Home
last modified time | relevance | path

Searched refs:vi3xACEG (Results 1 – 25 of 70) sorted by relevance

123

/external/XNNPACK/src/f32-dwconv2d-chw/gen/
D5x5s2p2-minmax-neon-1x4-acc4.c168 const float32x4_t vi3xACEG = vextq_f32(vi3x8ACE9BDF.val[0], vi3xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc4() local
179 vo0p1 = vmlaq_lane_f32(vo0p1, vi3xACEG, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc4()
267 const float32x4_t vi3xACEG = vextq_f32(vi3x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc4() local
276 vo0p1 = vmlaq_lane_f32(vo0p1, vi3xACEG, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc4()
D5x5s2p2-minmax-neon-1x4-acc3.c168 const float32x4_t vi3xACEG = vextq_f32(vi3x8ACE9BDF.val[0], vi3xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc3() local
179 vo0p1 = vmlaq_lane_f32(vo0p1, vi3xACEG, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc3()
266 const float32x4_t vi3xACEG = vextq_f32(vi3x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc3() local
275 vo0p1 = vmlaq_lane_f32(vo0p1, vi3xACEG, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc3()
D5x5s2p2-minmax-neonfma-1x4-acc4.c168 const float32x4_t vi3xACEG = vextq_f32(vi3x8ACE9BDF.val[0], vi3xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc4() local
179 vo0p1 = vfmaq_lane_f32(vo0p1, vi3xACEG, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc4()
267 const float32x4_t vi3xACEG = vextq_f32(vi3x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc4() local
276 vo0p1 = vfmaq_lane_f32(vo0p1, vi3xACEG, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc4()
D5x5s2p2-minmax-neonfma-1x4-acc2.c168 const float32x4_t vi3xACEG = vextq_f32(vi3x8ACE9BDF.val[0], vi3xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2() local
179 vo0p1 = vfmaq_lane_f32(vo0p1, vi3xACEG, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2()
265 const float32x4_t vi3xACEG = vextq_f32(vi3x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2() local
274 vo0p1 = vfmaq_lane_f32(vo0p1, vi3xACEG, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2()
D5x5s2p2-minmax-neonfma-1x4-acc3.c168 const float32x4_t vi3xACEG = vextq_f32(vi3x8ACE9BDF.val[0], vi3xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc3() local
179 vo0p1 = vfmaq_lane_f32(vo0p1, vi3xACEG, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc3()
266 const float32x4_t vi3xACEG = vextq_f32(vi3x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc3() local
275 vo0p1 = vfmaq_lane_f32(vo0p1, vi3xACEG, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc3()
D5x5s2p2-minmax-neon-1x4.c168 const float32x4_t vi3xACEG = vextq_f32(vi3x8ACE9BDF.val[0], vi3xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4() local
179 vo0p0 = vmlaq_lane_f32(vo0p0, vi3xACEG, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
264 const float32x4_t vi3xACEG = vextq_f32(vi3x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4() local
273 vo0p0 = vmlaq_lane_f32(vo0p0, vi3xACEG, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
D5x5s2p2-minmax-neonfma-1x4.c168 const float32x4_t vi3xACEG = vextq_f32(vi3x8ACE9BDF.val[0], vi3xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4() local
179 vo0p0 = vfmaq_lane_f32(vo0p0, vi3xACEG, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4()
264 const float32x4_t vi3xACEG = vextq_f32(vi3x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4() local
273 vo0p0 = vfmaq_lane_f32(vo0p0, vi3xACEG, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4()
D5x5s2p2-minmax-neon-1x4-acc2.c168 const float32x4_t vi3xACEG = vextq_f32(vi3x8ACE9BDF.val[0], vi3xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc2() local
179 vo0p1 = vmlaq_lane_f32(vo0p1, vi3xACEG, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc2()
265 const float32x4_t vi3xACEG = vextq_f32(vi3x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc2() local
274 vo0p1 = vmlaq_lane_f32(vo0p1, vi3xACEG, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc2()
D5x5s2p2-minmax-neon-2x4-acc2.c216 const float32x4_t vi3xACEG = vextq_f32(vi3x8ACE9BDF.val[0], vi3xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc2() local
229 vo1p1 = vmlaq_lane_f32(vo1p1, vi3xACEG, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc2()
234 vo0p1 = vmlaq_lane_f32(vo0p1, vi3xACEG, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc2()
355 const float32x4_t vi3xACEG = vextq_f32(vi3x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc2() local
364 vo1p1 = vmlaq_lane_f32(vo1p1, vi3xACEG, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc2()
369 vo0p1 = vmlaq_lane_f32(vo0p1, vi3xACEG, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc2()
D5x5s2p2-minmax-neonfma-2x4-acc2.c216 const float32x4_t vi3xACEG = vextq_f32(vi3x8ACE9BDF.val[0], vi3xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc2() local
229 vo1p1 = vfmaq_lane_f32(vo1p1, vi3xACEG, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc2()
234 vo0p1 = vfmaq_lane_f32(vo0p1, vi3xACEG, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc2()
355 const float32x4_t vi3xACEG = vextq_f32(vi3x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc2() local
