From 97c4e85e0ec04a33822eea11ed06728813e0d757 Mon Sep 17 00:00:00 2001 From: Nigel Tao Date: Wed, 28 Jun 2023 12:37:25 +1000 Subject: [PATCH] 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. --- script/print-jpeg-idct-code.go | 150 +++++++++++----- std/jpeg/decode_idct_default.wuffs | 264 +++++++++++++++++++++++++++++ 2 files changed, 376 insertions(+), 38 deletions(-) diff --git a/script/print-jpeg-idct-code.go b/script/print-jpeg-idct-code.go index 5aa40379..8a2c6201 100644 --- a/script/print-jpeg-idct-code.go +++ b/script/print-jpeg-idct-code.go @@ -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. $bounds_check$ diff --git a/std/jpeg/decode_idct_default.wuffs b/std/jpeg/decode_idct_default.wuffs index bd300f65..a2a20df3 100644 --- a/std/jpeg/decode_idct_default.wuffs +++ b/std/jpeg/decode_idct_default.wuffs @@ -101,10 +101,14 @@ pri func decoder.decode_idct!(dst_buffer: slice base.u8, dst_stride: base.u64, q // p0_298631336 = 0x098E = 2446 // p0_390180644 = 0x0C7C = 3196 + // p0_509795579 = 0x1051 = 4177 // p0_541196100 = 0x1151 = 4433 + // p0_601344887 = 0x133E = 4926 // p0_765366865 = 0x187E = 6270 + // p0_785694958 = 0x1925 = 6437 // p0_899976223 = 0x1CCD = 7373 // p1_175875602 = 0x25A1 = 9633 + // p1_306562965 = 0x29CF = 10703 // p1_501321110 = 0x300B = 12299 // p1_847759065 = 0x3B21 = 15137 // p1_961570560 = 0x3EC5 = 16069 @@ -113,7 +117,11 @@ pri func decoder.decode_idct!(dst_buffer: slice base.u8, dst_stride: base.u64, q // p3_072711026 = 0x6254 = 25172 // // m0_390180644 = 0xFFFFF384 = 4294964100 + // m0_509795579 = 0xFFFFEFB0 = 4294963120 + // m0_601344887 = 0xFFFFECC1 = 4294962369 + // m0_785694958 = 0xFFFFE6DC = 4294960860 // m0_899976223 = 0xFFFFE333 = 4294959923 + // m1_306562965 = 0xFFFFD630 = 4294956592 // m1_961570560 = 0xFFFFC13B = 4294951227 // m2_562915447 = 0xFFFFADFD = 4294946301 @@ -146,11 +154,17 @@ pri func decoder.decode_idct!(dst_buffer: slice base.u8, dst_stride: base.u64, q bq2 = this.util.sign_extend_convert_u16_u32(a: this.mcu_blocks[0][0x10]) ~mod* (this.quant_tables[args.q][0x10] as base.u32) bq6 = this.util.sign_extend_convert_u16_u32(a: this.mcu_blocks[0][0x30]) ~mod* (this.quant_tables[args.q][0x30] as base.u32) + // This code... ca = (bq2 ~mod+ bq6) ~mod* 0x1151 cb2 = ca ~mod+ (bq2 ~mod* 0x187E) cb6 = ca ~mod- (bq6 ~mod* 0x3B21) + // ...is equivalent to this more-SIMD-like code. + // + // cb2 = (bq2 ~mod* 0x29CF) ~mod+ (bq6 ~mod* 0x1151) + // cb6 = (bq2 ~mod* 0x1151) ~mod+ (bq6 ~mod* 0xFFFFD630) + bq0 = this.util.sign_extend_convert_u16_u32(a: this.mcu_blocks[0][0x00]) ~mod* (this.quant_tables[args.q][0x00] as base.u32) bq4 = this.util.sign_extend_convert_u16_u32(a: this.mcu_blocks[0][0x20]) ~mod* (this.quant_tables[args.q][0x20] as base.u32) @@ -174,6 +188,7 @@ pri func decoder.decode_idct!(dst_buffer: slice base.u8, dst_stride: base.u64, q ci71 = bq7 ~mod+ bq1 ci73 = bq7 ~mod+ bq3 + // This code... cj = (ci73 ~mod+ ci51) ~mod* 0x25A1 ck1 = bq1 ~mod* 0x300B @@ -194,6 +209,15 @@ pri func decoder.decode_idct!(dst_buffer: slice base.u8, dst_stride: base.u64, q ck5 ~mod+= ci53 ~mod+ cl51 ck7 ~mod+= ci71 ~mod+ cl73 + // ...is equivalent to this more-SIMD-like code. + // + // cl73 = (ci73 ~mod* 0xFFFFE6DC) ~mod+ (ci51 ~mod* 0x25A1) + // cl51 = (ci73 ~mod* 0x25A1) ~mod+ (ci51 ~mod* 0x1925) + // ck1 = cl51 ~mod+ ((bq1 ~mod* 0x133E) ~mod+ (bq7 ~mod* 0xFFFFE333)) + // ck3 = cl73 ~mod+ ((bq3 ~mod* 0x1051) ~mod+ (bq5 ~mod* 0xFFFFADFD)) + // ck5 = cl51 ~mod+ ((bq3 ~mod* 0xFFFFADFD) ~mod+ (bq5 ~mod* 0xFFFFEFB0)) + // ck7 = cl73 ~mod+ ((bq1 ~mod* 0xFFFFE333) ~mod+ (bq7 ~mod* 0xFFFFECC1)) + // Combine rows. intermediate[0x00] = this.util.sign_extend_rshift_u32(a: (cd0 ~mod+ ck1) ~mod+ (1 << 10), n: 11) @@ -235,11 +259,17 @@ pri func decoder.decode_idct!