std/jpeg: add more-SIMD-like IDCT algorithm

This adds commented-out code that is equivalent to the existing IDCT
algorithm. It's slower as is (and hence commented out) but it is closer
to a follow-up commit's SIMD-optimized algorithm.
This commit is contained in:
Nigel Tao committed 2023-06-28 13:45:25 +10:00
1 parent 49f50cca91
commit 97c4e85e0e
2 files changed
+376 -38

No files matched your search

+112 -38
View File
@@ -40,10 +40,14 @@ func main1() error {
fmt.Printf("// p0_298631336 = 0x%04X = %5d\n", round(p0_298631336), round(p0_298631336))
fmt.Printf("// p0_390180644 = 0x%04X = %5d\n", round(p0_390180644), round(p0_390180644))
fmt.Printf("// p0_509795579 = 0x%04X = %5d\n", round(p0_509795579), round(p0_509795579))
fmt.Printf("// p0_541196100 = 0x%04X = %5d\n", round(p0_541196100), round(p0_541196100))
fmt.Printf("// p0_601344887 = 0x%04X = %5d\n", round(p0_601344887), round(p0_601344887))
fmt.Printf("// p0_765366865 = 0x%04X = %5d\n", round(p0_765366865), round(p0_765366865))
fmt.Printf("// p0_785694958 = 0x%04X = %5d\n", round(p0_785694958), round(p0_785694958))
fmt.Printf("// p0_899976223 = 0x%04X = %5d\n", round(p0_899976223), round(p0_899976223))
fmt.Printf("// p1_175875602 = 0x%04X = %5d\n", round(p1_175875602), round(p1_175875602))
fmt.Printf("// p1_306562965 = 0x%04X = %5d\n", round(p1_306562965), round(p1_306562965))
fmt.Printf("// p1_501321110 = 0x%04X = %5d\n", round(p1_501321110), round(p1_501321110))
fmt.Printf("// p1_847759065 = 0x%04X = %5d\n", round(p1_847759065), round(p1_847759065))
fmt.Printf("// p1_961570560 = 0x%04X = %5d\n", round(p1_961570560), round(p1_961570560))
@@ -52,40 +56,32 @@ func main1() error {
fmt.Printf("// p3_072711026 = 0x%04X = %5d\n", round(p3_072711026), round(p3_072711026))
fmt.Printf("//\n")
fmt.Printf("// m0_390180644 = 0x%08X = %10d\n", round(m0_390180644), round(m0_390180644))
fmt.Printf("// m0_509795579 = 0x%08X = %10d\n", round(m0_509795579), round(m0_509795579))
fmt.Printf("// m0_601344887 = 0x%08X = %10d\n", round(m0_601344887), round(m0_601344887))
fmt.Printf("// m0_785694958 = 0x%08X = %10d\n", round(m0_785694958), round(m0_785694958))
fmt.Printf("// m0_899976223 = 0x%08X = %10d\n", round(m0_899976223), round(m0_899976223))
fmt.Printf("// m1_306562965 = 0x%08X = %10d\n", round(m1_306562965), round(m1_306562965))
fmt.Printf("// m1_961570560 = 0x%08X = %10d\n", round(m1_961570560), round(m1_961570560))
fmt.Printf("// m2_562915447 = 0x%08X = %10d\n", round(m2_562915447), round(m2_562915447))
fmt.Println()
for x := 0; x < 8; x++ {
fmt.Println(strings.TrimSpace(replace(pass0, map[string]string{
"$colX$": fmt.Sprint(x),
"$row0colX$": fmt.Sprintf("0x%02X", (8*0)|x),
"$row1colX$": fmt.Sprintf("0x%02X", (8*1)|x),
"$row2colX$": fmt.Sprintf("0x%02X", (8*2)|x),
"$row3colX$": fmt.Sprintf("0x%02X", (8*3)|x),
"$row4colX$": fmt.Sprintf("0x%02X", (8*4)|x),
"$row5colX$": fmt.Sprintf("0x%02X", (8*5)|x),
"$row6colX$": fmt.Sprintf("0x%02X", (8*6)|x),
"$row7colX$": fmt.Sprintf("0x%02X", (8*7)|x),
"$p0_298631336$": fmt.Sprintf("0x%04X", round(p0_298631336)),
"$p0_541196100$": fmt.Sprintf("0x%04X", round(p0_541196100)),
"$p0_765366865$": fmt.Sprintf("0x%04X", round(p0_765366865)),
"$p1_175875602$": fmt.Sprintf("0x%04X", round(p1_175875602)),
"$p1_501321110$": fmt.Sprintf("0x%04X", round(p1_501321110)),
"$p1_847759065$": fmt.Sprintf("0x%04X", round(p1_847759065)),
"$p2_053119869$": fmt.Sprintf("0x%04X", round(p2_053119869)),
"$p3_072711026$": fmt.Sprintf("0x%04X", round(p3_072711026)),
"$m0_390180644$": fmt.Sprintf("0x%08X", round(m0_390180644)),
