Home
last modified time | relevance | path

Searched refs:vi2x7BDF (Results 1 – 25 of 57) sorted by relevance

123

/external/XNNPACK/src/f32-dwconv2d-chw/gen/
D3x3s2p1-minmax-neonfma-2x4-acc2.c102 const float32x4_t vi2x7BDF = vextq_f32(vi2x1357, vi2x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4_acc2() local
110 vo1p1 = vfmaq_lane_f32(vo1p1, vi2x7BDF, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4_acc2()
115 vo0p1 = vfmaq_lane_f32(vo0p1, vi2x7BDF, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4_acc2()
173 const float32x4_t vi2x7BDF = vextq_f32(vi2x1357, vi2x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4_acc2() local
178 vo1p1 = vfmaq_lane_f32(vo1p1, vi2x7BDF, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4_acc2()
183 vo0p1 = vfmaq_lane_f32(vo0p1, vi2x7BDF, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4_acc2()
D3x3s2p1-minmax-neon-2x4.c102 const float32x4_t vi2x7BDF = vextq_f32(vi2x1357, vi2x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_2x4() local
110 vo1p0 = vmlaq_lane_f32(vo1p0, vi2x7BDF, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_2x4()
115 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x7BDF, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_2x4()
171 const float32x4_t vi2x7BDF = vextq_f32(vi2x1357, vi2x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_2x4() local
176 vo1p0 = vmlaq_lane_f32(vo1p0, vi2x7BDF, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_2x4()
181 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x7BDF, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_2x4()
D3x3s2p1-minmax-neonfma-2x4.c102 const float32x4_t vi2x7BDF = vextq_f32(vi2x1357, vi2x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4() local
110 vo1p0 = vfmaq_lane_f32(vo1p0, vi2x7BDF, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4()
115 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x7BDF, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4()
171 const float32x4_t vi2x7BDF = vextq_f32(vi2x1357, vi2x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4() local
176 vo1p0 = vfmaq_lane_f32(vo1p0, vi2x7BDF, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4()
181 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x7BDF, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4()
D3x3s2p1-minmax-neonfma-1x4.c83 const float32x4_t vi2x7BDF = vextq_f32(vi2x1357, vi2x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4() local
90 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x7BDF, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4()
129 const float32x4_t vi2x7BDF = vextq_f32(vi2x1357, vi2x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4() local
135 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x7BDF, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4()
D3x3s2p1-minmax-neonfma-1x4-acc2.c83 const float32x4_t vi2x7BDF = vextq_f32(vi2x1357, vi2x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4_acc2() local
90 vo0p1 = vfmaq_lane_f32(vo0p1, vi2x7BDF, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4_acc2()
130 const float32x4_t vi2x7BDF = vextq_f32(vi2x1357, vi2x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4_acc2() local
136 vo0p1 = vfmaq_lane_f32(vo0p1, vi2x7BDF, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4_acc2()
D3x3s2p1-minmax-neonfma-1x4-acc4.c83 const float32x4_t vi2x7BDF = vextq_f32(vi2x1357, vi2x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4_acc4() local
90 vo0p3 = vfmaq_lane_f32(vo0p3, vi2x7BDF, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4_acc4()
132 const float32x4_t vi2x7BDF = vextq_f32(vi2x1357, vi2x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4_acc4() local
138 vo0p3 = vfmaq_lane_f32(vo0p3, vi2x7BDF, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4_acc4()
D3x3s2p1-minmax-neon-1x4-acc2.c83 const float32x4_t vi2x7BDF = vextq_f32(vi2x1357, vi2x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_1x4_acc2() local
90 vo0p1 = vmlaq_lane_f32(vo0p1, vi2x7BDF, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_1x4_acc2()
130 const float32x4_t vi2x7BDF = vextq_f32(vi2x1357, vi2x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_1x4_acc2() local
136 vo0p1 = vmlaq_lane_f32(vo0p1, vi2x7BDF, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_1x4_acc2()
D3x3s2p1-minmax-neon-1x4.c83 const float32x4_t vi2x7BDF = vextq_f32(vi2x1357, vi2x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_1x4() local
90 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x7BDF, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_1x4()
129 const float32x4_t vi2x7BDF = vextq_f32(vi2x1357, vi2x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_1x4() local
135 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x7BDF, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_1x4()
D3x3s2p1-minmax-neon-1x4-acc3.c83 const float32x4_t vi2x7BDF = vextq_f32(vi2x1357, vi2x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_1x4_acc3() local
90 vo0p1 = vmlaq_lane_f32(vo0p1, vi2x7BDF, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_1x4_acc3()
131 const float32x4_t vi2x7BDF = vextq_f32(vi2x1357, vi2x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_1x4_acc3() local
