Home
last modified time | relevance | path

Searched refs:vi2xACEG (Results 1 – 25 of 70) sorted by relevance

123

/external/XNNPACK/src/f32-dwconv2d-chw/gen/
D5x5s2p2-minmax-neon-1x4-acc4.c166 const float32x4_t vi2xACEG = vextq_f32(vi2x8ACE9BDF.val[0], vi2xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc4() local
177 vo0p0 = vmlaq_lane_f32(vo0p0, vi2xACEG, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc4()
266 const float32x4_t vi2xACEG = vextq_f32(vi2x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc4() local
274 vo0p0 = vmlaq_lane_f32(vo0p0, vi2xACEG, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc4()
D5x5s2p2-minmax-neon-1x4-acc3.c166 const float32x4_t vi2xACEG = vextq_f32(vi2x8ACE9BDF.val[0], vi2xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc3() local
177 vo0p0 = vmlaq_lane_f32(vo0p0, vi2xACEG, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc3()
265 const float32x4_t vi2xACEG = vextq_f32(vi2x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc3() local
273 vo0p0 = vmlaq_lane_f32(vo0p0, vi2xACEG, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc3()
D5x5s2p2-minmax-neonfma-1x4-acc4.c166 const float32x4_t vi2xACEG = vextq_f32(vi2x8ACE9BDF.val[0], vi2xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc4() local
177 vo0p0 = vfmaq_lane_f32(vo0p0, vi2xACEG, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc4()
266 const float32x4_t vi2xACEG = vextq_f32(vi2x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc4() local
274 vo0p0 = vfmaq_lane_f32(vo0p0, vi2xACEG, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc4()
D5x5s2p2-minmax-neonfma-1x4-acc2.c166 const float32x4_t vi2xACEG = vextq_f32(vi2x8ACE9BDF.val[0], vi2xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2() local
177 vo0p0 = vfmaq_lane_f32(vo0p0, vi2xACEG, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2()
264 const float32x4_t vi2xACEG = vextq_f32(vi2x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2() local
272 vo0p0 = vfmaq_lane_f32(vo0p0, vi2xACEG, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2()
D5x5s2p2-minmax-neonfma-1x4-acc3.c166 const float32x4_t vi2xACEG = vextq_f32(vi2x8ACE9BDF.val[0], vi2xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc3() local
177 vo0p0 = vfmaq_lane_f32(vo0p0, vi2xACEG, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc3()
265 const float32x4_t vi2xACEG = vextq_f32(vi2x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc3() local
273 vo0p0 = vfmaq_lane_f32(vo0p0, vi2xACEG, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc3()
D5x5s2p2-minmax-neon-1x4.c166 const float32x4_t vi2xACEG = vextq_f32(vi2x8ACE9BDF.val[0], vi2xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4() local
177 vo0p0 = vmlaq_lane_f32(vo0p0, vi2xACEG, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
263 const float32x4_t vi2xACEG = vextq_f32(vi2x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4() local
271 vo0p0 = vmlaq_lane_f32(vo0p0, vi2xACEG, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
D5x5s2p2-minmax-neonfma-1x4.c166 const float32x4_t vi2xACEG = vextq_f32(vi2x8ACE9BDF.val[0], vi2xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4() local
177 vo0p0 = vfmaq_lane_f32(vo0p0, vi2xACEG, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4()
263 const float32x4_t vi2xACEG = vextq_f32(vi2x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4() local
271 vo0p0 = vfmaq_lane_f32(vo0p0, vi2xACEG, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4()
D5x5s2p2-minmax-neon-1x4-acc2.c166 const float32x4_t vi2xACEG = vextq_f32(vi2x8ACE9BDF.val[0], vi2xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc2() local
177 vo0p0 = vmlaq_lane_f32(vo0p0, vi2xACEG, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc2()
264 const float32x4_t vi2xACEG = vextq_f32(vi2x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc2() local
272 vo0p0 = vmlaq_lane_f32(vo0p0, vi2xACEG, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc2()
D5x5s2p2-minmax-neon-2x4-acc2.c214 const float32x4_t vi2xACEG = vextq_f32(vi2x8ACE9BDF.val[0], vi2xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc2() local
226 vo1p0 = vmlaq_lane_f32(vo1p0, vi2xACEG, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc2()
231 vo0p0 = vmlaq_lane_f32(vo0p0, vi2xACEG, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc2()
354 const float32x4_t vi2xACEG = vextq_f32(vi2x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc2() local
361 vo1p0 = vmlaq_lane_f32(vo1p0, vi2xACEG, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc2()
366 vo0p0 = vmlaq_lane_f32(vo0p0, vi2xACEG, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc2()
D5x5s2p2-minmax-neonfma-2x4-acc2.c214 const float32x4_t vi2xACEG = vextq_f32(vi2x8ACE9BDF.val[0], vi2xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc2() local
226 vo1p0 = vfmaq_lane_f32(vo1p0, vi2xACEG, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc2()
231 vo0p0 = vfmaq_lane_f32(vo0p0, vi2xACEG, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc2()
354 const float32x4_t vi2xACEG = vextq_f32(vi2x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc2() local
