| /external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
| D | 5x5s2p2-minmax-neon-1x4-acc4.c | 170 const float32x4_t vi4xACEG = vextq_f32(vi4x8ACE9BDF.val[0], vi4xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc4() local
181 vo0p2 = vmlaq_lane_f32(vo0p2, vi4xACEG, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc4()
268 const float32x4_t vi4xACEG = vextq_f32(vi4x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc4() local
278 vo0p2 = vmlaq_lane_f32(vo0p2, vi4xACEG, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc4()
|
| D | 5x5s2p2-minmax-neon-1x4-acc3.c | 170 const float32x4_t vi4xACEG = vextq_f32(vi4x8ACE9BDF.val[0], vi4xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc3() local
181 vo0p2 = vmlaq_lane_f32(vo0p2, vi4xACEG, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc3()
267 const float32x4_t vi4xACEG = vextq_f32(vi4x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc3() local
277 vo0p2 = vmlaq_lane_f32(vo0p2, vi4xACEG, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc3()
|
| D | 5x5s2p2-minmax-neonfma-1x4-acc4.c | 170 const float32x4_t vi4xACEG = vextq_f32(vi4x8ACE9BDF.val[0], vi4xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc4() local
181 vo0p2 = vfmaq_lane_f32(vo0p2, vi4xACEG, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc4()
268 const float32x4_t vi4xACEG = vextq_f32(vi4x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc4() local
278 vo0p2 = vfmaq_lane_f32(vo0p2, vi4xACEG, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc4()
|
| D | 5x5s2p2-minmax-neonfma-1x4-acc2.c | 170 const float32x4_t vi4xACEG = vextq_f32(vi4x8ACE9BDF.val[0], vi4xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2() local
181 vo0p0 = vfmaq_lane_f32(vo0p0, vi4xACEG, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2()
266 const float32x4_t vi4xACEG = vextq_f32(vi4x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2() local
276 vo0p0 = vfmaq_lane_f32(vo0p0, vi4xACEG, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2()
|
| D | 5x5s2p2-minmax-neonfma-1x4-acc3.c | 170 const float32x4_t vi4xACEG = vextq_f32(vi4x8ACE9BDF.val[0], vi4xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc3() local
181 vo0p2 = vfmaq_lane_f32(vo0p2, vi4xACEG, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc3()
267 const float32x4_t vi4xACEG = vextq_f32(vi4x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc3() local
277 vo0p2 = vfmaq_lane_f32(vo0p2, vi4xACEG, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc3()
|
| D | 5x5s2p2-minmax-neon-1x4.c | 170 const float32x4_t vi4xACEG = vextq_f32(vi4x8ACE9BDF.val[0], vi4xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4() local
181 vo0p0 = vmlaq_lane_f32(vo0p0, vi4xACEG, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
265 const float32x4_t vi4xACEG = vextq_f32(vi4x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4() local
275 vo0p0 = vmlaq_lane_f32(vo0p0, vi4xACEG, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4()
|
| D | 5x5s2p2-minmax-neonfma-1x4.c | 170 const float32x4_t vi4xACEG = vextq_f32(vi4x8ACE9BDF.val[0], vi4xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4() local
181 vo0p0 = vfmaq_lane_f32(vo0p0, vi4xACEG, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4()
265 const float32x4_t vi4xACEG = vextq_f32(vi4x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4() local
275 vo0p0 = vfmaq_lane_f32(vo0p0, vi4xACEG, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4()
|
| D | 5x5s2p2-minmax-neon-1x4-acc2.c | 170 const float32x4_t vi4xACEG = vextq_f32(vi4x8ACE9BDF.val[0], vi4xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc2() local
181 vo0p0 = vmlaq_lane_f32(vo0p0, vi4xACEG, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc2()
266 const float32x4_t vi4xACEG = vextq_f32(vi4x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc2() local
276 vo0p0 = vmlaq_lane_f32(vo0p0, vi4xACEG, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc2()
|
| D | 5x5s2p2-minmax-neonfma-3x4-acc2.c | 265 const float32x4_t vi4xACEG = vextq_f32(vi4x8ACE9BDF.val[0], vi4xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_3x4_acc2() local
278 vo2p0 = vfmaq_lane_f32(vo2p0, vi4xACEG, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_3x4_acc2()
285 vo1p0 = vfmaq_lane_f32(vo1p0, vi4xACEG, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_3x4_acc2()
292 vo0p0 = vfmaq_lane_f32(vo0p0, vi4xACEG, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_3x4_acc2()
