Home
last modified time | relevance | path

Searched refs:vi3x79BD (Results 1 – 25 of 70) sorted by relevance

123

/external/XNNPACK/src/f32-dwconv2d-chw/gen/
D5x5s2p2-minmax-neon-1x4-acc4.c141 const float32x4_t vi3x79BD = vextq_f32(vi3x1357, vi3x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc4() local
158 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x79BD, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc4()
250 const float32x4_t vi3x79BD = vextq_f32(vi3x1357, vi3x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc4() local
259 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x79BD, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc4()
D5x5s2p2-minmax-neon-1x4-acc3.c141 const float32x4_t vi3x79BD = vextq_f32(vi3x1357, vi3x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc3() local
158 vo0p2 = vmlaq_lane_f32(vo0p2, vi3x79BD, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc3()
249 const float32x4_t vi3x79BD = vextq_f32(vi3x1357, vi3x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc3() local
258 vo0p2 = vmlaq_lane_f32(vo0p2, vi3x79BD, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc3()
D5x5s2p2-minmax-neonfma-1x4-acc4.c141 const float32x4_t vi3x79BD = vextq_f32(vi3x1357, vi3x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc4() local
158 vo0p0 = vfmaq_lane_f32(vo0p0, vi3x79BD, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc4()
250 const float32x4_t vi3x79BD = vextq_f32(vi3x1357, vi3x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc4() local
259 vo0p0 = vfmaq_lane_f32(vo0p0, vi3x79BD, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc4()
D5x5s2p2-minmax-neonfma-1x4-acc2.c141 const float32x4_t vi3x79BD = vextq_f32(vi3x1357, vi3x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2() local
158 vo0p0 = vfmaq_lane_f32(vo0p0, vi3x79BD, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2()
248 const float32x4_t vi3x79BD = vextq_f32(vi3x1357, vi3x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2() local
257 vo0p0 = vfmaq_lane_f32(vo0p0, vi3x79BD, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2()
D5x5s2p2-minmax-neonfma-1x4-acc3.c141 const float32x4_t vi3x79BD = vextq_f32(vi3x1357, vi3x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc3() local
158 vo0p2 = vfmaq_lane_f32(vo0p2, vi3x79BD, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc3()
249 const float32x4_t vi3x79BD = vextq_f32(vi3x1357, vi3x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc3() local
258 vo0p2 = vfmaq_lane_f32(vo0p2, vi3x79BD, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc3()
D5x5s2p2-minmax-neon-1x4.c141 const float32x4_t vi3x79BD = vextq_f32(vi3x1357, vi3x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4() local
158 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x79BD, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
247 const float32x4_t vi3x79BD = vextq_f32(vi3x1357, vi3x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4() local
256 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x79BD, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
D5x5s2p2-minmax-neonfma-1x4.c141 const float32x4_t vi3x79BD = vextq_f32(vi3x1357, vi3x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4() local
158 vo0p0 = vfmaq_lane_f32(vo0p0, vi3x79BD, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4()
247 const float32x4_t vi3x79BD = vextq_f32(vi3x1357, vi3x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4() local
256 vo0p0 = vfmaq_lane_f32(vo0p0, vi3x79BD, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4()
D5x5s2p2-minmax-neon-1x4-acc2.c141 const float32x4_t vi3x79BD = vextq_f32(vi3x1357, vi3x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc2() local
158 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x79BD, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc2()
248 const float32x4_t vi3x79BD = vextq_f32(vi3x1357, vi3x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc2() local
257 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x79BD, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc2()
D5x5s2p2-minmax-neon-2x4-acc2.c178 const float32x4_t vi3x79BD = vextq_f32(vi3x1357, vi3x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc2() local
199 vo1p0 = vmlaq_lane_f32(vo1p0, vi3x79BD, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc2()
204 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x79BD, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc2()
331 const float32x4_t vi3x79BD = vextq_f32(vi3x1357, vi3x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc2() local
340 vo1p0 = vmlaq_lane_f32(vo1p0, vi3x79BD, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc2()
345 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x79BD, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc2()
D5x5s2p2-minmax-neonfma-2x4-acc2.c178 const float32x4_t vi3x79BD = vextq_f32(vi3x1357, vi3x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc2() local
199 vo1p0 = vfmaq_lane_f32(vo1p0, vi3x79BD, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc2()
204 vo0p0 = vfmaq_lane_f32(vo0p0, vi3x79BD, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc2()
331 const float32x4_t vi3x79BD = vextq_f32(vi3x1357, vi3x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc2() local
340 vo1p0 = vfmaq_lane_f32(vo1p0, vi3x79BD, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc2()
345 vo0p0 = vfmaq_lane_f32(vo0p0, vi3x79BD, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc2()
D5x5s2p2-minmax-neon-2x4.c178 const float32x4_t vi3x79BD = vextq_f32(vi3x1357, vi3x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4() local
199 vo1p0 = vmlaq_lane_f32(vo1p0, vi3x79BD, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4()
204 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x79BD, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4()
329 const float32x4_t vi3x79BD = vextq_f32(vi3x1357, vi3x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4() local
338 vo1p0 = vmlaq_lane_f32(vo1p0, vi3x79BD, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4()
343 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x79BD, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4()
D5x5s2p2-minmax-neon-2x4-acc3.c178 const float32x4_t vi3x79BD = vextq_f32(vi3x1357, vi3x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc3() local
199 vo1p0 = vmlaq_lane_f32(vo1p0, vi3x79BD, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc3()
204 vo0p2 = vmlaq_lane_f32(vo0p2, vi3x79BD, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc3()
333 const float32x4_t vi3x79BD = vextq_f32(vi3x1357, vi3x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc3() local
342 vo1p0 = vmlaq_lane_f32(vo1p0, vi3x79BD, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc3()
347 vo0p2 = vmlaq_lane_f32(vo0p2, vi3x79BD, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc3()
D5x5s2p2-minmax-neonfma-2x4-acc3.c178 const float32x4_t vi3x79BD = vextq_f32(vi3x1357, vi3x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc3() local
199 vo1p0 = vfmaq_lane_f32(vo1p0, vi3x79BD, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc3()
204 vo0p2 = vfmaq_lane_f32(vo0p2, vi3x79BD, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc3()
333 const float32x4_t vi3x79BD = vextq_f32(vi3x1357, vi3x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc3() local
342 vo1p0 = vfmaq_lane_f32(vo1p0, vi3x79BD, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc3()
347 vo0p2 = vfmaq_lane_f32(vo0p2, vi3x79BD, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc3()
