1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/media/libyuv/source/compare_posix.cc Wed Dec 31 06:09:35 2014 +0100
1.3 @@ -0,0 +1,158 @@
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(__x86_64__) || defined(__i386__))
1.23 +
1.24 +uint32 SumSquareError_SSE2(const uint8* src_a, const uint8* src_b, int count) {
1.25 + uint32 sse;
1.26 + asm volatile ( // NOLINT
1.27 + "pxor %%xmm0,%%xmm0 \n"
1.28 + "pxor %%xmm5,%%xmm5 \n"
1.29 + LABELALIGN
1.30 + "1: \n"
1.31 + "movdqa " MEMACCESS(0) ",%%xmm1 \n"
1.32 + "lea " MEMLEA(0x10, 0) ",%0 \n"
1.33 + "movdqa " MEMACCESS(1) ",%%xmm2 \n"
1.34 + "lea " MEMLEA(0x10, 1) ",%1 \n"
1.35 + "sub $0x10,%2 \n"
1.36 + "movdqa %%xmm1,%%xmm3 \n"
1.37 + "psubusb %%xmm2,%%xmm1 \n"
1.38 + "psubusb %%xmm3,%%xmm2 \n"
1.39 + "por %%xmm2,%%xmm1 \n"
1.40 + "movdqa %%xmm1,%%xmm2 \n"
1.41 + "punpcklbw %%xmm5,%%xmm1 \n"
1.42 + "punpckhbw %%xmm5,%%xmm2 \n"
1.43 + "pmaddwd %%xmm1,%%xmm1 \n"
1.44 + "pmaddwd %%xmm2,%%xmm2 \n"
1.45 + "paddd %%xmm1,%%xmm0 \n"
1.46 + "paddd %%xmm2,%%xmm0 \n"
1.47 + "jg 1b \n"
1.48 +
1.49 + "pshufd $0xee,%%xmm0,%%xmm1 \n"
1.50 + "paddd %%xmm1,%%xmm0 \n"
1.51 + "pshufd $0x1,%%xmm0,%%xmm1 \n"
1.52 + "paddd %%xmm1,%%xmm0 \n"
1.53 + "movd %%xmm0,%3 \n"
1.54 +
1.55 + : "+r"(src_a), // %0
1.56 + "+r"(src_b), // %1
1.57 + "+r"(count), // %2
1.58 + "=g"(sse) // %3
1.59 + :
1.60 + : "memory", "cc"
1.61 +#if defined(__SSE2__)
1.62 + , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
1.63 +#endif
1.64 + ); // NOLINT
1.65 + return sse;
1.66 +}
1.67 +
1.68 +#endif // defined(__x86_64__) || defined(__i386__)
1.69 +
1.70 +#if !defined(LIBYUV_DISABLE_X86) && \
1.71 + (defined(__x86_64__) || (defined(__i386__) && !defined(__pic__)))
1.72 +#define HAS_HASHDJB2_SSE41
1.73 +static uvec32 kHash16x33 = { 0x92d9e201, 0, 0, 0 }; // 33 ^ 16
1.74 +static uvec32 kHashMul0 = {
1.75 + 0x0c3525e1, // 33 ^ 15
1.76 + 0xa3476dc1, // 33 ^ 14
1.77 + 0x3b4039a1, // 33 ^ 13
1.78 + 0x4f5f0981, // 33 ^ 12
1.79 +};
1.80 +static uvec32 kHashMul1 = {
1.81 + 0x30f35d61, // 33 ^ 11
1.82 + 0x855cb541, // 33 ^ 10
1.83 + 0x040a9121, // 33 ^ 9
1.84 + 0x747c7101, // 33 ^ 8
1.85 +};
1.86 +static uvec32 kHashMul2 = {
1.87 + 0xec41d4e1, // 33 ^ 7
1.88 + 0x4cfa3cc1, // 33 ^ 6
