| /external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
| D | 3x3p1-minmax-wasmsimd-x86-loadsplat-6x4.c | 205 const v128_t vi5x5678 = wasm_v32x4_shuffle(vi5x4567, vi5x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4() local
214 vo5p0 = wasm_f32x4_add(vo5p0, wasm_f32x4_mul(vi5x5678, vk02)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4()
219 vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi5x5678, vk12)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4()
224 vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x5678, vk22)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4()
329 const v128_t vi5x5678 = wasm_v32x4_shuffle(vi5x4567, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4() local
338 vo5p0 = wasm_f32x4_add(vo5p0, wasm_f32x4_mul(vi5x5678, vk02)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4()
343 vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi5x5678, vk12)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4()
348 vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x5678, vk22)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4()
|
| D | 3x3p1-minmax-wasmsimd-arm-loadsplat-6x4.c | 205 const v128_t vi5x5678 = wasm_v32x4_shuffle(vi5x4567, vi5x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_6x4() local
214 vo5p0 = wasm_f32x4_add(vo5p0, wasm_f32x4_mul(vi5x5678, vk02)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_6x4()
219 vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi5x5678, vk12)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_6x4()
224 vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x5678, vk22)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_6x4()
329 const v128_t vi5x5678 = wasm_v32x4_shuffle(vi5x4567, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_6x4() local
338 vo5p0 = wasm_f32x4_add(vo5p0, wasm_f32x4_mul(vi5x5678, vk02)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_6x4()
343 vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi5x5678, vk12)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_6x4()
348 vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x5678, vk22)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_6x4()
|
| D | 3x3p1-minmax-neon-6x4.c | 186 const float32x4_t vi5x5678 = vextq_f32(vi5x4567, vi5x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_6x4() local
195 vo5p0 = vmlaq_lane_f32(vo5p0, vi5x5678, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_6x4()
201 vo4p0 = vmlaq_lane_f32(vo4p0, vi5x5678, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_6x4()
207 vo3p0 = vmlaq_lane_f32(vo3p0, vi5x5678, vw89, 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_6x4()
319 const float32x4_t vi5x5678 = vextq_f32(vi5x4567, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_6x4() local
328 vo5p0 = vmlaq_lane_f32(vo5p0, vi5x5678, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_6x4()
334 vo4p0 = vmlaq_lane_f32(vo4p0, vi5x5678, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_6x4()
340 vo3p0 = vmlaq_lane_f32(vo3p0, vi5x5678, vw89, 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_6x4()
|
| D | 3x3p1-minmax-ssse3-6x4.c | 198 …const __m128 vi5x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi5x89AB), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_6x4() local
207 vo5p0 = _mm_add_ps(vo5p0, _mm_mul_ps(vi5x5678, vk02)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_6x4()
212 vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi5x5678, vk12)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_6x4()
217 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x5678, vk22)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_6x4()
324 … const __m128 vi5x5678 = _mm_castsi128_ps(_mm_alignr_epi8(vzero, _mm_castps_si128(vi5x4567), 4)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_6x4() local
333 vo5p0 = _mm_add_ps(vo5p0, _mm_mul_ps(vi5x5678, vk02)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_6x4()
338 vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi5x5678, vk12)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_6x4()
343 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x5678, vk22)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_6x4()
|
| D | 3x3p1-minmax-neonfma-6x4.c | 186 const float32x4_t vi5x5678 = vextq_f32(vi5x4567, vi5x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_6x4() local
195 vo5p0 = vfmaq_lane_f32(vo5p0, vi5x5678, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_6x4()
201 vo4p0 = vfmaq_lane_f32(vo4p0, vi5x5678, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_6x4()
207 vo3p0 = vfmaq_lane_f32(vo3p0, vi5x5678, vw89, 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_6x4()
319 const float32x4_t vi5x5678 = vextq_f32(vi5x4567, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_6x4() local
328 vo5p0 = vfmaq_lane_f32(vo5p0, vi5x5678, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_6x4()
334 vo4p0 = vfmaq_lane_f32(vo4p0, vi5x5678, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_6x4()
340 vo3p0 = vfmaq_lane_f32(vo3p0, vi5x5678, vw89, 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_6x4()
|
| D | 3x3p1-minmax-sse-6x4.c | 262 const __m128 vi5x5678 = _mm_shuffle_ps(vi5x8567, vi5x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_6x4() local
273 vo5p0 = _mm_add_ps(vo5p0, _mm_mul_ps(vi5x5678, vk02)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_6x4()
278 vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi5x5678, vk12)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_6x4()
283 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x5678, vk22)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_6x4()
