std/jpeg: call slice.bulk_{load,save}_host_endian

Updates https://bugs.chromium.org/p/oss-fuzz/issues/detail?id=59540

name                                             old speed     new speed     delta

wuffs_jpeg_decode_30k_24bpp_progressive/clang11  131MB/s ± 0%  168MB/s ± 0%  +28.44%  (p=0.008 n=5+5)
wuffs_jpeg_decode_30k_24bpp_progressive/gcc10    167MB/s ± 0%  175MB/s ± 0%   +4.80%  (p=0.008 n=5+5)
This commit is contained in:
Nigel Tao committed 2023-06-05 12:51:50 +10:00
1 parent 4a164cda64
commit 2865c5bfbd
7 files changed
+162 -154

No files matched your search

+26
View File
@@ -214,6 +214,19 @@ wuffs_base__slice_u8__copy_from_slice(wuffs_base__slice_u8 dst,
return len;
}
static inline wuffs_base__empty_struct //
wuffs_base__bulk_load_host_endian(void* ptr,
size_t len,
wuffs_base__slice_u8 src) {
if (len > src.len) {
len = src.len;
}
if (len) {
memmove(ptr, src.ptr, len);
}
return wuffs_base__make_empty_struct();
}
static inline wuffs_base__empty_struct //
wuffs_base__bulk_memset(void* ptr, size_t len, uint8_t byte_value) {
if (len) {
@@ -222,6 +235,19 @@ wuffs_base__bulk_memset(void* ptr, size_t len, uint8_t byte_value) {
return wuffs_base__make_empty_struct();
}
static inline wuffs_base__empty_struct //
wuffs_base__bulk_save_host_endian(void* ptr,
size_t len,
wuffs_base__slice_u8 dst) {
if (len > dst.len) {
len = dst.len;
}
if (len) {
memmove(dst.ptr, ptr, len);
}
return wuffs_base__make_empty_struct();
}
// --------
static inline wuffs_base__slice_u8 //
+77 -61
View File
@@ -872,68 +872,26 @@ func (g *gen) writeBuiltinNumType(b *buffer, recv *a.Expr, method t.ID, args []*
func (g *gen) writeBuiltinSlice(b *buffer, recv *a.Expr, method t.ID, args []*a.Node, sideEffectsOnly bool, depth uint32) error {
switch method {
case t.IDBulkMemset:
if arrayOrSlice, lo, hi, ok := recv.IsSlice(); !ok {
return fmt.Errorf("TODO: bulk_memset for general expressions")
} else if hi == nil {
return fmt.Errorf("TODO: bulk_memset for general expressions")
} else if arrayOrSlice.MType().IsBulkNumType() {
b.writes("wuffs_base__bulk_memset(&")
if err := g.writeExpr(b, arrayOrSlice, false, depth); err != nil {
return err
}
if lo == nil {
b.writes("[0], ")
if err := g.writeExpr(b, hi, false, depth); err != nil {
return err
}
} else {
b.writes("[")
if err := g.writeExpr(b, lo, false, depth); err != nil {
return err
}
b.writes("], (")
if err := g.writeExpr(b, hi, false, depth); err != nil {
return err
}
b.writes(" - ")
if err := g.writeExpr(b, lo, false, depth); err != nil {
return err
}
b.writes(")")
}
if bs := arrayOrSlice.MType().Inner().BulkSize(); (bs != nil) && (bs.Cmp(one) != 0) {
b.writes(" * (size_t)")
b.writes(bs.String())
b.writes("u")
}
b.writes(", ")
return g.writeArgs(b, args, depth)
} else if arrayOrSlice.MType().IsContainerOfSpecificNumType(t.IDSlice, t.IDU8) {
b.writes("wuffs_base__bulk_memset(")
if err := g.writeExpr(b, recv, false, depth); err != nil {
return err
}
b.writes(".ptr, ")
if err := g.writeExpr(b, recv, false, depth); err != nil {
return err
}
b.writes(".len")
if bs := arrayOrSlice.MType().Inner().BulkSize(); (bs != nil) && (bs.Cmp(one) != 0) {
b.writes(" * (size_t)")
b.writes(bs.String())
b.writes("u")
}
b.writes(", ")
return g.writeArgs(b, args, depth)
} else {
return fmt.Errorf("TODO: bulk_memset for general expressions")
case t.IDBulkLoadHostEndian:
b.writes("wuffs_base__bulk_load_host_endian(")
if err := g.writeBuiltinSliceBulkMethodRecv(b, recv, depth); err != nil {
return err
}
return g.writeArgs(b, args, depth)
case t.IDBulkMemset:
b.writes("wuffs_base__bulk_memset(")
if err := g.writeBuiltinSliceBulkMethodRecv(b, recv, depth); err != nil {
return err
}
return g.writeArgs(b, args, depth)
case t.IDBulkSaveHostEndian:
b.writes("wuffs_base__bulk_save_host_endian(")
if err := g.writeBuiltinSliceBulkMethodRecv(b, recv, depth); err != nil {
return err
}
return g.writeArgs(b, args, depth)
case t.IDCopyFromSlice:
if err := g.writeBuiltinSliceCopyFromSlice8(b, recv, method, args, depth); err != errOptimizationNotApplicable {
@@ -1013,6 +971,64 @@ func (g *gen) writeBuiltinSlice(b *buffer, recv *a.Expr, method t.ID, args []*a.
