mirror of
https://github.com/ddidderr/wuffs.git
synced 2026-11-19 00:38:14 +01:00
std/jpeg: check up-front for sufficient AC bits
The micro-benchmarks are mixed for now, but this sets us up to remove an inner-loop bounds check in a future commit. name old speed new speed delta wuffs_jpeg_decode_19k_8bpp/clang11 226MB/s ± 0% 224MB/s ± 1% -1.03% (p=0.029 n=4+4) wuffs_jpeg_decode_30k_24bpp_progressive/clang11 210MB/s ± 0% 206MB/s ± 0% -1.59% (p=0.029 n=4+4) wuffs_jpeg_decode_30k_24bpp_sequential/clang11 379MB/s ± 0% 374MB/s ± 0% -1.25% (p=0.016 n=4+5) wuffs_jpeg_decode_77k_24bpp/clang11 326MB/s ± 0% 319MB/s ± 0% -2.35% (p=0.029 n=4+4) wuffs_jpeg_decode_552k_24bpp_420/clang11 338MB/s ± 0% 329MB/s ± 0% -2.53% (p=0.008 n=5+5) wuffs_jpeg_decode_552k_24bpp_444/clang11 468MB/s ± 0% 465MB/s ± 0% -0.58% (p=0.008 n=5+5) wuffs_jpeg_decode_4002k_24bpp/clang11 347MB/s ± 0% 332MB/s ± 0% -4.41% (p=0.016 n=4+5) wuffs_jpeg_decode_19k_8bpp/gcc10 174MB/s ± 0% 176MB/s ± 0% +1.18% (p=0.008 n=5+5) wuffs_jpeg_decode_30k_24bpp_progressive/gcc10 247MB/s ± 0% 246MB/s ± 0% -0.55% (p=0.008 n=5+5) wuffs_jpeg_decode_30k_24bpp_sequential/gcc10 477MB/s ± 0% 472MB/s ± 0% -0.90% (p=0.008 n=5+5) wuffs_jpeg_decode_77k_24bpp/gcc10 328MB/s ± 0% 334MB/s ± 0% +1.69% (p=0.008 n=5+5) wuffs_jpeg_decode_552k_24bpp_420/gcc10 346MB/s ± 0% 351MB/s ± 0% +1.57% (p=0.008 n=5+5) wuffs_jpeg_decode_552k_24bpp_444/gcc10 424MB/s ± 0% 425MB/s ± 0% +0.39% (p=0.008 n=5+5) wuffs_jpeg_decode_4002k_24bpp/gcc10 350MB/s ± 0% 349MB/s ± 0% -0.23% (p=0.008 n=5+5)
This commit is contained in:
1 parent
2ca61b521f
commit
143991a301
5 files changed
+46
-8
No files matched your search
@@ -41028,7 +41028,7 @@ wuffs_jpeg__decoder__fill_bitstream(
|
||||
label__0__break:;
|
||||
if (((uint64_t)(io2_a_src - iop_a_src)) > 1u) {
|
||||
if ((wuffs_base__peek_u8be__no_bounds_check(iop_a_src) >= 255u) && ((wuffs_base__peek_u16le__no_bounds_check(iop_a_src) >> 8u) > 0u)) {
|
||||
v_new_wi = (wuffs_base__u32__min(v_wi, 2016u) + 32u);
|
||||
v_new_wi = (wuffs_base__u32__min(v_wi, 1784u) + 264u);
|
||||
if (v_wi < v_new_wi) {
|
||||
wuffs_base__bulk_memset(&self->private_data.f_bitstream_buffer[v_wi], (v_new_wi - v_wi), 0u);
|
||||
}
|
||||
@@ -41676,11 +41676,15 @@ wuffs_jpeg__decoder__decode_mcu__choosy_default(
|
||||
goto label__dc_component__break;
|
||||
}
|
||||
label__dc_component__break:;
|
||||
if (((uint64_t)(io2_v_r - iop_v_r)) < 264u) {
|
||||
v_ret = 1u;
|
||||
goto label__goto_done__break;
|
||||
}
|
||||
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 (true) {
|
||||
if (((uint64_t)(io2_v_r - iop_v_r)) < 8u) {
|
||||
v_ret = 1u;
|
||||
v_ret = 2u;
|
||||
goto label__goto_done__break;
|
||||
}
|
||||
v_bits |= (wuffs_base__peek_u64be__no_bounds_check(iop_v_r) >> (v_n_bits & 63u));
|
||||
@@ -41837,11 +41841,15 @@ wuffs_jpeg__decoder__decode_mcu_progressive_ac_high_bits(
|
||||
((uint64_t)(self->private_impl.f_bitstream_ri)));
|
||||
while (true) {
|
||||
while (true) {
|
||||
if (((uint64_t)(io2_v_r - iop_v_r)) < 264u) {
|
||||
v_ret = 1u;
|
||||
goto label__goto_done__break;
|
||||
}
|
||||
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 (true) {
|
||||
if (((uint64_t)(io2_v_r - iop_v_r)) < 8u) {
|
||||
v_ret = 1u;
|
||||
v_ret = 2u;
|
||||
goto label__goto_done__break;
|
||||
}
|
||||
v_bits |= (wuffs_base__peek_u64be__no_bounds_check(iop_v_r) >> (v_n_bits & 63u));
|
||||
|
||||
@@ -1514,11 +1514,11 @@ pri func decoder.fill_bitstream!(src: base.io_reader) {
|
||||
} endwhile
|
||||
|
||||
// If we hit a "\xFF\x??" marker then pad the bitstream. In this method
|
||||
// invocation, insert up to 32 bytes of implicit zeroes. Repeated calls to
|
||||
// invocation, insert up to 264 bytes of implicit zeroes. Repeated calls to
|
||||
// this method can insert even more.