(dst_buffer: slice base.u8, dst_stride: base.u64, q bq2 = this.util.sign_extend_convert_u16_u32(a: this.mcu_blocks[0][0x11]) ~mod* (this.quant_tables[args.q][0x11] as base.u32) bq6 = this.util.sign_extend_convert_u16_u32(a: this.mcu_blocks[0][0x31]) ~mod* (this.quant_tables[args.q][0x31] as base.u32) + // This code... ca = (bq2 ~mod+ bq6) ~mod* 0x1151 cb2 = ca ~mod+ (bq2 ~mod* 0x187E) cb6 = ca ~mod- (bq6 ~mod* 0x3B21) + // ...is equivalent to this more-SIMD-like code. + // + // cb2 = (bq2 ~mod* 0x29CF) ~mod+ (bq6 ~mod* 0x1151) + // cb6 = (bq2 ~mod* 0x1151) ~mod+ (bq6 ~mod* 0xFFFFD630) + bq0 = this.util.sign_extend_convert_u16_u32(a: this.mcu_blocks[0][0x01]) ~mod* (this.quant_tables[args.q][0x01] as base.u32) bq4 = this.util.sign_extend_convert_u16_u32(a: this.mcu_blocks[0][0x21]) ~mod* (this.quant_tables[args.q][0x21] as base.u32) @@ -263,6 +293,7 @@ pri func decoder.decode_idct!(dst_buffer: slice base.u8, dst_stride: base.u64, q ci71 = bq7 ~mod+ bq1 ci73 = bq7 ~mod+ bq3 + // This code... cj = (ci73 ~mod+ ci51) ~mod* 0x25A1 ck1 = bq1 ~mod* 0x300B @@ -283,6 +314,15 @@ pri func decoder.decode_idct!(dst_buffer: slice base.u8, dst_stride: base.u64, q ck5 ~mod+= ci53 ~mod+ cl51 ck7 ~mod+= ci71 ~mod+ cl73 + // ...is equivalent to this more-SIMD-like code. + // + // cl73 = (ci73 ~mod* 0xFFFFE6DC) ~mod+ (ci51 ~mod* 0x25A1) + // cl51 = (ci73 ~mod* 0x25A1) ~mod+ (ci51 ~mod* 0x1925) + // ck1 = cl51 ~mod+ ((bq1 ~mod* 0x133E) ~mod+ (bq7 ~mod* 0xFFFFE333)) + // ck3 = cl73 ~mod+ ((bq3 ~mod* 0x1051) ~mod+ (bq5 ~mod* 0xFFFFADFD)) + // ck5 = cl51 ~mod+ ((bq3 ~mod* 0xFFFFADFD) ~mod+ (bq5 ~mod* 0xFFFFEFB0)) + // ck7 = cl73 ~mod+ ((bq1 ~mod* 0xFFFFE333) ~mod+ (bq7 ~mod* 0xFFFFECC1)) + // Combine rows. intermediate[0x01] = this.util.sign_extend_rshift_u32(a: (cd0 ~mod+ ck1) ~mod+ (1 << 10), n: 11) @@ -324,11 +364,17 @@ pri func decoder.decode_idct!(dst_buffer: slice base.u8, dst_stride: base.u64, q bq2 = this.util.sign_extend_convert_u16_u32(a: this.mcu_blocks[0][0x12]) ~mod* (this.quant_tables[args.q][0x12] as base.u32) bq6 = this.util.sign_extend_convert_u16_u32(a: this.mcu_blocks[0][0x32]) ~mod* (this.quant_tables[args.q][0x32] as base.u32) + // This code... ca = (bq2 ~mod+ bq6) ~mod* 0x1151 cb2 = ca ~mod+ (bq2 ~mod* 0x187E) cb6 = ca ~mod- (bq6 ~mod* 0x3B21) + // ...is equivalent to this more-SIMD-like code. + // + // cb2 = (bq2 ~mod* 0x29CF) ~mod+ (bq6 ~mod* 0x1151) + // cb6 = (bq2 ~mod* 0x1151) ~mod+ (bq6 ~mod* 0xFFFFD630) + bq0 = this.util.sign_extend_convert_u16_u32(a: this.mcu_blocks[0][0x02]) ~mod* (this.quant_tables[args.q][0x02] as base.u32) bq4 = this.util.sign_extend_convert_u16_u32(a: this.mcu_blocks[0][0x22]) ~mod* (this.quant_tables[args.q][0x22] as base.u32) @@ -352,6 +398,7 @@ pri func decoder.decode_idct!(dst_buffer: slice base.u8, dst_stride: base.u64, q ci71 = bq7 ~mod+ bq1 ci73 = bq7 ~mod+ bq3 + // This code... cj = (ci73 ~mod+ ci51) ~mod* 0x25A1 ck1 = bq1 ~mod* 0x300B @@ -372,6 +419,15 @@ pri func decoder.decode_idct!