Optimize std/deflate decoding for RLE sections

name                                                      old speed      new speed      delta

wuffs_deflate_decode_1k_full_init/clang9                   104MB/s ± 1%   104MB/s ± 1%     ~     (p=0.161 n=9+9)
wuffs_deflate_decode_1k_part_init/clang9                   121MB/s ± 1%   119MB/s ± 1%   -0.92%  (p=0.000 n=9+9)
wuffs_deflate_decode_10k_full_init/clang9                  223MB/s ± 1%   223MB/s ± 1%     ~     (p=0.796 n=9+9)
wuffs_deflate_decode_10k_part_init/clang9                  230MB/s ± 1%   228MB/s ± 2%     ~     (p=0.065 n=10+9)
wuffs_deflate_decode_100k_just_one_read/clang9             279MB/s ± 1%   292MB/s ± 0%   +4.78%  (p=0.000 n=9+8)
wuffs_deflate_decode_100k_many_big_reads/clang9            178MB/s ± 5%   189MB/s ± 2%   +6.41%  (p=0.000 n=10+10)

wuffs_deflate_decode_1k_full_init/gcc10                    103MB/s ± 1%   102MB/s ± 1%   -0.57%  (p=0.022 n=10+9)
wuffs_deflate_decode_1k_part_init/gcc10                    122MB/s ± 1%   121MB/s ± 3%   -1.46%  (p=0.000 n=10+9)
wuffs_deflate_decode_10k_full_init/gcc10                   235MB/s ± 1%   231MB/s ± 1%   -1.34%  (p=0.000 n=10+8)
wuffs_deflate_decode_10k_part_init/gcc10                   242MB/s ± 1%   239MB/s ± 0%   -1.40%  (p=0.001 n=9+8)
wuffs_deflate_decode_100k_just_one_read/gcc10              300MB/s ± 4%   297MB/s ± 2%     ~     (p=0.156 n=10+9)
wuffs_deflate_decode_100k_many_big_reads/gcc10             193MB/s ± 2%   197MB/s ± 1%   +2.29%  (p=0.000 n=10+9)

wuffs_gzip_decode_10k/clang9                               220MB/s ± 1%   219MB/s ± 1%     ~     (p=0.605 n=9+9)
wuffs_gzip_decode_100k/clang9                              269MB/s ± 2%   283MB/s ± 2%   +4.98%  (p=0.000 n=9+10)

wuffs_gzip_decode_10k/gcc10                                227MB/s ± 1%   224MB/s ± 1%   -1.42%  (p=0.000 n=10+9)
wuffs_gzip_decode_100k/gcc10                               283MB/s ± 1%   285MB/s ± 4%     ~     (p=0.182 n=9+10)

wuffs_zlib_decode_10k/clang9                               226MB/s ± 1%   220MB/s ± 4%   -2.66%  (p=0.000 n=9+10)
wuffs_zlib_decode_100k/clang9                              272MB/s ± 2%   285MB/s ± 1%   +4.77%  (p=0.000 n=10+8)

wuffs_zlib_decode_10k/gcc10                                227MB/s ± 1%   212MB/s ± 1%   -6.80%  (p=0.000 n=10+9)
wuffs_zlib_decode_100k/gcc10                               282MB/s ± 1%   273MB/s ± 0%   -3.32%  (p=0.000 n=9+8)

wuffs_png_decode_image_19k_8bpp/clang9                     148MB/s ± 1%   138MB/s ± 2%   -6.79%  (p=0.000 n=9+10)
wuffs_png_decode_image_40k_24bpp/clang9                    170MB/s ± 1%   171MB/s ± 2%     ~     (p=0.133 n=9+10)
wuffs_png_decode_image_77k_8bpp/clang9                     525MB/s ± 1%   495MB/s ± 4%   -5.81%  (p=0.000 n=8+9)
wuffs_png_decode_image_552k_32bpp_ignore_checksum/clang9   400MB/s ± 2%   457MB/s ± 4%  +14.16%  (p=0.000 n=9+9)
wuffs_png_decode_image_552k_32bpp_verify_checksum/clang9   388MB/s ± 1%   435MB/s ± 5%  +12.17%  (p=0.000 n=9+10)
wuffs_png_decode_image_4002k_24bpp/clang9                  175MB/s ± 1%   172MB/s ± 2%   -1.78%  (p=0.005 n=10+10)