364 vo1p1 = vfmaq_lane_f32(vo1p1, vi3xACEG, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc2()
369 vo0p1 = vfmaq_lane_f32(vo0p1, vi3xACEG, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc2()
D5x5s2p2-minmax-neon-2x4.c216 const float32x4_t vi3xACEG = vextq_f32(vi3x8ACE9BDF.val[0], vi3xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4() local
229 vo1p0 = vmlaq_lane_f32(vo1p0, vi3xACEG, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4()
234 vo0p0 = vmlaq_lane_f32(vo0p0, vi3xACEG, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4()
353 const float32x4_t vi3xACEG = vextq_f32(vi3x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4() local
362 vo1p0 = vmlaq_lane_f32(vo1p0, vi3xACEG, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4()
367 vo0p0 = vmlaq_lane_f32(vo0p0, vi3xACEG, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4()
D5x5s2p2-minmax-neon-2x4-acc3.c216 const float32x4_t vi3xACEG = vextq_f32(vi3x8ACE9BDF.val[0], vi3xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc3() local
229 vo1p2 = vmlaq_lane_f32(vo1p2, vi3xACEG, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc3()
234 vo0p1 = vmlaq_lane_f32(vo0p1, vi3xACEG, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc3()
357 const float32x4_t vi3xACEG = vextq_f32(vi3x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc3() local
366 vo1p2 = vmlaq_lane_f32(vo1p2, vi3xACEG, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc3()
371 vo0p1 = vmlaq_lane_f32(vo0p1, vi3xACEG, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc3()
D5x5s2p2-minmax-neonfma-2x4-acc3.c216 const float32x4_t vi3xACEG = vextq_f32(vi3x8ACE9BDF.val[0], vi3xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc3() local
229 vo1p2 = vfmaq_lane_f32(vo1p2, vi3xACEG, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc3()
234 vo0p1 = vfmaq_lane_f32(vo0p1, vi3xACEG, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc3()
357 const float32x4_t vi3xACEG = vextq_f32(vi3x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc3() local
366 vo1p2 = vfmaq_lane_f32(vo1p2, vi3xACEG, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc3()
371 vo0p1 = vfmaq_lane_f32(vo0p1, vi3xACEG, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc3()
D5x5s2p2-minmax-neonfma-2x4.c216 const float32x4_t vi3xACEG = vextq_f32(vi3x8ACE9BDF.val[0], vi3xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4() local
229 vo1p0 = vfmaq_lane_f32(vo1p0, vi3xACEG, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4()
234 vo0p0 = vfmaq_lane_f32(vo0p0, vi3xACEG, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4()
353 const float32x4_t vi3xACEG = vextq_f32(vi3x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4() local
362 vo1p0 = vfmaq_lane_f32(vo1p0, vi3xACEG, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4()
367 vo0p0 = vfmaq_lane_f32(vo0p0, vi3xACEG, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4()
D5x5s2p2-minmax-neon-1x4-acc5.c168 const float32x4_t vi3xACEG = vextq_f32(vi3x8ACE9BDF.val[0], vi3xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc5() local
179 vo0p0 = vmlaq_lane_f32(vo0p0, vi3xACEG, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc5()
268 const float32x4_t vi3xACEG = vextq_f32(vi3x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc5() local
277 vo0p0 = vmlaq_lane_f32(vo0p0, vi3xACEG, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc5()
D5x5s2p2-minmax-neonfma-1x4-acc5.c168 const float32x4_t vi3xACEG = vextq_f32(vi3x8ACE9BDF.val[0], vi3xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc5() local
179 vo0p0 = vfmaq_lane_f32(vo0p0, vi3xACEG, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc5()
268 const float32x4_t vi3xACEG = vextq_f32(vi3x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc5() local
277 vo0p0 = vfmaq_lane_f32(vo0p0, vi3xACEG, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc5()
D5x5s2p2-minmax-wasmsimd-arm-loadsplat-2x4.c300 const v128_t vi3xACEG = wasm_v32x4_shuffle(vi3x8ACE, vi3xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4() local
317 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3xACEG, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4()
322 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3xACEG, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4()
439 const v128_t vi3xACEG = wasm_v32x4_shuffle(vi3x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4() local
448 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3xACEG, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4()
453 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3xACEG, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4()
D5x5s2p2-minmax-wasmsimd-x86-loadsplat-2x4-acc2.c300 const v128_t vi3xACEG = wasm_v32x4_shuffle(vi3x8ACE, vi3xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc2() local
317 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi3xACEG, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc2()
322 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi3xACEG, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc2()
