smooth.cpp | 289 __m128i _s0 = _mm_add_epi32(_mm_loadu_si128((const __m128i*)(SUM+i)), local 294 __m128i _s0T = _mm_cvtps_epi32(_mm_mul_ps(scale4, _mm_cvtepi32_ps(_s0))); 301 _mm_storeu_si128((__m128i*)(SUM+i), _mm_sub_epi32(_s0,_sm)); 340 __m128i _s0 = _mm_add_epi32(_mm_loadu_si128((const __m128i*)(SUM+i)), local 345 __m128i _s0T = _mm_packs_epi32(_s0, _s01); 349 _mm_storeu_si128((__m128i*)(SUM+i), _mm_sub_epi32(_s0,_sm)); 463 __m128i _s0 = _mm_add_epi32(_mm_loadu_si128((const __m128i*)(SUM+i)), local 468 __m128i _s0T = _mm_cvtps_epi32(_mm_mul_ps(scale4, _mm_cvtepi32_ps(_s0))); 473 _mm_storeu_si128((__m128i*)(SUM+i),_mm_sub_epi32(_s0,_sm)); 511 __m128i _s0 = _mm_add_epi32(_mm_loadu_si128((const __m128i*)(SUM+i)) local 632 __m128i _s0 = _mm_add_epi32(_mm_loadu_si128((const __m128i*)(SUM+i)), local 678 __m128i _s0 = _mm_add_epi32(_mm_loadu_si128((const __m128i*)(SUM+i)), local 796 __m128i _s0 = _mm_add_epi32(_mm_loadu_si128((const __m128i*)(SUM+i)), local 833 __m128i _s0 = _mm_add_epi32(_mm_loadu_si128((const __m128i*)(SUM+i)), local 951 __m128i _s0 = _mm_add_epi32(_mm_loadu_si128((const __m128i*)(SUM+i)), local 990 __m128i _s0 = _mm_add_epi32(_mm_loadu_si128((const __m128i*)(SUM+i)), local [all...] |