|
||||
if args.src.length() > 1 {
|
||||
if (args.src.peek_u8() >= 0xFF) and ((args.src.peek_u16le() >> 8) > 0) {
|
||||
new_wi = wi.min(no_more_than: 0x800 - 32) + 32
|
||||
new_wi = wi.min(no_more_than: 0x800 - 264) + 264
|
||||
if wi < new_wi {
|
||||
this.bitstream_buffer[wi .. new_wi].bulk_memset!(byte_value: 0)
|
||||
}
|
||||
|
||||
@@ -133,6 +133,18 @@ pri func decoder.decode_mcu!(workbuf: slice base.u8, mx: base.u32[..= 0x1FFF], m
|
||||
break.dc_component
|
||||
} endwhile.dc_component
|
||||
|
||||
// Ensure that we have enough bits for the entire
|
||||
// while.ac_components loop.
|
||||
//
|
||||
// Worst case, there are 63 AC components (round that up to 64) and
|
||||
// each one needs (16 + 15) bits (round that up to 4 bytes), so we
|
||||
// need (64 * 4) = 256 bytes of buffer. Add 8 bytes slack so that
|
||||
// we can always call peek_u64be.
|
||||
if r.length() < 264 {
|
||||
ret = 1 // Request another fill_bitstream call.
|
||||
break.goto_done
|
||||
}
|
||||
|
||||
ac_h = this.mcu_blocks_ac_hselector[this.mcu_current_block]
|
||||
ac_huff_table_fast = this.huff_tables_fast[ac_h][..] as ptr array[256] base.u16
|
||||
while.ac_components true,
|
||||
@@ -140,8 +152,10 @@ pri func decoder.decode_mcu!(workbuf: slice base.u8, mx: base.u32[..= 0x1FFF], m
|
||||
inv ac_huff_table_fast <> nullptr,
|
||||
{
|
||||
// Load at least 31 bits, by keeping 56 .. 64 loaded.
|
||||
//
|
||||
// TODO: eliminate this bounds check entirely.
|
||||
if r.length() < 8 {
|
||||
ret = 1 // Request another fill_bitstream call.
|
||||
ret = 2 // Internal error.
|
||||
break.goto_done
|
||||
}
|
||||
bits |= r.peek_u64be() >> (n_bits & 63)
|
||||
|
||||
@@ -52,14 +52,28 @@ pri func decoder.decode_mcu_progressive_ac_high_bits!(workbuf: slice base.u8, mx
|
||||
while.goto_done true {{
|
||||
|
||||
while.block true {
|
||||
// Ensure that we have enough bits for the entire
|
||||
// while.ac_components loop.
|
||||
//
|
||||
// Worst case, there are 63 AC components (round that up to 64) and
|
||||
// each one needs (16 + 15) bits (round that up to 4 bytes), so we
|
||||
// need (64 * 4) = 256 bytes of buffer. Add 8 bytes slack so that
|
||||
// we can always call peek_u64be.
|
||||
if r.length() < 264 {
|
||||
ret = 1 // Request another fill_bitstream call.
|
||||
break.goto_done
|
||||
}
|
||||
|
||||
ac_h = this.mcu_blocks_ac_hselector[0]
|
||||
ac_huff_table_fast = this.huff_tables_fast[ac_h][..] as ptr array[256] base.u16
|
||||
while.ac_components true,
|
||||
inv ac_huff_table_fast <> nullptr,
|
||||
{
|
||||
// Load at least 31 bits, by keeping 56 .. 64 loaded.
|
||||
//
|
||||
// TODO: eliminate this bounds check entirely.
|
||||
if r.length() < 8 {
|
||||
ret = 1 // Request another fill_bitstream call.
|
||||
ret = 2 // Internal error.
|
||||
break.goto_done
|
||||
}
|
||||
bits |= r.peek_u64be() >> (n_bits & 63)
|
||||
|
||||
@@ -68,7 +68,9 @@ pri func decoder.decode_mcu_progressive_ac_low_bit!(workbuf: slice base.u8, mx:
|
||||
// and processing the ac_value (up to 15 bits).
|
||||
//
|
||||
// After that, it's simpler for this "AC low bit" flavor if we
|
||||
// never suspend for the rest of the while.ac_components loop.
|
||||
// never suspend for the rest of this iteration of the
|
||||
// while.ac_components loop.
|
||||
//
|
||||
// Specifically, we want to have enough bits in the bitstream
|
||||
// so that we don't suspend inside the while.refine_non_zeroes
|
||||
// loop. In the worst case, we need another 8 bytes for
|
||||
|
||||
Reference in new issue
Block a user