Lines Matching refs:BPS
204 dst0 = _mm_loadl_epi64((__m128i*)(dst + 0 * BPS)); in Transform()
205 dst1 = _mm_loadl_epi64((__m128i*)(dst + 1 * BPS)); in Transform()
206 dst2 = _mm_loadl_epi64((__m128i*)(dst + 2 * BPS)); in Transform()
207 dst3 = _mm_loadl_epi64((__m128i*)(dst + 3 * BPS)); in Transform()
210 dst0 = _mm_cvtsi32_si128(WebPMemToUint32(dst + 0 * BPS)); in Transform()
211 dst1 = _mm_cvtsi32_si128(WebPMemToUint32(dst + 1 * BPS)); in Transform()
212 dst2 = _mm_cvtsi32_si128(WebPMemToUint32(dst + 2 * BPS)); in Transform()
213 dst3 = _mm_cvtsi32_si128(WebPMemToUint32(dst + 3 * BPS)); in Transform()
233 _mm_storel_epi64((__m128i*)(dst + 0 * BPS), dst0); in Transform()
234 _mm_storel_epi64((__m128i*)(dst + 1 * BPS), dst1); in Transform()
235 _mm_storel_epi64((__m128i*)(dst + 2 * BPS), dst2); in Transform()
236 _mm_storel_epi64((__m128i*)(dst + 3 * BPS), dst3); in Transform()
239 WebPUint32ToMem(dst + 0 * BPS, _mm_cvtsi128_si32(dst0)); in Transform()
240 WebPUint32ToMem(dst + 1 * BPS, _mm_cvtsi128_si32(dst1)); in Transform()
241 WebPUint32ToMem(dst + 2 * BPS, _mm_cvtsi128_si32(dst2)); in Transform()
242 WebPUint32ToMem(dst + 3 * BPS, _mm_cvtsi128_si32(dst3)); in Transform()
265 __m128i dst0 = _mm_cvtsi32_si128(WebPMemToUint32(dst + 0 * BPS)); in TransformAC3()
266 __m128i dst1 = _mm_cvtsi32_si128(WebPMemToUint32(dst + 1 * BPS)); in TransformAC3()
267 __m128i dst2 = _mm_cvtsi32_si128(WebPMemToUint32(dst + 2 * BPS)); in TransformAC3()
268 __m128i dst3 = _mm_cvtsi32_si128(WebPMemToUint32(dst + 3 * BPS)); in TransformAC3()
285 WebPUint32ToMem(dst + 0 * BPS, _mm_cvtsi128_si32(dst0)); in TransformAC3()
286 WebPUint32ToMem(dst + 1 * BPS, _mm_cvtsi128_si32(dst1)); in TransformAC3()
287 WebPUint32ToMem(dst + 2 * BPS, _mm_cvtsi128_si32(dst2)); in TransformAC3()
288 WebPUint32ToMem(dst + 3 * BPS, _mm_cvtsi128_si32(dst3)); in TransformAC3()
936 #define DST(x, y) dst[(x) + (y) * BPS]
949 const __m128i ABCDEFGH = _mm_loadl_epi64((__m128i*)(dst - BPS - 1)); in VE4()
959 WebPUint32ToMem(dst + i * BPS, vals); in VE4()
965 const __m128i ABCDEFGH = _mm_loadl_epi64((__m128i*)(dst - BPS)); in LD4()
968 const __m128i CDEFGHH0 = _mm_insert_epi16(CDEFGH00, dst[-BPS + 7], 3); in LD4()
973 WebPUint32ToMem(dst + 0 * BPS, _mm_cvtsi128_si32( abcdefg )); in LD4()
974 WebPUint32ToMem(dst + 1 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 1))); in LD4()
975 WebPUint32ToMem(dst + 2 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 2))); in LD4()
976 WebPUint32ToMem(dst + 3 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 3))); in LD4()
981 const int I = dst[-1 + 0 * BPS]; in VR4()
982 const int J = dst[-1 + 1 * BPS]; in VR4()
983 const int K = dst[-1 + 2 * BPS]; in VR4()
984 const int X = dst[-1 - BPS]; in VR4()
