/external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
D | 5x5p2-minmax-sse-2x4.c | 127 const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4() local 263 const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4() local 409 const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4() local
|
D | 3x3p1-minmax-sse-4x4.c | 138 const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_4x4() local 274 const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_4x4() local
|
D | 5x5p2-minmax-sse-2x4-acc2.c | 127 const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc2() local 265 const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc2() local 413 const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc2() local
|
D | 5x5p2-minmax-sse-2x4-acc3.c | 127 const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc3() local 267 const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc3() local 417 const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc3() local
|
D | 3x3p1-minmax-sse-5x4.c | 151 const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_5x4() local 311 const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_5x4() local
|
D | 5x5p2-minmax-sse-3x4.c | 141 const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4() local 316 const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4() local 502 const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4() local
|
D | 5x5p2-minmax-sse-3x4-acc2.c | 141 const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2() local 319 const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2() local 508 const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2() local
|
D | 3x3p1-minmax-sse-6x4.c | 164 const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_6x4() local 348 const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_6x4() local
|
D | 5x5p2-minmax-sse-4x4.c | 155 const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4() local 369 const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4() local 595 const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4() local
|
D | 5x5p2-minmax-sse-4x4-acc2.c | 155 const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2() local 373 const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2() local 603 const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2() local
|
D | 5x5p2-minmax-sse-5x4.c | 169 const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4() local 422 const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4() local 688 const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4() local
|
/external/XNNPACK/src/amalgam/ |
D | sse.c | 3035 const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4() local 3249 const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4() local 3475 const __m128 vi5x7456 = _mm_shuffle_ps(vi5x4567, vi5x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4() local
|