(dst_buffer: slice base.u8, dst_stride: base.u64, q ck5 ~mod+= ci53 ~mod+ cl51 ck7 ~mod+= ci71 ~mod+ cl73 + // ...is equivalent to this more-SIMD-like code. + // + // cl73 = (ci73 ~mod* 0xFFFFE6DC) ~mod+ (ci51 ~mod* 0x25A1) + // cl51 = (ci73 ~mod* 0x25A1) ~mod+ (ci51 ~mod* 0x1925) + // ck1 = cl51 ~mod+ ((bq1 ~mod* 0x133E) ~mod+ (bq7 ~mod* 0xFFFFE333)) + // ck3 = cl73 ~mod+ ((bq3 ~mod* 0x1051) ~mod+ (bq5 ~mod* 0xFFFFADFD)) + // ck5 = cl51 ~mod+ ((bq3 ~mod* 0xFFFFADFD) ~mod+ (bq5 ~mod* 0xFFFFEFB0)) + // ck7 = cl73 ~mod+ ((bq1 ~mod* 0xFFFFE333) ~mod+ (bq7 ~mod* 0xFFFFECC1)) + // Combine rows. intermediate[0x02] = this.util.sign_extend_rshift_u32(a: (cd0 ~mod+ ck1) ~mod+ (1 << 10), n: 11) @@ -413,11 +469,17 @@ pri func decoder.decode_idct!(dst_buffer: slice base.u8, dst_stride: base.u64, q bq2 = this.util.sign_extend_convert_u16_u32(a: this.mcu_blocks[0][0x13]) ~mod* (this.quant_tables[args.q][0x13] as base.u32) bq6 = this.util.sign_extend_convert_u16_u32(a: this.mcu_blocks[0][0x33]) ~mod* (this.quant_tables[args.q][0x33] as base.u32) + // This code... ca = (bq2 ~mod+ bq6) ~mod* 0x1151 cb2 = ca ~mod+ (bq2 ~mod* 0x187E) cb6 = ca ~mod- (bq6 ~mod* 0x3B21) + // ...is equivalent to this more-SIMD-like code. + // + // cb2 = (bq2 ~mod* 0x29CF) ~mod+ (bq6 ~mod* 0x1151) + // cb6 = (bq2 ~mod* 0x1151) ~mod+ (bq6 ~mod* 0xFFFFD630) + bq0 = this.util.sign_extend_convert_u16_u32(a: this.mcu_blocks[0][0x03]) ~mod* (this.quant_tables[args.q][0x03] as base.u32) bq4 = this.util.sign_extend_convert_u16_u32(a: this.mcu_blocks[0][0x23]) ~mod* (this.quant_tables[args.q][0x23] as base.u32) @@ -441,6 +503,7 @@ pri func decoder.decode_idct!(dst_buffer: slice base.u8, dst_stride: base.u64, q ci71 = bq7 ~mod+ bq1 ci73 = bq7 ~mod+ bq3 + // This code... cj = (ci73 ~mod+ ci51) ~mod* 0x25A1 ck1 = bq1 ~mod* 0x300B @@ -461,6 +524,15 @@ pri func decoder.decode_idct!(dst_buffer: slice base.u8, dst_stride: base.u64, q ck5 ~mod+= ci53 ~mod+ cl51 ck7 ~mod+= ci71 ~mod+ cl73 + // ...is equivalent to this more-SIMD-like code. + // + // cl73 = (ci73 ~mod* 0xFFFFE6DC) ~mod+ (ci51 ~mod* 0x25A1) + // cl51 = (ci73 ~mod* 0x25A1) ~mod+ (ci51 ~mod* 0x1925) + // ck1 = cl51 ~mod+ ((bq1 ~mod* 0x133E) ~mod+ (bq7 ~mod* 0xFFFFE333)) + // ck3 = cl73 ~mod+ ((bq3 ~mod* 0x1051) ~mod+ (bq5 ~mod* 0xFFFFADFD)) + // ck5 = cl51 ~mod+ ((bq3 ~mod* 0xFFFFADFD) ~mod+ (bq5 ~mod* 0xFFFFEFB0)) + // ck7 = cl73 ~mod+ ((bq1 ~mod* 0xFFFFE333) ~mod+ (bq7 ~mod* 0xFFFFECC1)) + // Combine rows. intermediate[0x03] = this.util.sign_extend_rshift_u32(a: (cd0 ~mod+ ck1) ~mod+ (1 << 10), n: 11) @@ -502,11 +574,17 @@ pri func decoder.decode_idct!(dst_buffer: slice base.u8, dst_stride: base.u64, q bq2 = this.util.sign_extend_convert_u16_u32(a: this.mcu_blocks[0][0x14]) ~mod* (this.quant_tables[args.q][0x14] as base.u32) bq6 = this.util.sign_extend_convert_u16_u32(a: this.mcu_blocks[0][0x34]) ~mod* (this.quant_tables[args.q][0x34] as base.u32) + // This code... ca = (bq2 ~mod+ bq6) ~mod* 0x1151 cb2 = ca ~mod+ (bq2 ~mod* 0x187E) cb6 = ca ~mod- (bq6 ~mod* 0x3B21) + // ...is equivalent to this more-SIMD-like code. + // + // cb2 = (bq2 ~mod* 0x29CF) ~mod+ (bq6 ~mod* 0x1151) + // cb6 = (bq2 ~mod* 0x1151) ~mod+ (bq6 ~mod* 0xFFFFD630) + bq0 = this.util.sign_extend_convert_u16_u32(a: this.mcu_blocks[0][0x04]) ~mod* (this.quant_tables[args.q][0x04] as base.u32) bq4 = this.util.sign_extend_convert_u16_u32(a: this.mcu_blocks[0][0x24]) ~mod* (this.quant_tables[args.q][0x24] as base.u32) @@ -530,6 +608,7 @@ pri func decoder.decode_idct!(dst_buffer: slice base.u8, dst_stride: base.u64, q ci71 = bq7 ~mod+ bq1 ci73 = bq7 ~mod+ bq3 + // This code... cj = (ci73 ~mod+ ci51) ~mod* 0x25A1 ck1 = bq1 ~mod* 0x300B @@ -550,6 +629,15 @@ pri func decoder.decode_idct!