Home
last modified time | relevance | path

Searched refs:vi2x79BD (Results 1 – 25 of 70) sorted by relevance

123

/external/XNNPACK/src/f32-dwconv2d-chw/gen/
D5x5s2p2-minmax-neon-1x4-acc4.c139 const float32x4_t vi2x79BD = vextq_f32(vi2x1357, vi2x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc4() local
156 vo0p3 = vmlaq_lane_f32(vo0p3, vi2x79BD, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc4()
249 const float32x4_t vi2x79BD = vextq_f32(vi2x1357, vi2x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc4() local
257 vo0p3 = vmlaq_lane_f32(vo0p3, vi2x79BD, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc4()
D5x5s2p2-minmax-neon-1x4-acc3.c139 const float32x4_t vi2x79BD = vextq_f32(vi2x1357, vi2x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc3() local
156 vo0p1 = vmlaq_lane_f32(vo0p1, vi2x79BD, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc3()
248 const float32x4_t vi2x79BD = vextq_f32(vi2x1357, vi2x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc3() local
256 vo0p1 = vmlaq_lane_f32(vo0p1, vi2x79BD, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc3()
D5x5s2p2-minmax-neonfma-1x4-acc4.c139 const float32x4_t vi2x79BD = vextq_f32(vi2x1357, vi2x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc4() local
156 vo0p3 = vfmaq_lane_f32(vo0p3, vi2x79BD, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc4()
249 const float32x4_t vi2x79BD = vextq_f32(vi2x1357, vi2x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc4() local
257 vo0p3 = vfmaq_lane_f32(vo0p3, vi2x79BD, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc4()
D5x5s2p2-minmax-neonfma-1x4-acc2.c139 const float32x4_t vi2x79BD = vextq_f32(vi2x1357, vi2x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2() local
156 vo0p1 = vfmaq_lane_f32(vo0p1, vi2x79BD, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2()
247 const float32x4_t vi2x79BD = vextq_f32(vi2x1357, vi2x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2() local
255 vo0p1 = vfmaq_lane_f32(vo0p1, vi2x79BD, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2()
D5x5s2p2-minmax-neonfma-1x4-acc3.c139 const float32x4_t vi2x79BD = vextq_f32(vi2x1357, vi2x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc3() local
156 vo0p1 = vfmaq_lane_f32(vo0p1, vi2x79BD, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc3()
248 const float32x4_t vi2x79BD = vextq_f32(vi2x1357, vi2x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc3() local
256 vo0p1 = vfmaq_lane_f32(vo0p1, vi2x79BD, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc3()
D5x5s2p2-minmax-neon-1x4.c139 const float32x4_t vi2x79BD = vextq_f32(vi2x1357, vi2x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4() local
156 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x79BD, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
246 const float32x4_t vi2x79BD = vextq_f32(vi2x1357, vi2x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4() local
254 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x79BD, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
D5x5s2p2-minmax-neonfma-1x4.c139 const float32x4_t vi2x79BD = vextq_f32(vi2x1357, vi2x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4() local
156 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x79BD, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4()
246 const float32x4_t vi2x79BD = vextq_f32(vi2x1357, vi2x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4() local
254 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x79BD, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4()
D5x5s2p2-minmax-neon-1x4-acc2.c139 const float32x4_t vi2x79BD = vextq_f32(vi2x1357, vi2x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc2() local
156 vo0p1 = vmlaq_lane_f32(vo0p1, vi2x79BD, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc2()
247 const float32x4_t vi2x79BD = vextq_f32(vi2x1357, vi2x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc2() local
255 vo0p1 = vmlaq_lane_f32(vo0p1, vi2x79BD, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc2()
D5x5s2p2-minmax-neon-2x4-acc2.c176 const float32x4_t vi2x79BD = vextq_f32(vi2x1357, vi2x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc2() local
196 vo1p1 = vmlaq_lane_f32(vo1p1, vi2x79BD, vget_high_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc2()
201 vo0p1 = vmlaq_lane_f32(vo0p1, vi2x79BD, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc2()
330 const float32x4_t vi2x79BD = vextq_f32(vi2x1357, vi2x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc2() local
337 vo1p1 = vmlaq_lane_f32(vo1p1, vi2x79BD, vget_high_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc2()
342 vo0p1 = vmlaq_lane_f32(vo0p1, vi2x79BD, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc2()
D5x5s2p2-minmax-neonfma-2x4-acc2.c176 const float32x4_t vi2x79BD = vextq_f32(vi2x1357, vi2x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc2() local
196 vo1p1 = vfmaq_lane_f32(vo1p1, vi2x79BD, vget_high_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc2()
201 vo0p1 = vfmaq_lane_f32(vo0p1, vi2x79BD, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc2()
330 const float32x4_t vi2x79BD = vextq_f32(vi2x1357, vi2x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc2() local
