std/jpeg: read Huffman look-up table earlier

This commit is the first of two. It just changes code, and is the
minimal code change that leads to these performance improvements. The
follow-up commit will add comments that will also include a reference to
this one's commit hash. Even later commits may do some more refactoring.

name                                             old speed     new speed     delta

wuffs_jpeg_decode_19k_8bpp/clang11               224MB/s ± 0%  236MB/s ± 0%  +5.38%  (p=0.016 n=5+4)
wuffs_jpeg_decode_30k_24bpp_progressive/clang11  211MB/s ± 0%  210MB/s ± 0%  -0.55%  (p=0.008 n=5+5)
wuffs_jpeg_decode_30k_24bpp_sequential/clang11   377MB/s ± 0%  385MB/s ± 0%  +1.91%  (p=0.008 n=5+5)
wuffs_jpeg_decode_77k_24bpp/clang11              328MB/s ± 0%  338MB/s ± 0%  +3.23%  (p=0.008 n=5+5)
wuffs_jpeg_decode_552k_24bpp_420/clang11         341MB/s ± 0%  353MB/s ± 0%  +3.47%  (p=0.008 n=5+5)
wuffs_jpeg_decode_552k_24bpp_444/clang11         472MB/s ± 0%  481MB/s ± 0%  +1.93%  (p=0.008 n=5+5)
wuffs_jpeg_decode_4002k_24bpp/clang11            347MB/s ± 0%  360MB/s ± 0%  +3.70%  (p=0.008 n=5+5)

wuffs_jpeg_decode_19k_8bpp/gcc10                 225MB/s ± 0%  244MB/s ± 0%  +8.27%  (p=0.008 n=5+5)
wuffs_jpeg_decode_30k_24bpp_progressive/gcc10    246MB/s ± 0%  251MB/s ± 0%  +1.67%  (p=0.008 n=5+5)
wuffs_jpeg_decode_30k_24bpp_sequential/gcc10     498MB/s ± 0%  509MB/s ± 0%  +2.18%  (p=0.008 n=5+5)
wuffs_jpeg_decode_77k_24bpp/gcc10                408MB/s ± 0%  431MB/s ± 0%  +5.86%  (p=0.008 n=5+5)
wuffs_jpeg_decode_552k_24bpp_420/gcc10           428MB/s ± 0%  455MB/s ± 0%  +6.38%  (p=0.008 n=5+5)
wuffs_jpeg_decode_552k_24bpp_444/gcc10           470MB/s ± 0%  481MB/s ± 0%  +2.46%  (p=0.016 n=5+4)
wuffs_jpeg_decode_4002k_24bpp/gcc10              433MB/s ± 0%  463MB/s ± 0%  +7.03%  (p=0.008 n=5+5)
This commit is contained in:
Nigel Tao committed 2023-06-21 21:56:54 +10:00
1 parent bedf2bd7c3
commit d9c7740ea6
3 files changed
+20 -4

