/external/XNNPACK/src/f32-dwconv/gen/ |
D | up16x4-avx.c | 65 const __m256 vk1x01234567 = _mm256_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up16x4__avx() local 112 const __m256 vk1x01234567 = _mm256_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up16x4__avx() local 148 const __m256 vk1x01234567 = _mm256_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up16x4__avx() local
|
D | up16x4-fma3.c | 65 const __m256 vk1x01234567 = _mm256_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up16x4__fma3() local 112 const __m256 vk1x01234567 = _mm256_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up16x4__fma3() local 148 const __m256 vk1x01234567 = _mm256_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up16x4__fma3() local
|
D | up16x4-avx-acc2.c | 65 const __m256 vk1x01234567 = _mm256_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up16x4__avx_acc2() local 115 const __m256 vk1x01234567 = _mm256_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up16x4__avx_acc2() local 153 const __m256 vk1x01234567 = _mm256_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up16x4__avx_acc2() local
|
D | up16x4-fma3-acc2.c | 65 const __m256 vk1x01234567 = _mm256_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up16x4__fma3_acc2() local 115 const __m256 vk1x01234567 = _mm256_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up16x4__fma3_acc2() local 153 const __m256 vk1x01234567 = _mm256_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up16x4__fma3_acc2() local
|
D | up8x4-fma3.c | 60 const __m256 vk1x01234567 = _mm256_load_ps(w + 16); in xnn_f32_dwconv_ukernel_up8x4__fma3() local 96 const __m256 vk1x01234567 = _mm256_load_ps(w + 16); in xnn_f32_dwconv_ukernel_up8x4__fma3() local
|
D | up8x4-avx.c | 60 const __m256 vk1x01234567 = _mm256_load_ps(w + 16); in xnn_f32_dwconv_ukernel_up8x4__avx() local 96 const __m256 vk1x01234567 = _mm256_load_ps(w + 16); in xnn_f32_dwconv_ukernel_up8x4__avx() local
|
D | up8x4-avx-acc2.c | 60 const __m256 vk1x01234567 = _mm256_load_ps(w + 16); in xnn_f32_dwconv_ukernel_up8x4__avx_acc2() local 98 const __m256 vk1x01234567 = _mm256_load_ps(w + 16); in xnn_f32_dwconv_ukernel_up8x4__avx_acc2() local
|
D | up8x4-fma3-acc2.c | 60 const __m256 vk1x01234567 = _mm256_load_ps(w + 16); in xnn_f32_dwconv_ukernel_up8x4__fma3_acc2() local 98 const __m256 vk1x01234567 = _mm256_load_ps(w + 16); in xnn_f32_dwconv_ukernel_up8x4__fma3_acc2() local
|
D | up16x9-fma3-acc2.c | 75 const __m256 vk1x01234567 = _mm256_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up16x9__fma3_acc2() local 170 const __m256 vk1x01234567 = _mm256_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up16x9__fma3_acc2() local 238 const __m256 vk1x01234567 = _mm256_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up16x9__fma3_acc2() local
|
D | up16x9-avx-acc2.c | 75 const __m256 vk1x01234567 = _mm256_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up16x9__avx_acc2() local 170 const __m256 vk1x01234567 = _mm256_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up16x9__avx_acc2() local 238 const __m256 vk1x01234567 = _mm256_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up16x9__avx_acc2() local
|
D | up16x9-fma3.c | 75 const __m256 vk1x01234567 = _mm256_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up16x9__fma3() local 167 const __m256 vk1x01234567 = _mm256_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up16x9__fma3() local 233 const __m256 vk1x01234567 = _mm256_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up16x9__fma3() local
|
