
thread_valid[i+offset[5:4]]?thread_in[i+offset[5:4]]:0; thread_out[i+3] =
thread_valid[i+offset[7:6]]?thread_in[i+offset[7:6]]:0; }

} else { // offset[15] == 0 // limited data sharing within 32 consecutive threads xor_mask =

offset[14:10]; or_mask = offset[9:5]; and_mask = offset[4:0]; for (i = 0; i < 64; i++) { j = (((i & 0x1f)
& and_mask) \| or_mask) ^ xor_mask; j \|= (i & 0x20); // which group of 32 thread_out[i] =
thread_valid[j] ? thread_in[j] : 0; } }

12.13.2. LDS Instruction Limitations

Some of the DS instructions are available only to GDS, not LDS. These are:

• DS_GWS_SEMA_RELEASE_ALL
• DS_GWS_INIT
• DS_GWS_SEMA_V
• DS_GWS_SEMA_BR
• DS_GWS_SEMA_P
• DS_GWS_BARRIER
• DS_ORDERED_COUNT

12.14. MUBUF Instructions

The bitfield map of the MUBUF format is:

12.14. MUBUF Instructions 209 of 283

"RDNA 2" Instruction Set Architecture

  where:

  OFFSET = Unsigned immediate byte offset.

  OFFEN = Send offset either as VADDR or as zero..
  IDXEN = Send index either as VADDR or as zero.
  GLC = Global coherency.
  LDS = Data read from/written to LDS or VGPR.
  OP = Instruction Opcode.
  VADDR = VGPR address source.
  VDATA = Destination vector GPR.
  SRSRC = Scalar GPR that specifies resource constant.
  SLC = System level coherent.
  TFE = Texture fail enable.
  SOFFSET = Byte offset added to the memory address of an SGPR.

Opcode Name Description

0 BUFFER_LOAD_FORMAT_X Untyped buffer load 1 dword with format conversion.

1 BUFFER_LOAD_FORMAT_XY Untyped buffer load 2 dwords with format conversion.

2 BUFFER_LOAD_FORMAT_XYZ Untyped buffer load 3 dwords with format conversion.

3 BUFFER_LOAD_FORMAT_XYZW Untyped buffer load 4 dwords with format conversion.

4 BUFFER_STORE_FORMAT_X Untyped buffer store 1 dword with format conversion.

5 BUFFER_STORE_FORMAT_XY Untyped buffer store 2 dwords with format conversion.

6 BUFFER_STORE_FORMAT_XYZ Untyped buffer store 3 dwords with format conversion.

7 BUFFER_STORE_FORMAT_XYZW Untyped buffer store 4 dwords with format conversion.

8 BUFFER_LOAD_UBYTE Untyped buffer load unsigned byte (zero extend to VGPR

destination).

9 BUFFER_LOAD_SBYTE Untyped buffer load signed byte (sign extend to VGPR

destination).

10 BUFFER_LOAD_USHORT Untyped buffer load unsigned short (zero extend to VGPR

destination).

11 BUFFER_LOAD_SSHORT Untyped buffer load signed short (sign extend to VGPR

destination).

12 BUFFER_LOAD_DWORD Untyped buffer load dword.

13 BUFFER_LOAD_DWORDX2 Untyped buffer load 2 dwords.

14 BUFFER_LOAD_DWORDX4 Untyped buffer load 4 dwords.

15 BUFFER_LOAD_DWORDX3 Untyped buffer load 3 dwords.

24 BUFFER_STORE_BYTE Untyped buffer store byte. Stores S0[7:0].

25 BUFFER_STORE_BYTE_D16_HI Untyped buffer store byte. Stores S0[23:16].

26 BUFFER_STORE_SHORT Untyped buffer store short. Stores S0[15:0].

27 BUFFER_STORE_SHORT_D16_HI Untyped buffer store short. Stores S0[31:16].

28 BUFFER_STORE_DWORD Untyped buffer store dword.

29 BUFFER_STORE_DWORDX2 Untyped buffer store 2 dwords.

12.14. MUBUF Instructions 210 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

30 BUFFER_STORE_DWORDX4 Untyped buffer store 4 dwords.

31 BUFFER_STORE_DWORDX3 Untyped buffer store 3 dwords.

32 BUFFER_LOAD_UBYTE_D16   D0[15:0] = {8'h0, MEM[ADDR]}.

Untyped buffer load unsigned byte.

33 BUFFER_LOAD_UBYTE_D16_HI   D0[31:16] = {8'h0, MEM[ADDR]}.

Untyped buffer load unsigned byte.

34 BUFFER_LOAD_SBYTE_D16   D0[15:0] = signext(MEM[ADDR]).

Untyped buffer load signed byte.

35 BUFFER_LOAD_SBYTE_D16_HI   D0[31:16] = signext(MEM[ADDR]).

Untyped buffer load signed byte.

36 BUFFER_LOAD_SHORT_D16   D0[15:0] = MEM[ADDR].

Untyped buffer load short.

37 BUFFER_LOAD_SHORT_D16_HI   D0[31:16] = MEM[ADDR].

Untyped buffer load short.

38 BUFFER_LOAD_FORMAT_D16_HI   D0[31:16] = MEM[ADDR].

_X
Untyped buffer load 1 dword with format conversion.

39 BUFFER_STORE_FORMAT_D16_ Untyped buffer store 1 dword with format conversion.

HI_X

48 BUFFER_ATOMIC_SWAP   // 32bit
  tmp = MEM[ADDR];
  MEM[ADDR] = DATA;
  RETURN_DATA = tmp.

49 BUFFER_ATOMIC_CMPSWAP   // 32bit
  tmp = MEM[ADDR];
  src = DATA[0];
  cmp = DATA[1];
  MEM[ADDR] = (tmp == cmp) ? src : tmp;
  RETURN_DATA[0] = tmp.

50 BUFFER_ATOMIC_ADD   // 32bit
  tmp = MEM[ADDR];
  MEM[ADDR] += DATA;
  RETURN_DATA = tmp.

51 BUFFER_ATOMIC_SUB   // 32bit
  tmp = MEM[ADDR];
  MEM[ADDR] -= DATA;
  RETURN_DATA = tmp.

12.14. MUBUF Instructions 211 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

52 BUFFER_ATOMIC_CSUB   // 32bit
  old_value = MEM[ADDR];
  if old_value < DATA then
  new_value = 0;
  else
  new_value = old_value - DATA;
  endif;
  MEM[addr] = new_value;
  RETURN_DATA = old_value.

53 BUFFER_ATOMIC_SMIN   // 32bit
  tmp = MEM[ADDR];
  MEM[ADDR] = (DATA < tmp) ? DATA : tmp; // signed compare
  RETURN_DATA = tmp.

54 BUFFER_ATOMIC_UMIN   // 32bit
  tmp = MEM[ADDR];
  MEM[ADDR] = (DATA < tmp) ? DATA : tmp; // unsigned
compare
  RETURN_DATA = tmp.

55 BUFFER_ATOMIC_SMAX   // 32bit
  tmp = MEM[ADDR];
  MEM[ADDR] = (DATA > tmp) ? DATA : tmp; // signed compare
  RETURN_DATA = tmp.

56 BUFFER_ATOMIC_UMAX   // 32bit
  tmp = MEM[ADDR];
  MEM[ADDR] = (DATA > tmp) ? DATA : tmp; // unsigned
compare
  RETURN_DATA = tmp.

57 BUFFER_ATOMIC_AND   // 32bit
  tmp = MEM[ADDR];
  MEM[ADDR] &= DATA;
  RETURN_DATA = tmp.

58 BUFFER_ATOMIC_OR   // 32bit
  tmp = MEM[ADDR];
  MEM[ADDR] |= DATA;
  RETURN_DATA = tmp.

59 BUFFER_ATOMIC_XOR   // 32bit
  tmp = MEM[ADDR];
  MEM[ADDR] ^= DATA;
  RETURN_DATA = tmp.

60 BUFFER_ATOMIC_INC   // 32bit
  tmp = MEM[ADDR];
  MEM[ADDR] = (tmp >= DATA) ? 0 : tmp + 1; // unsigned
compare
  RETURN_DATA = tmp.

61 BUFFER_ATOMIC_DEC   // 32bit
  tmp = MEM[ADDR];
  MEM[ADDR] = (tmp == 0 || tmp > DATA) ? DATA : tmp - 1;
// unsigned compare
  RETURN_DATA = tmp.

12.14. MUBUF Instructions 212 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

62 BUFFER_ATOMIC_FCMPSWAP   // 32bit
  tmp = MEM[ADDR];
  src = DATA[0];
  cmp = DATA[1];
  MEM[ADDR] = (tmp == cmp) ? src : tmp;
  RETURN_DATA[0] = tmp.

Floating-point compare swap (handles NaN/INF/denorm).

63 BUFFER_ATOMIC_FMIN   // 32bit
  tmp = MEM[ADDR];
  src = DATA[0];
  MEM[ADDR] = (src < tmp) ? src : tmp;
  RETURN_DATA[0] = tmp.

Floating-point compare (handles NaN/INF/denorm).

64 BUFFER_ATOMIC_FMAX   // 32bit
  tmp = MEM[ADDR];
  src = DATA[0];
  MEM[ADDR] = (src > tmp) ? src : tmp;
  RETURN_DATA[0] = tmp.

Floating-point compare (handles NaN/INF/denorm).

80 BUFFER_ATOMIC_SWAP_X2   // 64bit
  tmp = MEM[ADDR];
  MEM[ADDR] = DATA[0:1];
  RETURN_DATA[0:1] = tmp.

81 BUFFER_ATOMIC_CMPSWAP_X2   // 64bit
  tmp = MEM[ADDR];
  src = DATA[0:1];
  cmp = DATA[2:3];
  MEM[ADDR] = (tmp == cmp) ? src : tmp;
  RETURN_DATA[0:1] = tmp.

82 BUFFER_ATOMIC_ADD_X2   // 64bit
  tmp = MEM[ADDR];
  MEM[ADDR] += DATA[0:1];
  RETURN_DATA[0:1] = tmp.

83 BUFFER_ATOMIC_SUB_X2   // 64bit
  tmp = MEM[ADDR];
  MEM[ADDR] -= DATA[0:1];
  RETURN_DATA[0:1] = tmp.

85 BUFFER_ATOMIC_SMIN_X2   // 64bit
  tmp = MEM[ADDR];
  MEM[ADDR] = (DATA[0:1] < tmp) ? DATA[0:1] : tmp; //
signed compare
  RETURN_DATA[0:1] = tmp.

86 BUFFER_ATOMIC_UMIN_X2   // 64bit
  tmp = MEM[ADDR];
  MEM[ADDR] = (DATA[0:1] < tmp) ? DATA[0:1] : tmp; //
unsigned compare
  RETURN_DATA[0:1] = tmp.

12.14. MUBUF Instructions 213 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

87 BUFFER_ATOMIC_SMAX_X2   // 64bit
  tmp = MEM[ADDR];
  MEM[ADDR] = (DATA[0:1] > tmp) ? DATA[0:1] : tmp; //
signed compare
  RETURN_DATA[0:1] = tmp.

88 BUFFER_ATOMIC_UMAX_X2   // 64bit
  tmp = MEM[ADDR];
  MEM[ADDR] = (DATA[0:1] > tmp) ? DATA[0:1] : tmp; //
unsigned compare
  RETURN_DATA[0:1] = tmp.

89 BUFFER_ATOMIC_AND_X2   // 64bit
  tmp = MEM[ADDR];
  MEM[ADDR] &= DATA[0:1];
  RETURN_DATA[0:1] = tmp.

90 BUFFER_ATOMIC_OR_X2   // 64bit
  tmp = MEM[ADDR];
  MEM[ADDR] |= DATA[0:1];
  RETURN_DATA[0:1] = tmp.

91 BUFFER_ATOMIC_XOR_X2   // 64bit
  tmp = MEM[ADDR];
  MEM[ADDR] ^= DATA[0:1];
  RETURN_DATA[0:1] = tmp.

92 BUFFER_ATOMIC_INC_X2   // 64bit
  tmp = MEM[ADDR];
  MEM[ADDR] = (tmp >= DATA[0:1]) ? 0 : tmp + 1; //
unsigned compare
  RETURN_DATA[0:1] = tmp.

93 BUFFER_ATOMIC_DEC_X2   // 64bit
  tmp = MEM[ADDR];
  MEM[ADDR] = (tmp == 0 || tmp > DATA[0:1]) ? DATA[0:1] :
tmp - 1; // unsigned compare
  RETURN_DATA[0:1] = tmp.

94 BUFFER_ATOMIC_FCMPSWAP_X   // 64bit
2   tmp = MEM[ADDR];
  src = DATA[0];
  cmp = DATA[1];
  MEM[ADDR] = (tmp == cmp) ? src : tmp;
  RETURN_DATA[0] = tmp.

Floating-point compare swap (handles NaN/INF/denorm).

95 BUFFER_ATOMIC_FMIN_X2   // 64bit
  tmp = MEM[ADDR];
  src = DATA[0];
  MEM[ADDR] = (src < tmp) ? src : tmp;
  RETURN_DATA[0] = tmp.

Floating-point compare (handles NaN/INF/denorm).

12.14. MUBUF Instructions 214 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

96 BUFFER_ATOMIC_FMAX_X2   // 64bit
  tmp = MEM[ADDR];
  src = DATA[0];
  MEM[ADDR] = (src > tmp) ? src : tmp;
  RETURN_DATA[0] = tmp.

Floating-point compare (handles NaN/INF/denorm).

113 BUFFER_GL0_INV Write back and invalidate the shader L0. Returns ACK to
shader.

114 BUFFER_GL1_INV Invalidate the GL1 cache only. Returns ACK to shader.

128 BUFFER_LOAD_FORMAT_D16_X Untyped buffer load 1 dword with format conversion.

D0[15:0] = MEM[ADDR].

129 BUFFER_LOAD_FORMAT_D16_XY Untyped buffer load 1 dword with format conversion.

130 BUFFER_LOAD_FORMAT_D16_XY Untyped buffer load 2 dwords with format conversion.

Z

131 BUFFER_LOAD_FORMAT_D16_XY Untyped buffer load 2 dwords with format conversion.

ZW

132 BUFFER_STORE_FORMAT_D16_X Untyped buffer store 1 dword with format conversion.

133 BUFFER_STORE_FORMAT_D16_ Untyped buffer store 1 dword with format conversion.

XY

134 BUFFER_STORE_FORMAT_D16_ Untyped buffer store 2 dwords with format conversion.

XYZ

135 BUFFER_STORE_FORMAT_D16_ Untyped buffer store 2 dwords with format conversion.

XYZW

12.15. MTBUF Instructions

The bitfield map of the MTBUF format is:

  where:

  OFFSET = Unsigned immediate byte offset.

  OFFEN = Send offset either as VADDR or as zero.
  IDXEN = Send index either as VADDR or as zero.
  GLC = Global coherency.
  OP = Instruction Opcode.
  FORMAT = Data format for typed buffer.
  VADDR = VGPR address source.
  VDATA = Vector GPR for read/write result.
  SRSRC = Scalar GPR that specifies resource constant.
  SOFFSET = Unsigned byte offset from an SGPR.

12.15. MTBUF Instructions 215 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

0 TBUFFER_LOAD_FORMAT_X Typed buffer load 1 dword with format conversion.

1 TBUFFER_LOAD_FORMAT_XY Typed buffer load 2 dwords with format conversion.

2 TBUFFER_LOAD_FORMAT_XYZ Typed buffer load 3 dwords with format conversion.

3 TBUFFER_LOAD_FORMAT_XYZW Typed buffer load 4 dwords with format conversion.

4 TBUFFER_STORE_FORMAT_X Typed buffer store 1 dword with format conversion.

5 TBUFFER_STORE_FORMAT_XY Typed buffer store 2 dwords with format conversion.

6 TBUFFER_STORE_FORMAT_XYZ Typed buffer store 3 dwords with format conversion.

7 TBUFFER_STORE_FORMAT_XYZW Typed buffer store 4 dwords with format conversion.

8 TBUFFER_LOAD_FORMAT_D16_X Typed buffer load 1 dword with format conversion.

9 TBUFFER_LOAD_FORMAT_D16_XY Typed buffer load 1 dword with format conversion.

10 TBUFFER_LOAD_FORMAT_D16_XY Typed buffer load 2 dwords with format conversion.

Z

11 TBUFFER_LOAD_FORMAT_D16_XY Typed buffer load 2 dwords with format conversion.

ZW

12 TBUFFER_STORE_FORMAT_D16_X Typed buffer store 1 dword with format conversion.

13 TBUFFER_STORE_FORMAT_D16_X Typed buffer store 1 dword with format conversion.

Y

14 TBUFFER_STORE_FORMAT_D16_X Typed buffer store 2 dwords with format conversion.

YZ

15 TBUFFER_STORE_FORMAT_D16_X Typed buffer store 2 dwords with format conversion.

YZW

12.16. MIMG Instructions

The bitfield map of the MIMG format is:

12.16. MIMG Instructions 216 of 283

"RDNA 2" Instruction Set Architecture

  where:

  DMASK = Enable mask for image read/write data components.

  UNRM = Force address to be unnormalized.
  GLC = Global coherency.
  DA = Declare an array.
  A16 = Texture address component size.
  TFE = Texture fail enable.
  LWE = LOD warning enable.
  OP = Instruction Opcode.
  SLC = System level coherent.
  VADDR = VGPR address source.
  VDATA = Vector GPR for read/write result.
  SRSRC = Scalar GPR that specifies resource constant.
  SSAMP = Scalar GPR that specifies sampler constant.
  D16 = Data in VGPRs is 16 bits, not 32 bits.

Opcode Name Description

0 IMAGE_LOAD Load element from largest miplevel in resource view,

with format conversion specified in the resource
constant. No sampler.

1 IMAGE_LOAD_MIP Load element from user-specified miplevel in resource

view, with format conversion specified in the resource
constant. No sampler.

2 IMAGE_LOAD_PCK Load element from largest miplevel in resource view,

without format conversion. 8- and 16-bit elements are
not sign-extended. No sampler.

3 IMAGE_LOAD_PCK_SGN Load element from largest miplevel in resource view,

without format conversion. 8- and 16-bit elements are
sign-extended. No sampler.

4 IMAGE_LOAD_MIP_PCK Load element from user-supplied miplevel in resource

view, without format conversion. 8- and 16-bit elements
are not sign-extended. No sampler.

5 IMAGE_LOAD_MIP_PCK_SGN Load element from user-supplied miplevel in resource

view, without format conversion. 8- and 16-bit elements
are sign-extended. No sampler.

8 IMAGE_STORE Store element to largest miplevel in resource view, with

format conversion specified in resource constant. No
sampler.

9 IMAGE_STORE_MIP Store element to user-specified miplevel in resource

view, with format conversion specified in resource
constant. No sampler.

10 IMAGE_STORE_PCK Store element to largest miplevel in resource view,

without format conversion. No sampler.

11 IMAGE_STORE_MIP_PCK Store element to user-specified miplevel in resource

view, without format conversion. No sampler.

