Home
last modified time | relevance | path

Searched refs:vi5xACEG (Results 1 – 25 of 35) sorted by relevance

12

/external/XNNPACK/src/f32-dwconv2d-chw/gen/
D5x5s2p2-minmax-neonfma-3x4-acc2.c267 const float32x4_t vi5xACEG = vextq_f32(vi5x8ACE9BDF.val[0], vi5xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_3x4_acc2() local
282 vo2p1 = vfmaq_lane_f32(vo2p1, vi5xACEG, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_3x4_acc2()
289 vo1p1 = vfmaq_lane_f32(vo1p1, vi5xACEG, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_3x4_acc2()
446 const float32x4_t vi5xACEG = vextq_f32(vi5x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_3x4_acc2() local
457 vo2p1 = vfmaq_lane_f32(vo2p1, vi5xACEG, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_3x4_acc2()
464 vo1p1 = vfmaq_lane_f32(vo1p1, vi5xACEG, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_3x4_acc2()
D5x5s2p2-minmax-neon-3x4.c267 const float32x4_t vi5xACEG = vextq_f32(vi5x8ACE9BDF.val[0], vi5xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_3x4() local
282 vo2p0 = vmlaq_lane_f32(vo2p0, vi5xACEG, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_3x4()
289 vo1p0 = vmlaq_lane_f32(vo1p0, vi5xACEG, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_3x4()
443 const float32x4_t vi5xACEG = vextq_f32(vi5x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_3x4() local
454 vo2p0 = vmlaq_lane_f32(vo2p0, vi5xACEG, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_3x4()
461 vo1p0 = vmlaq_lane_f32(vo1p0, vi5xACEG, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_3x4()
D5x5s2p2-minmax-neonfma-3x4.c267 const float32x4_t vi5xACEG = vextq_f32(vi5x8ACE9BDF.val[0], vi5xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_3x4() local
282 vo2p0 = vfmaq_lane_f32(vo2p0, vi5xACEG, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_3x4()
289 vo1p0 = vfmaq_lane_f32(vo1p0, vi5xACEG, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_3x4()
443 const float32x4_t vi5xACEG = vextq_f32(vi5x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_3x4() local
454 vo2p0 = vfmaq_lane_f32(vo2p0, vi5xACEG, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_3x4()
461 vo1p0 = vfmaq_lane_f32(vo1p0, vi5xACEG, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_3x4()
D5x5s2p2-minmax-neon-3x4-acc2.c267 const float32x4_t vi5xACEG = vextq_f32(vi5x8ACE9BDF.val[0], vi5xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_3x4_acc2() local
282 vo2p1 = vmlaq_lane_f32(vo2p1, vi5xACEG, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_3x4_acc2()
289 vo1p1 = vmlaq_lane_f32(vo1p1, vi5xACEG, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_3x4_acc2()
446 const float32x4_t vi5xACEG = vextq_f32(vi5x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_3x4_acc2() local
457 vo2p1 = vmlaq_lane_f32(vo2p1, vi5xACEG, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_3x4_acc2()
464 vo1p1 = vmlaq_lane_f32(vo1p1, vi5xACEG, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_3x4_acc2()
D5x5s2p2-minmax-wasmsimd-x86-loadsplat-3x4-acc2.c369 const v128_t vi5xACEG = wasm_v32x4_shuffle(vi5x8ACE, vi5xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_3x4_acc2() local
388 vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi5xACEG, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_3x4_acc2()
395 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi5xACEG, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_3x4_acc2()
550 const v128_t vi5xACEG = wasm_v32x4_shuffle(vi5x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_3x4_acc2() local
561 vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi5xACEG, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_3x4_acc2()
568 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi5xACEG, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_3x4_acc2()
D5x5s2p2-minmax-wasmsimd-x86-loadsplat-3x4.c369 const v128_t vi5xACEG = wasm_v32x4_shuffle(vi5x8ACE, vi5xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_3x4() local
388 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5xACEG, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_3x4()
395 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5xACEG, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_3x4()
547 const v128_t vi5xACEG = wasm_v32x4_shuffle(vi5x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_3x4() local
558 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5xACEG, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_3x4()
565 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5xACEG, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_3x4()
D5x5s2p2-minmax-wasmsimd-arm-loadsplat-3x4-acc2.c369 const v128_t vi5xACEG = wasm_v32x4_shuffle(vi5x8ACE, vi5xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_3x4_acc2() local
388 vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi5xACEG, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_3x4_acc2()
395 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi5xACEG, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_3x4_acc2()
550 const v128_t vi5xACEG = wasm_v32x4_shuffle(vi5x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_3x4_acc2() local
561 vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi5xACEG, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_3x4_acc2()
568 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi5xACEG, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_3x4_acc2()
