Unroll wuffs_base__pixel_swizzler__copy_4_1

It's a mystery why some non-bgra benchmarks improve (e.g. std/lzw
improves, std/deflate does not), but I'll take it.

name                                    old speed      new speed      delta

wuffs_lzw_decode_20k/clang               242MB/s ± 0%   251MB/s ± 0%  +3.75%  (p=0.008 n=5+5)
wuffs_lzw_decode_100k/clang              419MB/s ± 0%   432MB/s ± 0%  +3.16%  (p=0.008 n=5+5)

wuffs_lzw_decode_20k/gcc                 260MB/s ± 0%   264MB/s ± 0%  +1.40%  (p=0.008 n=5+5)
wuffs_lzw_decode_100k/gcc                449MB/s ± 1%   474MB/s ± 0%  +5.45%  (p=0.008 n=5+5)

wuffs_deflate_decode_1k/clang            152MB/s ± 1%   154MB/s ± 1%  +0.97%  (p=0.032 n=5+5)
wuffs_deflate_decode_10k/clang           242MB/s ± 0%   243MB/s ± 1%    ~     (p=0.548 n=5+5)
wuffs_deflate_decode_100k/clang          276MB/s ± 0%   275MB/s ± 1%    ~     (p=0.222 n=5+5)

wuffs_deflate_decode_1k/gcc              157MB/s ± 0%   156MB/s ± 1%  -0.89%  (p=0.008 n=5+5)
wuffs_deflate_decode_10k/gcc             261MB/s ± 0%   260MB/s ± 0%    ~     (p=0.222 n=5+5)
wuffs_deflate_decode_100k/gcc            317MB/s ± 0%   315MB/s ± 0%  -0.66%  (p=0.008 n=5+5)

wuffs_gif_decode_1k_bw/clang             242MB/s ± 0%   242MB/s ± 1%    ~     (p=0.222 n=5+5)
wuffs_gif_decode_1k_color/clang          135MB/s ± 0%   139MB/s ± 0%  +3.12%  (p=0.008 n=5+5)
wuffs_gif_decode_10k_bgra/clang          663MB/s ± 0%   683MB/s ± 0%  +3.03%  (p=0.008 n=5+5)
wuffs_gif_decode_10k_indexed/clang       177MB/s ± 0%   182MB/s ± 1%  +2.91%  (p=0.008 n=5+5)
wuffs_gif_decode_100k_artificial/clang   513MB/s ± 0%   529MB/s ± 0%  +3.09%  (p=0.008 n=5+5)
wuffs_gif_decode_100k_realistic/clang    205MB/s ± 0%   211MB/s ± 0%  +2.66%  (p=0.008 n=5+5)
wuffs_gif_decode_1000k/clang             209MB/s ± 0%   214MB/s ± 2%  +2.62%  (p=0.008 n=5+5)
wuffs_gif_decode_anim_screencap/clang   1.05GB/s ± 0%  1.09GB/s ± 0%  +3.40%  (p=0.008 n=5+5)

wuffs_gif_decode_1k_bw/gcc               266MB/s ± 1%   265MB/s ± 0%  -0.46%  (p=0.016 n=5+5)
wuffs_gif_decode_1k_color/gcc            149MB/s ± 1%   150MB/s ± 0%  +0.93%  (p=0.008 n=5+5)
wuffs_gif_decode_10k_bgra/gcc            597MB/s ± 1%   609MB/s ± 0%  +1.96%  (p=0.008 n=5+5)
wuffs_gif_decode_10k_indexed/gcc         186MB/s ± 0%   189MB/s ± 0%  +1.89%  (p=0.008 n=5+5)
wuffs_gif_decode_100k_artificial/gcc     520MB/s ± 0%   518MB/s ± 0%  -0.39%  (p=0.032 n=5+5)
wuffs_gif_decode_100k_realistic/gcc      212MB/s ± 0%   213MB/s ± 0%  +0.55%  (p=0.008 n=5+5)
wuffs_gif_decode_1000k/gcc               216MB/s ± 0%   217MB/s ± 0%  +0.56%  (p=0.016 n=5+5)
wuffs_gif_decode_anim_screencap/gcc     1.09GB/s ± 1%  1.09GB/s ± 0%    ~     (p=0.548 n=5+5)
This commit is contained in:
Nigel Tao committed 2018-12-26 13:40:43 +11:00
1 parent b2d598aac5
commit 80bd4061b5
3 files changed
+65 -28

