1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/media/libyuv/source/compare_win.cc Wed Dec 31 06:09:35 2014 +0100
1.3 @@ -0,0 +1,232 @@
1.4 +/*
1.5 + * Copyright 2012 The LibYuv Project Authors. All rights reserved.
1.6 + *
1.7 + * Use of this source code is governed by a BSD-style license
1.8 + * that can be found in the LICENSE file in the root of the source
1.9 + * tree. An additional intellectual property rights grant can be found
1.10 + * in the file PATENTS. All contributing project authors may
1.11 + * be found in the AUTHORS file in the root of the source tree.
1.12 + */
1.13 +
1.14 +#include "libyuv/basic_types.h"
1.15 +#include "libyuv/row.h"
1.16 +
1.17 +#ifdef __cplusplus
1.18 +namespace libyuv {
1.19 +extern "C" {
1.20 +#endif
1.21 +
1.22 +#if !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86) && defined(_MSC_VER)
1.23 +
1.24 +__declspec(naked) __declspec(align(16))
1.25 +uint32 SumSquareError_SSE2(const uint8* src_a, const uint8* src_b, int count) {
1.26 + __asm {
1.27 + mov eax, [esp + 4] // src_a
1.28 + mov edx, [esp + 8] // src_b
1.29 + mov ecx, [esp + 12] // count
1.30 + pxor xmm0, xmm0
1.31 + pxor xmm5, xmm5
1.32 +
1.33 + align 4
1.34 + wloop:
1.35 + movdqa xmm1, [eax]
1.36 + lea eax, [eax + 16]
1.37 + movdqa xmm2, [edx]
1.38 + lea edx, [edx + 16]
1.39 + sub ecx, 16
1.40 + movdqa xmm3, xmm1 // abs trick
1.41 + psubusb xmm1, xmm2
1.42 + psubusb xmm2, xmm3
1.43 + por xmm1, xmm2
1.44 + movdqa xmm2, xmm1
1.45 + punpcklbw xmm1, xmm5
1.46 + punpckhbw xmm2, xmm5
1.47 + pmaddwd xmm1, xmm1
1.48 + pmaddwd xmm2, xmm2
1.49 + paddd xmm0, xmm1
1.50 + paddd xmm0, xmm2
1.51 + jg wloop
1.52 +
1.53 + pshufd xmm1, xmm0, 0xee
1.54 + paddd xmm0, xmm1
1.55 + pshufd xmm1, xmm0, 0x01
1.56 + paddd xmm0, xmm1
1.57 + movd eax, xmm0
1.58 + ret
1.59 + }
1.60 +}
1.61 +
1.62 +// Visual C 2012 required for AVX2.
1.63 +#if _MSC_VER >= 1700
1.64 +// C4752: found Intel(R) Advanced Vector Extensions; consider using /arch:AVX.
1.65 +#pragma warning(disable: 4752)
1.66 +__declspec(naked) __declspec(align(16))
1.67 +uint32 SumSquareError_AVX2(const uint8* src_a, const uint8* src_b, int count) {
1.68 + __asm {
1.69 + mov eax, [esp + 4] // src_a
1.70 + mov edx, [esp + 8] // src_b
1.71 + mov ecx, [esp + 12] // count
1.72 + vpxor ymm0, ymm0, ymm0 // sum
1.73 + vpxor ymm5, ymm5, ymm5 // constant 0 for unpck
1.74 + sub edx, eax
1.75 +
1.76 + align 4
1.77 + wloop:
1.78 + vmovdqu ymm1, [eax]
1.79 + vmovdqu ymm2, [eax + edx]
1.80 + lea eax, [eax + 32]
1.81 + sub ecx, 32
1.82 + vpsubusb ymm3, ymm1, ymm2 // abs difference trick
1.83 + vpsubusb ymm2, ymm2, ymm1
1.84 + vpor ymm1, ymm2, ymm3
1.85 + vpunpcklbw ymm2, ymm1, ymm5 // u16. mutates order.