337 vo1p1 = vfmaq_lane_f32(vo1p1, vi2x79BD, vget_high_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc2()
342 vo0p1 = vfmaq_lane_f32(vo0p1, vi2x79BD, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc2()
D5x5s2p2-minmax-neon-2x4.c176 const float32x4_t vi2x79BD = vextq_f32(vi2x1357, vi2x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4() local
196 vo1p0 = vmlaq_lane_f32(vo1p0, vi2x79BD, vget_high_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4()
201 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x79BD, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4()
328 const float32x4_t vi2x79BD = vextq_f32(vi2x1357, vi2x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4() local
335 vo1p0 = vmlaq_lane_f32(vo1p0, vi2x79BD, vget_high_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4()
340 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x79BD, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4()
D5x5s2p2-minmax-neon-2x4-acc3.c176 const float32x4_t vi2x79BD = vextq_f32(vi2x1357, vi2x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc3() local
196 vo1p2 = vmlaq_lane_f32(vo1p2, vi2x79BD, vget_high_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc3()
201 vo0p1 = vmlaq_lane_f32(vo0p1, vi2x79BD, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc3()
332 const float32x4_t vi2x79BD = vextq_f32(vi2x1357, vi2x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc3() local
339 vo1p2 = vmlaq_lane_f32(vo1p2, vi2x79BD, vget_high_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc3()
344 vo0p1 = vmlaq_lane_f32(vo0p1, vi2x79BD, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc3()
D5x5s2p2-minmax-neonfma-2x4-acc3.c176 const float32x4_t vi2x79BD = vextq_f32(vi2x1357, vi2x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc3() local
196 vo1p2 = vfmaq_lane_f32(vo1p2, vi2x79BD, vget_high_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc3()
201 vo0p1 = vfmaq_lane_f32(vo0p1, vi2x79BD, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc3()
332 const float32x4_t vi2x79BD = vextq_f32(vi2x1357, vi2x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc3() local
339 vo1p2 = vfmaq_lane_f32(vo1p2, vi2x79BD, vget_high_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc3()
344 vo0p1 = vfmaq_lane_f32(vo0p1, vi2x79BD, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc3()
D5x5s2p2-minmax-neonfma-2x4.c176 const float32x4_t vi2x79BD = vextq_f32(vi2x1357, vi2x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4() local
196 vo1p0 = vfmaq_lane_f32(vo1p0, vi2x79BD, vget_high_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4()
201 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x79BD, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4()
328 const float32x4_t vi2x79BD = vextq_f32(vi2x1357, vi2x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4() local
335 vo1p0 = vfmaq_lane_f32(vo1p0, vi2x79BD, vget_high_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4()
340 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x79BD, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4()
D5x5s2p2-minmax-neon-1x4-acc5.c139 const float32x4_t vi2x79BD = vextq_f32(vi2x1357, vi2x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc5() local
156 vo0p4 = vmlaq_lane_f32(vo0p4, vi2x79BD, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc5()
250 const float32x4_t vi2x79BD = vextq_f32(vi2x1357, vi2x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc5() local
258 vo0p4 = vmlaq_lane_f32(vo0p4, vi2x79BD, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc5()
D5x5s2p2-minmax-neonfma-1x4-acc5.c139 const float32x4_t vi2x79BD = vextq_f32(vi2x1357, vi2x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc5() local
156 vo0p4 = vfmaq_lane_f32(vo0p4, vi2x79BD, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc5()
250 const float32x4_t vi2x79BD = vextq_f32(vi2x1357, vi2x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc5() local
258 vo0p4 = vfmaq_lane_f32(vo0p4, vi2x79BD, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc5()
D5x5s2p2-minmax-wasmsimd-arm-loadsplat-2x4.c228 const v128_t vi2x79BD = wasm_v32x4_shuffle(vi2x1357, vi2x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4() local
277 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x79BD, vk01)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4()
282 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x79BD, vk21)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4()
415 const v128_t vi2x79BD = wasm_v32x4_shuffle(vi2x1357, vi2x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4() local
422 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x79BD, vk01)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4()
427 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x79BD, vk21)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4()
D5x5s2p2-minmax-wasmsimd-x86-loadsplat-2x4-acc2.c228 const v128_t vi2x79BD = wasm_v32x4_shuffle(vi2x1357, vi2x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc2() local
277 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi2x79BD, vk01)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc2()
282 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x79BD, vk21)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc2()
