Home
last modified time | relevance | path

Searched refs:vi3x7BDF (Results 1 – 25 of 29) sorted by relevance

12

/external/XNNPACK/src/f32-dwconv2d-chw/gen/
D3x3s2p1-minmax-neonfma-2x4-acc2.c104 const float32x4_t vi3x7BDF = vextq_f32(vi3x1357, vi3x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4_acc2() local
113 vo1p0 = vfmaq_lane_f32(vo1p0, vi3x7BDF, vget_low_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4_acc2()
174 const float32x4_t vi3x7BDF = vextq_f32(vi3x1357, vi3x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4_acc2() local
181 vo1p0 = vfmaq_lane_f32(vo1p0, vi3x7BDF, vget_low_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4_acc2()
D3x3s2p1-minmax-neon-2x4.c104 const float32x4_t vi3x7BDF = vextq_f32(vi3x1357, vi3x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_2x4() local
113 vo1p0 = vmlaq_lane_f32(vo1p0, vi3x7BDF, vget_low_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_2x4()
172 const float32x4_t vi3x7BDF = vextq_f32(vi3x1357, vi3x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_2x4() local
179 vo1p0 = vmlaq_lane_f32(vo1p0, vi3x7BDF, vget_low_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_2x4()
D3x3s2p1-minmax-neonfma-2x4.c104 const float32x4_t vi3x7BDF = vextq_f32(vi3x1357, vi3x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4() local
113 vo1p0 = vfmaq_lane_f32(vo1p0, vi3x7BDF, vget_low_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4()
172 const float32x4_t vi3x7BDF = vextq_f32(vi3x1357, vi3x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4() local
179 vo1p0 = vfmaq_lane_f32(vo1p0, vi3x7BDF, vget_low_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4()
D3x3s2p1-minmax-wasmsimd-x86-loadsplat-2x4.c138 const v128_t vi3x7BDF = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_2x4() local
147 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x7BDF, vk10)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_2x4()
210 const v128_t vi3x7BDF = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_2x4() local
217 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x7BDF, vk10)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_2x4()
D3x3s2p1-minmax-wasmsimd-x86-loadsplat-2x4-acc2.c138 const v128_t vi3x7BDF = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_2x4_acc2() local
147 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x7BDF, vk10)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_2x4_acc2()
212 const v128_t vi3x7BDF = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_2x4_acc2() local
219 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x7BDF, vk10)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_2x4_acc2()
D3x3s2p1-minmax-wasmsimd-arm-loadsplat-2x4-acc2.c138 const v128_t vi3x7BDF = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_2x4_acc2() local
147 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x7BDF, vk10)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_2x4_acc2()
212 const v128_t vi3x7BDF = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_2x4_acc2() local
219 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x7BDF, vk10)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_2x4_acc2()
D3x3s2p1-minmax-wasmsimd-arm-splat-2x4-acc2.c128 const v128_t vi3x7BDF = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_2x4_acc2() local
137 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x7BDF, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_2x4_acc2()
202 const v128_t vi3x7BDF = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_2x4_acc2() local
209 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x7BDF, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_2x4_acc2()
D3x3s2p1-minmax-wasmsimd-arm-splat-2x4.c128 const v128_t vi3x7BDF = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_2x4() local
137 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x7BDF, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_2x4()
200 const v128_t vi3x7BDF = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_2x4() local
207 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x7BDF, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_2x4()
D3x3s2p1-minmax-wasmsimd-arm-loadsplat-2x4.c138 const v128_t vi3x7BDF = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_2x4() local
147 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x7BDF, vk10)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_2x4()
210 const v128_t vi3x7BDF = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_2x4() local
217 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x7BDF, vk10)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_2x4()
D3x3s2p1-minmax-sse-2x4-acc2.c137 const __m128 vi3x7BDF = _mm_move_ss(vi3xF9BD, vi3x7531); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_2x4_acc2() local
149 vo1p1 = _mm_add_ps(vo1p1, _mm_mul_ps(vi3x7BDF, vk10)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_2x4_acc2()
215 const __m128 vi3x7BDF = _mm_move_ss(vi3xF9BD, vi3x7531); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_2x4_acc2() local
227 vo1p1 = _mm_add_ps(vo1p1, _mm_mul_ps(vi3x7BDF, vk10)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_2x4_acc2()
D3x3s2p1-minmax-sse-2x4.c137 const __m128 vi3x7BDF = _mm_move_ss(vi3xF9BD, vi3x7531); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_2x4() local
149 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x7BDF, vk10)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_2x4()
213 const __m128 vi3x7BDF = _mm_move_ss(vi3xF9BD, vi3x7531); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_2x4() local
225 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x7BDF, vk10)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_2x4()
D3x3s2p1-minmax-wasmsimd-x86-splat-2x4.c128 const v128_t vi3x7BDF = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_2x4() local
137 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x7BDF, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_2x4()
200 const v128_t vi3x7BDF = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_2x4() local
207 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x7BDF, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_2x4()
D3x3s2p1-minmax-wasmsimd-x86-splat-2x4-acc2.c128 const v128_t vi3x7BDF = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_2x4_acc2() local
