micropython/py/persistentcode.c

Ignoring revisions in .git-blame-ignore-revs. Click here to bypass and see the normal blame view.

662 lines
22 KiB
C
Raw Normal View History

/*
* This file is part of the MicroPython project, http://micropython.org/
*
* The MIT License (MIT)
*
* Copyright (c) 2013-2020 Damien P. George
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to deal
* in the Software without restriction, including without limitation the rights
* to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
* copies of the Software, and to permit persons to whom the Software is
* furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in
* all copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
* THE SOFTWARE.
*/
#include <stdint.h>
#include <stdio.h>
#include <string.h>
#include <assert.h>
#include "py/reader.h"
#include "py/nativeglue.h"
#include "py/persistentcode.h"
#include "py/bc0.h"
#include "py/objstr.h"
#include "py/mpthread.h"
#if MICROPY_PERSISTENT_CODE_LOAD || MICROPY_PERSISTENT_CODE_SAVE
#include "py/smallint.h"
#define QSTR_LAST_STATIC MP_QSTR_zip
#if MICROPY_DYNAMIC_COMPILER
#define MPY_FEATURE_ARCH_DYNAMIC mp_dynamic_compiler.native_arch
#else
#define MPY_FEATURE_ARCH_DYNAMIC MPY_FEATURE_ARCH
#endif
typedef struct _bytecode_prelude_t {
uint n_state;
uint n_exc_stack;
uint scope_flags;
uint n_pos_args;
uint n_kwonly_args;
uint n_def_pos_args;
uint code_info_size;
} bytecode_prelude_t;
#endif // MICROPY_PERSISTENT_CODE_LOAD || MICROPY_PERSISTENT_CODE_SAVE
#if MICROPY_PERSISTENT_CODE_LOAD
#include "py/parsenum.h"
STATIC int read_byte(mp_reader_t *reader);
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
STATIC size_t read_uint(mp_reader_t *reader);
#if MICROPY_EMIT_MACHINE_CODE
typedef struct _reloc_info_t {
mp_reader_t *reader;
mp_module_context_t *context;
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
uint8_t *rodata;
uint8_t *bss;
} reloc_info_t;
void mp_native_relocate(void *ri_in, uint8_t *text, uintptr_t reloc_text) {
// Relocate native code
reloc_info_t *ri = ri_in;
uint8_t op;
uintptr_t *addr_to_adjust = NULL;
while ((op = read_byte(ri->reader)) != 0xff) {
if (op & 1) {
// Point to new location to make adjustments
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
size_t addr = read_uint(ri->reader);
if ((addr & 1) == 0) {
// Point to somewhere in text
addr_to_adjust = &((uintptr_t *)text)[addr >> 1];
} else {
// Point to somewhere in rodata
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
addr_to_adjust = &((uintptr_t *)ri->rodata)[addr >> 1];
}
}
op >>= 1;
uintptr_t dest;
size_t n = 1;
if (op <= 5) {
if (op & 1) {
// Read in number of adjustments to make
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
n = read_uint(ri->reader);
}
op >>= 1;
if (op == 0) {
// Destination is text
dest = reloc_text;
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
} else if (op == 1) {
// Destination is rodata
dest = (uintptr_t)ri->rodata;
} else {
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
// Destination is bss
dest = (uintptr_t)ri->bss;
}
} else if (op == 6) {
// Destination is qstr_table
dest = (uintptr_t)ri->context->constants.qstr_table;
} else if (op == 7) {
// Destination is obj_table
dest = (uintptr_t)ri->context->constants.obj_table;
} else if (op == 8) {
// Destination is mp_fun_table itself
dest = (uintptr_t)&mp_fun_table;
} else {
// Destination is an entry in mp_fun_table
dest = ((uintptr_t *)&mp_fun_table)[op - 9];
}
while (n--) {
*addr_to_adjust++ += dest;
}
}
}
#endif
STATIC int read_byte(mp_reader_t *reader) {
return reader->readbyte(reader->data);
}
STATIC void read_bytes(mp_reader_t *reader, byte *buf, size_t len) {
while (len-- > 0) {
*buf++ = reader->readbyte(reader->data);
}
}
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
STATIC size_t read_uint(mp_reader_t *reader) {
size_t unum = 0;
for (;;) {
byte b = reader->readbyte(reader->data);
unum = (unum << 7) | (b & 0x7f);
if ((b & 0x80) == 0) {
break;
}
}
return unum;
}
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
STATIC qstr load_qstr(mp_reader_t *reader) {
size_t len = read_uint(reader);
if (len & 1) {
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
// static qstr
return len >> 1;
}
len >>= 1;
