micropython/py/makecompresseddata.py

206 lines
6.6 KiB
Python
Raw Normal View History

py: Implement "common word" compression scheme for error messages. The idea here is that there's a moderate amount of ROM used up by exception text. Obviously we try to keep the messages short, and the code can enable terse errors, but it still adds up. Listed below is the total string data size for various ports: bare-arm 2860 minimal 2876 stm32 8926 (PYBV11) cc3200 3751 esp32 5721 This commit implements compression of these strings. It takes advantage of the fact that these strings are all 7-bit ascii and extracts the top 128 frequently used words from the messages and stores them packed (dropping their null-terminator), then uses (0x80 | index) inside strings to refer to these common words. Spaces are automatically added around words, saving more bytes. This happens transparently in the build process, mirroring the steps that are used to generate the QSTR data. The MP_COMPRESSED_ROM_TEXT macro wraps any literal string that should compressed, and it's automatically decompressed in mp_decompress_rom_string. There are many schemes that could be used for the compression, and some are included in py/makecompresseddata.py for reference (space, Huffman, ngram, common word). Results showed that the common-word compression gets better results. This is before counting the increased cost of the Huffman decoder. This might be slightly counter-intuitive, but this data is extremely repetitive at a word-level, and the byte-level entropy coder can't quite exploit that as efficiently. Ideally one would combine both approaches, but for now the common-word approach is the one that is used. For additional comparison, the size of the raw data compressed with gzip and zlib is calculated, as a sort of proxy for a lower entropy bound. With this scheme we come within 15% on stm32, and 30% on bare-arm (i.e. we use x% more bytes than the data compressed with gzip -- not counting the code overhead of a decoder, and how this would be hypothetically implemented). The feature is disabled by default and can be enabled by setting MICROPY_ROM_TEXT_COMPRESSION at the Makefile-level.
2019-09-26 13:19:29 +01:00
from __future__ import print_function
import collections
import re
import sys
import gzip
import zlib
_COMPRESSED_MARKER = 0xFF
def check_non_ascii(msg):
for c in msg:
if ord(c) >= 0x80:
print(
'Unable to generate compressed data: message "{}" contains a non-ascii character "{}".'.format(
msg, c
),
file=sys.stderr,
)
sys.exit(1)
# Replace <char><space> with <char | 0x80>.
# Trival scheme to demo/test.
def space_compression(error_strings):
for line in error_strings:
check_non_ascii(line)
result = ""
for i in range(len(line)):
if i > 0 and line[i] == " ":
result = result[:-1]
result += "\\{:03o}".format(ord(line[i - 1]))
else:
result += line[i]
error_strings[line] = result
return None
# Replace common words with <0x80 | index>.
# Index is into a table of words stored as aaaaa<0x80|a>bbb<0x80|b>...
# Replaced words are assumed to have spaces either side to avoid having to store the spaces in the compressed strings.
def word_compression(error_strings):
topn = collections.Counter()
for line in error_strings.keys():
check_non_ascii(line)
for word in line.split(" "):
topn[word] += 1
# Order not just by frequency, but by expected saving. i.e. prefer a longer string that is used less frequently.
# Use the word itself for ties so that compression is deterministic.
py: Implement "common word" compression scheme for error messages. The idea here is that there's a moderate amount of ROM used up by exception text. Obviously we try to keep the messages short, and the code can enable terse errors, but it still adds up. Listed below is the total string data size for various ports: bare-arm 2860 minimal 2876 stm32 8926 (PYBV11) cc3200 3751 esp32 5721 This commit implements compression of these strings. It takes advantage of the fact that these strings are all 7-bit ascii and extracts the top 128 frequently used words from the messages and stores them packed (dropping their null-terminator), then uses (0x80 | index) inside strings to refer to these common words. Spaces are automatically added around words, saving more bytes. This happens transparently in the build process, mirroring the steps that are used to generate the QSTR data. The MP_COMPRESSED_ROM_TEXT macro wraps any literal string that should compressed, and it's automatically decompressed in mp_decompress_rom_string. There are many schemes that could be used for the compression, and some are included in py/makecompresseddata.py for reference (space, Huffman, ngram, common word). Results showed that the common-word compression gets better results. This is before counting the increased cost of the Huffman decoder. This might be slightly counter-intuitive, but this data is extremely repetitive at a word-level, and the byte-level entropy coder can't quite exploit that as efficiently. Ideally one would combine both approaches, but for now the common-word approach is the one that is used. For additional comparison, the size of the raw data compressed with gzip and zlib is calculated, as a sort of proxy for a lower entropy bound. With this scheme we come within 15% on stm32, and 30% on bare-arm (i.e. we use x% more bytes than the data compressed with gzip -- not counting the code overhead of a decoder, and how this would be hypothetically implemented). The feature is disabled by default and can be enabled by setting MICROPY_ROM_TEXT_COMPRESSION at the Makefile-level.
