/external/XNNPACK/src/f32-dwconv/gen/ |
D | up16x4-avx.c | 61 const __m256 vi1x01234567 = _mm256_loadu_ps(i1); in xnn_f32_dwconv_ukernel_up16x4__avx() local 109 const __m256 vi1x01234567 = _mm256_loadu_ps(i1); in xnn_f32_dwconv_ukernel_up16x4__avx() local 147 const __m256 vi1x01234567 = _mm256_maskload_ps(i1, vmask); in xnn_f32_dwconv_ukernel_up16x4__avx() local
|
D | up16x4-fma3.c | 61 const __m256 vi1x01234567 = _mm256_loadu_ps(i1); in xnn_f32_dwconv_ukernel_up16x4__fma3() local 109 const __m256 vi1x01234567 = _mm256_loadu_ps(i1); in xnn_f32_dwconv_ukernel_up16x4__fma3() local 147 const __m256 vi1x01234567 = _mm256_maskload_ps(i1, vmask); in xnn_f32_dwconv_ukernel_up16x4__fma3() local
|
D | up16x4-avx-acc2.c | 61 const __m256 vi1x01234567 = _mm256_loadu_ps(i1); in xnn_f32_dwconv_ukernel_up16x4__avx_acc2() local 112 const __m256 vi1x01234567 = _mm256_loadu_ps(i1); in xnn_f32_dwconv_ukernel_up16x4__avx_acc2() local 152 const __m256 vi1x01234567 = _mm256_maskload_ps(i1, vmask); in xnn_f32_dwconv_ukernel_up16x4__avx_acc2() local
|
D | up16x4-fma3-acc2.c | 61 const __m256 vi1x01234567 = _mm256_loadu_ps(i1); in xnn_f32_dwconv_ukernel_up16x4__fma3_acc2() local 112 const __m256 vi1x01234567 = _mm256_loadu_ps(i1); in xnn_f32_dwconv_ukernel_up16x4__fma3_acc2() local 152 const __m256 vi1x01234567 = _mm256_maskload_ps(i1, vmask); in xnn_f32_dwconv_ukernel_up16x4__fma3_acc2() local
|
D | up8x4-fma3.c | 57 const __m256 vi1x01234567 = _mm256_loadu_ps(i1); in xnn_f32_dwconv_ukernel_up8x4__fma3() local 95 const __m256 vi1x01234567 = _mm256_maskload_ps(i1, vmask); in xnn_f32_dwconv_ukernel_up8x4__fma3() local
|
D | up8x4-avx.c | 57 const __m256 vi1x01234567 = _mm256_loadu_ps(i1); in xnn_f32_dwconv_ukernel_up8x4__avx() local 95 const __m256 vi1x01234567 = _mm256_maskload_ps(i1, vmask); in xnn_f32_dwconv_ukernel_up8x4__avx() local
|
D | up8x4-avx-acc2.c | 57 const __m256 vi1x01234567 = _mm256_loadu_ps(i1); in xnn_f32_dwconv_ukernel_up8x4__avx_acc2() local 97 const __m256 vi1x01234567 = _mm256_maskload_ps(i1, vmask); in xnn_f32_dwconv_ukernel_up8x4__avx_acc2() local
|
D | up8x4-fma3-acc2.c | 57 const __m256 vi1x01234567 = _mm256_loadu_ps(i1); in xnn_f32_dwconv_ukernel_up8x4__fma3_acc2() local 97 const __m256 vi1x01234567 = _mm256_maskload_ps(i1, vmask); in xnn_f32_dwconv_ukernel_up8x4__fma3_acc2() local
|
D | up16x9-fma3-acc2.c | 71 const __m256 vi1x01234567 = _mm256_loadu_ps(i1); in xnn_f32_dwconv_ukernel_up16x9__fma3_acc2() local 167 const __m256 vi1x01234567 = _mm256_loadu_ps(i1); in xnn_f32_dwconv_ukernel_up16x9__fma3_acc2() local 237 const __m256 vi1x01234567 = _mm256_maskload_ps(i1, vmask); in xnn_f32_dwconv_ukernel_up16x9__fma3_acc2() local
|
D | up16x9-avx-acc2.c | 71 const __m256 vi1x01234567 = _mm256_loadu_ps(i1); in xnn_f32_dwconv_ukernel_up16x9__avx_acc2() local 167 const __m256 vi1x01234567 = _mm256_loadu_ps(i1); in xnn_f32_dwconv_ukernel_up16x9__avx_acc2() local 237 const __m256 vi1x01234567 = _mm256_maskload_ps(i1, vmask); in xnn_f32_dwconv_ukernel_up16x9__avx_acc2() local
|
D | up16x9-fma3.c | 71 const __m256 vi1x01234567 = _mm256_loadu_ps(i1); in xnn_f32_dwconv_ukernel_up16x9__fma3() local 164 const __m256 vi1x01234567 = _mm256_loadu_ps(i1); in xnn_f32_dwconv_ukernel_up16x9__fma3() local 232 const __m256 vi1x01234567 = _mm256_maskload_ps(i1, vmask); in xnn_f32_dwconv_ukernel_up16x9__fma3() local
|