14 IMAGE_GET_RESINFO Return resource info for a given mip level specified in the
address vgpr. No sampler. Returns 4 integer values into VGPRs
3-0: {num_mip_levels, depth, height, width}.

12.16. MIMG Instructions 217 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

15 IMAGE_ATOMIC_SWAP   // 32bit
  tmp = MEM[ADDR];
  MEM[ADDR] = DATA;
  RETURN_DATA = tmp.

16 IMAGE_ATOMIC_CMPSWAP   // 32bit
  tmp = MEM[ADDR];
  src = DATA[0];
  cmp = DATA[1];
  MEM[ADDR] = (tmp == cmp) ? src : tmp;
  RETURN_DATA[0] = tmp.

17 IMAGE_ATOMIC_ADD   // 32bit
  tmp = MEM[ADDR];
  MEM[ADDR] += DATA;
  RETURN_DATA = tmp.

18 IMAGE_ATOMIC_SUB   // 32bit
  tmp = MEM[ADDR];
  MEM[ADDR] -= DATA;
  RETURN_DATA = tmp.

20 IMAGE_ATOMIC_SMIN   // 32bit
  tmp = MEM[ADDR];
  MEM[ADDR] = (DATA < tmp) ? DATA : tmp; // signed compare
  RETURN_DATA = tmp.

21 IMAGE_ATOMIC_UMIN   // 32bit
  tmp = MEM[ADDR];
  MEM[ADDR] = (DATA < tmp) ? DATA : tmp; // unsigned compare
  RETURN_DATA = tmp.

22 IMAGE_ATOMIC_SMAX   // 32bit
  tmp = MEM[ADDR];
  MEM[ADDR] = (DATA > tmp) ? DATA : tmp; // signed compare
  RETURN_DATA = tmp.

23 IMAGE_ATOMIC_UMAX   // 32bit
  tmp = MEM[ADDR];
  MEM[ADDR] = (DATA > tmp) ? DATA : tmp; // unsigned compare
  RETURN_DATA = tmp.

24 IMAGE_ATOMIC_AND   // 32bit
  tmp = MEM[ADDR];
  MEM[ADDR] &= DATA;
  RETURN_DATA = tmp.

25 IMAGE_ATOMIC_OR   // 32bit
  tmp = MEM[ADDR];
  MEM[ADDR] |= DATA;
  RETURN_DATA = tmp.

26 IMAGE_ATOMIC_XOR   // 32bit
  tmp = MEM[ADDR];
  MEM[ADDR] ^= DATA;
  RETURN_DATA = tmp.

12.16. MIMG Instructions 218 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

27 IMAGE_ATOMIC_INC   // 32bit
  tmp = MEM[ADDR];
  MEM[ADDR] = (tmp >= DATA) ? 0 : tmp + 1; // unsigned
compare
  RETURN_DATA = tmp.

28 IMAGE_ATOMIC_DEC   // 32bit
  tmp = MEM[ADDR];
  MEM[ADDR] = (tmp == 0 || tmp > DATA) ? DATA : tmp - 1; //
unsigned compare
  RETURN_DATA = tmp.

29 IMAGE_ATOMIC_FCMPSWAP   // 32bit
  tmp = MEM[ADDR];
  src = DATA[0];
  cmp = DATA[1];
  MEM[ADDR] = (tmp == cmp) ? src : tmp;
  RETURN_DATA[0] = tmp.

Floating-point compare swap (handles NaN/INF/denorm).

30 IMAGE_ATOMIC_FMIN   // 32bit
  tmp = MEM[ADDR];
  src = DATA[0];
  MEM[ADDR] = (src < tmp) ? src : tmp;
  RETURN_DATA[0] = tmp.

Floating-point compare (handles NaN/INF/denorm).

31 IMAGE_ATOMIC_FMAX   // 32bit
  tmp = MEM[ADDR];
  src = DATA[0];
  MEM[ADDR] = (src > tmp) ? src : tmp;
  RETURN_DATA[0] = tmp.

Floating-point compare (handles NaN/INF/denorm).

32 IMAGE_SAMPLE sample texture map.

33 IMAGE_SAMPLE_CL sample texture map, with LOD clamp specified in shader.

34 IMAGE_SAMPLE_D sample texture map, with user derivatives

35 IMAGE_SAMPLE_D_CL sample texture map, with LOD clamp specified in shader, with
user derivatives.

36 IMAGE_SAMPLE_L sample texture map, with user LOD.

37 IMAGE_SAMPLE_B sample texture map, with lod bias.

38 IMAGE_SAMPLE_B_CL sample texture map, with LOD clamp specified in shader, with
lod bias.

39 IMAGE_SAMPLE_LZ sample texture map, from level 0.

40 IMAGE_SAMPLE_C sample texture map, with PCF.

41 IMAGE_SAMPLE_C_CL SAMPLE_C, with LOD clamp specified in shader.

42 IMAGE_SAMPLE_C_D SAMPLE_C, with user derivatives.

12.16. MIMG Instructions 219 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

43 IMAGE_SAMPLE_C_D_CL SAMPLE_C, with LOD clamp specified in shader, with user

derivatives.

44 IMAGE_SAMPLE_C_L SAMPLE_C, with user LOD.

45 IMAGE_SAMPLE_C_B SAMPLE_C, with lod bias.

46 IMAGE_SAMPLE_C_B_CL SAMPLE_C, with LOD clamp specified in shader, with lod bias.

47 IMAGE_SAMPLE_C_LZ SAMPLE_C, from level 0.

48 IMAGE_SAMPLE_O sample texture map, with user offsets.

49 IMAGE_SAMPLE_CL_O SAMPLE_O with LOD clamp specified in shader.

50 IMAGE_SAMPLE_D_O SAMPLE_O, with user derivatives.

51 IMAGE_SAMPLE_D_CL_O SAMPLE_O, with LOD clamp specified in shader, with user

derivatives.

52 IMAGE_SAMPLE_L_O SAMPLE_O, with user LOD.

53 IMAGE_SAMPLE_B_O SAMPLE_O, with lod bias.

54 IMAGE_SAMPLE_B_CL_O SAMPLE_O, with LOD clamp specified in shader, with lod bias.

55 IMAGE_SAMPLE_LZ_O SAMPLE_O, from level 0.

56 IMAGE_SAMPLE_C_O SAMPLE_C with user specified offsets.

57 IMAGE_SAMPLE_C_CL_O SAMPLE_C_O, with LOD clamp specified in shader.

58 IMAGE_SAMPLE_C_D_O SAMPLE_C_O, with user derivatives.

59 IMAGE_SAMPLE_C_D_CL_O SAMPLE_C_O, with LOD clamp specified in shader, with user

derivatives.

60 IMAGE_SAMPLE_C_L_O SAMPLE_C_O, with user LOD.

61 IMAGE_SAMPLE_C_B_O SAMPLE_C_O, with lod bias.

62 IMAGE_SAMPLE_C_B_CL_O SAMPLE_C_O, with LOD clamp specified in shader, with lod bias.

63 IMAGE_SAMPLE_C_LZ_O SAMPLE_C_O, from level 0.

64 IMAGE_GATHER4 gather 4 single component elements (2x2).

65 IMAGE_GATHER4_CL gather 4 single component elements (2x2) with user LOD clamp.

68 IMAGE_GATHER4_L gather 4 single component elements (2x2) with user LOD.

69 IMAGE_GATHER4_B gather 4 single component elements (2x2) with user bias.

70 IMAGE_GATHER4_B_CL gather 4 single component elements (2x2) with user bias and
clamp.

71 IMAGE_GATHER4_LZ gather 4 single component elements (2x2) at level 0.

72 IMAGE_GATHER4_C gather 4 single component elements (2x2) with PCF.

73 IMAGE_GATHER4_C_CL gather 4 single component elements (2x2) with user LOD clamp
and PCF.

76 IMAGE_GATHER4_C_L gather 4 single component elements (2x2) with user LOD and
PCF.

12.16. MIMG Instructions 220 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

77 IMAGE_GATHER4_C_B gather 4 single component elements (2x2) with user bias and
PCF.

78 IMAGE_GATHER4_C_B_CL gather 4 single component elements (2x2) with user bias, clamp
and PCF.

79 IMAGE_GATHER4_C_LZ gather 4 single component elements (2x2) at level 0, with PCF.

80 IMAGE_GATHER4_O GATHER4, with user offsets.

81 IMAGE_GATHER4_CL_O GATHER4_CL, with user offsets.

84 IMAGE_GATHER4_L_O GATHER4_L, with user offsets.

85 IMAGE_GATHER4_B_O GATHER4_B, with user offsets.

86 IMAGE_GATHER4_B_CL_O GATHER4_B_CL, with user offsets.

87 IMAGE_GATHER4_LZ_O GATHER4_LZ, with user offsets.

88 IMAGE_GATHER4_C_O GATHER4_C, with user offsets.

89 IMAGE_GATHER4_C_CL_O GATHER4_C_CL, with user offsets.

92 IMAGE_GATHER4_C_L_O GATHER4_C_L, with user offsets.

93 IMAGE_GATHER4_C_B_O GATHER4_B, with user offsets.

94 IMAGE_GATHER4_C_B_CL_O GATHER4_B_CL, with user offsets.

95 IMAGE_GATHER4_C_LZ_O GATHER4_C_LZ, with user offsets.

96 IMAGE_GET_LOD   VDATA[0] = clampedLOD;

  VDATA[1] = rawLOD.

Return calculated LOD as two 32-bit floating point values.

97 IMAGE_GATHER4H Fetch 1 component per texel from 4x1 texels. DMASK

selects which component to read (R,G,B,A) and must have
only one bit set to 1.

128 IMAGE_MSAA_LOAD Load up to 4 samples of 1 component from an MSAA resource

with a user-specified fragment ID. No sampler.

162 IMAGE_SAMPLE_D_G16 SAMPLE_D with 16-bit floating point derivatives (gradients)

163 IMAGE_SAMPLE_D_CL_G16 SAMPLE_D_CL with 16-bit floating point derivatives (gradients)

170 IMAGE_SAMPLE_C_D_G16 SAMPLE_C_D with 16-bit floating point derivatives (gradients)

171 IMAGE_SAMPLE_C_D_CL_G16 SAMPLE_C_D_CL with 16-bit floating point derivatives

(gradients)

178 IMAGE_SAMPLE_D_O_G16 SAMPLE_D_O with 16-bit floating point derivatives (gradients)

179 IMAGE_SAMPLE_D_CL_O_G16 SAMPLE_D_CL_O with 16-bit floating point derivatives

(gradients)

186 IMAGE_SAMPLE_C_D_O_G16 SAMPLE_C_D_O with 16-bit floating point derivatives

(gradients)

187 IMAGE_SAMPLE_C_D_CL_O_G SAMPLE_C_D_CL_O with 16-bit floating point derivatives

16 (gradients)

12.16. MIMG Instructions 221 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

230 IMAGE_BVH_INTERSECT_RAY Intersection test on bound volume hierarchy nodes for

ray tracing acceleration. 32-bit node pointer. No
sampler.

DATA:

The destination VGPRs contain the results of

intersection testing. The values returned here are
different depending on the type of BVH node that was
fetched.

For box nodes the results contain the 4 pointers of the

children boxes in intersection time sorted order.

For triangle BVH nodes the results contain the

intersection time and triangle ID of the triangle
tested.

ADDR:

11 address VGPRs contain the ray data and BVH node

pointer for the intersection test. The data is laid out
as follows:

  vgpr_a[0] = node_pointer (uint32)

  vgpr_a[1] = ray_extent (float32)
  vgpr_a[2] = ray_origin.x (float32)
  vgpr_a[3] = ray_origin.y (float32)
  vgpr_a[4] = ray_origin.z (float32)
  vgpr_a[5] = ray_dir.x (float32)
  vgpr_a[6] = ray_dir.y (float32)
  vgpr_a[7] = ray_dir.z (float32)
  vgpr_a[8] = ray_inv_dir.x (float32)
  vgpr_a[9] = ray_inv_dir.y (float32)
  vgpr_a[10]= ray_inv_dir.z (float32)

For performance and power optimization,

the instruction can be encoded to use 16 bit
floats for ray_dir and ray_inv_dir by setting A16 to 1.
When the instruction is encoded with 16 bit addresses only
8 address VGPRs are used as follows:

  vgpr_a[0] = node_pointer (uint32)

  vgpr_a[1] = ray_extent (float32)
  vgpr_a[2] = ray_origin.x (float32)
  vgpr_a[3] = ray_origin.y (float32)
  vgpr_a[4] = ray_origin.z (float32)
  vgpr_a[5] = {ray_dir.x,ray_dir.y}(2x float16)
  vgpr_a[6] = {ray_dir.z,ray_inv_dir.x}(2x float16)
  vgpr_a[7] = {ray_inv_dir.y,ray_inv_dir.z}(2x float16)

RSRC:

The resource is the texture descriptor for the

operation. The instruction must e encoded with r128=1.

RESTRICTIONS:

The image_bvh_intersect_ray and

image_bvh64_intersect_ray opcode do not support all of
12.16. MIMG Instructions 222 of 283
"RDNA 2" Instruction Set Architecture

Opcode Name Description

231 IMAGE_BVH64_INTERSECT_RA Intersection test on bound volume hierarchy nodes for

Y ray tracing acceleration. 64-bit node pointer. No
sampler.

This instruction allows support for very large BVHs

(larger than 32 GBs) that may occur in workstation
workloads. See IMAGE_BVH_INTERSECT_RAY for basic
information including restrictions.

ADDR:

12 address VGPRs contain the ray data and BVH node

pointer for the intersection test. The data is laid out
as follows:

  vgpr_a[0] = node_pointer[31:0] (first part of uint64)

  vgpr_a[1] = node_pointer[63:32] (second part of uint64)
  vgpr_a[2] = ray_extent (float32)
  vgpr_a[3] = ray_origin.x (float32)
  vgpr_a[4] = ray_origin.y (float32)
  vgpr_a[5] = ray_origin.z (float32)
  vgpr_a[6] = ray_dir.x (float32)
  vgpr_a[7] = ray_dir.y (float32)
  vgpr_a[8] = ray_dir.z (float32)
  vgpr_a[9] = ray_inv_dir.x (float32)
  vgpr_a[10]= ray_inv_dir.y (float32)
  vgpr_a[11]= ray_inv_dir.z (float32)

For performance and power optimization,

the instruction can be encoded to use 16 bit
floats for ray_dir and ray_inv_dir by setting A16 to 1.
When the instruction is encoded with 16 bit addresses only
9 address VGPRs are used as follows:

  vgpr_a[0] = node_pointer[31:0] (first part of uint64)

  vgpr_a[1] = node_pointer[63:32] (second part of uint64)
  vgpr_a[2] = ray_extent (float32)
  vgpr_a[3] = ray_origin.x (float32)
  vgpr_a[4] = ray_origin.y (float32)
  vgpr_a[5] = ray_origin.z (float32)
  vgpr_a[6] = {ray_dir.x,ray_dir.y} (2x float16)
  vgpr_a[7] = {ray_dir.z,ray_inv_dir.x} (2x float16)
  vgpr_a[8] = {ray_inv_dir.y,ray_inv_dir.z} (2x float16)

12.17. EXPORT Instructions

Transfer vertex position, vertex parameter, pixel color, or pixel depth information to the output
buffer. Every pixel shader must do at least one export to a color, depth or NULL target with the
VM bit set to 1. This communicates the pixel-valid mask to the color and depth buffers. Every
pixel does only one of the above export types with the DONE bit set to 1. Vertex shaders must
do one or more position exports, and at least one parameter export. The final position export
must have the DONE bit set to 1.

12.17. EXPORT Instructions 223 of 283

"RDNA 2" Instruction Set Architecture

12.18. FLAT, Scratch and Global Instructions

The bitfield map of the FLAT format is:

  where:

  GLC = Global coherency.

  SLC = System level coherency.
  OP = Instruction Opcode.
  ADDR = Source of flat address VGPR.
  DATA = Source data.
  VDST = Destination VGPR.
  SADDR = SGPR holding address or offset
  SEG = Instruction type: Flat, Scratch, or Global
  LDS = Data is transferred between LDS and Memory, not VGPRs.
  OFFSET = Immediate address byte-offset.

12.18.1. Flat Instructions

Flat instructions look at the per-workitem address and determine for each work item if the target
memory address is in global, private or scratch memory.

Opcode Name Description

8 FLAT_LOAD_UBYTE Untyped buffer load unsigned byte (zero extend to VGPR

destination).

9 FLAT_LOAD_SBYTE Untyped buffer load signed byte (sign extend to VGPR

destination).

10 FLAT_LOAD_USHORT Untyped buffer load unsigned short (zero extend to VGPR

destination).

11 FLAT_LOAD_SSHORT Untyped buffer load signed short (sign extend to VGPR

destination).

12 FLAT_LOAD_DWORD Untyped buffer load dword.

13 FLAT_LOAD_DWORDX2 Untyped buffer load 2 dwords.

14 FLAT_LOAD_DWORDX4 Untyped buffer load 4 dwords.

15 FLAT_LOAD_DWORDX3 Untyped buffer load 3 dwords.

24 FLAT_STORE_BYTE Untyped buffer store byte. Stores S0[7:0].

12.18. FLAT, Scratch and Global Instructions 224 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

25 FLAT_STORE_BYTE_D16_HI Untyped buffer store byte. Stores S0[23:16].

26 FLAT_STORE_SHORT Untyped buffer store short. Stores S0[15:0].

27 FLAT_STORE_SHORT_D16_HI Untyped buffer store short. Stores S0[31:16].

28 FLAT_STORE_DWORD Untyped buffer store dword.

29 FLAT_STORE_DWORDX2 Untyped buffer store 2 dwords.

30 FLAT_STORE_DWORDX4 Untyped buffer store 4 dwords.

31 FLAT_STORE_DWORDX3 Untyped buffer store 3 dwords.

32 FLAT_LOAD_UBYTE_D16   D0[15:0] = {8'h0, MEM[ADDR]}.

Untyped buffer load unsigned byte.

33 FLAT_LOAD_UBYTE_D16_HI   D0[31:16] = {8'h0, MEM[ADDR]}.

Untyped buffer load unsigned byte.

34 FLAT_LOAD_SBYTE_D16   D0[15:0] = signext(MEM[ADDR]).

Untyped buffer load signed byte.

35 FLAT_LOAD_SBYTE_D16_HI   D0[31:16] = signext(MEM[ADDR]).

Untyped buffer load signed byte.

36 FLAT_LOAD_SHORT_D16   D0[15:0] = MEM[ADDR].

Untyped buffer load short.

37 FLAT_LOAD_SHORT_D16_HI   D0[31:16] = MEM[ADDR].

Untyped buffer load short.

48 FLAT_ATOMIC_SWAP   // 32bit
  tmp = MEM[ADDR];
  MEM[ADDR] = DATA;
  RETURN_DATA = tmp.

49 FLAT_ATOMIC_CMPSWAP   // 32bit
  tmp = MEM[ADDR];
  src = DATA[0];
  cmp = DATA[1];
  MEM[ADDR] = (tmp == cmp) ? src : tmp;
  RETURN_DATA[0] = tmp.