438 const __m128 vi5x5678 = _mm_shuffle_ps(vi5x8567, vi5x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_6x4() local
449 vo5p0 = _mm_add_ps(vo5p0, _mm_mul_ps(vi5x5678, vk02)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_6x4()
454 vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi5x5678, vk12)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_6x4()
459 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x5678, vk22)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_6x4()
|
| D | 5x5p2-minmax-sse-5x4.c | 298 const __m128 vi5x5678 = _mm_shuffle_ps(vi5x8567, vi5x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4() local
312 vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi5x5678, vk13)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
316 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x5678, vk23)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
320 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi5x5678, vk33)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
324 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi5x5678, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
334 const __m128 vi5x6789 = _mm_shuffle_ps(vi5x5678, vi5x89AB, _MM_SHUFFLE(1, 0, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
551 const __m128 vi5x5678 = _mm_shuffle_ps(vi5x8567, vi5x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4() local
565 vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi5x5678, vk13)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
569 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x5678, vk23)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
573 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi5x5678, vk33)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
[all …]
|
| D | 3x3p1-minmax-wasmsimd-arm-loadsplat-5x4.c | 186 const v128_t vi5x5678 = wasm_v32x4_shuffle(vi5x4567, vi5x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_5x4() local
198 vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi5x5678, vk12)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_5x4()
202 vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x5678, vk22)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_5x4()
293 const v128_t vi5x5678 = wasm_v32x4_shuffle(vi5x4567, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_5x4() local
305 vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi5x5678, vk12)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_5x4()
309 vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x5678, vk22)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_5x4()
|
| D | 3x3p1-minmax-neon-5x4.c | 168 const float32x4_t vi5x5678 = vextq_f32(vi5x4567, vi5x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_5x4() local
181 vo4p0 = vmlaq_lane_f32(vo4p0, vi5x5678, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_5x4()
186 vo3p0 = vmlaq_lane_f32(vo3p0, vi5x5678, vw89, 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_5x4()
284 const float32x4_t vi5x5678 = vextq_f32(vi5x4567, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_5x4() local
297 vo4p0 = vmlaq_lane_f32(vo4p0, vi5x5678, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_5x4()
302 vo3p0 = vmlaq_lane_f32(vo3p0, vi5x5678, vw89, 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_5x4()
|
| D | 3x3p1-minmax-neonfma-5x4.c | 168 const float32x4_t vi5x5678 = vextq_f32(vi5x4567, vi5x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_5x4() local
181 vo4p0 = vfmaq_lane_f32(vo4p0, vi5x5678, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_5x4()
186 vo3p0 = vfmaq_lane_f32(vo3p0, vi5x5678, vw89, 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_5x4()
284 const float32x4_t vi5x5678 = vextq_f32(vi5x4567, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_5x4() local
297 vo4p0 = vfmaq_lane_f32(vo4p0, vi5x5678, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_5x4()
302 vo3p0 = vfmaq_lane_f32(vo3p0, vi5x5678, vw89, 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_5x4()
|
| D | 5x5p2-minmax-sse-3x4-acc2.c | 234 const __m128 vi5x5678 = _mm_shuffle_ps(vi5x8567, vi5x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2() local
248 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi5x5678, vk33)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2()
250 vo1p1 = _mm_add_ps(vo1p1, _mm_mul_ps(vi5x5678, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2()
258 const __m128 vi5x6789 = _mm_shuffle_ps(vi5x5678, vi5x89AB, _MM_SHUFFLE(1, 0, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2()
412 const __m128 vi5x5678 = _mm_shuffle_ps(vi5x8567, vi5x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2() local
426 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi5x5678, vk33)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2()
428 vo1p1 = _mm_add_ps(vo1p1, _mm_mul_ps(vi5x5678, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2()
436 const __m128 vi5x6789 = _mm_shuffle_ps(vi5x5678, vi5x89AB, _MM_SHUFFLE(1, 0, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2()
573 const __m128 vi5x5678 = _mm_shuffle_ps(vi5x8567, vi5x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2() local
587 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi5x5678, vk33)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2()
[all …]
|
| D | 5x5p2-minmax-sse-3x4.c | 234 const __m128 vi5x5678 = _mm_shuffle_ps(vi5x8567, vi5x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4() local
248 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi5x5678, vk33)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4()