(dst_buffer: slice base.u8, dst_stride: base.u64, q ck5 ~mod+= ci53 ~mod+ cl51 ck7 ~mod+= ci71 ~mod+ cl73 + // ...is equivalent to this more-SIMD-like code. + // + // cl73 = (ci73 ~mod* 0xFFFFE6DC) ~mod+ (ci51 ~mod* 0x25A1) + // cl51 = (ci73 ~mod* 0x25A1) ~mod+ (ci51 ~mod* 0x1925) + // ck1 = cl51 ~mod+ ((bq1 ~mod* 0x133E) ~mod+ (bq7 ~mod* 0xFFFFE333)) + // ck3 = cl73 ~mod+ ((bq3 ~mod* 0x1051) ~mod+ (bq5 ~mod* 0xFFFFADFD)) + // ck5 = cl51 ~mod+ ((bq3 ~mod* 0xFFFFADFD) ~mod+ (bq5 ~mod* 0xFFFFEFB0)) + // ck7 = cl73 ~mod+ ((bq1 ~mod* 0xFFFFE333) ~mod+ (bq7 ~mod* 0xFFFFECC1)) + // Combine rows. intermediate[0x04] = this.util.sign_extend_rshift_u32(a: (cd0 ~mod+ ck1) ~mod+ (1 << 10), n: 11) @@ -591,11 +679,17 @@ pri func decoder.decode_idct!(dst_buffer: slice base.u8, dst_stride: base.u64, q bq2 = this.util.sign_extend_convert_u16_u32(a: this.mcu_blocks[0][0x15]) ~mod* (this.quant_tables[args.q][0x15] as base.u32) bq6 = this.util.sign_extend_convert_u16_u32(a: this.mcu_blocks[0][0x35]) ~mod* (this.quant_tables[args.q][0x35] as base.u32) + // This code... ca = (bq2 ~mod+ bq6) ~mod* 0x1151 cb2 = ca ~mod+ (bq2 ~mod* 0x187E) cb6 = ca ~mod- (bq6 ~mod* 0x3B21) + // ...is equivalent to this more-SIMD-like code. + // + // cb2 = (bq2 ~mod* 0x29CF) ~mod+ (bq6 ~mod* 0x1151) + // cb6 = (bq2 ~mod* 0x1151) ~mod+ (bq6 ~mod* 0xFFFFD630) + bq0 = this.util.sign_extend_convert_u16_u32(a: this.mcu_blocks[0][0x05]) ~mod* (this.quant_tables[args.q][0x05] as base.u32) bq4 = this.util.sign_extend_convert_u16_u32(a: this.mcu_blocks[0][0x25]) ~mod* (this.quant_tables[args.q][0x25] as base.u32) @@ -619,6 +713,7 @@ pri func decoder.decode_idct!(dst_buffer: slice base.u8, dst_stride: base.u64, q ci71 = bq7 ~mod+ bq1 ci73 = bq7 ~mod+ bq3 + // This code... cj = (ci73 ~mod+ ci51) ~mod* 0x25A1 ck1 = bq1 ~mod* 0x300B @@ -639,6 +734,15 @@ pri func decoder.decode_idct!(dst_buffer: slice base.u8, dst_stride: base.u64, q ck5 ~mod+= ci53 ~mod+ cl51 ck7 ~mod+= ci71 ~mod+ cl73 + // ...is equivalent to this more-SIMD-like code. + // + // cl73 = (ci73 ~mod* 0xFFFFE6DC) ~mod+ (ci51 ~mod* 0x25A1) + // cl51 = (ci73 ~mod* 0x25A1) ~mod+ (ci51 ~mod* 0x1925) + // ck1 = cl51 ~mod+ ((bq1 ~mod* 0x133E) ~mod+ (bq7 ~mod* 0xFFFFE333)) + // ck3 = cl73 ~mod+ ((bq3 ~mod* 0x1051) ~mod+ (bq5 ~mod* 0xFFFFADFD)) + // ck5 = cl51 ~mod+ ((bq3 ~mod* 0xFFFFADFD) ~mod+ (bq5 ~mod* 0xFFFFEFB0)) + // ck7 = cl73 ~mod+ ((bq1 ~mod* 0xFFFFE333) ~mod+ (bq7 ~mod* 0xFFFFECC1)) + // Combine rows. intermediate[0x05] = this.util.sign_extend_rshift_u32(a: (cd0 ~mod+ ck1) ~mod+ (1 << 10), n: 11) @@ -680,11 +784,17 @@ pri func decoder.decode_idct!(dst_buffer: slice base.u8, dst_stride: base.u64, q bq2 = this.util.sign_extend_convert_u16_u32(a: this.mcu_blocks[0][0x16]) ~mod* (this.quant_tables[args.q][0x16] as base.u32) bq6 = this.util.sign_extend_convert_u16_u32(a: this.mcu_blocks[0][0x36]) ~mod* (this.quant_tables[args.q][0x36] as base.u32) + // This code... ca = (bq2 ~mod+ bq6) ~mod* 0x1151 cb2 = ca ~mod+ (bq2 ~mod* 0x187E) cb6 = ca ~mod- (bq6 ~mod* 0x3B21) + // ...is equivalent to this more-SIMD-like code. + // + // cb2 = (bq2 ~mod* 0x29CF) ~mod+ (bq6 ~mod* 0x1151) + // cb6 = (bq2 ~mod* 0x1151) ~mod+ (bq6 ~mod* 0xFFFFD630) + bq0 = this.util.sign_extend_convert_u16_u32(a: this.mcu_blocks[0][0x06]) ~mod* (this.quant_tables[args.q][0x06] as base.u32) bq4 = this.util.sign_extend_convert_u16_u32(a: this.mcu_blocks[0][0x26]) ~mod* (this.quant_tables[args.q][0x26] as base.u32) @@ -708,6 +818,7 @@ pri func decoder.decode_idct!