return errNoSuchBuiltin
}
func (g *gen) writeBuiltinSliceBulkMethodRecv(b *buffer, recv *a.Expr, depth uint32) error {
if arrayOrSlice, lo, hi, ok := recv.IsSlice(); !ok || (hi == nil) {
// No-op.
} else if arrayOrSlice.MType().IsBulkNumType() {
b.writes("&")
if err := g.writeExpr(b, arrayOrSlice, false, depth); err != nil {
return err
}
if lo == nil {
b.writes("[0], ")
if err := g.writeExpr(b, hi, false, depth); err != nil {
return err
}
} else {
b.writes("[")
if err := g.writeExpr(b, lo, false, depth); err != nil {
return err
}
b.writes("], (")
if err := g.writeExpr(b, hi, false, depth); err != nil {
return err
}
b.writes(" - ")
if err := g.writeExpr(b, lo, false, depth); err != nil {
return err
}
b.writes(")")
}
if bs := arrayOrSlice.MType().Inner().BulkSize(); (bs != nil) && (bs.Cmp(one) != 0) {
b.writes(" * (size_t)")
b.writes(bs.String())
b.writes("u")
}
b.writes(", ")
return nil
} else if arrayOrSlice.MType().IsContainerOfSpecificNumType(t.IDSlice, t.IDU8) {
if err := g.writeExpr(b, recv, false, depth); err != nil {
return err
}
b.writes(".ptr, ")
if err := g.writeExpr(b, recv, false, depth); err != nil {
return err
}
b.writes(".len")
if bs := arrayOrSlice.MType().Inner().BulkSize(); (bs != nil) && (bs.Cmp(one) != 0) {
b.writes(" * (size_t)")
b.writes(bs.String())
b.writes("u")
}
b.writes(", ")
return nil
}
return fmt.Errorf("TODO: bulk_memset for general expressions")
}
// writeBuiltinSliceCopyFromSlice8 writes an optimized version of:
//
// foo[fIndex .. fIndex + 8].copy_from_slice!(s:bar[bIndex .. bIndex + 8])
+3 -1
View File
@@ -965,10 +965,12 @@ var SliceFuncs = []string{
"GENERIC T1.uintptr_low_12_bits() u32[..= 4095]",
"GENERIC T1.suffix(up_to: u64) T1",
// The bulk_etc methods are further restricted to those slices whose
// The bulk_etc methods are further restricted to those T1 slices whose
// elements are bulk-numeric types: numerics (e.g. base.u32), arrays of
// numerics, arrays of arrays of numerics, etc.