char *str = m_new(char, len);
read_bytes(reader, (byte *)str, len);
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
read_byte(reader); // read and discard null terminator
qstr qst = qstr_from_strn(str, len);
m_del(char, str, len);
return qst;
}
STATIC mp_obj_t load_obj(mp_reader_t *reader) {
byte obj_type = read_byte(reader);
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
#if MICROPY_EMIT_MACHINE_CODE
if (obj_type == MP_PERSISTENT_OBJ_FUN_TABLE) {
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
return MP_OBJ_FROM_PTR(&mp_fun_table);
} else
#endif
if (obj_type == MP_PERSISTENT_OBJ_NONE) {
return mp_const_none;
} else if (obj_type == MP_PERSISTENT_OBJ_FALSE) {
return mp_const_false;
} else if (obj_type == MP_PERSISTENT_OBJ_TRUE) {
return mp_const_true;
} else if (obj_type == MP_PERSISTENT_OBJ_ELLIPSIS) {
return MP_OBJ_FROM_PTR(&mp_const_ellipsis_obj);
} else {
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
size_t len = read_uint(reader);
if (len == 0 && obj_type == MP_PERSISTENT_OBJ_BYTES) {
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
read_byte(reader); // skip null terminator
return mp_const_empty_bytes;
} else if (obj_type == MP_PERSISTENT_OBJ_TUPLE) {
mp_obj_tuple_t *tuple = MP_OBJ_TO_PTR(mp_obj_new_tuple(len, NULL));
for (size_t i = 0; i < len; ++i) {
tuple->items[i] = load_obj(reader);
}
return MP_OBJ_FROM_PTR(tuple);
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
}
vstr_t vstr;
vstr_init_len(&vstr, len);
read_bytes(reader, (byte *)vstr.buf, len);
if (obj_type == MP_PERSISTENT_OBJ_STR || obj_type == MP_PERSISTENT_OBJ_BYTES) {
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
read_byte(reader); // skip null terminator
if (obj_type == MP_PERSISTENT_OBJ_STR) {
return mp_obj_new_str_from_utf8_vstr(&vstr);
} else {
return mp_obj_new_bytes_from_vstr(&vstr);
}
} else if (obj_type == MP_PERSISTENT_OBJ_INT) {
return mp_parse_num_integer(vstr.buf, vstr.len, 10, NULL);
} else {
assert(obj_type == MP_PERSISTENT_OBJ_FLOAT || obj_type == MP_PERSISTENT_OBJ_COMPLEX);
return mp_parse_num_float(vstr.buf, vstr.len, obj_type == MP_PERSISTENT_OBJ_COMPLEX, NULL);
}
}
}
STATIC mp_raw_code_t *load_raw_code(mp_reader_t *reader, mp_module_context_t *context) {
// Load function kind and data length
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
size_t kind_len = read_uint(reader);
int kind = (kind_len & 3) + MP_CODE_BYTECODE;
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
bool has_children = !!(kind_len & 4);
size_t fun_data_len = kind_len >> 3;
#if !MICROPY_EMIT_MACHINE_CODE
if (kind != MP_CODE_BYTECODE) {
mp_raise_ValueError(MP_ERROR_TEXT("incompatible .mpy file"));
}
#endif
uint8_t *fun_data = NULL;
#if MICROPY_EMIT_MACHINE_CODE
size_t prelude_offset = 0;
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
mp_uint_t native_scope_flags = 0;
mp_uint_t native_n_pos_args = 0;
mp_uint_t native_type_sig = 0;
#endif
if (kind == MP_CODE_BYTECODE) {
// Allocate memory for the bytecode
fun_data = m_new(uint8_t, fun_data_len);
// Load bytecode
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
read_bytes(reader, fun_data, fun_data_len);
#if MICROPY_EMIT_MACHINE_CODE
} else {
// Allocate memory for native data and load it
size_t fun_alloc;
MP_PLAT_ALLOC_EXEC(fun_data_len, (void **)&fun_data, &fun_alloc);
read_bytes(reader, fun_data, fun_data_len);
if (kind == MP_CODE_NATIVE_PY) {
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
// Read prelude offset within fun_data, and extract scope flags.
prelude_offset = read_uint(reader);
const byte *ip = fun_data + prelude_offset;
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
MP_BC_PRELUDE_SIG_DECODE(ip);
native_scope_flags = scope_flags;
} else {
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
// Load basic scope info for viper and asm.
native_scope_flags = read_uint(reader);
if (kind == MP_CODE_NATIVE_ASM) {
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
native_n_pos_args = read_uint(reader);
native_type_sig = read_uint(reader);
}
}
#endif
}
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
size_t n_children = 0;
mp_raw_code_t **children = NULL;
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
#if MICROPY_EMIT_MACHINE_CODE
// Load optional BSS/rodata for viper.
