Add base.arm_crc32_u32 type

On a Raspberry Pi 4 (32-bit armv7l) with -march=native and -mfpu=neon
("native" means "armv8-a+crc+simd"):

wuffs_crc32_ieee_10k/clang9                                825MB/s ± 0%  6802MB/s ± 0%  +724.00%  (p=0.008 n=5+5)
wuffs_crc32_ieee_100k/clang9                               837MB/s ± 0%  6199MB/s ± 0%  +640.25%  (p=0.008 n=5+5)

wuffs_crc32_ieee_10k/gcc8                                  960MB/s ± 0%  4495MB/s ± 0%  +368.35%  (p=0.016 n=4+5)
wuffs_crc32_ieee_100k/gcc8                                 961MB/s ± 0%  4447MB/s ± 0%  +363.01%  (p=0.016 n=5+4)

mimic_crc32_ieee_10k                                       638MB/s ± 0%   631MB/s ± 0%    -1.17%  (p=0.008 n=5+5)
mimic_crc32_ieee_100k                                      639MB/s ± 0%   639MB/s ± 0%    +0.02%  (p=0.008 n=5+5)

wuffs_gzip_decode_10k/clang9                              83.0MB/s ± 0%  90.7MB/s ± 0%    +9.26%  (p=0.008 n=5+5)
wuffs_gzip_decode_100k/clang9                              100MB/s ± 0%   112MB/s ± 0%   +11.91%  (p=0.008 n=5+5)

wuffs_gzip_decode_10k/gcc8                                86.1MB/s ± 0%  91.9MB/s ± 0%    +6.72%  (p=0.008 n=5+5)
wuffs_gzip_decode_100k/gcc8                                102MB/s ± 0%   112MB/s ± 0%    +9.34%  (p=0.008 n=5+5)

mimic_gzip_decode_10k                                     98.8MB/s ± 0%  98.2MB/s ± 0%    -0.60%  (p=0.008 n=5+5)
mimic_gzip_decode_100k                                     128MB/s ± 0%   128MB/s ± 0%    -0.08%  (p=0.008 n=5+5)

wuffs_png_decode_image_19k_8bpp/clang9                    67.6MB/s ± 0%  70.8MB/s ± 0%    +4.72%  (p=0.008 n=5+5)
wuffs_png_decode_image_40k_24bpp/clang9                   92.8MB/s ± 0%  98.8MB/s ± 0%    +6.49%  (p=0.008 n=5+5)
wuffs_png_decode_image_77k_8bpp/clang9                     255MB/s ± 0%   269MB/s ± 0%    +5.44%  (p=0.008 n=5+5)
wuffs_png_decode_image_552k_32bpp_ignore_checksum/clang9   123MB/s ± 0%   123MB/s ± 0%    +0.13%  (p=0.008 n=5+5)
wuffs_png_decode_image_552k_32bpp_verify_checksum/clang9   110MB/s ± 0%   112MB/s ± 0%    +1.46%  (p=0.008 n=5+5)
wuffs_png_decode_image_4002k_24bpp/clang9                 88.4MB/s ± 0%  92.8MB/s ± 0%    +4.90%  (p=0.008 n=5+5)

wuffs_png_decode_image_19k_8bpp/gcc8                      67.2MB/s ± 0%  70.8MB/s ± 0%    +5.37%  (p=0.008 n=5+5)
wuffs_png_decode_image_40k_24bpp/gcc8                     94.0MB/s ± 0%  98.2MB/s ± 0%    +4.45%  (p=0.008 n=5+5)
wuffs_png_decode_image_77k_8bpp/gcc8                       236MB/s ± 0%   248MB/s ± 0%    +4.84%  (p=0.008 n=5+5)
wuffs_png_decode_image_552k_32bpp_ignore_checksum/gcc8     114MB/s ± 0%   115MB/s ± 0%    +1.11%  (p=0.008 n=5+5)
wuffs_png_decode_image_552k_32bpp_verify_checksum/gcc8     102MB/s ± 0%   104MB/s ± 0%    +2.17%  (p=0.008 n=5+5)
wuffs_png_decode_image_4002k_24bpp/gcc8                   88.8MB/s ± 0%  94.3MB/s ± 0%    +6.21%  (p=0.008 n=5+5)