250 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi5x5678, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4()
258 const __m128 vi5x6789 = _mm_shuffle_ps(vi5x5678, vi5x89AB, _MM_SHUFFLE(1, 0, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4()
409 const __m128 vi5x5678 = _mm_shuffle_ps(vi5x8567, vi5x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4() local
423 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi5x5678, vk33)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4()
425 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi5x5678, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4()
433 const __m128 vi5x6789 = _mm_shuffle_ps(vi5x5678, vi5x89AB, _MM_SHUFFLE(1, 0, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4()
567 const __m128 vi5x5678 = _mm_shuffle_ps(vi5x8567, vi5x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4() local
581 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi5x5678, vk33)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4()
[all …]
|
| D | 5x5p2-minmax-neon-4x4.c | 222 const float32x4_t vi5x5678 = vextq_f32(vi5x4567, vi5x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4() local
239 vo3p0 = vmlaq_lane_f32(vo3p0, vi5x5678, vget_high_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()
243 vo2p0 = vmlaq_lane_f32(vo2p0, vi5x5678, vget_high_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()
247 vo1p0 = vmlaq_lane_f32(vo1p0, vi5x5678, vwOP, 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()
440 const float32x4_t vi5x5678 = vextq_f32(vi5x4567, vi5x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4() local
457 vo3p0 = vmlaq_lane_f32(vo3p0, vi5x5678, vget_high_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()
461 vo2p0 = vmlaq_lane_f32(vo2p0, vi5x5678, vget_high_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()
465 vo1p0 = vmlaq_lane_f32(vo1p0, vi5x5678, vwOP, 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()
645 const float32x4_t vi5x5678 = vextq_f32(vi5x4567, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4() local
662 vo3p0 = vmlaq_lane_f32(vo3p0, vi5x5678, vget_high_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()
[all …]
|
| D | 5x5p2-minmax-neon-4x4-acc2.c | 222 const float32x4_t vi5x5678 = vextq_f32(vi5x4567, vi5x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4_acc2() local
239 vo3p1 = vmlaq_lane_f32(vo3p1, vi5x5678, vget_high_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4_acc2()
243 vo2p0 = vmlaq_lane_f32(vo2p0, vi5x5678, vget_high_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4_acc2()
247 vo1p1 = vmlaq_lane_f32(vo1p1, vi5x5678, vwOP, 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4_acc2()
444 const float32x4_t vi5x5678 = vextq_f32(vi5x4567, vi5x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4_acc2() local
461 vo3p1 = vmlaq_lane_f32(vo3p1, vi5x5678, vget_high_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4_acc2()
465 vo2p0 = vmlaq_lane_f32(vo2p0, vi5x5678, vget_high_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4_acc2()
469 vo1p1 = vmlaq_lane_f32(vo1p1, vi5x5678, vwOP, 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4_acc2()
653 const float32x4_t vi5x5678 = vextq_f32(vi5x4567, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4_acc2() local
670 vo3p1 = vmlaq_lane_f32(vo3p1, vi5x5678, vget_high_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4_acc2()
[all …]
|
| D | 5x5p2-minmax-neonfma-4x4.c | 222 const float32x4_t vi5x5678 = vextq_f32(vi5x4567, vi5x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4() local
239 vo3p0 = vfmaq_lane_f32(vo3p0, vi5x5678, vget_high_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4()
243 vo2p0 = vfmaq_lane_f32(vo2p0, vi5x5678, vget_high_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4()
247 vo1p0 = vfmaq_lane_f32(vo1p0, vi5x5678, vwOP, 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4()
440 const float32x4_t vi5x5678 = vextq_f32(vi5x4567, vi5x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4() local
457 vo3p0 = vfmaq_lane_f32(vo3p0, vi5x5678, vget_high_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4()
461 vo2p0 = vfmaq_lane_f32(vo2p0, vi5x5678, vget_high_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4()
465 vo1p0 = vfmaq_lane_f32(vo1p0, vi5x5678, vwOP, 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4()
645 const float32x4_t vi5x5678 = vextq_f32(vi5x4567, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4() local
662 vo3p0 = vfmaq_lane_f32(vo3p0, vi5x5678, vget_high_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4()
[all …]
|
| D | 5x5p2-minmax-neonfma-4x4-acc2.c | 222 const float32x4_t vi5x5678 = vextq_f32(vi5x4567, vi5x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4_acc2() local
239 vo3p1 = vfmaq_lane_f32(vo3p1, vi5x5678, vget_high_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4_acc2()
243 vo2p0 = vfmaq_lane_f32(vo2p0, vi5x5678, vget_high_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4_acc2()
247 vo1p1 = vfmaq_lane_f32(vo1p1, vi5x5678, vwOP, 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4_acc2()
444 const float32x4_t vi5x5678 = vextq_f32(vi5x4567, vi5x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4_acc2() local
461 vo3p1 = vfmaq_lane_f32(vo3p1, vi5x5678, vget_high_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4_acc2()
465 vo2p0 = vfmaq_lane_f32(vo2p0, vi5x5678, vget_high_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4_acc2()
469 vo1p1 = vfmaq_lane_f32(vo1p1, vi5x5678, vwOP, 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4_acc2()
