| /external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
| D | 5x5p2-minmax-neon-4x4.c | 259 const float32x4_t vi4x6789 = vextq_f32(vi4x4567, vi4x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4() local
276 vo3p0 = vmlaq_lane_f32(vo3p0, vi4x6789, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()
280 vo2p0 = vmlaq_lane_f32(vo2p0, vi4x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()
284 vo1p0 = vmlaq_lane_f32(vo1p0, vi4x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()
288 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()
477 const float32x4_t vi4x6789 = vextq_f32(vi4x4567, vi4x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4() local
494 vo3p0 = vmlaq_lane_f32(vo3p0, vi4x6789, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()
498 vo2p0 = vmlaq_lane_f32(vo2p0, vi4x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()
502 vo1p0 = vmlaq_lane_f32(vo1p0, vi4x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()
506 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()
[all …]
|
| D | 5x5p2-minmax-neon-4x4-acc2.c | 259 const float32x4_t vi4x6789 = vextq_f32(vi4x4567, vi4x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4_acc2() local
276 vo3p1 = vmlaq_lane_f32(vo3p1, vi4x6789, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4_acc2()
280 vo2p0 = vmlaq_lane_f32(vo2p0, vi4x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4_acc2()
284 vo1p1 = vmlaq_lane_f32(vo1p1, vi4x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4_acc2()
288 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4_acc2()
481 const float32x4_t vi4x6789 = vextq_f32(vi4x4567, vi4x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4_acc2() local
498 vo3p1 = vmlaq_lane_f32(vo3p1, vi4x6789, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4_acc2()
502 vo2p0 = vmlaq_lane_f32(vo2p0, vi4x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4_acc2()
506 vo1p1 = vmlaq_lane_f32(vo1p1, vi4x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4_acc2()
510 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4_acc2()
[all …]
|
| D | 5x5p2-minmax-neonfma-4x4.c | 259 const float32x4_t vi4x6789 = vextq_f32(vi4x4567, vi4x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4() local
276 vo3p0 = vfmaq_lane_f32(vo3p0, vi4x6789, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4()
280 vo2p0 = vfmaq_lane_f32(vo2p0, vi4x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4()
284 vo1p0 = vfmaq_lane_f32(vo1p0, vi4x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4()
288 vo0p0 = vfmaq_lane_f32(vo0p0, vi4x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4()
477 const float32x4_t vi4x6789 = vextq_f32(vi4x4567, vi4x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4() local
494 vo3p0 = vfmaq_lane_f32(vo3p0, vi4x6789, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4()
498 vo2p0 = vfmaq_lane_f32(vo2p0, vi4x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4()
502 vo1p0 = vfmaq_lane_f32(vo1p0, vi4x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4()
506 vo0p0 = vfmaq_lane_f32(vo0p0, vi4x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4()
[all …]
|
| D | 5x5p2-minmax-neonfma-4x4-acc2.c | 259 const float32x4_t vi4x6789 = vextq_f32(vi4x4567, vi4x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4_acc2() local
276 vo3p1 = vfmaq_lane_f32(vo3p1, vi4x6789, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4_acc2()
280 vo2p0 = vfmaq_lane_f32(vo2p0, vi4x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4_acc2()
284 vo1p1 = vfmaq_lane_f32(vo1p1, vi4x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4_acc2()
288 vo0p0 = vfmaq_lane_f32(vo0p0, vi4x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4_acc2()
481 const float32x4_t vi4x6789 = vextq_f32(vi4x4567, vi4x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4_acc2() local
498 vo3p1 = vfmaq_lane_f32(vo3p1, vi4x6789, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4_acc2()
502 vo2p0 = vfmaq_lane_f32(vo2p0, vi4x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4_acc2()
506 vo1p1 = vfmaq_lane_f32(vo1p1, vi4x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4_acc2()
510 vo0p0 = vfmaq_lane_f32(vo0p0, vi4x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4_acc2()
[all …]
|
| D | 5x5p2-minmax-neonfma-1x4-acc2.c | 157 const float32x4_t vi4x6789 = vextq_f32(vi4x4567, vi4x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2() local
168 vo0p0 = vfmaq_lane_f32(vo0p0, vi4x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2()
265 const float32x4_t vi4x6789 = vextq_f32(vi4x4567, vi4x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2() local
276 vo0p0 = vfmaq_lane_f32(vo0p0, vi4x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2()
362 const float32x4_t vi4x6789 = vextq_f32(vi4x5678, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2() local
372 vo0p0 = vfmaq_lane_f32(vo0p0, vi4x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc2()
|
| D | 5x5p2-minmax-neonfma-1x4-acc3.c | 157 const float32x4_t vi4x6789 = vextq_f32(vi4x4567, vi4x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3() local
168 vo0p2 = vfmaq_lane_f32(vo0p2, vi4x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3()
266 const float32x4_t vi4x6789 = vextq_f32(vi4x4567, vi4x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3() local
277 vo0p2 = vfmaq_lane_f32(vo0p2, vi4x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3()
364 const float32x4_t vi4x6789 = vextq_f32(vi4x5678, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3() local
374 vo0p2 = vfmaq_lane_f32(vo0p2, vi4x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4_acc3()
|
| D | 5x5p2-minmax-neon-1x4.c | 157 const float32x4_t vi4x6789 = vextq_f32(vi4x4567, vi4x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4() local
168 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
264 const float32x4_t vi4x6789 = vextq_f32(vi4x4567, vi4x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4() local
275 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
360 const float32x4_t vi4x6789 = vextq_f32(vi4x5678, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4() local