wuffs_png_decode_image_19k_8bpp/gcc10                      154MB/s ± 3%   157MB/s ± 1%   +1.72%  (p=0.001 n=9+9)
wuffs_png_decode_image_40k_24bpp/gcc10                     182MB/s ± 1%   187MB/s ± 1%   +2.56%  (p=0.000 n=9+10)
wuffs_png_decode_image_77k_8bpp/gcc10                      552MB/s ± 3%   554MB/s ± 1%     ~     (p=0.863 n=9+9)
wuffs_png_decode_image_552k_32bpp_ignore_checksum/gcc10    426MB/s ± 2%   502MB/s ± 1%  +17.85%  (p=0.000 n=10+10)
wuffs_png_decode_image_552k_32bpp_verify_checksum/gcc10    410MB/s ± 2%   481MB/s ± 1%  +17.17%  (p=0.000 n=9+8)
wuffs_png_decode_image_4002k_24bpp/gcc10                   186MB/s ± 1%   189MB/s ± 2%   +1.33%  (p=0.001 n=9+10)
This commit is contained in:
Nigel Tao committed 2021-04-09 14:31:52 +10:00
1 parent a7ef82f7db
commit 1dd7ad4413
9 files changed
+157 -25

No files matched your search

+40
View File
@@ -228,6 +228,46 @@ wuffs_base__io_writer__limited_copy_u32_from_history_fast(uint8_t** ptr_iop_w,
return length;
}
// wuffs_base__io_writer__limited_copy_u32_from_history_8_byte_chunks_distance_1_fast
// copies the previous byte (the one immediately before *ptr_iop_w), copying 8
// byte chunks at a time. Each chunk contains 8 repetitions of the same byte.
//
// In terms of number of bytes copied, length is rounded up to a multiple of 8.
// As a special case, a zero length rounds up to 8 (even though 0 is already a
// multiple of 8), since there is always at least one 8 byte chunk copied.
//
// In terms of advancing *ptr_iop_w, length is not rounded up.
//
// The caller needs to prove that:
// - (length + 8) <= (io2_w - *ptr_iop_w)
// - distance == 1
// - distance <= (*ptr_iop_w - io1_w)
static inline uint32_t //
wuffs_base__io_writer__limited_copy_u32_from_history_8_byte_chunks_distance_1_fast(
uint8_t** ptr_iop_w,
uint8_t* io1_w,
uint8_t* io2_w,
uint32_t length,
uint32_t distance) {
uint8_t* p = *ptr_iop_w;
uint64_t x = p[-1];
x |= x << 8;
x |= x << 16;
x |= x << 32;
uint32_t n = length;
while (1) {
wuffs_base__poke_u64le__no_bounds_check(p, x);
if (n <= 8) {
p += n;
break;
}
p += 8;
n -= 8;
}
*ptr_iop_w = p;
return length;
}
// wuffs_base__io_writer__limited_copy_u32_from_history_8_byte_chunks_fast is
// like the wuffs_base__io_writer__limited_copy_u32_from_history_fast function
// above, but copies 8 byte chunks at a time.
+1
View File
@@ -308,6 +308,7 @@ func (g *gen) writeBuiltinIOWriter(b *buffer, recv *a.Expr, method t.ID, args []
switch method {
case t.IDLimitedCopyU32FromHistory,
t.IDLimitedCopyU32FromHistory8ByteChunksDistance1Fast,
t.IDLimitedCopyU32FromHistory8ByteChunksFast,
t.IDLimitedCopyU32FromHistoryFast:
b.printf("wuffs_base__io_writer__%s(\n&%s%s, %s%s, %s%s",
+6 -4
View File
@@ -223,10 +223,12 @@ const BaseIOPrivateH = "" +
"" +
