std/jpeg: tweak UNZIG array

We'll hopefully claw back the performance in a follow-up commit.

name                                             old speed     new speed     delta

wuffs_jpeg_decode_19k_8bpp/clang11               222MB/s ± 0%  224MB/s ± 0%   +1.11%  (p=0.008 n=5+5)
wuffs_jpeg_decode_30k_24bpp_progressive/clang11  213MB/s ± 0%  210MB/s ± 0%   -1.53%  (p=0.008 n=5+5)
wuffs_jpeg_decode_30k_24bpp_sequential/clang11   376MB/s ± 0%  378MB/s ± 0%   +0.37%  (p=0.008 n=5+5)
wuffs_jpeg_decode_77k_24bpp/clang11              325MB/s ± 0%  328MB/s ± 0%   +0.66%  (p=0.008 n=5+5)
wuffs_jpeg_decode_552k_24bpp_420/clang11         339MB/s ± 0%  340MB/s ± 0%   +0.46%  (p=0.008 n=5+5)
wuffs_jpeg_decode_552k_24bpp_444/clang11         468MB/s ± 0%  469MB/s ± 0%   +0.27%  (p=0.008 n=5+5)
wuffs_jpeg_decode_4002k_24bpp/clang11            344MB/s ± 0%  346MB/s ± 0%   +0.37%  (p=0.008 n=5+5)

wuffs_jpeg_decode_19k_8bpp/gcc10                 212MB/s ± 0%  213MB/s ± 0%   +0.46%  (p=0.008 n=5+5)
wuffs_jpeg_decode_30k_24bpp_progressive/gcc10    246MB/s ± 0%  221MB/s ± 0%   -9.98%  (p=0.008 n=5+5)
wuffs_jpeg_decode_30k_24bpp_sequential/gcc10     484MB/s ± 0%  484MB/s ± 0%     ~     (p=0.056 n=5+5)
wuffs_jpeg_decode_77k_24bpp/gcc10                390MB/s ± 0%  392MB/s ± 0%   +0.47%  (p=0.008 n=5+5)
wuffs_jpeg_decode_552k_24bpp_420/gcc10           415MB/s ± 0%  414MB/s ± 0%   -0.17%  (p=0.008 n=5+5)
wuffs_jpeg_decode_552k_24bpp_444/gcc10           463MB/s ± 0%  459MB/s ± 0%   -0.78%  (p=0.008 n=5+5)
wuffs_jpeg_decode_4002k_24bpp/gcc10              420MB/s ± 0%  419MB/s ± 0%   -0.19%  (p=0.008 n=5+5)
This commit is contained in:
Nigel Tao committed 2023-06-18 20:57:50 +10:00
1 parent 19e5cccee3
commit 9b97c27ec6
6 files changed
+35 -33