uint8_t *rodata = NULL;
uint8_t *bss = NULL;
if (kind == MP_CODE_NATIVE_VIPER) {
size_t rodata_size = 0;
if (native_scope_flags & MP_SCOPE_FLAG_VIPERRODATA) {
rodata_size = read_uint(reader);
}
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
size_t bss_size = 0;
if (native_scope_flags & MP_SCOPE_FLAG_VIPERBSS) {
bss_size = read_uint(reader);
}
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
if (rodata_size + bss_size != 0) {
bss_size = (uintptr_t)MP_ALIGN(bss_size, sizeof(uintptr_t));
uint8_t *data = m_new0(uint8_t, bss_size + rodata_size);
bss = data;
rodata = bss + bss_size;
if (native_scope_flags & MP_SCOPE_FLAG_VIPERRODATA) {
read_bytes(reader, rodata, rodata_size);
}
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
// Viper code with BSS/rodata should not have any children.
// Reuse the children pointer to reference the BSS/rodata
// memory so that it is not reclaimed by the GC.
assert(!has_children);
children = (void *)data;
}
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
}
#endif
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
// Load children if any.
if (has_children) {
n_children = read_uint(reader);
py/emitnative: Put a pointer to the native prelude in child_table array. Some architectures (like esp32 xtensa) cannot read byte-wise from executable memory. This means the prelude for native functions -- which is usually located after the machine code for the native function -- must be placed in separate memory that can be read byte-wise. Prior to this commit this was achieved by enabling N_PRELUDE_AS_BYTES_OBJ for the emitter and MICROPY_EMIT_NATIVE_PRELUDE_AS_BYTES_OBJ for the runtime. The prelude was then placed in a bytes object, pointed to by the module's constant table. This behaviour is changed by this commit so that a pointer to the prelude is stored either in mp_obj_fun_bc_t.child_table, or in mp_obj_fun_bc_t.child_table[num_children] if num_children > 0. The reasons for doing this are: 1. It decouples the native emitter from runtime requirements, the emitted code no longer needs to know if the system it runs on can/can't read byte-wise from executable memory. 2. It makes all ports have the same emitter behaviour, there is no longer the N_PRELUDE_AS_BYTES_OBJ option. 3. The module's constant table is now used only for actual constants in the Python code. This allows further optimisations to be done with the constants (eg constant deduplication). Code size change for those ports that enable the native emitter: unix x64: +80 +0.015% stm32: +24 +0.004% PYBV10 esp8266: +88 +0.013% GENERIC esp32: -20 -0.002% GENERIC[incl -112(data)] rp2: +32 +0.005% PICO Signed-off-by: Damien George <damien@micropython.org>
2022-05-10 04:56:24 +01:00
children = m_new(mp_raw_code_t *, n_children + (kind == MP_CODE_NATIVE_PY));
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
for (size_t i = 0; i < n_children; ++i) {
children[i] = load_raw_code(reader, context);
}
}
// Create raw_code and return it
mp_raw_code_t *rc = mp_emit_glue_new_raw_code();
if (kind == MP_CODE_BYTECODE) {
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
const byte *ip = fun_data;
MP_BC_PRELUDE_SIG_DECODE(ip);
// Assign bytecode to raw code object
mp_emit_glue_assign_bytecode(rc, fun_data,
#if MICROPY_PERSISTENT_CODE_SAVE || MICROPY_DEBUG_PRINTERS
fun_data_len,
#endif
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
children,
#if MICROPY_PERSISTENT_CODE_SAVE
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
n_children,
#endif
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
scope_flags);
#if MICROPY_EMIT_MACHINE_CODE
} else {
py/emitnative: Put a pointer to the native prelude in child_table array. Some architectures (like esp32 xtensa) cannot read byte-wise from executable memory. This means the prelude for native functions -- which is usually located after the machine code for the native function -- must be placed in separate memory that can be read byte-wise. Prior to this commit this was achieved by enabling N_PRELUDE_AS_BYTES_OBJ for the emitter and MICROPY_EMIT_NATIVE_PRELUDE_AS_BYTES_OBJ for the runtime. The prelude was then placed in a bytes object, pointed to by the module's constant table. This behaviour is changed by this commit so that a pointer to the prelude is stored either in mp_obj_fun_bc_t.child_table, or in mp_obj_fun_bc_t.child_table[num_children] if num_children > 0. The reasons for doing this are: 1. It decouples the native emitter from runtime requirements, the emitted code no longer needs to know if the system it runs on can/can't read byte-wise from executable memory. 2. It makes all ports have the same emitter behaviour, there is no longer the N_PRELUDE_AS_BYTES_OBJ option. 3. The module's constant table is now used only for actual constants in the Python code. This allows further optimisations to be done with the constants (eg constant deduplication). Code size change for those ports that enable the native emitter: unix x64: +80 +0.015% stm32: +24 +0.004% PYBV10 esp8266: +88 +0.013% GENERIC esp32: -20 -0.002% GENERIC[incl -112(data)] rp2: +32 +0.005% PICO Signed-off-by: Damien George <damien@micropython.org>
2022-05-10 04:56:24 +01:00
const uint8_t *prelude_ptr;
#if MICROPY_EMIT_NATIVE_PRELUDE_SEPARATE_FROM_MACHINE_CODE
if (kind == MP_CODE_NATIVE_PY) {
// Executable code cannot be accessed byte-wise on this architecture, so copy
// the prelude to a separate memory region that is byte-wise readable.
