Home
last modified time | relevance | path

Searched refs:vi3x68AC (Results 1 – 25 of 70) sorted by relevance

123

/external/XNNPACK/src/f32-dwconv2d-chw/gen/
D5x5s2p2-minmax-neon-1x4-acc4.c120 const float32x4_t vi3x68AC = vextq_f32(vi3x0246, vi3x8ACE9BDF.val[0], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc4() local
131 vo0p3 = vmlaq_lane_f32(vo0p3, vi3x68AC, vget_low_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc4()
234 const float32x4_t vi3x68AC = vextq_f32(vi3x0246, vi3x8ACE, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc4() local
243 vo0p3 = vmlaq_lane_f32(vo0p3, vi3x68AC, vget_low_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc4()
D5x5s2p2-minmax-neon-1x4-acc3.c120 const float32x4_t vi3x68AC = vextq_f32(vi3x0246, vi3x8ACE9BDF.val[0], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc3() local
131 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x68AC, vget_low_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc3()
233 const float32x4_t vi3x68AC = vextq_f32(vi3x0246, vi3x8ACE, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc3() local
242 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x68AC, vget_low_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc3()
D5x5s2p2-minmax-neonfma-1x4-acc4.c120 const float32x4_t vi3x68AC = vextq_f32(vi3x0246, vi3x8ACE9BDF.val[0], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc4() local
131 vo0p3 = vfmaq_lane_f32(vo0p3, vi3x68AC, vget_low_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc4()
234 const float32x4_t vi3x68AC = vextq_f32(vi3x0246, vi3x8ACE, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc4() local
243 vo0p3 = vfmaq_lane_f32(vo0p3, vi3x68AC, vget_low_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc4()
D5x5s2p2-minmax-neonfma-1x4-acc2.c120 const float32x4_t vi3x68AC = vextq_f32(vi3x0246, vi3x8ACE9BDF.val[0], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2() local
131 vo0p1 = vfmaq_lane_f32(vo0p1, vi3x68AC, vget_low_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2()
232 const float32x4_t vi3x68AC = vextq_f32(vi3x0246, vi3x8ACE, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2() local
241 vo0p1 = vfmaq_lane_f32(vo0p1, vi3x68AC, vget_low_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2()
D5x5s2p2-minmax-neonfma-1x4-acc3.c120 const float32x4_t vi3x68AC = vextq_f32(vi3x0246, vi3x8ACE9BDF.val[0], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc3() local
131 vo0p0 = vfmaq_lane_f32(vo0p0, vi3x68AC, vget_low_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc3()
233 const float32x4_t vi3x68AC = vextq_f32(vi3x0246, vi3x8ACE, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc3() local
242 vo0p0 = vfmaq_lane_f32(vo0p0, vi3x68AC, vget_low_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc3()
D5x5s2p2-minmax-neon-1x4.c120 const float32x4_t vi3x68AC = vextq_f32(vi3x0246, vi3x8ACE9BDF.val[0], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4() local
131 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x68AC, vget_low_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
231 const float32x4_t vi3x68AC = vextq_f32(vi3x0246, vi3x8ACE, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4() local
240 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x68AC, vget_low_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
D5x5s2p2-minmax-neonfma-1x4.c120 const float32x4_t vi3x68AC = vextq_f32(vi3x0246, vi3x8ACE9BDF.val[0], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4() local
131 vo0p0 = vfmaq_lane_f32(vo0p0, vi3x68AC, vget_low_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4()
231 const float32x4_t vi3x68AC = vextq_f32(vi3x0246, vi3x8ACE, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4() local
240 vo0p0 = vfmaq_lane_f32(vo0p0, vi3x68AC, vget_low_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4()
D5x5s2p2-minmax-neon-1x4-acc2.c120 const float32x4_t vi3x68AC = vextq_f32(vi3x0246, vi3x8ACE9BDF.val[0], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc2() local
131 vo0p1 = vmlaq_lane_f32(vo0p1, vi3x68AC, vget_low_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc2()
232 const float32x4_t vi3x68AC = vextq_f32(vi3x0246, vi3x8ACE, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc2() local
241 vo0p1 = vmlaq_lane_f32(vo0p1, vi3x68AC, vget_low_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc2()
D5x5s2p2-minmax-neon-2x4-acc2.c148 const float32x4_t vi3x68AC = vextq_f32(vi3x0246, vi3x8ACE9BDF.val[0], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc2() local
161 vo1p1 = vmlaq_lane_f32(vo1p1, vi3x68AC, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc2()
166 vo0p1 = vmlaq_lane_f32(vo0p1, vi3x68AC, vget_low_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc2()
308 const float32x4_t vi3x68AC = vextq_f32(vi3x0246, vi3x8ACE, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc2() local
317 vo1p1 = vmlaq_lane_f32(vo1p1, vi3x68AC, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc2()
322 vo0p1 = vmlaq_lane_f32(vo0p1, vi3x68AC, vget_low_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc2()
D5x5s2p2-minmax-neonfma-2x4-acc2.c148 const float32x4_t vi3x68AC = vextq_f32(vi3x0246, vi3x8ACE9BDF.val[0], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc2() local
161 vo1p1 = vfmaq_lane_f32(vo1p1, vi3x68AC, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc2()
166 vo0p1 = vfmaq_lane_f32(vo0p1, vi3x68AC, vget_low_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc2()