370 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4()
|
| D | 5x5p2-minmax-neonfma-1x4.c | 157 const float32x4_t vi4x6789 = vextq_f32(vi4x4567, vi4x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4() local
168 vo0p0 = vfmaq_lane_f32(vo0p0, vi4x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4()
264 const float32x4_t vi4x6789 = vextq_f32(vi4x4567, vi4x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4() local
275 vo0p0 = vfmaq_lane_f32(vo0p0, vi4x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4()
360 const float32x4_t vi4x6789 = vextq_f32(vi4x5678, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4() local
370 vo0p0 = vfmaq_lane_f32(vo0p0, vi4x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_1x4()
|
| D | 5x5p2-minmax-neon-1x4-acc3.c | 157 const float32x4_t vi4x6789 = vextq_f32(vi4x4567, vi4x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3() local
168 vo0p2 = vmlaq_lane_f32(vo0p2, vi4x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3()
266 const float32x4_t vi4x6789 = vextq_f32(vi4x4567, vi4x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3() local
277 vo0p2 = vmlaq_lane_f32(vo0p2, vi4x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3()
364 const float32x4_t vi4x6789 = vextq_f32(vi4x5678, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3() local
374 vo0p2 = vmlaq_lane_f32(vo0p2, vi4x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc3()
|
| D | 5x5p2-minmax-neon-1x4-acc2.c | 157 const float32x4_t vi4x6789 = vextq_f32(vi4x4567, vi4x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2() local
168 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2()
265 const float32x4_t vi4x6789 = vextq_f32(vi4x4567, vi4x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2() local
276 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2()
362 const float32x4_t vi4x6789 = vextq_f32(vi4x5678, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2() local
372 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4_acc2()
|
| D | 5x5p2-minmax-neon-2x4-acc2.c | 191 const float32x4_t vi4x6789 = vextq_f32(vi4x4567, vi4x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2() local
206 vo1p1 = vmlaq_lane_f32(vo1p1, vi4x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2()
208 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2()
337 const float32x4_t vi4x6789 = vextq_f32(vi4x4567, vi4x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2() local
352 vo1p1 = vmlaq_lane_f32(vo1p1, vi4x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2()
354 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2()
470 const float32x4_t vi4x6789 = vextq_f32(vi4x5678, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2() local
483 vo1p1 = vmlaq_lane_f32(vo1p1, vi4x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2()
485 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4_acc2()
|
| D | 5x5p2-minmax-neonfma-2x4.c | 191 const float32x4_t vi4x6789 = vextq_f32(vi4x4567, vi4x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4() local
206 vo1p0 = vfmaq_lane_f32(vo1p0, vi4x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4()
208 vo0p0 = vfmaq_lane_f32(vo0p0, vi4x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4()
335 const float32x4_t vi4x6789 = vextq_f32(vi4x4567, vi4x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4() local
350 vo1p0 = vfmaq_lane_f32(vo1p0, vi4x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4()
352 vo0p0 = vfmaq_lane_f32(vo0p0, vi4x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4()
466 const float32x4_t vi4x6789 = vextq_f32(vi4x5678, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4() local
479 vo1p0 = vfmaq_lane_f32(vo1p0, vi4x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4()
481 vo0p0 = vfmaq_lane_f32(vo0p0, vi4x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4()
|
| D | 5x5p2-minmax-neon-2x4.c | 191 const float32x4_t vi4x6789 = vextq_f32(vi4x4567, vi4x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4() local
206 vo1p0 = vmlaq_lane_f32(vo1p0, vi4x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
208 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
335 const float32x4_t vi4x6789 = vextq_f32(vi4x4567, vi4x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4() local
350 vo1p0 = vmlaq_lane_f32(vo1p0, vi4x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
352 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
466 const float32x4_t vi4x6789 = vextq_f32(vi4x5678, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4() local
479 vo1p0 = vmlaq_lane_f32(vo1p0, vi4x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
481 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_2x4()
|
| D | 5x5p2-minmax-neonfma-2x4-acc2.c | 191 const float32x4_t vi4x6789 = vextq_f32(vi4x4567, vi4x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc2() local
206 vo1p1 = vfmaq_lane_f32(vo1p1, vi4x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc2()
208 vo0p0 = vfmaq_lane_f32(vo0p0, vi4x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc2()
337 const float32x4_t vi4x6789 = vextq_f32(vi4x4567, vi4x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc2() local
352 vo1p1 = vfmaq_lane_f32(vo1p1, vi4x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc2()
354 vo0p0 = vfmaq_lane_f32(vo0p0, vi4x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc2()
470 const float32x4_t vi4x6789 = vextq_f32(vi4x5678, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc2() local
483 vo1p1 = vfmaq_lane_f32(vo1p1, vi4x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc2()
485 vo0p0 = vfmaq_lane_f32(vo0p0, vi4x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_2x4_acc2()
|
| D | 5x5p2-minmax-neonfma-3x4.c | 225 const float32x4_t vi4x6789 = vextq_f32(vi4x4567, vi4x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4() local
242 vo2p0 = vfmaq_lane_f32(vo2p0, vi4x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4()
245 vo1p0 = vfmaq_lane_f32(vo1p0, vi4x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4()