No files matched your search

+8 -2
View File
@@ -42108,11 +42108,15 @@ wuffs_jpeg__decoder__decode_mcu__choosy_default(
v_ret = 1u;
goto label__goto_done__break;
}
if (v_n_bits < 16u) {
v_bits |= (wuffs_base__peek_u64be__no_bounds_check(iop_v_r) >> (v_n_bits & 63u));
}
v_z = 1u;
self->private_impl.f_mcu_zig_index = 0u;
v_ac_h = self->private_impl.f_mcu_blocks_ac_hselector[self->private_impl.f_mcu_current_block];
v_ac_huff_table_fast = &self->private_impl.f_huff_tables_fast[v_ac_h][0u];
while (v_z < 64u) {
v_ac_ht_fast = ((uint32_t)(v_ac_huff_table_fast[(v_bits >> 56u)]));
if (((uint64_t)(io2_v_r - iop_v_r)) < 8u) {
v_ret = 2u;
goto label__goto_done__break;
@@ -42120,7 +42124,6 @@ wuffs_jpeg__decoder__decode_mcu__choosy_default(
v_bits |= (wuffs_base__peek_u64be__no_bounds_check(iop_v_r) >> (v_n_bits & 63u));
iop_v_r += ((63u - (v_n_bits & 63u)) >> 3u);
v_n_bits |= 56u;
v_ac_ht_fast = ((uint32_t)(v_ac_huff_table_fast[(v_bits >> 56u)]));
v_ac_bl = (v_ac_ht_fast >> 8u);
if (v_n_bits >= v_ac_bl) {
v_ac_symbol = (255u & v_ac_ht_fast);
@@ -42267,11 +42270,15 @@ wuffs_jpeg__decoder__decode_mcu_progressive_ac_high_bits(
v_ret = 1u;
goto label__goto_done__break;
}
if (v_n_bits < 16u) {
v_bits |= (wuffs_base__peek_u64be__no_bounds_check(iop_v_r) >> (v_n_bits & 63u));
}
v_z = self->private_impl.f_mcu_zig_index;
self->private_impl.f_mcu_zig_index = 0u;
v_ac_h = self->private_impl.f_mcu_blocks_ac_hselector[0u];
v_ac_huff_table_fast = &self->private_impl.f_huff_tables_fast[v_ac_h][0u];
while (v_z <= ((uint32_t)(self->private_impl.f_scan_se))) {
v_ac_ht_fast = ((uint32_t)(v_ac_huff_table_fast[(v_bits >> 56u)]));
if (((uint64_t)(io2_v_r - iop_v_r)) < 8u) {
v_ret = 2u;
goto label__goto_done__break;
@@ -42279,7 +42286,6 @@ wuffs_jpeg__decoder__decode_mcu_progressive_ac_high_bits(
v_bits |= (wuffs_base__peek_u64be__no_bounds_check(iop_v_r) >> (v_n_bits & 63u));
iop_v_r += ((63u - (v_n_bits & 63u)) >> 3u);
v_n_bits |= 56u;
v_ac_ht_fast = ((uint32_t)(v_ac_huff_table_fast[(v_bits >> 56u)]));
v_ac_bl = (v_ac_ht_fast >> 8u);
if (v_n_bits >= v_ac_bl) {
v_ac_symbol = (255u & v_ac_ht_fast);
+6 -1
View File
@@ -149,6 +149,10 @@ pri func decoder.decode_mcu!(workbuf: slice base.u8, mx: base.u32[..= 0x1FFF], m
break.goto_done
}
if n_bits < 16 {
bits |= r.peek_u64be() >> (n_bits & 63)
}
z = 1
this.mcu_zig_index = 0
ac_h = this.mcu_blocks_ac_hselector[this.mcu_current_block]
@@ -157,6 +161,8 @@ pri func decoder.decode_mcu!(workbuf: slice base.u8, mx: base.u32[..= 0x1FFF], m
inv this.mcu_current_block < 10,
inv ac_huff_table_fast <> nullptr,
{
ac_ht_fast = ac_huff_table_fast[bits >> 56] as base.u32
// Load at least 31 bits, by keeping 56 .. 64 loaded.
//
// TODO: eliminate this bounds check entirely.
@@ -169,7 +175,6 @@ pri func decoder.decode_mcu!(workbuf: slice base.u8, mx: base.u32[..= 0x1FFF], m
n_bits |= 56
// Read the Huffman-encoded ac_symbol, up to 16 bits long.
ac_ht_fast = ac_huff_table_fast[bits >> 56] as base.u32
ac_bl = ac_ht_fast >> 8
if n_bits >= ac_bl {
ac_symbol = 0xFF & ac_ht_fast
@@ -65,6 +65,10 @@ pri func decoder.decode_mcu_progressive_ac_high_bits!(workbuf: slice base.u8, mx
break.goto_done
}
if n_bits < 16 {
bits |= r.peek_u64be() >> (n_bits & 63)
}
z = this.mcu_zig_index
this.mcu_zig_index = 0
ac_h = this.mcu_blocks_ac_hselector[0]
@@ -74,6 +78,8 @@ pri func decoder.decode_mcu_progressive_ac_high_bits!(workbuf: slice base.u8, mx
{
assert z < 64 via "a < b: a <= c; c < b"(c: (this.scan_se as base.u32))
ac_ht_fast = ac_huff_table_fast[bits >> 56] as base.u32
// Load at least 31 bits, by keeping 56 .. 64 loaded.
//
// TODO: eliminate this bounds check entirely.
@@ -86,7 +92,6 @@ pri func decoder.decode_mcu_progressive_ac_high_bits!(workbuf: slice base.u8, mx
n_bits |= 56
// Read the Huffman-encoded ac_symbol, up to 16 bits long.
ac_ht_fast = ac_huff_table_fast[bits >> 56] as base.u32
ac_bl = ac_ht_fast >> 8
if n_bits >= ac_bl {
ac_symbol = 0xFF & ac_ht_fast