std/jpeg: add an EXTEND look-up table

name                                             old speed     new speed     delta

wuffs_jpeg_decode_19k_8bpp/clang11               223MB/s ± 0%  222MB/s ± 0%  -0.81%  (p=0.008 n=5+5)
wuffs_jpeg_decode_30k_24bpp_progressive/clang11  208MB/s ± 0%  213MB/s ± 0%  +2.32%  (p=0.008 n=5+5)
wuffs_jpeg_decode_30k_24bpp_sequential/clang11   375MB/s ± 0%  376MB/s ± 0%  +0.34%  (p=0.016 n=4+5)
wuffs_jpeg_decode_77k_24bpp/clang11              319MB/s ± 0%  325MB/s ± 0%  +1.92%  (p=0.016 n=4+5)
wuffs_jpeg_decode_552k_24bpp_420/clang11         329MB/s ± 0%  339MB/s ± 0%  +3.00%  (p=0.008 n=5+5)
wuffs_jpeg_decode_552k_24bpp_444/clang11         465MB/s ± 0%  468MB/s ± 0%  +0.65%  (p=0.008 n=5+5)
wuffs_jpeg_decode_4002k_24bpp/clang11            331MB/s ± 0%  344MB/s ± 0%  +4.02%  (p=0.008 n=5+5)

wuffs_jpeg_decode_19k_8bpp/gcc10                 208MB/s ± 0%  212MB/s ± 0%  +1.76%  (p=0.008 n=5+5)
wuffs_jpeg_decode_30k_24bpp_progressive/gcc10    243MB/s ± 0%  246MB/s ± 0%  +1.09%  (p=0.008 n=5+5)
wuffs_jpeg_decode_30k_24bpp_sequential/gcc10     479MB/s ± 0%  484MB/s ± 0%  +1.17%  (p=0.008 n=5+5)
wuffs_jpeg_decode_77k_24bpp/gcc10                385MB/s ± 0%  390MB/s ± 0%  +1.40%  (p=0.008 n=5+5)
wuffs_jpeg_decode_552k_24bpp_420/gcc10           407MB/s ± 0%  415MB/s ± 0%  +1.98%  (p=0.008 n=5+5)
wuffs_jpeg_decode_552k_24bpp_444/gcc10           457MB/s ± 0%  463MB/s ± 0%  +1.29%  (p=0.008 n=5+5)
wuffs_jpeg_decode_4002k_24bpp/gcc10              411MB/s ± 0%  420MB/s ± 0%  +2.06%  (p=0.008 n=5+5)
This commit is contained in:
Nigel Tao committed 2023-06-18 18:44:41 +10:00
1 parent 44e745dd0e
commit 19e5cccee3
5 files changed
+72 -16

No files matched your search

+22 -4
View File
@@ -37515,6 +37515,12 @@ WUFFS_JPEG__BIAS_AND_CLAMP[1024] WUFFS_BASE__POTENTIALLY_UNUSED = {
120, 121, 122, 123, 124, 125, 126, 127,
};
static const uint32_t
WUFFS_JPEG__EXTEND[16] WUFFS_BASE__POTENTIALLY_UNUSED = {
0, 4294967295, 4294967293, 4294967289, 4294967281, 4294967265, 4294967233, 4294967169,
4294967041, 4294966785, 4294966273, 4294965249, 4294963201, 4294959105, 4294950913, 4294934529,
};
// ---------------- Private Initializer Prototypes
// ---------------- Private Function Prototypes
@@ -41589,6 +41595,7 @@ wuffs_jpeg__decoder__decode_mcu__choosy_default(
uint32_t v_dc_blm1 = 0;
uint32_t v_dc_ht_slow = 0;
uint32_t v_dc_value = 0;
uint32_t v_dc_extend = 0;
const uint16_t* v_ac_huff_table_fast = NULL;
uint8_t v_ac_h = 0;
uint32_t v_ac_symbol = 0;
@@ -41598,6 +41605,7 @@ wuffs_jpeg__decoder__decode_mcu__choosy_default(
uint32_t v_ac_blm1 = 0;
uint32_t v_ac_ht_slow = 0;
uint32_t v_ac_value = 0;
uint32_t v_ac_extend = 0;
uint32_t v_ac_rrrr = 0;
uint32_t v_ac_ssss = 0;
uint32_t v_z = 0;
@@ -41642,6 +41650,7 @@ wuffs_jpeg__decoder__decode_mcu__choosy_default(
v_dc_bl = (v_dc_ht_fast >> 8u);
if (v_n_bits >= v_dc_bl) {
v_dc_symbol = (15u & v_dc_ht_fast);
v_dc_extend = WUFFS_JPEG__EXTEND[v_dc_symbol];
v_bits <<= (v_dc_bl & 63u);
v_n_bits -= v_dc_bl;
} else {
@@ -41653,6 +41662,7 @@ wuffs_jpeg__decoder__decode_mcu__choosy_default(
v_dc_ht_slow = self->private_impl.f_huff_tables_slow[v_dc_h][v_dc_blm1];
if (v_dc_code < (v_dc_ht_slow >> 8u)) {
v_dc_symbol = (15u & ((uint32_t)(self->private_impl.f_huff_tables_symbols[v_dc_h][(255u & ((uint32_t)(v_dc_code + v_dc_ht_slow)))])));