(dst_buffer: slice base.u8, dst_stride: base.u64, q ci71 = bq7 ~mod+ bq1 ci73 = bq7 ~mod+ bq3 + // This code... cj = (ci73 ~mod+ ci51) ~mod* 0x25A1 ck1 = bq1 ~mod* 0x300B @@ -728,6 +839,15 @@ pri func decoder.decode_idct!(dst_buffer: slice base.u8, dst_stride: base.u64, q ck5 ~mod+= ci53 ~mod+ cl51 ck7 ~mod+= ci71 ~mod+ cl73 + // ...is equivalent to this more-SIMD-like code. + // + // cl73 = (ci73 ~mod* 0xFFFFE6DC) ~mod+ (ci51 ~mod* 0x25A1) + // cl51 = (ci73 ~mod* 0x25A1) ~mod+ (ci51 ~mod* 0x1925) + // ck1 = cl51 ~mod+ ((bq1 ~mod* 0x133E) ~mod+ (bq7 ~mod* 0xFFFFE333)) + // ck3 = cl73 ~mod+ ((bq3 ~mod* 0x1051) ~mod+ (bq5 ~mod* 0xFFFFADFD)) + // ck5 = cl51 ~mod+ ((bq3 ~mod* 0xFFFFADFD) ~mod+ (bq5 ~mod* 0xFFFFEFB0)) + // ck7 = cl73 ~mod+ ((bq1 ~mod* 0xFFFFE333) ~mod+ (bq7 ~mod* 0xFFFFECC1)) + // Combine rows. intermediate[0x06] = this.util.sign_extend_rshift_u32(a: (cd0 ~mod+ ck1) ~mod+ (1 << 10), n: 11) @@ -769,11 +889,17 @@ pri func decoder.decode_idct!(dst_buffer: slice base.u8, dst_stride: base.u64, q bq2 = this.util.sign_extend_convert_u16_u32(a: this.mcu_blocks[0][0x17]) ~mod* (this.quant_tables[args.q][0x17] as base.u32) bq6 = this.util.sign_extend_convert_u16_u32(a: this.mcu_blocks[0][0x37]) ~mod* (this.quant_tables[args.q][0x37] as base.u32) + // This code... ca = (bq2 ~mod+ bq6) ~mod* 0x1151 cb2 = ca ~mod+ (bq2 ~mod* 0x187E) cb6 = ca ~mod- (bq6 ~mod* 0x3B21) + // ...is equivalent to this more-SIMD-like code. + // + // cb2 = (bq2 ~mod* 0x29CF) ~mod+ (bq6 ~mod* 0x1151) + // cb6 = (bq2 ~mod* 0x1151) ~mod+ (bq6 ~mod* 0xFFFFD630) + bq0 = this.util.sign_extend_convert_u16_u32(a: this.mcu_blocks[0][0x07]) ~mod* (this.quant_tables[args.q][0x07] as base.u32) bq4 = this.util.sign_extend_convert_u16_u32(a: this.mcu_blocks[0][0x27]) ~mod* (this.quant_tables[args.q][0x27] as base.u32) @@ -797,6 +923,7 @@ pri func decoder.decode_idct!(dst_buffer: slice base.u8, dst_stride: base.u64, q ci71 = bq7 ~mod+ bq1 ci73 = bq7 ~mod+ bq3 + // This code... cj = (ci73 ~mod+ ci51) ~mod* 0x25A1 ck1 = bq1 ~mod* 0x300B @@ -817,6 +944,15 @@ pri func decoder.decode_idct!(dst_buffer: slice base.u8, dst_stride: base.u64, q ck5 ~mod+= ci53 ~mod+ cl51 ck7 ~mod+= ci71 ~mod+ cl73 + // ...is equivalent to this more-SIMD-like code. + // + // cl73 = (ci73 ~mod* 0xFFFFE6DC) ~mod+ (ci51 ~mod* 0x25A1) + // cl51 = (ci73 ~mod* 0x25A1) ~mod+ (ci51 ~mod* 0x1925) + // ck1 = cl51 ~mod+ ((bq1 ~mod* 0x133E) ~mod+ (bq7 ~mod* 0xFFFFE333)) + // ck3 = cl73 ~mod+ ((bq3 ~mod* 0x1051) ~mod+ (bq5 ~mod* 0xFFFFADFD)) + // ck5 = cl51 ~mod+ ((bq3 ~mod* 0xFFFFADFD) ~mod+ (bq5 ~mod* 0xFFFFEFB0)) + // ck7 = cl73 ~mod+ ((bq1 ~mod* 0xFFFFE333) ~mod+ (bq7 ~mod* 0xFFFFECC1)) + // Combine rows. intermediate[0x07] = this.util.sign_extend_rshift_u32(a: (cd0 ~mod+ ck1) ~mod+ (1 << 10), n: 11) @@ -863,11 +999,17 @@ pri func decoder.decode_idct!(dst_buffer: slice base.u8, dst_stride: base.u64, q in2 = intermediate[0x02] in6 = intermediate[0x06] + // This code... ra = (in2 ~mod+ in6) ~mod* 0x1151 rb2 = ra ~mod+ (in2 ~mod* 0x187E) rb6 = ra ~mod- (in6 ~mod* 0x3B21) + // ...is equivalent to this more-SIMD-like code. + // + // rb2 = (in2 ~mod* 0x29CF) ~mod+ (in6 ~mod* 0x1151) + // rb6 = (in2 ~mod* 0x1151) ~mod+ (in6 ~mod* 0xFFFFD630) + in0 = intermediate[0x00] in4 = intermediate[0x04] @@ -891,6 +1033,7 @@ pri func decoder.decode_idct!(dst_buffer: slice base.u8, dst_stride: base.u64, q ri71 = in7 ~mod+ in1 ri73 = in7 ~mod+ in3 + // This code... rj = (ri73 ~mod+ ri51) ~mod* 0x25A1 rk1 = in1 ~mod* 0x300B @@ -911,6 +1054,15 @@ pri func decoder.decode_idct!