| /external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
| D | 5x5p2-minmax-neon-4x4.c | 261 const float32x4_t vi5x6789 = vextq_f32(vi5x4567, vi5x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4() local
281 vo3p0 = vmlaq_lane_f32(vo3p0, vi5x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()
285 vo2p0 = vmlaq_lane_f32(vo2p0, vi5x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()
289 vo1p0 = vmlaq_lane_f32(vo1p0, vi5x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()
479 const float32x4_t vi5x6789 = vextq_f32(vi5x4567, vi5x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4() local
499 vo3p0 = vmlaq_lane_f32(vo3p0, vi5x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()
503 vo2p0 = vmlaq_lane_f32(vo2p0, vi5x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()
507 vo1p0 = vmlaq_lane_f32(vo1p0, vi5x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()
679 const float32x4_t vi5x6789 = vextq_f32(vi5x5678, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4() local
696 vo3p0 = vmlaq_lane_f32(vo3p0, vi5x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()
[all …]
|
| D | 5x5p2-minmax-neon-4x4-acc2.c | 261 const float32x4_t vi5x6789 = vextq_f32(vi5x4567, vi5x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4_acc2() local
281 vo3p0 = vmlaq_lane_f32(vo3p0, vi5x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4_acc2()
285 vo2p1 = vmlaq_lane_f32(vo2p1, vi5x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4_acc2()
289 vo1p0 = vmlaq_lane_f32(vo1p0, vi5x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4_acc2()
483 const float32x4_t vi5x6789 = vextq_f32(vi5x4567, vi5x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4_acc2() local
503 vo3p0 = vmlaq_lane_f32(vo3p0, vi5x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4_acc2()
507 vo2p1 = vmlaq_lane_f32(vo2p1, vi5x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4_acc2()
511 vo1p0 = vmlaq_lane_f32(vo1p0, vi5x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4_acc2()
687 const float32x4_t vi5x6789 = vextq_f32(vi5x5678, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4_acc2() local
704 vo3p0 = vmlaq_lane_f32(vo3p0, vi5x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4_acc2()
[all …]
|
| D | 5x5p2-minmax-neonfma-4x4.c | 261 const float32x4_t vi5x6789 = vextq_f32(vi5x4567, vi5x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4() local
281 vo3p0 = vfmaq_lane_f32(vo3p0, vi5x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4()
285 vo2p0 = vfmaq_lane_f32(vo2p0, vi5x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4()
289 vo1p0 = vfmaq_lane_f32(vo1p0, vi5x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4()
479 const float32x4_t vi5x6789 = vextq_f32(vi5x4567, vi5x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4() local
499 vo3p0 = vfmaq_lane_f32(vo3p0, vi5x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4()
503 vo2p0 = vfmaq_lane_f32(vo2p0, vi5x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4()
507 vo1p0 = vfmaq_lane_f32(vo1p0, vi5x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4()
679 const float32x4_t vi5x6789 = vextq_f32(vi5x5678, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4() local
696 vo3p0 = vfmaq_lane_f32(vo3p0, vi5x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4()
[all …]
|
| D | 5x5p2-minmax-neonfma-4x4-acc2.c | 261 const float32x4_t vi5x6789 = vextq_f32(vi5x4567, vi5x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4_acc2() local
281 vo3p0 = vfmaq_lane_f32(vo3p0, vi5x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4_acc2()
285 vo2p1 = vfmaq_lane_f32(vo2p1, vi5x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4_acc2()
289 vo1p0 = vfmaq_lane_f32(vo1p0, vi5x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4_acc2()
483 const float32x4_t vi5x6789 = vextq_f32(vi5x4567, vi5x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4_acc2() local
503 vo3p0 = vfmaq_lane_f32(vo3p0, vi5x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4_acc2()
507 vo2p1 = vfmaq_lane_f32(vo2p1, vi5x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4_acc2()
511 vo1p0 = vfmaq_lane_f32(vo1p0, vi5x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4_acc2()
687 const float32x4_t vi5x6789 = vextq_f32(vi5x5678, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4_acc2() local
704 vo3p0 = vfmaq_lane_f32(vo3p0, vi5x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4_acc2()
[all …]
|
| D | 5x5p2-minmax-wasmsimd-x86-loadsplat-5x4.c | 324 const v128_t vi5x6789 = wasm_v32x4_shuffle(vi5x4567, vi5x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4() local
343 vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi5x6789, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
348 vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x6789, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
353 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x6789, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
358 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x6789, vk44)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
578 const v128_t vi5x6789 = wasm_v32x4_shuffle(vi5x4567, vi5x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4() local
597 vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi5x6789, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
602 vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x6789, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
607 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x6789, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
612 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x6789, vk44)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
[all …]
|
| D | 5x5p2-minmax-neonfma-5x4.c | 295 const float32x4_t vi5x6789 = vextq_f32(vi5x4567, vi5x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4() local
