/external/libvpx/libvpx/vpx_dsp/x86/ |
highbd_idct4x4_add_sse4.c | 51 __m128i io_short[2]; local 53 io_short[0] = _mm_packs_epi32(io[0], io[1]); 54 io_short[1] = _mm_packs_epi32(io[2], io[3]); 55 idct4_sse2(io_short); 56 idct4_sse2(io_short); 57 io_short[0] = _mm_add_epi16(io_short[0], _mm_set1_epi16(8)); 58 io_short[1] = _mm_add_epi16(io_short[1], _mm_set1_epi16(8)); 59 io[0] = _mm_srai_epi16(io_short[0], 4) [all...] |
highbd_idct8x8_add_sse2.c | 108 __m128i io_short[8]; local 110 io_short[0] = _mm_packs_epi32(io[0], io[4]); 111 io_short[1] = _mm_packs_epi32(io[1], io[5]); 112 io_short[2] = _mm_packs_epi32(io[2], io[6]); 113 io_short[3] = _mm_packs_epi32(io[3], io[7]); 122 io_short[4] = _mm_packs_epi32(io[8], io[12]); 123 io_short[5] = _mm_packs_epi32(io[9], io[13]); 124 io_short[6] = _mm_packs_epi32(io[10], io[14]); 125 io_short[7] = _mm_packs_epi32(io[11], io[15]); 127 idct8_sse2(io_short); 178 __m128i io_short[8]; local [all...] |
highbd_idct8x8_add_sse4.c | 110 __m128i io_short[8]; local 112 io_short[0] = _mm_packs_epi32(io[0], io[4]); 113 io_short[1] = _mm_packs_epi32(io[1], io[5]); 114 io_short[2] = _mm_packs_epi32(io[2], io[6]); 115 io_short[3] = _mm_packs_epi32(io[3], io[7]); 124 io_short[4] = _mm_packs_epi32(io[8], io[12]); 125 io_short[5] = _mm_packs_epi32(io[9], io[13]); 126 io_short[6] = _mm_packs_epi32(io[10], io[14]); 127 io_short[7] = _mm_packs_epi32(io[11], io[15]); 129 idct8_sse2(io_short); 180 __m128i io_short[8]; local [all...] |
highbd_idct4x4_add_sse2.c | 94 __m128i io[4], io_short[2]; local 101 io_short[0] = _mm_packs_epi32(io[0], io[1]); 102 io_short[1] = _mm_packs_epi32(io[2], io[3]); 107 max_input = _mm_max_epi16(io_short[0], io_short[1]); 108 min_input = _mm_min_epi16(io_short[0], io_short[1]); 120 idct4_sse2(io_short); 121 idct4_sse2(io_short); 122 io_short[0] = _mm_add_epi16(io_short[0], _mm_set1_epi16(8)) [all...] |