Home
last modified time | relevance | path

Searched refs:vxb45 (Results 1 – 9 of 9) sorted by relevance

/external/XNNPACK/src/qs8-gemm/gen/
D3x8c8-minmax-avx2.c104 const __m256i vxb45 = _mm256_cvtepi8_epi16(vb45); in xnn_qs8_gemm_minmax_ukernel_3x8c8__avx2() local
106 vacc0x45 = _mm256_add_epi32(vacc0x45, _mm256_madd_epi16(vxa0, vxb45)); in xnn_qs8_gemm_minmax_ukernel_3x8c8__avx2()
107 vacc1x45 = _mm256_add_epi32(vacc1x45, _mm256_madd_epi16(vxa1, vxb45)); in xnn_qs8_gemm_minmax_ukernel_3x8c8__avx2()
108 vacc2x45 = _mm256_add_epi32(vacc2x45, _mm256_madd_epi16(vxa2, vxb45)); in xnn_qs8_gemm_minmax_ukernel_3x8c8__avx2()
D3x8c8-xw-minmax-avx2.c101 … const __m256i vxb45 = _mm256_load_si256((const __m256i*) ((uintptr_t) w + 32 * sizeof(int16_t))); in xnn_qs8_gemm_xw_minmax_ukernel_3x8c8__avx2() local
103 vacc0x45 = _mm256_add_epi32(vacc0x45, _mm256_madd_epi16(vxa0, vxb45)); in xnn_qs8_gemm_xw_minmax_ukernel_3x8c8__avx2()
104 vacc1x45 = _mm256_add_epi32(vacc1x45, _mm256_madd_epi16(vxa1, vxb45)); in xnn_qs8_gemm_xw_minmax_ukernel_3x8c8__avx2()
105 vacc2x45 = _mm256_add_epi32(vacc2x45, _mm256_madd_epi16(vxa2, vxb45)); in xnn_qs8_gemm_xw_minmax_ukernel_3x8c8__avx2()
D2x8c8-xw-minmax-avx2.c86 … const __m256i vxb45 = _mm256_load_si256((const __m256i*) ((uintptr_t) w + 32 * sizeof(int16_t))); in xnn_qs8_gemm_xw_minmax_ukernel_2x8c8__avx2() local
88 vacc0x45 = _mm256_add_epi32(vacc0x45, _mm256_madd_epi16(vxa0, vxb45)); in xnn_qs8_gemm_xw_minmax_ukernel_2x8c8__avx2()
89 vacc1x45 = _mm256_add_epi32(vacc1x45, _mm256_madd_epi16(vxa1, vxb45)); in xnn_qs8_gemm_xw_minmax_ukernel_2x8c8__avx2()
D2x8c8-minmax-avx2.c89 const __m256i vxb45 = _mm256_cvtepi8_epi16(vb45); in xnn_qs8_gemm_minmax_ukernel_2x8c8__avx2() local
91 vacc0x45 = _mm256_add_epi32(vacc0x45, _mm256_madd_epi16(vxa0, vxb45)); in xnn_qs8_gemm_minmax_ukernel_2x8c8__avx2()
92 vacc1x45 = _mm256_add_epi32(vacc1x45, _mm256_madd_epi16(vxa1, vxb45)); in xnn_qs8_gemm_minmax_ukernel_2x8c8__avx2()
D1x8c8-xw-minmax-avx2.c71 … const __m256i vxb45 = _mm256_load_si256((const __m256i*) ((uintptr_t) w + 32 * sizeof(int16_t))); in xnn_qs8_gemm_xw_minmax_ukernel_1x8c8__avx2() local
73 vacc0x45 = _mm256_add_epi32(vacc0x45, _mm256_madd_epi16(vxa0, vxb45)); in xnn_qs8_gemm_xw_minmax_ukernel_1x8c8__avx2()
D1x8c8-minmax-avx2.c74 const __m256i vxb45 = _mm256_cvtepi8_epi16(vb45); in xnn_qs8_gemm_minmax_ukernel_1x8c8__avx2() local
76 vacc0x45 = _mm256_add_epi32(vacc0x45, _mm256_madd_epi16(vxa0, vxb45)); in xnn_qs8_gemm_minmax_ukernel_1x8c8__avx2()
/external/XNNPACK/src/qs8-igemm/gen/
D3x8c8-minmax-avx2.c119 const __m256i vxb45 = _mm256_cvtepi8_epi16(vb45); in xnn_qs8_igemm_minmax_ukernel_3x8c8__avx2() local
121 vacc0x45 = _mm256_add_epi32(vacc0x45, _mm256_madd_epi16(vxa0, vxb45)); in xnn_qs8_igemm_minmax_ukernel_3x8c8__avx2()
122 vacc1x45 = _mm256_add_epi32(vacc1x45, _mm256_madd_epi16(vxa1, vxb45)); in xnn_qs8_igemm_minmax_ukernel_3x8c8__avx2()
123 vacc2x45 = _mm256_add_epi32(vacc2x45, _mm256_madd_epi16(vxa2, vxb45)); in xnn_qs8_igemm_minmax_ukernel_3x8c8__avx2()
D2x8c8-minmax-avx2.c102 const __m256i vxb45 = _mm256_cvtepi8_epi16(vb45); in xnn_qs8_igemm_minmax_ukernel_2x8c8__avx2() local
104 vacc0x45 = _mm256_add_epi32(vacc0x45, _mm256_madd_epi16(vxa0, vxb45)); in xnn_qs8_igemm_minmax_ukernel_2x8c8__avx2()
105 vacc1x45 = _mm256_add_epi32(vacc1x45, _mm256_madd_epi16(vxa1, vxb45)); in xnn_qs8_igemm_minmax_ukernel_2x8c8__avx2()
D1x8c8-minmax-avx2.c85 const __m256i vxb45 = _mm256_cvtepi8_epi16(vb45); in xnn_qs8_igemm_minmax_ukernel_1x8c8__avx2() local
87 vacc0x45 = _mm256_add_epi32(vacc0x45, _mm256_madd_epi16(vxa0, vxb45)); in xnn_qs8_igemm_minmax_ukernel_1x8c8__avx2()