50 FLAT_ATOMIC_ADD   // 32bit
  tmp = MEM[ADDR];
  MEM[ADDR] += DATA;
  RETURN_DATA = tmp.

51 FLAT_ATOMIC_SUB   // 32bit
  tmp = MEM[ADDR];
  MEM[ADDR] -= DATA;
  RETURN_DATA = tmp.

53 FLAT_ATOMIC_SMIN   // 32bit
  tmp = MEM[ADDR];
  MEM[ADDR] = (DATA < tmp) ? DATA : tmp; // signed compare
  RETURN_DATA = tmp.

12.18. FLAT, Scratch and Global Instructions 225 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

54 FLAT_ATOMIC_UMIN   // 32bit
  tmp = MEM[ADDR];
  MEM[ADDR] = (DATA < tmp) ? DATA : tmp; // unsigned compare
  RETURN_DATA = tmp.

55 FLAT_ATOMIC_SMAX   // 32bit
  tmp = MEM[ADDR];
  MEM[ADDR] = (DATA > tmp) ? DATA : tmp; // signed compare
  RETURN_DATA = tmp.

56 FLAT_ATOMIC_UMAX   // 32bit
  tmp = MEM[ADDR];
  MEM[ADDR] = (DATA > tmp) ? DATA : tmp; // unsigned compare
  RETURN_DATA = tmp.

57 FLAT_ATOMIC_AND   // 32bit
  tmp = MEM[ADDR];
  MEM[ADDR] &= DATA;
  RETURN_DATA = tmp.

58 FLAT_ATOMIC_OR   // 32bit
  tmp = MEM[ADDR];
  MEM[ADDR] |= DATA;
  RETURN_DATA = tmp.

59 FLAT_ATOMIC_XOR   // 32bit
  tmp = MEM[ADDR];
  MEM[ADDR] ^= DATA;
  RETURN_DATA = tmp.

60 FLAT_ATOMIC_INC   // 32bit
  tmp = MEM[ADDR];
  MEM[ADDR] = (tmp >= DATA) ? 0 : tmp + 1; // unsigned
compare
  RETURN_DATA = tmp.

61 FLAT_ATOMIC_DEC   // 32bit
  tmp = MEM[ADDR];
  MEM[ADDR] = (tmp == 0 || tmp > DATA) ? DATA : tmp - 1; //
unsigned compare
  RETURN_DATA = tmp.

62 FLAT_ATOMIC_FCMPSWAP   // 32bit
  tmp = MEM[ADDR];
  src = DATA[0];
  cmp = DATA[1];
  MEM[ADDR] = (tmp == cmp) ? src : tmp;
  RETURN_DATA[0] = tmp.

Floating-point compare swap (handles NaN/INF/denorm).

63 FLAT_ATOMIC_FMIN   // 32bit
  tmp = MEM[ADDR];
  src = DATA[0];
  MEM[ADDR] = (src < tmp) ? src : tmp;
  RETURN_DATA[0] = tmp.

Floating-point compare (handles NaN/INF/denorm).

12.18. FLAT, Scratch and Global Instructions 226 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

64 FLAT_ATOMIC_FMAX   // 32bit
  tmp = MEM[ADDR];
  src = DATA[0];
  MEM[ADDR] = (src > tmp) ? src : tmp;
  RETURN_DATA[0] = tmp.

Floating-point compare (handles NaN/INF/denorm).

80 FLAT_ATOMIC_SWAP_X2   // 64bit
  tmp = MEM[ADDR];
  MEM[ADDR] = DATA[0:1];
  RETURN_DATA[0:1] = tmp.

81 FLAT_ATOMIC_CMPSWAP_X2   // 64bit
  tmp = MEM[ADDR];
  src = DATA[0:1];
  cmp = DATA[2:3];
  MEM[ADDR] = (tmp == cmp) ? src : tmp;
  RETURN_DATA[0:1] = tmp.

82 FLAT_ATOMIC_ADD_X2   // 64bit
  tmp = MEM[ADDR];
  MEM[ADDR] += DATA[0:1];
  RETURN_DATA[0:1] = tmp.

83 FLAT_ATOMIC_SUB_X2   // 64bit
  tmp = MEM[ADDR];
  MEM[ADDR] -= DATA[0:1];
  RETURN_DATA[0:1] = tmp.

85 FLAT_ATOMIC_SMIN_X2   // 64bit
  tmp = MEM[ADDR];
  MEM[ADDR] = (DATA[0:1] < tmp) ? DATA[0:1] : tmp; // signed
compare
  RETURN_DATA[0:1] = tmp.

86 FLAT_ATOMIC_UMIN_X2   // 64bit
  tmp = MEM[ADDR];
  MEM[ADDR] = (DATA[0:1] < tmp) ? DATA[0:1] : tmp; //
unsigned compare
  RETURN_DATA[0:1] = tmp.

87 FLAT_ATOMIC_SMAX_X2   // 64bit
  tmp = MEM[ADDR];
  MEM[ADDR] = (DATA[0:1] > tmp) ? DATA[0:1] : tmp; // signed
compare
  RETURN_DATA[0:1] = tmp.

88 FLAT_ATOMIC_UMAX_X2   // 64bit
  tmp = MEM[ADDR];
  MEM[ADDR] = (DATA[0:1] > tmp) ? DATA[0:1] : tmp; //
unsigned compare
  RETURN_DATA[0:1] = tmp.

89 FLAT_ATOMIC_AND_X2   // 64bit
  tmp = MEM[ADDR];
  MEM[ADDR] &= DATA[0:1];
  RETURN_DATA[0:1] = tmp.

12.18. FLAT, Scratch and Global Instructions 227 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

90 FLAT_ATOMIC_OR_X2   // 64bit
  tmp = MEM[ADDR];
  MEM[ADDR] |= DATA[0:1];
  RETURN_DATA[0:1] = tmp.

91 FLAT_ATOMIC_XOR_X2   // 64bit
  tmp = MEM[ADDR];
  MEM[ADDR] ^= DATA[0:1];
  RETURN_DATA[0:1] = tmp.

92 FLAT_ATOMIC_INC_X2   // 64bit
  tmp = MEM[ADDR];
  MEM[ADDR] = (tmp >= DATA[0:1]) ? 0 : tmp + 1; // unsigned
compare
  RETURN_DATA[0:1] = tmp.

93 FLAT_ATOMIC_DEC_X2   // 64bit
  tmp = MEM[ADDR];
  MEM[ADDR] = (tmp == 0 || tmp > DATA[0:1]) ? DATA[0:1] : tmp
- 1; // unsigned compare
  RETURN_DATA[0:1] = tmp.

94 FLAT_ATOMIC_FCMPSWAP_X2   // 64bit
  tmp = MEM[ADDR];
  src = DATA[0];
  cmp = DATA[1];
  MEM[ADDR] = (tmp == cmp) ? src : tmp;
  RETURN_DATA[0] = tmp.

Floating-point compare swap (handles NaN/INF/denorm).

95 FLAT_ATOMIC_FMIN_X2   // 64bit
  tmp = MEM[ADDR];
  src = DATA[0];
  MEM[ADDR] = (src < tmp) ? src : tmp;
  RETURN_DATA[0] = tmp.

Floating-point compare (handles NaN/INF/denorm).

96 FLAT_ATOMIC_FMAX_X2   // 64bit
  tmp = MEM[ADDR];
  src = DATA[0];
  MEM[ADDR] = (src > tmp) ? src : tmp;
  RETURN_DATA[0] = tmp.

Floating-point compare (handles NaN/INF/denorm).

12.18.2. Scratch Instructions

Scratch instructions are like Flat, but assume all workitem addresses fall in scratch (private)
space.

Opcode Name Description

8 SCRATCH_LOAD_UBYTE Untyped buffer load unsigned byte (zero extend to VGPR

destination).

12.18. FLAT, Scratch and Global Instructions 228 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

9 SCRATCH_LOAD_SBYTE Untyped buffer load signed byte (sign extend to VGPR

destination).

10 SCRATCH_LOAD_USHORT Untyped buffer load unsigned short (zero extend to VGPR

destination).

11 SCRATCH_LOAD_SSHORT Untyped buffer load signed short (sign extend to VGPR

destination).

12 SCRATCH_LOAD_DWORD Untyped buffer load dword.

13 SCRATCH_LOAD_DWORDX2 Untyped buffer load 2 dwords.

14 SCRATCH_LOAD_DWORDX4 Untyped buffer load 4 dwords.

15 SCRATCH_LOAD_DWORDX3 Untyped buffer load 3 dwords.

24 SCRATCH_STORE_BYTE Untyped buffer store byte. Stores S0[7:0].

25 SCRATCH_STORE_BYTE_D16_ Untyped buffer store byte. Stores S0[23:16].

HI

26 SCRATCH_STORE_SHORT Untyped buffer store short. Stores S0[15:0].

27 SCRATCH_STORE_SHORT_D16 Untyped buffer store short. Stores S0[31:16].

_HI

28 SCRATCH_STORE_DWORD Untyped buffer store dword.

29 SCRATCH_STORE_DWORDX2 Untyped buffer store 2 dwords.

30 SCRATCH_STORE_DWORDX4 Untyped buffer store 4 dwords.

31 SCRATCH_STORE_DWORDX3 Untyped buffer store 3 dwords.

32 SCRATCH_LOAD_UBYTE_D16   D0[15:0] = {8'h0, MEM[ADDR]}.

Untyped buffer load unsigned byte.

33 SCRATCH_LOAD_UBYTE_D16_   D0[31:16] = {8'h0, MEM[ADDR]}.

HI
Untyped buffer load unsigned byte.

34 SCRATCH_LOAD_SBYTE_D16   D0[15:0] = signext(MEM[ADDR]).

Untyped buffer load signed byte.

35 SCRATCH_LOAD_SBYTE_D16_   D0[31:16] = signext(MEM[ADDR]).

HI
Untyped buffer load signed byte.

36 SCRATCH_LOAD_SHORT_D16   D0[15:0] = MEM[ADDR].

Untyped buffer load short.

37 SCRATCH_LOAD_SHORT_D16_   D0[31:16] = MEM[ADDR].

HI
Untyped buffer load short.

12.18.3. Global Instructions

Global instructions are like Flat, but assume all workitem addresses fall in global memory space.

12.18. FLAT, Scratch and Global Instructions 229 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

8 GLOBAL_LOAD_UBYTE Untyped buffer load unsigned byte (zero extend to VGPR

destination).

9 GLOBAL_LOAD_SBYTE Untyped buffer load signed byte (sign extend to VGPR

destination).

10 GLOBAL_LOAD_USHORT Untyped buffer load unsigned short (zero extend to VGPR

destination).

11 GLOBAL_LOAD_SSHORT Untyped buffer load signed short (sign extend to VGPR

destination).

12 GLOBAL_LOAD_DWORD Untyped buffer load dword.

13 GLOBAL_LOAD_DWORDX2 Untyped buffer load 2 dwords.

14 GLOBAL_LOAD_DWORDX4 Untyped buffer load 4 dwords.

15 GLOBAL_LOAD_DWORDX3 Untyped buffer load 3 dwords.

22 GLOBAL_LOAD_DWORD_ADDTI Untyped buffer load dword. No VGPR address is supplied

D in this instruction. TID is added to the address as shown
below:

memory_Addr = sgpr_addr(64) + inst_offset(12) + tid*4

23 GLOBAL_STORE_DWORD_ADD Untyped buffer store dword. No VGPR address is supplied

TID in this instruction. TID is added to the address as shown
below:

memory_Addr = sgpr_addr(64) + inst_offset(12) + tid*4

24 GLOBAL_STORE_BYTE Untyped buffer store byte. Stores S0[7:0].

25 GLOBAL_STORE_BYTE_D16_HI Untyped buffer store byte. Stores S0[23:16].

26 GLOBAL_STORE_SHORT Untyped buffer store short. Stores S0[15:0].

27 GLOBAL_STORE_SHORT_D16_ Untyped buffer store short. Stores S0[31:16].

HI

28 GLOBAL_STORE_DWORD Untyped buffer store dword.

29 GLOBAL_STORE_DWORDX2 Untyped buffer store 2 dwords.

30 GLOBAL_STORE_DWORDX4 Untyped buffer store 4 dwords.

31 GLOBAL_STORE_DWORDX3 Untyped buffer store 3 dwords.

32 GLOBAL_LOAD_UBYTE_D16   D0[15:0] = {8'h0, MEM[ADDR]}.

Untyped buffer load unsigned byte.

33 GLOBAL_LOAD_UBYTE_D16_HI   D0[31:16] = {8'h0, MEM[ADDR]}.

Untyped buffer load unsigned byte.

34 GLOBAL_LOAD_SBYTE_D16   D0[15:0] = signext(MEM[ADDR]).

Untyped buffer load signed byte.

35 GLOBAL_LOAD_SBYTE_D16_HI   D0[31:16] = signext(MEM[ADDR]).

Untyped buffer load signed byte.

12.18. FLAT, Scratch and Global Instructions 230 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

36 GLOBAL_LOAD_SHORT_D16   D0[15:0] = MEM[ADDR].

Untyped buffer load short.

37 GLOBAL_LOAD_SHORT_D16_HI   D0[31:16] = MEM[ADDR].

Untyped buffer load short.

48 GLOBAL_ATOMIC_SWAP   // 32bit
  tmp = MEM[ADDR];
  MEM[ADDR] = DATA;
  RETURN_DATA = tmp.

49 GLOBAL_ATOMIC_CMPSWAP   // 32bit
  tmp = MEM[ADDR];
  src = DATA[0];
  cmp = DATA[1];
  MEM[ADDR] = (tmp == cmp) ? src : tmp;
  RETURN_DATA[0] = tmp.

50 GLOBAL_ATOMIC_ADD   // 32bit
  tmp = MEM[ADDR];
  MEM[ADDR] += DATA;
  RETURN_DATA = tmp.

51 GLOBAL_ATOMIC_SUB   // 32bit
  tmp = MEM[ADDR];
  MEM[ADDR] -= DATA;
  RETURN_DATA = tmp.

52 GLOBAL_ATOMIC_CSUB   // 32bit
  old_value = MEM[ADDR];
  if old_value < DATA then
  new_value = 0;
  else
  new_value = old_value - DATA;
  endif;
  MEM[addr] = new_value;
  RETURN_DATA = old_value.

53 GLOBAL_ATOMIC_SMIN   // 32bit
  tmp = MEM[ADDR];
  MEM[ADDR] = (DATA < tmp) ? DATA : tmp; // signed compare
  RETURN_DATA = tmp.

54 GLOBAL_ATOMIC_UMIN   // 32bit
  tmp = MEM[ADDR];
  MEM[ADDR] = (DATA < tmp) ? DATA : tmp; // unsigned compare
  RETURN_DATA = tmp.

55 GLOBAL_ATOMIC_SMAX   // 32bit
  tmp = MEM[ADDR];
  MEM[ADDR] = (DATA > tmp) ? DATA : tmp; // signed compare
  RETURN_DATA = tmp.

56 GLOBAL_ATOMIC_UMAX   // 32bit
  tmp = MEM[ADDR];
  MEM[ADDR] = (DATA > tmp) ? DATA : tmp; // unsigned compare
  RETURN_DATA = tmp.

12.18. FLAT, Scratch and Global Instructions 231 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

57 GLOBAL_ATOMIC_AND   // 32bit
  tmp = MEM[ADDR];
  MEM[ADDR] &= DATA;
  RETURN_DATA = tmp.

58 GLOBAL_ATOMIC_OR   // 32bit
  tmp = MEM[ADDR];
  MEM[ADDR] |= DATA;
  RETURN_DATA = tmp.

59 GLOBAL_ATOMIC_XOR   // 32bit
  tmp = MEM[ADDR];
  MEM[ADDR] ^= DATA;
  RETURN_DATA = tmp.

60 GLOBAL_ATOMIC_INC   // 32bit
  tmp = MEM[ADDR];
  MEM[ADDR] = (tmp >= DATA) ? 0 : tmp + 1; // unsigned
compare
  RETURN_DATA = tmp.

61 GLOBAL_ATOMIC_DEC   // 32bit
  tmp = MEM[ADDR];
  MEM[ADDR] = (tmp == 0 || tmp > DATA) ? DATA : tmp - 1; //
unsigned compare
  RETURN_DATA = tmp.

62 GLOBAL_ATOMIC_FCMPSWAP   // 32bit
  tmp = MEM[ADDR];
  src = DATA[0];
  cmp = DATA[1];
  MEM[ADDR] = (tmp == cmp) ? src : tmp;
  RETURN_DATA[0] = tmp.

Floating-point compare swap (handles NaN/INF/denorm).

63 GLOBAL_ATOMIC_FMIN   // 32bit
  tmp = MEM[ADDR];
  src = DATA[0];
  MEM[ADDR] = (src < tmp) ? src : tmp;
  RETURN_DATA[0] = tmp.

Floating-point compare (handles NaN/INF/denorm).

64 GLOBAL_ATOMIC_FMAX   // 32bit
  tmp = MEM[ADDR];
  src = DATA[0];
  MEM[ADDR] = (src > tmp) ? src : tmp;
  RETURN_DATA[0] = tmp.

Floating-point compare (handles NaN/INF/denorm).

80 GLOBAL_ATOMIC_SWAP_X2   // 64bit
  tmp = MEM[ADDR];
  MEM[ADDR] = DATA[0:1];
  RETURN_DATA[0:1] = tmp.

12.18. FLAT, Scratch and Global Instructions 232 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

81 GLOBAL_ATOMIC_CMPSWAP_   // 64bit
X2   tmp = MEM[ADDR];
  src = DATA[0:1];
  cmp = DATA[2:3];
  MEM[ADDR] = (tmp == cmp) ? src : tmp;
  RETURN_DATA[0:1] = tmp.

82 GLOBAL_ATOMIC_ADD_X2   // 64bit
  tmp = MEM[ADDR];
  MEM[ADDR] += DATA[0:1];
  RETURN_DATA[0:1] = tmp.

83 GLOBAL_ATOMIC_SUB_X2   // 64bit
  tmp = MEM[ADDR];
  MEM[ADDR] -= DATA[0:1];
  RETURN_DATA[0:1] = tmp.

85 GLOBAL_ATOMIC_SMIN_X2   // 64bit
  tmp = MEM[ADDR];
  MEM[ADDR] = (DATA[0:1] < tmp) ? DATA[0:1] : tmp; // signed
compare
  RETURN_DATA[0:1] = tmp.

86 GLOBAL_ATOMIC_UMIN_X2   // 64bit
  tmp = MEM[ADDR];
  MEM[ADDR] = (DATA[0:1] < tmp) ? DATA[0:1] : tmp; //
unsigned compare
  RETURN_DATA[0:1] = tmp.

87 GLOBAL_ATOMIC_SMAX_X2   // 64bit
  tmp = MEM[ADDR];
  MEM[ADDR] = (DATA[0:1] > tmp) ? DATA[0:1] : tmp; // signed
compare
  RETURN_DATA[0:1] = tmp.

88 GLOBAL_ATOMIC_UMAX_X2   // 64bit
  tmp = MEM[ADDR];
  MEM[ADDR] = (DATA[0:1] > tmp) ? DATA[0:1] : tmp; //
unsigned compare
  RETURN_DATA[0:1] = tmp.

