Home
last modified time | relevance | path

Searched refs:vi4x7BDF (Results 1 – 25 of 29) sorted by relevance

12

/external/XNNPACK/src/f32-dwconv2d-chw/gen/
D3x3s2p1-minmax-neonfma-3x4.c124 const float32x4_t vi4x7BDF = vextq_f32(vi4x1357, vi4x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_3x4() local
133 vo2p0 = vfmaq_lane_f32(vo2p0, vi4x7BDF, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_3x4()
140 vo1p0 = vfmaq_lane_f32(vo1p0, vi4x7BDF, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_3x4()
214 const float32x4_t vi4x7BDF = vextq_f32(vi4x1357, vi4x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_3x4() local
220 vo2p0 = vfmaq_lane_f32(vo2p0, vi4x7BDF, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_3x4()
227 vo1p0 = vfmaq_lane_f32(vo1p0, vi4x7BDF, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_3x4()
D3x3s2p1-minmax-neon-3x4.c124 const float32x4_t vi4x7BDF = vextq_f32(vi4x1357, vi4x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_3x4() local
133 vo2p0 = vmlaq_lane_f32(vo2p0, vi4x7BDF, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_3x4()
140 vo1p0 = vmlaq_lane_f32(vo1p0, vi4x7BDF, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_3x4()
214 const float32x4_t vi4x7BDF = vextq_f32(vi4x1357, vi4x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_3x4() local
220 vo2p0 = vmlaq_lane_f32(vo2p0, vi4x7BDF, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_3x4()
227 vo1p0 = vmlaq_lane_f32(vo1p0, vi4x7BDF, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_3x4()
D3x3s2p1-minmax-wasmsimd-x86-loadsplat-3x4.c166 const v128_t vi4x7BDF = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_3x4() local
175 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x7BDF, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_3x4()
182 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x7BDF, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_3x4()
262 const v128_t vi4x7BDF = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_3x4() local
268 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x7BDF, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_3x4()
275 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x7BDF, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_3x4()
D3x3s2p1-minmax-wasmsimd-arm-loadsplat-3x4.c166 const v128_t vi4x7BDF = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_3x4() local
175 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x7BDF, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_3x4()
182 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x7BDF, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_3x4()
262 const v128_t vi4x7BDF = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_3x4() local
268 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x7BDF, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_3x4()
275 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x7BDF, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_3x4()
D3x3s2p1-minmax-wasmsimd-arm-splat-3x4.c156 const v128_t vi4x7BDF = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_3x4() local
165 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x7BDF, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_3x4()
172 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x7BDF, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_3x4()
252 const v128_t vi4x7BDF = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_3x4() local
258 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x7BDF, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_3x4()
265 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x7BDF, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_3x4()
D3x3s2p1-minmax-neonfma-2x4-acc2.c106 const float32x4_t vi4x7BDF = vextq_f32(vi4x1357, vi4x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4_acc2() local
116 vo1p1 = vfmaq_lane_f32(vo1p1, vi4x7BDF, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4_acc2()
175 const float32x4_t vi4x7BDF = vextq_f32(vi4x1357, vi4x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4_acc2() local
184 vo1p1 = vfmaq_lane_f32(vo1p1, vi4x7BDF, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4_acc2()
D3x3s2p1-minmax-neon-2x4.c106 const float32x4_t vi4x7BDF = vextq_f32(vi4x1357, vi4x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_2x4() local
116 vo1p0 = vmlaq_lane_f32(vo1p0, vi4x7BDF, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_2x4()
173 const float32x4_t vi4x7BDF = vextq_f32(vi4x1357, vi4x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_2x4() local
182 vo1p0 = vmlaq_lane_f32(vo1p0, vi4x7BDF, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_2x4()
D3x3s2p1-minmax-neonfma-2x4.c106 const float32x4_t vi4x7BDF = vextq_f32(vi4x1357, vi4x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4() local
116 vo1p0 = vfmaq_lane_f32(vo1p0, vi4x7BDF, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4()
173 const float32x4_t vi4x7BDF = vextq_f32(vi4x1357, vi4x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4() local
182 vo1p0 = vfmaq_lane_f32(vo1p0, vi4x7BDF, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4()
D3x3s2p1-minmax-wasmsimd-x86-splat-3x4.c156 const v128_t vi4x7BDF = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_3x4() local
165 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x7BDF, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_3x4()
172 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x7BDF, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_3x4()
252 const v128_t vi4x7BDF = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_3x4() local
258 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x7BDF, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_3x4()
265 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x7BDF, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_3x4()
D3x3s2p1-minmax-neonfma-4x4.c142 const float32x4_t vi4x7BDF = vextq_f32(vi4x1357, vi4x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_4x4() local
155 vo2p0 = vfmaq_lane_f32(vo2p0, vi4x7BDF, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_4x4()
164 vo1p0 = vfmaq_lane_f32(vo1p0, vi4x7BDF, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_4x4()
255 const float32x4_t vi4x7BDF = vextq_f32(vi4x1357, vi4x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_4x4() local
263 vo2p0 = vfmaq_lane_f32(vo2p0, vi4x7BDF, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_4x4()
272 vo1p0 = vfmaq_lane_f32(vo1p0, vi4x7BDF, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_4x4()
D3x3s2p1-minmax-neon-4x4.c142 const float32x4_t vi4x7BDF = vextq_f32(vi4x1357, vi4x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_4x4() local
155 vo2p0 = vmlaq_lane_f32(vo2p0, vi4x7BDF, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_4x4()
164 vo1p0 = vmlaq_lane_f32(vo1p0, vi4x7BDF, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_4x4()
255 const float32x4_t vi4x7BDF = vextq_f32(vi4x1357, vi4x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_4x4() local
263 vo2p0 = vmlaq_lane_f32(vo2p0, vi4x7BDF, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_4x4()
272 vo1p0 = vmlaq_lane_f32(vo1p0, vi4x7BDF, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_4x4()
D3x3s2p1-minmax-sse-3x4.c168 const __m128 vi4x7BDF = _mm_move_ss(vi4xF9BD, vi4x7531); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_3x4() local
182 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi4x7BDF, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_3x4()
187 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi4x7BDF, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_3x4()