No files matched your search

+30 -12
View File
@@ -36,19 +36,37 @@ wuffs_base__pixel_swizzler__copy_4_1(wuffs_base__slice_u8 dst,
uint8_t* s = src.ptr;
size_t n = len;
while (n--) {
uint8_t* p = dst_palette.ptr + ((size_t)(*s) * 4);
uint8_t b0 = p[0];
uint8_t b1 = p[1];
uint8_t b2 = p[2];
uint8_t b3 = p[3];
d[0] = b0;
d[1] = b1;
d[2] = b2;
d[3] = b3;
s += 1;
d += 4;
const int N = 4;
while (n >= N) {
wuffs_base__store_u32le(
d + (0 * 4),
wuffs_base__load_u32le(dst_palette.ptr + ((uint32_t)(s[0]) * 4)));
wuffs_base__store_u32le(
d + (1 * 4),
wuffs_base__load_u32le(dst_palette.ptr + ((uint32_t)(s[1]) * 4)));
wuffs_base__store_u32le(
d + (2 * 4),
wuffs_base__load_u32le(dst_palette.ptr + ((uint32_t)(s[2]) * 4)));
wuffs_base__store_u32le(
d + (3 * 4),
wuffs_base__load_u32le(dst_palette.ptr + ((uint32_t)(s[3]) * 4)));
s += 1 * N;
d += 4 * N;
n -= 1 * N;
}
while (n >= 1) {
wuffs_base__store_u32le(
d + (0 * 4),
wuffs_base__load_u32le(dst_palette.ptr + ((uint32_t)(s[0]) * 4)));
s += 1 * 1;
d += 4 * 1;
n -= 1 * 1;
}
return len;
}
+5 -4
View File
@@ -24,10 +24,11 @@ const baseAllImplC = "" +
""
const baseImageImplC = "" +
"// ---------------- Images\n\nstatic uint64_t //\nwuffs_base__pixel_swizzler__copy_1_1(wuffs_base__slice_u8 dst,\n wuffs_base__slice_u8 dst_palette,\n wuffs_base__slice_u8 src) {\n return wuffs_base__slice_u8__copy_from_slice(dst, src);\n}\n\nstatic uint64_t //\nwuffs_base__pixel_swizzler__copy_4_1(wuffs_base__slice_u8 dst,\n wuffs_base__slice_u8 dst_palette,\n wuffs_base__slice_u8 src) {\n if (dst_palette.len != 1024) {\n return 0;\n }\n size_t dst_len4 = dst.len / 4;\n size_t len = dst_len4 < src.len ? dst_len4 : src.len;\n uint8_t* d = dst.ptr;\n uint8_t* s = src.ptr;\n\n size_t n = len;\n while (n--) {\n uint8_t* p = dst_palette.ptr + ((size_t)(*s) * 4);\n uint8_t b0 = p[0];\n uint8_t b1 = p[1];\n uint8_t b2 = p[2];\n uint8_t b3 = p[3];\n d[0] = b0;\n d[1] = b1;\n d[2] = b2;\n d[3] = b3;\n s += 1;\n d += 4;\n }\n return len;\n}\n\nstatic uint64_t //\nwu" +
"ffs_base__pixel_swizzler__swap_rgbx_bgrx(wuffs_base__slice_u8 dst,\n wuffs_base__slice_u8 src) {\n size_t len4 = (dst.len < src.len ? dst.len : src.len) / 4;\n uint8_t* d = dst.ptr;\n uint8_t* s = src.ptr;\n\n size_t n = len4;\n while (n--) {\n uint8_t b0 = s[0];\n uint8_t b1 = s[1];\n uint8_t b2 = s[2];\n uint8_t b3 = s[3];\n d[0] = b2;\n d[1] = b1;\n d[2] = b0;\n d[3] = b3;\n s += 4;\n d += 4;\n }\n return len4 * 4;\n}\n\nvoid //\nwuffs_base__pixel_swizzler__initialize(wuffs_base__pixel_swizzler* p,\n wuffs_base__pixel_format dst_format,\n wuffs_base__slice_u8 dst_palette,\n wuffs_base__pixel_format src_format,\n wuffs_base__slice_u8 src_palette) {\n if (!p) {\n return;\n }\n\n // TODO: support many more formats.\n\n uint64_t (*func)(wuffs_base__slice_u8 dst, wuffs_base__slice_u8 dst_palette,\n " +
