std/jpeg: tighten EXTEND element type s/u32/u16/

I'm not sure of the underlying reason for the gcc benchmark anomaly (the
clang numbers don't change).

name                                             old speed     new speed     delta

wuffs_jpeg_decode_19k_8bpp/clang11               227MB/s ± 0%  227MB/s ± 0%     ~     (p=0.151 n=5+5)
wuffs_jpeg_decode_30k_24bpp_progressive/clang11  209MB/s ± 0%  213MB/s ± 0%   +1.62%  (p=0.008 n=5+5)
wuffs_jpeg_decode_30k_24bpp_sequential/clang11   379MB/s ± 0%  380MB/s ± 0%   +0.13%  (p=0.008 n=5+5)
wuffs_jpeg_decode_77k_24bpp/clang11              329MB/s ± 0%  329MB/s ± 0%   +0.06%  (p=0.008 n=5+5)
wuffs_jpeg_decode_552k_24bpp_420/clang11         342MB/s ± 0%  342MB/s ± 0%     ~     (p=0.690 n=5+5)
wuffs_jpeg_decode_552k_24bpp_444/clang11         468MB/s ± 0%  469MB/s ± 0%   +0.10%  (p=0.008 n=5+5)
wuffs_jpeg_decode_4002k_24bpp/clang11            348MB/s ± 0%  348MB/s ± 0%   +0.08%  (p=0.016 n=5+5)

wuffs_jpeg_decode_19k_8bpp/gcc10                 213MB/s ± 0%  213MB/s ± 0%     ~     (p=0.151 n=5+5)
wuffs_jpeg_decode_30k_24bpp_progressive/gcc10    221MB/s ± 0%  246MB/s ± 0%  +11.51%  (p=0.008 n=5+5)
wuffs_jpeg_decode_30k_24bpp_sequential/gcc10     487MB/s ± 0%  482MB/s ± 0%   -0.98%  (p=0.008 n=5+5)
wuffs_jpeg_decode_77k_24bpp/gcc10                392MB/s ± 0%  392MB/s ± 0%     ~     (p=1.000 n=5+5)
wuffs_jpeg_decode_552k_24bpp_420/gcc10           413MB/s ± 0%  412MB/s ± 0%     ~     (p=0.190 n=4+5)
wuffs_jpeg_decode_552k_24bpp_444/gcc10           459MB/s ± 0%  460MB/s ± 0%   +0.09%  (p=0.008 n=5+5)
wuffs_jpeg_decode_4002k_24bpp/gcc10              419MB/s ± 0%  419MB/s ± 0%   +0.04%  (p=0.008 n=5+5)
This commit is contained in:
Nigel Tao committed 2023-06-18 21:35:49 +10:00
1 parent e4902dbd45
commit 36323abf87
6 files changed
+59 -59

No files matched your search

+27 -27
View File
@@ -37515,10 +37515,10 @@ WUFFS_JPEG__BIAS_AND_CLAMP[1024] WUFFS_BASE__POTENTIALLY_UNUSED = {
120, 121, 122, 123, 124, 125, 126, 127,
};
static const uint32_t
static const uint16_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,
0, 65535, 65533, 65529, 65521, 65505, 65473, 65409,
65281, 65025, 64513, 63489, 61441, 57345, 49153, 32769,
};
// ---------------- Private Initializer Prototypes
@@ -41594,8 +41594,8 @@ wuffs_jpeg__decoder__decode_mcu__choosy_default(
uint32_t v_dc_code = 0;
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;
uint16_t v_dc_value = 0;
uint16_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;
@@ -41604,8 +41604,8 @@ wuffs_jpeg__decoder__decode_mcu__choosy_default(
uint32_t v_ac_code = 0;
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;
uint16_t v_ac_value = 0;
uint16_t v_ac_extend = 0;
uint32_t v_ac_rrrr = 0;
uint32_t v_ac_ssss = 0;
uint32_t v_z = 0;
@@ -41677,12 +41677,12 @@ wuffs_jpeg__decoder__decode_mcu__choosy_default(
}
label__0__break:;
}
v_dc_value = ((uint32_t)(((v_bits >> 32u) >> (32u - v_dc_symbol))));
v_dc_value += (v_dc_extend & (((uint32_t)(wuffs_base__utility__sign_extend_rshift_u64(v_bits, 63u))) ^ 4294967295u));
v_dc_value = ((uint16_t)(((v_bits >> 32u) >> (32u - v_dc_symbol))));
v_dc_value += (v_dc_extend & (((uint16_t)(wuffs_base__utility__sign_extend_rshift_u64(v_bits, 63u))) ^ 65535u));
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]];
self->private_impl.f_mcu_previous_dc_values[v_csel] += ((uint16_t)(v_dc_value));