1.86 + vpunpckhbw ymm1, ymm1, ymm5
1.87 + vpmaddwd ymm2, ymm2, ymm2 // square + hadd to u32.
1.88 + vpmaddwd ymm1, ymm1, ymm1
1.89 + vpaddd ymm0, ymm0, ymm1
1.90 + vpaddd ymm0, ymm0, ymm2
1.91 + jg wloop
1.92 +
1.93 + vpshufd ymm1, ymm0, 0xee // 3, 2 + 1, 0 both lanes.
1.94 + vpaddd ymm0, ymm0, ymm1
1.95 + vpshufd ymm1, ymm0, 0x01 // 1 + 0 both lanes.
1.96 + vpaddd ymm0, ymm0, ymm1
1.97 + vpermq ymm1, ymm0, 0x02 // high + low lane.
1.98 + vpaddd ymm0, ymm0, ymm1
1.99 + vmovd eax, xmm0
1.100 + vzeroupper
1.101 + ret
1.102 + }
1.103 +}
1.104 +#endif // _MSC_VER >= 1700
1.105 +
1.106 +#define HAS_HASHDJB2_SSE41
1.107 +static uvec32 kHash16x33 = { 0x92d9e201, 0, 0, 0 }; // 33 ^ 16
1.108 +static uvec32 kHashMul0 = {
1.109 + 0x0c3525e1, // 33 ^ 15
1.110 + 0xa3476dc1, // 33 ^ 14
1.111 + 0x3b4039a1, // 33 ^ 13
1.112 + 0x4f5f0981, // 33 ^ 12
1.113 +};
1.114 +static uvec32 kHashMul1 = {
1.115 + 0x30f35d61, // 33 ^ 11
1.116 + 0x855cb541, // 33 ^ 10
1.117 + 0x040a9121, // 33 ^ 9
1.118 + 0x747c7101, // 33 ^ 8
1.119 +};
1.120 +static uvec32 kHashMul2 = {
1.121 + 0xec41d4e1, // 33 ^ 7
1.122 + 0x4cfa3cc1, // 33 ^ 6
1.123 + 0x025528a1, // 33 ^ 5
1.124 + 0x00121881, // 33 ^ 4
1.125 +};
1.126 +static uvec32 kHashMul3 = {
1.127 + 0x00008c61, // 33 ^ 3
1.128 + 0x00000441, // 33 ^ 2
1.129 + 0x00000021, // 33 ^ 1
1.130 + 0x00000001, // 33 ^ 0
1.131 +};
1.132 +
1.133 +// 27: 66 0F 38 40 C6 pmulld xmm0,xmm6
1.134 +// 44: 66 0F 38 40 DD pmulld xmm3,xmm5
1.135 +// 59: 66 0F 38 40 E5 pmulld xmm4,xmm5
1.136 +// 72: 66 0F 38 40 D5 pmulld xmm2,xmm5
1.137 +// 83: 66 0F 38 40 CD pmulld xmm1,xmm5
1.138 +#define pmulld(reg) _asm _emit 0x66 _asm _emit 0x0F _asm _emit 0x38 \
1.139 + _asm _emit 0x40 _asm _emit reg
1.140 +
1.141 +__declspec(naked) __declspec(align(16))
1.142 +uint32 HashDjb2_SSE41(const uint8* src, int count, uint32 seed) {
1.143 + __asm {
1.144 + mov eax, [esp + 4] // src
1.145 + mov ecx, [esp + 8] // count
1.146 + movd xmm0, [esp + 12] // seed
1.147 +
1.148 + pxor xmm7, xmm7 // constant 0 for unpck
1.149 + movdqa xmm6, kHash16x33
1.150 +
1.151 + align 4
1.152 + wloop:
1.153 + movdqu xmm1, [eax] // src[0-15]
1.154 + lea eax, [eax + 16]
1.155 + pmulld(0xc6) // pmulld xmm0,xmm6 hash *= 33 ^ 16
1.156 + movdqa xmm5, kHashMul0
1.157 + movdqa xmm2, xmm1
1.158 + punpcklbw xmm2, xmm7 // src[0-7]
1.159 + movdqa xmm3, xmm2
