qstr: Separate hash and len from string data. #7209

tyomitch · 2021-05-03T18:33:59Z

This allows the compiler to merge strings: e.g. "update",
"difference_update" and "symmetric_difference_update"
will all point to the same memory.

No functional change.

   bare-arm:   -12 -0.021%
minimal x86:   +38 +0.026% [incl +40(data)]
   unix x64:  -576 -0.114% [incl +32(data)]
unix nanbox:  -696 -0.158%
      stm32: -1336 -0.344% PYBV10
     cc3200:  -440 -0.240%
    esp8266: -1092 -0.159% GENERIC
      esp32: -1332 -0.093% GENERIC[incl -1408(data)]
        nrf:  -404 -0.273% pca10040
        rp2: -1024 -0.213% PICO
       samd:  -196 -0.190% ADAFRUIT_ITSYBITSY_M4_EXPRESS

dpgeorge · 2021-05-03T23:46:42Z

Wow, great find!

Need to run the performance benchmark to see if it has any impact there. Also see what the change is for RAM usage.

py/qstr.h

This allows the compiler to merge strings: e.g. "update", "difference_update" and "symmetric_difference_update" will all point to the same memory. No functional change. ``` bare-arm: -12 -0.021% minimal x86: +38 +0.026% [incl +40(data)] unix x64: -576 -0.114% [incl +32(data)] unix nanbox: -696 -0.158% stm32: -1336 -0.344% PYBV10 cc3200: -440 -0.240% esp8266: -1092 -0.159% GENERIC esp32: -1332 -0.093% GENERIC[incl -1408(data)] nrf: -404 -0.273% pca10040 rp2: -1024 -0.213% PICO samd: -196 -0.190% ADAFRUIT_ITSYBITSY_M4_EXPRESS ``` Originally at adafruit#4583 Signed-off-by: Artyom Skrobov <tyomitch@gmail.com>

Originally at adafruit#4707 Signed-off-by: Artyom Skrobov <tyomitch@gmail.com>

dpgeorge · 2021-05-11T13:44:52Z

I ran the performance benchmark on this PR, against master at 70f50c4. This is the result (perf0 is master, perf1 is this PR):

diff of scores (higher is better)
N=100 M=100                   perf0 ->      perf1         diff      diff% (error%)
bm_chaos.py                  308.47 ->     303.92 :      -4.55 =  -1.475% (+/-0.01%)
bm_fannkuch.py                77.06 ->      76.96 :      -0.10 =  -0.130% (+/-0.00%)
bm_fft.py                   2453.01 ->    2463.74 :     +10.73 =  +0.437% (+/-0.00%)
bm_float.py                 4909.52 ->    4826.24 :     -83.28 =  -1.696% (+/-0.00%)
bm_hexiom.py                  35.16 ->      35.47 :      +0.31 =  +0.882% (+/-0.00%)
bm_nqueens.py               4195.59 ->    4190.63 :      -4.96 =  -0.118% (+/-0.00%)
bm_pidigits.py               648.14 ->     647.14 :      -1.00 =  -0.154% (+/-0.29%)
misc_aes.py                  365.74 ->     367.25 :      +1.51 =  +0.413% (+/-0.00%)
misc_mandel.py              3010.53 ->    3019.65 :      +9.12 =  +0.303% (+/-0.00%)
misc_pystone.py             1938.93 ->    1943.80 :      +4.87 =  +0.251% (+/-0.01%)
misc_raytrace.py             309.49 ->     304.47 :      -5.02 =  -1.622% (+/-0.00%)
viper_call0.py               584.31 ->     584.83 :      +0.52 =  +0.089% (+/-0.11%)
viper_call1a.py              556.62 ->     557.64 :      +1.02 =  +0.183% (+/-0.13%)
viper_call1b.py              442.75 ->     442.90 :      +0.15 =  +0.034% (+/-0.02%)
viper_call1c.py              447.41 ->     447.64 :      +0.23 =  +0.051% (+/-0.03%)
viper_call2a.py              542.77 ->     543.22 :      +0.45 =  +0.083% (+/-0.09%)
viper_call2b.py              385.41 ->     385.91 :      +0.50 =  +0.130% (+/-0.10%)

That's more or less no change.

py/qstr.c

tyomitch · 2021-06-19T13:26:36Z

Ping?

