Home
last modified time | relevance | path

Searched refs:vxb67 (Results 1 – 9 of 9) sorted by relevance

/external/XNNPACK/src/qs8-gemm/gen/
D3x8c8-minmax-avx2.c110 const __m256i vxb67 = _mm256_cvtepi8_epi16(vb67); in xnn_qs8_gemm_minmax_ukernel_3x8c8__avx2() local
112 vacc0x67 = _mm256_add_epi32(vacc0x67, _mm256_madd_epi16(vxa0, vxb67)); in xnn_qs8_gemm_minmax_ukernel_3x8c8__avx2()
113 vacc1x67 = _mm256_add_epi32(vacc1x67, _mm256_madd_epi16(vxa1, vxb67)); in xnn_qs8_gemm_minmax_ukernel_3x8c8__avx2()
114 vacc2x67 = _mm256_add_epi32(vacc2x67, _mm256_madd_epi16(vxa2, vxb67)); in xnn_qs8_gemm_minmax_ukernel_3x8c8__avx2()
D3x8c8-xw-minmax-avx2.c106 … const __m256i vxb67 = _mm256_load_si256((const __m256i*) ((uintptr_t) w + 48 * sizeof(int16_t))); in xnn_qs8_gemm_xw_minmax_ukernel_3x8c8__avx2() local
108 vacc0x67 = _mm256_add_epi32(vacc0x67, _mm256_madd_epi16(vxa0, vxb67)); in xnn_qs8_gemm_xw_minmax_ukernel_3x8c8__avx2()
109 vacc1x67 = _mm256_add_epi32(vacc1x67, _mm256_madd_epi16(vxa1, vxb67)); in xnn_qs8_gemm_xw_minmax_ukernel_3x8c8__avx2()
110 vacc2x67 = _mm256_add_epi32(vacc2x67, _mm256_madd_epi16(vxa2, vxb67)); in xnn_qs8_gemm_xw_minmax_ukernel_3x8c8__avx2()
D2x8c8-xw-minmax-avx2.c90 … const __m256i vxb67 = _mm256_load_si256((const __m256i*) ((uintptr_t) w + 48 * sizeof(int16_t))); in xnn_qs8_gemm_xw_minmax_ukernel_2x8c8__avx2() local
92 vacc0x67 = _mm256_add_epi32(vacc0x67, _mm256_madd_epi16(vxa0, vxb67)); in xnn_qs8_gemm_xw_minmax_ukernel_2x8c8__avx2()
93 vacc1x67 = _mm256_add_epi32(vacc1x67, _mm256_madd_epi16(vxa1, vxb67)); in xnn_qs8_gemm_xw_minmax_ukernel_2x8c8__avx2()
D2x8c8-minmax-avx2.c94 const __m256i vxb67 = _mm256_cvtepi8_epi16(vb67); in xnn_qs8_gemm_minmax_ukernel_2x8c8__avx2() local
96 vacc0x67 = _mm256_add_epi32(vacc0x67, _mm256_madd_epi16(vxa0, vxb67)); in xnn_qs8_gemm_minmax_ukernel_2x8c8__avx2()
97 vacc1x67 = _mm256_add_epi32(vacc1x67, _mm256_madd_epi16(vxa1, vxb67)); in xnn_qs8_gemm_minmax_ukernel_2x8c8__avx2()
D1x8c8-xw-minmax-avx2.c74 … const __m256i vxb67 = _mm256_load_si256((const __m256i*) ((uintptr_t) w + 48 * sizeof(int16_t))); in xnn_qs8_gemm_xw_minmax_ukernel_1x8c8__avx2() local
76 vacc0x67 = _mm256_add_epi32(vacc0x67, _mm256_madd_epi16(vxa0, vxb67)); in xnn_qs8_gemm_xw_minmax_ukernel_1x8c8__avx2()
D1x8c8-minmax-avx2.c78 const __m256i vxb67 = _mm256_cvtepi8_epi16(vb67); in xnn_qs8_gemm_minmax_ukernel_1x8c8__avx2() local
80 vacc0x67 = _mm256_add_epi32(vacc0x67, _mm256_madd_epi16(vxa0, vxb67)); in xnn_qs8_gemm_minmax_ukernel_1x8c8__avx2()
/external/XNNPACK/src/qs8-igemm/gen/
D3x8c8-minmax-avx2.c125 const __m256i vxb67 = _mm256_cvtepi8_epi16(vb67); in xnn_qs8_igemm_minmax_ukernel_3x8c8__avx2() local
127 vacc0x67 = _mm256_add_epi32(vacc0x67, _mm256_madd_epi16(vxa0, vxb67)); in xnn_qs8_igemm_minmax_ukernel_3x8c8__avx2()
128 vacc1x67 = _mm256_add_epi32(vacc1x67, _mm256_madd_epi16(vxa1, vxb67)); in xnn_qs8_igemm_minmax_ukernel_3x8c8__avx2()
129 vacc2x67 = _mm256_add_epi32(vacc2x67, _mm256_madd_epi16(vxa2, vxb67)); in xnn_qs8_igemm_minmax_ukernel_3x8c8__avx2()
D2x8c8-minmax-avx2.c107 const __m256i vxb67 = _mm256_cvtepi8_epi16(vb67); in xnn_qs8_igemm_minmax_ukernel_2x8c8__avx2() local
109 vacc0x67 = _mm256_add_epi32(vacc0x67, _mm256_madd_epi16(vxa0, vxb67)); in xnn_qs8_igemm_minmax_ukernel_2x8c8__avx2()
110 vacc1x67 = _mm256_add_epi32(vacc1x67, _mm256_madd_epi16(vxa1, vxb67)); in xnn_qs8_igemm_minmax_ukernel_2x8c8__avx2()
D1x8c8-minmax-avx2.c89 const __m256i vxb67 = _mm256_cvtepi8_epi16(vb67); in xnn_qs8_igemm_minmax_ukernel_1x8c8__avx2() local
91 vacc0x67 = _mm256_add_epi32(vacc0x67, _mm256_madd_epi16(vxa0, vxb67)); in xnn_qs8_igemm_minmax_ukernel_1x8c8__avx2()