89 GLOBAL_ATOMIC_AND_X2   // 64bit
  tmp = MEM[ADDR];
  MEM[ADDR] &= DATA[0:1];
  RETURN_DATA[0:1] = tmp.

90 GLOBAL_ATOMIC_OR_X2   // 64bit
  tmp = MEM[ADDR];
  MEM[ADDR] |= DATA[0:1];
  RETURN_DATA[0:1] = tmp.

91 GLOBAL_ATOMIC_XOR_X2   // 64bit
  tmp = MEM[ADDR];
  MEM[ADDR] ^= DATA[0:1];
  RETURN_DATA[0:1] = tmp.

12.18. FLAT, Scratch and Global Instructions 233 of 283

"RDNA 2" Instruction Set Architecture

Opcode Name Description

92 GLOBAL_ATOMIC_INC_X2   // 64bit
  tmp = MEM[ADDR];
  MEM[ADDR] = (tmp >= DATA[0:1]) ? 0 : tmp + 1; // unsigned
compare
  RETURN_DATA[0:1] = tmp.

93 GLOBAL_ATOMIC_DEC_X2   // 64bit
  tmp = MEM[ADDR];
  MEM[ADDR] = (tmp == 0 || tmp > DATA[0:1]) ? DATA[0:1] : tmp
- 1; // unsigned compare
  RETURN_DATA[0:1] = tmp.

94 GLOBAL_ATOMIC_FCMPSWAP   // 64bit
_X2   tmp = MEM[ADDR];
  src = DATA[0];
  cmp = DATA[1];
  MEM[ADDR] = (tmp == cmp) ? src : tmp;
  RETURN_DATA[0] = tmp.

Floating-point compare swap (handles NaN/INF/denorm).

95 GLOBAL_ATOMIC_FMIN_X2   // 64bit
  tmp = MEM[ADDR];
  src = DATA[0];
  MEM[ADDR] = (src < tmp) ? src : tmp;
  RETURN_DATA[0] = tmp.

Floating-point compare (handles NaN/INF/denorm).

96 GLOBAL_ATOMIC_FMAX_X2   // 64bit
  tmp = MEM[ADDR];
  src = DATA[0];
  MEM[ADDR] = (src > tmp) ? src : tmp;
  RETURN_DATA[0] = tmp.

Floating-point compare (handles NaN/INF/denorm).

12.19. Instruction Limitations

12.19.1. DPP
The following instructions cannot use DPP:

• V_FMAMK_F32
• V_FMAAK_F32
• V_FMAMK_F16
• V_FMAAK_F16
• V_READFIRSTLANE_B32
• V_CVT_I32_F64
• V_CVT_F64_I32

12.19. Instruction Limitations 234 of 283

"RDNA 2" Instruction Set Architecture

• V_CVT_F32_F64
• V_CVT_F64_F32
• V_CVT_U32_F64
• V_CVT_F64_U32
• V_TRUNC_F64
• V_CEIL_F64
• V_RNDNE_F64
• V_FLOOR_F64
• V_RCP_F64
• V_RSQ_F64
• V_SQRT_F64
• V_FREXP_EXP_I32_F64
• V_FREXP_MANT_F64
• V_FRACT_F64
• V_CLREXCP
• V_SWAP_B32
• V_CMP_CLASS_F64
• V_CMPX_CLASS_F64
• V_CMP_*_F64
• V_CMPX_*_F64
• V_CMP_*_I64
• V_CMP_*_U64
• V_CMPX_*_I64
• V_CMPX_*_U64

12.19.2. SDWA
The following instructions cannot use SDWA:

• V_FMAC_F32
• V_FMAMK_F32
• V_FMAAK_F32
• V_FMAC_F16
• V_FMAMK_F16
• V_FMAAK_F16
• V_READFIRSTLANE_B32
• V_CLREXCP
• V_SWAP_B32

12.19. Instruction Limitations 235 of 283

"RDNA 2" Instruction Set Architecture

Chapter 13. Microcode Formats

This section specifies the microcode formats. The definitions can be used to simplify compilation
by providing standard templates and enumeration names for the various instruction formats.

Endian Order - The RDNA architecture addresses memory and registers using littleendian byte-
ordering and bit-ordering. Multi-byte values are stored with their least-significant (low-order) byte
(LSB) at the lowest byte address, and they are illustrated with their LSB at the right side. Byte
values are stored with their least-significant (low-order) bit (lsb) at the lowest bit address, and
they are illustrated with their lsb at the right side.

The table below summarizes the microcode formats and their widths. The sections that follow
provide details

Table 62. Summary of Microcode Formats

Microcode Formats Reference Width (bits)

Scalar ALU and Control Formats

SOP2 SOP2 32

SOP1 SOP1

SOPK SOPK

SOPP SOPP

SOPC SOPC

Scalar Memory Format

SMEM SMEM 64

Vector ALU Format

VOP1 VOP1 32

VOP2 VOP2 32

VOPC VOPC 32

VOP3A VOP3A 64

VOP3B VOP3B 64

VOP3P VOP3P 64

DPP DPP 32

SDWA VOP2 32

Vector Parameter Interpolation Format

VINTRP VINTRP 32

LDS/GDS Format

DS DS 64

Vector Memory Buffer Formats

236 of 283
"RDNA 2" Instruction Set Architecture

Microcode Formats Reference Width (bits)

MTBUF MTBUF 64

MUBUF MUBUF 64

Vector Memory Image Format

MIMG MIMG 64

Export Format

EXP EXP 64

Flat Formats

FLAT FLAT 64

GLOBAL GLOBAL 64

SCRATCH SCRATCH 64

The field-definition tables that accompany the descriptions in the sections below use the
following notation.

• int(2) - A two-bit field that specifies an unsigned integer value.

• enum(7) - A seven-bit field that specifies an enumerated set of values (in this case, a set of
up to 27 values). The number of valid values can be less than the maximum.

The default value of all fields is zero. Any bitfield not identified is assumed to be reserved.

Instruction Suffixes

Most instructions include a suffix which indicates the data type the instruction handles. This
suffix may also include a number which indicate the size of the data.

For example: "F32" indicates "32-bit floating point data", or "B16" is "16-bit binary data".

• B = binary
• F = floating point
• U = unsigned integer
• S = signed integer

When more than one data-type specifier occurs in an instruction, the last one is the result type
and size, and the earlier one(s) is/are input data type and size.

13.1. Scalar ALU and Control Formats

13.1. Scalar ALU and Control Formats 237 of 283

"RDNA 2" Instruction Set Architecture

13.1.1. SOP2
Scalar format with Two inputs, one output

Format SOP2

Description This is a scalar instruction with two inputs and one output. Can be followed
by a 32-bit literal constant.

Table 63. SOP2 Fields

Field Name Bits Format or Description

SSRC0 [7:0] Source 0. First operand for the instruction.

0 - 105 SGPR0 to SGPR105: Scalar general-purpose registers.
106 VCC_LO: vcc[31:0].
107 VCC_HI: vcc[63:32].
108-123 TTMP0 - TTMP15: Trap handler temporary register.
124 M0. Memory register 0.
125 NULL
126 EXEC_LO: exec[31:0].
127 EXEC_HI: exec[63:32].
128 0.
129-192 Signed integer 1 to 64.
193-208 Signed integer -1 to -16.
209-234 Reserved.
235 SHARED_BASE (Memory Aperture definition).
236 SHARED_LIMIT (Memory Aperture definition).
237 PRIVATE_BASE (Memory Aperture definition).
238 PRIVATE_LIMIT (Memory Aperture definition).
239 POPS_EXITING_WAVE_ID .
240 0.5.
241 -0.5.
242 1.0.
243 -1.0.
244 2.0.
245 -2.0.
246 4.0.
247 -4.0.
248 1/(2*PI).
249 - 250 Reserved.
251 VCCZ.
252 EXECZ.
253 SCC.
254 Reserved.
255 Literal constant.

SSRC1 [15:8] Second scalar source operand.

Same codes as SSRC0, above.

13.1. Scalar ALU and Control Formats 238 of 283

"RDNA 2" Instruction Set Architecture

Field Name Bits Format or Description

SDST [22:16] Scalar destination.

Same codes as SSRC0, above except only codes 0-127 are valid.

OP [29:23] See Opcode table below.

ENCODING [31:30] 10

Table 64. SOP2 Opcodes

Opcode # Name Opcode # Name

0 S_ADD_U32 28 S_XNOR_B32

1 S_SUB_U32 29 S_XNOR_B64

2 S_ADD_I32 30 S_LSHL_B32

3 S_SUB_I32 31 S_LSHL_B64

4 S_ADDC_U32 32 S_LSHR_B32

5 S_SUBB_U32 33 S_LSHR_B64

6 S_MIN_I32 34 S_ASHR_I32

7 S_MIN_U32 35 S_ASHR_I64

8 S_MAX_I32 36 S_BFM_B32

9 S_MAX_U32 37 S_BFM_B64

10 S_CSELECT_B32 38 S_MUL_I32

11 S_CSELECT_B64 39 S_BFE_U32

14 S_AND_B32 40 S_BFE_I32

15 S_AND_B64 41 S_BFE_U64

16 S_OR_B32 42 S_BFE_I64

17 S_OR_B64 44 S_ABSDIFF_I32

18 S_XOR_B32 46 S_LSHL1_ADD_U32

19 S_XOR_B64 47 S_LSHL2_ADD_U32

20 S_ANDN2_B32 48 S_LSHL3_ADD_U32

21 S_ANDN2_B64 49 S_LSHL4_ADD_U32

22 S_ORN2_B32 50 S_PACK_LL_B32_B16

23 S_ORN2_B64 51 S_PACK_LH_B32_B16

24 S_NAND_B32 52 S_PACK_HH_B32_B16

25 S_NAND_B64 53 S_MUL_HI_U32

26 S_NOR_B32 54 S_MUL_HI_I32

27 S_NOR_B64

13.1. Scalar ALU and Control Formats 239 of 283

"RDNA 2" Instruction Set Architecture

13.1.2. SOPK

Format SOPK

Description This is a scalar instruction with one 16-bit signed immediate (SIMM16)
input and a single destination. Instructions which take 2 inputs use the
destination as the second input.

Table 65. SOPK Fields

Field Name Bits Format or Description

SIMM16 [15:0] Signed immediate 16-bit value.

SDST [22:16] Scalar destination, and can provide second source operand.
0 - 105 SGPR0 to SGPR105: Scalar general-purpose registers.
106 VCC_LO: vcc[31:0].
107 VCC_HI: vcc[63:32].
108-123 TTMP0 - TTMP15: Trap handler temporary register.
124 M0. Memory register 0.
125 NULL
126 EXEC_LO: exec[31:0].
127 EXEC_HI: exec[63:32].

OP [27:23] See Opcode table below.

ENCODING [31:28] 1011

Table 66. SOPK Opcodes

Opcode # Name Opcode # Name

0 S_MOVK_I32 14 S_CMPK_LE_U32

1 S_VERSION 15 S_ADDK_I32

2 S_CMOVK_I32 16 S_MULK_I32

3 S_CMPK_EQ_I32 18 S_GETREG_B32

4 S_CMPK_LG_I32 19 S_SETREG_B32

5 S_CMPK_GT_I32 21 S_SETREG_IMM32_B32

6 S_CMPK_GE_I32 22 S_CALL_B64

7 S_CMPK_LT_I32 23 S_WAITCNT_VSCNT

8 S_CMPK_LE_I32 24 S_WAITCNT_VMCNT

9 S_CMPK_EQ_U32 25 S_WAITCNT_EXPCNT

10 S_CMPK_LG_U32 26 S_WAITCNT_LGKMCNT

11 S_CMPK_GT_U32 27 S_SUBVECTOR_LOOP_BEGIN

13.1. Scalar ALU and Control Formats 240 of 283

"RDNA 2" Instruction Set Architecture

Opcode # Name Opcode # Name

12 S_CMPK_GE_U32 28 S_SUBVECTOR_LOOP_END

13 S_CMPK_LT_U32

13.1.3. SOP1

Format SOP1

Description This is a scalar instruction with two inputs and one output. Can be followed
by a 32-bit literal constant.

Table 67. SOP1 Fields

13.1. Scalar ALU and Control Formats 241 of 283

"RDNA 2" Instruction Set Architecture

Field Name Bits Format or Description

SSRC0 [7:0] Source 0. First operand for the instruction.

0 - 105 SGPR0 to SGPR105: Scalar general-purpose registers.
106 VCC_LO: vcc[31:0].
107 VCC_HI: vcc[63:32].
108-123 TTMP0 - TTMP15: Trap handler temporary register.
124 M0. Memory register 0.
125 NULL
126 EXEC_LO: exec[31:0].
127 EXEC_HI: exec[63:32].
128 0.
129-192 Signed integer 1 to 64.
193-208 Signed integer -1 to -16.
209-234 Reserved.
235 SHARED_BASE (Memory Aperture definition).
236 SHARED_LIMIT (Memory Aperture definition).
237 PRIVATE_BASE (Memory Aperture definition).
238 PRIVATE_LIMIT (Memory Aperture definition).
239 POPS_EXITING_WAVE_ID .
240 0.5.
241 -0.5.
242 1.0.
243 -1.0.
244 2.0.
245 -2.0.
246 4.0.
247 -4.0.
248 1/(2*PI).
249 - 250 Reserved.
251 VCCZ.
252 EXECZ.
253 SCC.
254 Reserved.
255 Literal constant.

OP [15:8] See Opcode table below.

SDST [22:16] Scalar destination.

Same codes as SSRC0, above except only codes 0-127 are valid.

ENCODING [31:23] 10_1111101

Table 68. SOP1 Opcodes

Opcode # Name Opcode # Name

3 S_MOV_B32 37 S_OR_SAVEEXEC_B64

4 S_MOV_B64 38 S_XOR_SAVEEXEC_B64

5 S_CMOV_B32 39 S_ANDN2_SAVEEXEC_B64

6 S_CMOV_B64 40 S_ORN2_SAVEEXEC_B64

7 S_NOT_B32 41 S_NAND_SAVEEXEC_B64

8 S_NOT_B64 42 S_NOR_SAVEEXEC_B64

13.1. Scalar ALU and Control Formats 242 of 283

"RDNA 2" Instruction Set Architecture

Opcode # Name Opcode # Name

9 S_WQM_B32 43 S_XNOR_SAVEEXEC_B64

10 S_WQM_B64 44 S_QUADMASK_B32

11 S_BREV_B32 45 S_QUADMASK_B64

12 S_BREV_B64 46 S_MOVRELS_B32

13 S_BCNT0_I32_B32 47 S_MOVRELS_B64

14 S_BCNT0_I32_B64 48 S_MOVRELD_B32

15 S_BCNT1_I32_B32 49 S_MOVRELD_B64

16 S_BCNT1_I32_B64 52 S_ABS_I32

17 S_FF0_I32_B32 55 S_ANDN1_SAVEEXEC_B64

18 S_FF0_I32_B64 56 S_ORN1_SAVEEXEC_B64

19 S_FF1_I32_B32 57 S_ANDN1_WREXEC_B64

20 S_FF1_I32_B64 58 S_ANDN2_WREXEC_B64

21 S_FLBIT_I32_B32 59 S_BITREPLICATE_B64_B32

22 S_FLBIT_I32_B64 60 S_AND_SAVEEXEC_B32

23 S_FLBIT_I32 61 S_OR_SAVEEXEC_B32

24 S_FLBIT_I32_I64 62 S_XOR_SAVEEXEC_B32

25 S_SEXT_I32_I8 63 S_ANDN2_SAVEEXEC_B32

26 S_SEXT_I32_I16 64 S_ORN2_SAVEEXEC_B32

27 S_BITSET0_B32 65 S_NAND_SAVEEXEC_B32

28 S_BITSET0_B64 66 S_NOR_SAVEEXEC_B32

29 S_BITSET1_B32 67 S_XNOR_SAVEEXEC_B32

30 S_BITSET1_B64 68 S_ANDN1_SAVEEXEC_B32

31 S_GETPC_B64 69 S_ORN1_SAVEEXEC_B32

32 S_SETPC_B64 70 S_ANDN1_WREXEC_B32

33 S_SWAPPC_B64 71 S_ANDN2_WREXEC_B32

34 S_RFE_B64 73 S_MOVRELSD_2_B32

36 S_AND_SAVEEXEC_B64

13.1.4. SOPC

Format SOPC

13.1. Scalar ALU and Control Formats 243 of 283

"RDNA 2" Instruction Set Architecture

Description This is a scalar instruction with two inputs which are compared and
produce SCC as a result. Can be followed by a 32-bit literal constant.

Table 69. SOPC Fields

Field Name Bits Format or Description

SSRC0 [7:0] Source 0. First operand for the instruction.

0 - 105 SGPR0 to SGPR105: Scalar general-purpose registers.
106 VCC_LO: vcc[31:0].
107 VCC_HI: vcc[63:32].
108-123 TTMP0 - TTMP15: Trap handler temporary register.
124 M0. Memory register 0.
125 NULL
126 EXEC_LO: exec[31:0].
127 EXEC_HI: exec[63:32].
128 0.
129-192 Signed integer 1 to 64.
193-208 Signed integer -1 to -16.
209-234 Reserved.
235 SHARED_BASE (Memory Aperture definition).
236 SHARED_LIMIT (Memory Aperture definition).
237 PRIVATE_BASE (Memory Aperture definition).
238 PRIVATE_LIMIT (Memory Aperture definition).
239 POPS_EXITING_WAVE_ID .
240 0.5.
241 -0.5.
242 1.0.
243 -1.0.
244 2.0.
245 -2.0.
246 4.0.
247 -4.0.
248 1/(2*PI).
249 - 250 Reserved.
251 VCCZ.
252 EXECZ.
253 SCC.
254 Reserved.
255 Literal constant.

SSRC1 [15:8] Second scalar source operand.

Same codes as SSRC0, above.

OP [22:16] See Opcode table below.

ENCODING [31:23] 10_1111110

Table 70. SOPC Opcodes

Opcode # Name Opcode # Name

0 S_CMP_EQ_I32 9 S_CMP_GE_U32

1 S_CMP_LG_I32 10 S_CMP_LT_U32

13.1. Scalar ALU and Control Formats 244 of 283

"RDNA 2" Instruction Set Architecture

Opcode # Name Opcode # Name

2 S_CMP_GT_I32 11 S_CMP_LE_U32

3 S_CMP_GE_I32 12 S_BITCMP0_B32

4 S_CMP_LT_I32 13 S_BITCMP1_B32

5 S_CMP_LE_I32 14 S_BITCMP0_B64

6 S_CMP_EQ_U32 15 S_BITCMP1_B64

7 S_CMP_LG_U32 18 S_CMP_EQ_U64

8 S_CMP_GT_U32 19 S_CMP_LG_U64

13.1.5. SOPP

Format SOPP

Description This is a scalar instruction with one 16-bit signed immediate (SIMM16)
input.

Table 71. SOPP Fields

Field Name Bits Format or Description

SIMM16 [15:0] Signed immediate 16-bit value.

