mirror of
https://github.com/ddidderr/wuffs.git
synced 2026-11-17 16:32:11 +01:00
name old speed new speed delta wuffs_crc32_ieee_10k/clang14 13.5GB/s ± 0% 21.9GB/s ± 0% +62.09% (p=0.016 n=5+4) wuffs_crc32_ieee_100k/clang14 22.4GB/s ± 9% 29.7GB/s ± 0% +32.76% (p=0.016 n=5+4) wuffs_crc32_ieee_10k/gcc12 14.2GB/s ± 2% 22.2GB/s ± 0% +56.92% (p=0.008 n=5+5) wuffs_crc32_ieee_100k/gcc12 21.3GB/s ± 3% 29.6GB/s ± 1% +39.18% (p=0.008 n=5+5) wuffs_gzip_decode_10k/clang14 366MB/s ± 0% 372MB/s ± 0% +1.41% (p=0.008 n=5+5) wuffs_gzip_decode_100k/clang14 482MB/s ± 0% 494MB/s ± 0% +2.57% (p=0.008 n=5+5) wuffs_gzip_decode_10k/gcc12 398MB/s ± 0% 418MB/s ± 0% +5.19% (p=0.008 n=5+5) wuffs_gzip_decode_100k/gcc12 510MB/s ± 0% 537MB/s ± 0% +5.21% (p=0.008 n=5+5) wuffs_png_decode_image_19k_8bpp/clang14 263MB/s ± 0% 264MB/s ± 0% +0.39% (p=0.008 n=5+5) wuffs_png_decode_image_40k_24bpp/clang14 297MB/s ± 0% 298MB/s ± 0% +0.16% (p=0.008 n=5+5) wuffs_png_decode_image_77k_8bpp/clang14 932MB/s ± 0% 945MB/s ± 0% +1.33% (p=0.008 n=5+5) wuffs_png_decode_image_552k_32bpp_ignore_checksum/clang14 831MB/s ± 0% 833MB/s ± 0% +0.25% (p=0.008 n=5+5) wuffs_png_decode_image_552k_32bpp_verify_checksum/clang14 799MB/s ± 0% 802MB/s ± 0% +0.37% (p=0.008 n=5+5) wuffs_png_decode_image_4002k_24bpp/clang14 306MB/s ± 0% 307MB/s ± 0% +0.56% (p=0.008 n=5+5) wuffs_png_decode_image_19k_8bpp/gcc12 283MB/s ± 0% 270MB/s ± 0% -4.84% (p=0.008 n=5+5) wuffs_png_decode_image_40k_24bpp/gcc12 330MB/s ± 0% 329MB/s ± 0% -0.29% (p=0.016 n=5+4) wuffs_png_decode_image_77k_8bpp/gcc12 992MB/s ± 0% 957MB/s ± 0% -3.56% (p=0.008 n=5+5) wuffs_png_decode_image_552k_32bpp_ignore_checksum/gcc12 920MB/s ± 0% 908MB/s ± 0% -1.30% (p=0.008 n=5+5) wuffs_png_decode_image_552k_32bpp_verify_checksum/gcc12 875MB/s ± 0% 867MB/s ± 0% -0.95% (p=0.008 n=5+5) wuffs_png_decode_image_4002k_24bpp/gcc12 341MB/s ± 0% 342MB/s ± 0% +0.25% (p=0.008 n=5+5)
168 lines
7.4 KiB
Go
168 lines
7.4 KiB
Go
// Copyright 2024 The Wuffs Authors.
|
|
//
|
|
// Licensed under the Apache License, Version 2.0 <LICENSE-APACHE or
|
|
// https://www.apache.org/licenses/LICENSE-2.0> or the MIT license
|
|
// <LICENSE-MIT or https://opensource.org/licenses/MIT>, at your
|
|
// option. This file may not be copied, modified, or distributed
|
|
// except according to those terms.