"$m0_899976223$": fmt.Sprintf("0x%08X", round(m0_899976223)),
"$m1_961570560$": fmt.Sprintf("0x%08X", round(m1_961570560)),
"$m2_562915447$": fmt.Sprintf("0x%08X", round(m2_562915447)),
})))
fmt.Println(strings.TrimSpace(replace(pass0, addConsts(map[string]string{
"$colX$": fmt.Sprint(x),
"$row0colX$": fmt.Sprintf("0x%02X", (8*0)|x),
"$row1colX$": fmt.Sprintf("0x%02X", (8*1)|x),
"$row2colX$": fmt.Sprintf("0x%02X", (8*2)|x),
"$row3colX$": fmt.Sprintf("0x%02X", (8*3)|x),
"$row4colX$": fmt.Sprintf("0x%02X", (8*4)|x),
"$row5colX$": fmt.Sprintf("0x%02X", (8*5)|x),
"$row6colX$": fmt.Sprintf("0x%02X", (8*6)|x),
"$row7colX$": fmt.Sprintf("0x%02X", (8*7)|x),
}))))
fmt.Println()
}
for y := 0; y < 8; y++ {
fmt.Println(strings.TrimSpace(replace(pass1, map[string]string{
fmt.Println(strings.TrimSpace(replace(pass1, addConsts(map[string]string{
"$rowY$": fmt.Sprint(y),
"$rowYcol0$": fmt.Sprintf("0x%02X", (8*y)|0),
"$rowYcol1$": fmt.Sprintf("0x%02X", (8*y)|1),
@@ -95,21 +91,9 @@ func main1() error {
"$rowYcol5$": fmt.Sprintf("0x%02X", (8*y)|5),
"$rowYcol6$": fmt.Sprintf("0x%02X", (8*y)|6),
"$rowYcol7$": fmt.Sprintf("0x%02X", (8*y)|7),
"$p0_298631336$": fmt.Sprintf("0x%04X", round(p0_298631336)),
"$p0_541196100$": fmt.Sprintf("0x%04X", round(p0_541196100)),
"$p0_765366865$": fmt.Sprintf("0x%04X", round(p0_765366865)),
"$p1_175875602$": fmt.Sprintf("0x%04X", round(p1_175875602)),
"$p1_501321110$": fmt.Sprintf("0x%04X", round(p1_501321110)),
"$p1_847759065$": fmt.Sprintf("0x%04X", round(p1_847759065)),
"$p2_053119869$": fmt.Sprintf("0x%04X", round(p2_053119869)),
"$p3_072711026$": fmt.Sprintf("0x%04X", round(p3_072711026)),
"$m0_390180644$": fmt.Sprintf("0x%08X", round(m0_390180644)),
"$m0_899976223$": fmt.Sprintf("0x%08X", round(m0_899976223)),
"$m1_961570560$": fmt.Sprintf("0x%08X", round(m1_961570560)),
"$m2_562915447$": fmt.Sprintf("0x%08X", round(m2_562915447)),
"$bounds_check$": boundsCheck(y == 7),
"$advance$": advance(y == 7),
})))
}))))
fmt.Println()
}
@@ -124,8 +108,56 @@ func replace(s string, m map[string]string) string {
return s
}
func addConsts(m map[string]string) map[string]string {
m["$p0_298631336$"] = fmt.Sprintf("0x%04X", round(p0_298631336))
m["$p0_509795579$"] = fmt.Sprintf("0x%04X", round(p0_509795579))
m["$p0_541196100$"] = fmt.Sprintf("0x%04X", round(p0_541196100))
m["$p0_601344887$"] = fmt.Sprintf("0x%04X", round(p0_601344887))
m["$p0_765366865$"] = fmt.Sprintf("0x%04X", round(p0_765366865))
m["$p0_785694958$"] = fmt.Sprintf("0x%04X", round(p0_785694958))
m["$p1_175875602$"] = fmt.Sprintf("0x%04X", round(p1_175875602))
m["$p1_306562965$"] = fmt.Sprintf("0x%04X", round(p1_306562965))
m["$p1_501321110$"] = fmt.Sprintf("0x%04X", round(p1_501321110))
m["$p1_847759065$"] = fmt.Sprintf("0x%04X", round(p1_847759065))
m["$p2_053119869$"] = fmt.Sprintf("0x%04X", round(p2_053119869))
m["$p3_072711026$"] = fmt.Sprintf("0x%04X", round(p3_072711026))
m["$m0_390180644$"] = fmt.Sprintf("0x%08X", round(m0_390180644))
m["$m0_509795579$"] = fmt.Sprintf("0x%08X", round(m0_509795579))
m["$m0_601344887$"] = fmt.Sprintf("0x%08X", round(m0_601344887))
m["$m0_785694958$"] = fmt.Sprintf("0x%08X", round(m0_785694958))
m["$m0_899976223$"] = fmt.Sprintf("0x%08X", round(m0_899976223))
m["$m1_306562965$"] = fmt.Sprintf("0x%08X", round(m1_306562965))
m["$m1_961570560$"] = fmt.Sprintf("0x%08X", round(m1_961570560))
m["$m2_562915447$"] = fmt.Sprintf("0x%08X", round(m2_562915447))
return m
}
func round(x float64) uint32 {
return uint32(0.5 + (x * (1 << 13)))
// Hard-code some rounding errors to match libjpeg-turbo exactly.