OP [22:16] See Opcode table below.

ENCODING [31:23] 10_1111111

Table 72. SOPP Opcodes

Opcode # Name Opcode # Name

0 S_NOP 18 S_TRAP

1 S_ENDPGM 19 S_ICACHE_INV

2 S_BRANCH 20 S_INCPERFLEVEL

3 S_WAKEUP 21 S_DECPERFLEVEL

4 S_CBRANCH_SCC0 22 S_TTRACEDATA

5 S_CBRANCH_SCC1 23 S_CBRANCH_CDBGSYS

6 S_CBRANCH_VCCZ 24 S_CBRANCH_CDBGUSER

7 S_CBRANCH_VCCNZ 25 S_CBRANCH_CDBGSYS_OR_USER

8 S_CBRANCH_EXECZ 26 S_CBRANCH_CDBGSYS_AND_USER

9 S_CBRANCH_EXECNZ 27 S_ENDPGM_SAVED

13.1. Scalar ALU and Control Formats 245 of 283

"RDNA 2" Instruction Set Architecture

Opcode # Name Opcode # Name

10 S_BARRIER 30 S_ENDPGM_ORDERED_PS_DONE

11 S_SETKILL 31 S_CODE_END

12 S_WAITCNT 32 S_INST_PREFETCH

13 S_SETHALT 33 S_CLAUSE

14 S_SLEEP 35 S_WAITCNT_DEPCTR

15 S_SETPRIO 36 S_ROUND_MODE

16 S_SENDMSG 37 S_DENORM_MODE

17 S_SENDMSGHALT 40 S_TTRACEDATA_IMM

13.2. Scalar Memory Format

13.2.1. SMEM

Format SMEM

Description Scalar Memory data load/store

Table 73. SMEM Fields

Field Name Bits Format or Description

SBASE [5:0] SGPR-pair which provides base address or SGPR-quad which provides V#.
(LSB of SGPR address is omitted).

SDATA [12:6] SGPR which provides write data or accepts return data.

DLC [14] Device level coherent.

GLC [16] Globally memory Coherent. Force bypass of L1 cache, or for atomics, cause
pre-op value to be returned.

OP [25:18] See Opcode table below.

ENCODING [31:26] 111101

OFFSET [52:32] An immediate signed byte offset. Signed offsets only work with
S_LOAD/STORE.

SOFFSET [63:57] SGPR which supplies an unsigned byte offset. Disabled if set to NULL.

Table 74. SMEM Opcodes

13.2. Scalar Memory Format 246 of 283

"RDNA 2" Instruction Set Architecture

Opcode # Name Opcode # Name

0 S_LOAD_DWORD 11 S_BUFFER_LOAD_DWORDX8

1 S_LOAD_DWORDX2 12 S_BUFFER_LOAD_DWORDX16

2 S_LOAD_DWORDX4 31 S_GL1_INV

3 S_LOAD_DWORDX8 32 S_DCACHE_INV

4 S_LOAD_DWORDX16 36 S_MEMTIME

8 S_BUFFER_LOAD_DWORD 37 S_MEMREALTIME

9 S_BUFFER_LOAD_DWORDX2 38 S_ATC_PROBE

10 S_BUFFER_LOAD_DWORDX4 39 S_ATC_PROBE_BUFFER

13.3. Vector ALU Formats

13.3.1. VOP2

Format VOP2

Description Vector ALU format with two operands

Table 75. VOP2 Fields

13.3. Vector ALU Formats 247 of 283

"RDNA 2" Instruction Set Architecture

Field Name Bits Format or Description

SRC0 [8:0] Source 0. First operand for the instruction.

0 - 105 SGPR0 to SGPR105: Scalar general-purpose registers.
106 VCC_LO: vcc[31:0].
107 VCC_HI: vcc[63:32].
108-123 TTMP0 - TTMP15: Trap handler temporary register.
124 M0. Memory register 0.
125 NULL
126 EXEC_LO: exec[31:0].
127 EXEC_HI: exec[63:32].
128 0.
129-192 Signed integer 1 to 64.
193-208 Signed integer -1 to -16.
209-232 Reserved.
233 DPP8
234 DPP8FI
235 SHARED_BASE (Memory Aperture definition).
236 SHARED_LIMIT (Memory Aperture definition).
237 PRIVATE_BASE (Memory Aperture definition).
238 PRIVATE_LIMIT (Memory Aperture definition).
239 POPS_EXITING_WAVE_ID .
240 0.5.
241 -0.5.
242 1.0.
243 -1.0.
244 2.0.
245 -2.0.
246 4.0.
247 -4.0.
248 1/(2*PI).
249 SDWA
250 DPP16
251 VCCZ.
252 EXECZ.
253 SCC.
254 Reserved.
255 Literal constant.
256 - 511 VGPR 0 - 255

VSRC1 [16:9] VGPR which provides the second operand.

VDST [24:17] Destination VGPR.

OP [30:25] See Opcode table below.

ENCODING [31] 0

Table 76. VOP2 Opcodes

Opcode # Name Opcode # Name

1 V_CNDMASK_B32 29 V_XOR_B32

2 V_DOT2C_F32_F16 30 V_XNOR_B32

3 V_ADD_F32 37 V_ADD_NC_U32

4 V_SUB_F32 38 V_SUB_NC_U32

13.3. Vector ALU Formats 248 of 283

"RDNA 2" Instruction Set Architecture

Opcode # Name Opcode # Name

5 V_SUBREV_F32 39 V_SUBREV_NC_U32

6 V_FMAC_LEGACY_F32 40 V_ADD_CO_CI_U32

7 V_MUL_LEGACY_F32 41 V_SUB_CO_CI_U32

8 V_MUL_F32 42 V_SUBREV_CO_CI_U32

9 V_MUL_I32_I24 43 V_FMAC_F32

10 V_MUL_HI_I32_I24 44 V_FMAMK_F32

11 V_MUL_U32_U24 45 V_FMAAK_F32

12 V_MUL_HI_U32_U24 47 V_CVT_PKRTZ_F16_F32

13 V_DOT4C_I32_I8 50 V_ADD_F16

15 V_MIN_F32 51 V_SUB_F16

16 V_MAX_F32 52 V_SUBREV_F16

17 V_MIN_I32 53 V_MUL_F16

18 V_MAX_I32 54 V_FMAC_F16

19 V_MIN_U32 55 V_FMAMK_F16

20 V_MAX_U32 56 V_FMAAK_F16

22 V_LSHRREV_B32 57 V_MAX_F16

24 V_ASHRREV_I32 58 V_MIN_F16

26 V_LSHLREV_B32 59 V_LDEXP_F16

27 V_AND_B32 60 V_PK_FMAC_F16

28 V_OR_B32

13.3.2. VOP1

Format VOP1

Description Vector ALU format with one operand

Table 77. VOP1 Fields

13.3. Vector ALU Formats 249 of 283

"RDNA 2" Instruction Set Architecture

Field Name Bits Format or Description

SRC0 [8:0] Source 0. First operand for the instruction.

0 - 105 SGPR0 to SGPR105: Scalar general-purpose registers.
106 VCC_LO: vcc[31:0].
107 VCC_HI: vcc[63:32].
108-123 TTMP0 - TTMP15: Trap handler temporary register.
124 M0. Memory register 0.
125 NULL
126 EXEC_LO: exec[31:0].
127 EXEC_HI: exec[63:32].
128 0.
129-192 Signed integer 1 to 64.
193-208 Signed integer -1 to -16.
209-232 Reserved.
233 DPP8
234 DPP8FI
235 SHARED_BASE (Memory Aperture definition).
236 SHARED_LIMIT (Memory Aperture definition).
237 PRIVATE_BASE (Memory Aperture definition).
238 PRIVATE_LIMIT (Memory Aperture definition).
239 POPS_EXITING_WAVE_ID .
240 0.5.
241 -0.5.
242 1.0.
243 -1.0.
244 2.0.
245 -2.0.
246 4.0.
247 -4.0.
248 1/(2*PI).
249 SDWA
250 DPP16
251 VCCZ.
252 EXECZ.
253 SCC.
254 Reserved.
255 Literal constant.
256 - 511 VGPR 0 - 255

OP [16:9] See Opcode table below.

VDST [24:17] Destination VGPR.

ENCODING [31:25] 0_111111

Table 78. VOP1 Opcodes

Opcode # Name Opcode # Name

0 V_NOP 53 V_SIN_F32

1 V_MOV_B32 54 V_COS_F32

2 V_READFIRSTLANE_B32 55 V_NOT_B32

3 V_CVT_I32_F64 56 V_BFREV_B32

4 V_CVT_F64_I32 57 V_FFBH_U32

13.3. Vector ALU Formats 250 of 283

"RDNA 2" Instruction Set Architecture

Opcode # Name Opcode # Name

5 V_CVT_F32_I32 58 V_FFBL_B32

6 V_CVT_F32_U32 59 V_FFBH_I32

7 V_CVT_U32_F32 60 V_FREXP_EXP_I32_F64

8 V_CVT_I32_F32 61 V_FREXP_MANT_F64

10 V_CVT_F16_F32 62 V_FRACT_F64

11 V_CVT_F32_F16 63 V_FREXP_EXP_I32_F32

12 V_CVT_RPI_I32_F32 64 V_FREXP_MANT_F32

13 V_CVT_FLR_I32_F32 65 V_CLREXCP

14 V_CVT_OFF_F32_I4 66 V_MOVRELD_B32

15 V_CVT_F32_F64 67 V_MOVRELS_B32

16 V_CVT_F64_F32 68 V_MOVRELSD_B32

17 V_CVT_F32_UBYTE0 72 V_MOVRELSD_2_B32

18 V_CVT_F32_UBYTE1 80 V_CVT_F16_U16

19 V_CVT_F32_UBYTE2 81 V_CVT_F16_I16

20 V_CVT_F32_UBYTE3 82 V_CVT_U16_F16

21 V_CVT_U32_F64 83 V_CVT_I16_F16

22 V_CVT_F64_U32 84 V_RCP_F16

23 V_TRUNC_F64 85 V_SQRT_F16

24 V_CEIL_F64 86 V_RSQ_F16

25 V_RNDNE_F64 87 V_LOG_F16

26 V_FLOOR_F64 88 V_EXP_F16

27 V_PIPEFLUSH 89 V_FREXP_MANT_F16

32 V_FRACT_F32 90 V_FREXP_EXP_I16_F16

33 V_TRUNC_F32 91 V_FLOOR_F16

34 V_CEIL_F32 92 V_CEIL_F16

35 V_RNDNE_F32 93 V_TRUNC_F16

36 V_FLOOR_F32 94 V_RNDNE_F16

37 V_EXP_F32 95 V_FRACT_F16

39 V_LOG_F32 96 V_SIN_F16

42 V_RCP_F32 97 V_COS_F16

43 V_RCP_IFLAG_F32 98 V_SAT_PK_U8_I16

46 V_RSQ_F32 99 V_CVT_NORM_I16_F16

47 V_RCP_F64 100 V_CVT_NORM_U16_F16

49 V_RSQ_F64 101 V_SWAP_B32

13.3. Vector ALU Formats 251 of 283

"RDNA 2" Instruction Set Architecture

Opcode # Name Opcode # Name

51 V_SQRT_F32 104 V_SWAPREL_B32

52 V_SQRT_F64

13.3.3. VOPC

Format VOPC

Description Vector instruction taking two inputs and producing a comparison result. Can
be followed by a 32- bit literal constant. Vector Comparison operations are
divided into three groups:

• those which can use any one of 16 comparison operations,

• those which can use any one of 8, and
• those which have only a single comparison operation.

The final opcode number is determined by adding the base for the opcode family plus the offset
from the compare op. Compare instructions write a result to VCC (for VOPC) or an SGPR (for
VOP3). Additionally, every compare instruction has a variant that writes to the EXEC mask
instead of VCC or SGPR. The destination of the compare result is VCC or EXEC when encoded
using the VOPC format, and can be an arbitrary SGPR when only encoded in the VOP3 format.

Comparison Operations

Table 79. Comparison Operations

Compare Operation Opcode Description
Offset

Sixteen Compare Operations (OP16)

F 0 D.u = 0

LT 1 D.u = (S0 < S1)

EQ 2 D.u = (S0 == S1)

LE 3 D.u = (S0 <= S1)

GT 4 D.u = (S0 > S1)

LG 5 D.u = (S0 <> S1)

GE 6 D.u = (S0 >= S1)

O 7 D.u = (!isNaN(S0) && !isNaN(S1))

U 8 D.u = (!isNaN(S0) || !isNaN(S1))

13.3. Vector ALU Formats 252 of 283

"RDNA 2" Instruction Set Architecture

Compare Operation Opcode Description

Offset

NGE 9 D.u = !(S0 >= S1)

NLG 10 D.u = !(S0 <> S1)

NGT 11 D.u = !(S0 > S1)

NLE 12 D.u = !(S0 <= S1)

NEQ 13 D.u = !(S0 == S1)

NLT 14 D.u = !(S0 < S1)

TRU 15 D.u = 1

Eight Compare Operations (OP8)

F 0 D.u = 0

LT 1 D.u = (S0 < S1)

EQ 2 D.u = (S0 == S1)

LE 3 D.u = (S0 <= S1)

GT 4 D.u = (S0 > S1)

LG 5 D.u = (S0 <> S1)

GE 6 D.u = (S0 >= S1)

TRU 7 D.u = 1

Table 80. VOPC Fields

13.3. Vector ALU Formats 253 of 283

"RDNA 2" Instruction Set Architecture

Field Name Bits Format or Description

SRC0 [8:0] Source 0. First operand for the instruction.

0 - 105 SGPR0 to SGPR105: Scalar general-purpose registers.
106 VCC_LO: vcc[31:0].
107 VCC_HI: vcc[63:32].
108-123 TTMP0 - TTMP15: Trap handler temporary register.
124 M0. Memory register 0.
125 NULL
126 EXEC_LO: exec[31:0].
127 EXEC_HI: exec[63:32].
128 0.
129-192 Signed integer 1 to 64.
193-208 Signed integer -1 to -16.
209-232 Reserved.
233 DPP8
234 DPP8FI
235 SHARED_BASE (Memory Aperture definition).
236 SHARED_LIMIT (Memory Aperture definition).
237 PRIVATE_BASE (Memory Aperture definition).
238 PRIVATE_LIMIT (Memory Aperture definition).
239 POPS_EXITING_WAVE_ID .
240 0.5.
241 -0.5.
242 1.0.
243 -1.0.
244 2.0.
245 -2.0.
246 4.0.
247 -4.0.
248 1/(2*PI).
249 SDWA
250 DPP16
251 VCCZ.
252 EXECZ.
253 SCC.
254 Reserved.
255 Literal constant.
256 - 511 VGPR 0 - 255

VSRC1 [16:9] VGPR which provides the second operand.

OP [24:17] See Opcode table below.