"// --------\n\nstatic inline uint64_t //\nwuffs_base__io_writer__copy_from_slice(uint8_t** ptr_iop_w,\n uint8_t* io2_w,\n wuffs_base__slice_u8 src) {\n uint8_t* iop_w = *ptr_iop_w;\n size_t n = src.len;\n if (n > ((size_t)(io2_w - iop_w))) {\n n = (size_t)(io2_w - iop_w);\n }\n if (n > 0) {\n memmove(iop_w, src.ptr, n);\n *ptr_iop_w += n;\n }\n return (uint64_t)(n);\n}\n\nstatic inline void //\nwuffs_base__io_writer__limit(uint8_t** ptr_io2_w,\n uint8_t* iop_w,\n uint64_t limit) {\n if (((uint64_t)(*ptr_io2_w - iop_w)) > limit) {\n *ptr_io2_w = iop_w + limit;\n }\n}\n\nstatic inline uint32_t //\nwuffs_base__io_writer__limited_copy_u32_from_history(uint8_t** ptr_iop_w,\n uint8_t* io1_w,\n uint8_t* io2_w,\n uint32_t length,\n " +
" uint32_t distance) {\n if (!distance) {\n return 0;\n }\n uint8_t* p = *ptr_iop_w;\n if ((size_t)(p - io1_w) < (size_t)(distance)) {\n return 0;\n }\n uint8_t* q = p - distance;\n size_t n = (size_t)(io2_w - p);\n if ((size_t)(length) > n) {\n length = (uint32_t)(n);\n } else {\n n = (size_t)(length);\n }\n // TODO: unrolling by 3 seems best for the std/deflate benchmarks, but that\n // is mostly because 3 is the minimum length for the deflate format. This\n // function implementation shouldn't overfit to that one format. Perhaps the\n // limited_copy_u32_from_history Wuffs method should also take an unroll hint\n // argument, and the cgen can look if that argument is the constant\n // expression '3'.\n //\n // See also wuffs_base__io_writer__limited_copy_u32_from_history_fast below.\n for (; n >= 3; n -= 3) {\n *p++ = *q++;\n *p++ = *q++;\n *p++ = *q++;\n }\n for (; n; n--) {\n *p++ = *q++;\n }\n *ptr_iop_w = p;\n return length;\n}\n\n// wuffs_base__io_w" +
"riter__limited_copy_u32_from_history_fast is like the\n// wuffs_base__io_writer__limited_copy_u32_from_history function above, but has\n// stronger pre-conditions.\n//\n// The caller needs to prove that:\n// - length <= (io2_w - *ptr_iop_w)\n// - distance >= 1\n// - distance <= (*ptr_iop_w - io1_w)\nstatic inline uint32_t //\nwuffs_base__io_writer__limited_copy_u32_from_history_fast(uint8_t** ptr_iop_w,\n uint8_t* io1_w,\n uint8_t* io2_w,\n uint32_t length,\n uint32_t distance) {\n uint8_t* p = *ptr_iop_w;\n uint8_t* q = p - distance;\n uint32_t n = length;\n for (; n >= 3; n -= 3) {\n *p++ = *q++;\n *p++ = *q++;\n *p++ = *q++;\n }\n for (; n; n--) {\n *p++ = *q++;\n }\n *ptr_iop_w = p;\n return length;\n}\n\n// wuffs_base__io_writer__limited_copy_u32_from_history_8_byte_chunks_fast" +
" is\n// like the wuffs_base__io_writer__limited_copy_u32_from_history_fast function\n// above, but copies 8 byte chunks at a time.\n//\n// In terms of number of bytes copied, length is rounded up to a multiple of 8.\n// As a special case, a zero length rounds up to 8 (even though 0 is already a\n// multiple of 8), since there is always at least one 8 byte chunk copied.\n//\n// In terms of advancing *ptr_iop_w, length is not rounded up.\n//\n// The caller needs to prove that:\n// - (length + 8) <= (io2_w - *ptr_iop_w)\n// - distance >= 8\n// - distance <= (*ptr_iop_w - io1_w)\nstatic inline uint32_t //\nwuffs_base__io_writer__limited_copy_u32_from_history_8_byte_chunks_fast(\n uint8_t** ptr_iop_w,\n uint8_t* io1_w,\n uint8_t* io2_w,\n uint32_t length,\n uint32_t distance) {\n uint8_t* p = *ptr_iop_w;\n uint8_t* q = p - distance;\n uint32_t n = length;\n while (1) {\n memcpy(p, q, 8);\n if (n <= 8) {\n p += n;\n break;\n }\n p += 8;\n q += 8;\n n -= 8;\n }\n *ptr_iop_w = p;\n ret" +