2019-09-26 13:19:29 +01:00
def bytes_saved(item):
w, n = item
return -((len(w) + 1) * (n - 1)), w
py: Implement "common word" compression scheme for error messages. The idea here is that there's a moderate amount of ROM used up by exception text. Obviously we try to keep the messages short, and the code can enable terse errors, but it still adds up. Listed below is the total string data size for various ports: bare-arm 2860 minimal 2876 stm32 8926 (PYBV11) cc3200 3751 esp32 5721 This commit implements compression of these strings. It takes advantage of the fact that these strings are all 7-bit ascii and extracts the top 128 frequently used words from the messages and stores them packed (dropping their null-terminator), then uses (0x80 | index) inside strings to refer to these common words. Spaces are automatically added around words, saving more bytes. This happens transparently in the build process, mirroring the steps that are used to generate the QSTR data. The MP_COMPRESSED_ROM_TEXT macro wraps any literal string that should compressed, and it's automatically decompressed in mp_decompress_rom_string. There are many schemes that could be used for the compression, and some are included in py/makecompresseddata.py for reference (space, Huffman, ngram, common word). Results showed that the common-word compression gets better results. This is before counting the increased cost of the Huffman decoder. This might be slightly counter-intuitive, but this data is extremely repetitive at a word-level, and the byte-level entropy coder can't quite exploit that as efficiently. Ideally one would combine both approaches, but for now the common-word approach is the one that is used. For additional comparison, the size of the raw data compressed with gzip and zlib is calculated, as a sort of proxy for a lower entropy bound. With this scheme we come within 15% on stm32, and 30% on bare-arm (i.e. we use x% more bytes than the data compressed with gzip -- not counting the code overhead of a decoder, and how this would be hypothetically implemented). The feature is disabled by default and can be enabled by setting MICROPY_ROM_TEXT_COMPRESSION at the Makefile-level.
2019-09-26 13:19:29 +01:00
top128 = sorted(topn.items(), key=bytes_saved)[:128]
index = [w for w, _ in top128]
index_lookup = {w: i for i, w in enumerate(index)}
for line in error_strings.keys():
result = ""
need_space = False
for word in line.split(" "):
if word in index_lookup:
result += "\\{:03o}".format(0b10000000 | index_lookup[word])
need_space = False
else:
if need_space:
result += " "
need_space = True
result += word
error_strings[line] = result.strip()
return "".join(w[:-1] + "\\{:03o}".format(0b10000000 | ord(w[-1])) for w in index)
# Replace chars in text with variable length bit sequence.
# For comparison only (the table is not emitted).
def huffman_compression(error_strings):
# https://github.com/tannewt/huffman
import huffman
all_strings = "".join(error_strings)
cb = huffman.codebook(collections.Counter(all_strings).items())
for line in error_strings:
b = "1"
for c in line:
b += cb[c]
n = len(b)
if n % 8 != 0:
n += 8 - (n % 8)
result = ""
for i in range(0, n, 8):
result += "\\{:03o}".format(int(b[i : i + 8], 2))
if len(result) > len(line) * 4:
result = line
error_strings[line] = result
# TODO: This would be the prefix lengths and the table ordering.
return "_" * (10 + len(cb))
# Replace common N-letter sequences with <0x80 | index>, where
# the common sequences are stored in a separate table.
# This isn't very useful, need a smarter way to find top-ngrams.
def ngram_compression(error_strings):
topn = collections.Counter()
N = 2
for line in error_strings.keys():
check_non_ascii(line)
if len(line) < N:
continue
for i in range(0, len(line) - N, N):
topn[line[i : i + N]] += 1
def bytes_saved(item):
w, n = item
return -(len(w) * (n - 1))
top128 = sorted(topn.items(), key=bytes_saved)[:128]
index = [w for w, _ in top128]
index_lookup = {w: i for i, w in enumerate(index)}
for line in error_strings.keys():
result = ""
for i in range(0, len(line) - N + 1, N):
word = line[i : i + N]
if word in index_lookup:
result += "\\{:03o}".format(0b10000000 | index_lookup[word])
else:
result += word
if len(line) % N != 0:
result += line[len(line) - len(line) % N :]
error_strings[line] = result.strip()
return "".join(index)
def main(collected_path, fn):
error_strings = collections.OrderedDict()
py: Implement "common word" compression scheme for error messages. The idea here is that there's a moderate amount of ROM used up by exception text. Obviously we try to keep the messages short, and the code can enable terse errors, but it still adds up. Listed below is the total string data size for various ports: bare-arm 2860 minimal 2876 stm32 8926 (PYBV11) cc3200 3751 esp32 5721 This commit implements compression of these strings. It takes advantage of the fact that these strings are all 7-bit ascii and extracts the top 128 frequently used words from the messages and stores them packed (dropping their null-terminator), then uses (0x80 | index) inside strings to refer to these common words. Spaces are automatically added around words, saving more bytes. This happens transparently in the build process, mirroring the steps that are used to generate the QSTR data. The MP_COMPRESSED_ROM_TEXT macro wraps any literal string that should compressed, and it's automatically decompressed in mp_decompress_rom_string. There are many schemes that could be used for the compression, and some are included in py/makecompresseddata.py for reference (space, Huffman, ngram, common word). Results showed that the common-word compression gets better results. This is before counting the increased cost of the Huffman decoder. This might be slightly counter-intuitive, but this data is extremely repetitive at a word-level, and the byte-level entropy coder can't quite exploit that as efficiently. Ideally one would combine both approaches, but for now the common-word approach is the one that is used. For additional comparison, the size of the raw data compressed with gzip and zlib is calculated, as a sort of proxy for a lower entropy bound. With this scheme we come within 15% on stm32, and 30% on bare-arm (i.e. we use x% more bytes than the data compressed with gzip -- not counting the code overhead of a decoder, and how this would be hypothetically implemented). The feature is disabled by default and can be enabled by setting MICROPY_ROM_TEXT_COMPRESSION at the Makefile-level.