248 vo0p0 = vfmaq_lane_f32(vo0p0, vi4x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4()
406 const float32x4_t vi4x6789 = vextq_f32(vi4x4567, vi4x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4() local
423 vo2p0 = vfmaq_lane_f32(vo2p0, vi4x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4()
426 vo1p0 = vfmaq_lane_f32(vo1p0, vi4x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4()
429 vo0p0 = vfmaq_lane_f32(vo0p0, vi4x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4()
572 const float32x4_t vi4x6789 = vextq_f32(vi4x5678, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4() local
586 vo2p0 = vfmaq_lane_f32(vo2p0, vi4x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4()
[all …]
|
| D | 5x5p2-minmax-wasmsimd-arm-loadsplat-3x4.c | 254 const v128_t vi4x6789 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4() local
271 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x6789, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4()
274 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x6789, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4()
277 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x6789, vk44)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4()
434 const v128_t vi4x6789 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4() local
451 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x6789, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4()
454 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x6789, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4()
457 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x6789, vk44)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4()
598 const v128_t vi4x6789 = wasm_v32x4_shuffle(vi4x5678, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4() local
612 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x6789, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4()
[all …]
|
| D | 5x5p2-minmax-wasmsimd-arm-loadsplat-3x4-acc2.c | 254 const v128_t vi4x6789 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4_acc2() local
271 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x6789, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4_acc2()
274 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi4x6789, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4_acc2()
277 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x6789, vk44)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4_acc2()
437 const v128_t vi4x6789 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4_acc2() local
454 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x6789, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4_acc2()
457 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi4x6789, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4_acc2()
460 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x6789, vk44)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4_acc2()
604 const v128_t vi4x6789 = wasm_v32x4_shuffle(vi4x5678, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4_acc2() local
618 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x6789, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4_acc2()
[all …]
|
| D | 5x5p2-minmax-wasmsimd-x86-loadsplat-3x4.c | 254 const v128_t vi4x6789 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4() local
271 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x6789, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
274 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x6789, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
277 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x6789, vk44)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
434 const v128_t vi4x6789 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4() local
451 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x6789, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
454 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x6789, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
457 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x6789, vk44)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
598 const v128_t vi4x6789 = wasm_v32x4_shuffle(vi4x5678, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4() local
612 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x6789, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
[all …]
|
| D | 5x5p2-minmax-neon-3x4-acc2.c | 225 const float32x4_t vi4x6789 = vextq_f32(vi4x4567, vi4x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4_acc2() local
242 vo2p0 = vmlaq_lane_f32(vo2p0, vi4x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4_acc2()
245 vo1p1 = vmlaq_lane_f32(vo1p1, vi4x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4_acc2()
248 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4_acc2()
409 const float32x4_t vi4x6789 = vextq_f32(vi4x4567, vi4x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4_acc2() local
426 vo2p0 = vmlaq_lane_f32(vo2p0, vi4x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4_acc2()
429 vo1p1 = vmlaq_lane_f32(vo1p1, vi4x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4_acc2()
432 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4_acc2()
578 const float32x4_t vi4x6789 = vextq_f32(vi4x5678, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4_acc2() local
592 vo2p0 = vmlaq_lane_f32(vo2p0, vi4x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4_acc2()
[all …]
|
| D | 5x5p2-minmax-neon-3x4.c | 225 const float32x4_t vi4x6789 = vextq_f32(vi4x4567, vi4x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4() local
242 vo2p0 = vmlaq_lane_f32(vo2p0, vi4x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
245 vo1p0 = vmlaq_lane_f32(vo1p0, vi4x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
248 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
406 const float32x4_t vi4x6789 = vextq_f32(vi4x4567, vi4x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4() local
423 vo2p0 = vmlaq_lane_f32(vo2p0, vi4x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