void *buf = fun_data + prelude_offset;
size_t n = fun_data_len - prelude_offset;
prelude_ptr = memcpy(m_new(uint8_t, n), buf, n);
}
#endif
// Relocate and commit code to executable address space
reloc_info_t ri = {reader, context, rodata, bss};
#if defined(MP_PLAT_COMMIT_EXEC)
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
void *opt_ri = (native_scope_flags & MP_SCOPE_FLAG_VIPERRELOC) ? &ri : NULL;
fun_data = MP_PLAT_COMMIT_EXEC(fun_data, fun_data_len, opt_ri);
#else
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
if (native_scope_flags & MP_SCOPE_FLAG_VIPERRELOC) {
#if MICROPY_PERSISTENT_CODE_TRACK_RELOC_CODE
// If native code needs relocations then it's not guaranteed that a pointer to
// the head of `buf` (containing the machine code) will be retained for the GC
// to trace. This is because native functions can start inside `buf` and so
// it's possible that the only GC-reachable pointers are pointers inside `buf`.
// So put this `buf` on a list of reachable root pointers.
if (MP_STATE_PORT(track_reloc_code_list) == MP_OBJ_NULL) {
MP_STATE_PORT(track_reloc_code_list) = mp_obj_new_list(0, NULL);
}
mp_obj_list_append(MP_STATE_PORT(track_reloc_code_list), MP_OBJ_FROM_PTR(fun_data));
#endif
// Do the relocations.
mp_native_relocate(&ri, fun_data, (uintptr_t)fun_data);
}
#endif
py/emitnative: Put a pointer to the native prelude in child_table array. Some architectures (like esp32 xtensa) cannot read byte-wise from executable memory. This means the prelude for native functions -- which is usually located after the machine code for the native function -- must be placed in separate memory that can be read byte-wise. Prior to this commit this was achieved by enabling N_PRELUDE_AS_BYTES_OBJ for the emitter and MICROPY_EMIT_NATIVE_PRELUDE_AS_BYTES_OBJ for the runtime. The prelude was then placed in a bytes object, pointed to by the module's constant table. This behaviour is changed by this commit so that a pointer to the prelude is stored either in mp_obj_fun_bc_t.child_table, or in mp_obj_fun_bc_t.child_table[num_children] if num_children > 0. The reasons for doing this are: 1. It decouples the native emitter from runtime requirements, the emitted code no longer needs to know if the system it runs on can/can't read byte-wise from executable memory. 2. It makes all ports have the same emitter behaviour, there is no longer the N_PRELUDE_AS_BYTES_OBJ option. 3. The module's constant table is now used only for actual constants in the Python code. This allows further optimisations to be done with the constants (eg constant deduplication). Code size change for those ports that enable the native emitter: unix x64: +80 +0.015% stm32: +24 +0.004% PYBV10 esp8266: +88 +0.013% GENERIC esp32: -20 -0.002% GENERIC[incl -112(data)] rp2: +32 +0.005% PICO Signed-off-by: Damien George <damien@micropython.org>
2022-05-10 04:56:24 +01:00
if (kind == MP_CODE_NATIVE_PY) {
#if !MICROPY_EMIT_NATIVE_PRELUDE_SEPARATE_FROM_MACHINE_CODE
prelude_ptr = fun_data + prelude_offset;
#endif
if (n_children == 0) {
children = (void *)prelude_ptr;
} else {
children[n_children] = (void *)prelude_ptr;
}
}
// Assign native code to raw code object
mp_emit_glue_assign_native(rc, kind,
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
fun_data, fun_data_len,
children,
#if MICROPY_PERSISTENT_CODE_SAVE
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
n_children,
prelude_offset,
#endif
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
native_scope_flags, native_n_pos_args, native_type_sig
);
#endif
}
return rc;
}
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
mp_compiled_module_t mp_raw_code_load(mp_reader_t *reader, mp_module_context_t *context) {
byte header[4];