"urn length;\n}\n\nstatic inline uint32_t //\nwuffs_base__io_writer__limited_copy_u32_from_reader(uint8_t** ptr_iop_w,\n uint8_t* io2_w,\n uint32_t length,\n const uint8_t** ptr_iop_r,\n const uint8_t* io2_r) {\n uint8_t* iop_w = *ptr_iop_w;\n size_t n = length;\n if (n > ((size_t)(io2_w - iop_w))) {\n n = (size_t)(io2_w - iop_w);\n }\n const uint8_t* iop_r = *ptr_iop_r;\n if (n > ((size_t)(io2_r - iop_r))) {\n n = (size_t)(io2_r - iop_r);\n }\n if (n > 0) {\n memmove(iop_w, iop_r, n);\n *ptr_iop_w += n;\n *ptr_iop_r += n;\n }\n return (uint32_t)(n);\n}\n\nstatic inline uint32_t //\nwuffs_base__io_writer__limited_copy_u32_from_slice(uint8_t** ptr_iop_w,\n uint8_t* io2_w,\n uint32_t length,\n " +
" wuffs_base__slice_u8 src) {\n uint8_t* iop_w = *ptr_iop_w;\n size_t n = src.len;\n if (n > length) {\n n = length;\n }\n if (n > ((size_t)(io2_w - iop_w))) {\n n = (size_t)(io2_w - iop_w);\n }\n if (n > 0) {\n memmove(iop_w, src.ptr, n);\n *ptr_iop_w += n;\n }\n return (uint32_t)(n);\n}\n\nstatic inline wuffs_base__io_buffer* //\nwuffs_base__io_writer__set(wuffs_base__io_buffer* b,\n uint8_t** ptr_iop_w,\n uint8_t** ptr_io0_w,\n uint8_t** ptr_io1_w,\n uint8_t** ptr_io2_w,\n wuffs_base__slice_u8 data) {\n b->data = data;\n b->meta.wi = 0;\n b->meta.ri = 0;\n b->meta.pos = 0;\n b->meta.closed = false;\n\n *ptr_iop_w = data.ptr;\n *ptr_io0_w = data.ptr;\n *ptr_io1_w = data.ptr;\n *ptr_io2_w = data.ptr + data.len;\n\n return b;\n}\n\n" +
"riter__limited_copy_u32_from_history_fast is like the\n// wuffs_base__io_writer__limited_copy_u32_from_history function above, but has\n// stronger pre-conditions.\n//\n// The caller needs to prove that:\n// - length <= (io2_w - *ptr_iop_w)\n// - distance >= 1\n// - distance <= (*ptr_iop_w - io1_w)\nstatic inline uint32_t //\nwuffs_base__io_writer__limited_copy_u32_from_history_fast(uint8_t** ptr_iop_w,\n uint8_t* io1_w,\n uint8_t* io2_w,\n uint32_t length,\n uint32_t distance) {\n uint8_t* p = *ptr_iop_w;\n uint8_t* q = p - distance;\n uint32_t n = length;\n for (; n >= 3; n -= 3) {\n *p++ = *q++;\n *p++ = *q++;\n *p++ = *q++;\n }\n for (; n; n--) {\n *p++ = *q++;\n }\n *ptr_iop_w = p;\n return length;\n}\n\n// wuffs_base__io_writer__limited_copy_u32_from_history_8_byte_chunks_dist" +
"ance_1_fast\n// copies the previous byte (the one immediately before *ptr_iop_w), copying 8\n// byte chunks at a time. Each chunk contains 8 repetitions of the same byte.\n//\n// In terms of number of bytes copied, length is rounded up to a multiple of 8.\n// As a special case, a zero length rounds up to 8 (even though 0 is already a\n// multiple of 8), since there is always at least one 8 byte chunk copied.\n//\n// In terms of advancing *ptr_iop_w, length is not rounded up.\n//\n// The caller needs to prove that:\n// - (length + 8) <= (io2_w - *ptr_iop_w)\n// - distance == 1\n// - distance <= (*ptr_iop_w - io1_w)\nstatic inline uint32_t //\nwuffs_base__io_writer__limited_copy_u32_from_history_8_byte_chunks_distance_1_fast(\n uint8_t** ptr_iop_w,\n uint8_t* io1_w,\n uint8_t* io2_w,\n uint32_t length,\n uint32_t distance) {\n uint8_t* p = *ptr_iop_w;\n uint64_t x = p[-1];\n x |= x << 8;\n x |= x << 16;\n x |= x << 32;\n uint32_t n = length;\n while (1) {\n wuffs_base__poke_u64le__no_bounds_check(" +