314 vo4p0 = vfmaq_lane_f32(vo4p0, vi5x6789, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4()
319 vo3p0 = vfmaq_lane_f32(vo3p0, vi5x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4()
324 vo2p0 = vfmaq_lane_f32(vo2p0, vi5x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4()
329 vo1p0 = vfmaq_lane_f32(vo1p0, vi5x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4()
550 const float32x4_t vi5x6789 = vextq_f32(vi5x4567, vi5x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4() local
569 vo4p0 = vfmaq_lane_f32(vo4p0, vi5x6789, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4()
574 vo3p0 = vfmaq_lane_f32(vo3p0, vi5x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4()
579 vo2p0 = vfmaq_lane_f32(vo2p0, vi5x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4()
584 vo1p0 = vfmaq_lane_f32(vo1p0, vi5x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4()
[all …]
|
| D | 5x5p2-minmax-neon-5x4.c | 295 const float32x4_t vi5x6789 = vextq_f32(vi5x4567, vi5x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4() local
314 vo4p0 = vmlaq_lane_f32(vo4p0, vi5x6789, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
319 vo3p0 = vmlaq_lane_f32(vo3p0, vi5x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
324 vo2p0 = vmlaq_lane_f32(vo2p0, vi5x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
329 vo1p0 = vmlaq_lane_f32(vo1p0, vi5x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
550 const float32x4_t vi5x6789 = vextq_f32(vi5x4567, vi5x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4() local
569 vo4p0 = vmlaq_lane_f32(vo4p0, vi5x6789, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
574 vo3p0 = vmlaq_lane_f32(vo3p0, vi5x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
579 vo2p0 = vmlaq_lane_f32(vo2p0, vi5x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
584 vo1p0 = vmlaq_lane_f32(vo1p0, vi5x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
[all …]
|
| D | 5x5p2-minmax-wasmsimd-arm-loadsplat-5x4.c | 324 const v128_t vi5x6789 = wasm_v32x4_shuffle(vi5x4567, vi5x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4() local
343 vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi5x6789, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4()
348 vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x6789, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4()
353 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x6789, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4()
358 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x6789, vk44)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4()
578 const v128_t vi5x6789 = wasm_v32x4_shuffle(vi5x4567, vi5x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4() local
597 vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi5x6789, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4()
602 vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x6789, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4()
607 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x6789, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4()
612 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x6789, vk44)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4()
[all …]
|
| D | 5x5p2-minmax-sse-5x4.c | 334 const __m128 vi5x6789 = _mm_shuffle_ps(vi5x5678, vi5x89AB, _MM_SHUFFLE(1, 0, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4() local
348 vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi5x6789, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
352 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x6789, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
356 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi5x6789, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
360 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi5x6789, vk44)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
587 const __m128 vi5x6789 = _mm_shuffle_ps(vi5x5678, vi5x89AB, _MM_SHUFFLE(1, 0, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4() local
601 vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi5x6789, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
605 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x6789, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
609 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi5x6789, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
613 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi5x6789, vk44)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
[all …]
|
| D | 5x5p2-minmax-wasmsimd-x86-loadsplat-4x4-acc2.c | 290 const v128_t vi5x6789 = wasm_v32x4_shuffle(vi5x4567, vi5x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_4x4_acc2() local
310 vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x6789, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_4x4_acc2()
314 vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi5x6789, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_4x4_acc2()
318 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x6789, vk44)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_4x4_acc2()
511 const v128_t vi5x6789 = wasm_v32x4_shuffle(vi5x4567, vi5x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_4x4_acc2() local
531 vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x6789, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_4x4_acc2()
535 vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi5x6789, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_4x4_acc2()
539 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x6789, vk44)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_4x4_acc2()
713 const v128_t vi5x6789 = wasm_v32x4_shuffle(vi5x5678, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_4x4_acc2() local