308 const float32x4_t vi3x68AC = vextq_f32(vi3x0246, vi3x8ACE, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc2() local
317 vo1p1 = vfmaq_lane_f32(vo1p1, vi3x68AC, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc2()
322 vo0p1 = vfmaq_lane_f32(vo0p1, vi3x68AC, vget_low_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc2()
D5x5s2p2-minmax-neon-2x4.c148 const float32x4_t vi3x68AC = vextq_f32(vi3x0246, vi3x8ACE9BDF.val[0], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4() local
161 vo1p0 = vmlaq_lane_f32(vo1p0, vi3x68AC, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4()
166 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x68AC, vget_low_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4()
306 const float32x4_t vi3x68AC = vextq_f32(vi3x0246, vi3x8ACE, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4() local
315 vo1p0 = vmlaq_lane_f32(vo1p0, vi3x68AC, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4()
320 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x68AC, vget_low_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4()
D5x5s2p2-minmax-neon-2x4-acc3.c148 const float32x4_t vi3x68AC = vextq_f32(vi3x0246, vi3x8ACE9BDF.val[0], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc3() local
161 vo1p1 = vmlaq_lane_f32(vo1p1, vi3x68AC, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc3()
166 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x68AC, vget_low_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc3()
310 const float32x4_t vi3x68AC = vextq_f32(vi3x0246, vi3x8ACE, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc3() local
319 vo1p1 = vmlaq_lane_f32(vo1p1, vi3x68AC, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc3()
324 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x68AC, vget_low_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc3()
D5x5s2p2-minmax-neonfma-2x4-acc3.c148 const float32x4_t vi3x68AC = vextq_f32(vi3x0246, vi3x8ACE9BDF.val[0], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc3() local
161 vo1p1 = vfmaq_lane_f32(vo1p1, vi3x68AC, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc3()
166 vo0p0 = vfmaq_lane_f32(vo0p0, vi3x68AC, vget_low_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc3()
310 const float32x4_t vi3x68AC = vextq_f32(vi3x0246, vi3x8ACE, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc3() local
319 vo1p1 = vfmaq_lane_f32(vo1p1, vi3x68AC, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc3()
324 vo0p0 = vfmaq_lane_f32(vo0p0, vi3x68AC, vget_low_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc3()
D5x5s2p2-minmax-neonfma-2x4.c148 const float32x4_t vi3x68AC = vextq_f32(vi3x0246, vi3x8ACE9BDF.val[0], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4() local
161 vo1p0 = vfmaq_lane_f32(vo1p0, vi3x68AC, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4()
166 vo0p0 = vfmaq_lane_f32(vo0p0, vi3x68AC, vget_low_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4()
306 const float32x4_t vi3x68AC = vextq_f32(vi3x0246, vi3x8ACE, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4() local
315 vo1p0 = vfmaq_lane_f32(vo1p0, vi3x68AC, vget_high_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4()
320 vo0p0 = vfmaq_lane_f32(vo0p0, vi3x68AC, vget_low_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4()
D5x5s2p2-minmax-neon-1x4-acc5.c120 const float32x4_t vi3x68AC = vextq_f32(vi3x0246, vi3x8ACE9BDF.val[0], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc5() local
131 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x68AC, vget_low_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc5()
235 const float32x4_t vi3x68AC = vextq_f32(vi3x0246, vi3x8ACE, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc5() local
244 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x68AC, vget_low_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc5()
D5x5s2p2-minmax-neonfma-1x4-acc5.c120 const float32x4_t vi3x68AC = vextq_f32(vi3x0246, vi3x8ACE9BDF.val[0], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc5() local
131 vo0p0 = vfmaq_lane_f32(vo0p0, vi3x68AC, vget_low_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc5()
235 const float32x4_t vi3x68AC = vextq_f32(vi3x0246, vi3x8ACE, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc5() local
244 vo0p0 = vfmaq_lane_f32(vo0p0, vi3x68AC, vget_low_f32(vwGHIJ), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc5()
D5x5s2p2-minmax-wasmsimd-arm-loadsplat-2x4.c200 const v128_t vi3x68AC = wasm_v32x4_shuffle(vi3x0246, vi3x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4() local
213 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x68AC, vk10)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4()
218 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x68AC, vk30)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4()
393 const v128_t vi3x68AC = wasm_v32x4_shuffle(vi3x0246, vi3x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4() local
402 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x68AC, vk10)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4()
407 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x68AC, vk30)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4()
D5x5s2p2-minmax-wasmsimd-x86-loadsplat-2x4-acc2.c200 const v128_t vi3x68AC = wasm_v32x4_shuffle(vi3x0246, vi3x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc2() local
213 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi3x68AC, vk10)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc2()