D | up16x9-avx.c | 75 const __m256 vk1x01234567 = _mm256_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up16x9__avx() local 167 const __m256 vk1x01234567 = _mm256_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up16x9__avx() local 233 const __m256 vk1x01234567 = _mm256_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up16x9__avx() local
|
D | up8x9-avx-acc2.c | 70 const __m256 vk1x01234567 = _mm256_load_ps(w + 16); in xnn_f32_dwconv_ukernel_up8x9__avx_acc2() local 138 const __m256 vk1x01234567 = _mm256_load_ps(w + 16); in xnn_f32_dwconv_ukernel_up8x9__avx_acc2() local
|
D | up8x9-fma3-acc2.c | 70 const __m256 vk1x01234567 = _mm256_load_ps(w + 16); in xnn_f32_dwconv_ukernel_up8x9__fma3_acc2() local 138 const __m256 vk1x01234567 = _mm256_load_ps(w + 16); in xnn_f32_dwconv_ukernel_up8x9__fma3_acc2() local
|
D | up8x9-fma3.c | 70 const __m256 vk1x01234567 = _mm256_load_ps(w + 16); in xnn_f32_dwconv_ukernel_up8x9__fma3() local 136 const __m256 vk1x01234567 = _mm256_load_ps(w + 16); in xnn_f32_dwconv_ukernel_up8x9__fma3() local
|
D | up8x9-avx.c | 70 const __m256 vk1x01234567 = _mm256_load_ps(w + 16); in xnn_f32_dwconv_ukernel_up8x9__avx() local 136 const __m256 vk1x01234567 = _mm256_load_ps(w + 16); in xnn_f32_dwconv_ukernel_up8x9__avx() local
|
D | up8x25-fma3-acc2.c | 102 const __m256 vk1x01234567 = _mm256_load_ps(w + 16); in xnn_f32_dwconv_ukernel_up8x25__fma3_acc2() local 266 const __m256 vk1x01234567 = _mm256_load_ps(w + 16); in xnn_f32_dwconv_ukernel_up8x25__fma3_acc2() local
|
D | up8x25-avx.c | 102 const __m256 vk1x01234567 = _mm256_load_ps(w + 16); in xnn_f32_dwconv_ukernel_up8x25__avx() local 264 const __m256 vk1x01234567 = _mm256_load_ps(w + 16); in xnn_f32_dwconv_ukernel_up8x25__avx() local
|
D | up8x25-avx-acc2.c | 102 const __m256 vk1x01234567 = _mm256_load_ps(w + 16); in xnn_f32_dwconv_ukernel_up8x25__avx_acc2() local 266 const __m256 vk1x01234567 = _mm256_load_ps(w + 16); in xnn_f32_dwconv_ukernel_up8x25__avx_acc2() local
|
D | up8x25-fma3.c | 102 const __m256 vk1x01234567 = _mm256_load_ps(w + 16); in xnn_f32_dwconv_ukernel_up8x25__fma3() local 264 const __m256 vk1x01234567 = _mm256_load_ps(w + 16); in xnn_f32_dwconv_ukernel_up8x25__fma3() local
|
D | up16x25-avx-acc2.c | 107 const __m256 vk1x01234567 = _mm256_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up16x25__avx_acc2() local 346 const __m256 vk1x01234567 = _mm256_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up16x25__avx_acc2() local 510 const __m256 vk1x01234567 = _mm256_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up16x25__avx_acc2() local
|
D | up16x25-fma3.c | 107 const __m256 vk1x01234567 = _mm256_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up16x25__fma3() local 343 const __m256 vk1x01234567 = _mm256_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up16x25__fma3() local 505 const __m256 vk1x01234567 = _mm256_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up16x25__fma3() local
|
D | up16x25-fma3-acc2.c | 107 const __m256 vk1x01234567 = _mm256_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up16x25__fma3_acc2() local 346 const __m256 vk1x01234567 = _mm256_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up16x25__fma3_acc2() local 510 const __m256 vk1x01234567 = _mm256_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up16x25__fma3_acc2() local
|
D | up16x25-avx.c | 107 const __m256 vk1x01234567 = _mm256_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up16x25__avx() local 343 const __m256 vk1x01234567 = _mm256_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up16x25__avx() local 505 const __m256 vk1x01234567 = _mm256_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up16x25__avx() local
|