read_bytes(reader, header, sizeof(header));
byte arch = MPY_FEATURE_DECODE_ARCH(header[2]);
if (header[0] != 'M'
|| header[1] != MPY_VERSION
|| (arch != MP_NATIVE_ARCH_NONE && MPY_FEATURE_DECODE_SUB_VERSION(header[2]) != MPY_SUB_VERSION)
|| header[3] > MP_SMALL_INT_BITS) {
mp_raise_ValueError(MP_ERROR_TEXT("incompatible .mpy file"));
}
if (MPY_FEATURE_DECODE_ARCH(header[2]) != MP_NATIVE_ARCH_NONE) {
if (!MPY_FEATURE_ARCH_TEST(arch)) {
if (MPY_FEATURE_ARCH_TEST(MP_NATIVE_ARCH_NONE)) {
// On supported ports this can be resolved by enabling feature, eg
// mpconfigboard.h: MICROPY_EMIT_THUMB (1)
mp_raise_ValueError(MP_ERROR_TEXT("native code in .mpy unsupported"));
} else {
mp_raise_ValueError(MP_ERROR_TEXT("incompatible .mpy arch"));
}
}
}
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
size_t n_qstr = read_uint(reader);
size_t n_obj = read_uint(reader);
mp_module_context_alloc_tables(context, n_qstr, n_obj);
// Load qstrs.
for (size_t i = 0; i < n_qstr; ++i) {
context->constants.qstr_table[i] = load_qstr(reader);
}
// Load constant objects.
for (size_t i = 0; i < n_obj; ++i) {
context->constants.obj_table[i] = load_obj(reader);
}
// Load top-level module.
mp_compiled_module_t cm2;
cm2.rc = load_raw_code(reader, context);
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
cm2.context = context;
#if MICROPY_PERSISTENT_CODE_SAVE
cm2.has_native = MPY_FEATURE_DECODE_ARCH(header[2]) != MP_NATIVE_ARCH_NONE;
cm2.n_qstr = n_qstr;
cm2.n_obj = n_obj;
#endif
reader->close(reader->data);
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
return cm2;
}
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
mp_compiled_module_t mp_raw_code_load_mem(const byte *buf, size_t len, mp_module_context_t *context) {
mp_reader_t reader;
mp_reader_new_mem(&reader, buf, len, 0);
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
return mp_raw_code_load(&reader, context);
}
#if MICROPY_HAS_FILE_READER
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
mp_compiled_module_t mp_raw_code_load_file(const char *filename, mp_module_context_t *context) {
mp_reader_t reader;
mp_reader_new_file(&reader, filename);
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
return mp_raw_code_load(&reader, context);
}
#endif // MICROPY_HAS_FILE_READER
#endif // MICROPY_PERSISTENT_CODE_LOAD
#if MICROPY_PERSISTENT_CODE_SAVE
#include "py/objstr.h"
STATIC void mp_print_bytes(mp_print_t *print, const byte *data, size_t len) {
print->print_strn(print->data, (const char *)data, len);
}
#define BYTES_FOR_INT ((MP_BYTES_PER_OBJ_WORD * 8 + 6) / 7)
STATIC void mp_print_uint(mp_print_t *print, size_t n) {
byte buf[BYTES_FOR_INT];
byte *p = buf + sizeof(buf);
*--p = n & 0x7f;
n >>= 7;
for (; n != 0; n >>= 7) {
*--p = 0x80 | (n & 0x7f);
}
print->print_strn(print->data, (char *)p, buf + sizeof(buf) - p);
}
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
STATIC void save_qstr(mp_print_t *print, qstr qst) {
if (qst <= QSTR_LAST_STATIC) {
// encode static qstr
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
mp_print_uint(print, qst << 1 | 1);
return;
}
size_t len;
const byte *str = qstr_data(qst, &len);
mp_print_uint(print, len << 1);
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
mp_print_bytes(print, str, len + 1); // +1 to store null terminator
}
STATIC void save_obj(mp_print_t *print, mp_obj_t o) {
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
#if MICROPY_EMIT_MACHINE_CODE
if (o == MP_OBJ_FROM_PTR(&mp_fun_table)) {
byte obj_type = MP_PERSISTENT_OBJ_FUN_TABLE;
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
mp_print_bytes(print, &obj_type, 1);
} else
#endif
if (mp_obj_is_str_or_bytes(o)) {
byte obj_type;
if (mp_obj_is_str(o)) {
obj_type = MP_PERSISTENT_OBJ_STR;
} else {
