Home
last modified time | relevance | path

Searched refs:vi1xACEG (Results 1 – 25 of 70) sorted by relevance

123

/external/XNNPACK/src/f32-dwconv2d-chw/gen/
D5x5s2p2-minmax-neon-1x4-acc4.c164 const float32x4_t vi1xACEG = vextq_f32(vi1x8ACE9BDF.val[0], vi1xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc4() local
175 vo0p3 = vmlaq_lane_f32(vo0p3, vi1xACEG, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc4()
265 const float32x4_t vi1xACEG = vextq_f32(vi1x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc4() local
272 vo0p3 = vmlaq_lane_f32(vo0p3, vi1xACEG, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc4()
D5x5s2p2-minmax-neon-1x4-acc3.c164 const float32x4_t vi1xACEG = vextq_f32(vi1x8ACE9BDF.val[0], vi1xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc3() local
175 vo0p2 = vmlaq_lane_f32(vo0p2, vi1xACEG, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc3()
264 const float32x4_t vi1xACEG = vextq_f32(vi1x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc3() local
271 vo0p2 = vmlaq_lane_f32(vo0p2, vi1xACEG, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc3()
D5x5s2p2-minmax-neonfma-1x4-acc4.c164 const float32x4_t vi1xACEG = vextq_f32(vi1x8ACE9BDF.val[0], vi1xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc4() local
175 vo0p3 = vfmaq_lane_f32(vo0p3, vi1xACEG, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc4()
265 const float32x4_t vi1xACEG = vextq_f32(vi1x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc4() local
272 vo0p3 = vfmaq_lane_f32(vo0p3, vi1xACEG, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc4()
D5x5s2p2-minmax-neonfma-1x4-acc2.c164 const float32x4_t vi1xACEG = vextq_f32(vi1x8ACE9BDF.val[0], vi1xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2() local
175 vo0p1 = vfmaq_lane_f32(vo0p1, vi1xACEG, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2()
263 const float32x4_t vi1xACEG = vextq_f32(vi1x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2() local
270 vo0p1 = vfmaq_lane_f32(vo0p1, vi1xACEG, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2()
D5x5s2p2-minmax-neonfma-1x4-acc3.c164 const float32x4_t vi1xACEG = vextq_f32(vi1x8ACE9BDF.val[0], vi1xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc3() local
175 vo0p2 = vfmaq_lane_f32(vo0p2, vi1xACEG, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc3()
264 const float32x4_t vi1xACEG = vextq_f32(vi1x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc3() local
271 vo0p2 = vfmaq_lane_f32(vo0p2, vi1xACEG, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc3()
D5x5s2p2-minmax-neon-1x4.c164 const float32x4_t vi1xACEG = vextq_f32(vi1x8ACE9BDF.val[0], vi1xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4() local
175 vo0p0 = vmlaq_lane_f32(vo0p0, vi1xACEG, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
262 const float32x4_t vi1xACEG = vextq_f32(vi1x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4() local
269 vo0p0 = vmlaq_lane_f32(vo0p0, vi1xACEG, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
D5x5s2p2-minmax-neonfma-1x4.c164 const float32x4_t vi1xACEG = vextq_f32(vi1x8ACE9BDF.val[0], vi1xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4() local
175 vo0p0 = vfmaq_lane_f32(vo0p0, vi1xACEG, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4()
262 const float32x4_t vi1xACEG = vextq_f32(vi1x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4() local
269 vo0p0 = vfmaq_lane_f32(vo0p0, vi1xACEG, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4()
D5x5s2p2-minmax-neon-1x4-acc2.c164 const float32x4_t vi1xACEG = vextq_f32(vi1x8ACE9BDF.val[0], vi1xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc2() local
175 vo0p1 = vmlaq_lane_f32(vo0p1, vi1xACEG, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc2()
263 const float32x4_t vi1xACEG = vextq_f32(vi1x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc2() local
270 vo0p1 = vmlaq_lane_f32(vo0p1, vi1xACEG, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc2()
D5x5s2p2-minmax-neon-1x4-acc5.c164 const float32x4_t vi1xACEG = vextq_f32(vi1x8ACE9BDF.val[0], vi1xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc5() local
175 vo0p3 = vmlaq_lane_f32(vo0p3, vi1xACEG, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc5()
266 const float32x4_t vi1xACEG = vextq_f32(vi1x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc5() local
273 vo0p3 = vmlaq_lane_f32(vo0p3, vi1xACEG, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc5()
D5x5s2p2-minmax-neonfma-1x4-acc5.c164 const float32x4_t vi1xACEG = vextq_f32(vi1x8ACE9BDF.val[0], vi1xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc5() local
175 vo0p3 = vfmaq_lane_f32(vo0p3, vi1xACEG, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc5()
266 const float32x4_t vi1xACEG = vextq_f32(vi1x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc5() local
273 vo0p3 = vfmaq_lane_f32(vo0p3, vi1xACEG, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc5()
D5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc3.c230 const v128_t vi1xACEG = wasm_v32x4_shuffle(vi1x8ACE, vi1xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3() local
245 vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi1xACEG, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3()
332 const v128_t vi1xACEG = wasm_v32x4_shuffle(vi1x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3() local
339 vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi1xACEG, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3()
D5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc2.c230 const v128_t vi1xACEG = wasm_v32x4_shuffle(vi1x8ACE, vi1xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2() local
245 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi1xACEG, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2()
331 const v128_t vi1xACEG = wasm_v32x4_shuffle(vi1x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2() local
338 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi1xACEG, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2()
D5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4.c230 const v128_t vi1xACEG = wasm_v32x4_shuffle(vi1x8ACE, vi1xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4() local
245 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1xACEG, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4()
330 const v128_t vi1xACEG = wasm_v32x4_shuffle(vi1x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4() local
337 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1xACEG, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4()
