| /external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
| D | 5x5s2p2-minmax-neon-1x4-acc4.c | 143 const float32x4_t vi4x79BD = vextq_f32(vi4x1357, vi4x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc4() local
160 vo0p1 = vmlaq_lane_f32(vo0p1, vi4x79BD, vget_high_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc4()
251 const float32x4_t vi4x79BD = vextq_f32(vi4x1357, vi4x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc4() local
261 vo0p1 = vmlaq_lane_f32(vo0p1, vi4x79BD, vget_high_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc4()
|
| D | 5x5s2p2-minmax-neon-1x4-acc3.c | 143 const float32x4_t vi4x79BD = vextq_f32(vi4x1357, vi4x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc3() local
160 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x79BD, vget_high_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc3()
250 const float32x4_t vi4x79BD = vextq_f32(vi4x1357, vi4x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc3() local
260 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x79BD, vget_high_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc3()
|
| D | 5x5s2p2-minmax-neonfma-1x4-acc4.c | 143 const float32x4_t vi4x79BD = vextq_f32(vi4x1357, vi4x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc4() local
160 vo0p1 = vfmaq_lane_f32(vo0p1, vi4x79BD, vget_high_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc4()
251 const float32x4_t vi4x79BD = vextq_f32(vi4x1357, vi4x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc4() local
261 vo0p1 = vfmaq_lane_f32(vo0p1, vi4x79BD, vget_high_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc4()
|
| D | 5x5s2p2-minmax-neonfma-1x4-acc2.c | 143 const float32x4_t vi4x79BD = vextq_f32(vi4x1357, vi4x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2() local
160 vo0p1 = vfmaq_lane_f32(vo0p1, vi4x79BD, vget_high_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2()
249 const float32x4_t vi4x79BD = vextq_f32(vi4x1357, vi4x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2() local
259 vo0p1 = vfmaq_lane_f32(vo0p1, vi4x79BD, vget_high_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2()
|
| D | 5x5s2p2-minmax-neonfma-1x4-acc3.c | 143 const float32x4_t vi4x79BD = vextq_f32(vi4x1357, vi4x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc3() local
160 vo0p0 = vfmaq_lane_f32(vo0p0, vi4x79BD, vget_high_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc3()
250 const float32x4_t vi4x79BD = vextq_f32(vi4x1357, vi4x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc3() local
260 vo0p0 = vfmaq_lane_f32(vo0p0, vi4x79BD, vget_high_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc3()
|
| D | 5x5s2p2-minmax-neon-1x4.c | 143 const float32x4_t vi4x79BD = vextq_f32(vi4x1357, vi4x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4() local
160 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x79BD, vget_high_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
248 const float32x4_t vi4x79BD = vextq_f32(vi4x1357, vi4x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4() local
258 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x79BD, vget_high_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
|
| D | 5x5s2p2-minmax-neonfma-1x4.c | 143 const float32x4_t vi4x79BD = vextq_f32(vi4x1357, vi4x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4() local
160 vo0p0 = vfmaq_lane_f32(vo0p0, vi4x79BD, vget_high_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4()
248 const float32x4_t vi4x79BD = vextq_f32(vi4x1357, vi4x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4() local
258 vo0p0 = vfmaq_lane_f32(vo0p0, vi4x79BD, vget_high_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4()
|
| D | 5x5s2p2-minmax-neon-1x4-acc2.c | 143 const float32x4_t vi4x79BD = vextq_f32(vi4x1357, vi4x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc2() local
160 vo0p1 = vmlaq_lane_f32(vo0p1, vi4x79BD, vget_high_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc2()
249 const float32x4_t vi4x79BD = vextq_f32(vi4x1357, vi4x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc2() local
259 vo0p1 = vmlaq_lane_f32(vo0p1, vi4x79BD, vget_high_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc2()
|
| D | 5x5s2p2-minmax-neonfma-3x4-acc2.c | 216 const float32x4_t vi4x79BD = vextq_f32(vi4x1357, vi4x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_3x4_acc2() local
239 vo2p1 = vfmaq_lane_f32(vo2p1, vi4x79BD, vget_high_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_3x4_acc2()
246 vo1p1 = vfmaq_lane_f32(vo1p1, vi4x79BD, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_3x4_acc2()
253 vo0p1 = vfmaq_lane_f32(vo0p1, vi4x79BD, vget_high_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_3x4_acc2()
414 const float32x4_t vi4x79BD = vextq_f32(vi4x1357, vi4x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_3x4_acc2() local
422 vo2p1 = vfmaq_lane_f32(vo2p1, vi4x79BD, vget_high_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_3x4_acc2()
429 vo1p1 = vfmaq_lane_f32(vo1p1, vi4x79BD, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_3x4_acc2()
436 vo0p1 = vfmaq_lane_f32(vo0p1, vi4x79BD, vget_high_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_3x4_acc2()
|
| D | 5x5s2p2-minmax-neon-3x4.c | 216 const float32x4_t vi4x79BD = vextq_f32(vi4x1357, vi4x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_3x4() local
239 vo2p0 = vmlaq_lane_f32(vo2p0, vi4x79BD, vget_high_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_3x4()
246 vo1p0 = vmlaq_lane_f32(vo1p0, vi4x79BD, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_3x4()
253 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x79BD, vget_high_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_3x4()
411 const float32x4_t vi4x79BD = vextq_f32(vi4x1357, vi4x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_3x4() local
419 vo2p0 = vmlaq_lane_f32(vo2p0, vi4x79BD, vget_high_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_3x4()
426 vo1p0 = vmlaq_lane_f32(vo1p0, vi4x79BD, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_3x4()
433 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x79BD, vget_high_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_3x4()
