std/jpeg: optimize IDCT when AC terms are zero

name                                             old speed     new speed     delta

wuffs_jpeg_decode_19k_8bpp/clang11               199MB/s ± 0%  211MB/s ± 0%   +5.69%  (p=0.008 n=5+5)
wuffs_jpeg_decode_30k_24bpp_progressive/clang11  169MB/s ± 1%  176MB/s ± 0%   +4.32%  (p=0.008 n=5+5)
wuffs_jpeg_decode_30k_24bpp_sequential/clang11   269MB/s ± 0%  291MB/s ± 0%   +7.89%  (p=0.008 n=5+5)
wuffs_jpeg_decode_77k_24bpp/clang11              250MB/s ± 0%  254MB/s ± 0%   +1.90%  (p=0.008 n=5+5)
wuffs_jpeg_decode_552k_24bpp_420/clang11         258MB/s ± 0%  262MB/s ± 0%   +1.51%  (p=0.008 n=5+5)
wuffs_jpeg_decode_552k_24bpp_444/clang11         275MB/s ± 0%  356MB/s ± 0%  +29.59%  (p=0.016 n=5+4)
wuffs_jpeg_decode_4002k_24bpp/clang11            261MB/s ± 0%  263MB/s ± 1%   +0.61%  (p=0.008 n=5+5)

wuffs_jpeg_decode_19k_8bpp/gcc10                 153MB/s ± 0%  165MB/s ± 0%   +8.09%  (p=0.008 n=5+5)
wuffs_jpeg_decode_30k_24bpp_progressive/gcc10    180MB/s ± 0%  197MB/s ± 0%   +9.56%  (p=0.008 n=5+5)
wuffs_jpeg_decode_30k_24bpp_sequential/gcc10     288MB/s ± 0%  337MB/s ± 0%  +16.81%  (p=0.008 n=5+5)
wuffs_jpeg_decode_77k_24bpp/gcc10                237MB/s ± 0%  245MB/s ± 0%   +3.47%  (p=0.008 n=5+5)
wuffs_jpeg_decode_552k_24bpp_420/gcc10           247MB/s ± 0%  256MB/s ± 0%   +3.71%  (p=0.016 n=4+5)
wuffs_jpeg_decode_552k_24bpp_444/gcc10           249MB/s ± 0%  348MB/s ± 0%  +39.88%  (p=0.008 n=5+5)
wuffs_jpeg_decode_4002k_24bpp/gcc10              247MB/s ± 0%  254MB/s ± 0%   +2.81%  (p=0.008 n=5+5)
This commit is contained in:
Nigel Tao committed 2023-06-08 22:03:46 +10:00
1 parent ca2909b6d7
commit d6cc525059
3 files changed
+2281 -1524

No files matched your search

+48
View File
@@ -186,6 +186,28 @@ const (
const pass0 = `// ==== First pass, column $colX$.
if (0 == (
this.mcu_blocks[args.b][$row1colX$] |
this.mcu_blocks[args.b][$row2colX$] |
this.mcu_blocks[args.b][$row3colX$] |
this.mcu_blocks[args.b][$row4colX$] |
this.mcu_blocks[args.b][$row5colX$] |
this.mcu_blocks[args.b][$row6colX$] |
this.mcu_blocks[args.b][$row7colX$])) {
// Fast path when the 1-dimensional AC terms are all zero.
intermediate[$row0colX$] =
(this.util.sign_extend_convert_u16_u32(a: this.mcu_blocks[args.b][$row0colX$]) ~mod*
(this.quant_tables[args.q][$row0colX$] as base.u32)) ~mod<< 2
intermediate[$row1colX$] = intermediate[$row0colX$]
intermediate[$row2colX$] = intermediate[$row0colX$]
intermediate[$row3colX$] = intermediate[$row0colX$]
intermediate[$row4colX$] = intermediate[$row0colX$]
intermediate[$row5colX$] = intermediate[$row0colX$]
intermediate[$row6colX$] = intermediate[$row0colX$]
intermediate[$row7colX$] = intermediate[$row0colX$]
} else {
// Even rows.
bq2 = this.util.sign_extend_convert_u16_u32(a: this.mcu_blocks[args.b][$row2colX$]) ~mod* (this.quant_tables[args.q][$row2colX$] as base.u32)
@@ -249,10 +271,35 @@ intermediate[$row2colX$] = this.util.sign_extend_rshift_u32(a: (cd2 ~mod+ ck5) ~
intermediate[$row5colX$] = this.util.sign_extend_rshift_u32(a: (cd2 ~mod- ck5) ~mod+ (1 << 10), n: 11)
intermediate[$row3colX$] = this.util.sign_extend_rshift_u32(a: (cd3 ~mod+ ck7) ~mod+ (1 << 10), n: 11)
intermediate[$row4colX$] = this.util.sign_extend_rshift_u32(a: (cd3 ~mod- ck7) ~mod+ (1 << 10), n: 11)
}
`
const pass1 = `// ==== Second pass, row $rowY$.
if (0 == (
intermediate[$rowYcol1$] |
intermediate[$rowYcol2$] |
intermediate[$rowYcol3$] |
intermediate[$rowYcol4$] |
intermediate[$rowYcol5$] |
intermediate[$rowYcol6$] |
intermediate[$rowYcol7$])) {
// Fast path when the 1-dimensional AC terms are all zero.
$bounds_check$
args.dst_buffer[0] = BIAS_AND_CLAMP[((intermediate[$rowYcol0$] ~mod+ (1 << 4)) >> 5) & 1023]
args.dst_buffer[1] = args.dst_buffer[0]
args.dst_buffer[2] = args.dst_buffer[0]
args.dst_buffer[3] = args.dst_buffer[0]
args.dst_buffer[4] = args.dst_buffer[0]
args.dst_buffer[5] = args.dst_buffer[0]
args.dst_buffer[6] = args.dst_buffer[0]
args.dst_buffer[7] = args.dst_buffer[0]
$advance$
} else {
// Even columns.
in2 = intermediate[$rowYcol2$]
@@ -320,4 +367,5 @@ args.dst_buffer[3] = BIAS_AND_CLAMP[(((rd3 ~mod+ rk7) ~mod+ (1 << 17)) >> 18) &
args.dst_buffer[4] = BIAS_AND_CLAMP[(((rd3 ~mod- rk7) ~mod+ (1 << 17)) >> 18) & 1023]
$advance$
}
`