/external/XNNPACK/src/f32-dwconv/gen/ |
D | up16x4-avx.c | 83 const __m256 vk3x01234567 = _mm256_load_ps(w + 64); in xnn_f32_dwconv_ukernel_up16x4__avx() local 124 const __m256 vk3x01234567 = _mm256_load_ps(w + 64); in xnn_f32_dwconv_ukernel_up16x4__avx() local 156 const __m256 vk3x01234567 = _mm256_load_ps(w + 64); in xnn_f32_dwconv_ukernel_up16x4__avx() local
|
D | up16x4-fma3.c | 83 const __m256 vk3x01234567 = _mm256_load_ps(w + 64); in xnn_f32_dwconv_ukernel_up16x4__fma3() local 124 const __m256 vk3x01234567 = _mm256_load_ps(w + 64); in xnn_f32_dwconv_ukernel_up16x4__fma3() local 156 const __m256 vk3x01234567 = _mm256_load_ps(w + 64); in xnn_f32_dwconv_ukernel_up16x4__fma3() local
|
D | up16x4-avx-acc2.c | 83 const __m256 vk3x01234567 = _mm256_load_ps(w + 64); in xnn_f32_dwconv_ukernel_up16x4__avx_acc2() local 127 const __m256 vk3x01234567 = _mm256_load_ps(w + 64); in xnn_f32_dwconv_ukernel_up16x4__avx_acc2() local 161 const __m256 vk3x01234567 = _mm256_load_ps(w + 64); in xnn_f32_dwconv_ukernel_up16x4__avx_acc2() local
|
D | up16x4-fma3-acc2.c | 83 const __m256 vk3x01234567 = _mm256_load_ps(w + 64); in xnn_f32_dwconv_ukernel_up16x4__fma3_acc2() local 127 const __m256 vk3x01234567 = _mm256_load_ps(w + 64); in xnn_f32_dwconv_ukernel_up16x4__fma3_acc2() local 161 const __m256 vk3x01234567 = _mm256_load_ps(w + 64); in xnn_f32_dwconv_ukernel_up16x4__fma3_acc2() local
|
D | up8x4-fma3.c | 72 const __m256 vk3x01234567 = _mm256_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up8x4__fma3() local 104 const __m256 vk3x01234567 = _mm256_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up8x4__fma3() local
|
D | up8x4-avx.c | 72 const __m256 vk3x01234567 = _mm256_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up8x4__avx() local 104 const __m256 vk3x01234567 = _mm256_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up8x4__avx() local
|
D | up8x4-avx-acc2.c | 72 const __m256 vk3x01234567 = _mm256_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up8x4__avx_acc2() local 106 const __m256 vk3x01234567 = _mm256_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up8x4__avx_acc2() local
|
D | up8x4-fma3-acc2.c | 72 const __m256 vk3x01234567 = _mm256_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up8x4__fma3_acc2() local 106 const __m256 vk3x01234567 = _mm256_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up8x4__fma3_acc2() local
|
D | up16x9-fma3-acc2.c | 93 const __m256 vk3x01234567 = _mm256_load_ps(w + 64); in xnn_f32_dwconv_ukernel_up16x9__fma3_acc2() local 182 const __m256 vk3x01234567 = _mm256_load_ps(w + 64); in xnn_f32_dwconv_ukernel_up16x9__fma3_acc2() local 246 const __m256 vk3x01234567 = _mm256_load_ps(w + 64); in xnn_f32_dwconv_ukernel_up16x9__fma3_acc2() local
|
D | up16x9-avx-acc2.c | 93 const __m256 vk3x01234567 = _mm256_load_ps(w + 64); in xnn_f32_dwconv_ukernel_up16x9__avx_acc2() local 182 const __m256 vk3x01234567 = _mm256_load_ps(w + 64); in xnn_f32_dwconv_ukernel_up16x9__avx_acc2() local 246 const __m256 vk3x01234567 = _mm256_load_ps(w + 64); in xnn_f32_dwconv_ukernel_up16x9__avx_acc2() local
|
D | up16x9-fma3.c | 93 const __m256 vk3x01234567 = _mm256_load_ps(w + 64); in xnn_f32_dwconv_ukernel_up16x9__fma3() local 179 const __m256 vk3x01234567 = _mm256_load_ps(w + 64); in xnn_f32_dwconv_ukernel_up16x9__fma3() local 241 const __m256 vk3x01234567 = _mm256_load_ps(w + 64); in xnn_f32_dwconv_ukernel_up16x9__fma3() local
|