361 vo1p0 = vfmaq_lane_f32(vo1p0, vi2xACEG, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc2()
366 vo0p0 = vfmaq_lane_f32(vo0p0, vi2xACEG, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc2()
D5x5s2p2-minmax-neon-2x4.c214 const float32x4_t vi2xACEG = vextq_f32(vi2x8ACE9BDF.val[0], vi2xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4() local
226 vo1p0 = vmlaq_lane_f32(vo1p0, vi2xACEG, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4()
231 vo0p0 = vmlaq_lane_f32(vo0p0, vi2xACEG, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4()
352 const float32x4_t vi2xACEG = vextq_f32(vi2x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4() local
359 vo1p0 = vmlaq_lane_f32(vo1p0, vi2xACEG, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4()
364 vo0p0 = vmlaq_lane_f32(vo0p0, vi2xACEG, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4()
D5x5s2p2-minmax-neon-2x4-acc3.c214 const float32x4_t vi2xACEG = vextq_f32(vi2x8ACE9BDF.val[0], vi2xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc3() local
226 vo1p1 = vmlaq_lane_f32(vo1p1, vi2xACEG, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc3()
231 vo0p0 = vmlaq_lane_f32(vo0p0, vi2xACEG, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc3()
356 const float32x4_t vi2xACEG = vextq_f32(vi2x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc3() local
363 vo1p1 = vmlaq_lane_f32(vo1p1, vi2xACEG, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc3()
368 vo0p0 = vmlaq_lane_f32(vo0p0, vi2xACEG, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc3()
D5x5s2p2-minmax-neonfma-2x4-acc3.c214 const float32x4_t vi2xACEG = vextq_f32(vi2x8ACE9BDF.val[0], vi2xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc3() local
226 vo1p1 = vfmaq_lane_f32(vo1p1, vi2xACEG, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc3()
231 vo0p0 = vfmaq_lane_f32(vo0p0, vi2xACEG, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc3()
356 const float32x4_t vi2xACEG = vextq_f32(vi2x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc3() local
363 vo1p1 = vfmaq_lane_f32(vo1p1, vi2xACEG, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc3()
368 vo0p0 = vfmaq_lane_f32(vo0p0, vi2xACEG, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc3()
D5x5s2p2-minmax-neonfma-2x4.c214 const float32x4_t vi2xACEG = vextq_f32(vi2x8ACE9BDF.val[0], vi2xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4() local
226 vo1p0 = vfmaq_lane_f32(vo1p0, vi2xACEG, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4()
231 vo0p0 = vfmaq_lane_f32(vo0p0, vi2xACEG, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4()
352 const float32x4_t vi2xACEG = vextq_f32(vi2x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4() local
359 vo1p0 = vfmaq_lane_f32(vo1p0, vi2xACEG, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4()
364 vo0p0 = vfmaq_lane_f32(vo0p0, vi2xACEG, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4()
D5x5s2p2-minmax-neon-1x4-acc5.c166 const float32x4_t vi2xACEG = vextq_f32(vi2x8ACE9BDF.val[0], vi2xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc5() local
177 vo0p4 = vmlaq_lane_f32(vo0p4, vi2xACEG, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc5()
267 const float32x4_t vi2xACEG = vextq_f32(vi2x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc5() local
275 vo0p4 = vmlaq_lane_f32(vo0p4, vi2xACEG, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc5()
D5x5s2p2-minmax-neonfma-1x4-acc5.c166 const float32x4_t vi2xACEG = vextq_f32(vi2x8ACE9BDF.val[0], vi2xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc5() local
177 vo0p4 = vfmaq_lane_f32(vo0p4, vi2xACEG, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc5()
267 const float32x4_t vi2xACEG = vextq_f32(vi2x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc5() local
275 vo0p4 = vfmaq_lane_f32(vo0p4, vi2xACEG, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc5()
D5x5s2p2-minmax-wasmsimd-arm-loadsplat-2x4.c297 const v128_t vi2xACEG = wasm_v32x4_shuffle(vi2x8ACE, vi2xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4() local
314 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2xACEG, vk04)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4()
319 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2xACEG, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4()
438 const v128_t vi2xACEG = wasm_v32x4_shuffle(vi2x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4() local
445 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2xACEG, vk04)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4()
450 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2xACEG, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4()
D5x5s2p2-minmax-wasmsimd-x86-loadsplat-2x4-acc2.c297 const v128_t vi2xACEG = wasm_v32x4_shuffle(vi2x8ACE, vi2xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc2() local
314 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2xACEG, vk04)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc2()
319 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2xACEG, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc2()