self->private_impl.f_mcu_previous_dc_values[v_csel] += v_dc_value;
self->private_data.f_mcu_blocks[0u][0u] = self->private_impl.f_mcu_previous_dc_values[v_csel];
self->private_impl.f_mcu_zig_index = 1u;
goto label__dc_component__break;
@@ -41736,12 +41736,12 @@ wuffs_jpeg__decoder__decode_mcu__choosy_default(
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 += (v_ac_extend & (((uint32_t)(wuffs_base__utility__sign_extend_rshift_u64(v_bits, 63u))) ^ 4294967295u));
v_ac_value = ((uint16_t)((v_bits >> (64u - v_ac_ssss))));
v_ac_value += (v_ac_extend & (((uint16_t)(wuffs_base__utility__sign_extend_rshift_u64(v_bits, 63u))) ^ 65535u));
v_bits <<= v_ac_ssss;
v_n_bits -= v_ac_ssss;
v_z += (v_ac_rrrr + 1u);
self->private_data.f_mcu_blocks[0u][WUFFS_JPEG__UNZIG[v_z]] = ((uint16_t)(v_ac_value));
self->private_data.f_mcu_blocks[0u][WUFFS_JPEG__UNZIG[v_z]] = v_ac_value;
} else if (v_ac_rrrr < 15u) {
goto label__ac_components__break;
} else {
@@ -41808,8 +41808,8 @@ wuffs_jpeg__decoder__decode_mcu_progressive_ac_high_bits(
uint32_t v_ac_code = 0;
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;
uint16_t v_ac_value = 0;
uint16_t v_ac_extend = 0;
uint32_t v_ac_rrrr = 0;
uint32_t v_ac_ssss = 0;
uint32_t v_z = 0;
@@ -41896,12 +41896,12 @@ wuffs_jpeg__decoder__decode_mcu_progressive_ac_high_bits(
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 += (v_ac_extend & (((uint32_t)(wuffs_base__utility__sign_extend_rshift_u64(v_bits, 63u))) ^ 4294967295u));
v_ac_value = ((uint16_t)((v_bits >> (64u - v_ac_ssss))));
v_ac_value += (v_ac_extend & (((uint16_t)(wuffs_base__utility__sign_extend_rshift_u64(v_bits, 63u))) ^ 65535u));
v_bits <<= v_ac_ssss;
v_n_bits -= v_ac_ssss;
v_z += (v_ac_rrrr + 1u);
self->private_data.f_mcu_blocks[0u][WUFFS_JPEG__UNZIG[v_z]] = ((uint16_t)(((uint32_t)(v_ac_value << self->private_impl.f_scan_al))));
self->private_data.f_mcu_blocks[0u][WUFFS_JPEG__UNZIG[v_z]] = ((uint16_t)(((uint16_t)(v_ac_value << self->private_impl.f_scan_al))));
} else if (v_ac_rrrr < 15u) {
self->private_impl.f_eob_run = ((uint16_t)(((((uint16_t)(1u)) << v_ac_rrrr) - 1u)));
if (v_ac_rrrr > 0u) {
@@ -41965,7 +41965,7 @@ wuffs_jpeg__decoder__decode_mcu_progressive_ac_low_bit(
uint32_t v_ac_code = 0;
uint32_t v_ac_blm1 = 0;
uint32_t v_ac_ht_slow = 0;
uint32_t v_ac_value = 0;
uint16_t v_ac_value = 0;
uint32_t v_ac_rrrr = 0;
uint32_t v_ac_ssss = 0;
uint8_t v_unzig = 0;
@@ -42041,9 +42041,9 @@ wuffs_jpeg__decoder__decode_mcu_progressive_ac_low_bit(
v_ac_ssss = (v_ac_symbol & 15u);
v_ac_value = 0u;
if (v_ac_ssss > 0u) {
v_ac_value = (((uint32_t)(1u)) << self->private_impl.f_scan_al);
v_ac_value = (((uint16_t)(1u)) << self->private_impl.f_scan_al);
if ((v_bits >> 63u) == 0u) {
v_ac_value = ((uint32_t)(((uint32_t)(4294967295u)) << self->private_impl.f_scan_al));
v_ac_value = ((uint16_t)(((uint16_t)(65535u)) << self->private_impl.f_scan_al));
}
v_bits <<= 1u;
v_n_bits -= 1u;
@@ -42104,7 +42104,7 @@ wuffs_jpeg__decoder__decode_mcu_progressive_ac_low_bit(
}
label__refine_non_zeroes__break:;
if (v_ac_value != 0u) {
self->private_data.f_mcu_blocks[0u][WUFFS_JPEG__UNZIG[(1u + self->private_impl.f_mcu_zig_index)]] = ((uint16_t)(v_ac_value));
self->private_data.f_mcu_blocks[0u][WUFFS_JPEG__UNZIG[(1u + self->private_impl.f_mcu_zig_index)]] = v_ac_value;
}