mimic_png_decode_image_19k_8bpp                           43.7MB/s ± 1%  43.4MB/s ± 1%      ~     (p=0.095 n=5+5)
mimic_png_decode_image_40k_24bpp                          54.6MB/s ± 0%  53.9MB/s ± 0%    -1.25%  (p=0.008 n=5+5)
mimic_png_decode_image_77k_8bpp                            125MB/s ± 0%   124MB/s ± 0%    -0.73%  (p=0.008 n=5+5)
mimic_png_decode_image_552k_32bpp_ignore_checksum               skipped
mimic_png_decode_image_552k_32bpp_verify_checksum          102MB/s ± 0%   101MB/s ± 0%    -0.95%  (p=0.008 n=5+5)
mimic_png_decode_image_4002k_24bpp                        81.3MB/s ± 0%  81.0MB/s ± 0%    -0.46%  (p=0.016 n=4+5)
This commit is contained in:
Nigel Tao committed 2021-02-13 22:54:47 +11:00
1 parent 81ba41e251
commit 84f3592dbe
10 files changed
+94 -106

No files matched your search

+7 -28
View File
@@ -72,45 +72,24 @@ pri func ieee_hasher.up!(x: slice base.u8),
pri func ieee_hasher.up_arm_crc32!(x: slice base.u8),
choose cpu_arch >= arm_crc32,
{
var s : base.u32
var s : base.arm_crc32_u32
var p : slice base.u8
s = 0xFFFF_FFFF ^ this.state
s = s.create(a: 0xFFFF_FFFF ^ this.state)
// Align to a 16-byte boundary.
while (args.x.length() > 0) and ((15 & args.x.uintptr_low_12_bits()) <> 0) {
s = IEEE_TABLE[0][((s & 0xFF) as base.u8) ^ args.x[0]] ^ (s >> 8)
s = s._crc32b(b: args.x[0])
args.x = args.x[1 ..]
} endwhile
// See "Multi-Byte Lookup Tables" in std/crc32/README.md for more detail on
// the slicing-by-M algorithm. We use an M of 16.
iterate (p = args.x)(length: 16, advance: 16, unroll: 2) {
s ^= ((p[0x00] as base.u32) << 0) |
((p[0x01] as base.u32) << 8) |
((p[0x02] as base.u32) << 16) |
((p[0x03] as base.u32) << 24)
s = IEEE_TABLE[0x00][p[0x0F]] ^
IEEE_TABLE[0x01][p[0x0E]] ^
IEEE_TABLE[0x02][p[0x0D]] ^
IEEE_TABLE[0x03][p[0x0C]] ^
IEEE_TABLE[0x04][p[0x0B]] ^
IEEE_TABLE[0x05][p[0x0A]] ^
IEEE_TABLE[0x06][p[0x09]] ^
IEEE_TABLE[0x07][p[0x08]] ^
IEEE_TABLE[0x08][p[0x07]] ^
IEEE_TABLE[0x09][p[0x06]] ^
IEEE_TABLE[0x0A][p[0x05]] ^
IEEE_TABLE[0x0B][p[0x04]] ^
IEEE_TABLE[0x0C][0xFF & (s >> 24)] ^
IEEE_TABLE[0x0D][0xFF & (s >> 16)] ^
IEEE_TABLE[0x0E][0xFF & (s >> 8)] ^
IEEE_TABLE[0x0F][0xFF & (s >> 0)]
iterate (p = args.x)(length: 8, advance: 8, unroll: 16) {
s = s._crc32d(b: p.peek_u64le())
} else (length: 1, advance: 1, unroll: 1) {
s = IEEE_TABLE[0][((s & 0xFF) as base.u8) ^ p[0]] ^ (s >> 8)
s = s._crc32b(b: p[0])
}
this.state = 0xFFFF_FFFF ^ s
this.state = 0xFFFF_FFFF ^ s.value()
}
// The table below was created by script/print-crc32-magic-numbers.go.