ENCODING [31:25] 0_111110

Table 81. VOPC Opcodes

Opcode # Name Opcode # Name

0 V_CMP_F_F32 159 V_CMPX_CLASS_F16

1 V_CMP_LT_F32 160 V_CMP_F_I64

2 V_CMP_EQ_F32 161 V_CMP_LT_I64

3 V_CMP_LE_F32 162 V_CMP_EQ_I64

4 V_CMP_GT_F32 163 V_CMP_LE_I64

13.3. Vector ALU Formats 254 of 283

"RDNA 2" Instruction Set Architecture

Opcode # Name Opcode # Name

5 V_CMP_LG_F32 164 V_CMP_GT_I64

6 V_CMP_GE_F32 165 V_CMP_NE_I64

7 V_CMP_O_F32 166 V_CMP_GE_I64

8 V_CMP_U_F32 167 V_CMP_T_I64

9 V_CMP_NGE_F32 168 V_CMP_CLASS_F64

10 V_CMP_NLG_F32 169 V_CMP_LT_U16

11 V_CMP_NGT_F32 170 V_CMP_EQ_U16

12 V_CMP_NLE_F32 171 V_CMP_LE_U16

13 V_CMP_NEQ_F32 172 V_CMP_GT_U16

14 V_CMP_NLT_F32 173 V_CMP_NE_U16

15 V_CMP_TRU_F32 174 V_CMP_GE_U16

16 V_CMPX_F_F32 176 V_CMPX_F_I64

17 V_CMPX_LT_F32 177 V_CMPX_LT_I64

18 V_CMPX_EQ_F32 178 V_CMPX_EQ_I64

19 V_CMPX_LE_F32 179 V_CMPX_LE_I64

20 V_CMPX_GT_F32 180 V_CMPX_GT_I64

21 V_CMPX_LG_F32 181 V_CMPX_NE_I64

22 V_CMPX_GE_F32 182 V_CMPX_GE_I64

23 V_CMPX_O_F32 183 V_CMPX_T_I64

24 V_CMPX_U_F32 184 V_CMPX_CLASS_F64

25 V_CMPX_NGE_F32 185 V_CMPX_LT_U16

26 V_CMPX_NLG_F32 186 V_CMPX_EQ_U16

27 V_CMPX_NGT_F32 187 V_CMPX_LE_U16

28 V_CMPX_NLE_F32 188 V_CMPX_GT_U16

29 V_CMPX_NEQ_F32 189 V_CMPX_NE_U16

30 V_CMPX_NLT_F32 190 V_CMPX_GE_U16

31 V_CMPX_TRU_F32 192 V_CMP_F_U32

32 V_CMP_F_F64 193 V_CMP_LT_U32

33 V_CMP_LT_F64 194 V_CMP_EQ_U32

34 V_CMP_EQ_F64 195 V_CMP_LE_U32

35 V_CMP_LE_F64 196 V_CMP_GT_U32

36 V_CMP_GT_F64 197 V_CMP_NE_U32

37 V_CMP_LG_F64 198 V_CMP_GE_U32

38 V_CMP_GE_F64 199 V_CMP_T_U32

13.3. Vector ALU Formats 255 of 283

"RDNA 2" Instruction Set Architecture

Opcode # Name Opcode # Name

39 V_CMP_O_F64 200 V_CMP_F_F16

40 V_CMP_U_F64 201 V_CMP_LT_F16

41 V_CMP_NGE_F64 202 V_CMP_EQ_F16

42 V_CMP_NLG_F64 203 V_CMP_LE_F16

43 V_CMP_NGT_F64 204 V_CMP_GT_F16

44 V_CMP_NLE_F64 205 V_CMP_LG_F16

45 V_CMP_NEQ_F64 206 V_CMP_GE_F16

46 V_CMP_NLT_F64 207 V_CMP_O_F16

47 V_CMP_TRU_F64 208 V_CMPX_F_U32

48 V_CMPX_F_F64 209 V_CMPX_LT_U32

49 V_CMPX_LT_F64 210 V_CMPX_EQ_U32

50 V_CMPX_EQ_F64 211 V_CMPX_LE_U32

51 V_CMPX_LE_F64 212 V_CMPX_GT_U32

52 V_CMPX_GT_F64 213 V_CMPX_NE_U32

53 V_CMPX_LG_F64 214 V_CMPX_GE_U32

54 V_CMPX_GE_F64 215 V_CMPX_T_U32

55 V_CMPX_O_F64 216 V_CMPX_F_F16

56 V_CMPX_U_F64 217 V_CMPX_LT_F16

57 V_CMPX_NGE_F64 218 V_CMPX_EQ_F16

58 V_CMPX_NLG_F64 219 V_CMPX_LE_F16

59 V_CMPX_NGT_F64 220 V_CMPX_GT_F16

60 V_CMPX_NLE_F64 221 V_CMPX_LG_F16

61 V_CMPX_NEQ_F64 222 V_CMPX_GE_F16

62 V_CMPX_NLT_F64 223 V_CMPX_O_F16

63 V_CMPX_TRU_F64 224 V_CMP_F_U64

128 V_CMP_F_I32 225 V_CMP_LT_U64

129 V_CMP_LT_I32 226 V_CMP_EQ_U64

130 V_CMP_EQ_I32 227 V_CMP_LE_U64

131 V_CMP_LE_I32 228 V_CMP_GT_U64

132 V_CMP_GT_I32 229 V_CMP_NE_U64

133 V_CMP_NE_I32 230 V_CMP_GE_U64

134 V_CMP_GE_I32 231 V_CMP_T_U64

135 V_CMP_T_I32 232 V_CMP_U_F16

136 V_CMP_CLASS_F32 233 V_CMP_NGE_F16

13.3. Vector ALU Formats 256 of 283

"RDNA 2" Instruction Set Architecture

Opcode # Name Opcode # Name

137 V_CMP_LT_I16 234 V_CMP_NLG_F16

138 V_CMP_EQ_I16 235 V_CMP_NGT_F16

139 V_CMP_LE_I16 236 V_CMP_NLE_F16

140 V_CMP_GT_I16 237 V_CMP_NEQ_F16

141 V_CMP_NE_I16 238 V_CMP_NLT_F16

142 V_CMP_GE_I16 239 V_CMP_TRU_F16

143 V_CMP_CLASS_F16 240 V_CMPX_F_U64

144 V_CMPX_F_I32 241 V_CMPX_LT_U64

145 V_CMPX_LT_I32 242 V_CMPX_EQ_U64

146 V_CMPX_EQ_I32 243 V_CMPX_LE_U64

147 V_CMPX_LE_I32 244 V_CMPX_GT_U64

148 V_CMPX_GT_I32 245 V_CMPX_NE_U64

149 V_CMPX_NE_I32 246 V_CMPX_GE_U64

150 V_CMPX_GE_I32 247 V_CMPX_T_U64

151 V_CMPX_T_I32 248 V_CMPX_U_F16

152 V_CMPX_CLASS_F32 249 V_CMPX_NGE_F16

153 V_CMPX_LT_I16 250 V_CMPX_NLG_F16

154 V_CMPX_EQ_I16 251 V_CMPX_NGT_F16

155 V_CMPX_LE_I16 252 V_CMPX_NLE_F16

156 V_CMPX_GT_I16 253 V_CMPX_NEQ_F16

157 V_CMPX_NE_I16 254 V_CMPX_NLT_F16

158 V_CMPX_GE_I16 255 V_CMPX_TRU_F16

13.3.4. VOP3A

Format VOP3A

Description Vector ALU format with three operands

Table 82. VOP3A Fields

13.3. Vector ALU Formats 257 of 283

"RDNA 2" Instruction Set Architecture

Field Name Bits Format or Description

VDST [7:0] Destination VGPR

ABS [10:8] Absolute value of input. [8] = src0, [9] = src1, [10] = src2

OPSEL [14:11] Operand select for 16-bit data. 0 = select low half, 1 = select high half. [11] =
src0, [12] = src1, [13] = src2, [14] = dest.

CLMP [15] Clamp output

OP [25:16] Opcode. See next table.

ENCODING [31:26] 110101

SRC0 [40:32] Source 0. First operand for the instruction.

0 - 105 SGPR0 to SGPR105: Scalar general-purpose registers.
106 VCC_LO: vcc[31:0].
107 VCC_HI: vcc[63:32].
108-123 TTMP0 - TTMP15: Trap handler temporary register.
124 M0. Memory register 0.
125 NULL
126 EXEC_LO: exec[31:0].
127 EXEC_HI: exec[63:32].
128 0.
129-192 Signed integer 1 to 64.
193-208 Signed integer -1 to -16.
209-232 Reserved.
233 DPP8
234 DPP8FI
235 SHARED_BASE (Memory Aperture definition).
236 SHARED_LIMIT (Memory Aperture definition).
237 PRIVATE_BASE (Memory Aperture definition).
238 PRIVATE_LIMIT (Memory Aperture definition).
239 POPS_EXITING_WAVE_ID .
240 0.5.
241 -0.5.
242 1.0.
243 -1.0.
244 2.0.
245 -2.0.
246 4.0.
247 -4.0.
248 1/(2*PI).
249 SDWA
250 DPP16
251 VCCZ.
252 EXECZ.
253 SCC.
254 Reserved.
255 Literal constant.
256 - 511 VGPR 0 - 255

SRC1 [49:41] Second input operand. Same options as SRC0.

SRC2 [58:50] Third input operand. Same options as SRC0.

OMOD [60:59] Output Modifier: 0=none, 1=*2, 2=*4, 3=div-2

NEG [63:61] Negate input. [61] = src0, [62] = src1, [63] = src2

13.3. Vector ALU Formats 258 of 283

"RDNA 2" Instruction Set Architecture

Table 83. VOP3A Opcodes

Opcode # Name Opcode # Name

320 V_FMA_LEGACY_F32 775 V_LSHRREV_B16

322 V_MAD_I32_I24 776 V_ASHRREV_I16

323 V_MAD_U32_U24 777 V_MAX_U16

324 V_CUBEID_F32 778 V_MAX_I16

325 V_CUBESC_F32 779 V_MIN_U16

326 V_CUBETC_F32 780 V_MIN_I16

327 V_CUBEMA_F32 781 V_ADD_NC_I16

328 V_BFE_U32 782 V_SUB_NC_I16

329 V_BFE_I32 785 V_PACK_B32_F16

330 V_BFI_B32 786 V_CVT_PKNORM_I16_F16

331 V_FMA_F32 787 V_CVT_PKNORM_U16_F16

332 V_FMA_F64 788 V_LSHLREV_B16

333 V_LERP_U8 832 V_MAD_U16

334 V_ALIGNBIT_B32 834 V_INTERP_P1LL_F16

335 V_ALIGNBYTE_B32 835 V_INTERP_P1LV_F16

336 V_MULLIT_F32 836 V_PERM_B32

337 V_MIN3_F32 837 V_XAD_U32

338 V_MIN3_I32 838 V_LSHL_ADD_U32

339 V_MIN3_U32 839 V_ADD_LSHL_U32

340 V_MAX3_F32 843 V_FMA_F16

341 V_MAX3_I32 849 V_MIN3_F16

342 V_MAX3_U32 850 V_MIN3_I16

343 V_MED3_F32 851 V_MIN3_U16

344 V_MED3_I32 852 V_MAX3_F16

345 V_MED3_U32 853 V_MAX3_I16

346 V_SAD_U8 854 V_MAX3_U16

347 V_SAD_HI_U8 855 V_MED3_F16

348 V_SAD_U16 856 V_MED3_I16

349 V_SAD_U32 857 V_MED3_U16

350 V_CVT_PK_U8_F32 858 V_INTERP_P2_F16

351 V_DIV_FIXUP_F32 862 V_MAD_I16

352 V_DIV_FIXUP_F64 863 V_DIV_FIXUP_F16

356 V_ADD_F64 864 V_READLANE_B32

13.3. Vector ALU Formats 259 of 283

"RDNA 2" Instruction Set Architecture

Opcode # Name Opcode # Name

357 V_MUL_F64 865 V_WRITELANE_B32

358 V_MIN_F64 866 V_LDEXP_F32

359 V_MAX_F64 867 V_BFM_B32

360 V_LDEXP_F64 868 V_BCNT_U32_B32

361 V_MUL_LO_U32 869 V_MBCNT_LO_U32_B32

362 V_MUL_HI_U32 870 V_MBCNT_HI_U32_B32

364 V_MUL_HI_I32 872 V_CVT_PKNORM_I16_F32

367 V_DIV_FMAS_F32 873 V_CVT_PKNORM_U16_F32

368 V_DIV_FMAS_F64 874 V_CVT_PK_U16_U32

369 V_MSAD_U8 875 V_CVT_PK_I16_I32

370 V_QSAD_PK_U16_U8 877 V_ADD3_U32

371 V_MQSAD_PK_U16_U8 879 V_LSHL_OR_B32

372 V_TRIG_PREOP_F64 881 V_AND_OR_B32

373 V_MQSAD_U32_U8 882 V_OR3_B32

376 V_XOR3_B32 883 V_MAD_U32_U16

767 V_LSHLREV_B64 885 V_MAD_I32_I16

768 V_LSHRREV_B64 886 V_SUB_NC_I32

769 V_ASHRREV_I64 887 V_PERMLANE16_B32

771 V_ADD_NC_U16 888 V_PERMLANEX16_B32

772 V_SUB_NC_U16 895 V_ADD_NC_I32

773 V_MUL_LO_U16

13.3.5. VOP3B

Format VOP3B

Description Vector ALU format with three operands and a scalar result. This encoding
is used only for a few opcodes.

This encoding allows specifying a unique scalar destination, and is used only for the opcodes
listed below. All other opcodes use VOP3A.

13.3. Vector ALU Formats 260 of 283

"RDNA 2" Instruction Set Architecture

• V_ADD_CO_U32
• V_SUB_CO_U32
• V_SUBREV_CO_U32
• V_ADDC_CO_U32
• V_SUBB_CO_U32
• V_SUBBREV_CO_U32
• V_DIV_SCALE_F32
• V_DIV_SCALE_F64
• V_MAD_U64_U32
• V_MAD_I64_I32

Table 84. VOP3B Fields

Field Name Bits Format or Description

VDST [7:0] Destination VGPR

SDST [14:8] Scalar destination

CLMP [15] Clamp result

OP [25:16] Opcode. see next table.

ENCODING [31:26] 110101

13.3. Vector ALU Formats 261 of 283

"RDNA 2" Instruction Set Architecture

Field Name Bits Format or Description

SRC0 [40:32] Source 0. First operand for the instruction.

0 - 105 SGPR0 to SGPR105: Scalar general-purpose registers.
106 VCC_LO: vcc[31:0].
107 VCC_HI: vcc[63:32].
108-123 TTMP0 - TTMP15: Trap handler temporary register.
124 M0. Memory register 0.
125 NULL
126 EXEC_LO: exec[31:0].
127 EXEC_HI: exec[63:32].
128 0.
129-192 Signed integer 1 to 64.
193-208 Signed integer -1 to -16.
209-232 Reserved.
233 DPP8
234 DPP8FI
235 SHARED_BASE (Memory Aperture definition).
236 SHARED_LIMIT (Memory Aperture definition).
237 PRIVATE_BASE (Memory Aperture definition).
238 PRIVATE_LIMIT (Memory Aperture definition).
239 POPS_EXITING_WAVE_ID .
240 0.5.
241 -0.5.
242 1.0.
243 -1.0.
244 2.0.
245 -2.0.
246 4.0.
247 -4.0.
248 1/(2*PI).
249 SDWA
250 DPP16
251 VCCZ.
252 EXECZ.
253 SCC.
254 Reserved.
255 Literal constant.
256 - 511 VGPR 0 - 255

SRC1 [49:41] Second input operand. Same options as SRC0.

SRC2 [58:50] Third input operand. Same options as SRC0.

OMOD [60:59] Output Modifier: 0=none, 1=*2, 2=*4, 3=div-2

NEG [63:61] Negate input. [61] = src0, [62] = src1, [63] = src2

Table 85. VOP3B Opcodes

Opcode # Name Opcode # Name

365 V_DIV_SCALE_F32 783 V_ADD_CO_U32

366 V_DIV_SCALE_F64 784 V_SUB_CO_U32

374 V_MAD_U64_U32 793 V_SUBREV_CO_U32

375 V_MAD_I64_I32

13.3. Vector ALU Formats 262 of 283

"RDNA 2" Instruction Set Architecture

13.3.6. VOP3P

Format VOP3P

Description Vector ALU format taking one, two or three pairs of 16 bit inputs and
producing two 16-bit outputs (packed into 1 dword).

Table 86. VOP3P Fields

Field Name Bits Format or Description

VDST [7:0] Destination VGPR

NEG_HI [10:8] Negate sources 0,1,2 of the high 16-bits.

OPSEL [13:11] Select low or high for low sources 0=[11], 1=[12], 2=[13].

OPSEL_HI2 [14] Select low or high for high sources 0=[14], 1=[60], 2=[59].

CLMP [15] 1 = clamp result.

OP [22:16] Opcode. see next table.

ENCODING [31:26] 110011

13.3. Vector ALU Formats 263 of 283

"RDNA 2" Instruction Set Architecture

Field Name Bits Format or Description

SRC0 [40:32] Source 0. First operand for the instruction.

0 - 105 SGPR0 to SGPR105: Scalar general-purpose registers.
106 VCC_LO: vcc[31:0].
107 VCC_HI: vcc[63:32].
108-123 TTMP0 - TTMP15: Trap handler temporary register.
124 M0. Memory register 0.
125 NULL
126 EXEC_LO: exec[31:0].
127 EXEC_HI: exec[63:32].
128 0.
129-192 Signed integer 1 to 64.
193-208 Signed integer -1 to -16.
209-232 Reserved.
233 DPP8
234 DPP8FI
235 SHARED_BASE (Memory Aperture definition).
236 SHARED_LIMIT (Memory Aperture definition).
237 PRIVATE_BASE (Memory Aperture definition).
238 PRIVATE_LIMIT (Memory Aperture definition).
239 POPS_EXITING_WAVE_ID .
240 0.5.
241 -0.5.
242 1.0.
243 -1.0.
244 2.0.
245 -2.0.
246 4.0.
247 -4.0.
248 1/(2*PI).
249 SDWA
250 DPP16
251 VCCZ.
252 EXECZ.
253 SCC.
254 Reserved.
255 Literal constant.
256 - 511 VGPR 0 - 255

SRC1 [49:41] Second input operand. Same options as SRC0.

SRC2 [58:50] Third input operand. Same options as SRC0.

OPSEL_HI [60:59] See OP_SEL_HI2.

NEG [63:61] Negate input for low 16-bits of sources. [61] = src0, [62] = src1, [63] = src2

Table 87. VOP3P Opcodes

Opcode # Name Opcode # Name

0 V_PK_MAD_I16 15 V_PK_ADD_F16

1 V_PK_MUL_LO_U16 16 V_PK_MUL_F16

2 V_PK_ADD_I16 17 V_PK_MIN_F16

3 V_PK_SUB_I16 18 V_PK_MAX_F16

13.3. Vector ALU Formats 264 of 283

"RDNA 2" Instruction Set Architecture

Opcode # Name Opcode # Name

4 V_PK_LSHLREV_B16 19 V_DOT2_F32_F16

5 V_PK_LSHRREV_B16 20 V_DOT2_I32_I16

6 V_PK_ASHRREV_I16 21 V_DOT2_U32_U16

7 V_PK_MAX_I16 22 V_DOT4_I32_I8

8 V_PK_MIN_I16 23 V_DOT4_U32_U8

9 V_PK_MAD_U16 24 V_DOT8_I32_I4

10 V_PK_ADD_U16 25 V_DOT8_U32_U4

11 V_PK_SUB_U16 32 V_FMA_MIX_F32

12 V_PK_MAX_U16 33 V_FMA_MIXLO_F16

13 V_PK_MIN_U16 34 V_FMA_MIXHI_F16

14 V_PK_FMA_F16

13.3.7. SDWA

Format SDWA

Description Sub-Dword Addressing. This is a second dword which can follow VOP1 or
VOP2 instructions (in place of a literal constant) to control selection of sub-
dword (8-bit and 16-bit) operands. Use of SDWA is indicated by assigning
the SRC0 field to SDWA, and then the actual VGPR used as source-zero is
determined in SDWA instruction word.

Table 88. SDWA Fields

Field Name Bits Format or Description

SRC0 [39:32] Real SRC0 operand (VGPR).

DST_SEL [42:40] Select the data destination:

0 = data[7:0]
1 = data[15:8]
2 = data[23:16]
3 = data[31:24]
4 = data[15:0]
5 = data[31:16]
6 = data[31:0]
7 = reserved

13.3. Vector ALU Formats 265 of 283

"RDNA 2" Instruction Set Architecture

Field Name Bits Format or Description

DST_U [44:43] Destination format: what do with the bits in the VGPR that are not selected by
DST_SEL:
0 = pad with zeros + 1 = sign extend upper / zero lower
2 = preserve (don’t modify)
3 = reserved

CLMP [45] 1 = clamp result

OMOD [47:46] Output modifiers (see VOP3). [46] = low half, [47] = high half

SRC0_SEL [50:48] Source 0 select. Same options as DST_SEL.

SRC0_SEXT [51] Sign extend modifier for source 0.

SRC0_NEG [52] 1 = negate source 0.

SRC0_ABS [53] 1 = Absolute value of source 0.

S0 [55] 0 = source 0 is VGPR, 1 = is SGPR.

SRC1_SEL [58:56] Same options as SRC0_SEL.

SRC1_SEXT [59] Sign extend modifier for source 1.

SRC1_NEG [60] 1 = negate source 1.

SRC1_ABS [61] 1 = Absolute value of source 1.

S1 [63] 0 = source 1 is VGPR, 1 = is SGPR.

13.3.8. SDWAB

Format SDWAB

Description Sub-Dword Addressing. This is a second dword which can follow VOPC
instructions (in place of a literal constant) to control selection of sub-dword
(8-bit and 16-bit) operands. Use of SDWA is indicated by assigning the
SRC0 field to SDWA, and then the actual VGPR used as source-zero is
determined in SDWA instruction word. This version has a scalar
destination.

Table 89. SDWAB Fields

Field Name Bits Format or Description

SRC0 [39:32] Real SRC0 operand (VGPR).

SDST [46:40] Scalar GPR destination.

SD [47] Scalar destination type: 0 = VCC, 1 = normal SGPR.

SRC0_SEL [50:48] Source 0 select. Same options as DST_SEL.

13.3. Vector ALU Formats 266 of 283

"RDNA 2" Instruction Set Architecture

Field Name Bits Format or Description

SRC0_SEXT [51] Sign extend modifier for source 0.

SRC0_NEG [52] 1 = negate source 0.

SRC0_ABS [53] 1 = Absolute value of source 0.

S0 [55] 0 = source 0 is VGPR, 1 = is SGPR.

SRC1_SEL [58:56] Same options as SRC0_SEL.

SRC1_SEXT [59] Sign extend modifier for source 1.

SRC1_NEG [60] 1 = negate source 1.

SRC1_ABS [61] 1 = Absolute value of source 1.

S1 [63] 0 = source 1 is VGPR, 1 = is SGPR.

13.3.9. DPP16

Format DPP16

Description Data Parallel Primitives over 16 lanes. This is a second dword which can
follow VOP1, VOP2 or VOPC instructions (in place of a literal constant) to
control selection of data from other lanes.