obj_type = MP_PERSISTENT_OBJ_BYTES;
}
size_t len;
const char *str = mp_obj_str_get_data(o, &len);
mp_print_bytes(print, &obj_type, 1);
mp_print_uint(print, len);
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
mp_print_bytes(print, (const byte *)str, len + 1); // +1 to store null terminator
} else if (o == mp_const_none) {
byte obj_type = MP_PERSISTENT_OBJ_NONE;
mp_print_bytes(print, &obj_type, 1);
} else if (o == mp_const_false) {
byte obj_type = MP_PERSISTENT_OBJ_FALSE;
mp_print_bytes(print, &obj_type, 1);
} else if (o == mp_const_true) {
byte obj_type = MP_PERSISTENT_OBJ_TRUE;
mp_print_bytes(print, &obj_type, 1);
} else if (MP_OBJ_TO_PTR(o) == &mp_const_ellipsis_obj) {
byte obj_type = MP_PERSISTENT_OBJ_ELLIPSIS;
mp_print_bytes(print, &obj_type, 1);
} else if (mp_obj_is_type(o, &mp_type_tuple)) {
size_t len;
mp_obj_t *items;
mp_obj_tuple_get(o, &len, &items);
byte obj_type = MP_PERSISTENT_OBJ_TUPLE;
mp_print_bytes(print, &obj_type, 1);
mp_print_uint(print, len);
for (size_t i = 0; i < len; ++i) {
save_obj(print, items[i]);
}
} else {
// we save numbers using a simplistic text representation
// TODO could be improved
byte obj_type;
if (mp_obj_is_int(o)) {
obj_type = MP_PERSISTENT_OBJ_INT;
#if MICROPY_PY_BUILTINS_COMPLEX
} else if (mp_obj_is_type(o, &mp_type_complex)) {
obj_type = MP_PERSISTENT_OBJ_COMPLEX;
#endif
} else {
assert(mp_obj_is_float(o));
obj_type = MP_PERSISTENT_OBJ_FLOAT;
}
vstr_t vstr;
mp_print_t pr;
vstr_init_print(&vstr, 10, &pr);
mp_obj_print_helper(&pr, o, PRINT_REPR);
mp_print_bytes(print, &obj_type, 1);
mp_print_uint(print, vstr.len);
mp_print_bytes(print, (const byte *)vstr.buf, vstr.len);
vstr_clear(&vstr);
}
}
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
STATIC void save_raw_code(mp_print_t *print, const mp_raw_code_t *rc) {
// Save function kind and data length
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
mp_print_uint(print, (rc->fun_data_len << 3) | ((rc->n_children != 0) << 2) | (rc->kind - MP_CODE_BYTECODE));
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
// Save function code.
mp_print_bytes(print, rc->fun_data, rc->fun_data_len);
#if MICROPY_EMIT_MACHINE_CODE
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
if (rc->kind == MP_CODE_NATIVE_PY) {
// Save prelude size
mp_print_uint(print, rc->prelude_offset);
} else if (rc->kind == MP_CODE_NATIVE_VIPER || rc->kind == MP_CODE_NATIVE_ASM) {
// Save basic scope info for viper and asm
mp_print_uint(print, rc->scope_flags & MP_SCOPE_FLAG_ALL_SIG);
if (rc->kind == MP_CODE_NATIVE_ASM) {
mp_print_uint(print, rc->n_pos_args);
mp_print_uint(print, rc->type_sig);
}
}
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
#endif
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
if (rc->n_children) {
mp_print_uint(print, rc->n_children);
for (size_t i = 0; i < rc->n_children; ++i) {
save_raw_code(print, rc->children[i]);
}
}
}
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
void mp_raw_code_save(mp_compiled_module_t *cm, mp_print_t *print) {
// header contains:
// byte 'M'
// byte version
// byte native arch (and sub-version if native)
// byte number of bits in a small int
byte header[4] = {
'M',
MPY_VERSION,
cm->has_native ? MPY_FEATURE_ENCODE_SUB_VERSION(MPY_SUB_VERSION) | MPY_FEATURE_ENCODE_ARCH(MPY_FEATURE_ARCH_DYNAMIC) : 0,
#if MICROPY_DYNAMIC_COMPILER
mp_dynamic_compiler.small_int_bits,
#else
MP_SMALL_INT_BITS,
#endif
};
mp_print_bytes(print, header, sizeof(header));
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
// Number of entries in constant table.
mp_print_uint(print, cm->n_qstr);
mp_print_uint(print, cm->n_obj);
// Save qstrs.
for (size_t i = 0; i < cm->n_qstr; ++i) {
save_qstr(print, cm->context->constants.qstr_table[i]);
}
// Save constant objects.