271 const __m128 vi4x7BDF = _mm_move_ss(vi4xF9BD, vi4x7531); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_3x4() local
285 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi4x7BDF, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_3x4()
290 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi4x7BDF, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_3x4()
D3x3s2p1-minmax-wasmsimd-x86-loadsplat-2x4.c140 const v128_t vi4x7BDF = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_2x4() local
150 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x7BDF, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_2x4()
211 const v128_t vi4x7BDF = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_2x4() local
220 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x7BDF, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_2x4()
D3x3s2p1-minmax-wasmsimd-x86-loadsplat-2x4-acc2.c140 const v128_t vi4x7BDF = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_2x4_acc2() local
150 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi4x7BDF, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_2x4_acc2()
213 const v128_t vi4x7BDF = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_2x4_acc2() local
222 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi4x7BDF, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_2x4_acc2()
D3x3s2p1-minmax-wasmsimd-arm-loadsplat-2x4-acc2.c140 const v128_t vi4x7BDF = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_2x4_acc2() local
150 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi4x7BDF, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_2x4_acc2()
213 const v128_t vi4x7BDF = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_2x4_acc2() local
222 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi4x7BDF, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_2x4_acc2()
D3x3s2p1-minmax-wasmsimd-arm-splat-2x4-acc2.c130 const v128_t vi4x7BDF = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_2x4_acc2() local
140 …vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi4x7BDF, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_2x4_acc2()
203 const v128_t vi4x7BDF = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_2x4_acc2() local
212 …vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi4x7BDF, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_2x4_acc2()
D3x3s2p1-minmax-wasmsimd-arm-splat-2x4.c130 const v128_t vi4x7BDF = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_2x4() local
140 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x7BDF, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_2x4()
201 const v128_t vi4x7BDF = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_2x4() local
210 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x7BDF, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_2x4()
D3x3s2p1-minmax-wasmsimd-arm-loadsplat-2x4.c140 const v128_t vi4x7BDF = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_2x4() local
150 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x7BDF, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_2x4()
211 const v128_t vi4x7BDF = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_2x4() local
220 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x7BDF, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_2x4()
D3x3s2p1-minmax-wasmsimd-arm-loadsplat-4x4.c192 const v128_t vi4x7BDF = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_4x4() local
205 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x7BDF, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_4x4()
214 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x7BDF, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_4x4()
313 const v128_t vi4x7BDF = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_4x4() local
321 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x7BDF, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_4x4()
330 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x7BDF, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_4x4()
D3x3s2p1-minmax-sse-2x4-acc2.c138 const __m128 vi4x7BDF = _mm_move_ss(vi4xF9BD, vi4x7531); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_2x4_acc2() local
151 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi4x7BDF, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_2x4_acc2()
216 const __m128 vi4x7BDF = _mm_move_ss(vi4xF9BD, vi4x7531); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_2x4_acc2() local
229 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi4x7BDF, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_2x4_acc2()
D3x3s2p1-minmax-sse-2x4.c138 const __m128 vi4x7BDF = _mm_move_ss(vi4xF9BD, vi4x7531); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_2x4() local
151 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi4x7BDF, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_2x4()
214 const __m128 vi4x7BDF = _mm_move_ss(vi4xF9BD, vi4x7531); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_2x4() local
227 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi4x7BDF, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_2x4()
D3x3s2p1-minmax-wasmsimd-x86-splat-2x4.c130 const v128_t vi4x7BDF = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_2x4() local
140 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x7BDF, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_2x4()
201 const v128_t vi4x7BDF = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_2x4() local
210 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x7BDF, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_2x4()
D3x3s2p1-minmax-wasmsimd-x86-splat-2x4-acc2.c130 const v128_t vi4x7BDF = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_2x4_acc2() local
140 …vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi4x7BDF, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_2x4_acc2()
203 const v128_t vi4x7BDF = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_2x4_acc2() local
212 …vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi4x7BDF, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_2x4_acc2()
D3x3s2p1-minmax-wasmsimd-x86-loadsplat-4x4.c192 const v128_t vi4x7BDF = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_4x4() local
205 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x7BDF, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_4x4()
214 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x7BDF, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_4x4()
313 const v128_t vi4x7BDF = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_4x4() local
321 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x7BDF, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_4x4()
330 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x7BDF, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_4x4()
D3x3s2p1-minmax-wasmsimd-arm-splat-4x4.c182 const v128_t vi4x7BDF = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_4x4() local
195 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x7BDF, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_4x4()
204 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x7BDF, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_4x4()
303 const v128_t vi4x7BDF = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_4x4() local
311 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x7BDF, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_4x4()
320 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x7BDF, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_4x4()

12