Table 90. DPP16 Fields

Field Name Bits Format or Description

SRC0 [39:32] Real SRC0 operand (VGPR).

DPP_CTRL [48:40] See next table: "DPP_CTRL Enumeration"

FI [50] Fetch invalid data: 0 = read zero for any inactive lanes; 1 = read VGPRs even
for invalid lanes.

BC [51] Bounds Control: 0 = do not write when source is out of range, 1 = write.

SRC0_NEG [52] 1 = negate source 0.

SRC0_ABS [53] 1 = Absolute value of source 0.

SRC1_NEG [54] 1 = negate source 1.

SRC1_ABS [55] 1 = Absolute value of source 1.

13.3. Vector ALU Formats 267 of 283

"RDNA 2" Instruction Set Architecture

Field Name Bits Format or Description

BANK_MASK [59:56] Bank Mask Applies to the VGPR destination write only, does not impact the
thread mask when fetching source VGPR data.
27==0: lanes[12:15, 28:31, 44:47, 60:63] are disabled
26==0: lanes[8:11, 24:27, 40:43, 56:59] are disabled
25==0: lanes[4:7, 20:23, 36:39, 52:55] are disabled
24==0: lanes[0:3, 16:19, 32:35, 48:51] are disabled
Notice: the term "bank" here is not the same as was used for the VGPR bank.

ROW_MASK [63:60] Row Mask Applies to the VGPR destination write only, does not impact the
thread mask when fetching source VGPR data.
31==0: lanes[63:48] are disabled (wave 64 only)
30==0: lanes[47:32] are disabled (wave 64 only)
29==0: lanes[31:16] are disabled
28==0: lanes[15:0] are disabled

Table 91. DPP_CTRL Enumeration

DPP_Cntl Hex Function Description
Enumeration Value

DPP_QUAD_PER 000- pix[n].srca = pix[(n&0x3c)+ dpp_cntl[n%4*2+1 : n%4*2]].srca Permute of four

M* 0FF threads.

DPP_UNUSED 100 Undefined Reserved.

DPP_ROW_SL* 101- if ((n&0xf) < (16-cntl[3:0])) pix[n].srca = pix[n+ cntl[3:0]].srca Row shift left by 1-15
10F else use bound_cntl threads.

DPP_ROW_SR* 111- if \((n&0xf) >= cntl[3:0]) pix[n].srca = pix[n - cntl[3:0]].srca Row shift right by 1-15
11F else use bound_cntl threads.

DPP_ROW_RR* 121- if \((n&0xf) >= cnt[3:0]) pix[n].srca = pix[n - cntl[3:0]].srca Row rotate right by 1-15
12F else pix[n].srca = pix[n + 16 - cntl[3:0]].srca threads.

DPP_ROW_MIRR 140 pix[n].srca = pix[15-(n&f)].srca Mirror threads within

OR* row.

DPP_ROW_HALF_ 141 pix[n].srca = pix[7-(n&7)].srca Mirror threads within

MIRROR* row (8 threads).

13.3.10. DPP8

Format DPP8

Description Data Parallel Primitives over 8 lanes. This is a second dword which can
follow VOP1, VOP2 or VOPC instructions (in place of a literal constant) to
control selection of data from other lanes.

Table 92. DPP8 Fields

13.3. Vector ALU Formats 268 of 283

"RDNA 2" Instruction Set Architecture

Field Name Bits Format or Description

SRC0 [39:32] Real SRC0 operand (VGPR).

LANE_SEL0 42:40 Which lane to read for 1st output lane per 8-lane group

LANE_SEL1 45:43 Which lane to read for 2nd output lane per 8-lane group

LANE_SEL2 48:46 Which lane to read for 3rd output lane per 8-lane group

LANE_SEL3 51:49 Which lane to read for 4th output lane per 8-lane group

LANE_SEL4 54:52 Which lane to read for 5th output lane per 8-lane group

LANE_SEL5 57:55 Which lane to read for 6th output lane per 8-lane group

LANE_SEL6 60:58 Which lane to read for 7th output lane per 8-lane group

LANE_SEL7 63:61 Which lane to read for 8th output lane per 8-lane group

13.4. Vector Parameter Interpolation Format

13.4.1. VINTRP

Format VINTRP

Description Vector Parameter Interpolation.

These opcodes perform parameter interpolation using vertex data in pixel
shaders.

Table 93. VINTRP Fields

Field Name Bits Format or Description

VSRC [7:0] SRC0 operand (VGPR).

ATTR_CHAN [9:8] Attribute channel: 0=X, 1=Y, 2=Z, 3=W

ATTR [15:10] Attribute number: 0 - 32.

OP [17:16] Opcode:
0: v_interp_p1_f32 : VDST = P10 * VSRC + P0
1: v_interp_p2_f32: VDST = P20 * VSRC + VDST
2: v_interp_mov_f32: VDST = (P0, P10 or P20 selected by VSRC[1:0])

VDST [25:18] Destination VGPR

ENCODING [31:26] 110010

 User must set VSRC to be different from VDST.

13.4. Vector Parameter Interpolation Format 269 of 283

"RDNA 2" Instruction Set Architecture

13.5. LDS and GDS format

13.5.1. DS

Format LDS and GDS

Description Local and Global Data Sharing instructions

Table 94. DS Fields

Field Name Bits Format or Description

OFFSET0 [7:0] First address offset

OFFSET1 [15:8] Second address offset. For some opcodes this is concatenated with OFFSET0.

GDS [16] 1=GDS, 0=LDS operation.

OP [24:17] See Opcode table below.

ENCODING [31:26] 110110

ADDR [39:32] VGPR which supplies the address.

DATA0 [47:40] First data VGPR.

DATA1 [55:48] Second data VGPR.

VDST [63:56] Destination VGPR when results returned to VGPRs.

Table 95. DS Opcodes

Opcode # Name Opcode # Name

0 DS_ADD_U32 64 DS_ADD_U64

1 DS_SUB_U32 65 DS_SUB_U64

2 DS_RSUB_U32 66 DS_RSUB_U64

3 DS_INC_U32 67 DS_INC_U64

4 DS_DEC_U32 68 DS_DEC_U64

5 DS_MIN_I32 69 DS_MIN_I64

6 DS_MAX_I32 70 DS_MAX_I64

7 DS_MIN_U32 71 DS_MIN_U64

8 DS_MAX_U32 72 DS_MAX_U64

9 DS_AND_B32 73 DS_AND_B64

10 DS_OR_B32 74 DS_OR_B64

13.5. LDS and GDS format 270 of 283

"RDNA 2" Instruction Set Architecture

Opcode # Name Opcode # Name

11 DS_XOR_B32 75 DS_XOR_B64

12 DS_MSKOR_B32 76 DS_MSKOR_B64

13 DS_WRITE_B32 77 DS_WRITE_B64

14 DS_WRITE2_B32 78 DS_WRITE2_B64

15 DS_WRITE2ST64_B32 79 DS_WRITE2ST64_B64

16 DS_CMPST_B32 80 DS_CMPST_B64

17 DS_CMPST_F32 81 DS_CMPST_F64

18 DS_MIN_F32 82 DS_MIN_F64

19 DS_MAX_F32 83 DS_MAX_F64

20 DS_NOP 85 DS_ADD_RTN_F32

21 DS_ADD_F32 96 DS_ADD_RTN_U64

24 DS_GWS_SEMA_RELEASE_ALL 97 DS_SUB_RTN_U64

25 DS_GWS_INIT 98 DS_RSUB_RTN_U64

26 DS_GWS_SEMA_V 99 DS_INC_RTN_U64

27 DS_GWS_SEMA_BR 100 DS_DEC_RTN_U64

28 DS_GWS_SEMA_P 101 DS_MIN_RTN_I64

29 DS_GWS_BARRIER 102 DS_MAX_RTN_I64

30 DS_WRITE_B8 103 DS_MIN_RTN_U64

31 DS_WRITE_B16 104 DS_MAX_RTN_U64

32 DS_ADD_RTN_U32 105 DS_AND_RTN_B64

33 DS_SUB_RTN_U32 106 DS_OR_RTN_B64

34 DS_RSUB_RTN_U32 107 DS_XOR_RTN_B64

35 DS_INC_RTN_U32 108 DS_MSKOR_RTN_B64

36 DS_DEC_RTN_U32 109 DS_WRXCHG_RTN_B64

37 DS_MIN_RTN_I32 110 DS_WRXCHG2_RTN_B64

38 DS_MAX_RTN_I32 111 DS_WRXCHG2ST64_RTN_B64

39 DS_MIN_RTN_U32 112 DS_CMPST_RTN_B64

40 DS_MAX_RTN_U32 113 DS_CMPST_RTN_F64

41 DS_AND_RTN_B32 114 DS_MIN_RTN_F64

42 DS_OR_RTN_B32 115 DS_MAX_RTN_F64

43 DS_XOR_RTN_B32 118 DS_READ_B64

44 DS_MSKOR_RTN_B32 119 DS_READ2_B64

45 DS_WRXCHG_RTN_B32 120 DS_READ2ST64_B64

46 DS_WRXCHG2_RTN_B32 126 DS_CONDXCHG32_RTN_B64

13.5. LDS and GDS format 271 of 283

"RDNA 2" Instruction Set Architecture

Opcode # Name Opcode # Name

47 DS_WRXCHG2ST64_RTN_B32 160 DS_WRITE_B8_D16_HI

48 DS_CMPST_RTN_B32 161 DS_WRITE_B16_D16_HI

49 DS_CMPST_RTN_F32 162 DS_READ_U8_D16

50 DS_MIN_RTN_F32 163 DS_READ_U8_D16_HI

51 DS_MAX_RTN_F32 164 DS_READ_I8_D16

52 DS_WRAP_RTN_B32 165 DS_READ_I8_D16_HI

53 DS_SWIZZLE_B32 166 DS_READ_U16_D16

54 DS_READ_B32 167 DS_READ_U16_D16_HI

55 DS_READ2_B32 176 DS_WRITE_ADDTID_B32

56 DS_READ2ST64_B32 177 DS_READ_ADDTID_B32

57 DS_READ_I8 178 DS_PERMUTE_B32

58 DS_READ_U8 179 DS_BPERMUTE_B32

59 DS_READ_I16 222 DS_WRITE_B96

60 DS_READ_U16 223 DS_WRITE_B128

61 DS_CONSUME 254 DS_READ_B96

62 DS_APPEND 255 DS_READ_B128

63 DS_ORDERED_COUNT

13.6. Vector Memory Buffer Formats

There are two memory buffer instruction formats:

MTBUF
typed buffer access (data type is defined by the instruction)

MUBUF
untyped buffer access (data type is defined by the buffer / resource-constant)

13.6.1. MTBUF

Format MTBUF

13.6. Vector Memory Buffer Formats 272 of 283

"RDNA 2" Instruction Set Architecture

Description Memory Typed-Buffer Instructions

Table 96. MTBUF Fields

Field Name Bits Format or Description

OFFSET [11:0] Address offset, unsigned byte.

OFFEN [12] 1 = enable offset VGPR, 0 = use zero for address offset

IDXEN [13] 1 = enable index VGPR, 0 = use zero for address index

GLC [14] 0 = normal, 1 = globally coherent (bypass L0 cache) or for atomics, return pre-
op value to VGPR.

DLC [15] 0 = normal, 1 = Device Coherent

OP [53],[18:16 Opcode. See table below. (combined bits 53 with 18-16 to form opcode)
]

DFMT 25:19 Data Format of data in memory buffer. See chapter 8 for encoding. Buffer
Image format Table

ENCODING [31:26] 111010

VADDR [39:32] Address of VGPR to supply first component of address (offset or index). When
both index and offset are used, index is in the first VGPR and offset in the
second.

VDATA [47:40] Address of VGPR to supply first component of write data or receive first
component of read-data.

SRSRC [52:48] SGPR to supply V# (resource constant) in 4 or 8 consecutive SGPRs. It is

missing 2 LSB’s of SGPR-address since it is aligned to 4 SGPRs.

SLC [54] System Level Coherent. Used in conjunction with DLC to determine L2 cache
policies.

TFE [55] Partially resident texture, texture fail enable.

SOFFSET [63:56] Address offset, unsigned byte.

Table 97. MTBUF Opcodes

Opcode # Name Opcode # Name

0 TBUFFER_LOAD_FORMAT_X 8 TBUFFER_LOAD_FORMAT_D16_X

1 TBUFFER_LOAD_FORMAT_XY 9 TBUFFER_LOAD_FORMAT_D16_XY

2 TBUFFER_LOAD_FORMAT_XYZ 10 TBUFFER_LOAD_FORMAT_D16_XYZ

3 TBUFFER_LOAD_FORMAT_XYZW 11 TBUFFER_LOAD_FORMAT_D16_XYZW

4 TBUFFER_STORE_FORMAT_X 12 TBUFFER_STORE_FORMAT_D16_X

5 TBUFFER_STORE_FORMAT_XY 13 TBUFFER_STORE_FORMAT_D16_XY

6 TBUFFER_STORE_FORMAT_XYZ 14 TBUFFER_STORE_FORMAT_D16_XYZ

7 TBUFFER_STORE_FORMAT_XYZW 15 TBUFFER_STORE_FORMAT_D16_XYZW

13.6. Vector Memory Buffer Formats 273 of 283

"RDNA 2" Instruction Set Architecture

13.6.2. MUBUF

Format MUBUF

Description Memory Untyped-Buffer Instructions

Table 98. MUBUF Fields

Field Name Bits Format or Description

OFFSET [11:0] Address offset, unsigned byte.

OFFEN [12] 1 = enable offset VGPR, 0 = use zero for address offset

IDXEN [13] 1 = enable index VGPR, 0 = use zero for address index

GLC [14] 0 = normal, 1 = globally coherent (bypass L0 cache) or for atomics, return pre-
op value to VGPR.

DLC [15] 0 = normal, 1 = Device Coherent

LDS [16] 0 = normal, 1 = transfer data between LDS and memory instead of VGPRs and
memory.

OP [25:18] Opcode. See table below.

ENCODING [31:26] 111000

VADDR [39:32] Address of VGPR to supply first component of address (offset or index). When
both index and offset are used, index is in the first VGPR and offset in the
second.

VDATA [47:40] Address of VGPR to supply first component of write data or receive first
component of read-data.

SRSRC [52:48] SGPR to supply V# (resource constant) in 4 or 8 consecutive SGPRs. It is

missing 2 LSB’s of SGPR-address since it is aligned to 4 SGPRs.

SLC [54] System Level Coherent. Used in conjunction with DLC to determine L2 cache
policies.

TFE [55] Partially resident texture, texture fail enable.

SOFFSET [63:56] Address offset, unsigned byte.

Table 99. MUBUF Opcodes

Opcode # Name Opcode # Name

0 BUFFER_LOAD_FORMAT_X 54 BUFFER_ATOMIC_UMIN

1 BUFFER_LOAD_FORMAT_XY 55 BUFFER_ATOMIC_SMAX

2 BUFFER_LOAD_FORMAT_XYZ 56 BUFFER_ATOMIC_UMAX

3 BUFFER_LOAD_FORMAT_XYZW 57 BUFFER_ATOMIC_AND

13.6. Vector Memory Buffer Formats 274 of 283

"RDNA 2" Instruction Set Architecture

Opcode # Name Opcode # Name

4 BUFFER_STORE_FORMAT_X 58 BUFFER_ATOMIC_OR

5 BUFFER_STORE_FORMAT_XY 59 BUFFER_ATOMIC_XOR

6 BUFFER_STORE_FORMAT_XYZ 60 BUFFER_ATOMIC_INC

7 BUFFER_STORE_FORMAT_XYZW 61 BUFFER_ATOMIC_DEC

8 BUFFER_LOAD_UBYTE 62 BUFFER_ATOMIC_FCMPSWAP

9 BUFFER_LOAD_SBYTE 63 BUFFER_ATOMIC_FMIN

10 BUFFER_LOAD_USHORT 64 BUFFER_ATOMIC_FMAX

11 BUFFER_LOAD_SSHORT 80 BUFFER_ATOMIC_SWAP_X2

12 BUFFER_LOAD_DWORD 81 BUFFER_ATOMIC_CMPSWAP_X2

13 BUFFER_LOAD_DWORDX2 82 BUFFER_ATOMIC_ADD_X2

14 BUFFER_LOAD_DWORDX4 83 BUFFER_ATOMIC_SUB_X2

15 BUFFER_LOAD_DWORDX3 85 BUFFER_ATOMIC_SMIN_X2

24 BUFFER_STORE_BYTE 86 BUFFER_ATOMIC_UMIN_X2

25 BUFFER_STORE_BYTE_D16_HI 87 BUFFER_ATOMIC_SMAX_X2

26 BUFFER_STORE_SHORT 88 BUFFER_ATOMIC_UMAX_X2

27 BUFFER_STORE_SHORT_D16_HI 89 BUFFER_ATOMIC_AND_X2

28 BUFFER_STORE_DWORD 90 BUFFER_ATOMIC_OR_X2

29 BUFFER_STORE_DWORDX2 91 BUFFER_ATOMIC_XOR_X2

30 BUFFER_STORE_DWORDX4 92 BUFFER_ATOMIC_INC_X2

31 BUFFER_STORE_DWORDX3 93 BUFFER_ATOMIC_DEC_X2

32 BUFFER_LOAD_UBYTE_D16 94 BUFFER_ATOMIC_FCMPSWAP_X2

33 BUFFER_LOAD_UBYTE_D16_HI 95 BUFFER_ATOMIC_FMIN_X2

34 BUFFER_LOAD_SBYTE_D16 96 BUFFER_ATOMIC_FMAX_X2

35 BUFFER_LOAD_SBYTE_D16_HI 113 BUFFER_GL0_INV

36 BUFFER_LOAD_SHORT_D16 114 BUFFER_GL1_INV

37 BUFFER_LOAD_SHORT_D16_HI 128 BUFFER_LOAD_FORMAT_D16_X

38 BUFFER_LOAD_FORMAT_D16_HI_X 129 BUFFER_LOAD_FORMAT_D16_XY

39 BUFFER_STORE_FORMAT_D16_HI_X 130 BUFFER_LOAD_FORMAT_D16_XYZ

48 BUFFER_ATOMIC_SWAP 131 BUFFER_LOAD_FORMAT_D16_XYZW

49 BUFFER_ATOMIC_CMPSWAP 132 BUFFER_STORE_FORMAT_D16_X

50 BUFFER_ATOMIC_ADD 133 BUFFER_STORE_FORMAT_D16_XY

51 BUFFER_ATOMIC_SUB 134 BUFFER_STORE_FORMAT_D16_XYZ

52 BUFFER_ATOMIC_CSUB 135 BUFFER_STORE_FORMAT_D16_XYZW

53 BUFFER_ATOMIC_SMIN

13.6. Vector Memory Buffer Formats 275 of 283

"RDNA 2" Instruction Set Architecture

13.7. Vector Memory Image Format

13.7.1. MIMG

Format MIMG

Description Memory Image Instructions

Memory Image instructions (MIMG format) can be betwen 2 and 5 dwords. There are two
variations of the instruction:

• Normal, where the address VGPRs are specified in the "ADDR" field, and are a contiguous
set of VGPRs. This is a 2-dword instruction.
• Non-Sequential-Address (NSA), where each address VGPR is specified individually and the
address VGPRs can be scattered. This version uses 1-3 extra dwords to specify the
individual address VGPRs.