"GENERIC T1.bulk_load_host_endian!(src: slice u8)",
"GENERIC T1.bulk_memset!(byte_value: u8)",
"GENERIC T1.bulk_save_host_endian!(dst: slice u8)",
}
var SliceU8Funcs = []string{
+2
View File
@@ -732,7 +732,9 @@ func (q *checker) tcheckDot(n *a.Expr, depth uint32) error {
if q.c.isBuiltInSliceFunc(qqid, lTyp) {
bulky := false
switch qqid[2] {
case t.IDBulkLoadHostEndian:
case t.IDBulkMemset:
case t.IDBulkSaveHostEndian:
bulky = true
}
+20 -14
View File
@@ -668,19 +668,22 @@ const (
IDIsOK = ID(0x231)
IDIsSuspension = ID(0x232)
IDBulkMemset = ID(0x240)
IDData = ID(0x241)
IDHeight = ID(0x242)
IDIO = ID(0x243)
IDLimit = ID(0x244)
IDPrefix = ID(0x245)
IDRowU32 = ID(0x246)
IDStride = ID(0x247)
IDSubtable = ID(0x248)
IDSuffix = ID(0x249)
IDUintptrLow12Bits = ID(0x24A)
IDValidUTF8Length = ID(0x24B)
IDWidth = ID(0x24C)
IDBulkLoadHostEndian = ID(0x238)
IDBulkMemset = ID(0x239)
IDBulkSaveHostEndian = ID(0x23A)
IDData = ID(0x240)
IDHeight = ID(0x241)
IDIO = ID(0x242)
IDLimit = ID(0x243)
IDPrefix = ID(0x244)
IDRowU32 = ID(0x245)
IDStride = ID(0x246)
IDSubtable = ID(0x247)
IDSuffix = ID(0x248)
IDUintptrLow12Bits = ID(0x249)
IDValidUTF8Length = ID(0x24A)
IDWidth = ID(0x24B)
IDLimitedSwizzleU32InterleavedFromReader = ID(0x280)
IDSwizzleInterleavedFromReader = ID(0x281)
@@ -1100,7 +1103,10 @@ var builtInsByID = [nBuiltInIDs]string{
IDIsOK: "is_ok",
IDIsSuspension: "is_suspension",
IDBulkMemset: "bulk_memset",
IDBulkLoadHostEndian: "bulk_load_host_endian",
IDBulkMemset: "bulk_memset",
IDBulkSaveHostEndian: "bulk_save_host_endian",
IDData: "data",
IDHeight: "height",
IDIO: "io",
+29 -33
View File
@@ -12277,6 +12277,19 @@ wuffs_base__slice_u8__copy_from_slice(wuffs_base__slice_u8 dst,
return len;
}
static inline wuffs_base__empty_struct //
wuffs_base__bulk_load_host_endian(void* ptr,
size_t len,
wuffs_base__slice_u8 src) {
if (len > src.len) {
len = src.len;
}
if (len) {
memmove(ptr, src.ptr, len);
}
return wuffs_base__make_empty_struct();
}
static inline wuffs_base__empty_struct //
wuffs_base__bulk_memset(void* ptr, size_t len, uint8_t byte_value) {
if (len) {
@@ -12285,6 +12298,19 @@ wuffs_base__bulk_memset(void* ptr, size_t len, uint8_t byte_value) {
return wuffs_base__make_empty_struct();
}
static inline wuffs_base__empty_struct //
wuffs_base__bulk_save_host_endian(void* ptr,
size_t len,
wuffs_base__slice_u8 dst) {
if (len > dst.len) {
len = dst.len;
}
if (len) {
memmove(dst.ptr, ptr, len);
}
return wuffs_base__make_empty_struct();
}
// --------
static inline wuffs_base__slice_u8 //
@@ -40016,21 +40042,12 @@ wuffs_jpeg__decoder__load_mcu_blocks_for_single_component(
uint32_t a_csel) {
uint64_t v_stride16 = 0;
uint64_t v_offset = 0;
wuffs_base__slice_u8 v_s = {0};
uint32_t v_i = 0;
while (true) {
v_stride16 = ((uint64_t)((self->private_impl.f_components_workbuf_widths[a_csel] * 16u)));
v_offset = (self->private_impl.f_components_workbuf_offsets[(a_csel | 4u)] + (((uint64_t)(a_mx)) * 128u) + (((uint64_t)(a_my)) * v_stride16));
if (v_offset <= ((uint64_t)(a_workbuf.len))) {
v_s = wuffs_base__slice_u8__subslice_i(a_workbuf, v_offset);
if (((uint64_t)(v_s.len)) >= 128u) {
v_i = 0u;
while (v_i < 64u) {