445 const float32x4_t vi4xACEG = vextq_f32(vi4x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_3x4_acc2() local
453 vo2p0 = vfmaq_lane_f32(vo2p0, vi4xACEG, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_3x4_acc2()
460 vo1p0 = vfmaq_lane_f32(vo1p0, vi4xACEG, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_3x4_acc2()
467 vo0p0 = vfmaq_lane_f32(vo0p0, vi4xACEG, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_3x4_acc2()
|
| D | 5x5s2p2-minmax-neon-3x4.c | 265 const float32x4_t vi4xACEG = vextq_f32(vi4x8ACE9BDF.val[0], vi4xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_3x4() local
278 vo2p0 = vmlaq_lane_f32(vo2p0, vi4xACEG, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_3x4()
285 vo1p0 = vmlaq_lane_f32(vo1p0, vi4xACEG, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_3x4()
292 vo0p0 = vmlaq_lane_f32(vo0p0, vi4xACEG, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_3x4()
442 const float32x4_t vi4xACEG = vextq_f32(vi4x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_3x4() local
450 vo2p0 = vmlaq_lane_f32(vo2p0, vi4xACEG, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_3x4()
457 vo1p0 = vmlaq_lane_f32(vo1p0, vi4xACEG, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_3x4()
464 vo0p0 = vmlaq_lane_f32(vo0p0, vi4xACEG, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_3x4()
|
| D | 5x5s2p2-minmax-neonfma-3x4.c | 265 const float32x4_t vi4xACEG = vextq_f32(vi4x8ACE9BDF.val[0], vi4xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_3x4() local
278 vo2p0 = vfmaq_lane_f32(vo2p0, vi4xACEG, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_3x4()
285 vo1p0 = vfmaq_lane_f32(vo1p0, vi4xACEG, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_3x4()
292 vo0p0 = vfmaq_lane_f32(vo0p0, vi4xACEG, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_3x4()
442 const float32x4_t vi4xACEG = vextq_f32(vi4x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_3x4() local
450 vo2p0 = vfmaq_lane_f32(vo2p0, vi4xACEG, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_3x4()
457 vo1p0 = vfmaq_lane_f32(vo1p0, vi4xACEG, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_3x4()
464 vo0p0 = vfmaq_lane_f32(vo0p0, vi4xACEG, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_3x4()
|
| D | 5x5s2p2-minmax-neon-3x4-acc2.c | 265 const float32x4_t vi4xACEG = vextq_f32(vi4x8ACE9BDF.val[0], vi4xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_3x4_acc2() local
278 vo2p0 = vmlaq_lane_f32(vo2p0, vi4xACEG, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_3x4_acc2()
285 vo1p0 = vmlaq_lane_f32(vo1p0, vi4xACEG, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_3x4_acc2()
292 vo0p0 = vmlaq_lane_f32(vo0p0, vi4xACEG, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_3x4_acc2()
445 const float32x4_t vi4xACEG = vextq_f32(vi4x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_3x4_acc2() local
453 vo2p0 = vmlaq_lane_f32(vo2p0, vi4xACEG, vget_low_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_3x4_acc2()
460 vo1p0 = vmlaq_lane_f32(vo1p0, vi4xACEG, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_3x4_acc2()
467 vo0p0 = vmlaq_lane_f32(vo0p0, vi4xACEG, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_3x4_acc2()
|
| D | 5x5s2p2-minmax-neon-2x4-acc2.c | 218 const float32x4_t vi4xACEG = vextq_f32(vi4x8ACE9BDF.val[0], vi4xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc2() local
232 vo1p0 = vmlaq_lane_f32(vo1p0, vi4xACEG, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc2()
237 vo0p0 = vmlaq_lane_f32(vo0p0, vi4xACEG, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc2()
356 const float32x4_t vi4xACEG = vextq_f32(vi4x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc2() local
367 vo1p0 = vmlaq_lane_f32(vo1p0, vi4xACEG, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc2()
372 vo0p0 = vmlaq_lane_f32(vo0p0, vi4xACEG, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc2()
|
| D | 5x5s2p2-minmax-neonfma-2x4-acc2.c | 218 const float32x4_t vi4xACEG = vextq_f32(vi4x8ACE9BDF.val[0], vi4xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc2() local
232 vo1p0 = vfmaq_lane_f32(vo1p0, vi4xACEG, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc2()
237 vo0p0 = vfmaq_lane_f32(vo0p0, vi4xACEG, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc2()
356 const float32x4_t vi4xACEG = vextq_f32(vi4x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc2() local
367 vo1p0 = vfmaq_lane_f32(vo1p0, vi4xACEG, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc2()
372 vo0p0 = vfmaq_lane_f32(vo0p0, vi4xACEG, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc2()