|
| D | 5x5s2p2-minmax-neonfma-3x4.c | 216 const float32x4_t vi4x79BD = vextq_f32(vi4x1357, vi4x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_3x4() local
239 vo2p0 = vfmaq_lane_f32(vo2p0, vi4x79BD, vget_high_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_3x4()
246 vo1p0 = vfmaq_lane_f32(vo1p0, vi4x79BD, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_3x4()
253 vo0p0 = vfmaq_lane_f32(vo0p0, vi4x79BD, vget_high_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_3x4()
411 const float32x4_t vi4x79BD = vextq_f32(vi4x1357, vi4x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_3x4() local
419 vo2p0 = vfmaq_lane_f32(vo2p0, vi4x79BD, vget_high_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_3x4()
426 vo1p0 = vfmaq_lane_f32(vo1p0, vi4x79BD, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_3x4()
433 vo0p0 = vfmaq_lane_f32(vo0p0, vi4x79BD, vget_high_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_3x4()
|
| D | 5x5s2p2-minmax-neon-3x4-acc2.c | 216 const float32x4_t vi4x79BD = vextq_f32(vi4x1357, vi4x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_3x4_acc2() local
239 vo2p1 = vmlaq_lane_f32(vo2p1, vi4x79BD, vget_high_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_3x4_acc2()
246 vo1p1 = vmlaq_lane_f32(vo1p1, vi4x79BD, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_3x4_acc2()
253 vo0p1 = vmlaq_lane_f32(vo0p1, vi4x79BD, vget_high_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_3x4_acc2()
414 const float32x4_t vi4x79BD = vextq_f32(vi4x1357, vi4x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_3x4_acc2() local
422 vo2p1 = vmlaq_lane_f32(vo2p1, vi4x79BD, vget_high_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_3x4_acc2()
429 vo1p1 = vmlaq_lane_f32(vo1p1, vi4x79BD, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_3x4_acc2()
436 vo0p1 = vmlaq_lane_f32(vo0p1, vi4x79BD, vget_high_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_3x4_acc2()
|
| D | 5x5s2p2-minmax-neon-2x4-acc2.c | 180 const float32x4_t vi4x79BD = vextq_f32(vi4x1357, vi4x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc2() local
202 vo1p1 = vmlaq_lane_f32(vo1p1, vi4x79BD, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc2()
207 vo0p1 = vmlaq_lane_f32(vo0p1, vi4x79BD, vget_high_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc2()
332 const float32x4_t vi4x79BD = vextq_f32(vi4x1357, vi4x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc2() local
343 vo1p1 = vmlaq_lane_f32(vo1p1, vi4x79BD, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc2()
348 vo0p1 = vmlaq_lane_f32(vo0p1, vi4x79BD, vget_high_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc2()
|
| D | 5x5s2p2-minmax-neonfma-2x4-acc2.c | 180 const float32x4_t vi4x79BD = vextq_f32(vi4x1357, vi4x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc2() local
202 vo1p1 = vfmaq_lane_f32(vo1p1, vi4x79BD, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc2()
207 vo0p1 = vfmaq_lane_f32(vo0p1, vi4x79BD, vget_high_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc2()
332 const float32x4_t vi4x79BD = vextq_f32(vi4x1357, vi4x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc2() local
343 vo1p1 = vfmaq_lane_f32(vo1p1, vi4x79BD, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc2()
348 vo0p1 = vfmaq_lane_f32(vo0p1, vi4x79BD, vget_high_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc2()
|
| D | 5x5s2p2-minmax-neon-2x4.c | 180 const float32x4_t vi4x79BD = vextq_f32(vi4x1357, vi4x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4() local
202 vo1p0 = vmlaq_lane_f32(vo1p0, vi4x79BD, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4()
207 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x79BD, vget_high_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4()
330 const float32x4_t vi4x79BD = vextq_f32(vi4x1357, vi4x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4() local
341 vo1p0 = vmlaq_lane_f32(vo1p0, vi4x79BD, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4()
346 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x79BD, vget_high_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4()
|
| D | 5x5s2p2-minmax-neon-2x4-acc3.c | 180 const float32x4_t vi4x79BD = vextq_f32(vi4x1357, vi4x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc3() local
202 vo1p1 = vmlaq_lane_f32(vo1p1, vi4x79BD, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc3()
207 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x79BD, vget_high_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc3()
334 const float32x4_t vi4x79BD = vextq_f32(vi4x1357, vi4x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc3() local
345 vo1p1 = vmlaq_lane_f32(vo1p1, vi4x79BD, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc3()
350 vo0p0 = vmlaq_lane_f32(vo0p0, vi4x79BD, vget_high_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc3()
|
| D | 5x5s2p2-minmax-neonfma-2x4-acc3.c | 180 const float32x4_t vi4x79BD = vextq_f32(vi4x1357, vi4x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc3() local
202 vo1p1 = vfmaq_lane_f32(vo1p1, vi4x79BD, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc3()
207 vo0p0 = vfmaq_lane_f32(vo0p0, vi4x79BD, vget_high_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc3()
334 const float32x4_t vi4x79BD = vextq_f32(vi4x1357, vi4x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc3() local
345 vo1p1 = vfmaq_lane_f32(vo1p1, vi4x79BD, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc3()
350 vo0p0 = vfmaq_lane_f32(vo0p0, vi4x79BD, vget_high_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc3()
|
| D | 5x5s2p2-minmax-neonfma-2x4.c | 180 const float32x4_t vi4x79BD = vextq_f32(vi4x1357, vi4x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4() local
202 vo1p0 = vfmaq_lane_f32(vo1p0, vi4x79BD, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4()
207 vo0p0 = vfmaq_lane_f32(vo0p0, vi4x79BD, vget_high_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4()
330 const float32x4_t vi4x79BD = vextq_f32(vi4x1357, vi4x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4() local