No files matched your search

+18 -18
View File
@@ -37371,14 +37371,14 @@ const char wuffs_jpeg__error__internal_error_inconsistent_decoder_state[] = "#jp
static const uint8_t
WUFFS_JPEG__UNZIG[80] WUFFS_BASE__POTENTIALLY_UNUSED = {
0, 1, 8, 16, 9, 2, 3, 10,
17, 24, 32, 25, 18, 11, 4, 5,
12, 19, 26, 33, 40, 48, 41, 34,
27, 20, 13, 6, 7, 14, 21, 28,
35, 42, 49, 56, 57, 50, 43, 36,
29, 22, 15, 23, 30, 37, 44, 51,
58, 59, 52, 45, 38, 31, 39, 46,
53, 60, 61, 54, 47, 55, 62, 63,
0, 0, 1, 8, 16, 9, 2, 3,
10, 17, 24, 32, 25, 18, 11, 4,
5, 12, 19, 26, 33, 40, 48, 41,
34, 27, 20, 13, 6, 7, 14, 21,
28, 35, 42, 49, 56, 57, 50, 43,
36, 29, 22, 15, 23, 30, 37, 44,
51, 58, 59, 52, 45, 38, 31, 39,
46, 53, 60, 61, 54, 47, 55, 62,
63, 63, 63, 63, 63, 63, 63, 63,
63, 63, 63, 63, 63, 63, 63, 63,
};
@@ -39326,6 +39326,7 @@ wuffs_jpeg__decoder__decode_dqt(
self->private_impl.f_payload_length -= 64u;
v_i = 0u;
while (v_i < 64u) {
v_i += 1u;
{
WUFFS_BASE__COROUTINE_SUSPENSION_POINT(2);
if (WUFFS_BASE__UNLIKELY(iop_a_src == io2_a_src)) {
@@ -39335,7 +39336,6 @@ wuffs_jpeg__decoder__decode_dqt(
uint8_t t_1 = *iop_a_src++;
self->private_impl.f_quant_tables[v_q][WUFFS_JPEG__UNZIG[v_i]] = t_1;
}
v_i += 1u;
}
self->private_impl.f_seen_dqt[v_q] = true;
if (self->private_impl.f_sof_marker == 0u) {
@@ -41738,12 +41738,12 @@ wuffs_jpeg__decoder__decode_mcu__choosy_default(
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);
v_z = (self->private_impl.f_mcu_zig_index + v_ac_rrrr + 1u);
self->private_data.f_mcu_blocks[0u][WUFFS_JPEG__UNZIG[v_z]] = ((uint16_t)(v_ac_value));
if ((v_z + 1u) > 63u) {
if (v_z > 63u) {
goto label__ac_components__break;
}
self->private_impl.f_mcu_zig_index = (v_z + 1u);
self->private_impl.f_mcu_zig_index = v_z;
} else if (v_ac_rrrr < 15u) {
goto label__ac_components__break;
} else if ((self->private_impl.f_mcu_zig_index + 16u) > 63u) {
@@ -41903,12 +41903,12 @@ wuffs_jpeg__decoder__decode_mcu_progressive_ac_high_bits(
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);
v_z = (self->private_impl.f_mcu_zig_index + 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))));
if ((v_z + 1u) > ((uint32_t)(self->private_impl.f_scan_se))) {
if (v_z > ((uint32_t)(self->private_impl.f_scan_se))) {
goto label__ac_components__break;
}
self->private_impl.f_mcu_zig_index = (v_z + 1u);
self->private_impl.f_mcu_zig_index = v_z;
} 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) {
@@ -42067,7 +42067,7 @@ wuffs_jpeg__decoder__decode_mcu_progressive_ac_low_bit(
goto label__goto_do_eob__break;
}
while (true) {
v_unzig = WUFFS_JPEG__UNZIG[self->private_impl.f_mcu_zig_index];
v_unzig = WUFFS_JPEG__UNZIG[(1u + self->private_impl.f_mcu_zig_index)];
if (self->private_data.f_mcu_blocks[0u][v_unzig] != 0u) {
if (v_n_bits == 0u) {
if (((uint64_t)(io2_v_r - iop_v_r)) < 8u) {
@@ -42114,7 +42114,7 @@ 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[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)]] = ((uint16_t)(v_ac_value));
}
if (self->private_impl.f_mcu_zig_index >= ((uint32_t)(self->private_impl.f_scan_se))) {
goto label__ac_components__break;
@@ -42130,7 +42130,7 @@ wuffs_jpeg__decoder__decode_mcu_progressive_ac_low_bit(
goto label__goto_done__break;
}
while (true) {
v_unzig = WUFFS_JPEG__UNZIG[self->private_impl.f_mcu_zig_index];
v_unzig = WUFFS_JPEG__UNZIG[(1u + self->private_impl.f_mcu_zig_index)];
if (self->private_data.f_mcu_blocks[0u][v_unzig] != 0u) {
if (v_n_bits == 0u) {
if (((uint64_t)(io2_v_r - iop_v_r)) < 8u) {
+4 -2
View File
@@ -16,6 +16,8 @@
// UNZIG[3] is the column and row of the fourth element in zig-zag order. The
// value is 16, which means first column (16%8 == 0) and third row (16/8 == 2).
pri const UNZIG : roarray[80] base.u8[..= 63] = [
0x00,
// The 8x8 indexes are below.