(dst_buffer: slice base.u8, dst_stride: base.u64, q rk5 ~mod+= ri53 ~mod+ rl51 rk7 ~mod+= ri71 ~mod+ rl73 + // ...is equivalent to this more-SIMD-like code. + // + // rl73 = (ri73 ~mod* 0xFFFFE6DC) ~mod+ (ri51 ~mod* 0x25A1) + // rl51 = (ri73 ~mod* 0x25A1) ~mod+ (ri51 ~mod* 0x1925) + // rk1 = rl51 ~mod+ ((in1 ~mod* 0x133E) ~mod+ (in7 ~mod* 0xFFFFE333)) + // rk3 = rl73 ~mod+ ((in3 ~mod* 0x1051) ~mod+ (in5 ~mod* 0xFFFFADFD)) + // rk5 = rl51 ~mod+ ((in3 ~mod* 0xFFFFADFD) ~mod+ (in5 ~mod* 0xFFFFEFB0)) + // rk7 = rl73 ~mod+ ((in1 ~mod* 0xFFFFE333) ~mod+ (in7 ~mod* 0xFFFFECC1)) + // Combine columns. if args.dst_stride > args.dst_buffer.length() { @@ -964,11 +1116,17 @@ pri func decoder.decode_idct!(dst_buffer: slice base.u8, dst_stride: base.u64, q in2 = intermediate[0x0A] in6 = intermediate[0x0E] + // This code... ra = (in2 ~mod+ in6) ~mod* 0x1151 rb2 = ra ~mod+ (in2 ~mod* 0x187E) rb6 = ra ~mod- (in6 ~mod* 0x3B21) + // ...is equivalent to this more-SIMD-like code. + // + // rb2 = (in2 ~mod* 0x29CF) ~mod+ (in6 ~mod* 0x1151) + // rb6 = (in2 ~mod* 0x1151) ~mod+ (in6 ~mod* 0xFFFFD630) + in0 = intermediate[0x08] in4 = intermediate[0x0C] @@ -992,6 +1150,7 @@ pri func decoder.decode_idct!(dst_buffer: slice base.u8, dst_stride: base.u64, q ri71 = in7 ~mod+ in1 ri73 = in7 ~mod+ in3 + // This code... rj = (ri73 ~mod+ ri51) ~mod* 0x25A1 rk1 = in1 ~mod* 0x300B @@ -1012,6 +1171,15 @@ pri func decoder.decode_idct!(dst_buffer: slice base.u8, dst_stride: base.u64, q rk5 ~mod+= ri53 ~mod+ rl51 rk7 ~mod+= ri71 ~mod+ rl73 + // ...is equivalent to this more-SIMD-like code. + // + // rl73 = (ri73 ~mod* 0xFFFFE6DC) ~mod+ (ri51 ~mod* 0x25A1) + // rl51 = (ri73 ~mod* 0x25A1) ~mod+ (ri51 ~mod* 0x1925) + // rk1 = rl51 ~mod+ ((in1 ~mod* 0x133E) ~mod+ (in7 ~mod* 0xFFFFE333)) + // rk3 = rl73 ~mod+ ((in3 ~mod* 0x1051) ~mod+ (in5 ~mod* 0xFFFFADFD)) + // rk5 = rl51 ~mod+ ((in3 ~mod* 0xFFFFADFD) ~mod+ (in5 ~mod* 0xFFFFEFB0)) + // rk7 = rl73 ~mod+ ((in1 ~mod* 0xFFFFE333) ~mod+ (in7 ~mod* 0xFFFFECC1)) + // Combine columns. if args.dst_stride > args.dst_buffer.length() { @@ -1065,11 +1233,17 @@ pri func decoder.decode_idct!(dst_buffer: slice base.u8, dst_stride: base.u64, q in2 = intermediate[0x12] in6 = intermediate[0x16] + // This code... ra = (in2 ~mod+ in6) ~mod* 0x1151 rb2 = ra ~mod+ (in2 ~mod* 0x187E) rb6 = ra ~mod- (in6 ~mod* 0x3B21) + // ...is equivalent to this more-SIMD-like code. + // + // rb2 = (in2 ~mod* 0x29CF) ~mod+ (in6 ~mod* 0x1151) + // rb6 = (in2 ~mod* 0x1151) ~mod+ (in6 ~mod* 0xFFFFD630) + in0 = intermediate[0x10] in4 = intermediate[0x14] @@ -1093,6 +1267,7 @@ pri func decoder.decode_idct!(dst_buffer: slice base.u8, dst_stride: base.u64, q ri71 = in7 ~mod+ in1 ri73 = in7 ~mod+ in3 + // This code... rj = (ri73 ~mod+ ri51) ~mod* 0x25A1 rk1 = in1 ~mod* 0x300B @@ -1113,6 +1288,15 @@ pri func decoder.decode_idct!(dst_buffer: slice base.u8, dst_stride: base.u64, q rk5 ~mod+= ri53 ~mod+ rl51 rk7 ~mod+= ri71 ~mod+ rl73 + // ...is equivalent to this more-SIMD-like code. + // + // rl73 = (ri73 ~mod* 0xFFFFE6DC) ~mod+ (ri51 ~mod* 0x25A1) + // rl51 = (ri73 ~mod* 0x25A1) ~mod+ (ri51 ~mod* 0x1925) + // rk1 = rl51 ~mod+ ((in1 ~mod* 0x133E) ~mod+ (in7 ~mod* 0xFFFFE333)) + // rk3 = rl73 ~mod+ ((in3 ~mod* 0x1051) ~mod+ (in5 ~mod* 0xFFFFADFD)) + // rk5 = rl51 ~mod+ ((in3 ~mod* 0xFFFFADFD) ~mod+ (in5 ~mod* 0xFFFFEFB0)) + // rk7 = rl73 ~mod+ ((in1 ~mod* 0xFFFFE333) ~mod+ (in7 ~mod* 0xFFFFECC1)) + // Combine columns. if args.dst_stride > args.dst_buffer.length() { @@ -1166,11 +1350,17 @@ pri func decoder.decode_idct!