441 const v128_t vi3xACEG = wasm_v32x4_shuffle(vi3x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc2() local
450 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi3xACEG, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc2()
455 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi3xACEG, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc2()
D5x5s2p2-minmax-wasmsimd-arm-loadsplat-2x4-acc2.c300 const v128_t vi3xACEG = wasm_v32x4_shuffle(vi3x8ACE, vi3xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc2() local
317 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi3xACEG, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc2()
322 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi3xACEG, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc2()
441 const v128_t vi3xACEG = wasm_v32x4_shuffle(vi3x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc2() local
450 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi3xACEG, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc2()
455 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi3xACEG, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc2()
D5x5s2p2-minmax-wasmsimd-x86-loadsplat-2x4.c300 const v128_t vi3xACEG = wasm_v32x4_shuffle(vi3x8ACE, vi3xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4() local
317 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3xACEG, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4()
322 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3xACEG, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4()
439 const v128_t vi3xACEG = wasm_v32x4_shuffle(vi3x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4() local
448 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3xACEG, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4()
453 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3xACEG, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4()
D5x5s2p2-minmax-wasmsimd-arm-loadsplat-2x4-acc3.c300 const v128_t vi3xACEG = wasm_v32x4_shuffle(vi3x8ACE, vi3xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc3() local
317 vo1p2 = wasm_f32x4_add(vo1p2, wasm_f32x4_mul(vi3xACEG, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc3()
322 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi3xACEG, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc3()
443 const v128_t vi3xACEG = wasm_v32x4_shuffle(vi3x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc3() local
452 vo1p2 = wasm_f32x4_add(vo1p2, wasm_f32x4_mul(vi3xACEG, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc3()
457 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi3xACEG, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc3()
D5x5s2p2-minmax-wasmsimd-x86-loadsplat-2x4-acc3.c300 const v128_t vi3xACEG = wasm_v32x4_shuffle(vi3x8ACE, vi3xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc3() local
317 vo1p2 = wasm_f32x4_add(vo1p2, wasm_f32x4_mul(vi3xACEG, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc3()
322 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi3xACEG, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc3()
443 const v128_t vi3xACEG = wasm_v32x4_shuffle(vi3x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc3() local
452 vo1p2 = wasm_f32x4_add(vo1p2, wasm_f32x4_mul(vi3xACEG, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc3()
457 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi3xACEG, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc3()
D5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc3.c236 const v128_t vi3xACEG = wasm_v32x4_shuffle(vi3x8ACE, vi3xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3() local
249 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi3xACEG, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3()
334 const v128_t vi3xACEG = wasm_v32x4_shuffle(vi3x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3() local
343 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi3xACEG, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3()
D5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc2.c236 const v128_t vi3xACEG = wasm_v32x4_shuffle(vi3x8ACE, vi3xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2() local
249 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi3xACEG, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2()
333 const v128_t vi3xACEG = wasm_v32x4_shuffle(vi3x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2() local
342 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi3xACEG, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2()
D5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4.c236 const v128_t vi3xACEG = wasm_v32x4_shuffle(vi3x8ACE, vi3xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4() local
249 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3xACEG, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4()
332 const v128_t vi3xACEG = wasm_v32x4_shuffle(vi3x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4() local
341 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3xACEG, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4()

123