730 vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x6789, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_4x4_acc2()
[all …]
|
| D | 5x5p2-minmax-wasmsimd-arm-loadsplat-4x4.c | 290 const v128_t vi5x6789 = wasm_v32x4_shuffle(vi5x4567, vi5x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4() local
310 vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x6789, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4()
314 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x6789, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4()
318 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x6789, vk44)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4()
507 const v128_t vi5x6789 = wasm_v32x4_shuffle(vi5x4567, vi5x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4() local
527 vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x6789, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4()
531 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x6789, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4()
535 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x6789, vk44)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4()
705 const v128_t vi5x6789 = wasm_v32x4_shuffle(vi5x5678, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4() local
722 vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x6789, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4()
[all …]
|
| D | 5x5p2-minmax-wasmsimd-arm-loadsplat-4x4-acc2.c | 290 const v128_t vi5x6789 = wasm_v32x4_shuffle(vi5x4567, vi5x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4_acc2() local
310 vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x6789, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4_acc2()
314 vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi5x6789, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4_acc2()
318 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x6789, vk44)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4_acc2()
511 const v128_t vi5x6789 = wasm_v32x4_shuffle(vi5x4567, vi5x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4_acc2() local
531 vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x6789, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4_acc2()
535 vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi5x6789, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4_acc2()
539 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x6789, vk44)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4_acc2()
713 const v128_t vi5x6789 = wasm_v32x4_shuffle(vi5x5678, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4_acc2() local
730 vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x6789, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4_acc2()
[all …]
|
| D | 5x5p2-minmax-wasmsimd-x86-loadsplat-4x4.c | 290 const v128_t vi5x6789 = wasm_v32x4_shuffle(vi5x4567, vi5x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_4x4() local
310 vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x6789, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_4x4()
314 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x6789, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_4x4()
318 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x6789, vk44)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_4x4()
507 const v128_t vi5x6789 = wasm_v32x4_shuffle(vi5x4567, vi5x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_4x4() local
527 vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x6789, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_4x4()
531 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x6789, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_4x4()
535 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x6789, vk44)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_4x4()
705 const v128_t vi5x6789 = wasm_v32x4_shuffle(vi5x5678, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_4x4() local
722 vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x6789, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_4x4()
[all …]
|
| D | 5x5p2-minmax-neonfma-3x4.c | 227 const float32x4_t vi5x6789 = vextq_f32(vi5x4567, vi5x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4() local
246 vo2p0 = vfmaq_lane_f32(vo2p0, vi5x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4()
249 vo1p0 = vfmaq_lane_f32(vo1p0, vi5x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4()
408 const float32x4_t vi5x6789 = vextq_f32(vi5x4567, vi5x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4() local
427 vo2p0 = vfmaq_lane_f32(vo2p0, vi5x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4()
430 vo1p0 = vfmaq_lane_f32(vo1p0, vi5x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4()
573 const float32x4_t vi5x6789 = vextq_f32(vi5x5678, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4() local
590 vo2p0 = vfmaq_lane_f32(vo2p0, vi5x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4()
593 vo1p0 = vfmaq_lane_f32(vo1p0, vi5x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4()
|
| D | 5x5p2-minmax-wasmsimd-arm-loadsplat-3x4.c | 256 const v128_t vi5x6789 = wasm_v32x4_shuffle(vi5x4567, vi5x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4() local
275 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x6789, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4()
278 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x6789, vk44)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4()
436 const v128_t vi5x6789 = wasm_v32x4_shuffle(vi5x4567, vi5x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4() local
455 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x6789, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4()
458 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x6789, vk44)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4()