137 vo0p1 = vmlaq_lane_f32(vo0p1, vi2x7BDF, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_1x4_acc3()
D3x3s2p1-minmax-neonfma-1x4-acc3.c83 const float32x4_t vi2x7BDF = vextq_f32(vi2x1357, vi2x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4_acc3() local
90 vo0p1 = vfmaq_lane_f32(vo0p1, vi2x7BDF, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4_acc3()
131 const float32x4_t vi2x7BDF = vextq_f32(vi2x1357, vi2x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4_acc3() local
137 vo0p1 = vfmaq_lane_f32(vo0p1, vi2x7BDF, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_1x4_acc3()
D3x3s2p1-minmax-neon-1x4-acc4.c83 const float32x4_t vi2x7BDF = vextq_f32(vi2x1357, vi2x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_1x4_acc4() local
90 vo0p3 = vmlaq_lane_f32(vo0p3, vi2x7BDF, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_1x4_acc4()
132 const float32x4_t vi2x7BDF = vextq_f32(vi2x1357, vi2x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_1x4_acc4() local
138 vo0p3 = vmlaq_lane_f32(vo0p3, vi2x7BDF, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_1x4_acc4()
D3x3s2p1-minmax-wasmsimd-x86-loadsplat-2x4.c136 const v128_t vi2x7BDF = wasm_v32x4_shuffle(vi2x1357, vi2x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_2x4() local
144 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x7BDF, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_2x4()
149 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x7BDF, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_2x4()
209 const v128_t vi2x7BDF = wasm_v32x4_shuffle(vi2x1357, vi2x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_2x4() local
214 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x7BDF, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_2x4()
219 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x7BDF, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_2x4()
D3x3s2p1-minmax-wasmsimd-x86-loadsplat-2x4-acc2.c136 const v128_t vi2x7BDF = wasm_v32x4_shuffle(vi2x1357, vi2x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_2x4_acc2() local
144 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi2x7BDF, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_2x4_acc2()
149 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x7BDF, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_2x4_acc2()
211 const v128_t vi2x7BDF = wasm_v32x4_shuffle(vi2x1357, vi2x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_2x4_acc2() local
216 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi2x7BDF, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_2x4_acc2()
221 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x7BDF, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_2x4_acc2()
D3x3s2p1-minmax-wasmsimd-arm-loadsplat-2x4-acc2.c136 const v128_t vi2x7BDF = wasm_v32x4_shuffle(vi2x1357, vi2x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_2x4_acc2() local
144 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi2x7BDF, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_2x4_acc2()
149 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x7BDF, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_2x4_acc2()
211 const v128_t vi2x7BDF = wasm_v32x4_shuffle(vi2x1357, vi2x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_2x4_acc2() local
216 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi2x7BDF, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_2x4_acc2()
221 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x7BDF, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_2x4_acc2()
D3x3s2p1-minmax-wasmsimd-arm-splat-2x4-acc2.c126 const v128_t vi2x7BDF = wasm_v32x4_shuffle(vi2x1357, vi2x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_2x4_acc2() local
134 …vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi2x7BDF, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_2x4_acc2()
139 …vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x7BDF, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_2x4_acc2()
201 const v128_t vi2x7BDF = wasm_v32x4_shuffle(vi2x1357, vi2x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_2x4_acc2() local
206 …vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi2x7BDF, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_2x4_acc2()
211 …vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x7BDF, 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.c126 const v128_t vi2x7BDF = wasm_v32x4_shuffle(vi2x1357, vi2x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_2x4() local
134 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x7BDF, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_2x4()
139 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x7BDF, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_2x4()
199 const v128_t vi2x7BDF = wasm_v32x4_shuffle(vi2x1357, vi2x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_2x4() local
204 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x7BDF, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_2x4()
209 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x7BDF, 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.c136 const v128_t vi2x7BDF = wasm_v32x4_shuffle(vi2x1357, vi2x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_2x4() local