(dst_buffer: slice base.u8, dst_stride: base.u64, q in2 = intermediate[0x1A] in6 = intermediate[0x1E] + // This code... ra = (in2 ~mod+ in6) ~mod* 0x1151 rb2 = ra ~mod+ (in2 ~mod* 0x187E) rb6 = ra ~mod- (in6 ~mod* 0x3B21) + // ...is equivalent to this more-SIMD-like code. + // + // rb2 = (in2 ~mod* 0x29CF) ~mod+ (in6 ~mod* 0x1151) + // rb6 = (in2 ~mod* 0x1151) ~mod+ (in6 ~mod* 0xFFFFD630) + in0 = intermediate[0x18] in4 = intermediate[0x1C] @@ -1194,6 +1384,7 @@ pri func decoder.decode_idct!(dst_buffer: slice base.u8, dst_stride: base.u64, q ri71 = in7 ~mod+ in1 ri73 = in7 ~mod+ in3 + // This code... rj = (ri73 ~mod+ ri51) ~mod* 0x25A1 rk1 = in1 ~mod* 0x300B @@ -1214,6 +1405,15 @@ pri func decoder.decode_idct!(dst_buffer: slice base.u8, dst_stride: base.u64, q rk5 ~mod+= ri53 ~mod+ rl51 rk7 ~mod+= ri71 ~mod+ rl73 + // ...is equivalent to this more-SIMD-like code. + // + // rl73 = (ri73 ~mod* 0xFFFFE6DC) ~mod+ (ri51 ~mod* 0x25A1) + // rl51 = (ri73 ~mod* 0x25A1) ~mod+ (ri51 ~mod* 0x1925) + // rk1 = rl51 ~mod+ ((in1 ~mod* 0x133E) ~mod+ (in7 ~mod* 0xFFFFE333)) + // rk3 = rl73 ~mod+ ((in3 ~mod* 0x1051) ~mod+ (in5 ~mod* 0xFFFFADFD)) + // rk5 = rl51 ~mod+ ((in3 ~mod* 0xFFFFADFD) ~mod+ (in5 ~mod* 0xFFFFEFB0)) + // rk7 = rl73 ~mod+ ((in1 ~mod* 0xFFFFE333) ~mod+ (in7 ~mod* 0xFFFFECC1)) + // Combine columns. if args.dst_stride > args.dst_buffer.length() { @@ -1267,11 +1467,17 @@ pri func decoder.decode_idct!(dst_buffer: slice base.u8, dst_stride: base.u64, q in2 = intermediate[0x22] in6 = intermediate[0x26] + // This code... ra = (in2 ~mod+ in6) ~mod* 0x1151 rb2 = ra ~mod+ (in2 ~mod* 0x187E) rb6 = ra ~mod- (in6 ~mod* 0x3B21) + // ...is equivalent to this more-SIMD-like code. + // + // rb2 = (in2 ~mod* 0x29CF) ~mod+ (in6 ~mod* 0x1151) + // rb6 = (in2 ~mod* 0x1151) ~mod+ (in6 ~mod* 0xFFFFD630) + in0 = intermediate[0x20] in4 = intermediate[0x24] @@ -1295,6 +1501,7 @@ pri func decoder.decode_idct!(dst_buffer: slice base.u8, dst_stride: base.u64, q ri71 = in7 ~mod+ in1 ri73 = in7 ~mod+ in3 + // This code... rj = (ri73 ~mod+ ri51) ~mod* 0x25A1 rk1 = in1 ~mod* 0x300B @@ -1315,6 +1522,15 @@ pri func decoder.decode_idct!(dst_buffer: slice base.u8, dst_stride: base.u64, q rk5 ~mod+= ri53 ~mod+ rl51 rk7 ~mod+= ri71 ~mod+ rl73 + // ...is equivalent to this more-SIMD-like code. + // + // rl73 = (ri73 ~mod* 0xFFFFE6DC) ~mod+ (ri51 ~mod* 0x25A1) + // rl51 = (ri73 ~mod* 0x25A1) ~mod+ (ri51 ~mod* 0x1925) + // rk1 = rl51 ~mod+ ((in1 ~mod* 0x133E) ~mod+ (in7 ~mod* 0xFFFFE333)) + // rk3 = rl73 ~mod+ ((in3 ~mod* 0x1051) ~mod+ (in5 ~mod* 0xFFFFADFD)) + // rk5 = rl51 ~mod+ ((in3 ~mod* 0xFFFFADFD) ~mod+ (in5 ~mod* 0xFFFFEFB0)) + // rk7 = rl73 ~mod+ ((in1 ~mod* 0xFFFFE333) ~mod+ (in7 ~mod* 0xFFFFECC1)) + // Combine columns. if args.dst_stride > args.dst_buffer.length() { @@ -1368,11 +1584,17 @@ pri func decoder.decode_idct!(dst_buffer: slice base.u8, dst_stride: base.u64, q in2 = intermediate[0x2A] in6 = intermediate[0x2E] + // This code... ra = (in2 ~mod+ in6) ~mod* 0x1151 rb2 = ra ~mod+ (in2 ~mod* 0x187E) rb6 = ra ~mod- (in6 ~mod* 0x3B21) + // ...is equivalent to this more-SIMD-like code. + // + // rb2 = (in2 ~mod* 0x29CF) ~mod+ (in6 ~mod* 0x1151) + // rb6 = (in2 ~mod* 0x1151) ~mod+ (in6 ~mod* 0xFFFFD630) + in0 = intermediate[0x28] in4 = intermediate[0x2C] @@ -1396,6 +1618,7 @@ pri func decoder.decode_idct!(dst_buffer: slice base.u8, dst_stride: base.u64, q ri71 = in7 ~mod+ in1 ri73 = in7 ~mod+ in3 + // This code... rj = (ri73 ~mod+ ri51) ~mod* 0x25A1 rk1 = in1 ~mod* 0x300B @@ -1416,6 +1639,15 @@ pri func decoder.decode_idct!