Lines Matching refs:vo0p1
103 __m128 vo0p1 = _mm_mul_ps(vi1x4567, vk12); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc2() local
105 vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi3x4567, vk32)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc2()
131 vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi0x3456, vk01)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc2()
133 vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi2x3456, vk21)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc2()
135 vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi4x3456, vk41)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc2()
160 vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi1x2345, vk10)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc2()
162 vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi3x2345, vk30)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc2()
171 vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi0x5678, vk03)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc2()
173 vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi2x5678, vk23)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc2()
175 vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi4x5678, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc2()
184 vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi1x6789, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc2()
186 vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi3x6789, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc2()
189 vo0p0 = _mm_add_ps(vo0p0, vo0p1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc2()
201 __m128 vo0p1 = _mm_mul_ps(vi1x4567, vk12); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc2() local
203 vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi3x4567, vk32)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc2()
229 vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi0x3456, vk01)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc2()
231 vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi2x3456, vk21)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc2()
233 vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi4x3456, vk41)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc2()
258 vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi1x2345, vk10)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc2()
260 vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi3x2345, vk30)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc2()
269 vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi0x5678, vk03)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc2()
271 vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi2x5678, vk23)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc2()
273 vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi4x5678, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc2()
282 vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi1x6789, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc2()
284 vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi3x6789, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc2()
287 vo0p0 = _mm_add_ps(vo0p0, vo0p1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc2()
308 __m128 vo0p1 = _mm_mul_ps(vi1x4567, vk12); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc2() local
310 vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi3x4567, vk32)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc2()
325 vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi0x3456, vk01)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc2()
327 vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi2x3456, vk21)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc2()
329 vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi4x3456, vk41)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc2()
345 vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi1x2345, vk10)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc2()
347 vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi3x2345, vk30)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc2()
356 vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi0x5678, vk03)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc2()
358 vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi2x5678, vk23)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc2()
360 vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi4x5678, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc2()
369 vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi1x6789, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc2()
371 vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi3x6789, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc2()
374 vo0p0 = _mm_add_ps(vo0p0, vo0p1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_1x4_acc2()