//
// For example, p1_306562965 is defined precisely as (sqrt2 * (+cos2pi16)).
// Scaling by 8192 gives 10703.36380826727651620162087671551716... which
// rounds down (as a fixed precision approximation) to 10703 = 0x29CF.
//
// Conceptually, m1_306562965 is the negative of that 0x29CF, but it
// appears in libjpeg-turbo's algorithm as the negative of (p1_847759065 -
// p0_541196100). Those two constants (after scaling by 8192 and rounding
// *separately*) fold as (15137 - 4433) = 0x29D0, differing by 1.
adjustment := int32(0)
switch x {
case p0_509795579:
adjustment = +1
case p0_785694958:
adjustment = +1
case m0_601344887:
adjustment = -1
case m1_306562965:
adjustment = -1
}
return uint32(adjustment) + uint32(0.5+(x*(1<<13)))
}
func boundsCheck(final bool) string {
@@ -163,7 +195,11 @@ const (
const (
p0_541196100 = sqrt2 * (+cos6pi16)
p0_785694958 = sqrt2 * (+cos5pi16)
p1_175875602 = sqrt2 * (+cos3pi16)
p1_306562965 = sqrt2 * (+cos2pi16)
p0_601344887 = sqrt2 * (+cos1pi16 - cos5pi16)
p0_765366865 = sqrt2 * (+cos2pi16 - cos6pi16)
p1_847759065 = sqrt2 * (+cos2pi16 + cos6pi16)
@@ -173,13 +209,19 @@ const (
p1_961570560 = sqrt2 * (+cos3pi16 + cos5pi16)
p2_562915447 = sqrt2 * (+cos3pi16 + cos1pi16)
p0_509795579 = sqrt2 * (+cos5pi16 - cos7pi16)
p0_298631336 = sqrt2 * (-cos1pi16 + cos3pi16 + cos5pi16 - cos7pi16)
p1_501321110 = sqrt2 * (+cos1pi16 + cos3pi16 - cos5pi16 - cos7pi16)
p2_053119869 = sqrt2 * (+cos1pi16 + cos3pi16 - cos5pi16 + cos7pi16)
p3_072711026 = sqrt2 * (+cos1pi16 + cos3pi16 + cos5pi16 - cos7pi16)
m0_390180644 = (1 << (32 - 13)) - p0_390180644
m0_509795579 = (1 << (32 - 13)) - p0_509795579
m0_601344887 = (1 << (32 - 13)) - p0_601344887
m0_785694958 = (1 << (32 - 13)) - p0_785694958
m0_899976223 = (1 << (32 - 13)) - p0_899976223
m1_306562965 = (1 << (32 - 13)) - p1_306562965
m1_961570560 = (1 << (32 - 13)) - p1_961570560
m2_562915447 = (1 << (32 - 13)) - p2_562915447
)
@@ -213,11 +255,17 @@ intermediate[$row7colX$] = intermediate[$row0colX$]
bq2 = this.util.sign_extend_convert_u16_u32(a: this.mcu_blocks[0][$row2colX$]) ~mod* (this.quant_tables[args.q][$row2colX$] as base.u32)
bq6 = this.util.sign_extend_convert_u16_u32(a: this.mcu_blocks[0][$row6colX$]) ~mod* (this.quant_tables[args.q][$row6colX$] as base.u32)
// This code...
ca = (bq2 ~mod+ bq6) ~mod* $p0_541196100$
cb2 = ca ~mod+ (bq2 ~mod* $p0_765366865$)
cb6 = ca ~mod- (bq6 ~mod* $p1_847759065$)
// ...is equivalent to this more-SIMD-like code.