|
|
//
|
|
// SPDX-License-Identifier: Apache-2.0 OR MIT
|
|
|
|
//go:build ignore
|
|
// +build ignore
|
|
|
|
package main
|
|
|
|
// print-crc32-x86-sse42-code.go prints the std/crc32 x86/SSE4.2 Wuffs code
|
|
// based on some C code generated by https://github.com/corsix/fast-crc32/
|
|
//
|
|
// Usage: go run print-crc32-x86-sse42-code.go
|
|
|
|
import (
|
|
"fmt"
|
|
"regexp"
|
|
"strconv"
|
|
"strings"
|
|
)
|
|
|
|
func main() {
|
|
var (
|
|
reXEqLoadu = regexp.MustCompile(`^__m128i x(\d+) = _mm_loadu_si128`)
|
|
reKEqSetr = regexp.MustCompile(`^k = _mm_setr_epi32\(([^,]+), ([^,]+), ([^,]+), ([^\)]+)\);$`)
|
|
reYEqClmul = regexp.MustCompile(`^y(\d+) = clmul_lo\(x(\d+), k\), x(\d+) = clmul_hi\(x(\d+), k\);$`)
|
|
reYEqXorLoadu = regexp.MustCompile(`^y(\d+) = _mm_xor_si128\(y(\d+), _mm_loadu_si128`)
|
|
reYEqXorYX = regexp.MustCompile(`^y(\d+) = _mm_xor_si128\(y(\d+), x(\d+)\), x(\d+) = _mm_xor_si128\(x(\d+), y(\d+)\);$`)
|
|
)
|
|
|
|
fmt.Println("// BEGIN script/print-crc32-x86-sse42-code.go generated code.")
|
|
for src := srcSSECRC32V8; src != ""; {
|
|
i := strings.IndexByte(src, '\n')
|
|
line := strings.TrimSpace(src[:i])
|
|
src = src[i+1:]
|
|
|
|
if (line == "") || strings.HasPrefix(line, "/*") {
|
|
continue
|
|
|
|
} else if s := reXEqLoadu.FindStringSubmatch(line); len(s) > 0 {
|
|
n, _ := strconv.Atoi(s[1])
|
|
fmt.Printf("x%d = util.make_m128i_slice128(a: args.x[0x%02X .. 0x%02X])\n", n, 16*(n), 16*(n+1))
|
|
|
|
} else if line == "__m128i k;" {
|
|
continue
|
|
|
|
} else if s := reKEqSetr.FindStringSubmatch(line); len(s) > 0 {
|
|
fmt.Printf("kk = util.make_m128i_multiple_u32(a00: %s, a01: %s, a02: %s, a03: %s)\n", s[1], s[2], s[3], s[4])
|
|
|
|
} else if line == "x0 = _mm_xor_si128(_mm_cvtsi32_si128(crc0), x0);" {
|
|
fmt.Printf("x0 = x0._mm_xor_si128(b: util.make_m128i_single_u32(a: s))\n")
|
|
|
|
} else if line == "buf += 128;" {
|
|
fmt.Printf("args.x = args.x[128 ..]\n")
|
|
|
|
} else if line == "len -= 128;" {
|
|
continue
|
|
|
|
} else if line == "while (len >= 128) {" {
|
|
fmt.Printf("while args.x.length() >= 128 {\n")
|
|
|
|
} else if line == "}" {
|
|
fmt.Printf("} endwhile\n")
|
|
|
|
} else if s := reYEqClmul.FindStringSubmatch(line); len(s) > 0 {
|
|
fmt.Printf("y%s = x%s._mm_clmulepi64_si128(b: kk, imm8: 0x00)\n", s[1], s[2])
|
|
fmt.Printf("x%s = x%s._mm_clmulepi64_si128(b: kk, imm8: 0x11)\n", s[3], s[4])
|
|
|
|
} else if s := reYEqXorLoadu.FindStringSubmatch(line); len(s) > 0 {
|
|