0x00, 0x01, 0x08, 0x10, 0x09, 0x02, 0x03, 0x0A,
0x11, 0x18, 0x20, 0x19, 0x12, 0x0B, 0x04, 0x05,
0x0C, 0x13, 0x1A, 0x21, 0x28, 0x30, 0x29, 0x22,
@@ -24,11 +26,11 @@ pri const UNZIG : roarray[80] base.u8[..= 63] = [
0x1D, 0x16, 0x0F, 0x17, 0x1E, 0x25, 0x2C, 0x33,
0x3A, 0x3B, 0x34, 0x2D, 0x26, 0x1F, 0x27, 0x2E,
0x35, 0x3C, 0x3D, 0x36, 0x2F, 0x37, 0x3E, 0x3F,
// The 16 extra elements are for invalid JPEGs where "the z in UNZIG[z]
// The 15 extra elements are for invalid JPEGs where "the z in UNZIG[z]
// overflows 63, when saving the AC value". This isn't in the spec but
// matches how libjpeg-turbo treats "above 63".
0x3F, 0x3F, 0x3F, 0x3F, 0x3F, 0x3F, 0x3F, 0x3F,
0x3F, 0x3F, 0x3F, 0x3F, 0x3F, 0x3F, 0x3F, 0x3F,
0x3F, 0x3F, 0x3F, 0x3F, 0x3F, 0x3F, 0x3F,
]
// BIAS_AND_CLAMP maps 10-bit signed values (0 centered, in the range -512 ..
+1 -1
View File
@@ -499,8 +499,8 @@ pri func decoder.decode_dqt?(src: base.io_reader) {
i = 0
while i < 64 {
this.quant_tables[q][UNZIG[i]] = args.src.read_u8?()
i += 1
this.quant_tables[q][UNZIG[i]] = args.src.read_u8?()
} endwhile
this.seen_dqt[q] = true
+4 -4
View File
@@ -53,7 +53,7 @@ pri func decoder.decode_mcu!(workbuf: slice base.u8, mx: base.u32[..= 0x1FFF], m
var ac_rrrr : base.u32[..= 15]
var ac_ssss : base.u32[..= 15]
var z : base.u32[..= 78] // 78 = 63 + 15.
var z : base.u32[..= 79] // 79 = 63 + 15 + 1.
var mcb : base.u32[..= 9]
var stride : base.u64[..= 0x1_0008]
@@ -211,13 +211,13 @@ pri func decoder.decode_mcu!(workbuf: slice base.u8, mx: base.u32[..= 0x1FFF], m
(((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
z = this.mcu_zig_index + ac_rrrr + 1
this.mcu_blocks[0][UNZIG[z]] =
(ac_value & 0xFFFF) as base.u16
if (z + 1) > 63 {
if z > 63 {
break.ac_components
}
this.mcu_zig_index = z + 1
this.mcu_zig_index = z
} else if ac_rrrr < 15 {
break.ac_components
} else if (this.mcu_zig_index + 16) > 63 {
@@ -36,7 +36,7 @@ pri func decoder.decode_mcu_progressive_ac_high_bits!(workbuf: slice base.u8, mx
var ac_rrrr : base.u32[..= 15]
var ac_ssss : base.u32[..= 15]
var z : base.u32[..= 78] // 78 = 63 + 15.
var z : base.u32[..= 79] // 79 = 63 + 15 + 1.
if this.eob_run > 0 {
this.eob_run -= 1
@@ -126,14 +126,14 @@ pri func decoder.decode_mcu_progressive_ac_high_bits!(workbuf: slice base.u8, mx
(((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
z = this.mcu_zig_index + ac_rrrr + 1
this.mcu_blocks[0][UNZIG[z]] =
((ac_value ~mod<< this.scan_al) & 0xFFFF) as base.u16
if (z + 1) > (this.scan_se as base.u32) {
if z > (this.scan_se as base.u32) {
break.ac_components
}
assert (z + 1) <= 63 via "a <= b: a <= c; c <= b"(c: (this.scan_se as base.u32))
this.mcu_zig_index = z + 1
assert z <= 63 via "a <= b: a <= c; c <= b"(c: (this.scan_se as base.u32))
this.mcu_zig_index = z
} else if ac_rrrr < 15 {
this.eob_run = (((1 as base.u16) << ac_rrrr) - 1) as base.u16
if ac_rrrr > 0 {
@@ -144,7 +144,7 @@ pri func decoder.decode_mcu_progressive_ac_low_bit!(workbuf: slice base.u8, mx:
while.refine_non_zeroes true,
inv ac_huff_table_fast <> nullptr,
{
unzig = UNZIG[this.mcu_zig_index]
unzig = UNZIG[1 + this.mcu_zig_index]
if this.mcu_blocks[0][unzig] <> 0 {
if n_bits == 0 {
if r.length() < 8 {
@@ -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[this.mcu_zig_index]] = (ac_value & 0xFFFF) as base.u16
this.mcu_blocks[0][UNZIG[1 + this.mcu_zig_index]] = (ac_value & 0xFFFF) as base.u16
}
if this.mcu_zig_index >= (this.scan_se as base.u32) {
@@ -200,7 +200,7 @@ pri func decoder.decode_mcu_progressive_ac_low_bit!(workbuf: slice base.u8, mx:
while.refine_eob_non_zeroes true,
inv this.eob_run > 0,
{
unzig = UNZIG[this.mcu_zig_index]
unzig = UNZIG[1 + this.mcu_zig_index]
if this.mcu_blocks[0][unzig] <> 0 {
if n_bits == 0 {
// Load at least 1 bit, by keeping 56 .. 64 loaded.