/external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
D | 5x5s2p2-minmax-neon-1x4-acc4.c | 135 const float32x4_t vi0x79BD = vextq_f32(vi0x1357, vi0x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc4() local 152 vo0p1 = vmlaq_lane_f32(vo0p1, vi0x79BD, vget_high_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc4() 247 const float32x4_t vi0x79BD = vextq_f32(vi0x1357, vi0x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc4() local 253 vo0p1 = vmlaq_lane_f32(vo0p1, vi0x79BD, vget_high_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc4()
|
D | 5x5s2p2-minmax-neon-1x4-acc3.c | 135 const float32x4_t vi0x79BD = vextq_f32(vi0x1357, vi0x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc3() local 152 vo0p2 = vmlaq_lane_f32(vo0p2, vi0x79BD, vget_high_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc3() 246 const float32x4_t vi0x79BD = vextq_f32(vi0x1357, vi0x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc3() local 252 vo0p2 = vmlaq_lane_f32(vo0p2, vi0x79BD, vget_high_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc3()
|
D | 5x5s2p2-minmax-neonfma-1x4-acc4.c | 135 const float32x4_t vi0x79BD = vextq_f32(vi0x1357, vi0x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc4() local 152 vo0p1 = vfmaq_lane_f32(vo0p1, vi0x79BD, vget_high_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc4() 247 const float32x4_t vi0x79BD = vextq_f32(vi0x1357, vi0x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc4() local 253 vo0p1 = vfmaq_lane_f32(vo0p1, vi0x79BD, vget_high_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc4()
|
D | 5x5s2p2-minmax-neonfma-1x4-acc2.c | 135 const float32x4_t vi0x79BD = vextq_f32(vi0x1357, vi0x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2() local 152 vo0p1 = vfmaq_lane_f32(vo0p1, vi0x79BD, vget_high_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2() 245 const float32x4_t vi0x79BD = vextq_f32(vi0x1357, vi0x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2() local 251 vo0p1 = vfmaq_lane_f32(vo0p1, vi0x79BD, vget_high_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2()
|
D | 5x5s2p2-minmax-neonfma-1x4-acc3.c | 135 const float32x4_t vi0x79BD = vextq_f32(vi0x1357, vi0x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc3() local 152 vo0p2 = vfmaq_lane_f32(vo0p2, vi0x79BD, vget_high_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc3() 246 const float32x4_t vi0x79BD = vextq_f32(vi0x1357, vi0x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc3() local 252 vo0p2 = vfmaq_lane_f32(vo0p2, vi0x79BD, vget_high_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc3()
|
D | 5x5s2p2-minmax-neon-1x4.c | 135 const float32x4_t vi0x79BD = vextq_f32(vi0x1357, vi0x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4() local 152 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x79BD, vget_high_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4() 244 const float32x4_t vi0x79BD = vextq_f32(vi0x1357, vi0x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4() local 250 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x79BD, vget_high_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
|
D | 5x5s2p2-minmax-neonfma-1x4.c | 135 const float32x4_t vi0x79BD = vextq_f32(vi0x1357, vi0x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4() local 152 vo0p0 = vfmaq_lane_f32(vo0p0, vi0x79BD, vget_high_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4() 244 const float32x4_t vi0x79BD = vextq_f32(vi0x1357, vi0x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4() local 250 vo0p0 = vfmaq_lane_f32(vo0p0, vi0x79BD, vget_high_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4()
|
D | 5x5s2p2-minmax-neon-1x4-acc2.c | 135 const float32x4_t vi0x79BD = vextq_f32(vi0x1357, vi0x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc2() local 152 vo0p1 = vmlaq_lane_f32(vo0p1, vi0x79BD, vget_high_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc2() 245 const float32x4_t vi0x79BD = vextq_f32(vi0x1357, vi0x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc2() local 251 vo0p1 = vmlaq_lane_f32(vo0p1, vi0x79BD, vget_high_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc2()
|
D | 5x5s2p2-minmax-neon-1x4-acc5.c | 135 const float32x4_t vi0x79BD = vextq_f32(vi0x1357, vi0x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc5() local 152 vo0p2 = vmlaq_lane_f32(vo0p2, vi0x79BD, vget_high_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc5() 248 const float32x4_t vi0x79BD = vextq_f32(vi0x1357, vi0x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc5() local 254 vo0p2 = vmlaq_lane_f32(vo0p2, vi0x79BD, vget_high_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc5()
|
D | 5x5s2p2-minmax-neonfma-1x4-acc5.c | 135 const float32x4_t vi0x79BD = vextq_f32(vi0x1357, vi0x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc5() local 152 vo0p2 = vfmaq_lane_f32(vo0p2, vi0x79BD, vget_high_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc5() 248 const float32x4_t vi0x79BD = vextq_f32(vi0x1357, vi0x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc5() local 254 vo0p2 = vfmaq_lane_f32(vo0p2, vi0x79BD, vget_high_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc5()
|
D | 5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc3.c | 179 const v128_t vi0x79BD = wasm_v32x4_shuffle(vi0x1357, vi0x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3() local 217 vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi0x79BD, vk01)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3() 315 const v128_t vi0x79BD = wasm_v32x4_shuffle(vi0x1357, vi0x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3() local 321 vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi0x79BD, vk01)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3()
|
D | 5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc2.c | 179 const v128_t vi0x79BD = wasm_v32x4_shuffle(vi0x1357, vi0x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2() local 217 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi0x79BD, vk01)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2() 314 const v128_t vi0x79BD = wasm_v32x4_shuffle(vi0x1357, vi0x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2() local 320 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi0x79BD, vk01)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2()
|
D | 5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4.c | 179 const v128_t vi0x79BD = wasm_v32x4_shuffle(vi0x1357, vi0x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4() local 217 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x79BD, vk01)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4() 313 const v128_t vi0x79BD = wasm_v32x4_shuffle(vi0x1357, vi0x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4() local 319 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x79BD, vk01)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4()