653 const float32x4_t vi5x5678 = vextq_f32(vi5x4567, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4_acc2() local
670 vo3p1 = vfmaq_lane_f32(vo3p1, vi5x5678, vget_high_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4_acc2()
[all …]
|
| D | 5x5p2-minmax-sse-4x4.c | 266 const __m128 vi5x5678 = _mm_shuffle_ps(vi5x8567, vi5x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4() local
281 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x5678, vk23)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4()
284 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi5x5678, vk33)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4()
287 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi5x5678, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4()
296 const __m128 vi5x6789 = _mm_shuffle_ps(vi5x5678, vi5x89AB, _MM_SHUFFLE(1, 0, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4()
480 const __m128 vi5x5678 = _mm_shuffle_ps(vi5x8567, vi5x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4() local
495 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x5678, vk23)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4()
498 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi5x5678, vk33)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4()
501 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi5x5678, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4()
510 const __m128 vi5x6789 = _mm_shuffle_ps(vi5x5678, vi5x89AB, _MM_SHUFFLE(1, 0, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4()
[all …]
|
| D | 5x5p2-minmax-sse-4x4-acc2.c | 266 const __m128 vi5x5678 = _mm_shuffle_ps(vi5x8567, vi5x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2() local
281 vo3p1 = _mm_add_ps(vo3p1, _mm_mul_ps(vi5x5678, vk23)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2()
284 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi5x5678, vk33)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2()
287 vo1p1 = _mm_add_ps(vo1p1, _mm_mul_ps(vi5x5678, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2()
296 const __m128 vi5x6789 = _mm_shuffle_ps(vi5x5678, vi5x89AB, _MM_SHUFFLE(1, 0, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2()
484 const __m128 vi5x5678 = _mm_shuffle_ps(vi5x8567, vi5x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2() local
499 vo3p1 = _mm_add_ps(vo3p1, _mm_mul_ps(vi5x5678, vk23)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2()
502 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi5x5678, vk33)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2()
505 vo1p1 = _mm_add_ps(vo1p1, _mm_mul_ps(vi5x5678, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2()
514 const __m128 vi5x6789 = _mm_shuffle_ps(vi5x5678, vi5x89AB, _MM_SHUFFLE(1, 0, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2()
[all …]
|
| D | 3x3p1-minmax-wasmsimd-arm-splat-6x4.c | 189 const v128_t vi5x5678 = wasm_v32x4_shuffle(vi5x4567, vi5x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_6x4() local
198 …vo5p0 = wasm_f32x4_add(vo5p0, wasm_f32x4_mul(vi5x5678, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_6x4()
204 …vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi5x5678, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_6x4()
210 …vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x5678, wasm_v32x4_shuffle(vw89, vw89, 1, 1, 1, 1))… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_6x4()
320 const v128_t vi5x5678 = wasm_v32x4_shuffle(vi5x4567, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_6x4() local
329 …vo5p0 = wasm_f32x4_add(vo5p0, wasm_f32x4_mul(vi5x5678, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_6x4()
335 …vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi5x5678, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_6x4()
341 …vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x5678, wasm_v32x4_shuffle(vw89, vw89, 1, 1, 1, 1))… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_6x4()
|
| D | 3x3p1-minmax-wasmsimd-x86-splat-6x4.c | 189 const v128_t vi5x5678 = wasm_v32x4_shuffle(vi5x4567, vi5x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_6x4() local
198 …vo5p0 = wasm_f32x4_add(vo5p0, wasm_f32x4_mul(vi5x5678, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_6x4()
204 …vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi5x5678, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_6x4()
210 …vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x5678, wasm_v32x4_shuffle(vw89, vw89, 1, 1, 1, 1))… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_6x4()
320 const v128_t vi5x5678 = wasm_v32x4_shuffle(vi5x4567, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_6x4() local
329 …vo5p0 = wasm_f32x4_add(vo5p0, wasm_f32x4_mul(vi5x5678, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_6x4()
335 …vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi5x5678, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_6x4()
341 …vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x5678, wasm_v32x4_shuffle(vw89, vw89, 1, 1, 1, 1))… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_6x4()
|
| D | 5x5p2-minmax-wasmsimd-x86-loadsplat-5x4.c | 279 const v128_t vi5x5678 = wasm_v32x4_shuffle(vi5x4567, vi5x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4() local
294 vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi5x5678, vk13)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
299 vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x5678, vk23)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
304 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x5678, vk33)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