" wuffs_base__slice_u8 src) = NULL;\n\n switch (src_format) {\n case WUFFS_BASE__PIXEL_FORMAT__INDEXED__BGRA_BINARY:\n switch (dst_format) {\n case WUFFS_BASE__PIXEL_FORMAT__INDEXED__BGRA_NONPREMUL:\n case WUFFS_BASE__PIXEL_FORMAT__INDEXED__BGRA_PREMUL:\n case WUFFS_BASE__PIXEL_FORMAT__INDEXED__BGRA_BINARY:\n if (wuffs_base__slice_u8__copy_from_slice(dst_palette, src_palette) !=\n 1024) {\n break;\n }\n func = wuffs_base__pixel_swizzler__copy_1_1;\n break;\n case WUFFS_BASE__PIXEL_FORMAT__BGRA_NONPREMUL:\n case WUFFS_BASE__PIXEL_FORMAT__BGRA_PREMUL:\n case WUFFS_BASE__PIXEL_FORMAT__BGRA_BINARY:\n if (wuffs_base__slice_u8__copy_from_slice(dst_palette, src_palette) !=\n 1024) {\n break;\n }\n func = wuffs_base__pixel_swizzler__copy_4_1;\n break;\n case WUFFS_BASE__PIXEL_FORMAT__RGBA_NONPREMUL:\n case WUFFS_BASE__PIXEL_FORMAT__RGBA_PREMUL:\n ca" +
"se WUFFS_BASE__PIXEL_FORMAT__RGBA_BINARY:\n if (wuffs_base__pixel_swizzler__swap_rgbx_bgrx(dst_palette,\n src_palette) != 1024) {\n break;\n }\n func = wuffs_base__pixel_swizzler__copy_4_1;\n break;\n default:\n break;\n }\n break;\n\n default:\n break;\n }\n\n p->private_impl.func = func;\n}\n\nuint64_t //\nwuffs_base__pixel_swizzler__swizzle_packed(wuffs_base__pixel_swizzler* p,\n wuffs_base__slice_u8 dst,\n wuffs_base__slice_u8 dst_palette,\n wuffs_base__slice_u8 src) {\n if (p && p->private_impl.func) {\n return (*(p->private_impl.func))(dst, dst_palette, src);\n }\n return 0;\n}\n" +
"// ---------------- Images\n\nstatic uint64_t //\nwuffs_base__pixel_swizzler__copy_1_1(wuffs_base__slice_u8 dst,\n wuffs_base__slice_u8 dst_palette,\n wuffs_base__slice_u8 src) {\n return wuffs_base__slice_u8__copy_from_slice(dst, src);\n}\n\nstatic uint64_t //\nwuffs_base__pixel_swizzler__copy_4_1(wuffs_base__slice_u8 dst,\n wuffs_base__slice_u8 dst_palette,\n wuffs_base__slice_u8 src) {\n if (dst_palette.len != 1024) {\n return 0;\n }\n size_t dst_len4 = dst.len / 4;\n size_t len = dst_len4 < src.len ? dst_len4 : src.len;\n uint8_t* d = dst.ptr;\n uint8_t* s = src.ptr;\n\n size_t n = len;\n const int N = 4;\n\n while (n >= N) {\n wuffs_base__store_u32le(\n d + (0 * 4),\n wuffs_base__load_u32le(dst_palette.ptr + ((uint32_t)(s[0]) * 4)));\n wuffs_base__store_u32le(\n d + (1 * 4),\n wuffs_base__load_u32le(dst_palette.ptr + ((uint32_t)(s[1]) * 4)));\n " +
" wuffs_base__store_u32le(\n d + (2 * 4),\n wuffs_base__load_u32le(dst_palette.ptr + ((uint32_t)(s[2]) * 4)));\n wuffs_base__store_u32le(\n d + (3 * 4),\n wuffs_base__load_u32le(dst_palette.ptr + ((uint32_t)(s[3]) * 4)));\n\n s += 1 * N;\n d += 4 * N;\n n -= 1 * N;\n }\n\n while (n >= 1) {\n wuffs_base__store_u32le(\n d + (0 * 4),\n wuffs_base__load_u32le(dst_palette.ptr + ((uint32_t)(s[0]) * 4)));\n\n s += 1 * 1;\n d += 4 * 1;\n n -= 1 * 1;\n }\n\n return len;\n}\n\nstatic uint64_t //\nwuffs_base__pixel_swizzler__swap_rgbx_bgrx(wuffs_base__slice_u8 dst,\n wuffs_base__slice_u8 src) {\n size_t len4 = (dst.len < src.len ? dst.len : src.len) / 4;\n uint8_t* d = dst.ptr;\n uint8_t* s = src.ptr;\n\n size_t n = len4;\n while (n--) {\n uint8_t b0 = s[0];\n uint8_t b1 = s[1];\n uint8_t b2 = s[2];\n uint8_t b3 = s[3];\n d[0] = b2;\n d[1] = b1;\n d[2] = b0;\n d[3] = b3;\n s += 4;\n d += 4;\n }\n return len4 * 4;\n}\n" +