v_dc_extend = WUFFS_JPEG__EXTEND[v_dc_symbol];
goto label__0__break;
}
v_dc_code = (((uint32_t)(v_dc_code << 1u)) | ((uint32_t)((v_bits >> 63u))));
@@ -41661,13 +41671,14 @@ wuffs_jpeg__decoder__decode_mcu__choosy_default(
v_dc_blm1 = ((v_dc_blm1 + 1u) & 15u);
if (v_dc_blm1 == 0u) {
v_dc_symbol = 0u;
v_dc_extend = WUFFS_JPEG__EXTEND[v_dc_symbol];
goto label__0__break;
}
}
label__0__break:;
}
v_dc_value = ((uint32_t)(((v_bits >> 32u) >> (32u - 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_dc_value += (v_dc_extend & (((uint32_t)(wuffs_base__utility__sign_extend_rshift_u64(v_bits, 63u))) ^ 4294967295u));
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]];
@@ -41721,9 +41732,10 @@ wuffs_jpeg__decoder__decode_mcu__choosy_default(
}
v_ac_rrrr = (v_ac_symbol >> 4u);
v_ac_ssss = (v_ac_symbol & 15u);
v_ac_extend = WUFFS_JPEG__EXTEND[v_ac_ssss];
if (v_ac_ssss > 0u) {
v_ac_value = ((uint32_t)((v_bits >> (64u - 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_ac_value += (v_ac_extend & (((uint32_t)(wuffs_base__utility__sign_extend_rshift_u64(v_bits, 63u))) ^ 4294967295u));
v_bits <<= v_ac_ssss;
v_n_bits -= v_ac_ssss;
v_z = (self->private_impl.f_mcu_zig_index + v_ac_rrrr);
@@ -41802,6 +41814,7 @@ wuffs_jpeg__decoder__decode_mcu_progressive_ac_high_bits(
uint32_t v_ac_blm1 = 0;
uint32_t v_ac_ht_slow = 0;
uint32_t v_ac_value = 0;
uint32_t v_ac_extend = 0;
uint32_t v_ac_rrrr = 0;
uint32_t v_ac_ssss = 0;
uint32_t v_z = 0;
@@ -41884,9 +41897,10 @@ wuffs_jpeg__decoder__decode_mcu_progressive_ac_high_bits(
}
v_ac_rrrr = (v_ac_symbol >> 4u);
v_ac_ssss = (v_ac_symbol & 15u);
v_ac_extend = WUFFS_JPEG__EXTEND[v_ac_ssss];
if (v_ac_ssss > 0u) {
v_ac_value = ((uint32_t)((v_bits >> (64u - 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_ac_value += (v_ac_extend & (((uint32_t)(wuffs_base__utility__sign_extend_rshift_u64(v_bits, 63u))) ^ 4294967295u));
v_bits <<= v_ac_ssss;
v_n_bits -= v_ac_ssss;
v_z = (self->private_impl.f_mcu_zig_index + v_ac_rrrr);
@@ -42216,6 +42230,7 @@ wuffs_jpeg__decoder__decode_mcu_progressive_dc_high_bits(
uint32_t v_dc_blm1 = 0;
uint32_t v_dc_ht_slow = 0;
uint32_t v_dc_value = 0;
uint32_t v_dc_extend = 0;
v_bits = self->private_impl.f_bitstream_bits;
v_n_bits = self->private_impl.f_bitstream_n_bits;
@@ -42253,6 +42268,7 @@ wuffs_jpeg__decoder__decode_mcu_progressive_dc_high_bits(
v_dc_bl = (v_dc_ht_fast >> 8u);
if (v_n_bits >= v_dc_bl) {
v_dc_symbol = (15u & v_dc_ht_fast);
v_dc_extend = WUFFS_JPEG__EXTEND[v_dc_symbol];
v_bits <<= (v_dc_bl & 63u);
v_n_bits -= v_dc_bl;
} else {
@@ -42264,6 +42280,7 @@ wuffs_jpeg__decoder__decode_mcu_progressive_dc_high_bits(
v_dc_ht_slow = self->private_impl.f_huff_tables_slow[v_dc_h][v_dc_blm1];
if (v_dc_code < (v_dc_ht_slow >> 8u)) {
v_dc_symbol = (15u & ((uint32_t)(self->private_impl.f_huff_tables_symbols[v_dc_h][(255u & ((uint32_t)(v_dc_code + v_dc_ht_slow)))])));
v_dc_extend = WUFFS_JPEG__EXTEND[v_dc_symbol];
goto label__0__break;
}
v_dc_code = (((uint32_t)(v_dc_code << 1u)) | ((uint32_t)((v_bits >> 63u))));
@@ -42272,13 +42289,14 @@ wuffs_jpeg__decoder__decode_mcu_progressive_dc_high_bits(
v_dc_blm1 = ((v_dc_blm1 + 1u) & 15u);
if (v_dc_blm1 == 0u) {
v_dc_symbol = 0u;
v_dc_extend = WUFFS_JPEG__EXTEND[v_dc_symbol];