|
D | 5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4.c | 179 const v128_t vi0x79BD = wasm_v32x4_shuffle(vi0x1357, vi0x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4() local 217 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x79BD, vk01)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4() 313 const v128_t vi0x79BD = wasm_v32x4_shuffle(vi0x1357, vi0x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4() local 319 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x79BD, vk01)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
|
D | 5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4-acc4.c | 179 const v128_t vi0x79BD = wasm_v32x4_shuffle(vi0x1357, vi0x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc4() local 217 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi0x79BD, vk01)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc4() 316 const v128_t vi0x79BD = wasm_v32x4_shuffle(vi0x1357, vi0x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc4() local 322 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi0x79BD, vk01)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc4()
|
D | 5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc4.c | 159 const v128_t vi0x79BD = wasm_v32x4_shuffle(vi0x1357, vi0x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc4() local 197 …vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi0x79BD, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc4() 296 const v128_t vi0x79BD = wasm_v32x4_shuffle(vi0x1357, vi0x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc4() local 302 …vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi0x79BD, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc4()
|
D | 5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc5.c | 159 const v128_t vi0x79BD = wasm_v32x4_shuffle(vi0x1357, vi0x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc5() local 197 …vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi0x79BD, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc5() 297 const v128_t vi0x79BD = wasm_v32x4_shuffle(vi0x1357, vi0x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc5() local 303 …vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi0x79BD, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc5()
|
D | 5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc4.c | 179 const v128_t vi0x79BD = wasm_v32x4_shuffle(vi0x1357, vi0x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc4() local 217 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi0x79BD, vk01)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc4() 316 const v128_t vi0x79BD = wasm_v32x4_shuffle(vi0x1357, vi0x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc4() local 322 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi0x79BD, vk01)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc4()
|
D | 5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc2.c | 159 const v128_t vi0x79BD = wasm_v32x4_shuffle(vi0x1357, vi0x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc2() local 197 …vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi0x79BD, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc2() 294 const v128_t vi0x79BD = wasm_v32x4_shuffle(vi0x1357, vi0x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc2() local 300 …vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi0x79BD, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc2()
|
D | 5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc5.c | 159 const v128_t vi0x79BD = wasm_v32x4_shuffle(vi0x1357, vi0x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc5() local 197 …vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi0x79BD, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc5() 297 const v128_t vi0x79BD = wasm_v32x4_shuffle(vi0x1357, vi0x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc5() local 303 …vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi0x79BD, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc5()
|
D | 5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc2.c | 159 const v128_t vi0x79BD = wasm_v32x4_shuffle(vi0x1357, vi0x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc2() local 197 …vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi0x79BD, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc2() 294 const v128_t vi0x79BD = wasm_v32x4_shuffle(vi0x1357, vi0x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc2() local 300 …vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi0x79BD, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc2()
|
D | 5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc3.c | 159 const v128_t vi0x79BD = wasm_v32x4_shuffle(vi0x1357, vi0x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc3() local 197 …vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi0x79BD, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc3() 295 const v128_t vi0x79BD = wasm_v32x4_shuffle(vi0x1357, vi0x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc3() local 301 …vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi0x79BD, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc3()
|
D | 5x5s2p2-minmax-wasmsimd-x86-splat-1x4.c | 159 const v128_t vi0x79BD = wasm_v32x4_shuffle(vi0x1357, vi0x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4() local 197 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x79BD, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4() 293 const v128_t vi0x79BD = wasm_v32x4_shuffle(vi0x1357, vi0x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4() local 299 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x79BD, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
|
D | 5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc5.c | 179 const v128_t vi0x79BD = wasm_v32x4_shuffle(vi0x1357, vi0x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc5() local 217 vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi0x79BD, vk01)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc5() 317 const v128_t vi0x79BD = wasm_v32x4_shuffle(vi0x1357, vi0x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc5() local 323 vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi0x79BD, vk01)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc5()
|
D | 5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc4.c | 159 const v128_t vi0x79BD = wasm_v32x4_shuffle(vi0x1357, vi0x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc4() local 197 …vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi0x79BD, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc4() 296 const v128_t vi0x79BD = wasm_v32x4_shuffle(vi0x1357, vi0x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc4() local 302 …vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi0x79BD, wasm_v32x4_shuffle(vw0123, vw0123, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc4()
|