69     float32x4_t vi5x4567 = vld1q_f32(i5); i5 = (const float*) ((uintptr_t) i5 + input_tuple_stride);  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()  local
100       vo4567p20 = vfmaq_laneq_f32(vo4567p20, vi5x4567, vwGHIJ, 2);  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
103       vo4567p10 = vfmaq_laneq_f32(vo4567p10, vi5x4567, vwKLMN, 3);  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
112       const float32x4_t vi5x3456 = vextq_f32(vi5x0123, vi5x4567, 3);  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
140       const float32x4_t vi5x2345 = vextq_f32(vi5x0123, vi5x4567, 2);  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
168       vi5x0123 = vi5x4567;  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
176       const float32x4_t vi5x5678 = vextq_f32(vi5x4567, vi5x89AB, 1);  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
204       const float32x4_t vi5x6789 = vextq_f32(vi5x4567, vi5x89AB, 2);  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
232       vi5x4567 = vi5x89AB;  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
287       vo4567p20 = vfmaq_laneq_f32(vo4567p20, vi5x4567, vwGHIJ, 2);  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
290       vo4567p10 = vfmaq_laneq_f32(vo4567p10, vi5x4567, vwKLMN, 3);  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
299       const float32x4_t vi5x3456 = vextq_f32(vi5x0123, vi5x4567, 3);  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
328       const float32x4_t vi5x2345 = vextq_f32(vi5x0123, vi5x4567, 2);  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
357       vi5x0123 = vi5x4567;  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
365       const float32x4_t vi5x5678 = vextq_f32(vi5x4567, vi5x89AB, 1);  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
393       const float32x4_t vi5x6789 = vextq_f32(vi5x4567, vi5x89AB, 2);  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
421       vi5x4567 = vi5x89AB;  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
454       vi5x4567 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi5x4567)));  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
471       vo4567p20 = vfmaq_laneq_f32(vo4567p20, vi5x4567, vwGHIJ, 2);  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
474       vo4567p10 = vfmaq_laneq_f32(vo4567p10, vi5x4567, vwKLMN, 3);  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
483       const float32x4_t vi5x3456 = vextq_f32(vi5x0123, vi5x4567, 3);  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
511       const float32x4_t vi5x2345 = vextq_f32(vi5x0123, vi5x4567, 2);  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
541       const float32x4_t vi5x5678 = vextq_f32(vi5x4567, vzero, 1);  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
569       const float32x4_t vi5x6789 = vextq_f32(vi5x4567, vzero, 2);  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
655     float32x4_t vi5x4567 = vld1q_f32(i5); i5 = (const float*) ((uintptr_t) i5 + input_tuple_stride);  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()  local
682       vo4567p10 = vfmaq_laneq_f32(vo4567p10, vi5x4567, vwKLMN, 3);  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
690       const float32x4_t vi5x3456 = vextq_f32(vi5x0123, vi5x4567, 3);  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
712       const float32x4_t vi5x2345 = vextq_f32(vi5x0123, vi5x4567, 2);  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
734       vi5x0123 = vi5x4567;  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
741       const float32x4_t vi5x5678 = vextq_f32(vi5x4567, vi5x89AB, 1);  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
763       const float32x4_t vi5x6789 = vextq_f32(vi5x4567, vi5x89AB, 2);  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
785       vi5x4567 = vi5x89AB;  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
831       vo4567p10 = vfmaq_laneq_f32(vo4567p10, vi5x4567, vwKLMN, 3);  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
839       const float32x4_t vi5x3456 = vextq_f32(vi5x0123, vi5x4567, 3);  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
862       const float32x4_t vi5x2345 = vextq_f32(vi5x0123, vi5x4567, 2);  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
885       vi5x0123 = vi5x4567;  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
892       const float32x4_t vi5x5678 = vextq_f32(vi5x4567, vi5x89AB, 1);  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
914       const float32x4_t vi5x6789 = vextq_f32(vi5x4567, vi5x89AB, 2);  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
936       vi5x4567 = vi5x89AB;  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
963       vi5x4567 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi5x4567)));  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
978       vo4567p10 = vfmaq_laneq_f32(vo4567p10, vi5x4567, vwKLMN, 3);  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
986       const float32x4_t vi5x3456 = vextq_f32(vi5x0123, vi5x4567, 3);  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
1008       const float32x4_t vi5x2345 = vextq_f32(vi5x0123, vi5x4567, 2);  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
1031       const float32x4_t vi5x5678 = vextq_f32(vi5x4567, vzero, 1);  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()
1053       const float32x4_t vi5x6789 = vextq_f32(vi5x4567, vzero, 2);  in xnn_f32_dwconv_spchw_ukernel_5x5p2__neonfma()