417 const v128_t vi2x79BD = wasm_v32x4_shuffle(vi2x1357, vi2x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc2() local
424 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi2x79BD, vk01)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc2()
429 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x79BD, vk21)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc2()
D5x5s2p2-minmax-wasmsimd-arm-loadsplat-2x4-acc2.c228 const v128_t vi2x79BD = wasm_v32x4_shuffle(vi2x1357, vi2x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc2() local
277 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi2x79BD, vk01)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc2()
282 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x79BD, vk21)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc2()
417 const v128_t vi2x79BD = wasm_v32x4_shuffle(vi2x1357, vi2x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc2() local
424 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi2x79BD, vk01)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc2()
429 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x79BD, vk21)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc2()
D5x5s2p2-minmax-wasmsimd-x86-loadsplat-2x4.c228 const v128_t vi2x79BD = wasm_v32x4_shuffle(vi2x1357, vi2x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4() local
277 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x79BD, vk01)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4()
282 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x79BD, vk21)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4()
415 const v128_t vi2x79BD = wasm_v32x4_shuffle(vi2x1357, vi2x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4() local
422 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x79BD, vk01)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4()
427 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x79BD, vk21)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4()
D5x5s2p2-minmax-wasmsimd-arm-loadsplat-2x4-acc3.c228 const v128_t vi2x79BD = wasm_v32x4_shuffle(vi2x1357, vi2x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc3() local
277 vo1p2 = wasm_f32x4_add(vo1p2, wasm_f32x4_mul(vi2x79BD, vk01)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc3()
282 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x79BD, vk21)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc3()
419 const v128_t vi2x79BD = wasm_v32x4_shuffle(vi2x1357, vi2x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc3() local
426 vo1p2 = wasm_f32x4_add(vo1p2, wasm_f32x4_mul(vi2x79BD, vk01)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc3()
431 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x79BD, vk21)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc3()
D5x5s2p2-minmax-wasmsimd-x86-loadsplat-2x4-acc3.c228 const v128_t vi2x79BD = wasm_v32x4_shuffle(vi2x1357, vi2x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc3() local
277 vo1p2 = wasm_f32x4_add(vo1p2, wasm_f32x4_mul(vi2x79BD, vk01)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc3()
282 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x79BD, vk21)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc3()
419 const v128_t vi2x79BD = wasm_v32x4_shuffle(vi2x1357, vi2x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc3() local
426 vo1p2 = wasm_f32x4_add(vo1p2, wasm_f32x4_mul(vi2x79BD, vk01)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc3()
431 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x79BD, vk21)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc3()
D5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc3.c183 const v128_t vi2x79BD = wasm_v32x4_shuffle(vi2x1357, vi2x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3() local
221 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x79BD, vk21)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3()
317 const v128_t vi2x79BD = wasm_v32x4_shuffle(vi2x1357, vi2x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3() local
325 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x79BD, vk21)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3()
D5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc2.c183 const v128_t vi2x79BD = wasm_v32x4_shuffle(vi2x1357, vi2x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2() local
221 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x79BD, vk21)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2()
316 const v128_t vi2x79BD = wasm_v32x4_shuffle(vi2x1357, vi2x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2() local
324 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x79BD, vk21)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2()
D5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4.c183 const v128_t vi2x79BD = wasm_v32x4_shuffle(vi2x1357, vi2x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4() local
221 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x79BD, vk21)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4()
315 const v128_t vi2x79BD = wasm_v32x4_shuffle(vi2x1357, vi2x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4() local
323 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x79BD, vk21)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4()

123