if (self->private_impl.f_mcu_zig_index >= ((uint32_t)(self->private_impl.f_scan_se))) {
goto label__ac_components__break;
@@ -42219,8 +42219,8 @@ wuffs_jpeg__decoder__decode_mcu_progressive_dc_high_bits(
uint32_t v_dc_code = 0;
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;
uint16_t v_dc_value = 0;
uint16_t v_dc_extend = 0;
v_bits = self->private_impl.f_bitstream_bits;
v_n_bits = self->private_impl.f_bitstream_n_bits;
@@ -42285,12 +42285,12 @@ wuffs_jpeg__decoder__decode_mcu_progressive_dc_high_bits(
}
label__0__break:;
}
v_dc_value = ((uint32_t)(((v_bits >> 32u) >> (32u - v_dc_symbol))));
v_dc_value += (v_dc_extend & (((uint32_t)(wuffs_base__utility__sign_extend_rshift_u64(v_bits, 63u))) ^ 4294967295u));
v_dc_value = ((uint16_t)(((v_bits >> 32u) >> (32u - v_dc_symbol))));
v_dc_value += (v_dc_extend & (((uint16_t)(wuffs_base__utility__sign_extend_rshift_u64(v_bits, 63u))) ^ 65535u));
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]];
self->private_impl.f_mcu_previous_dc_values[v_csel] += ((uint16_t)(v_dc_value));
self->private_impl.f_mcu_previous_dc_values[v_csel] += v_dc_value;
self->private_data.f_mcu_blocks[self->private_impl.f_mcu_current_block][0u] = ((uint16_t)(self->private_impl.f_mcu_previous_dc_values[v_csel] << self->private_impl.f_scan_al));
goto label__dc_component__break;
}
+9 -9
View File
@@ -123,18 +123,18 @@ pri const BIAS_AND_CLAMP : roarray[1024] base.u8 = [
// )
//
// func main() {
// for i := 0; i < 16; i++ {
// for t := 0; t < 16; t++ {
// // "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 {
// v := (uint16(0xFFFF) << t) + 1
// fmt.Printf("0x%04X,", v)
// if (t & 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,
pri const EXTEND : roarray[16] base.u16 = [
0x0000, 0xFFFF, 0xFFFD, 0xFFF9,
0xFFF1, 0xFFE1, 0xFFC1, 0xFF81,
0xFF01, 0xFE01, 0xFC01, 0xF801,
0xF001, 0xE001, 0xC001, 0x8001,
]
+10 -10
View File
@@ -35,8 +35,8 @@ pri func decoder.decode_mcu!(workbuf: slice base.u8, mx: base.u32[..= 0x1FFF], m
var dc_code : base.u32
var dc_blm1 : base.u32[..= 15]
var dc_ht_slow : base.u32
var dc_value : base.u32
var dc_extend : base.u32
var dc_value : base.u16
var dc_extend : base.u16
var ac_huff_table_fast : nptr roarray[256] base.u16
@@ -47,8 +47,8 @@ pri func decoder.decode_mcu!(workbuf: slice base.u8, mx: base.u32[..= 0x1FFF], m
var ac_code : base.u32
var ac_blm1 : base.u32[..= 15]
var ac_ht_slow : base.u32
var ac_value : base.u32
var ac_extend : base.u32
var ac_value : base.u16
var ac_extend : base.u16
var ac_rrrr : base.u32[..= 15]
var ac_ssss : base.u32[..= 15]
@@ -123,13 +123,13 @@ 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 = (((bits >> 32) >> (32 - dc_symbol)) & 0xFFFF) as base.u16
dc_value ~mod+= dc_extend &
(((this.util.sign_extend_rshift_u64(a: bits, n: 63) & 0xFFFF_FFFF) as base.u32) ^ 0xFFFF_FFFF)
(((this.util.sign_extend_rshift_u64(a: bits, n: 63) & 0xFFFF) as base.u16) ^ 0xFFFF)
bits ~mod<<= dc_symbol
n_bits ~mod-= dc_symbol
csel = this.scan_comps_cselector[this.mcu_blocks_sselector[this.mcu_current_block]]
this.mcu_previous_dc_values[csel] ~mod+= (dc_value & 0xFFFF) as base.u16
this.mcu_previous_dc_values[csel] ~mod+= dc_value
this.mcu_blocks[0][0] =
this.mcu_previous_dc_values[csel]
@@ -209,14 +209,14 @@ 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
ac_value = ((bits >> (64 - ac_ssss)) & 0xFFFF) as base.u16