goto label__0__break;
}
}
label__0__break:;
}
v_dc_value = ((uint32_t)(((v_bits >> 32u) >> (32u - 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_dc_value += (v_dc_extend & (((uint32_t)(wuffs_base__utility__sign_extend_rshift_u64(v_bits, 63u))) ^ 4294967295u));
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]];
+30
View File
@@ -106,3 +106,33 @@ pri const BIAS_AND_CLAMP : roarray[1024] base.u8 = [
0x60, 0x61, 0x62, 0x63, 0x64, 0x65, 0x66, 0x67, 0x68, 0x69, 0x6A, 0x6B, 0x6C, 0x6D, 0x6E, 0x6F,
0x70, 0x71, 0x72, 0x73, 0x74, 0x75, 0x76, 0x77, 0x78, 0x79, 0x7A, 0x7B, 0x7C, 0x7D, 0x7E, 0x7F,
]
// EXTEND is a look-up table that helps implement Figure F.12 "Extending the
// sign bit of a decoded value in V".
//
// It was generated by this Go code (and then wuffsfmt'ed).
//
// ----
//
// package main
//
// import (
// "fmt"
// )
//
// func main() {
// for i := 0; i < 16; i++ {
// // "V = (SLL –1 T) + 1" and SLL is "Shift Left Logical".
// v := (uint32(0xFFFF_FFFF) << t) + 1
// fmt.Printf("0x%08X,", v)
// if (i & 3) == 3 {
// fmt.Println()
// }
// }
// }
pri const EXTEND : roarray[16] base.u32 = [
0x0000_0000, 0xFFFF_FFFF, 0xFFFF_FFFD, 0xFFFF_FFF9,
0xFFFF_FFF1, 0xFFFF_FFE1, 0xFFFF_FFC1, 0xFFFF_FF81,
0xFFFF_FF01, 0xFFFF_FE01, 0xFFFF_FC01, 0xFFFF_F801,
0xFFFF_F001, 0xFFFF_E001, 0xFFFF_C001, 0xFFFF_8001,
]
+10 -6
View File
@@ -36,6 +36,7 @@ pri func decoder.decode_mcu!(workbuf: slice base.u8, mx: base.u32[..= 0x1FFF], m
var dc_blm1 : base.u32[..= 15]
var dc_ht_slow : base.u32
var dc_value : base.u32
var dc_extend : base.u32
var ac_huff_table_fast : nptr roarray[256] base.u16
@@ -47,6 +48,7 @@ pri func decoder.decode_mcu!(workbuf: slice base.u8, mx: base.u32[..= 0x1FFF], m
var ac_blm1 : base.u32[..= 15]
var ac_ht_slow : base.u32
var ac_value : base.u32
var ac_extend : base.u32
var ac_rrrr : base.u32[..= 15]
var ac_ssss : base.u32[..= 15]
@@ -88,6 +90,7 @@ pri func decoder.decode_mcu!(workbuf: slice base.u8, mx: base.u32[..= 0x1FFF], m
dc_bl = dc_ht_fast >> 8
if n_bits >= dc_bl {
dc_symbol = 0x0F & dc_ht_fast
dc_extend = EXTEND[dc_symbol]
bits ~mod<<= (dc_bl & 63)
n_bits -= dc_bl
} else {
@@ -101,6 +104,7 @@ pri func decoder.decode_mcu!(workbuf: slice base.u8, mx: base.u32[..= 0x1FFF], m
dc_ht_slow = this.huff_tables_slow[dc_h][dc_blm1]
if dc_code < (dc_ht_slow >> 8) {
dc_symbol = 0x0F & (this.huff_tables_symbols[dc_h][0xFF & (dc_code ~mod+ dc_ht_slow)] as base.u32)
dc_extend = EXTEND[dc_symbol]
break
}
dc_code = (dc_code ~mod<< 1) | ((bits >> 63) as base.u32)
@@ -109,6 +113,7 @@ pri func decoder.decode_mcu!(workbuf: slice base.u8, mx: base.u32[..= 0x1FFF], m
dc_blm1 = (dc_blm1 + 1) & 15
if dc_blm1 == 0 {
dc_symbol = 0
dc_extend = EXTEND[dc_symbol]
break
}
} endwhile
@@ -119,9 +124,8 @@ 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
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))
dc_value ~mod+= dc_extend &
(((this.util.sign_extend_rshift_u64(a: bits, n: 63) & 0xFFFF_FFFF) as base.u32) ^ 0xFFFF_FFFF)
bits ~mod<<= dc_symbol
n_bits ~mod-= dc_symbol
csel = this.scan_comps_cselector[this.mcu_blocks_sselector[this.mcu_current_block]]
@@ -198,13 +202,13 @@ pri func decoder.decode_mcu!(workbuf: slice base.u8, mx: base.u32[..= 0x1FFF], m
// F.2.2.2 "Decoding procedure for AC coefficients".