D5x5s2p2-minmax-wasmsimd-arm-loadsplat-3x4.c369 const v128_t vi5xACEG = wasm_v32x4_shuffle(vi5x8ACE, vi5xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_3x4() local
388 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5xACEG, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_3x4()
395 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5xACEG, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_3x4()
547 const v128_t vi5xACEG = wasm_v32x4_shuffle(vi5x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_3x4() local
558 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5xACEG, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_3x4()
565 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5xACEG, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_3x4()
D5x5s2p2-minmax-neon-2x4-acc2.c220 const float32x4_t vi5xACEG = vextq_f32(vi5x8ACE9BDF.val[0], vi5xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc2() local
235 vo1p1 = vmlaq_lane_f32(vo1p1, vi5xACEG, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc2()
357 const float32x4_t vi5xACEG = vextq_f32(vi5x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc2() local
370 vo1p1 = vmlaq_lane_f32(vo1p1, vi5xACEG, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc2()
D5x5s2p2-minmax-neonfma-2x4-acc2.c220 const float32x4_t vi5xACEG = vextq_f32(vi5x8ACE9BDF.val[0], vi5xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc2() local
235 vo1p1 = vfmaq_lane_f32(vo1p1, vi5xACEG, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc2()
357 const float32x4_t vi5xACEG = vextq_f32(vi5x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc2() local
370 vo1p1 = vfmaq_lane_f32(vo1p1, vi5xACEG, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc2()
D5x5s2p2-minmax-neon-2x4.c220 const float32x4_t vi5xACEG = vextq_f32(vi5x8ACE9BDF.val[0], vi5xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4() local
235 vo1p0 = vmlaq_lane_f32(vo1p0, vi5xACEG, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4()
355 const float32x4_t vi5xACEG = vextq_f32(vi5x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4() local
368 vo1p0 = vmlaq_lane_f32(vo1p0, vi5xACEG, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4()
D5x5s2p2-minmax-neon-2x4-acc3.c220 const float32x4_t vi5xACEG = vextq_f32(vi5x8ACE9BDF.val[0], vi5xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc3() local
235 vo1p1 = vmlaq_lane_f32(vo1p1, vi5xACEG, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc3()
359 const float32x4_t vi5xACEG = vextq_f32(vi5x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc3() local
372 vo1p1 = vmlaq_lane_f32(vo1p1, vi5xACEG, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc3()
D5x5s2p2-minmax-neonfma-2x4-acc3.c220 const float32x4_t vi5xACEG = vextq_f32(vi5x8ACE9BDF.val[0], vi5xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc3() local
235 vo1p1 = vfmaq_lane_f32(vo1p1, vi5xACEG, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc3()
359 const float32x4_t vi5xACEG = vextq_f32(vi5x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc3() local
372 vo1p1 = vfmaq_lane_f32(vo1p1, vi5xACEG, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc3()
D5x5s2p2-minmax-neonfma-2x4.c220 const float32x4_t vi5xACEG = vextq_f32(vi5x8ACE9BDF.val[0], vi5xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4() local
235 vo1p0 = vfmaq_lane_f32(vo1p0, vi5xACEG, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4()
355 const float32x4_t vi5xACEG = vextq_f32(vi5x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4() local
368 vo1p0 = vfmaq_lane_f32(vo1p0, vi5xACEG, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4()
D5x5s2p2-minmax-sse-3x4.c383 const __m128 vi5xACEG = _mm_shuffle_ps(vi5xGACE, vi5xGACE, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_3x4() local
393 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi5xACEG, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_3x4()
398 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi5xACEG, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_3x4()
563 const __m128 vi5xACEG = _mm_shuffle_ps(vi5xGACE, vi5xGACE, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_3x4() local
573 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi5xACEG, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_3x4()
578 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi5xACEG, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_3x4()
D5x5s2p2-minmax-sse-3x4-acc2.c383 const __m128 vi5xACEG = _mm_shuffle_ps(vi5xGACE, vi5xGACE, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_3x4_acc2() local
393 vo2p1 = _mm_add_ps(vo2p1, _mm_mul_ps(vi5xACEG, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_3x4_acc2()
398 vo1p1 = _mm_add_ps(vo1p1, _mm_mul_ps(vi5xACEG, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_3x4_acc2()
566 const __m128 vi5xACEG = _mm_shuffle_ps(vi5xGACE, vi5xGACE, _MM_SHUFFLE(0, 3, 2, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_3x4_acc2() local
576 vo2p1 = _mm_add_ps(vo2p1, _mm_mul_ps(vi5xACEG, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_3x4_acc2()
581 vo1p1 = _mm_add_ps(vo1p1, _mm_mul_ps(vi5xACEG, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_3x4_acc2()
D5x5s2p2-minmax-wasmsimd-x86-splat-3x4-acc2.c349 const v128_t vi5xACEG = wasm_v32x4_shuffle(vi5x8ACE, vi5xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4_acc2() local
368 …vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi5xACEG, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4_acc2()