ac_value ~mod+= ac_extend &
(((this.util.sign_extend_rshift_u64(a: bits, n: 63) & 0xFFFF_FFFF) as base.u32) ^ 0xFFFF_FFFF)
(((this.util.sign_extend_rshift_u64(a: bits, n: 63) & 0xFFFF) as base.u16) ^ 0xFFFF)
bits ~mod<<= ac_ssss
n_bits ~mod-= ac_ssss
z += ac_rrrr + 1
this.mcu_blocks[0][UNZIG[z]] =
(ac_value & 0xFFFF) as base.u16
ac_value
} else if ac_rrrr < 15 {
break.ac_components
} else {
@@ -30,8 +30,8 @@ pri func decoder.decode_mcu_progressive_ac_high_bits!(workbuf: slice base.u8, mx
var ac_code : base.u32
var ac_blm1 : base.u32[..= 15]
var ac_ht_slow : base.u32
var ac_value : base.u32
var ac_extend : base.u32
var ac_value : base.u16
var ac_extend : base.u16
var ac_rrrr : base.u32[..= 15]
var ac_ssss : base.u32[..= 15]
@@ -126,9 +126,9 @@ pri func decoder.decode_mcu_progressive_ac_high_bits!(workbuf: slice base.u8, mx
// 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 = ((bits >> (64 - ac_ssss)) & 0xFFFF) as base.u16
ac_value ~mod+= ac_extend &
(((this.util.sign_extend_rshift_u64(a: bits, n: 63) & 0xFFFF_FFFF) as base.u32) ^ 0xFFFF_FFFF)
(((this.util.sign_extend_rshift_u64(a: bits, n: 63) & 0xFFFF) as base.u16) ^ 0xFFFF)
bits ~mod<<= ac_ssss
n_bits ~mod-= ac_ssss
z += ac_rrrr + 1
@@ -32,7 +32,7 @@ pri func decoder.decode_mcu_progressive_ac_low_bit!(workbuf: slice base.u8, mx:
var ac_code : base.u32
var ac_blm1 : base.u32[..= 15]
var ac_ht_slow : base.u32
var ac_value : base.u32
var ac_value : base.u16
var ac_rrrr : base.u32[..= 15]
var ac_ssss : base.u32[..= 15]
@@ -124,9 +124,9 @@ pri func decoder.decode_mcu_progressive_ac_low_bit!(workbuf: slice base.u8, mx:
// the eob_run in the next ac_rrrr (up to 15) bits.
ac_value = 0
if ac_ssss > 0 {
ac_value = (0x0000_0001 as base.u32) << this.scan_al
ac_value = (0x0001 as base.u16) << this.scan_al
if (bits >> 63) == 0 {
ac_value = (0xFFFF_FFFF as base.u32) ~mod<< this.scan_al
ac_value = (0xFFFF as base.u16) ~mod<< this.scan_al
}
bits ~mod<<= 1
n_bits ~mod-= 1
@@ -179,7 +179,7 @@ pri func decoder.decode_mcu_progressive_ac_low_bit!(workbuf: slice base.u8, mx:
} endwhile.refine_non_zeroes
if ac_value <> 0 {
this.mcu_blocks[0][UNZIG[1 + this.mcu_zig_index]] = (ac_value & 0xFFFF) as base.u16
this.mcu_blocks[0][UNZIG[1 + this.mcu_zig_index]] = ac_value
}
if this.mcu_zig_index >= (this.scan_se as base.u32) {
@@ -30,8 +30,8 @@ pri func decoder.decode_mcu_progressive_dc_high_bits!(workbuf: slice base.u8, mx
var dc_code : base.u32
var dc_blm1 : base.u32[..= 15]
var dc_ht_slow : base.u32
var dc_value : base.u32
var dc_extend : base.u32
var dc_value : base.u16
var dc_extend : base.u16
bits = this.bitstream_bits
n_bits = this.bitstream_n_bits
@@ -95,13 +95,13 @@ 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 = (((bits >> 32) >> (32 - dc_symbol)) & 0xFFFF) as base.u16
dc_value ~mod+= dc_extend &
(((this.util.sign_extend_rshift_u64(a: bits, n: 63) & 0xFFFF_FFFF) as base.u32) ^ 0xFFFF_FFFF)
(((this.util.sign_extend_rshift_u64(a: bits, n: 63) & 0xFFFF) as base.u16) ^ 0xFFFF)
bits ~mod<<= dc_symbol
n_bits ~mod-= dc_symbol
csel = this.scan_comps_cselector[this.mcu_blocks_sselector[this.mcu_current_block]]
this.mcu_previous_dc_values[csel] ~mod+= (dc_value & 0xFFFF) as base.u16
this.mcu_previous_dc_values[csel] ~mod+= dc_value
this.mcu_blocks[this.mcu_current_block][0] =
this.mcu_previous_dc_values[csel] ~mod<< this.scan_al