144 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x7BDF, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_2x4()
149 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x7BDF, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_2x4()
209 const v128_t vi2x7BDF = wasm_v32x4_shuffle(vi2x1357, vi2x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_2x4() local
214 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x7BDF, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_2x4()
219 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x7BDF, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_2x4()
D3x3s2p1-minmax-wasmsimd-arm-loadsplat-1x4.c109 const v128_t vi2x7BDF = wasm_v32x4_shuffle(vi2x1357, vi2x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_1x4() local
116 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x7BDF, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_1x4()
157 const v128_t vi2x7BDF = wasm_v32x4_shuffle(vi2x1357, vi2x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_1x4() local
163 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x7BDF, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_1x4()
D3x3s2p1-minmax-wasmsimd-arm-loadsplat-1x4-acc2.c109 const v128_t vi2x7BDF = wasm_v32x4_shuffle(vi2x1357, vi2x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_1x4_acc2() local
116 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x7BDF, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_1x4_acc2()
158 const v128_t vi2x7BDF = wasm_v32x4_shuffle(vi2x1357, vi2x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_1x4_acc2() local
164 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x7BDF, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_1x4_acc2()
D3x3s2p1-minmax-wasmsimd-x86-splat-1x4.c99 const v128_t vi2x7BDF = wasm_v32x4_shuffle(vi2x1357, vi2x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_1x4() local
106 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x7BDF, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_1x4()
147 const v128_t vi2x7BDF = wasm_v32x4_shuffle(vi2x1357, vi2x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_1x4() local
153 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x7BDF, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_1x4()
D3x3s2p1-minmax-wasmsimd-arm-splat-1x4.c99 const v128_t vi2x7BDF = wasm_v32x4_shuffle(vi2x1357, vi2x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_1x4() local
106 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x7BDF, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_1x4()
147 const v128_t vi2x7BDF = wasm_v32x4_shuffle(vi2x1357, vi2x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_1x4() local
153 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x7BDF, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_1x4()
D3x3s2p1-minmax-wasmsimd-arm-splat-1x4-acc3.c99 const v128_t vi2x7BDF = wasm_v32x4_shuffle(vi2x1357, vi2x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_1x4_acc3() local
106 …vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x7BDF, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_1x4_acc3()
149 const v128_t vi2x7BDF = wasm_v32x4_shuffle(vi2x1357, vi2x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_1x4_acc3() local
155 …vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x7BDF, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_1x4_acc3()
D3x3s2p1-minmax-wasmsimd-arm-splat-1x4-acc2.c99 const v128_t vi2x7BDF = wasm_v32x4_shuffle(vi2x1357, vi2x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_1x4_acc2() local
106 …vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x7BDF, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_1x4_acc2()
148 const v128_t vi2x7BDF = wasm_v32x4_shuffle(vi2x1357, vi2x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_1x4_acc2() local
154 …vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x7BDF, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_1x4_acc2()
D3x3s2p1-minmax-wasmsimd-x86-splat-1x4-acc2.c99 const v128_t vi2x7BDF = wasm_v32x4_shuffle(vi2x1357, vi2x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_1x4_acc2() local
106 …vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x7BDF, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_1x4_acc2()
148 const v128_t vi2x7BDF = wasm_v32x4_shuffle(vi2x1357, vi2x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_1x4_acc2() local
154 …vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x7BDF, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_1x4_acc2()
D3x3s2p1-minmax-wasmsimd-arm-splat-1x4-acc4.c99 const v128_t vi2x7BDF = wasm_v32x4_shuffle(vi2x1357, vi2x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_1x4_acc4() local
106 …vo0p3 = wasm_f32x4_add(vo0p3, wasm_f32x4_mul(vi2x7BDF, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_1x4_acc4()
150 const v128_t vi2x7BDF = wasm_v32x4_shuffle(vi2x1357, vi2x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_1x4_acc4() local
156 …vo0p3 = wasm_f32x4_add(vo0p3, wasm_f32x4_mul(vi2x7BDF, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_1x4_acc4()

123