//
// cb2 = (bq2 ~mod* $p1_306562965$) ~mod+ (bq6 ~mod* $p0_541196100$)
// cb6 = (bq2 ~mod* $p0_541196100$) ~mod+ (bq6 ~mod* $m1_306562965$)
bq0 = this.util.sign_extend_convert_u16_u32(a: this.mcu_blocks[0][$row0colX$]) ~mod* (this.quant_tables[args.q][$row0colX$] as base.u32)
bq4 = this.util.sign_extend_convert_u16_u32(a: this.mcu_blocks[0][$row4colX$]) ~mod* (this.quant_tables[args.q][$row4colX$] as base.u32)
@@ -241,6 +289,7 @@ ci53 = bq5 ~mod+ bq3
ci71 = bq7 ~mod+ bq1
ci73 = bq7 ~mod+ bq3
// This code...
cj = (ci73 ~mod+ ci51) ~mod* $p1_175875602$
ck1 = bq1 ~mod* $p1_501321110$
@@ -261,6 +310,15 @@ ck3 ~mod+= ci53 ~mod+ cl73
ck5 ~mod+= ci53 ~mod+ cl51
ck7 ~mod+= ci71 ~mod+ cl73
// ...is equivalent to this more-SIMD-like code.
//
// cl73 = (ci73 ~mod* $m0_785694958$) ~mod+ (ci51 ~mod* $p1_175875602$)
// cl51 = (ci73 ~mod* $p1_175875602$) ~mod+ (ci51 ~mod* $p0_785694958$)
// ck1 = cl51 ~mod+ ((bq1 ~mod* $p0_601344887$) ~mod+ (bq7 ~mod* $m0_899976223$))
// ck3 = cl73 ~mod+ ((bq3 ~mod* $p0_509795579$) ~mod+ (bq5 ~mod* $m2_562915447$))
// ck5 = cl51 ~mod+ ((bq3 ~mod* $m2_562915447$) ~mod+ (bq5 ~mod* $m0_509795579$))
// ck7 = cl73 ~mod+ ((bq1 ~mod* $m0_899976223$) ~mod+ (bq7 ~mod* $m0_601344887$))
// Combine rows.
intermediate[$row0colX$] = this.util.sign_extend_rshift_u32(a: (cd0 ~mod+ ck1) ~mod+ (1 << 10), n: 11)
@@ -305,11 +363,17 @@ $advance$
in2 = intermediate[$rowYcol2$]
in6 = intermediate[$rowYcol6$]
// This code...
ra = (in2 ~mod+ in6) ~mod* $p0_541196100$
rb2 = ra ~mod+ (in2 ~mod* $p0_765366865$)
rb6 = ra ~mod- (in6 ~mod* $p1_847759065$)
// ...is equivalent to this more-SIMD-like code.
//
// rb2 = (in2 ~mod* $p1_306562965$) ~mod+ (in6 ~mod* $p0_541196100$)
// rb6 = (in2 ~mod* $p0_541196100$) ~mod+ (in6 ~mod* $m1_306562965$)
in0 = intermediate[$rowYcol0$]
in4 = intermediate[$rowYcol4$]
@@ -333,6 +397,7 @@ ri53 = in5 ~mod+ in3
ri71 = in7 ~mod+ in1
ri73 = in7 ~mod+ in3
// This code...
rj = (ri73 ~mod+ ri51) ~mod* $p1_175875602$
rk1 = in1 ~mod* $p1_501321110$
@@ -353,6 +418,15 @@ rk3 ~mod+= ri53 ~mod+ rl73
rk5 ~mod+= ri53 ~mod+ rl51
rk7 ~mod+= ri71 ~mod+ rl73
// ...is equivalent to this more-SIMD-like code.
//
// rl73 = (ri73 ~mod* $m0_785694958$) ~mod+ (ri51 ~mod* $p1_175875602$)
// rl51 = (ri73 ~mod* $p1_175875602$) ~mod+ (ri51 ~mod* $p0_785694958$)
// rk1 = rl51 ~mod+ ((in1 ~mod* $p0_601344887$) ~mod+ (in7 ~mod* $m0_899976223$))
// rk3 = rl73 ~mod+ ((in3 ~mod* $p0_509795579$) ~mod+ (in5 ~mod* $m2_562915447$))
// rk5 = rl51 ~mod+ ((in3 ~mod* $m2_562915447$) ~mod+ (in5 ~mod* $m0_509795579$))
// rk7 = rl73 ~mod+ ((in1 ~mod* $m0_899976223$) ~mod+ (in7 ~mod* $m0_601344887$))
// Combine columns.
$bounds_check$