599 const v128_t vi5x6789 = wasm_v32x4_shuffle(vi5x5678, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4() local
616 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x6789, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4()
619 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x6789, vk44)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4()
|
| D | 5x5p2-minmax-wasmsimd-arm-loadsplat-3x4-acc2.c | 256 const v128_t vi5x6789 = wasm_v32x4_shuffle(vi5x4567, vi5x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4_acc2() local
275 vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi5x6789, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4_acc2()
278 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x6789, vk44)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4_acc2()
439 const v128_t vi5x6789 = wasm_v32x4_shuffle(vi5x4567, vi5x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4_acc2() local
458 vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi5x6789, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4_acc2()
461 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x6789, vk44)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4_acc2()
605 const v128_t vi5x6789 = wasm_v32x4_shuffle(vi5x5678, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4_acc2() local
622 vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi5x6789, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4_acc2()
625 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x6789, vk44)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4_acc2()
|
| D | 5x5p2-minmax-wasmsimd-x86-loadsplat-3x4.c | 256 const v128_t vi5x6789 = wasm_v32x4_shuffle(vi5x4567, vi5x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4() local
275 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x6789, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
278 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x6789, vk44)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
436 const v128_t vi5x6789 = wasm_v32x4_shuffle(vi5x4567, vi5x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4() local
455 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x6789, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
458 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x6789, vk44)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
599 const v128_t vi5x6789 = wasm_v32x4_shuffle(vi5x5678, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4() local
616 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x6789, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
619 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x6789, vk44)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
|
| D | 5x5p2-minmax-neon-3x4-acc2.c | 227 const float32x4_t vi5x6789 = vextq_f32(vi5x4567, vi5x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4_acc2() local
246 vo2p1 = vmlaq_lane_f32(vo2p1, vi5x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4_acc2()
249 vo1p0 = vmlaq_lane_f32(vo1p0, vi5x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4_acc2()
411 const float32x4_t vi5x6789 = vextq_f32(vi5x4567, vi5x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4_acc2() local
430 vo2p1 = vmlaq_lane_f32(vo2p1, vi5x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4_acc2()
433 vo1p0 = vmlaq_lane_f32(vo1p0, vi5x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4_acc2()
579 const float32x4_t vi5x6789 = vextq_f32(vi5x5678, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4_acc2() local
596 vo2p1 = vmlaq_lane_f32(vo2p1, vi5x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4_acc2()
599 vo1p0 = vmlaq_lane_f32(vo1p0, vi5x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4_acc2()
|
| D | 5x5p2-minmax-neon-3x4.c | 227 const float32x4_t vi5x6789 = vextq_f32(vi5x4567, vi5x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4() local
246 vo2p0 = vmlaq_lane_f32(vo2p0, vi5x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
249 vo1p0 = vmlaq_lane_f32(vo1p0, vi5x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
408 const float32x4_t vi5x6789 = vextq_f32(vi5x4567, vi5x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4() local
427 vo2p0 = vmlaq_lane_f32(vo2p0, vi5x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
430 vo1p0 = vmlaq_lane_f32(vo1p0, vi5x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
573 const float32x4_t vi5x6789 = vextq_f32(vi5x5678, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4() local
590 vo2p0 = vmlaq_lane_f32(vo2p0, vi5x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
593 vo1p0 = vmlaq_lane_f32(vo1p0, vi5x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
|
| D | 5x5p2-minmax-neonfma-3x4-acc2.c | 227 const float32x4_t vi5x6789 = vextq_f32(vi5x4567, vi5x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2() local
246 vo2p1 = vfmaq_lane_f32(vo2p1, vi5x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2()
249 vo1p0 = vfmaq_lane_f32(vo1p0, vi5x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2()
411 const float32x4_t vi5x6789 = vextq_f32(vi5x4567, vi5x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2() local
430 vo2p1 = vfmaq_lane_f32(vo2p1, vi5x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2()
433 vo1p0 = vfmaq_lane_f32(vo1p0, vi5x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2()
579 const float32x4_t vi5x6789 = vextq_f32(vi5x5678, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2() local
596 vo2p1 = vfmaq_lane_f32(vo2p1, vi5x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2()
599 vo1p0 = vfmaq_lane_f32(vo1p0, vi5x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2()
|
| D | 5x5p2-minmax-wasmsimd-x86-splat-5x4.c | 298 const v128_t vi5x6789 = wasm_v32x4_shuffle(vi5x4567, vi5x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4() local