D5x5s2p2-minmax-neonfma-2x4.c178 const float32x4_t vi3x79BD = vextq_f32(vi3x1357, vi3x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4() local
199 vo1p0 = vfmaq_lane_f32(vo1p0, vi3x79BD, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4()
204 vo0p0 = vfmaq_lane_f32(vo0p0, vi3x79BD, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4()
329 const float32x4_t vi3x79BD = vextq_f32(vi3x1357, vi3x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4() local
338 vo1p0 = vfmaq_lane_f32(vo1p0, vi3x79BD, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4()
343 vo0p0 = vfmaq_lane_f32(vo0p0, vi3x79BD, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4()
D5x5s2p2-minmax-neon-1x4-acc5.c141 const float32x4_t vi3x79BD = vextq_f32(vi3x1357, vi3x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc5() local
158 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x79BD, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc5()
251 const float32x4_t vi3x79BD = vextq_f32(vi3x1357, vi3x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc5() local
260 vo0p0 = vmlaq_lane_f32(vo0p0, vi3x79BD, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc5()
D5x5s2p2-minmax-neonfma-1x4-acc5.c141 const float32x4_t vi3x79BD = vextq_f32(vi3x1357, vi3x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc5() local
158 vo0p0 = vfmaq_lane_f32(vo0p0, vi3x79BD, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc5()
251 const float32x4_t vi3x79BD = vextq_f32(vi3x1357, vi3x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc5() local
260 vo0p0 = vfmaq_lane_f32(vo0p0, vi3x79BD, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc5()
D5x5s2p2-minmax-wasmsimd-arm-loadsplat-2x4.c230 const v128_t vi3x79BD = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4() local
280 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x79BD, vk11)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4()
285 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x79BD, vk31)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4()
416 const v128_t vi3x79BD = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4() local
425 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x79BD, vk11)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4()
430 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x79BD, vk31)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4()
D5x5s2p2-minmax-wasmsimd-x86-loadsplat-2x4-acc2.c230 const v128_t vi3x79BD = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc2() local
280 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x79BD, vk11)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc2()
285 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x79BD, vk31)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc2()
418 const v128_t vi3x79BD = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc2() local
427 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x79BD, vk11)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc2()
432 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x79BD, vk31)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc2()
D5x5s2p2-minmax-wasmsimd-arm-loadsplat-2x4-acc2.c230 const v128_t vi3x79BD = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc2() local
280 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x79BD, vk11)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc2()
285 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x79BD, vk31)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc2()
418 const v128_t vi3x79BD = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc2() local
427 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x79BD, vk11)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc2()
432 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x79BD, vk31)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc2()
D5x5s2p2-minmax-wasmsimd-x86-loadsplat-2x4.c230 const v128_t vi3x79BD = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4() local
280 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x79BD, vk11)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4()
285 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x79BD, vk31)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4()
416 const v128_t vi3x79BD = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4() local
425 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x79BD, vk11)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4()
430 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x79BD, vk31)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4()
D5x5s2p2-minmax-wasmsimd-arm-loadsplat-2x4-acc3.c230 const v128_t vi3x79BD = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc3() local
280 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x79BD, vk11)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc3()
285 vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi3x79BD, vk31)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc3()
420 const v128_t vi3x79BD = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc3() local
429 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x79BD, vk11)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc3()
434 vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi3x79BD, vk31)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4_acc3()
D5x5s2p2-minmax-wasmsimd-x86-loadsplat-2x4-acc3.c230 const v128_t vi3x79BD = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc3() local
280 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x79BD, vk11)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc3()
285 vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi3x79BD, vk31)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc3()
420 const v128_t vi3x79BD = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc3() local
429 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x79BD, vk11)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc3()
434 vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi3x79BD, vk31)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_2x4_acc3()
D5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc3.c185 const v128_t vi3x79BD = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3() local
223 vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi3x79BD, vk31)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3()
318 const v128_t vi3x79BD = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3() local
327 vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi3x79BD, vk31)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3()
D5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc2.c185 const v128_t vi3x79BD = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2() local
223 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x79BD, vk31)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2()
317 const v128_t vi3x79BD = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2() local
326 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x79BD, vk31)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2()
D5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4.c185 const v128_t vi3x79BD = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4() local
223 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x79BD, vk31)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4()
316 const v128_t vi3x79BD = wasm_v32x4_shuffle(vi3x1357, vi3x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4() local
325 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi3x79BD, vk31)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4()

123