"\nvoid //\nwuffs_base__pixel_swizzler__initialize(wuffs_base__pixel_swizzler* p,\n wuffs_base__pixel_format dst_format,\n wuffs_base__slice_u8 dst_palette,\n wuffs_base__pixel_format src_format,\n wuffs_base__slice_u8 src_palette) {\n if (!p) {\n return;\n }\n\n // TODO: support many more formats.\n\n uint64_t (*func)(wuffs_base__slice_u8 dst, wuffs_base__slice_u8 dst_palette,\n wuffs_base__slice_u8 src) = NULL;\n\n switch (src_format) {\n case WUFFS_BASE__PIXEL_FORMAT__INDEXED__BGRA_BINARY:\n switch (dst_format) {\n case WUFFS_BASE__PIXEL_FORMAT__INDEXED__BGRA_NONPREMUL:\n case WUFFS_BASE__PIXEL_FORMAT__INDEXED__BGRA_PREMUL:\n case WUFFS_BASE__PIXEL_FORMAT__INDEXED__BGRA_BINARY:\n if (wuffs_base__slice_u8__copy_from_slice(dst_palette, src_palette) !=\n 1024) {\n break;\n }\n func" +
" = wuffs_base__pixel_swizzler__copy_1_1;\n break;\n case WUFFS_BASE__PIXEL_FORMAT__BGRA_NONPREMUL:\n case WUFFS_BASE__PIXEL_FORMAT__BGRA_PREMUL:\n case WUFFS_BASE__PIXEL_FORMAT__BGRA_BINARY:\n if (wuffs_base__slice_u8__copy_from_slice(dst_palette, src_palette) !=\n 1024) {\n break;\n }\n func = wuffs_base__pixel_swizzler__copy_4_1;\n break;\n case WUFFS_BASE__PIXEL_FORMAT__RGBA_NONPREMUL:\n case WUFFS_BASE__PIXEL_FORMAT__RGBA_PREMUL:\n case WUFFS_BASE__PIXEL_FORMAT__RGBA_BINARY:\n if (wuffs_base__pixel_swizzler__swap_rgbx_bgrx(dst_palette,\n src_palette) != 1024) {\n break;\n }\n func = wuffs_base__pixel_swizzler__copy_4_1;\n break;\n default:\n break;\n }\n break;\n\n default:\n break;\n }\n\n p->private_impl.func = func;\n}\n\nuint64_t //\nwuffs_base__pixel_swizzler__swizzle_packed(wuffs_base_" +
"_pixel_swizzler* p,\n wuffs_base__slice_u8 dst,\n wuffs_base__slice_u8 dst_palette,\n wuffs_base__slice_u8 src) {\n if (p && p->private_impl.func) {\n return (*(p->private_impl.func))(dst, dst_palette, src);\n }\n return 0;\n}\n" +
""
const baseCorePrivateH = "" +
+30 -12
View File
@@ -4022,19 +4022,37 @@ wuffs_base__pixel_swizzler__copy_4_1(wuffs_base__slice_u8 dst,
uint8_t* s = src.ptr;
size_t n = len;
while (n--) {
uint8_t* p = dst_palette.ptr + ((size_t)(*s) * 4);
uint8_t b0 = p[0];
uint8_t b1 = p[1];
uint8_t b2 = p[2];
uint8_t b3 = p[3];
d[0] = b0;
d[1] = b1;
d[2] = b2;
d[3] = b3;
s += 1;
d += 4;
const int N = 4;
while (n >= N) {
wuffs_base__store_u32le(
d + (0 * 4),
wuffs_base__load_u32le(dst_palette.ptr + ((uint32_t)(s[0]) * 4)));
wuffs_base__store_u32le(
d + (1 * 4),
wuffs_base__load_u32le(dst_palette.ptr + ((uint32_t)(s[1]) * 4)));
wuffs_base__store_u32le(
d + (2 * 4),
wuffs_base__load_u32le(dst_palette.ptr + ((uint32_t)(s[2]) * 4)));
wuffs_base__store_u32le(
d + (3 * 4),
wuffs_base__load_u32le(dst_palette.ptr + ((uint32_t)(s[3]) * 4)));
s += 1 * N;
d += 4 * N;
n -= 1 * N;
}
while (n >= 1) {
wuffs_base__store_u32le(
d + (0 * 4),
wuffs_base__load_u32le(dst_palette.ptr + ((uint32_t)(s[0]) * 4)));
s += 1 * 1;
d += 4 * 1;
n -= 1 * 1;
}
return len;
}