for (size_t i = 0; i < cm->n_obj; ++i) {
save_obj(print, (mp_obj_t)cm->context->constants.obj_table[i]);
}
// Save outer raw code, which will save all its child raw codes.
save_raw_code(print, cm->rc);
}
#if MICROPY_PERSISTENT_CODE_SAVE_FILE
#include <unistd.h>
#include <sys/stat.h>
#include <fcntl.h>
STATIC void fd_print_strn(void *env, const char *str, size_t len) {
int fd = (intptr_t)env;
MP_THREAD_GIL_EXIT();
ssize_t ret = write(fd, str, len);
MP_THREAD_GIL_ENTER();
(void)ret;
}
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
void mp_raw_code_save_file(mp_compiled_module_t *cm, const char *filename) {
MP_THREAD_GIL_EXIT();
int fd = open(filename, O_WRONLY | O_CREAT | O_TRUNC, 0644);
MP_THREAD_GIL_ENTER();
if (fd < 0) {
mp_raise_OSError_with_filename(errno, filename);
}
mp_print_t fd_print = {(void *)(intptr_t)fd, fd_print_strn};
py: Rework bytecode and .mpy file format to be mostly static data. Background: .mpy files are precompiled .py files, built using mpy-cross, that contain compiled bytecode functions (and can also contain machine code). The benefit of using an .mpy file over a .py file is that they are faster to import and take less memory when importing. They are also smaller on disk. But the real benefit of .mpy files comes when they are frozen into the firmware. This is done by loading the .mpy file during compilation of the firmware and turning it into a set of big C data structures (the job of mpy-tool.py), which are then compiled and downloaded into the ROM of a device. These C data structures can be executed in-place, ie directly from ROM. This makes importing even faster because there is very little to do, and also means such frozen modules take up much less RAM (because their bytecode stays in ROM). The downside of frozen code is that it requires recompiling and reflashing the entire firmware. This can be a big barrier to entry, slows down development time, and makes it harder to do OTA updates of frozen code (because the whole firmware must be updated). This commit attempts to solve this problem by providing a solution that sits between loading .mpy files into RAM and freezing them into the firmware. The .mpy file format has been reworked so that it consists of data and bytecode which is mostly static and ready to run in-place. If these new .mpy files are located in flash/ROM which is memory addressable, the .mpy file can be executed (mostly) in-place. With this approach there is still a small amount of unpacking and linking of the .mpy file that needs to be done when it's imported, but it's still much better than loading an .mpy from disk into RAM (although not as good as freezing .mpy files into the firmware). The main trick to make static .mpy files is to adjust the bytecode so any qstrs that it references now go through a lookup table to convert from local qstr number in the module to global qstr number in the firmware. That means the bytecode does not need linking/rewriting of qstrs when it's loaded. Instead only a small qstr table needs to be built (and put in RAM) at import time. This means the bytecode itself is static/constant and can be used directly if it's in addressable memory. Also the qstr string data in the .mpy file, and some constant object data, can be used directly. Note that the qstr table is global to the module (ie not per function). In more detail, in the VM what used to be (schematically): qst = DECODE_QSTR_VALUE; is now (schematically): idx = DECODE_QSTR_INDEX; qst = qstr_table[idx]; That allows the bytecode to be fixed at compile time and not need relinking/rewriting of the qstr values. Only qstr_table needs to be linked when the .mpy is loaded. Incidentally, this helps to reduce the size of bytecode because what used to be 2-byte qstr values in the bytecode are now (mostly) 1-byte indices. If the module uses the same qstr more than two times then the bytecode is smaller than before. The following changes are measured for this commit compared to the previous (the baseline): - average 7%-9% reduction in size of .mpy files - frozen code size is reduced