341 vo1p0 = vfmaq_lane_f32(vo1p0, vi4x79BD, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4()
346 vo0p0 = vfmaq_lane_f32(vo0p0, vi4x79BD, vget_high_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4()
|
| D | 5x5s2p2-minmax-neon-1x4-acc5.c | 143 const float32x4_t vi4x79BD = vextq_f32(vi4x1357, vi4x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc5() local
160 vo0p1 = vmlaq_lane_f32(vo0p1, vi4x79BD, vget_high_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc5()
252 const float32x4_t vi4x79BD = vextq_f32(vi4x1357, vi4x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc5() local
262 vo0p1 = vmlaq_lane_f32(vo0p1, vi4x79BD, vget_high_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc5()
|
| D | 5x5s2p2-minmax-neonfma-1x4-acc5.c | 143 const float32x4_t vi4x79BD = vextq_f32(vi4x1357, vi4x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc5() local
160 vo0p1 = vfmaq_lane_f32(vo0p1, vi4x79BD, vget_high_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc5()
252 const float32x4_t vi4x79BD = vextq_f32(vi4x1357, vi4x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc5() local
262 vo0p1 = vfmaq_lane_f32(vo0p1, vi4x79BD, vget_high_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc5()
|
| D | 5x5s2p2-minmax-wasmsimd-x86-loadsplat-3x4-acc2.c | 276 const v128_t vi4x79BD = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_3x4_acc2() local
336 vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi4x79BD, vk01)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_3x4_acc2()
343 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi4x79BD, vk21)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_3x4_acc2()
350 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x79BD, vk41)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_3x4_acc2()
519 const v128_t vi4x79BD = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_3x4_acc2() local
527 vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi4x79BD, vk01)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_3x4_acc2()
534 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi4x79BD, vk21)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_3x4_acc2()
541 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x79BD, vk41)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_3x4_acc2()
|
| D | 5x5s2p2-minmax-wasmsimd-x86-loadsplat-3x4.c | 276 const v128_t vi4x79BD = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_3x4() local
336 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x79BD, vk01)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_3x4()
343 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x79BD, vk21)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_3x4()
350 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x79BD, vk41)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_3x4()
516 const v128_t vi4x79BD = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_3x4() local
524 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x79BD, vk01)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_3x4()
531 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x79BD, vk21)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_3x4()
538 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x79BD, vk41)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_3x4()
|
| D | 5x5s2p2-minmax-wasmsimd-arm-loadsplat-3x4-acc2.c | 276 const v128_t vi4x79BD = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_3x4_acc2() local
336 vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi4x79BD, vk01)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_3x4_acc2()
343 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi4x79BD, vk21)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_3x4_acc2()
350 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x79BD, vk41)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_3x4_acc2()
519 const v128_t vi4x79BD = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_3x4_acc2() local
527 vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi4x79BD, vk01)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_3x4_acc2()
534 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi4x79BD, vk21)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_3x4_acc2()
541 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x79BD, vk41)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_3x4_acc2()
|
| D | 5x5s2p2-minmax-wasmsimd-arm-loadsplat-3x4.c | 276 const v128_t vi4x79BD = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_3x4() local
336 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x79BD, vk01)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_3x4()
343 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x79BD, vk21)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_3x4()
350 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x79BD, vk41)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_3x4()
516 const v128_t vi4x79BD = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_3x4() local
524 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x79BD, vk01)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_3x4()
531 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x79BD, vk21)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_3x4()
538 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x79BD, vk41)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_3x4()
|
| D | 5x5s2p2-minmax-wasmsimd-arm-loadsplat-2x4.c | 232 const v128_t vi4x79BD = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4() local
283 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x79BD, vk21)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4()
288 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x79BD, vk41)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4()
417 const v128_t vi4x79BD = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4() local
428 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x79BD, vk21)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4()
433 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x79BD, vk41)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4()
|