self->private_data.f_mcu_blocks[0u][v_i] = ((uint16_t)(((((uint16_t)(v_s.ptr[((2u * v_i) + 0u)])) << 0u) | (((uint16_t)(v_s.ptr[((2u * v_i) + 1u)])) << 8u))));
v_i += 1u;
}
}
wuffs_base__bulk_load_host_endian(&self->private_data.f_mcu_blocks[0], 1u * (size_t)128u, wuffs_base__slice_u8__subslice_i(a_workbuf, v_offset));
}
goto label__0__break;
}
@@ -40052,8 +40069,6 @@ wuffs_jpeg__decoder__load_mcu_blocks(
uint64_t v_v = 0;
uint64_t v_stride16 = 0;
uint64_t v_offset = 0;
wuffs_base__slice_u8 v_s = {0};
uint32_t v_i = 0;
v_h = 1u;
v_v = 1u;
@@ -40067,14 +40082,7 @@ wuffs_jpeg__decoder__load_mcu_blocks(
v_stride16 = ((uint64_t)((self->private_impl.f_components_workbuf_widths[v_csel] * 16u)));
v_offset = (self->private_impl.f_components_workbuf_offsets[(v_csel | 4u)] + (((v_h * ((uint64_t)(a_mx))) + ((uint64_t)(self->private_impl.f_scan_comps_bx_offset[v_b]))) * 128u) + (((v_v * ((uint64_t)(a_my))) + ((uint64_t)(self->private_impl.f_scan_comps_by_offset[v_b]))) * v_stride16));
if (v_offset <= ((uint64_t)(a_workbuf.len))) {
v_s = wuffs_base__slice_u8__subslice_i(a_workbuf, v_offset);
if (((uint64_t)(v_s.len)) >= 128u) {
v_i = 0u;
while (v_i < 64u) {
self->private_data.f_mcu_blocks[v_b][v_i] = ((uint16_t)(((((uint16_t)(v_s.ptr[((2u * v_i) + 0u)])) << 0u) | (((uint16_t)(v_s.ptr[((2u * v_i) + 1u)])) << 8u))));
v_i += 1u;
}
}
wuffs_base__bulk_load_host_endian(&self->private_data.f_mcu_blocks[v_b], ((v_b + 1u) - v_b) * (size_t)128u, wuffs_base__slice_u8__subslice_i(a_workbuf, v_offset));
}
v_b += 1u;
}
@@ -40095,9 +40103,6 @@ wuffs_jpeg__decoder__save_mcu_blocks(
uint64_t v_v = 0;
uint64_t v_stride16 = 0;
uint64_t v_offset = 0;
wuffs_base__slice_u8 v_s = {0};
uint32_t v_i = 0;
uint16_t v_m = 0;
v_h = 1u;
v_v = 1u;
@@ -40111,16 +40116,7 @@ wuffs_jpeg__decoder__save_mcu_blocks(
v_stride16 = ((uint64_t)((self->private_impl.f_components_workbuf_widths[v_csel] * 16u)));
v_offset = (self->private_impl.f_components_workbuf_offsets[(v_csel | 4u)] + (((v_h * ((uint64_t)(a_mx))) + ((uint64_t)(self->private_impl.f_scan_comps_bx_offset[v_b]))) * 128u) + (((v_v * ((uint64_t)(a_my))) + ((uint64_t)(self->private_impl.f_scan_comps_by_offset[v_b]))) * v_stride16));
if (v_offset <= ((uint64_t)(a_workbuf.len))) {
v_s = wuffs_base__slice_u8__subslice_i(a_workbuf, v_offset);
if (((uint64_t)(v_s.len)) >= 128u) {
v_i = 0u;
while (v_i < 64u) {
v_m = self->private_data.f_mcu_blocks[v_b][v_i];
v_s.ptr[((2u * v_i) + 0u)] = ((uint8_t)((v_m >> 0u)));
v_s.ptr[((2u * v_i) + 1u)] = ((uint8_t)((v_m >> 8u)));
v_i += 1u;
}
}
wuffs_base__bulk_save_host_endian(&self->private_data.f_mcu_blocks[v_b], ((v_b + 1u) - v_b) * (size_t)128u, wuffs_base__slice_u8__subslice_i(a_workbuf, v_offset));
}
v_b += 1u;
}
+5 -45
View File
@@ -1547,8 +1547,6 @@ pri func decoder.fill_bitstream!(src: base.io_reader) {
pri func decoder.load_mcu_blocks_for_single_component!(mx: base.u32[..= 0x2000], my: base.u32[..= 0x2000], workbuf: slice base.u8, csel: base.u32[..= 3]) {
var stride16 : base.u64[..= 0x10_0080]
var offset : base.u64
var s : slice base.u8
var i : base.u32
// "while true" is redundant but minimizes the diff with load_mcu_blocks.