|
| D | 5x5s2p2-minmax-neon-2x4.c | 218 const float32x4_t vi4xACEG = vextq_f32(vi4x8ACE9BDF.val[0], vi4xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4() local
232 vo1p0 = vmlaq_lane_f32(vo1p0, vi4xACEG, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4()
237 vo0p0 = vmlaq_lane_f32(vo0p0, vi4xACEG, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4()
354 const float32x4_t vi4xACEG = vextq_f32(vi4x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4() local
365 vo1p0 = vmlaq_lane_f32(vo1p0, vi4xACEG, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4()
370 vo0p0 = vmlaq_lane_f32(vo0p0, vi4xACEG, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4()
|
| D | 5x5s2p2-minmax-neon-2x4-acc3.c | 218 const float32x4_t vi4xACEG = vextq_f32(vi4x8ACE9BDF.val[0], vi4xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc3() local
232 vo1p0 = vmlaq_lane_f32(vo1p0, vi4xACEG, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc3()
237 vo0p2 = vmlaq_lane_f32(vo0p2, vi4xACEG, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc3()
358 const float32x4_t vi4xACEG = vextq_f32(vi4x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc3() local
369 vo1p0 = vmlaq_lane_f32(vo1p0, vi4xACEG, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc3()
374 vo0p2 = vmlaq_lane_f32(vo0p2, vi4xACEG, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_2x4_acc3()
|
| D | 5x5s2p2-minmax-neonfma-2x4-acc3.c | 218 const float32x4_t vi4xACEG = vextq_f32(vi4x8ACE9BDF.val[0], vi4xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc3() local
232 vo1p0 = vfmaq_lane_f32(vo1p0, vi4xACEG, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc3()
237 vo0p2 = vfmaq_lane_f32(vo0p2, vi4xACEG, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc3()
358 const float32x4_t vi4xACEG = vextq_f32(vi4x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc3() local
369 vo1p0 = vfmaq_lane_f32(vo1p0, vi4xACEG, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc3()
374 vo0p2 = vfmaq_lane_f32(vo0p2, vi4xACEG, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4_acc3()
|
| D | 5x5s2p2-minmax-neonfma-2x4.c | 218 const float32x4_t vi4xACEG = vextq_f32(vi4x8ACE9BDF.val[0], vi4xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4() local
232 vo1p0 = vfmaq_lane_f32(vo1p0, vi4xACEG, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4()
237 vo0p0 = vfmaq_lane_f32(vo0p0, vi4xACEG, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4()
354 const float32x4_t vi4xACEG = vextq_f32(vi4x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4() local
365 vo1p0 = vfmaq_lane_f32(vo1p0, vi4xACEG, vget_high_f32(vwCDEF), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4()
370 vo0p0 = vfmaq_lane_f32(vo0p0, vi4xACEG, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_2x4()
|
| D | 5x5s2p2-minmax-neon-1x4-acc5.c | 170 const float32x4_t vi4xACEG = vextq_f32(vi4x8ACE9BDF.val[0], vi4xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc5() local
181 vo0p1 = vmlaq_lane_f32(vo0p1, vi4xACEG, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc5()
269 const float32x4_t vi4xACEG = vextq_f32(vi4x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc5() local
279 vo0p1 = vmlaq_lane_f32(vo0p1, vi4xACEG, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4_acc5()
|
| D | 5x5s2p2-minmax-neonfma-1x4-acc5.c | 170 const float32x4_t vi4xACEG = vextq_f32(vi4x8ACE9BDF.val[0], vi4xGIKMHJLN.val[0], 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc5() local
181 vo0p1 = vfmaq_lane_f32(vo0p1, vi4xACEG, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc5()
269 const float32x4_t vi4xACEG = vextq_f32(vi4x8ACE, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc5() local
279 vo0p1 = vfmaq_lane_f32(vo0p1, vi4xACEG, vwOP, 1); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc5()
|
| D | 5x5s2p2-minmax-wasmsimd-x86-loadsplat-3x4-acc2.c | 366 const v128_t vi4xACEG = wasm_v32x4_shuffle(vi4x8ACE, vi4xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_3x4_acc2() local
384 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4xACEG, vk04)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_3x4_acc2()
391 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4xACEG, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_3x4_acc2()
398 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4xACEG, vk44)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_3x4_acc2()
549 const v128_t vi4xACEG = wasm_v32x4_shuffle(vi4x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_3x4_acc2() local