1.89 + 0x025528a1, // 33 ^ 5
1.90 + 0x00121881, // 33 ^ 4
1.91 +};
1.92 +static uvec32 kHashMul3 = {
1.93 + 0x00008c61, // 33 ^ 3
1.94 + 0x00000441, // 33 ^ 2
1.95 + 0x00000021, // 33 ^ 1
1.96 + 0x00000001, // 33 ^ 0
1.97 +};
1.98 +
1.99 +uint32 HashDjb2_SSE41(const uint8* src, int count, uint32 seed) {
1.100 + uint32 hash;
1.101 + asm volatile ( // NOLINT
1.102 + "movd %2,%%xmm0 \n"
1.103 + "pxor %%xmm7,%%xmm7 \n"
1.104 + "movdqa %4,%%xmm6 \n"
1.105 + LABELALIGN
1.106 + "1: \n"
1.107 + "movdqu " MEMACCESS(0) ",%%xmm1 \n"
1.108 + "lea " MEMLEA(0x10, 0) ",%0 \n"
1.109 + "pmulld %%xmm6,%%xmm0 \n"
1.110 + "movdqa %5,%%xmm5 \n"
1.111 + "movdqa %%xmm1,%%xmm2 \n"
1.112 + "punpcklbw %%xmm7,%%xmm2 \n"
1.113 + "movdqa %%xmm2,%%xmm3 \n"
1.114 + "punpcklwd %%xmm7,%%xmm3 \n"
1.115 + "pmulld %%xmm5,%%xmm3 \n"
1.116 + "movdqa %6,%%xmm5 \n"
1.117 + "movdqa %%xmm2,%%xmm4 \n"
1.118 + "punpckhwd %%xmm7,%%xmm4 \n"
1.119 + "pmulld %%xmm5,%%xmm4 \n"
1.120 + "movdqa %7,%%xmm5 \n"
1.121 + "punpckhbw %%xmm7,%%xmm1 \n"
1.122 + "movdqa %%xmm1,%%xmm2 \n"
1.123 + "punpcklwd %%xmm7,%%xmm2 \n"
1.124 + "pmulld %%xmm5,%%xmm2 \n"
1.125 + "movdqa %8,%%xmm5 \n"
1.126 + "punpckhwd %%xmm7,%%xmm1 \n"
1.127 + "pmulld %%xmm5,%%xmm1 \n"
1.128 + "paddd %%xmm4,%%xmm3 \n"
1.129 + "paddd %%xmm2,%%xmm1 \n"
1.130 + "sub $0x10,%1 \n"
1.131 + "paddd %%xmm3,%%xmm1 \n"
1.132 + "pshufd $0xe,%%xmm1,%%xmm2 \n"
1.133 + "paddd %%xmm2,%%xmm1 \n"
1.134 + "pshufd $0x1,%%xmm1,%%xmm2 \n"
1.135 + "paddd %%xmm2,%%xmm1 \n"
1.136 + "paddd %%xmm1,%%xmm0 \n"
1.137 + "jg 1b \n"
1.138 + "movd %%xmm0,%3 \n"
1.139 + : "+r"(src), // %0
1.140 + "+r"(count), // %1
1.141 + "+rm"(seed), // %2
1.142 + "=g"(hash) // %3
1.143 + : "m"(kHash16x33), // %4
1.144 + "m"(kHashMul0), // %5
1.145 + "m"(kHashMul1), // %6
1.146 + "m"(kHashMul2), // %7
1.147 + "m"(kHashMul3) // %8
1.148 + : "memory", "cc"
1.149 +#if defined(__SSE2__)
1.150 + , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
1.151 +#endif
1.152 + ); // NOLINT
1.153 + return hash;
1.154 +}
1.155 +#endif // defined(__x86_64__) || (defined(__i386__) && !defined(__pic__)))
1.156 +
1.157 +#ifdef __cplusplus
1.158 +} // extern "C"
1.159 +} // namespace libyuv
1.160 +#endif
1.161 +