1.160 + punpcklwd xmm3, xmm7 // src[0-3]
1.161 + pmulld(0xdd) // pmulld xmm3, xmm5
1.162 + movdqa xmm5, kHashMul1
1.163 + movdqa xmm4, xmm2
1.164 + punpckhwd xmm4, xmm7 // src[4-7]
1.165 + pmulld(0xe5) // pmulld xmm4, xmm5
1.166 + movdqa xmm5, kHashMul2
1.167 + punpckhbw xmm1, xmm7 // src[8-15]
1.168 + movdqa xmm2, xmm1
1.169 + punpcklwd xmm2, xmm7 // src[8-11]
1.170 + pmulld(0xd5) // pmulld xmm2, xmm5
1.171 + movdqa xmm5, kHashMul3
1.172 + punpckhwd xmm1, xmm7 // src[12-15]
1.173 + pmulld(0xcd) // pmulld xmm1, xmm5
1.174 + paddd xmm3, xmm4 // add 16 results
1.175 + paddd xmm1, xmm2
1.176 + sub ecx, 16
1.177 + paddd xmm1, xmm3
1.178 +
1.179 + pshufd xmm2, xmm1, 0x0e // upper 2 dwords
1.180 + paddd xmm1, xmm2
1.181 + pshufd xmm2, xmm1, 0x01
1.182 + paddd xmm1, xmm2
1.183 + paddd xmm0, xmm1
1.184 + jg wloop
1.185 +
1.186 + movd eax, xmm0 // return hash
1.187 + ret
1.188 + }
1.189 +}
1.190 +
1.191 +// Visual C 2012 required for AVX2.
1.192 +#if _MSC_VER >= 1700
1.193 +__declspec(naked) __declspec(align(16))
1.194 +uint32 HashDjb2_AVX2(const uint8* src, int count, uint32 seed) {
1.195 + __asm {
1.196 + mov eax, [esp + 4] // src
1.197 + mov ecx, [esp + 8] // count
1.198 + movd xmm0, [esp + 12] // seed
1.199 + movdqa xmm6, kHash16x33
1.200 +
1.201 + align 4
1.202 + wloop:
1.203 + vpmovzxbd xmm3, dword ptr [eax] // src[0-3]
1.204 + pmulld xmm0, xmm6 // hash *= 33 ^ 16
1.205 + vpmovzxbd xmm4, dword ptr [eax + 4] // src[4-7]
1.206 + pmulld xmm3, kHashMul0
1.207 + vpmovzxbd xmm2, dword ptr [eax + 8] // src[8-11]
1.208 + pmulld xmm4, kHashMul1
1.209 + vpmovzxbd xmm1, dword ptr [eax + 12] // src[12-15]
1.210 + pmulld xmm2, kHashMul2
1.211 + lea eax, [eax + 16]
1.212 + pmulld xmm1, kHashMul3
1.213 + paddd xmm3, xmm4 // add 16 results
1.214 + paddd xmm1, xmm2
1.215 + sub ecx, 16
1.216 + paddd xmm1, xmm3
1.217 + pshufd xmm2, xmm1, 0x0e // upper 2 dwords
1.218 + paddd xmm1, xmm2
1.219 + pshufd xmm2, xmm1, 0x01
1.220 + paddd xmm1, xmm2
1.221 + paddd xmm0, xmm1
1.222 + jg wloop
1.223 +
1.224 + movd eax, xmm0 // return hash
1.225 + ret
1.226 + }
1.227 +}
1.228 +#endif // _MSC_VER >= 1700
1.229 +
1.230 +#endif // !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86) && defined(_MSC_VER)
1.231 +
1.232 +#ifdef __cplusplus
1.233 +} // extern "C"
1.234 +} // namespace libyuv
1.235 +#endif