557 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4xACEG, vk04)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_3x4_acc2()
564 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4xACEG, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_3x4_acc2()
571 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4xACEG, vk44)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_3x4_acc2()
|
| D | 5x5s2p2-minmax-wasmsimd-x86-loadsplat-3x4.c | 366 const v128_t vi4xACEG = wasm_v32x4_shuffle(vi4x8ACE, vi4xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_3x4() local
384 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4xACEG, vk04)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_3x4()
391 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4xACEG, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_3x4()
398 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4xACEG, vk44)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_3x4()
546 const v128_t vi4xACEG = wasm_v32x4_shuffle(vi4x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_3x4() local
554 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4xACEG, vk04)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_3x4()
561 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4xACEG, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_3x4()
568 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4xACEG, vk44)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_3x4()
|
| D | 5x5s2p2-minmax-wasmsimd-arm-loadsplat-3x4-acc2.c | 366 const v128_t vi4xACEG = wasm_v32x4_shuffle(vi4x8ACE, vi4xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_3x4_acc2() local
384 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4xACEG, vk04)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_3x4_acc2()
391 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4xACEG, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_3x4_acc2()
398 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4xACEG, vk44)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_3x4_acc2()
549 const v128_t vi4xACEG = wasm_v32x4_shuffle(vi4x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_3x4_acc2() local
557 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4xACEG, vk04)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_3x4_acc2()
564 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4xACEG, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_3x4_acc2()
571 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4xACEG, vk44)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_3x4_acc2()
|
| D | 5x5s2p2-minmax-wasmsimd-arm-loadsplat-3x4.c | 366 const v128_t vi4xACEG = wasm_v32x4_shuffle(vi4x8ACE, vi4xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_3x4() local
384 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4xACEG, vk04)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_3x4()
391 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4xACEG, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_3x4()
398 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4xACEG, vk44)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_3x4()
546 const v128_t vi4xACEG = wasm_v32x4_shuffle(vi4x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_3x4() local
554 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4xACEG, vk04)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_3x4()
561 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4xACEG, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_3x4()
568 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4xACEG, vk44)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_3x4()
|
| D | 5x5s2p2-minmax-wasmsimd-arm-loadsplat-2x4.c | 303 const v128_t vi4xACEG = wasm_v32x4_shuffle(vi4x8ACE, vi4xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4() local
320 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4xACEG, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4()
325 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4xACEG, vk44)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4()
440 const v128_t vi4xACEG = wasm_v32x4_shuffle(vi4x8ACE, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4() local
451 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4xACEG, vk24)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4()
456 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4xACEG, vk44)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_2x4()
|