D | up16x9-avx.c | 71 const __m256 vi1x01234567 = _mm256_loadu_ps(i1); in xnn_f32_dwconv_ukernel_up16x9__avx() local 164 const __m256 vi1x01234567 = _mm256_loadu_ps(i1); in xnn_f32_dwconv_ukernel_up16x9__avx() local 232 const __m256 vi1x01234567 = _mm256_maskload_ps(i1, vmask); in xnn_f32_dwconv_ukernel_up16x9__avx() local
|
D | up8x9-avx-acc2.c | 67 const __m256 vi1x01234567 = _mm256_loadu_ps(i1); in xnn_f32_dwconv_ukernel_up8x9__avx_acc2() local 137 const __m256 vi1x01234567 = _mm256_maskload_ps(i1, vmask); in xnn_f32_dwconv_ukernel_up8x9__avx_acc2() local
|
D | up8x9-fma3-acc2.c | 67 const __m256 vi1x01234567 = _mm256_loadu_ps(i1); in xnn_f32_dwconv_ukernel_up8x9__fma3_acc2() local 137 const __m256 vi1x01234567 = _mm256_maskload_ps(i1, vmask); in xnn_f32_dwconv_ukernel_up8x9__fma3_acc2() local
|
D | up8x9-fma3.c | 67 const __m256 vi1x01234567 = _mm256_loadu_ps(i1); in xnn_f32_dwconv_ukernel_up8x9__fma3() local 135 const __m256 vi1x01234567 = _mm256_maskload_ps(i1, vmask); in xnn_f32_dwconv_ukernel_up8x9__fma3() local
|
D | up8x9-avx.c | 67 const __m256 vi1x01234567 = _mm256_loadu_ps(i1); in xnn_f32_dwconv_ukernel_up8x9__avx() local 135 const __m256 vi1x01234567 = _mm256_maskload_ps(i1, vmask); in xnn_f32_dwconv_ukernel_up8x9__avx() local
|
D | up8x25-fma3-acc2.c | 99 const __m256 vi1x01234567 = _mm256_loadu_ps(i1); in xnn_f32_dwconv_ukernel_up8x25__fma3_acc2() local 265 const __m256 vi1x01234567 = _mm256_maskload_ps(i1, vmask); in xnn_f32_dwconv_ukernel_up8x25__fma3_acc2() local
|
D | up8x25-avx.c | 99 const __m256 vi1x01234567 = _mm256_loadu_ps(i1); in xnn_f32_dwconv_ukernel_up8x25__avx() local 263 const __m256 vi1x01234567 = _mm256_maskload_ps(i1, vmask); in xnn_f32_dwconv_ukernel_up8x25__avx() local
|
D | up8x25-avx-acc2.c | 99 const __m256 vi1x01234567 = _mm256_loadu_ps(i1); in xnn_f32_dwconv_ukernel_up8x25__avx_acc2() local 265 const __m256 vi1x01234567 = _mm256_maskload_ps(i1, vmask); in xnn_f32_dwconv_ukernel_up8x25__avx_acc2() local
|
D | up8x25-fma3.c | 99 const __m256 vi1x01234567 = _mm256_loadu_ps(i1); in xnn_f32_dwconv_ukernel_up8x25__fma3() local 263 const __m256 vi1x01234567 = _mm256_maskload_ps(i1, vmask); in xnn_f32_dwconv_ukernel_up8x25__fma3() local
|
D | up16x25-avx-acc2.c | 103 const __m256 vi1x01234567 = _mm256_loadu_ps(i1); in xnn_f32_dwconv_ukernel_up16x25__avx_acc2() local 343 const __m256 vi1x01234567 = _mm256_loadu_ps(i1); in xnn_f32_dwconv_ukernel_up16x25__avx_acc2() local 509 const __m256 vi1x01234567 = _mm256_maskload_ps(i1, vmask); in xnn_f32_dwconv_ukernel_up16x25__avx_acc2() local
|
D | up16x25-fma3.c | 103 const __m256 vi1x01234567 = _mm256_loadu_ps(i1); in xnn_f32_dwconv_ukernel_up16x25__fma3() local 340 const __m256 vi1x01234567 = _mm256_loadu_ps(i1); in xnn_f32_dwconv_ukernel_up16x25__fma3() local 504 const __m256 vi1x01234567 = _mm256_maskload_ps(i1, vmask); in xnn_f32_dwconv_ukernel_up16x25__fma3() local
|
D | up16x25-fma3-acc2.c | 103 const __m256 vi1x01234567 = _mm256_loadu_ps(i1); in xnn_f32_dwconv_ukernel_up16x25__fma3_acc2() local 343 const __m256 vi1x01234567 = _mm256_loadu_ps(i1); in xnn_f32_dwconv_ukernel_up16x25__fma3_acc2() local 509 const __m256 vi1x01234567 = _mm256_maskload_ps(i1, vmask); in xnn_f32_dwconv_ukernel_up16x25__fma3_acc2() local
|
D | up16x25-avx.c | 103 const __m256 vi1x01234567 = _mm256_loadu_ps(i1); in xnn_f32_dwconv_ukernel_up16x25__avx() local 340 const __m256 vi1x01234567 = _mm256_loadu_ps(i1); in xnn_f32_dwconv_ukernel_up16x25__avx() local 504 const __m256 vi1x01234567 = _mm256_maskload_ps(i1, vmask); in xnn_f32_dwconv_ukernel_up16x25__avx() local
|