(dst_buffer: slice base.u8, dst_stride: base.u64, q rk5 ~mod+= ri53 ~mod+ rl51 rk7 ~mod+= ri71 ~mod+ rl73 + // ...is equivalent to this more-SIMD-like code. + // + // rl73 = (ri73 ~mod* 0xFFFFE6DC) ~mod+ (ri51 ~mod* 0x25A1) + // rl51 = (ri73 ~mod* 0x25A1) ~mod+ (ri51 ~mod* 0x1925) + // rk1 = rl51 ~mod+ ((in1 ~mod* 0x133E) ~mod+ (in7 ~mod* 0xFFFFE333)) + // rk3 = rl73 ~mod+ ((in3 ~mod* 0x1051) ~mod+ (in5 ~mod* 0xFFFFADFD)) + // rk5 = rl51 ~mod+ ((in3 ~mod* 0xFFFFADFD) ~mod+ (in5 ~mod* 0xFFFFEFB0)) + // rk7 = rl73 ~mod+ ((in1 ~mod* 0xFFFFE333) ~mod+ (in7 ~mod* 0xFFFFECC1)) + // Combine columns. if args.dst_stride > args.dst_buffer.length() { @@ -1469,11 +1701,17 @@ pri func decoder.decode_idct!(dst_buffer: slice base.u8, dst_stride: base.u64, q in2 = intermediate[0x32] in6 = intermediate[0x36] + // This code... ra = (in2 ~mod+ in6) ~mod* 0x1151 rb2 = ra ~mod+ (in2 ~mod* 0x187E) rb6 = ra ~mod- (in6 ~mod* 0x3B21) + // ...is equivalent to this more-SIMD-like code. + // + // rb2 = (in2 ~mod* 0x29CF) ~mod+ (in6 ~mod* 0x1151) + // rb6 = (in2 ~mod* 0x1151) ~mod+ (in6 ~mod* 0xFFFFD630) + in0 = intermediate[0x30] in4 = intermediate[0x34] @@ -1497,6 +1735,7 @@ pri func decoder.decode_idct!(dst_buffer: slice base.u8, dst_stride: base.u64, q ri71 = in7 ~mod+ in1 ri73 = in7 ~mod+ in3 + // This code... rj = (ri73 ~mod+ ri51) ~mod* 0x25A1 rk1 = in1 ~mod* 0x300B @@ -1517,6 +1756,15 @@ pri func decoder.decode_idct!(dst_buffer: slice base.u8, dst_stride: base.u64, q rk5 ~mod+= ri53 ~mod+ rl51 rk7 ~mod+= ri71 ~mod+ rl73 + // ...is equivalent to this more-SIMD-like code. + // + // rl73 = (ri73 ~mod* 0xFFFFE6DC) ~mod+ (ri51 ~mod* 0x25A1) + // rl51 = (ri73 ~mod* 0x25A1) ~mod+ (ri51 ~mod* 0x1925) + // rk1 = rl51 ~mod+ ((in1 ~mod* 0x133E) ~mod+ (in7 ~mod* 0xFFFFE333)) + // rk3 = rl73 ~mod+ ((in3 ~mod* 0x1051) ~mod+ (in5 ~mod* 0xFFFFADFD)) + // rk5 = rl51 ~mod+ ((in3 ~mod* 0xFFFFADFD) ~mod+ (in5 ~mod* 0xFFFFEFB0)) + // rk7 = rl73 ~mod+ ((in1 ~mod* 0xFFFFE333) ~mod+ (in7 ~mod* 0xFFFFECC1)) + // Combine columns. if args.dst_stride > args.dst_buffer.length() { @@ -1567,11 +1815,17 @@ pri func decoder.decode_idct!(dst_buffer: slice base.u8, dst_stride: base.u64, q in2 = intermediate[0x3A] in6 = intermediate[0x3E] + // This code... ra = (in2 ~mod+ in6) ~mod* 0x1151 rb2 = ra ~mod+ (in2 ~mod* 0x187E) rb6 = ra ~mod- (in6 ~mod* 0x3B21) + // ...is equivalent to this more-SIMD-like code. + // + // rb2 = (in2 ~mod* 0x29CF) ~mod+ (in6 ~mod* 0x1151) + // rb6 = (in2 ~mod* 0x1151) ~mod+ (in6 ~mod* 0xFFFFD630) + in0 = intermediate[0x38] in4 = intermediate[0x3C] @@ -1595,6 +1849,7 @@ pri func decoder.decode_idct!(dst_buffer: slice base.u8, dst_stride: base.u64, q ri71 = in7 ~mod+ in1 ri73 = in7 ~mod+ in3 + // This code... rj = (ri73 ~mod+ ri51) ~mod* 0x25A1 rk1 = in1 ~mod* 0x300B @@ -1615,6 +1870,15 @@ pri func decoder.decode_idct!(dst_buffer: slice base.u8, dst_stride: base.u64, q rk5 ~mod+= ri53 ~mod+ rl51 rk7 ~mod+= ri71 ~mod+ rl73 + // ...is equivalent to this more-SIMD-like code. + // + // rl73 = (ri73 ~mod* 0xFFFFE6DC) ~mod+ (ri51 ~mod* 0x25A1) + // rl51 = (ri73 ~mod* 0x25A1) ~mod+ (ri51 ~mod* 0x1925) + // rk1 = rl51 ~mod+ ((in1 ~mod* 0x133E) ~mod+ (in7 ~mod* 0xFFFFE333)) + // rk3 = rl73 ~mod+ ((in3 ~mod* 0x1051) ~mod+ (in5 ~mod* 0xFFFFADFD)) + // rk5 = rl51 ~mod+ ((in3 ~mod* 0xFFFFADFD) ~mod+ (in5 ~mod* 0xFFFFEFB0)) + // rk7 = rl73 ~mod+ ((in1 ~mod* 0xFFFFE333) ~mod+ (in7 ~mod* 0xFFFFECC1)) + // Combine columns. if 8 > args.dst_buffer.length() {