ac_rrrr = ac_symbol >> 4
ac_ssss = ac_symbol & 15
ac_extend = EXTEND[ac_ssss]
// 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
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))
ac_value ~mod+= ac_extend &
(((this.util.sign_extend_rshift_u64(a: bits, n: 63) & 0xFFFF_FFFF) as base.u32) ^ 0xFFFF_FFFF)
bits ~mod<<= ac_ssss
n_bits ~mod-= ac_ssss
z = this.mcu_zig_index + ac_rrrr
@@ -31,6 +31,7 @@ pri func decoder.decode_mcu_progressive_ac_high_bits!(workbuf: slice base.u8, mx
var ac_blm1 : base.u32[..= 15]
var ac_ht_slow : base.u32
var ac_value : base.u32
var ac_extend : base.u32
var ac_rrrr : base.u32[..= 15]
var ac_ssss : base.u32[..= 15]
@@ -115,14 +116,14 @@ pri func decoder.decode_mcu_progressive_ac_high_bits!(workbuf: slice base.u8, mx
// F.2.2.2 "Decoding procedure for AC coefficients".
ac_rrrr = ac_symbol >> 4
ac_ssss = ac_symbol & 15
ac_extend = EXTEND[ac_ssss]
// Process the ac_value in the next ac_ssss (up to 15) bits or
// 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
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))
ac_value ~mod+= ac_extend &
(((this.util.sign_extend_rshift_u64(a: bits, n: 63) & 0xFFFF_FFFF) as base.u32) ^ 0xFFFF_FFFF)
bits ~mod<<= ac_ssss
n_bits ~mod-= ac_ssss
z = this.mcu_zig_index + ac_rrrr
@@ -31,6 +31,7 @@ pri func decoder.decode_mcu_progressive_dc_high_bits!(workbuf: slice base.u8, mx
var dc_blm1 : base.u32[..= 15]
var dc_ht_slow : base.u32
var dc_value : base.u32
var dc_extend : base.u32
bits = this.bitstream_bits
n_bits = this.bitstream_n_bits
@@ -61,6 +62,7 @@ pri func decoder.decode_mcu_progressive_dc_high_bits!(workbuf: slice base.u8, mx
dc_bl = dc_ht_fast >> 8
if n_bits >= dc_bl {
dc_symbol = 0x0F & dc_ht_fast
dc_extend = EXTEND[dc_symbol]
bits ~mod<<= (dc_bl & 63)
n_bits -= dc_bl
} else {
@@ -74,6 +76,7 @@ pri func decoder.decode_mcu_progressive_dc_high_bits!(workbuf: slice base.u8, mx
dc_ht_slow = this.huff_tables_slow[dc_h][dc_blm1]
if dc_code < (dc_ht_slow >> 8) {
dc_symbol = 0x0F & (this.huff_tables_symbols[dc_h][0xFF & (dc_code ~mod+ dc_ht_slow)] as base.u32)
dc_extend = EXTEND[dc_symbol]
break
}
dc_code = (dc_code ~mod<< 1) | ((bits >> 63) as base.u32)
@@ -82,6 +85,7 @@ pri func decoder.decode_mcu_progressive_dc_high_bits!(workbuf: slice base.u8, mx
dc_blm1 = (dc_blm1 + 1) & 15
if dc_blm1 == 0 {
dc_symbol = 0
dc_extend = EXTEND[dc_symbol]
break
}
} endwhile
@@ -92,9 +96,8 @@ 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
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))
dc_value ~mod+= dc_extend &
(((this.util.sign_extend_rshift_u64(a: bits, n: 63) & 0xFFFF_FFFF) as base.u32) ^ 0xFFFF_FFFF)
bits ~mod<<= dc_symbol
n_bits ~mod-= dc_symbol
csel = this.scan_comps_cselector[this.mcu_blocks_sselector[this.mcu_current_block]]