440 const v128_t vi2xACEG = wasm_v32x4_shuffle(vi2x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc2() local
447 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2xACEG, vk04)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc2()
452 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2xACEG, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc2()
D5x5s2p2-minmax-wasmsimd-arm-loadsplat-2x4-acc2.c297 const v128_t vi2xACEG = wasm_v32x4_shuffle(vi2x8ACE, vi2xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc2() local
314 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2xACEG, vk04)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc2()
319 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2xACEG, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc2()
440 const v128_t vi2xACEG = wasm_v32x4_shuffle(vi2x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc2() local
447 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2xACEG, vk04)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc2()
452 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2xACEG, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc2()
D5x5s2p2-minmax-wasmsimd-x86-loadsplat-2x4.c297 const v128_t vi2xACEG = wasm_v32x4_shuffle(vi2x8ACE, vi2xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4() local
314 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2xACEG, vk04)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4()
319 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2xACEG, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4()
438 const v128_t vi2xACEG = wasm_v32x4_shuffle(vi2x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4() local
445 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2xACEG, vk04)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4()
450 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2xACEG, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4()
D5x5s2p2-minmax-wasmsimd-arm-loadsplat-2x4-acc3.c297 const v128_t vi2xACEG = wasm_v32x4_shuffle(vi2x8ACE, vi2xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc3() local
314 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi2xACEG, vk04)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc3()
319 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2xACEG, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc3()
442 const v128_t vi2xACEG = wasm_v32x4_shuffle(vi2x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc3() local
449 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi2xACEG, vk04)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc3()
454 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2xACEG, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc3()
D5x5s2p2-minmax-wasmsimd-x86-loadsplat-2x4-acc3.c297 const v128_t vi2xACEG = wasm_v32x4_shuffle(vi2x8ACE, vi2xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc3() local
314 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi2xACEG, vk04)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc3()
319 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2xACEG, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc3()
442 const v128_t vi2xACEG = wasm_v32x4_shuffle(vi2x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc3() local
449 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi2xACEG, vk04)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc3()
454 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2xACEG, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc3()
D5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc3.c233 const v128_t vi2xACEG = wasm_v32x4_shuffle(vi2x8ACE, vi2xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3() local
247 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2xACEG, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3()
333 const v128_t vi2xACEG = wasm_v32x4_shuffle(vi2x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3() local
341 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2xACEG, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3()
D5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc2.c233 const v128_t vi2xACEG = wasm_v32x4_shuffle(vi2x8ACE, vi2xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2() local
247 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2xACEG, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2()
332 const v128_t vi2xACEG = wasm_v32x4_shuffle(vi2x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2() local
340 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2xACEG, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2()
D5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4.c233 const v128_t vi2xACEG = wasm_v32x4_shuffle(vi2x8ACE, vi2xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4() local
247 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2xACEG, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4()
331 const v128_t vi2xACEG = wasm_v32x4_shuffle(vi2x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4() local
339 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2xACEG, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4()

123