std/jpeg: optimize EXTEND'ing MCU bits

name                                             old speed     new speed     delta

wuffs_jpeg_decode_19k_8bpp/clang11               224MB/s ± 1%  223MB/s ± 0%     ~     (p=0.190 n=4+5)
wuffs_jpeg_decode_30k_24bpp_progressive/clang11  206MB/s ± 0%  208MB/s ± 0%   +0.86%  (p=0.016 n=4+5)
wuffs_jpeg_decode_30k_24bpp_sequential/clang11   374MB/s ± 0%  375MB/s ± 0%   +0.26%  (p=0.016 n=5+4)
wuffs_jpeg_decode_77k_24bpp/clang11              319MB/s ± 0%  319MB/s ± 0%   +0.16%  (p=0.029 n=4+4)
wuffs_jpeg_decode_552k_24bpp_420/clang11         329MB/s ± 0%  329MB/s ± 0%   -0.06%  (p=0.008 n=5+5)
wuffs_jpeg_decode_552k_24bpp_444/clang11         465MB/s ± 0%  465MB/s ± 0%   -0.12%  (p=0.008 n=5+5)
wuffs_jpeg_decode_4002k_24bpp/clang11            332MB/s ± 0%  331MB/s ± 0%   -0.15%  (p=0.008 n=5+5)

wuffs_jpeg_decode_19k_8bpp/gcc10                 176MB/s ± 0%  208MB/s ± 0%  +18.26%  (p=0.008 n=5+5)
wuffs_jpeg_decode_30k_24bpp_progressive/gcc10    246MB/s ± 0%  243MB/s ± 0%   -1.03%  (p=0.008 n=5+5)
wuffs_jpeg_decode_30k_24bpp_sequential/gcc10     472MB/s ± 0%  479MB/s ± 0%   +1.33%  (p=0.008 n=5+5)
wuffs_jpeg_decode_77k_24bpp/gcc10                334MB/s ± 0%  385MB/s ± 0%  +15.33%  (p=0.008 n=5+5)
wuffs_jpeg_decode_552k_24bpp_420/gcc10           351MB/s ± 0%  407MB/s ± 0%  +15.74%  (p=0.008 n=5+5)
wuffs_jpeg_decode_552k_24bpp_444/gcc10           425MB/s ± 0%  457MB/s ± 0%   +7.38%  (p=0.008 n=5+5)
wuffs_jpeg_decode_4002k_24bpp/gcc10              349MB/s ± 0%  411MB/s ± 0%  +17.74%  (p=0.008 n=5+5)
This commit is contained in:
Nigel Tao committed 2023-06-18 17:37:23 +10:00
1 parent 143991a301
commit 44e745dd0e
6 files changed
+23 -24