n, _ := strconv.Atoi(s[1])
|
|
fmt.Printf("y%d = y%d._mm_xor_si128(b: util.make_m128i_slice128(a: args.x[0x%02X .. 0x%02X]))\n", n, n, 16*(n), 16*(n+1))
|
|
fmt.Printf("x%d = x%d._mm_xor_si128(b: y%d)\n", n, n, n)
|
|
|
|
} else if s := reYEqXorYX.FindStringSubmatch(line); len(s) > 0 {
|
|
fmt.Printf("y%s = y%s._mm_xor_si128(b: x%s)\n", s[1], s[2], s[3])
|
|
fmt.Printf("x%s = x%s._mm_xor_si128(b: y%s)\n", s[4], s[5], s[6])
|
|
|
|
} else if line == "crc0 = crc_u64(0, _mm_extract_epi64(x0, 0));" {
|
|
fmt.Printf("kk = util.make_m128i_multiple_u32(a00: 0xF701_1641, a01: 0xB4E5_B025, a02: 0xDB71_0641, a03: 1)\n")
|
|
fmt.Printf("s = util.make_m128i_single_u64(a: x0._mm_extract_epi64(imm8: 0)).\n")
|
|
fmt.Printf(" _mm_clmulepi64_si128(b: kk, imm8: 0x00).\n")
|
|
fmt.Printf(" _mm_clmulepi64_si128(b: kk, imm8: 0x10).\n")
|
|
fmt.Printf(" _mm_extract_epi32(imm8: 2)\n")
|
|
|
|
// fmt.Printf("s = util.make_m128i_single_u64(a: (s as base.u64) ^ args.x.peek_u64le()).\n")
|
|
} else if line == "crc0 = crc_u64(crc0, _mm_extract_epi64(x0, 1));" {
|
|
fmt.Printf("kk = util.make_m128i_multiple_u32(a00: 0xF701_1641, a01: 0xB4E5_B025, a02: 0xDB71_0641, a03: 1)\n")
|
|
fmt.Printf("s = util.make_m128i_single_u64(a: x0._mm_extract_epi64(imm8: 1) ^ (s as base.u64)).\n")
|
|
fmt.Printf(" _mm_clmulepi64_si128(b: kk, imm8: 0x00).\n")
|
|
fmt.Printf(" _mm_clmulepi64_si128(b: kk, imm8: 0x10).\n")
|
|
fmt.Printf(" _mm_extract_epi32(imm8: 2)\n")
|
|
|
|
} else {
|
|
fmt.Printf("// Could not process %q.\n", line)
|
|
break
|
|
}
|
|
}
|
|
fmt.Println("// END script/print-crc32-x86-sse42-code.go generated code.")
|
|
}
|
|
|
|
// This is the core (inside "if (len >= 128)") of the code produced by
|
|
// generate.c in https://github.com/corsix/fast-crc32/ when parameterized by
|
|
// "./generate -i sse -p crc32 -a v8".
|
|
const srcSSECRC32V8 = `
|
|
/* First vector chunk. */
|
|
__m128i x0 = _mm_loadu_si128((const __m128i*)buf), y0;
|
|
__m128i x1 = _mm_loadu_si128((const __m128i*)(buf + 16)), y1;
|
|
__m128i x2 = _mm_loadu_si128((const __m128i*)(buf + 32)), y2;
|
|
__m128i x3 = _mm_loadu_si128((const __m128i*)(buf + 48)), y3;
|
|
__m128i x4 = _mm_loadu_si128((const __m128i*)(buf + 64)), y4;
|
|
__m128i x5 = _mm_loadu_si128((const __m128i*)(buf + 80)), y5;
|
|
__m128i x6 = _mm_loadu_si128((const __m128i*)(buf + 96)), y6;
|
|
__m128i x7 = _mm_loadu_si128((const __m128i*)(buf + 112)), y7;
|
|
__m128i k;
|
|
k = _mm_setr_epi32(0x33fff533, 0, 0x910eeec1, 0);
|
|