"p, x);\n if (n <= 8) {\n p += n;\n break;\n }\n p += 8;\n n -= 8;\n }\n *ptr_iop_w = p;\n return length;\n}\n\n// wuffs_base__io_writer__limited_copy_u32_from_history_8_byte_chunks_fast is\n// like the wuffs_base__io_writer__limited_copy_u32_from_history_fast function\n// above, but copies 8 byte chunks at a time.\n//\n// In terms of number of bytes copied, length is rounded up to a multiple of 8.\n// As a special case, a zero length rounds up to 8 (even though 0 is already a\n// multiple of 8), since there is always at least one 8 byte chunk copied.\n//\n// In terms of advancing *ptr_iop_w, length is not rounded up.\n//\n// The caller needs to prove that:\n// - (length + 8) <= (io2_w - *ptr_iop_w)\n// - distance >= 8\n// - distance <= (*ptr_iop_w - io1_w)\nstatic inline uint32_t //\nwuffs_base__io_writer__limited_copy_u32_from_history_8_byte_chunks_fast(\n uint8_t** ptr_iop_w,\n uint8_t* io1_w,\n uint8_t* io2_w,\n uint32_t length,\n uint32_t distance) {\n uint8_t* p = *ptr_iop_w;\n u" +
"int8_t* q = p - distance;\n uint32_t n = length;\n while (1) {\n memcpy(p, q, 8);\n if (n <= 8) {\n p += n;\n break;\n }\n p += 8;\n q += 8;\n n -= 8;\n }\n *ptr_iop_w = p;\n return length;\n}\n\nstatic inline uint32_t //\nwuffs_base__io_writer__limited_copy_u32_from_reader(uint8_t** ptr_iop_w,\n uint8_t* io2_w,\n uint32_t length,\n const uint8_t** ptr_iop_r,\n const uint8_t* io2_r) {\n uint8_t* iop_w = *ptr_iop_w;\n size_t n = length;\n if (n > ((size_t)(io2_w - iop_w))) {\n n = (size_t)(io2_w - iop_w);\n }\n const uint8_t* iop_r = *ptr_iop_r;\n if (n > ((size_t)(io2_r - iop_r))) {\n n = (size_t)(io2_r - iop_r);\n }\n if (n > 0) {\n memmove(iop_w, iop_r, n);\n *ptr_iop_w += n;\n *ptr_iop_r += n;\n }\n return (uint32_t)(n);\n}\n\nstatic inline uint32_t //\nwuffs_base__io_writer__limited_co" +
"py_u32_from_slice(uint8_t** ptr_iop_w,\n uint8_t* io2_w,\n uint32_t length,\n wuffs_base__slice_u8 src) {\n uint8_t* iop_w = *ptr_iop_w;\n size_t n = src.len;\n if (n > length) {\n n = length;\n }\n if (n > ((size_t)(io2_w - iop_w))) {\n n = (size_t)(io2_w - iop_w);\n }\n if (n > 0) {\n memmove(iop_w, src.ptr, n);\n *ptr_iop_w += n;\n }\n return (uint32_t)(n);\n}\n\nstatic inline wuffs_base__io_buffer* //\nwuffs_base__io_writer__set(wuffs_base__io_buffer* b,\n uint8_t** ptr_iop_w,\n uint8_t** ptr_io0_w,\n uint8_t** ptr_io1_w,\n uint8_t** ptr_io2_w,\n wuffs_base__slice_u8 data) {\n b->data = data;\n b->meta.wi = 0;\n b->meta.ri = 0;\n b->meta.pos = 0;\n b->meta.closed = false;\n\n *ptr_iop_w = data.ptr;\n *ptr_io0_w = data.ptr;\n *ptr_io1_" +
"w = data.ptr;\n *ptr_io2_w = data.ptr + data.len;\n\n return b;\n}\n\n" +
"" +
"// ---------------- I/O (Utility)\n\n#define wuffs_base__utility__empty_io_reader wuffs_base__empty_io_reader\n#define wuffs_base__utility__empty_io_writer wuffs_base__empty_io_writer\n" +
""