137 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x7BDF, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_2x4_acc2()
202 const v128_t vi3x7BDF = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_2x4_acc2() local
209 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x7BDF, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_2x4_acc2()
D3x3s2p1-minmax-neonfma-3x4.c122 const float32x4_t vi3x7BDF = vextq_f32(vi3x1357, vi3x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_3x4() local
136 vo1p0 = vfmaq_lane_f32(vo1p0, vi3x7BDF, vget_low_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_3x4()
213 const float32x4_t vi3x7BDF = vextq_f32(vi3x1357, vi3x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_3x4() local
223 vo1p0 = vfmaq_lane_f32(vo1p0, vi3x7BDF, vget_low_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_3x4()
D3x3s2p1-minmax-neon-3x4.c122 const float32x4_t vi3x7BDF = vextq_f32(vi3x1357, vi3x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_3x4() local
136 vo1p0 = vmlaq_lane_f32(vo1p0, vi3x7BDF, vget_low_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_3x4()
213 const float32x4_t vi3x7BDF = vextq_f32(vi3x1357, vi3x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_3x4() local
223 vo1p0 = vmlaq_lane_f32(vo1p0, vi3x7BDF, vget_low_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_3x4()
D3x3s2p1-minmax-wasmsimd-x86-loadsplat-3x4.c164 const v128_t vi3x7BDF = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_3x4() local
178 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x7BDF, vk10)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_3x4()
261 const v128_t vi3x7BDF = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_3x4() local
271 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x7BDF, vk10)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_3x4()
D3x3s2p1-minmax-wasmsimd-arm-loadsplat-3x4.c164 const v128_t vi3x7BDF = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_3x4() local
178 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x7BDF, vk10)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_3x4()
261 const v128_t vi3x7BDF = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_3x4() local
271 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x7BDF, vk10)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_3x4()
D3x3s2p1-minmax-wasmsimd-arm-splat-3x4.c154 const v128_t vi3x7BDF = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_3x4() local
168 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x7BDF, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_3x4()
251 const v128_t vi3x7BDF = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_3x4() local
261 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x7BDF, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_3x4()
D3x3s2p1-minmax-wasmsimd-x86-splat-3x4.c154 const v128_t vi3x7BDF = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_3x4() local
168 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x7BDF, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_3x4()
251 const v128_t vi3x7BDF = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_3x4() local
261 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x7BDF, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_3x4()
D3x3s2p1-minmax-neonfma-4x4.c140 const float32x4_t vi3x7BDF = vextq_f32(vi3x1357, vi3x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_4x4() local
159 vo1p0 = vfmaq_lane_f32(vo1p0, vi3x7BDF, vget_low_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_4x4()
254 const float32x4_t vi3x7BDF = vextq_f32(vi3x1357, vi3x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_4x4() local
267 vo1p0 = vfmaq_lane_f32(vo1p0, vi3x7BDF, vget_low_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_4x4()
D3x3s2p1-minmax-neon-4x4.c140 const float32x4_t vi3x7BDF = vextq_f32(vi3x1357, vi3x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_4x4() local
159 vo1p0 = vmlaq_lane_f32(vo1p0, vi3x7BDF, vget_low_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_4x4()
254 const float32x4_t vi3x7BDF = vextq_f32(vi3x1357, vi3x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_4x4() local
267 vo1p0 = vmlaq_lane_f32(vo1p0, vi3x7BDF, vget_low_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_4x4()
D3x3s2p1-minmax-sse-3x4.c167 const __m128 vi3x7BDF = _mm_move_ss(vi3xF9BD, vi3x7531); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_3x4() local
184 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x7BDF, vk10)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_3x4()
270 const __m128 vi3x7BDF = _mm_move_ss(vi3xF9BD, vi3x7531); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_3x4() local
287 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x7BDF, vk10)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_3x4()
D3x3s2p1-minmax-wasmsimd-arm-loadsplat-4x4.c190 const v128_t vi3x7BDF = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_4x4() local
209 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x7BDF, vk10)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_4x4()
312 const v128_t vi3x7BDF = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_4x4() local
325 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x7BDF, vk10)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_4x4()
D3x3s2p1-minmax-wasmsimd-x86-loadsplat-4x4.c190 const v128_t vi3x7BDF = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_4x4() local
209 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x7BDF, vk10)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_4x4()
312 const v128_t vi3x7BDF = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_4x4() local
325 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x7BDF, vk10)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_4x4()
D3x3s2p1-minmax-wasmsimd-arm-splat-4x4.c180 const v128_t vi3x7BDF = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_4x4() local
199 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x7BDF, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_4x4()
302 const v128_t vi3x7BDF = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_4x4() local
315 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x7BDF, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_4x4()

12