317 …vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi5x6789, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
322 …vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x6789, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
327 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x6789, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
332 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x6789, wasm_v32x4_shuffle(vwOP, vwOP, 1, 1, 1, 1))… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
552 const v128_t vi5x6789 = wasm_v32x4_shuffle(vi5x4567, vi5x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4() local
571 …vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi5x6789, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
576 …vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x6789, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
581 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x6789, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
586 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x6789, wasm_v32x4_shuffle(vwOP, vwOP, 1, 1, 1, 1))… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
[all …]
|
| D | 5x5p2-minmax-wasmsimd-arm-splat-5x4.c | 298 const v128_t vi5x6789 = wasm_v32x4_shuffle(vi5x4567, vi5x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_5x4() local
317 …vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi5x6789, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_5x4()
322 …vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x6789, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_5x4()
327 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x6789, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_5x4()
332 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x6789, wasm_v32x4_shuffle(vwOP, vwOP, 1, 1, 1, 1))… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_5x4()
552 const v128_t vi5x6789 = wasm_v32x4_shuffle(vi5x4567, vi5x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_5x4() local
571 …vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi5x6789, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_5x4()
576 …vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x6789, wasm_v32x4_shuffle(vwCDEF, vwCDEF, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_5x4()
581 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x6789, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_5x4()
586 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x6789, wasm_v32x4_shuffle(vwOP, vwOP, 1, 1, 1, 1))… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_5x4()
[all …]
|
| D | 5x5p2-minmax-sse-4x4.c | 296 const __m128 vi5x6789 = _mm_shuffle_ps(vi5x5678, vi5x89AB, _MM_SHUFFLE(1, 0, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4() local
311 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x6789, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4()
314 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi5x6789, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4()
317 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi5x6789, vk44)); 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() local
525 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x6789, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4()
528 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi5x6789, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4()
531 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi5x6789, vk44)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4()
704 const __m128 vi5x6789 = _mm_shuffle_ps(vi5x5678, vzero, _MM_SHUFFLE(1, 0, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4() local
719 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x6789, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4()
[all …]
|
| D | 5x5p2-minmax-sse-4x4-acc2.c | 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() local
311 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x6789, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2()
314 vo2p1 = _mm_add_ps(vo2p1, _mm_mul_ps(vi5x6789, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2()
317 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi5x6789, vk44)); 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() local
529 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x6789, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2()
532 vo2p1 = _mm_add_ps(vo2p1, _mm_mul_ps(vi5x6789, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2()
535 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi5x6789, vk44)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2()
712 const __m128 vi5x6789 = _mm_shuffle_ps(vi5x5678, vzero, _MM_SHUFFLE(1, 0, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2() local
727 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x6789, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4_acc2()
[all …]
|
| D | 5x5p2-minmax-sse-3x4-acc2.c | 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() local
272 vo2p1 = _mm_add_ps(vo2p1, _mm_mul_ps(vi5x6789, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2()
274 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi5x6789, vk44)); 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() local
450 vo2p1 = _mm_add_ps(vo2p1, _mm_mul_ps(vi5x6789, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2()
452 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi5x6789, vk44)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2()
597 const __m128 vi5x6789 = _mm_shuffle_ps(vi5x5678, vzero, _MM_SHUFFLE(1, 0, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2() local
611 vo2p1 = _mm_add_ps(vo2p1, _mm_mul_ps(vi5x6789, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2()
613 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi5x6789, vk44)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_3x4_acc2()
|