while true {
@@ -1557,18 +1555,7 @@ pri func decoder.load_mcu_blocks_for_single_component!(mx: base.u32[..= 0x2000],
((args.mx as base.u64) * 128) +
((args.my as base.u64) * stride16)
if offset <= args.workbuf.length() {
s = args.workbuf[offset ..]
if s.length() >= 128 {
i = 0
while i < 64,
inv s.length() >= 128,
{
this.mcu_blocks[0][i] = (
((s[(2 * i) + 0] as base.u16) << 0) |
((s[(2 * i) + 1] as base.u16) << 8)) as base.u16
i += 1
} endwhile
}
this.mcu_blocks[.. 1].bulk_load_host_endian!(src: args.workbuf[offset ..])
}
break
} endwhile
@@ -1581,8 +1568,6 @@ pri func decoder.load_mcu_blocks!(mx: base.u32[..= 0x2000], my: base.u32[..= 0x2
var v : base.u64[..= 4]
var stride16 : base.u64[..= 0x10_0080]
var offset : base.u64
var s : slice base.u8
var i : base.u32
h = 1
v = 1
@@ -1599,19 +1584,8 @@ pri func decoder.load_mcu_blocks!(mx: base.u32[..= 0x2000], my: base.u32[..= 0x2
(((h * (args.mx as base.u64)) + (this.scan_comps_bx_offset[b] as base.u64)) * 128) +
(((v * (args.my as base.u64)) + (this.scan_comps_by_offset[b] as base.u64)) * stride16)
if offset <= args.workbuf.length() {
s = args.workbuf[offset ..]
if s.length() >= 128 {
i = 0
while i < 64,
inv b < 10,
inv s.length() >= 128,
{
this.mcu_blocks[b][i] = (
((s[(2 * i) + 0] as base.u16) << 0) |
((s[(2 * i) + 1] as base.u16) << 8)) as base.u16
i += 1
} endwhile
}
assert b <= (b + 1) via "a <= (a + b): 0 <= b"()
this.mcu_blocks[b .. (b + 1)].bulk_load_host_endian!(src: args.workbuf[offset ..])
}
b += 1
} endwhile
@@ -1624,9 +1598,6 @@ pri func decoder.save_mcu_blocks!(mx: base.u32[..= 0x2000], my: base.u32[..= 0x2
var v : base.u64[..= 4]
var stride16 : base.u64[..= 0x10_0080]
var offset : base.u64
var s : slice base.u8
var i : base.u32
var m : base.u16
h = 1
v = 1
@@ -1643,19 +1614,8 @@ pri func decoder.save_mcu_blocks!(mx: base.u32[..= 0x2000], my: base.u32[..= 0x2
(((h * (args.mx as base.u64)) + (this.scan_comps_bx_offset[b] as base.u64)) * 128) +
(((v * (args.my as base.u64)) + (this.scan_comps_by_offset[b] as base.u64)) * stride16)
if offset <= args.workbuf.length() {
s = args.workbuf[offset ..]
if s.length() >= 128 {
i = 0
while i < 64,
inv b < 10,
inv s.length() >= 128,
{
m = this.mcu_blocks[b][i]
s[(2 * i) + 0] = ((m >> 0) & 0xFF) as base.u8
s[(2 * i) + 1] = ((m >> 8) & 0xFF) as base.u8
i += 1
} endwhile
}
assert b <= (b + 1) via "a <= (a + b): 0 <= b"()
this.mcu_blocks[b .. (b + 1)].bulk_save_host_endian!(dst: args.workbuf[offset ..])
}
b += 1
} endwhile