No files matched your search

+3
View File
@@ -178,6 +178,9 @@ wuffs_base__u64__sat_sub_indirect(uint64_t* x, uint64_t y) {
#define wuffs_base__utility__sign_extend_rshift_u32(a, n) \
((uint32_t)(((int32_t)(a)) >> (n)))
#define wuffs_base__utility__sign_extend_rshift_u64(a, n) \
((uint64_t)(((int64_t)(a)) >> (n)))
// ---------------- Slices and Tables
// wuffs_base__slice_u8__prefix returns up to the first up_to bytes of s.
+1
View File
@@ -388,6 +388,7 @@ var funcsOther = [...]string{
"min_incl_x: u32, min_incl_y: u32, max_excl_x: u32, max_excl_y: u32) rect_ie_u32",
"utility.sign_extend_convert_u16_u32(a: u16) u32",
"utility.sign_extend_rshift_u32(a: u32, n: u32[..= 31]) u32",
"utility.sign_extend_rshift_u64(a: u64, n: u32[..= 63]) u64",
// ---- ranges
+7 -12
View File
@@ -12274,6 +12274,9 @@ wuffs_base__u64__sat_sub_indirect(uint64_t* x, uint64_t y) {
#define wuffs_base__utility__sign_extend_rshift_u32(a, n) \
((uint32_t)(((int32_t)(a)) >> (n)))
#define wuffs_base__utility__sign_extend_rshift_u64(a, n) \
((uint64_t)(((int64_t)(a)) >> (n)))
// ---------------- Slices and Tables
// wuffs_base__slice_u8__prefix returns up to the first up_to bytes of s.
@@ -41664,9 +41667,7 @@ wuffs_jpeg__decoder__decode_mcu__choosy_default(
label__0__break:;
}
v_dc_value = ((uint32_t)(((v_bits >> 32u) >> (32u - v_dc_symbol))));
if ((v_bits >> 63u) == 0u) {
v_dc_value += ((uint32_t)(1u + ((uint32_t)(((uint32_t)(4294967295u)) << v_dc_symbol))));
}
v_dc_value += ((((uint32_t)(wuffs_base__utility__sign_extend_rshift_u64(v_bits, 63u))) ^ 4294967295u) & ((uint32_t)(1u + ((uint32_t)(((uint32_t)(4294967295u)) << v_dc_symbol)))));
v_bits <<= v_dc_symbol;
v_n_bits -= v_dc_symbol;
v_csel = self->private_impl.f_scan_comps_cselector[self->private_impl.f_mcu_blocks_sselector[self->private_impl.f_mcu_current_block]];
@@ -41722,9 +41723,7 @@ wuffs_jpeg__decoder__decode_mcu__choosy_default(
v_ac_ssss = (v_ac_symbol & 15u);
if (v_ac_ssss > 0u) {
v_ac_value = ((uint32_t)((v_bits >> (64u - v_ac_ssss))));
if ((v_bits >> 63u) == 0u) {
v_ac_value += ((uint32_t)(1u + ((uint32_t)(((uint32_t)(4294967295u)) << v_ac_ssss))));
}
v_ac_value += ((((uint32_t)(wuffs_base__utility__sign_extend_rshift_u64(v_bits, 63u))) ^ 4294967295u) & ((uint32_t)(1u + ((uint32_t)(((uint32_t)(4294967295u)) << v_ac_ssss)))));
v_bits <<= v_ac_ssss;
v_n_bits -= v_ac_ssss;
v_z = (self->private_impl.f_mcu_zig_index + v_ac_rrrr);
@@ -41887,9 +41886,7 @@ wuffs_jpeg__decoder__decode_mcu_progressive_ac_high_bits(
v_ac_ssss = (v_ac_symbol & 15u);
if (v_ac_ssss > 0u) {
v_ac_value = ((uint32_t)((v_bits >> (64u - v_ac_ssss))));
if ((v_bits >> 63u) == 0u) {
v_ac_value += ((uint32_t)(1u + ((uint32_t)(((uint32_t)(4294967295u)) << v_ac_ssss))));
}
v_ac_value += ((((uint32_t)(wuffs_base__utility__sign_extend_rshift_u64(v_bits, 63u))) ^ 4294967295u) & ((uint32_t)(1u + ((uint32_t)(((uint32_t)(4294967295u)) << v_ac_ssss)))));
v_bits <<= v_ac_ssss;
v_n_bits -= v_ac_ssss;
v_z = (self->private_impl.f_mcu_zig_index + v_ac_rrrr);
@@ -42281,9 +42278,7 @@ wuffs_jpeg__decoder__decode_mcu_progressive_dc_high_bits(
label__0__break:;
}
v_dc_value = ((uint32_t)(((v_bits >> 32u) >> (32u - v_dc_symbol))));
if ((v_bits >> 63u) == 0u) {