D5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4.c230 const v128_t vi1xACEG = wasm_v32x4_shuffle(vi1x8ACE, vi1xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4() local
245 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1xACEG, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
330 const v128_t vi1xACEG = wasm_v32x4_shuffle(vi1x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4() local
337 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1xACEG, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
D5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4-acc4.c230 const v128_t vi1xACEG = wasm_v32x4_shuffle(vi1x8ACE, vi1xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc4() local
245 vo0p3 = wasm_f32x4_add(vo0p3, wasm_f32x4_mul(vi1xACEG, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc4()
333 const v128_t vi1xACEG = wasm_v32x4_shuffle(vi1x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc4() local
340 vo0p3 = wasm_f32x4_add(vo0p3, wasm_f32x4_mul(vi1xACEG, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc4()
D5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc4.c210 const v128_t vi1xACEG = wasm_v32x4_shuffle(vi1x8ACE, vi1xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc4() local
225 …vo0p3 = wasm_f32x4_add(vo0p3, wasm_f32x4_mul(vi1xACEG, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc4()
313 const v128_t vi1xACEG = wasm_v32x4_shuffle(vi1x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc4() local
320 …vo0p3 = wasm_f32x4_add(vo0p3, wasm_f32x4_mul(vi1xACEG, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc4()
D5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc5.c210 const v128_t vi1xACEG = wasm_v32x4_shuffle(vi1x8ACE, vi1xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc5() local
225 …vo0p3 = wasm_f32x4_add(vo0p3, wasm_f32x4_mul(vi1xACEG, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc5()
314 const v128_t vi1xACEG = wasm_v32x4_shuffle(vi1x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc5() local
321 …vo0p3 = wasm_f32x4_add(vo0p3, wasm_f32x4_mul(vi1xACEG, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc5()
D5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc4.c230 const v128_t vi1xACEG = wasm_v32x4_shuffle(vi1x8ACE, vi1xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc4() local
245 vo0p3 = wasm_f32x4_add(vo0p3, wasm_f32x4_mul(vi1xACEG, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc4()
333 const v128_t vi1xACEG = wasm_v32x4_shuffle(vi1x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc4() local
340 vo0p3 = wasm_f32x4_add(vo0p3, wasm_f32x4_mul(vi1xACEG, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc4()
D5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc2.c210 const v128_t vi1xACEG = wasm_v32x4_shuffle(vi1x8ACE, vi1xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc2() local
225 …vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi1xACEG, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc2()
311 const v128_t vi1xACEG = wasm_v32x4_shuffle(vi1x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc2() local
318 …vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi1xACEG, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc2()
D5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc5.c210 const v128_t vi1xACEG = wasm_v32x4_shuffle(vi1x8ACE, vi1xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc5() local
225 …vo0p3 = wasm_f32x4_add(vo0p3, wasm_f32x4_mul(vi1xACEG, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc5()
314 const v128_t vi1xACEG = wasm_v32x4_shuffle(vi1x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc5() local
321 …vo0p3 = wasm_f32x4_add(vo0p3, wasm_f32x4_mul(vi1xACEG, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc5()
D5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc2.c210 const v128_t vi1xACEG = wasm_v32x4_shuffle(vi1x8ACE, vi1xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc2() local
225 …vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi1xACEG, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc2()
311 const v128_t vi1xACEG = wasm_v32x4_shuffle(vi1x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc2() local
318 …vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi1xACEG, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc2()
D5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc3.c210 const v128_t vi1xACEG = wasm_v32x4_shuffle(vi1x8ACE, vi1xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc3() local
225 …vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi1xACEG, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc3()
312 const v128_t vi1xACEG = wasm_v32x4_shuffle(vi1x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc3() local
319 …vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi1xACEG, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc3()
D5x5s2p2-minmax-wasmsimd-x86-splat-1x4.c210 const v128_t vi1xACEG = wasm_v32x4_shuffle(vi1x8ACE, vi1xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4() local
225 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1xACEG, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
310 const v128_t vi1xACEG = wasm_v32x4_shuffle(vi1x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4() local
317 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1xACEG, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
D5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc5.c230 const v128_t vi1xACEG = wasm_v32x4_shuffle(vi1x8ACE, vi1xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc5() local
245 vo0p3 = wasm_f32x4_add(vo0p3, wasm_f32x4_mul(vi1xACEG, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc5()
334 const v128_t vi1xACEG = wasm_v32x4_shuffle(vi1x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc5() local
341 vo0p3 = wasm_f32x4_add(vo0p3, wasm_f32x4_mul(vi1xACEG, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc5()
D5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc4.c210 const v128_t vi1xACEG = wasm_v32x4_shuffle(vi1x8ACE, vi1xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc4() local
225 …vo0p3 = wasm_f32x4_add(vo0p3, wasm_f32x4_mul(vi1xACEG, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc4()
313 const v128_t vi1xACEG = wasm_v32x4_shuffle(vi1x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc4() local
320 …vo0p3 = wasm_f32x4_add(vo0p3, wasm_f32x4_mul(vi1xACEG, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc4()

123