Changeset 5699 for pjproject/trunk/third_party/yuv/source/compare_gcc.cc
- Timestamp:
- Nov 21, 2017 9:25:11 AM (6 years ago)
- File:
-
- 1 edited
Legend:
- Unmodified
- Added
- Removed
-
pjproject/trunk/third_party/yuv/source/compare_gcc.cc
r5633 r5699 23 23 (defined(__x86_64__) || (defined(__i386__) && !defined(_MSC_VER))) 24 24 25 uint32 HammingDistance_X86(const uint8* src_a, const uint8* src_b, int count) { 25 #if defined(__x86_64__) 26 uint32 HammingDistance_SSE42(const uint8* src_a, 27 const uint8* src_b, 28 int count) { 29 uint64 diff = 0u; 30 31 asm volatile( 32 "xor %3,%3 \n" 33 "xor %%r8,%%r8 \n" 34 "xor %%r9,%%r9 \n" 35 "xor %%r10,%%r10 \n" 36 37 // Process 32 bytes per loop. 38 LABELALIGN 39 "1: \n" 40 "mov (%0),%%rcx \n" 41 "mov 0x8(%0),%%rdx \n" 42 "xor (%1),%%rcx \n" 43 "xor 0x8(%1),%%rdx \n" 44 "popcnt %%rcx,%%rcx \n" 45 "popcnt %%rdx,%%rdx \n" 46 "mov 0x10(%0),%%rsi \n" 47 "mov 0x18(%0),%%rdi \n" 48 "xor 0x10(%1),%%rsi \n" 49 "xor 0x18(%1),%%rdi \n" 50 "popcnt %%rsi,%%rsi \n" 51 "popcnt %%rdi,%%rdi \n" 52 "add $0x20,%0 \n" 53 "add $0x20,%1 \n" 54 "add %%rcx,%3 \n" 55 "add %%rdx,%%r8 \n" 56 "add %%rsi,%%r9 \n" 57 "add %%rdi,%%r10 \n" 58 "sub $0x20,%2 \n" 59 "jg 1b \n" 60 61 "add %%r8, %3 \n" 62 "add %%r9, %3 \n" 63 "add %%r10, %3 \n" 64 : "+r"(src_a), // %0 65 "+r"(src_b), // %1 66 "+r"(count), // %2 67 "=r"(diff) // %3 68 : 69 : "memory", "cc", "rcx", "rdx", "rsi", "rdi", "r8", "r9", "r10"); 70 71 return static_cast<uint32>(diff); 72 } 73 #else 74 uint32 HammingDistance_SSE42(const uint8* src_a, 75 const uint8* src_b, 76 int count) { 26 77 uint32 diff = 0u; 27 78 28 int i; 29 for (i = 0; i < count - 7; i += 8) { 30 uint64 x = *((uint64*)src_a) ^ *((uint64*)src_b); 31 src_a += 8; 32 src_b += 8; 33 diff += __builtin_popcountll(x); 34 } 79 asm volatile( 80 // Process 16 bytes per loop. 81 LABELALIGN 82 "1: \n" 83 "mov (%0),%%ecx \n" 84 "mov 0x4(%0),%%edx \n" 85 "xor (%1),%%ecx \n" 86 "xor 0x4(%1),%%edx \n" 87 "popcnt %%ecx,%%ecx \n" 88 "add %%ecx,%3 \n" 89 "popcnt %%edx,%%edx \n" 90 "add %%edx,%3 \n" 91 "mov 0x8(%0),%%ecx \n" 92 "mov 0xc(%0),%%edx \n" 93 "xor 0x8(%1),%%ecx \n" 94 "xor 0xc(%1),%%edx \n" 95 "popcnt %%ecx,%%ecx \n" 96 "add %%ecx,%3 \n" 97 "popcnt %%edx,%%edx \n" 98 "add %%edx,%3 \n" 99 "add $0x10,%0 \n" 100 "add $0x10,%1 \n" 101 "sub $0x10,%2 \n" 102 "jg 1b \n" 103 : "+r"(src_a), // %0 104 "+r"(src_b), // %1 105 "+r"(count), // %2 106 "+r"(diff) // %3 107 : 108 : "memory", "cc", "ecx", "edx"); 109 35 110 return diff; 36 111 } 112 #endif 113 114 static vec8 kNibbleMask = {15, 15, 15, 15, 15, 15, 15, 15, 115 15, 15, 15, 15, 15, 15, 15, 15}; 116 static vec8 kBitCount = {0, 1, 1, 2, 1, 2, 2, 3, 1, 2, 2, 3, 2, 3, 3, 4}; 117 118 uint32 HammingDistance_SSSE3(const uint8* src_a, 119 const uint8* src_b, 120 int count) { 121 uint32 diff = 0u; 122 123 asm volatile( 124 "movdqa %4,%%xmm2 \n" 125 "movdqa %5,%%xmm3 \n" 126 "pxor %%xmm0,%%xmm0 \n" 127 "pxor %%xmm1,%%xmm1 \n" 128 "sub %0,%1 \n" 129 130 LABELALIGN 131 "1: \n" 132 "movdqa (%0),%%xmm4 \n" 133 "movdqa 0x10(%0), %%xmm5 \n" 134 "pxor (%0,%1), %%xmm4 \n" 135 "movdqa %%xmm4,%%xmm6 \n" 136 "pand %%xmm2,%%xmm6 \n" 137 "psrlw $0x4,%%xmm4 \n" 138 "movdqa %%xmm3,%%xmm7 \n" 139 "pshufb %%xmm6,%%xmm7 \n" 140 "pand %%xmm2,%%xmm4 \n" 141 "movdqa %%xmm3,%%xmm6 \n" 142 "pshufb %%xmm4,%%xmm6 \n" 143 "paddb %%xmm7,%%xmm6 \n" 144 "pxor 0x10(%0,%1),%%xmm5 \n" 145 "add $0x20,%0 \n" 146 "movdqa %%xmm5,%%xmm4 \n" 147 "pand %%xmm2,%%xmm5 \n" 148 "psrlw $0x4,%%xmm4 \n" 149 "movdqa %%xmm3,%%xmm7 \n" 150 "pshufb %%xmm5,%%xmm7 \n" 151 "pand %%xmm2,%%xmm4 \n" 152 "movdqa %%xmm3,%%xmm5 \n" 153 "pshufb %%xmm4,%%xmm5 \n" 154 "paddb %%xmm7,%%xmm5 \n" 155 "paddb %%xmm5,%%xmm6 \n" 156 "psadbw %%xmm1,%%xmm6 \n" 157 "paddd %%xmm6,%%xmm0 \n" 158 "sub $0x20,%2 \n" 159 "jg 1b \n" 160 161 "pshufd $0xaa,%%xmm0,%%xmm1 \n" 162 "paddd %%xmm1,%%xmm0 \n" 163 "movd %%xmm0, %3 \n" 164 : "+r"(src_a), // %0 165 "+r"(src_b), // %1 166 "+r"(count), // %2 167 "=r"(diff) // %3 168 : "m"(kNibbleMask), // %4 169 "m"(kBitCount) // %5 170 : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", 171 "xmm7"); 172 173 return diff; 174 } 175 176 #ifdef HAS_HAMMINGDISTANCE_AVX2 177 uint32 HammingDistance_AVX2(const uint8* src_a, const uint8* src_b, int count) { 178 uint32 diff = 0u; 179 180 asm volatile( 181 "vbroadcastf128 %4,%%ymm2 \n" 182 "vbroadcastf128 %5,%%ymm3 \n" 183 "vpxor %%ymm0,%%ymm0,%%ymm0 \n" 184 "vpxor %%ymm1,%%ymm1,%%ymm1 \n" 185 "sub %0,%1 \n" 186 187 LABELALIGN 188 "1: \n" 189 "vmovdqa (%0),%%ymm4 \n" 190 "vmovdqa 0x20(%0), %%ymm5 \n" 191 "vpxor (%0,%1), %%ymm4, %%ymm4 \n" 192 "vpand %%ymm2,%%ymm4,%%ymm6 \n" 193 "vpsrlw $0x4,%%ymm4,%%ymm4 \n" 194 "vpshufb %%ymm6,%%ymm3,%%ymm6 \n" 195 "vpand %%ymm2,%%ymm4,%%ymm4 \n" 196 "vpshufb %%ymm4,%%ymm3,%%ymm4 \n" 197 "vpaddb %%ymm4,%%ymm6,%%ymm6 \n" 198 "vpxor 0x20(%0,%1),%%ymm5,%%ymm4 \n" 199 "add $0x40,%0 \n" 200 "vpand %%ymm2,%%ymm4,%%ymm5 \n" 201 "vpsrlw $0x4,%%ymm4,%%ymm4 \n" 202 "vpshufb %%ymm5,%%ymm3,%%ymm5 \n" 203 "vpand %%ymm2,%%ymm4,%%ymm4 \n" 204 "vpshufb %%ymm4,%%ymm3,%%ymm4 \n" 205 "vpaddb %%ymm5,%%ymm4,%%ymm4 \n" 206 "vpaddb %%ymm6,%%ymm4,%%ymm4 \n" 207 "vpsadbw %%ymm1,%%ymm4,%%ymm4 \n" 208 "vpaddd %%ymm0,%%ymm4,%%ymm0 \n" 209 "sub $0x40,%2 \n" 210 "jg 1b \n" 211 212 "vpermq $0xb1,%%ymm0,%%ymm1 \n" 213 "vpaddd %%ymm1,%%ymm0,%%ymm0 \n" 214 "vpermq $0xaa,%%ymm0,%%ymm1 \n" 215 "vpaddd %%ymm1,%%ymm0,%%ymm0 \n" 216 "vmovd %%xmm0, %3 \n" 217 "vzeroupper \n" 218 : "+r"(src_a), // %0 219 "+r"(src_b), // %1 220 "+r"(count), // %2 221 "=r"(diff) // %3 222 : "m"(kNibbleMask), // %4 223 "m"(kBitCount) // %5 224 : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"); 225 226 return diff; 227 } 228 #endif // HAS_HAMMINGDISTANCE_AVX2 37 229 38 230 uint32 SumSquareError_SSE2(const uint8* src_a, const uint8* src_b, int count) {
Note: See TracChangeset
for help on using the changeset viewer.