| /external/libaom/libaom/av1/common/x86/ |
| D | highbd_inv_txfm_avx2.c | 271 const __m256i *cospi24, const __m256i *cospi40, const __m256i *cospim24, in idct32_stage4_avx2() argument
282 temp1 = half_btf_avx2(cospim40, &bf1[21], cospi24, &bf1[26], rounding, bit); in idct32_stage4_avx2()
283 bf1[26] = half_btf_avx2(cospi24, &bf1[21], cospi40, &bf1[26], rounding, bit); in idct32_stage4_avx2()
287 bf1[25] = half_btf_avx2(cospim40, &bf1[22], cospi24, &bf1[25], rounding, bit); in idct32_stage4_avx2()
552 const __m256i cospi24 = _mm256_set1_epi32(cospi[24]); in idct32_low8_avx2() local
618 &cospi24, &cospi40, &cospim24, &rounding, bit); in idct32_low8_avx2()
677 const __m256i cospi24 = _mm256_set1_epi32(cospi[24]); in idct32_low16_avx2() local
757 bf1[6] = half_btf_0_avx2(&cospi24, &bf1[6], &rounding, bit); in idct32_low16_avx2()
765 &cospi24, &cospi40, &cospim24, &rounding, bit); in idct32_low16_avx2()
839 const __m256i cospi24 = _mm256_set1_epi32(cospi[24]); in idct32_avx2() local
[all …]
|
| D | highbd_inv_txfm_sse4.c | 151 const __m128i *cospi24, const __m128i *cospi40, const __m128i *cospim24, in idct32_stage4_sse4_1() argument
163 temp1 = half_btf_sse4_1(cospim40, &bf1[21], cospi24, &bf1[26], rounding, bit); in idct32_stage4_sse4_1()
165 half_btf_sse4_1(cospi24, &bf1[21], cospi40, &bf1[26], rounding, bit); in idct32_stage4_sse4_1()
171 half_btf_sse4_1(cospim40, &bf1[22], cospi24, &bf1[25], rounding, bit); in idct32_stage4_sse4_1()
779 const __m128i cospi24 = _mm_set1_epi32(cospi[24]); in idct8x8_sse4_1() local
822 x = _mm_mullo_epi32(in[5 * 2 + col], cospi24); in idct8x8_sse4_1()
829 y = _mm_mullo_epi32(in[3 * 2 + col], cospi24); in idct8x8_sse4_1()
1492 const __m128i cospi24 = _mm_set1_epi32(cospi[24]); in idct8x8_new_sse4_1() local
1528 x = _mm_mullo_epi32(in[5], cospi24); in idct8x8_new_sse4_1()
1535 y = _mm_mullo_epi32(in[3], cospi24); in idct8x8_new_sse4_1()
[all …]
|
| /external/libaom/libaom/av1/encoder/x86/ |
| D | highbd_fwd_txfm_sse4.c | 524 const __m128i cospi24 = _mm_set1_epi32(cospi[24]); in fdct4x8_sse4_1() local
611 v[0] = _mm_mullo_epi32(u[5], cospi24); in fdct4x8_sse4_1()
618 v[1] = _mm_mullo_epi32(u[6], cospi24); in fdct4x8_sse4_1()
1190 const __m128i cospi24 = _mm_set1_epi32(cospi[24]); in fdct16x16_sse4_1() local
1367 u[5] = _mm_mullo_epi32(v[5], cospi24); in fdct16x16_sse4_1()
1374 x = _mm_mullo_epi32(v[6], cospi24); in fdct16x16_sse4_1()
1477 const __m128i cospi24 = _mm_set1_epi32(cospi[24]); in fadst16x16_sse4_1() local
1649 v[10] = half_btf_sse4_1(&cospi40, &u[10], &cospi24, &u[11], &rnding, bit); in fadst16x16_sse4_1()
1650 v[11] = half_btf_sse4_1(&cospi24, &u[10], &cospim40, &u[11], &rnding, bit); in fadst16x16_sse4_1()
1654 v[15] = half_btf_sse4_1(&cospi40, &u[14], &cospi24, &u[15], &rnding, bit); in fadst16x16_sse4_1()
|
| D | highbd_fwd_txfm_avx2.c | 265 const __m256i cospi24 = _mm256_set1_epi32(cospi[24]); in av1_fdct8_avx2() local
339 v[0] = _mm256_mullo_epi32(u[5], cospi24); in av1_fdct8_avx2()
346 v[1] = _mm256_mullo_epi32(u[6], cospi24); in av1_fdct8_avx2()
741 const __m256i cospi24 = _mm256_set1_epi32(cospi[24]); in av1_fdct16_avx2() local
918 u[5] = _mm256_mullo_epi32(v[5], cospi24); in av1_fdct16_avx2()
925 x = _mm256_mullo_epi32(v[6], cospi24); in av1_fdct16_avx2()
1027 const __m256i cospi24 = _mm256_set1_epi32(cospi[24]); in av1_fadst16_avx2() local
1201 v[10] = av1_half_btf_avx2(&cospi40, &u[10], &cospi24, &u[11], &rnding, bit); in av1_fadst16_avx2()
1203 av1_half_btf_avx2(&cospi24, &u[10], &cospim40, &u[11], &rnding, bit); in av1_fadst16_avx2()
1208 v[15] = av1_half_btf_avx2(&cospi40, &u[14], &cospi24, &u[15], &rnding, bit); in av1_fadst16_avx2()
|