985 const __m128i XABCD = _mm_loadl_epi64((__m128i*)(dst - BPS - 1)); in VR4()
994 WebPUint32ToMem(dst + 0 * BPS, _mm_cvtsi128_si32( abcd )); in VR4()
995 WebPUint32ToMem(dst + 1 * BPS, _mm_cvtsi128_si32( efgh )); in VR4()
996 WebPUint32ToMem(dst + 2 * BPS, _mm_cvtsi128_si32(_mm_slli_si128(abcd, 1))); in VR4()
997 WebPUint32ToMem(dst + 3 * BPS, _mm_cvtsi128_si32(_mm_slli_si128(efgh, 1))); in VR4()
1006 const __m128i ABCDEFGH = _mm_loadl_epi64((__m128i*)(dst - BPS)); in VL4()
1019 WebPUint32ToMem(dst + 0 * BPS, _mm_cvtsi128_si32( avg1 )); in VL4()
1020 WebPUint32ToMem(dst + 1 * BPS, _mm_cvtsi128_si32( avg4 )); in VL4()
1021 WebPUint32ToMem(dst + 2 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(avg1, 1))); in VL4()
1022 WebPUint32ToMem(dst + 3 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(avg4, 1))); in VL4()
1031 const __m128i XABCD = _mm_loadl_epi64((__m128i*)(dst - BPS - 1)); in RD4()
1033 const uint32_t I = dst[-1 + 0 * BPS]; in RD4()
1034 const uint32_t J = dst[-1 + 1 * BPS]; in RD4()
1035 const uint32_t K = dst[-1 + 2 * BPS]; in RD4()
1036 const uint32_t L = dst[-1 + 3 * BPS]; in RD4()
1046 WebPUint32ToMem(dst + 3 * BPS, _mm_cvtsi128_si32( abcdefg )); in RD4()
1047 WebPUint32ToMem(dst + 2 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 1))); in RD4()
1048 WebPUint32ToMem(dst + 1 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 2))); in RD4()
1049 WebPUint32ToMem(dst + 0 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 3))); in RD4()
1059 const uint8_t* top = dst - BPS; in TrueMotion()
1065 for (y = 0; y < 4; ++y, dst += BPS) { in TrueMotion()
1074 for (y = 0; y < 8; ++y, dst += BPS) { in TrueMotion()
1084 for (y = 0; y < 16; ++y, dst += BPS) { in TrueMotion()
1100 const __m128i top = _mm_loadu_si128((const __m128i*)(dst - BPS)); in VE16()
1103 _mm_storeu_si128((__m128i*)(dst + j * BPS), top); in VE16()
1112 dst += BPS; in HE16()
1120 _mm_storeu_si128((__m128i*)(dst + j * BPS), values); in Put16()
1126 const __m128i top = _mm_loadu_si128((const __m128i*)(dst - BPS)); in DC16()
1133 left += dst[-1 + j * BPS]; in DC16()
1145 DC += dst[-1 + j * BPS]; in DC16NoTop()
1152 const __m128i top = _mm_loadu_si128((const __m128i*)(dst - BPS)); in DC16NoLeft()
1169 const __m128i top = _mm_loadl_epi64((const __m128i*)(dst - BPS)); in VE8uv()
1171 _mm_storel_epi64((__m128i*)(dst + j * BPS), top); in VE8uv()
1180 dst += BPS; in HE8uv()
1189 _mm_storel_epi64((__m128i*)(dst + j * BPS), values); in Put8x8uv()
1195 const __m128i top = _mm_loadl_epi64((const __m128i*)(dst - BPS)); in DC8uv()
1200 left += dst[-1 + j * BPS]; in DC8uv()
1210 const __m128i top = _mm_loadl_epi64((const __m128i*)(dst - BPS)); in DC8uvNoLeft()
1220 dc0 += dst[-1 + i * BPS]; in DC8uvNoTop()