2019-09-26 13:19:29 +01:00
max_uncompressed_len = 0
num_uses = 0
# Read in all MP_ERROR_TEXT strings.
with open(collected_path, "r") as f:
for line in f:
line = line.strip()
if not line:
continue
num_uses += 1
error_strings[line] = None
max_uncompressed_len = max(max_uncompressed_len, len(line))
# So that objexcept.c can figure out how big the buffer needs to be.
print("#define MP_MAX_UNCOMPRESSED_TEXT_LEN ({})".format(max_uncompressed_len))
# Run the compression.
compressed_data = fn(error_strings)
# Print the data table.
print('MP_COMPRESSED_DATA("{}")'.format(compressed_data))
# Print the replacements.
for uncomp, comp in error_strings.items():
if uncomp == comp:
prefix = ""
else:
prefix = "\\{:03o}".format(_COMPRESSED_MARKER)
print('MP_MATCH_COMPRESSED("{}", "{}{}")'.format(uncomp, prefix, comp))
py: Implement "common word" compression scheme for error messages. The idea here is that there's a moderate amount of ROM used up by exception text. Obviously we try to keep the messages short, and the code can enable terse errors, but it still adds up. Listed below is the total string data size for various ports: bare-arm 2860 minimal 2876 stm32 8926 (PYBV11) cc3200 3751 esp32 5721 This commit implements compression of these strings. It takes advantage of the fact that these strings are all 7-bit ascii and extracts the top 128 frequently used words from the messages and stores them packed (dropping their null-terminator), then uses (0x80 | index) inside strings to refer to these common words. Spaces are automatically added around words, saving more bytes. This happens transparently in the build process, mirroring the steps that are used to generate the QSTR data. The MP_COMPRESSED_ROM_TEXT macro wraps any literal string that should compressed, and it's automatically decompressed in mp_decompress_rom_string. There are many schemes that could be used for the compression, and some are included in py/makecompresseddata.py for reference (space, Huffman, ngram, common word). Results showed that the common-word compression gets better results. This is before counting the increased cost of the Huffman decoder. This might be slightly counter-intuitive, but this data is extremely repetitive at a word-level, and the byte-level entropy coder can't quite exploit that as efficiently. Ideally one would combine both approaches, but for now the common-word approach is the one that is used. For additional comparison, the size of the raw data compressed with gzip and zlib is calculated, as a sort of proxy for a lower entropy bound. With this scheme we come within 15% on stm32, and 30% on bare-arm (i.e. we use x% more bytes than the data compressed with gzip -- not counting the code overhead of a decoder, and how this would be hypothetically implemented). The feature is disabled by default and can be enabled by setting MICROPY_ROM_TEXT_COMPRESSION at the Makefile-level.
2019-09-26 13:19:29 +01:00
# Used to calculate the "true" length of the (escaped) compressed strings.
def unescape(s):
return re.sub(r"\\\d\d\d", "!", s)
# Stats. Note this doesn't include the cost of the decompressor code.
uncomp_len = sum(len(s) + 1 for s in error_strings.keys())
comp_len = sum(1 + len(unescape(s)) + 1 for s in error_strings.values())
data_len = len(compressed_data) + 1 if compressed_data else 0
print("// Total input length: {}".format(uncomp_len))
print("// Total compressed length: {}".format(comp_len))
print("// Total data length: {}".format(data_len))
print("// Predicted saving: {}".format(uncomp_len - comp_len - data_len))
# Somewhat meaningless comparison to zlib/gzip.
all_input_bytes = "\\0".join(error_strings.keys()).encode()
print()
if hasattr(gzip, "compress"):
gzip_len = len(gzip.compress(all_input_bytes)) + num_uses * 4
print("// gzip length: {}".format(gzip_len))
print("// Percentage of gzip: {:.1f}%".format(100 * (comp_len + data_len) / gzip_len))
if hasattr(zlib, "compress"):
zlib_len = len(zlib.compress(all_input_bytes)) + num_uses * 4
print("// zlib length: {}".format(zlib_len))
print("// Percentage of zlib: {:.1f}%".format(100 * (comp_len + data_len) / zlib_len))
if __name__ == "__main__":
main(sys.argv[1], word_compression)