D | up16x9-avx.c | 93 const __m256 vk3x01234567 = _mm256_load_ps(w + 64); in xnn_f32_dwconv_ukernel_up16x9__avx() local 179 const __m256 vk3x01234567 = _mm256_load_ps(w + 64); in xnn_f32_dwconv_ukernel_up16x9__avx() local 241 const __m256 vk3x01234567 = _mm256_load_ps(w + 64); in xnn_f32_dwconv_ukernel_up16x9__avx() local
|
D | up8x9-avx-acc2.c | 82 const __m256 vk3x01234567 = _mm256_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up8x9__avx_acc2() local 146 const __m256 vk3x01234567 = _mm256_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up8x9__avx_acc2() local
|
D | up8x9-fma3-acc2.c | 82 const __m256 vk3x01234567 = _mm256_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up8x9__fma3_acc2() local 146 const __m256 vk3x01234567 = _mm256_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up8x9__fma3_acc2() local
|
D | up8x9-fma3.c | 82 const __m256 vk3x01234567 = _mm256_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up8x9__fma3() local 144 const __m256 vk3x01234567 = _mm256_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up8x9__fma3() local
|
D | up8x9-avx.c | 82 const __m256 vk3x01234567 = _mm256_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up8x9__avx() local 144 const __m256 vk3x01234567 = _mm256_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up8x9__avx() local
|
D | up8x25-fma3-acc2.c | 114 const __m256 vk3x01234567 = _mm256_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up8x25__fma3_acc2() local 274 const __m256 vk3x01234567 = _mm256_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up8x25__fma3_acc2() local
|
D | up8x25-avx.c | 114 const __m256 vk3x01234567 = _mm256_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up8x25__avx() local 272 const __m256 vk3x01234567 = _mm256_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up8x25__avx() local
|
D | up8x25-avx-acc2.c | 114 const __m256 vk3x01234567 = _mm256_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up8x25__avx_acc2() local 274 const __m256 vk3x01234567 = _mm256_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up8x25__avx_acc2() local
|
D | up8x25-fma3.c | 114 const __m256 vk3x01234567 = _mm256_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up8x25__fma3() local 272 const __m256 vk3x01234567 = _mm256_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up8x25__fma3() local
|
D | up16x25-avx-acc2.c | 125 const __m256 vk3x01234567 = _mm256_load_ps(w + 64); in xnn_f32_dwconv_ukernel_up16x25__avx_acc2() local 358 const __m256 vk3x01234567 = _mm256_load_ps(w + 64); in xnn_f32_dwconv_ukernel_up16x25__avx_acc2() local 518 const __m256 vk3x01234567 = _mm256_load_ps(w + 64); in xnn_f32_dwconv_ukernel_up16x25__avx_acc2() local
|
D | up16x25-fma3.c | 125 const __m256 vk3x01234567 = _mm256_load_ps(w + 64); in xnn_f32_dwconv_ukernel_up16x25__fma3() local 355 const __m256 vk3x01234567 = _mm256_load_ps(w + 64); in xnn_f32_dwconv_ukernel_up16x25__fma3() local 513 const __m256 vk3x01234567 = _mm256_load_ps(w + 64); in xnn_f32_dwconv_ukernel_up16x25__fma3() local
|
D | up16x25-fma3-acc2.c | 125 const __m256 vk3x01234567 = _mm256_load_ps(w + 64); in xnn_f32_dwconv_ukernel_up16x25__fma3_acc2() local 358 const __m256 vk3x01234567 = _mm256_load_ps(w + 64); in xnn_f32_dwconv_ukernel_up16x25__fma3_acc2() local 518 const __m256 vk3x01234567 = _mm256_load_ps(w + 64); in xnn_f32_dwconv_ukernel_up16x25__fma3_acc2() local
|
D | up16x25-avx.c | 125 const __m256 vk3x01234567 = _mm256_load_ps(w + 64); in xnn_f32_dwconv_ukernel_up16x25__avx() local 355 const __m256 vk3x01234567 = _mm256_load_ps(w + 64); in xnn_f32_dwconv_ukernel_up16x25__avx() local 513 const __m256 vk3x01234567 = _mm256_load_ps(w + 64); in xnn_f32_dwconv_ukernel_up16x25__avx() local
|