dpgeorge · 2022-02-11T11:32:06Z

Against latest master, the size diff is now:

   bare-arm:    -4 -0.007% 
minimal x86:  +150 +0.092% [incl +48(data)]
   unix x64:  -608 -0.118% 
unix nanbox:  -572 -0.126% [incl +32(data)]
      stm32: -1392 -0.352% PYBV10
     cc3200:  -448 -0.244% 
    esp8266: -1208 -0.173% GENERIC
      esp32: -1028 -0.068% GENERIC[incl -1020(data)]
        nrf:  -440 -0.252% pca10040
        rp2: -1072 -0.217% PICO
       samd:  -368 -0.264% ADAFRUIT_ITSYBITSY_M4_EXPRESS

Performance change on PYBv1.1 is:

diff of scores (higher is better)
N=100 M=100                pybv10-perf0 -> pybv10-perf1         diff      diff% (error%)
bm_chaos.py                    365.21 ->     371.06 :      +5.85 =  +1.602% (+/-0.01%)
bm_fannkuch.py                  77.51 ->      78.72 :      +1.21 =  +1.561% (+/-0.00%)
bm_fft.py                     2559.69 ->    2591.71 :     +32.02 =  +1.251% (+/-0.00%)
bm_float.py                   5993.07 ->    6034.82 :     +41.75 =  +0.697% (+/-0.02%)
bm_hexiom.py                    48.01 ->      48.96 :      +0.95 =  +1.979% (+/-0.00%)
bm_nqueens.py                 4475.93 ->    4510.62 :     +34.69 =  +0.775% (+/-0.00%)
bm_pidigits.py                 647.87 ->     651.87 :      +4.00 =  +0.617% (+/-0.22%)
core_import_mpy_multi.py       516.38 ->     564.77 :     +48.39 =  +9.371% (+/-0.01%)
core_import_mpy_single.py       59.33 ->      68.67 :      +9.34 = +15.742% (+/-0.01%)
core_qstr.py                    49.99 ->      64.16 :     +14.17 = +28.346% (+/-0.00%)
core_yield_from.py             363.99 ->     362.58 :      -1.41 =  -0.387% (+/-0.00%)
misc_aes.py                    425.57 ->     429.69 :      +4.12 =  +0.968% (+/-0.01%)
misc_mandel.py                3462.57 ->    3485.14 :     +22.57 =  +0.652% (+/-0.00%)
misc_pystone.py               2465.46 ->    2496.52 :     +31.06 =  +1.260% (+/-0.01%)
misc_raytrace.py               380.58 ->     381.47 :      +0.89 =  +0.234% (+/-0.01%)
viper_call0.py                 576.49 ->     576.73 :      +0.24 =  +0.042% (+/-0.03%)
viper_call1a.py                550.03 ->     550.37 :      +0.34 =  +0.062% (+/-0.03%)
viper_call1b.py                435.40 ->     438.23 :      +2.83 =  +0.650% (+/-0.22%)
viper_call1c.py                441.73 ->     442.84 :      +1.11 =  +0.251% (+/-0.12%)
viper_call2a.py                535.37 ->     536.31 :      +0.94 =  +0.176% (+/-0.14%)
viper_call2b.py                380.97 ->     382.34 :      +1.37 =  +0.360% (+/-0.19%)

The new benchmarks that measure qstr interning now show quite good improvements. And other benchmarks are pretty much unchanged. That is good.

The esp32 port (GENERIC on TinyPICO) also shows similar performance results to PYBv1.1:

diff of scores (higher is better)
N=100 M=100                esp32-perf0 -> esp32-perf1         diff      diff% (error%)
bm_chaos.py                    301.74 ->     292.94 :      -8.80 =  -2.916% (+/-0.01%)
bm_fannkuch.py                  77.43 ->      77.45 :      +0.02 =  +0.026% (+/-0.01%)
bm_fft.py                     2340.59 ->    2381.94 :     +41.35 =  +1.767% (+/-0.00%)
bm_float.py                   5025.37 ->    4486.82 :    -538.55 = -10.717% (+/-0.18%)
bm_hexiom.py                    31.95 ->      35.90 :      +3.95 = +12.363% (+/-0.02%)
bm_nqueens.py                 2291.45 ->    2313.04 :     +21.59 =  +0.942% (+/-0.03%)
bm_pidigits.py                 662.89 ->     665.97 :      +3.08 =  +0.465% (+/-0.18%)
core_import_mpy_multi.py        98.04 ->     272.48 :    +174.44 = +177.927% (+/-0.48%)
core_import_mpy_single.py       13.00 ->      49.13 :     +36.13 = +277.923% (+/-0.21%)
core_qstr.py                    14.95 ->      76.87 :     +61.92 = +414.181% (+/-0.11%)
core_yield_from.py             109.32 ->     109.34 :      +0.02 =  +0.018% (+/-0.00%)
misc_aes.py                    345.72 ->     345.55 :      -0.17 =  -0.049% (+/-0.01%)
misc_mandel.py                2047.04 ->    2205.99 :    +158.95 =  +7.765% (+/-0.01%)
misc_pystone.py               1751.29 ->    1797.49 :     +46.20 =  +2.638% (+/-0.02%)
misc_raytrace.py               315.02 ->     319.06 :      +4.04 =  +1.282% (+/-0.01%)
viper_call0.py                 273.15 ->     273.16 :      +0.01 =  +0.004% (+/-0.00%)
viper_call1a.py                269.46 ->     269.47 :      +0.01 =  +0.004% (+/-0.00%)
viper_call1b.py                228.22 ->     228.23 :      +0.01 =  +0.004% (+/-0.00%)
viper_call1c.py                228.87 ->     228.88 :      +0.01 =  +0.004% (+/-0.00%)
viper_call2a.py                267.21 ->     267.22 :      +0.01 =  +0.004% (+/-0.00%)
viper_call2b.py                209.39 ->     209.40 :      +0.01 =  +0.005% (+/-0.00%)

The large improvements are probably because the qstr hashes are much more localised in flash and have better cache behaviour.

dpgeorge · 2022-02-11T11:58:16Z

Rebased and merged in 18b1ba0 and f46a714 (without any changes).

This improvement will also help with #8191, because the qstr pools will be able to reference string data directly in a memory-mapped .mpy file.

Thanks @tyomitch for the contribution, and for the clean commits!

Save code space by packing rgbw values into C union

dpgeorge added the py-core Relates to py/ directory in source label May 3, 2021

dpgeorge reviewed May 3, 2021

View reviewed changes

py/qstr.h Show resolved Hide resolved

tyomitch added 2 commits May 6, 2021 01:31

qstr: Use const consistently to avoid a cast.

1bfd117

Originally at adafruit#4707 Signed-off-by: Artyom Skrobov <tyomitch@gmail.com>

tyomitch force-pushed the patch-5 branch from c40360a to 1bfd117 Compare May 6, 2021 05:38

dpgeorge reviewed May 13, 2021

View reviewed changes

py/qstr.c Show resolved Hide resolved

tyomitch mentioned this pull request May 27, 2021

[qstr] Separate hash and len from string data adafruit/circuitpython#4583

Merged

dpgeorge mentioned this pull request Feb 10, 2022

tests: add performance tests for qstrs and importing .mpy files #8288

Merged

dpgeorge closed this Feb 11, 2022

tannewt pushed a commit to tannewt/circuitpython that referenced this pull request Nov 28, 2022

Merge pull request micropython#7209 from jepler/code-optimize-pixelbuf

f637332

Save code space by packing rgbw values into C union

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Uh oh!

qstr: Separate hash and len from string data. #7209

qstr: Separate hash and len from string data. #7209

Uh oh!

tyomitch commented May 3, 2021

Uh oh!

dpgeorge commented May 3, 2021

Uh oh!

Uh oh!

dpgeorge commented May 11, 2021

Uh oh!

Uh oh!

tyomitch commented Jun 19, 2021

Uh oh!

dpgeorge commented Feb 11, 2022

Uh oh!

dpgeorge commented Feb 11, 2022

Uh oh!

Uh oh!

Pfad - The Proxy pFad of © 2024 Garber Painting. All rights reserved.

Uh oh!

qstr: Separate hash and len from string data. #7209

qstr: Separate hash and len from string data. #7209

Uh oh!

Conversation

tyomitch commented May 3, 2021

Uh oh!

dpgeorge commented May 3, 2021

Uh oh!

Uh oh!

dpgeorge commented May 11, 2021

Uh oh!

Uh oh!

tyomitch commented Jun 19, 2021

Uh oh!

dpgeorge commented Feb 11, 2022

Uh oh!

dpgeorge commented Feb 11, 2022

Uh oh!

Uh oh!

Pfad - The Proxy pFad of © 2024 Garber Painting. All rights reserved.