v_dc_value += ((uint32_t)(1u + ((uint32_t)(((uint32_t)(4294967295u)) << v_dc_symbol))));
}
v_dc_value += ((((uint32_t)(wuffs_base__utility__sign_extend_rshift_u64(v_bits, 63u))) ^ 4294967295u) & ((uint32_t)(1u + ((uint32_t)(((uint32_t)(4294967295u)) << v_dc_symbol)))));
v_bits <<= v_dc_symbol;
v_n_bits -= v_dc_symbol;
v_csel = self->private_impl.f_scan_comps_cselector[self->private_impl.f_mcu_blocks_sselector[self->private_impl.f_mcu_current_block]];
+6 -6
View File
@@ -119,9 +119,9 @@ pri func decoder.decode_mcu!(workbuf: slice base.u8, mx: base.u32[..= 0x1FFF], m
// The dc_value is shifted by (64 - dc_symbol) in two steps,
// because we want to shift by 64 (not 0) when dc_symbol is 0.
dc_value = (((bits >> 32) >> (32 - dc_symbol)) & 0xFFFF_FFFF) as base.u32
if (bits >> 63) == 0 { // EXTEND per section F.2.2.1.
dc_value ~mod+= 1 ~mod+ ((0xFFFF_FFFF as base.u32) ~mod<< dc_symbol)
}
dc_value ~mod+= // EXTEND per section F.2.2.1.
(((this.util.sign_extend_rshift_u64(a: bits, n: 63) & 0xFFFF_FFFF) as base.u32) ^ 0xFFFF_FFFF) &
(1 ~mod+ ((0xFFFF_FFFF as base.u32) ~mod<< dc_symbol))
bits ~mod<<= dc_symbol
n_bits ~mod-= dc_symbol
csel = this.scan_comps_cselector[this.mcu_blocks_sselector[this.mcu_current_block]]
@@ -202,9 +202,9 @@ pri func decoder.decode_mcu!(workbuf: slice base.u8, mx: base.u32[..= 0x1FFF], m
// Process the ac_value in the next ac_ssss (up to 15) bits.
if ac_ssss > 0 {
ac_value = ((bits >> (64 - ac_ssss)) & 0xFFFF_FFFF) as base.u32
if (bits >> 63) == 0 { // EXTEND per section F.2.2.1.
ac_value ~mod+= 1 ~mod+ ((0xFFFF_FFFF as base.u32) ~mod<< ac_ssss)
}
ac_value ~mod+= // EXTEND per section F.2.2.1.
(((this.util.sign_extend_rshift_u64(a: bits, n: 63) & 0xFFFF_FFFF) as base.u32) ^ 0xFFFF_FFFF) &
(1 ~mod+ ((0xFFFF_FFFF as base.u32) ~mod<< ac_ssss))
bits ~mod<<= ac_ssss
n_bits ~mod-= ac_ssss
z = this.mcu_zig_index + ac_rrrr
@@ -120,9 +120,9 @@ pri func decoder.decode_mcu_progressive_ac_high_bits!(workbuf: slice base.u8, mx
// the eob_run in the next ac_rrrr (up to 15) bits.
if ac_ssss > 0 {
ac_value = ((bits >> (64 - ac_ssss)) & 0xFFFF_FFFF) as base.u32
if (bits >> 63) == 0 { // EXTEND per section F.2.2.1.
ac_value ~mod+= 1 ~mod+ ((0xFFFF_FFFF as base.u32) ~mod<< ac_ssss)
}
ac_value ~mod+= // EXTEND per section F.2.2.1.
(((this.util.sign_extend_rshift_u64(a: bits, n: 63) & 0xFFFF_FFFF) as base.u32) ^ 0xFFFF_FFFF) &
(1 ~mod+ ((0xFFFF_FFFF as base.u32) ~mod<< ac_ssss))
bits ~mod<<= ac_ssss
n_bits ~mod-= ac_ssss
z = this.mcu_zig_index + ac_rrrr
@@ -92,9 +92,9 @@ pri func decoder.decode_mcu_progressive_dc_high_bits!(workbuf: slice base.u8, mx
// The dc_value is shifted by (64 - dc_symbol) in two steps,
// because we want to shift by 64 (not 0) when dc_symbol is 0.
dc_value = (((bits >> 32) >> (32 - dc_symbol)) & 0xFFFF_FFFF) as base.u32
if (bits >> 63) == 0 { // EXTEND per section F.2.2.1.
dc_value ~mod+= 1 ~mod+ ((0xFFFF_FFFF as base.u32) ~mod<< dc_symbol)
}
dc_value ~mod+= // EXTEND per section F.2.2.1.
(((this.util.sign_extend_rshift_u64(a: bits, n: 63) & 0xFFFF_FFFF) as base.u32) ^ 0xFFFF_FFFF) &
(1 ~mod+ ((0xFFFF_FFFF as base.u32) ~mod<< dc_symbol))
bits ~mod<<= dc_symbol
n_bits ~mod-= dc_symbol
csel = this.scan_comps_cselector[this.mcu_blocks_sselector[this.mcu_current_block]]