/external/XNNPACK/src/f32-dwconv/gen/ |
D | up16x9-fma3-acc2.c | 125 const __m256 vi7x01234567 = _mm256_loadu_ps(i7); in xnn_f32_dwconv_ukernel_up16x9__fma3_acc2() local 203 const __m256 vi7x01234567 = _mm256_loadu_ps(i7); in xnn_f32_dwconv_ukernel_up16x9__fma3_acc2() local 261 const __m256 vi7x01234567 = _mm256_maskload_ps(i7, vmask); in xnn_f32_dwconv_ukernel_up16x9__fma3_acc2() local
|
D | up16x9-avx-acc2.c | 125 const __m256 vi7x01234567 = _mm256_loadu_ps(i7); in xnn_f32_dwconv_ukernel_up16x9__avx_acc2() local 203 const __m256 vi7x01234567 = _mm256_loadu_ps(i7); in xnn_f32_dwconv_ukernel_up16x9__avx_acc2() local 261 const __m256 vi7x01234567 = _mm256_maskload_ps(i7, vmask); in xnn_f32_dwconv_ukernel_up16x9__avx_acc2() local
|
D | up16x9-fma3.c | 125 const __m256 vi7x01234567 = _mm256_loadu_ps(i7); in xnn_f32_dwconv_ukernel_up16x9__fma3() local 200 const __m256 vi7x01234567 = _mm256_loadu_ps(i7); in xnn_f32_dwconv_ukernel_up16x9__fma3() local 256 const __m256 vi7x01234567 = _mm256_maskload_ps(i7, vmask); in xnn_f32_dwconv_ukernel_up16x9__fma3() local
|
D | up16x9-avx.c | 125 const __m256 vi7x01234567 = _mm256_loadu_ps(i7); in xnn_f32_dwconv_ukernel_up16x9__avx() local 200 const __m256 vi7x01234567 = _mm256_loadu_ps(i7); in xnn_f32_dwconv_ukernel_up16x9__avx() local 256 const __m256 vi7x01234567 = _mm256_maskload_ps(i7, vmask); in xnn_f32_dwconv_ukernel_up16x9__avx() local
|
D | up8x9-avx-acc2.c | 103 const __m256 vi7x01234567 = _mm256_loadu_ps(i7); in xnn_f32_dwconv_ukernel_up8x9__avx_acc2() local 161 const __m256 vi7x01234567 = _mm256_maskload_ps(i7, vmask); in xnn_f32_dwconv_ukernel_up8x9__avx_acc2() local
|
D | up8x9-fma3-acc2.c | 103 const __m256 vi7x01234567 = _mm256_loadu_ps(i7); in xnn_f32_dwconv_ukernel_up8x9__fma3_acc2() local 161 const __m256 vi7x01234567 = _mm256_maskload_ps(i7, vmask); in xnn_f32_dwconv_ukernel_up8x9__fma3_acc2() local
|
D | up8x9-fma3.c | 103 const __m256 vi7x01234567 = _mm256_loadu_ps(i7); in xnn_f32_dwconv_ukernel_up8x9__fma3() local 159 const __m256 vi7x01234567 = _mm256_maskload_ps(i7, vmask); in xnn_f32_dwconv_ukernel_up8x9__fma3() local
|
D | up8x9-avx.c | 103 const __m256 vi7x01234567 = _mm256_loadu_ps(i7); in xnn_f32_dwconv_ukernel_up8x9__avx() local 159 const __m256 vi7x01234567 = _mm256_maskload_ps(i7, vmask); in xnn_f32_dwconv_ukernel_up8x9__avx() local
|
D | up8x25-fma3-acc2.c | 135 const __m256 vi7x01234567 = _mm256_loadu_ps(i7); in xnn_f32_dwconv_ukernel_up8x25__fma3_acc2() local 289 const __m256 vi7x01234567 = _mm256_maskload_ps(i7, vmask); in xnn_f32_dwconv_ukernel_up8x25__fma3_acc2() local
|
D | up8x25-avx.c | 135 const __m256 vi7x01234567 = _mm256_loadu_ps(i7); in xnn_f32_dwconv_ukernel_up8x25__avx() local 287 const __m256 vi7x01234567 = _mm256_maskload_ps(i7, vmask); in xnn_f32_dwconv_ukernel_up8x25__avx() local
|
D | up8x25-avx-acc2.c | 135 const __m256 vi7x01234567 = _mm256_loadu_ps(i7); in xnn_f32_dwconv_ukernel_up8x25__avx_acc2() local 289 const __m256 vi7x01234567 = _mm256_maskload_ps(i7, vmask); in xnn_f32_dwconv_ukernel_up8x25__avx_acc2() local
|
D | up8x25-fma3.c | 135 const __m256 vi7x01234567 = _mm256_loadu_ps(i7); in xnn_f32_dwconv_ukernel_up8x25__fma3() local 287 const __m256 vi7x01234567 = _mm256_maskload_ps(i7, vmask); in xnn_f32_dwconv_ukernel_up8x25__fma3() local
|
D | up16x25-avx-acc2.c | 157 const __m256 vi7x01234567 = _mm256_loadu_ps(i7); in xnn_f32_dwconv_ukernel_up16x25__avx_acc2() local 379 const __m256 vi7x01234567 = _mm256_loadu_ps(i7); in xnn_f32_dwconv_ukernel_up16x25__avx_acc2() local 533 const __m256 vi7x01234567 = _mm256_maskload_ps(i7, vmask); in xnn_f32_dwconv_ukernel_up16x25__avx_acc2() local
|
D | up16x25-fma3.c | 157 const __m256 vi7x01234567 = _mm256_loadu_ps(i7); in xnn_f32_dwconv_ukernel_up16x25__fma3() local 376 const __m256 vi7x01234567 = _mm256_loadu_ps(i7); in xnn_f32_dwconv_ukernel_up16x25__fma3() local 528 const __m256 vi7x01234567 = _mm256_maskload_ps(i7, vmask); in xnn_f32_dwconv_ukernel_up16x25__fma3() local
|
D | up16x25-fma3-acc2.c | 157 const __m256 vi7x01234567 = _mm256_loadu_ps(i7); in xnn_f32_dwconv_ukernel_up16x25__fma3_acc2() local 379 const __m256 vi7x01234567 = _mm256_loadu_ps(i7); in xnn_f32_dwconv_ukernel_up16x25__fma3_acc2() local 533 const __m256 vi7x01234567 = _mm256_maskload_ps(i7, vmask); in xnn_f32_dwconv_ukernel_up16x25__fma3_acc2() local
|
D | up16x25-avx.c | 157 const __m256 vi7x01234567 = _mm256_loadu_ps(i7); in xnn_f32_dwconv_ukernel_up16x25__avx() local 376 const __m256 vi7x01234567 = _mm256_loadu_ps(i7); in xnn_f32_dwconv_ukernel_up16x25__avx() local 528 const __m256 vi7x01234567 = _mm256_maskload_ps(i7, vmask); in xnn_f32_dwconv_ukernel_up16x25__avx() local
|