Home
last modified time | relevance | path

Searched refs:vi1x8567 (Results 1 – 23 of 23) sorted by relevance

/external/XNNPACK/src/f32-dwconv2d-chw/gen/
D3x3p1-minmax-sse-1x4.c117 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_1x4() local
124 const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_1x4()
178 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_1x4() local
185 const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_1x4()
D3x3p1-minmax-sse-1x4-acc3.c117 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_1x4_acc3() local
124 const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_1x4_acc3()
180 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_1x4_acc3() local
187 const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_1x4_acc3()
D3x3p1-minmax-sse-1x4-acc2.c117 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_1x4_acc2() local
124 const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_1x4_acc2()
179 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_1x4_acc2() local
186 const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_1x4_acc2()
D3x3p1-minmax-sse-1x4-acc4.c117 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_1x4_acc4() local
124 const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_1x4_acc4()
181 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_1x4_acc4() local
188 const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_1x4_acc4()
D3x3p1-minmax-sse-2x4.c141 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_2x4() local
150 const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_2x4()
225 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_2x4() local
234 const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_2x4()
D3x3p1-minmax-sse-2x4-acc2.c141 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_2x4_acc2() local
150 const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_2x4_acc2()
227 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_2x4_acc2() local
236 const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_2x4_acc2()
D5x5p2-minmax-sse-1x4-acc4.c150 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc4() local
166 const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc4()
250 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc4() local
266 const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc4()
343 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc4() local
355 const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc4()
D5x5p2-minmax-sse-1x4.c150 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4() local
166 const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
247 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4() local
263 const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
337 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4() local
349 const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4()
D5x5p2-minmax-sse-1x4-acc2.c150 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc2() local
166 const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc2()
248 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc2() local
264 const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc2()
339 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc2() local
351 const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc2()
D5x5p2-minmax-sse-1x4-acc3.c150 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc3() local
166 const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc3()
249 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc3() local
265 const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc3()
341 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc3() local
353 const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc3()
D5x5p2-minmax-sse-1x4-acc5.c150 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc5() local
166 const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc5()
251 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc5() local
267 const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc5()
345 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc5() local
357 const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc5()
D3x3p1-minmax-sse-3x4.c165 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_3x4() local
176 const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_3x4()
272 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_3x4() local
283 const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_3x4()
D3x3p1-minmax-sse-4x4.c189 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_4x4() local
202 const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_4x4()
319 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_4x4() local
332 const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_4x4()
D5x5p2-minmax-sse-2x4.c175 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4() local
198 const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4()
311 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4() local
334 const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4()
438 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4() local
456 const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4()
D3x3p1-minmax-sse-5x4.c213 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_5x4() local
228 const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_5x4()
366 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_5x4() local
381 const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_5x4()
D5x5p2-minmax-sse-2x4-acc2.c175 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc2() local
198 const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc2()
313 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc2() local
336 const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc2()
442 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc2() local
460 const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc2()
D5x5p2-minmax-sse-2x4-acc3.c175 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc3() local
198 const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc3()
315 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc3() local
338 const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc3()
446 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc3() local
464 const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4_acc3()
D3x3p1-minmax-sse-6x4.c237 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_6x4() local
254 const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_6x4()
413 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_6x4() local
430 const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_6x4()
D5x5p2-minmax-sse-3x4-acc2.c200 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2() local
230 const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2()
378 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2() local
408 const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2()
545 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2() local
569 const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2()
D5x5p2-minmax-sse-3x4.c200 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4() local
230 const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4()
375 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4() local
405 const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4()
539 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4() local
563 const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4()
D5x5p2-minmax-sse-4x4.c225 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4() local
262 const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4()
439 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4() local
476 const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4()
640 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4() local
670 const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4()
D5x5p2-minmax-sse-4x4-acc2.c225 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2() local
262 const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2()
443 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2() local
480 const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2()
648 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2() local
678 const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2()
D5x5p2-minmax-sse-5x4.c250 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4() local
294 const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
503 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vi1x89AB); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4() local
547 const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
741 const __m128 vi1x8567 = _mm_move_ss(vi1x4567, vzero); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4() local
777 const __m128 vi1x5678 = _mm_shuffle_ps(vi1x8567, vi1x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()