309 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x5678, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
533 const v128_t vi5x5678 = wasm_v32x4_shuffle(vi5x4567, vi5x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4() local
548 vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi5x5678, vk13)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
553 vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x5678, vk23)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
558 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x5678, vk33)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
563 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x5678, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
[all …]
|
| D | 5x5p2-minmax-neonfma-5x4.c | 250 const float32x4_t vi5x5678 = vextq_f32(vi5x4567, vi5x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4() local
265 vo4p0 = vfmaq_lane_f32(vo4p0, vi5x5678, vget_low_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4()
270 vo3p0 = vfmaq_lane_f32(vo3p0, vi5x5678, vget_high_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4()
275 vo2p0 = vfmaq_lane_f32(vo2p0, vi5x5678, vget_high_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4()
280 vo1p0 = vfmaq_lane_f32(vo1p0, vi5x5678, vwOP, 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4()
505 const float32x4_t vi5x5678 = vextq_f32(vi5x4567, vi5x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4() local
520 vo4p0 = vfmaq_lane_f32(vo4p0, vi5x5678, vget_low_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4()
525 vo3p0 = vfmaq_lane_f32(vo3p0, vi5x5678, vget_high_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4()
530 vo2p0 = vfmaq_lane_f32(vo2p0, vi5x5678, vget_high_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4()
535 vo1p0 = vfmaq_lane_f32(vo1p0, vi5x5678, vwOP, 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4()
[all …]
|
| D | 5x5p2-minmax-neon-5x4.c | 250 const float32x4_t vi5x5678 = vextq_f32(vi5x4567, vi5x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4() local
265 vo4p0 = vmlaq_lane_f32(vo4p0, vi5x5678, vget_low_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
270 vo3p0 = vmlaq_lane_f32(vo3p0, vi5x5678, vget_high_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
275 vo2p0 = vmlaq_lane_f32(vo2p0, vi5x5678, vget_high_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
280 vo1p0 = vmlaq_lane_f32(vo1p0, vi5x5678, vwOP, 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
505 const float32x4_t vi5x5678 = vextq_f32(vi5x4567, vi5x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4() local
520 vo4p0 = vmlaq_lane_f32(vo4p0, vi5x5678, vget_low_f32(vw89AB), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
525 vo3p0 = vmlaq_lane_f32(vo3p0, vi5x5678, vget_high_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
530 vo2p0 = vmlaq_lane_f32(vo2p0, vi5x5678, vget_high_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
535 vo1p0 = vmlaq_lane_f32(vo1p0, vi5x5678, vwOP, 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
[all …]
|
| D | 5x5p2-minmax-wasmsimd-arm-loadsplat-5x4.c | 279 const v128_t vi5x5678 = wasm_v32x4_shuffle(vi5x4567, vi5x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4() local
294 vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi5x5678, vk13)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4()
299 vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x5678, vk23)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4()
304 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x5678, vk33)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4()
309 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x5678, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4()
533 const v128_t vi5x5678 = wasm_v32x4_shuffle(vi5x4567, vi5x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4() local
548 vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi5x5678, vk13)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4()
553 vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x5678, vk23)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4()
558 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x5678, vk33)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4()
563 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x5678, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4()
[all …]
|
| D | 5x5p2-minmax-sse-2x4.c | 202 const __m128 vi5x5678 = _mm_shuffle_ps(vi5x8567, vi5x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4() local
213 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi5x5678, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4()
220 const __m128 vi5x6789 = _mm_shuffle_ps(vi5x5678, vi5x89AB, _MM_SHUFFLE(1, 0, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4()
338 const __m128 vi5x5678 = _mm_shuffle_ps(vi5x8567, vi5x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4() local
349 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi5x5678, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4()
356 const __m128 vi5x6789 = _mm_shuffle_ps(vi5x5678, vi5x89AB, _MM_SHUFFLE(1, 0, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4()
460 const __m128 vi5x5678 = _mm_shuffle_ps(vi5x8567, vi5x8567, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4() local
471 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi5x5678, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4()
478 const __m128 vi5x6789 = _mm_shuffle_ps(vi5x5678, vzero, _MM_SHUFFLE(1, 0, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_2x4()
|