375 …vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi5xACEG, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4_acc2()
530 const v128_t vi5xACEG = wasm_v32x4_shuffle(vi5x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4_acc2() local
541 …vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi5xACEG, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4_acc2()
548 …vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi5xACEG, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4_acc2()
D5x5s2p2-minmax-wasmsimd-arm-splat-3x4-acc2.c349 const v128_t vi5xACEG = wasm_v32x4_shuffle(vi5x8ACE, vi5xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_3x4_acc2() local
368 …vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi5xACEG, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_3x4_acc2()
375 …vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi5xACEG, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_3x4_acc2()
530 const v128_t vi5xACEG = wasm_v32x4_shuffle(vi5x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_3x4_acc2() local
541 …vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi5xACEG, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_3x4_acc2()
548 …vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi5xACEG, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_3x4_acc2()
D5x5s2p2-minmax-wasmsimd-x86-splat-3x4.c349 const v128_t vi5xACEG = wasm_v32x4_shuffle(vi5x8ACE, vi5xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4() local
368 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5xACEG, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4()
375 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5xACEG, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4()
527 const v128_t vi5xACEG = wasm_v32x4_shuffle(vi5x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4() local
538 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5xACEG, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4()
545 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5xACEG, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4()
D5x5s2p2-minmax-wasmsimd-arm-splat-3x4.c349 const v128_t vi5xACEG = wasm_v32x4_shuffle(vi5x8ACE, vi5xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_3x4() local
368 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5xACEG, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_3x4()
375 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5xACEG, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_3x4()
527 const v128_t vi5xACEG = wasm_v32x4_shuffle(vi5x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_3x4() local
538 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5xACEG, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_3x4()
545 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5xACEG, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_3x4()
D5x5s2p2-minmax-wasmsimd-arm-loadsplat-2x4.c306 const v128_t vi5xACEG = wasm_v32x4_shuffle(vi5x8ACE, vi5xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4() local
323 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5xACEG, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4()
441 const v128_t vi5xACEG = wasm_v32x4_shuffle(vi5x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4() local
454 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5xACEG, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4()
D5x5s2p2-minmax-wasmsimd-x86-loadsplat-2x4-acc2.c306 const v128_t vi5xACEG = wasm_v32x4_shuffle(vi5x8ACE, vi5xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc2() local
323 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi5xACEG, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc2()
443 const v128_t vi5xACEG = wasm_v32x4_shuffle(vi5x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc2() local
456 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi5xACEG, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc2()
D5x5s2p2-minmax-wasmsimd-arm-loadsplat-2x4-acc2.c306 const v128_t vi5xACEG = wasm_v32x4_shuffle(vi5x8ACE, vi5xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc2() local
323 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi5xACEG, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc2()
443 const v128_t vi5xACEG = wasm_v32x4_shuffle(vi5x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc2() local
456 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi5xACEG, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc2()
D5x5s2p2-minmax-wasmsimd-x86-loadsplat-2x4.c306 const v128_t vi5xACEG = wasm_v32x4_shuffle(vi5x8ACE, vi5xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4() local
323 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5xACEG, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4()
441 const v128_t vi5xACEG = wasm_v32x4_shuffle(vi5x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4() local
454 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5xACEG, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4()
D5x5s2p2-minmax-wasmsimd-arm-loadsplat-2x4-acc3.c306 const v128_t vi5xACEG = wasm_v32x4_shuffle(vi5x8ACE, vi5xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc3() local
323 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi5xACEG, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc3()
445 const v128_t vi5xACEG = wasm_v32x4_shuffle(vi5x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc3() local
458 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi5xACEG, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc3()

12