218 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi3x68AC, vk30)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc2()
395 const v128_t vi3x68AC = wasm_v32x4_shuffle(vi3x0246, vi3x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc2() local
404 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi3x68AC, vk10)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc2()
409 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi3x68AC, vk30)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc2()
D5x5s2p2-minmax-wasmsimd-arm-loadsplat-2x4-acc2.c200 const v128_t vi3x68AC = wasm_v32x4_shuffle(vi3x0246, vi3x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc2() local
213 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi3x68AC, vk10)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc2()
218 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi3x68AC, vk30)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc2()
395 const v128_t vi3x68AC = wasm_v32x4_shuffle(vi3x0246, vi3x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc2() local
404 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi3x68AC, vk10)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc2()
409 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi3x68AC, vk30)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc2()
D5x5s2p2-minmax-wasmsimd-x86-loadsplat-2x4.c200 const v128_t vi3x68AC = wasm_v32x4_shuffle(vi3x0246, vi3x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4() local
213 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x68AC, vk10)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4()
218 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x68AC, vk30)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4()
393 const v128_t vi3x68AC = wasm_v32x4_shuffle(vi3x0246, vi3x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4() local
402 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x68AC, vk10)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4()
407 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x68AC, vk30)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4()
D5x5s2p2-minmax-wasmsimd-arm-loadsplat-2x4-acc3.c200 const v128_t vi3x68AC = wasm_v32x4_shuffle(vi3x0246, vi3x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc3() local
213 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi3x68AC, vk10)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc3()
218 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x68AC, vk30)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc3()
397 const v128_t vi3x68AC = wasm_v32x4_shuffle(vi3x0246, vi3x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc3() local
406 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi3x68AC, vk10)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc3()
411 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x68AC, vk30)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc3()
D5x5s2p2-minmax-wasmsimd-x86-loadsplat-2x4-acc3.c200 const v128_t vi3x68AC = wasm_v32x4_shuffle(vi3x0246, vi3x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc3() local
213 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi3x68AC, vk10)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc3()
218 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x68AC, vk30)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc3()
397 const v128_t vi3x68AC = wasm_v32x4_shuffle(vi3x0246, vi3x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc3() local
406 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi3x68AC, vk10)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc3()
411 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x68AC, vk30)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc3()
D5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc3.c164 const v128_t vi3x68AC = wasm_v32x4_shuffle(vi3x0246, vi3x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3() local
175 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x68AC, vk30)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3()
302 const v128_t vi3x68AC = wasm_v32x4_shuffle(vi3x0246, vi3x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3() local
311 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x68AC, vk30)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3()
D5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc2.c164 const v128_t vi3x68AC = wasm_v32x4_shuffle(vi3x0246, vi3x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2() local
175 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi3x68AC, vk30)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2()
301 const v128_t vi3x68AC = wasm_v32x4_shuffle(vi3x0246, vi3x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2() local
310 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi3x68AC, vk30)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2()
D5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4.c164 const v128_t vi3x68AC = wasm_v32x4_shuffle(vi3x0246, vi3x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4() local
175 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x68AC, vk30)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4()
300 const v128_t vi3x68AC = wasm_v32x4_shuffle(vi3x0246, vi3x8ACE, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4() local
309 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x68AC, vk30)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4()

123