426 vo1p0 = vmlaq_lane_f32(vo1p0, vi4x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
429 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
572 const float32x4_t vi4x6789 = vextq_f32(vi4x5678, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4() local
586 vo2p0 = vmlaq_lane_f32(vo2p0, vi4x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4()
[all …]
|
| D | 5x5p2-minmax-neonfma-3x4-acc2.c | 225 const float32x4_t vi4x6789 = vextq_f32(vi4x4567, vi4x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2() local
242 vo2p0 = vfmaq_lane_f32(vo2p0, vi4x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2()
245 vo1p1 = vfmaq_lane_f32(vo1p1, vi4x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2()
248 vo0p0 = vfmaq_lane_f32(vo0p0, vi4x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2()
409 const float32x4_t vi4x6789 = vextq_f32(vi4x4567, vi4x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2() local
426 vo2p0 = vfmaq_lane_f32(vo2p0, vi4x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2()
429 vo1p1 = vfmaq_lane_f32(vo1p1, vi4x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2()
432 vo0p0 = vfmaq_lane_f32(vo0p0, vi4x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2()
578 const float32x4_t vi4x6789 = vextq_f32(vi4x5678, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2() local
592 vo2p0 = vfmaq_lane_f32(vo2p0, vi4x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_3x4_acc2()
[all …]
|
| D | 5x5p2-minmax-wasmsimd-x86-loadsplat-5x4.c | 322 const v128_t vi4x6789 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4() local
337 vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi4x6789, vk04)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
342 vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x6789, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
347 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x6789, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
352 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x6789, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
357 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x6789, vk44)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
576 const v128_t vi4x6789 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4() local
591 vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi4x6789, vk04)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
596 vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x6789, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
601 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x6789, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
[all …]
|
| D | 5x5p2-minmax-neonfma-5x4.c | 293 const float32x4_t vi4x6789 = vextq_f32(vi4x4567, vi4x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4() local
308 vo4p0 = vfmaq_lane_f32(vo4p0, vi4x6789, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4()
313 vo3p0 = vfmaq_lane_f32(vo3p0, vi4x6789, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4()
318 vo2p0 = vfmaq_lane_f32(vo2p0, vi4x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4()
323 vo1p0 = vfmaq_lane_f32(vo1p0, vi4x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4()
328 vo0p0 = vfmaq_lane_f32(vo0p0, vi4x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4()
548 const float32x4_t vi4x6789 = vextq_f32(vi4x4567, vi4x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4() local
563 vo4p0 = vfmaq_lane_f32(vo4p0, vi4x6789, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4()
568 vo3p0 = vfmaq_lane_f32(vo3p0, vi4x6789, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4()
573 vo2p0 = vfmaq_lane_f32(vo2p0, vi4x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4()
[all …]
|
| D | 5x5p2-minmax-neon-5x4.c | 293 const float32x4_t vi4x6789 = vextq_f32(vi4x4567, vi4x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4() local
308 vo4p0 = vmlaq_lane_f32(vo4p0, vi4x6789, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
313 vo3p0 = vmlaq_lane_f32(vo3p0, vi4x6789, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
318 vo2p0 = vmlaq_lane_f32(vo2p0, vi4x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
323 vo1p0 = vmlaq_lane_f32(vo1p0, vi4x6789, vget_low_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
328 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x6789, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
548 const float32x4_t vi4x6789 = vextq_f32(vi4x4567, vi4x89AB, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4() local
563 vo4p0 = vmlaq_lane_f32(vo4p0, vi4x6789, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
568 vo3p0 = vmlaq_lane_f32(vo3p0, vi4x6789, vget_high_f32(vw89AB), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
573 vo2p0 = vmlaq_lane_f32(vo2p0, vi4x6789, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
[all …]
|
| D | 5x5p2-minmax-wasmsimd-arm-loadsplat-5x4.c | 322 const v128_t vi4x6789 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4() local
337 vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi4x6789, vk04)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4()
342 vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x6789, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4()
347 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x6789, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4()
352 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x6789, vk34)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4()
357 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x6789, vk44)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4()
576 const v128_t vi4x6789 = wasm_v32x4_shuffle(vi4x4567, vi4x89AB, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4() local
591 vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi4x6789, vk04)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4()
596 vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x6789, vk14)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4()
601 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x6789, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4()
[all …]
|