Table 100. MIMG Fields

Field Name Bits Format or Description

NSA [2:1] Non-sequential address. Specifies how many additional instruction dwords
exist (0-3).

DIM [5:3] Dimensionality of the resource constant. Set to bits [3:1] of the resource type
field.

DLC [7] 0 = normal, 1 = Device Coherent

DMASK [11:8] Data VGPR enable mask: 1 .. 4 consecutive VGPRs

Reads: defines which components are returned:
0=red,1=green,2=blue,3=alpha
Writes: defines which components are written with data from VGPRs (missing
components get 0).
Enabled components come from consecutive VGPRs.
E.G. dmask=1001 : Red is in VGPRn and alpha in VGPRn+1.
For D16 writes, DMASK is only used as a word count: each bit represents 16
bits of data to be written starting at the LSB’s of VDATA, then MSBs, then
VDATA+1 etc. Bit position is ignored.

UNRM [12] Force address to be un-normalized. User must set to 1 for Image stores &
atomics.

13.7. Vector Memory Image Format 276 of 283

"RDNA 2" Instruction Set Architecture

Field Name Bits Format or Description

GLC [13] 0 = normal, 1 = globally coherent (bypass L0 cache) or for atomics, return pre-
op value to VGPR.

R128 [15] Resource constant size: 1 = 128bit, 0 = 256bit

TFE [16] Partially resident texture, texture fail enable.

LWE [17] LOD Warning Enable. When set to 1, a texture fetch may return
"LOD_CLAMPED = 1".

OP [0],[24:18] Opcode. See table below. (combine bits zero and 18-24 to form opcode).

SLC [25] System Level Coherent. Used in conjunction with DLC to determine L2 cache
policies.

ENCODING [31:26] 111100

VADDR [39:32] Address of VGPR to supply first component of address (offset or index). When
both index and offset are used, index is in the first VGPR and offset in the
second.

VDATA [47:40] Address of VGPR to supply first component of write data or receive first
component of read-data.

SRSRC [52:48] SGPR to supply V# (resource constant) in 4 or 8 consecutive SGPRs. It is

missing 2 LSB’s of SGPR-address since it is aligned to 4 SGPRs.

SSAMP [57:53] SGPR to supply V# (resource constant) in 4 or 8 consecutive SGPRs. It is

missing 2 LSB’s of SGPR-address since it is aligned to 4 SGPRs.

A16 [62] Address components are 16-bits (instead of the usual 32 bits).
When set, all address components are 16 bits (packed into 2 per dword),
except:
Texel offsets (3 6bit UINT packed into 1 dword)
PCF reference (for "_C" instructions)
Address components are 16b uint for image ops without sampler; 16b float with
sampler.

D16 [63] Address offset, unsigned byte.

Table 101. MIMG Opcodes

Opcode # Name Opcode # Name

0 IMAGE_LOAD 53 IMAGE_SAMPLE_B_O

1 IMAGE_LOAD_MIP 54 IMAGE_SAMPLE_B_CL_O

2 IMAGE_LOAD_PCK 55 IMAGE_SAMPLE_LZ_O

3 IMAGE_LOAD_PCK_SGN 56 IMAGE_SAMPLE_C_O

4 IMAGE_LOAD_MIP_PCK 57 IMAGE_SAMPLE_C_CL_O

5 IMAGE_LOAD_MIP_PCK_SGN 58 IMAGE_SAMPLE_C_D_O

8 IMAGE_STORE 59 IMAGE_SAMPLE_C_D_CL_O

9 IMAGE_STORE_MIP 60 IMAGE_SAMPLE_C_L_O

10 IMAGE_STORE_PCK 61 IMAGE_SAMPLE_C_B_O

13.7. Vector Memory Image Format 277 of 283

"RDNA 2" Instruction Set Architecture

Opcode # Name Opcode # Name

11 IMAGE_STORE_MIP_PCK 62 IMAGE_SAMPLE_C_B_CL_O

14 IMAGE_GET_RESINFO 63 IMAGE_SAMPLE_C_LZ_O

15 IMAGE_ATOMIC_SWAP 64 IMAGE_GATHER4

16 IMAGE_ATOMIC_CMPSWAP 65 IMAGE_GATHER4_CL

17 IMAGE_ATOMIC_ADD 68 IMAGE_GATHER4_L

18 IMAGE_ATOMIC_SUB 69 IMAGE_GATHER4_B

20 IMAGE_ATOMIC_SMIN 70 IMAGE_GATHER4_B_CL

21 IMAGE_ATOMIC_UMIN 71 IMAGE_GATHER4_LZ

22 IMAGE_ATOMIC_SMAX 72 IMAGE_GATHER4_C

23 IMAGE_ATOMIC_UMAX 73 IMAGE_GATHER4_C_CL

24 IMAGE_ATOMIC_AND 76 IMAGE_GATHER4_C_L

25 IMAGE_ATOMIC_OR 77 IMAGE_GATHER4_C_B

26 IMAGE_ATOMIC_XOR 78 IMAGE_GATHER4_C_B_CL

27 IMAGE_ATOMIC_INC 79 IMAGE_GATHER4_C_LZ

28 IMAGE_ATOMIC_DEC 80 IMAGE_GATHER4_O

29 IMAGE_ATOMIC_FCMPSWAP 81 IMAGE_GATHER4_CL_O

30 IMAGE_ATOMIC_FMIN 84 IMAGE_GATHER4_L_O

31 IMAGE_ATOMIC_FMAX 85 IMAGE_GATHER4_B_O

32 IMAGE_SAMPLE 86 IMAGE_GATHER4_B_CL_O

33 IMAGE_SAMPLE_CL 87 IMAGE_GATHER4_LZ_O

34 IMAGE_SAMPLE_D 88 IMAGE_GATHER4_C_O

35 IMAGE_SAMPLE_D_CL 89 IMAGE_GATHER4_C_CL_O

36 IMAGE_SAMPLE_L 92 IMAGE_GATHER4_C_L_O

37 IMAGE_SAMPLE_B 93 IMAGE_GATHER4_C_B_O

38 IMAGE_SAMPLE_B_CL 94 IMAGE_GATHER4_C_B_CL_O

39 IMAGE_SAMPLE_LZ 95 IMAGE_GATHER4_C_LZ_O

40 IMAGE_SAMPLE_C 96 IMAGE_GET_LOD

41 IMAGE_SAMPLE_C_CL 97 IMAGE_GATHER4H

42 IMAGE_SAMPLE_C_D 128 IMAGE_MSAA_LOAD

43 IMAGE_SAMPLE_C_D_CL 162 IMAGE_SAMPLE_D_G16

44 IMAGE_SAMPLE_C_L 163 IMAGE_SAMPLE_D_CL_G16

45 IMAGE_SAMPLE_C_B 170 IMAGE_SAMPLE_C_D_G16

46 IMAGE_SAMPLE_C_B_CL 171 IMAGE_SAMPLE_C_D_CL_G16

47 IMAGE_SAMPLE_C_LZ 178 IMAGE_SAMPLE_D_O_G16

13.7. Vector Memory Image Format 278 of 283

"RDNA 2" Instruction Set Architecture

Opcode # Name Opcode # Name

48 IMAGE_SAMPLE_O 179 IMAGE_SAMPLE_D_CL_O_G16

49 IMAGE_SAMPLE_CL_O 186 IMAGE_SAMPLE_C_D_O_G16

50 IMAGE_SAMPLE_D_O 187 IMAGE_SAMPLE_C_D_CL_O_G16

51 IMAGE_SAMPLE_D_CL_O 230 IMAGE_BVH_INTERSECT_RAY

52 IMAGE_SAMPLE_L_O 231 IMAGE_BVH64_INTERSECT_RAY

13.8. Flat Formats

Flat memory instruction come in three versions: FLAT:: memory address (per work-item) may be
in global memory, scratch (private) memory or shared memory (LDS) GLOBAL:: same as FLAT,
but assumes all memory addresses are global memory. SCRATCH:: same as FLAT, but
assumes all memory addresses are scratch (private) memory.

The microcode format is identical for each, and only the value of the SEG (segment) field differs.

13.8.1. FLAT

Format FLAT

Description FLAT Memory Access

Table 102. FLAT Fields

Field Name Bits Format or Description

OFFSET [11:0] Address offset

Scratch, Global: 12-bit signed byte offset
FLAT: 11-bit unsigned offset (MSB is ignored)

DLC [12] 0 = normal, 1 = Device Coherent

LDS [13] 0 = normal, 1 = transfer data between LDS and memory instead of VGPRs and
memory.

SEG [15:14] Memory Segment (instruction type): 0 = flat, 1 = scratch, 2 = global.

GLC [16] 0 = normal, 1 = globally coherent (bypass L0 cache) or for atomics, return pre-
op value to VGPR.

SLC [17] System Level Coherent. Used in conjunction with DLC to determine L2 cache
policies.

13.8. Flat Formats 279 of 283

"RDNA 2" Instruction Set Architecture

Field Name Bits Format or Description

OP [24:18] Opcode. See tables below for FLAT, SCRATCH and GLOBAL opcodes.

ENCODING [31:26] 110111

ADDR [39:32] VGPR which holds address or offset. For 64-bit addresses, ADDR has the
LSB’s and ADDR+1 has the MSBs. For offset a single VGPR has a 32 bit
unsigned offset.
For FLAT_*: specifies an address.
For GLOBAL_* and SCRATCH_* when SADDR is NULL or 0x7f: specifies an
address.
For GLOBAL_* and SCRATCH_* when SADDR is not NULL or0x7f: specifies
an offset.

DATA [47:40] VGPR which supplies data.

SADDR [54:48] Scalar SGPR which provides an address of offset (unsigned). Set this field to
NULL or 0x7f to disable use.
Meaning of this field is different for Scratch and Global:
FLAT: Unused
Scratch: use an SGPR for the address instead of a VGPR
Global: use the SGPR to provide a base address and the VGPR provides a 32-
bit byte offset.

VDST [63:56] Destination VGPR for data returned from memory to VGPRs.

Table 103. FLAT Opcodes

Opcode # Name Opcode # Name

8 FLAT_LOAD_UBYTE 54 FLAT_ATOMIC_UMIN

9 FLAT_LOAD_SBYTE 55 FLAT_ATOMIC_SMAX

10 FLAT_LOAD_USHORT 56 FLAT_ATOMIC_UMAX

11 FLAT_LOAD_SSHORT 57 FLAT_ATOMIC_AND

12 FLAT_LOAD_DWORD 58 FLAT_ATOMIC_OR

13 FLAT_LOAD_DWORDX2 59 FLAT_ATOMIC_XOR

14 FLAT_LOAD_DWORDX4 60 FLAT_ATOMIC_INC

15 FLAT_LOAD_DWORDX3 61 FLAT_ATOMIC_DEC

24 FLAT_STORE_BYTE 62 FLAT_ATOMIC_FCMPSWAP

25 FLAT_STORE_BYTE_D16_HI 63 FLAT_ATOMIC_FMIN

26 FLAT_STORE_SHORT 64 FLAT_ATOMIC_FMAX

27 FLAT_STORE_SHORT_D16_HI 80 FLAT_ATOMIC_SWAP_X2

28 FLAT_STORE_DWORD 81 FLAT_ATOMIC_CMPSWAP_X2

29 FLAT_STORE_DWORDX2 82 FLAT_ATOMIC_ADD_X2

30 FLAT_STORE_DWORDX4 83 FLAT_ATOMIC_SUB_X2

31 FLAT_STORE_DWORDX3 85 FLAT_ATOMIC_SMIN_X2

32 FLAT_LOAD_UBYTE_D16 86 FLAT_ATOMIC_UMIN_X2

13.8. Flat Formats 280 of 283

"RDNA 2" Instruction Set Architecture

Opcode # Name Opcode # Name

33 FLAT_LOAD_UBYTE_D16_HI 87 FLAT_ATOMIC_SMAX_X2

34 FLAT_LOAD_SBYTE_D16 88 FLAT_ATOMIC_UMAX_X2

35 FLAT_LOAD_SBYTE_D16_HI 89 FLAT_ATOMIC_AND_X2

36 FLAT_LOAD_SHORT_D16 90 FLAT_ATOMIC_OR_X2

37 FLAT_LOAD_SHORT_D16_HI 91 FLAT_ATOMIC_XOR_X2

48 FLAT_ATOMIC_SWAP 92 FLAT_ATOMIC_INC_X2

49 FLAT_ATOMIC_CMPSWAP 93 FLAT_ATOMIC_DEC_X2

50 FLAT_ATOMIC_ADD 94 FLAT_ATOMIC_FCMPSWAP_X2

51 FLAT_ATOMIC_SUB 95 FLAT_ATOMIC_FMIN_X2

53 FLAT_ATOMIC_SMIN 96 FLAT_ATOMIC_FMAX_X2

13.8.2. GLOBAL
Table 104. GLOBAL Opcodes
Opcode # Name Opcode # Name

8 GLOBAL_LOAD_UBYTE 53 GLOBAL_ATOMIC_SMIN

9 GLOBAL_LOAD_SBYTE 54 GLOBAL_ATOMIC_UMIN

10 GLOBAL_LOAD_USHORT 55 GLOBAL_ATOMIC_SMAX

11 GLOBAL_LOAD_SSHORT 56 GLOBAL_ATOMIC_UMAX

12 GLOBAL_LOAD_DWORD 57 GLOBAL_ATOMIC_AND

13 GLOBAL_LOAD_DWORDX2 58 GLOBAL_ATOMIC_OR

14 GLOBAL_LOAD_DWORDX4 59 GLOBAL_ATOMIC_XOR

15 GLOBAL_LOAD_DWORDX3 60 GLOBAL_ATOMIC_INC

22 GLOBAL_LOAD_DWORD_ADDTID 61 GLOBAL_ATOMIC_DEC

23 GLOBAL_STORE_DWORD_ADDTID 62 GLOBAL_ATOMIC_FCMPSWAP

24 GLOBAL_STORE_BYTE 63 GLOBAL_ATOMIC_FMIN

25 GLOBAL_STORE_BYTE_D16_HI 64 GLOBAL_ATOMIC_FMAX

26 GLOBAL_STORE_SHORT 80 GLOBAL_ATOMIC_SWAP_X2

27 GLOBAL_STORE_SHORT_D16_HI 81 GLOBAL_ATOMIC_CMPSWAP_X2

28 GLOBAL_STORE_DWORD 82 GLOBAL_ATOMIC_ADD_X2

29 GLOBAL_STORE_DWORDX2 83 GLOBAL_ATOMIC_SUB_X2

30 GLOBAL_STORE_DWORDX4 85 GLOBAL_ATOMIC_SMIN_X2

31 GLOBAL_STORE_DWORDX3 86 GLOBAL_ATOMIC_UMIN_X2

32 GLOBAL_LOAD_UBYTE_D16 87 GLOBAL_ATOMIC_SMAX_X2

13.8. Flat Formats 281 of 283

"RDNA 2" Instruction Set Architecture

Opcode # Name Opcode # Name

33 GLOBAL_LOAD_UBYTE_D16_HI 88 GLOBAL_ATOMIC_UMAX_X2

34 GLOBAL_LOAD_SBYTE_D16 89 GLOBAL_ATOMIC_AND_X2

35 GLOBAL_LOAD_SBYTE_D16_HI 90 GLOBAL_ATOMIC_OR_X2

36 GLOBAL_LOAD_SHORT_D16 91 GLOBAL_ATOMIC_XOR_X2

37 GLOBAL_LOAD_SHORT_D16_HI 92 GLOBAL_ATOMIC_INC_X2

48 GLOBAL_ATOMIC_SWAP 93 GLOBAL_ATOMIC_DEC_X2

49 GLOBAL_ATOMIC_CMPSWAP 94 GLOBAL_ATOMIC_FCMPSWAP_X2

50 GLOBAL_ATOMIC_ADD 95 GLOBAL_ATOMIC_FMIN_X2

51 GLOBAL_ATOMIC_SUB 96 GLOBAL_ATOMIC_FMAX_X2

52 GLOBAL_ATOMIC_CSUB

13.8.3. SCRATCH
Table 105. SCRATCH Opcodes
Opcode # Name Opcode # Name

8 SCRATCH_LOAD_UBYTE 27 SCRATCH_STORE_SHORT_D16_HI

9 SCRATCH_LOAD_SBYTE 28 SCRATCH_STORE_DWORD

10 SCRATCH_LOAD_USHORT 29 SCRATCH_STORE_DWORDX2

11 SCRATCH_LOAD_SSHORT 30 SCRATCH_STORE_DWORDX4

12 SCRATCH_LOAD_DWORD 31 SCRATCH_STORE_DWORDX3

13 SCRATCH_LOAD_DWORDX2 32 SCRATCH_LOAD_UBYTE_D16

14 SCRATCH_LOAD_DWORDX4 33 SCRATCH_LOAD_UBYTE_D16_HI

15 SCRATCH_LOAD_DWORDX3 34 SCRATCH_LOAD_SBYTE_D16

24 SCRATCH_STORE_BYTE 35 SCRATCH_LOAD_SBYTE_D16_HI

25 SCRATCH_STORE_BYTE_D16_HI 36 SCRATCH_LOAD_SHORT_D16

26 SCRATCH_STORE_SHORT 37 SCRATCH_LOAD_SHORT_D16_HI

13.9. Export Format

13.9.1. EXP

13.9. Export Format 282 of 283

"RDNA 2" Instruction Set Architecture

Format EXP

Description EXPORT instructions

The export format has only a single opcode, "EXPORT".

Table 106. EXP Fields

Field Name Bits Format or Description

EN [3:0] COMPR==1: export half-dword enable. Valid values are: 0x0,3,c,f

[0] enables VSRC0 : R,G from one VGPR (R in low bits, G high)
[2] enables VSRC1 : B,A from one VGPR (B in low bits, A high)
COMPR==0: [0-3] = enables for VSRC0..3.
EN may be zero only for "NULL Pixel Shader" exports (used when exporting
only valid mask to NULL target).

TARGET [9:4] Export destination:

0-7: MRT 0..7
8: Z
9: Null pixel shader export (no data)
12-15: Position 0..3
20: Primitive data
32-63: Parameter 0..31

COMPR [10] Indicates that data is float-16/short/byte (compressed). Data is written to

consecutive components (rgba or xyzw).

DONE [11] Indicates that this is the last export from the shader. Used only for Position and
Pixel/color data.

VM [12] 1 = the exec mask IS the valid mask for this export. Can be sent multiple times,
but user must send at least once per pixel shader. This bit is only used for Pixel
Shaders.

ENCODING [31:26] 111110

VSRC0 [39:32] VGPR for source 0.

VSRC1 [47:40] VGPR for source 1.

VSRC2 [55:48] VGPR for source 2.

VSRC3 [63:56] VGPR for source 3.

13.9. Export Format 283 of 283