/external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
D | 3x3p1-minmax-sse-1x4.c | 91 const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_1x4() local 155 const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_1x4() local
|
D | 3x3p1-minmax-sse-1x4-acc2.c | 91 const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_1x4_acc2() local 156 const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_1x4_acc2() local
|
D | 3x3p1-minmax-sse-1x4-acc3.c | 91 const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_1x4_acc3() local 157 const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_1x4_acc3() local
|
D | 3x3p1-minmax-sse-1x4-acc4.c | 91 const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_1x4_acc4() local 158 const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_1x4_acc4() local
|
D | 3x3p1-minmax-sse-2x4.c | 104 const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_2x4() local 192 const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_2x4() local
|
D | 3x3p1-minmax-sse-2x4-acc2.c | 104 const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_2x4_acc2() local 194 const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_2x4_acc2() local
|
D | 5x5p2-minmax-sse-1x4.c | 109 const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4() local 206 const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4() local 312 const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4() local
|
D | 5x5p2-minmax-sse-1x4-acc2.c | 109 const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc2() local 207 const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc2() local 314 const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc2() local
|
D | 3x3p1-minmax-sse-3x4.c | 117 const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_3x4() local 229 const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_3x4() local
|
D | 5x5p2-minmax-sse-1x4-acc3.c | 109 const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc3() local 208 const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc3() local 316 const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc3() local
|
D | 5x5p2-minmax-sse-1x4-acc4.c | 109 const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc4() local 209 const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc4() local 318 const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc4() local
|
D | 5x5p2-minmax-sse-1x4-acc5.c | 109 const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc5() local 210 const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc5() local 320 const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc5() local
|
D | 5x5p2-minmax-sse-2x4.c | 123 const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4() local 259 const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4() local 405 const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4() local
|
D | 3x3p1-minmax-sse-4x4.c | 130 const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_4x4() local 266 const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_4x4() local
|
D | 5x5p2-minmax-sse-2x4-acc2.c | 123 const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc2() local 261 const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc2() local 409 const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc2() local
|
D | 5x5p2-minmax-sse-2x4-acc3.c | 123 const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc3() local 263 const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc3() local 413 const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc3() local
|
D | 3x3p1-minmax-sse-5x4.c | 143 const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_5x4() local 303 const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_5x4() local
|
D | 5x5p2-minmax-sse-3x4.c | 137 const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4() local 312 const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4() local 498 const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4() local
|
D | 5x5p2-minmax-sse-3x4-acc2.c | 137 const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2() local 315 const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2() local 504 const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2() local
|
D | 3x3p1-minmax-sse-6x4.c | 156 const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_6x4() local 340 const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_6x4() local
|
D | 5x5p2-minmax-sse-4x4.c | 151 const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4() local 365 const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4() local 591 const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4() local
|
D | 5x5p2-minmax-sse-4x4-acc2.c | 151 const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2() local 369 const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2() local 599 const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2() local
|
D | 5x5p2-minmax-sse-5x4.c | 165 const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4() local 418 const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4() local 684 const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4() local
|
/external/XNNPACK/src/amalgam/ |
D | sse.c | 2525 const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_2x4_acc2() local 2615 const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_2x4_acc2() local 3031 const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4() local 3245 const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4() local 3471 const __m128 vi1x7456 = _mm_shuffle_ps(vi1x4567, vi1x4567, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4() local
|