by about 5%-7% - importing .py files uses about 5% less RAM in total - importing .mpy files uses about 4% less RAM in total - importing .py and .mpy files takes about the same time as before The qstr indirection in the bytecode has only a small impact on VM performance. For stm32 on PYBv1.0 the performance change of this commit is: diff of scores (higher is better) N=100 M=100 baseline -> this-commit diff diff% (error%) bm_chaos.py 371.07 -> 357.39 : -13.68 = -3.687% (+/-0.02%) bm_fannkuch.py 78.72 -> 77.49 : -1.23 = -1.563% (+/-0.01%) bm_fft.py 2591.73 -> 2539.28 : -52.45 = -2.024% (+/-0.00%) bm_float.py 6034.93 -> 5908.30 : -126.63 = -2.098% (+/-0.01%) bm_hexiom.py 48.96 -> 47.93 : -1.03 = -2.104% (+/-0.00%) bm_nqueens.py 4510.63 -> 4459.94 : -50.69 = -1.124% (+/-0.00%) bm_pidigits.py 650.28 -> 644.96 : -5.32 = -0.818% (+/-0.23%) core_import_mpy_multi.py 564.77 -> 581.49 : +16.72 = +2.960% (+/-0.01%) core_import_mpy_single.py 68.67 -> 67.16 : -1.51 = -2.199% (+/-0.01%) core_qstr.py 64.16 -> 64.12 : -0.04 = -0.062% (+/-0.00%) core_yield_from.py 362.58 -> 354.50 : -8.08 = -2.228% (+/-0.00%) misc_aes.py 429.69 -> 405.59 : -24.10 = -5.609% (+/-0.01%) misc_mandel.py 3485.13 -> 3416.51 : -68.62 = -1.969% (+/-0.00%) misc_pystone.py 2496.53 -> 2405.56 : -90.97 = -3.644% (+/-0.01%) misc_raytrace.py 381.47 -> 374.01 : -7.46 = -1.956% (+/-0.01%) viper_call0.py 576.73 -> 572.49 : -4.24 = -0.735% (+/-0.04%) viper_call1a.py 550.37 -> 546.21 : -4.16 = -0.756% (+/-0.09%) viper_call1b.py 438.23 -> 435.68 : -2.55 = -0.582% (+/-0.06%) viper_call1c.py 442.84 -> 440.04 : -2.80 = -0.632% (+/-0.08%) viper_call2a.py 536.31 -> 532.35 : -3.96 = -0.738% (+/-0.06%) viper_call2b.py 382.34 -> 377.07 : -5.27 = -1.378% (+/-0.03%) And for unix on x64: diff of scores (higher is better) N=2000 M=2000 baseline -> this-commit diff diff% (error%) bm_chaos.py 13594.20 -> 13073.84 : -520.36 = -3.828% (+/-5.44%) bm_fannkuch.py 60.63 -> 59.58 : -1.05 = -1.732% (+/-3.01%) bm_fft.py 112009.15 -> 111603.32 : -405.83 = -0.362% (+/-4.03%) bm_float.py 246202.55 -> 247923.81 : +1721.26 = +0.699% (+/-2.79%) bm_hexiom.py 615.65 -> 617.21 : +1.56 = +0.253% (+/-1.64%) bm_nqueens.py 215807.95 -> 215600.96 : -206.99 = -0.096% (+/-3.52%) bm_pidigits.py 8246.74 -> 8422.82 : +176.08 = +2.135% (+/-3.64%) misc_aes.py 16133.00 -> 16452.74 : +319.74 = +1.982% (+/-1.50%) misc_mandel.py 128146.69 -> 130796.43 : +2649.74 = +2.068% (+/-3.18%) misc_pystone.py 83811.49 -> 83124.85 : -686.64 = -0.819% (+/-1.03%) misc_raytrace.py 21688.02 -> 21385.10 : -302.92 = -1.397% (+/-3.20%) The code size change is (firmware with a lot of frozen code benefits the most): bare-arm: +396 +0.697% minimal x86: +1595 +0.979% [incl +32(data)] unix x64: +2408 +0.470% [incl +800(data)] unix nanbox: +1396 +0.309% [incl -96(data)] stm32: -1256 -0.318% PYBV10 cc3200: +288 +0.157% esp8266: -260 -0.037% GENERIC esp32: -216 -0.014% GENERIC[incl -1072(data)] nrf: +116 +0.067% pca10040 rp2: -664 -0.135% PICO samd: +844 +0.607% ADAFRUIT_ITSYBITSY_M4_EXPRESS As part of this change the .mpy file format version is bumped to version 6. And mpy-tool.py has been improved to provide a good visualisation of the contents of .mpy files. In summary: this commit changes the bytecode to use qstr indirection, and reworks the .mpy file format to be simpler and allow .mpy files to be executed in-place. Performance is not impacted too much. Eventually it will be possible to store such .mpy files in a linear, read-only, memory- mappable filesystem so they can be executed from flash/ROM. This will essentially be able to replace frozen code for most applications. Signed-off-by: Damien George <damien@micropython.org>
2021-10-22 12:22:47 +01:00
mp_raw_code_save(cm, &fd_print);
MP_THREAD_GIL_EXIT();
close(fd);
MP_THREAD_GIL_ENTER();
}
#endif // MICROPY_PERSISTENT_CODE_SAVE_FILE
#endif // MICROPY_PERSISTENT_CODE_SAVE
#if MICROPY_PERSISTENT_CODE_TRACK_RELOC_CODE
// An mp_obj_list_t that tracks relocated native code to prevent the GC from reclaiming them.
MP_REGISTER_ROOT_POINTER(mp_obj_t track_reloc_code_list);
#endif