x0 = _mm_xor_si128(_mm_cvtsi32_si128(crc0), x0);
|
|
buf += 128;
|
|
len -= 128;
|
|
/* Main loop. */
|
|
while (len >= 128) {
|
|
y0 = clmul_lo(x0, k), x0 = clmul_hi(x0, k);
|
|
y1 = clmul_lo(x1, k), x1 = clmul_hi(x1, k);
|
|
y2 = clmul_lo(x2, k), x2 = clmul_hi(x2, k);
|
|
y3 = clmul_lo(x3, k), x3 = clmul_hi(x3, k);
|
|
y4 = clmul_lo(x4, k), x4 = clmul_hi(x4, k);
|
|
y5 = clmul_lo(x5, k), x5 = clmul_hi(x5, k);
|
|
y6 = clmul_lo(x6, k), x6 = clmul_hi(x6, k);
|
|
y7 = clmul_lo(x7, k), x7 = clmul_hi(x7, k);
|
|
y0 = _mm_xor_si128(y0, _mm_loadu_si128((const __m128i*)buf)), x0 = _mm_xor_si128(x0, y0);
|
|
y1 = _mm_xor_si128(y1, _mm_loadu_si128((const __m128i*)(buf + 16))), x1 = _mm_xor_si128(x1, y1);
|
|
y2 = _mm_xor_si128(y2, _mm_loadu_si128((const __m128i*)(buf + 32))), x2 = _mm_xor_si128(x2, y2);
|
|
y3 = _mm_xor_si128(y3, _mm_loadu_si128((const __m128i*)(buf + 48))), x3 = _mm_xor_si128(x3, y3);
|
|
y4 = _mm_xor_si128(y4, _mm_loadu_si128((const __m128i*)(buf + 64))), x4 = _mm_xor_si128(x4, y4);
|
|
y5 = _mm_xor_si128(y5, _mm_loadu_si128((const __m128i*)(buf + 80))), x5 = _mm_xor_si128(x5, y5);
|
|
y6 = _mm_xor_si128(y6, _mm_loadu_si128((const __m128i*)(buf + 96))), x6 = _mm_xor_si128(x6, y6);
|
|
y7 = _mm_xor_si128(y7, _mm_loadu_si128((const __m128i*)(buf + 112))), x7 = _mm_xor_si128(x7, y7);
|
|
buf += 128;
|
|
len -= 128;
|
|
}
|
|
/* Reduce x0 ... x7 to just x0. */
|
|
k = _mm_setr_epi32(0xae689191, 0, 0xccaa009e, 0);
|
|
y0 = clmul_lo(x0, k), x0 = clmul_hi(x0, k);
|
|
y2 = clmul_lo(x2, k), x2 = clmul_hi(x2, k);
|
|
y4 = clmul_lo(x4, k), x4 = clmul_hi(x4, k);
|
|
y6 = clmul_lo(x6, k), x6 = clmul_hi(x6, k);
|
|
y0 = _mm_xor_si128(y0, x1), x0 = _mm_xor_si128(x0, y0);
|
|
y2 = _mm_xor_si128(y2, x3), x2 = _mm_xor_si128(x2, y2);
|
|
y4 = _mm_xor_si128(y4, x5), x4 = _mm_xor_si128(x4, y4);
|
|
y6 = _mm_xor_si128(y6, x7), x6 = _mm_xor_si128(x6, y6);
|
|
k = _mm_setr_epi32(0xf1da05aa, 0, 0x81256527, 0);
|
|
y0 = clmul_lo(x0, k), x0 = clmul_hi(x0, k);
|
|
y4 = clmul_lo(x4, k), x4 = clmul_hi(x4, k);
|
|
y0 = _mm_xor_si128(y0, x2), x0 = _mm_xor_si128(x0, y0);
|
|
y4 = _mm_xor_si128(y4, x6), x4 = _mm_xor_si128(x4, y4);
|
|
k = _mm_setr_epi32(0x8f352d95, 0, 0x1d9513d7, 0);
|
|
y0 = clmul_lo(x0, k), x0 = clmul_hi(x0, k);
|
|
y0 = _mm_xor_si128(y0, x4), x0 = _mm_xor_si128(x0, y0);
|
|
/* Reduce 128 bits to 32 bits, and multiply by x^32. */
|
|
crc0 = crc_u64(0, _mm_extract_epi64(x0, 0));
|
|
crc0 = crc_u64(crc0, _mm_extract_epi64(x0, 1));
|
|
`
|