173         const float32x4_t vi2x1 = vld1q_f32(i2); i2 += 4;  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()  local
181         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c0x0123, vget_low_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
183         vo1c4567 = vfmaq_lane_f32(vo1c4567, vk01c0x4567, vget_low_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
196         vo0c0123 = vfmaq_lane_f32(vo0c0123, vk21c0x0123, vget_low_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
198         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk21c0x4567, vget_low_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
205         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c1x0123, vget_low_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
207         vo1c4567 = vfmaq_lane_f32(vo1c4567, vk01c1x4567, vget_low_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
220         vo0c0123 = vfmaq_lane_f32(vo0c0123, vk21c1x0123, vget_low_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
222         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk21c1x4567, vget_low_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
229         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c2x0123, vget_high_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
231         vo1c4567 = vfmaq_lane_f32(vo1c4567, vk01c2x4567, vget_high_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
244         vo0c0123 = vfmaq_lane_f32(vo0c0123, vk21c2x0123, vget_high_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
246         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk21c2x4567, vget_high_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
253         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk02c0x0123, vget_high_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
255         vo1c4567 = vfmaq_lane_f32(vo1c4567, vk02c0x4567, vget_high_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
268         vo0c0123 = vfmaq_lane_f32(vo0c0123, vk22c0x0123, vget_high_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
270         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk22c0x4567, vget_high_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
330         vi2x0 = vcombine_f32(vget_high_f32(vi2x1), vi2x2);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
460         const float32x4_t vi2x1 = vld1q_f32(i2); i2 += 3;  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()  local
468         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c0x0123, vget_low_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
470         vo1c4567 = vfmaq_lane_f32(vo1c4567, vk01c0x4567, vget_low_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
483         vo0c0123 = vfmaq_lane_f32(vo0c0123, vk21c0x0123, vget_low_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
485         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk21c0x4567, vget_low_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
492         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c1x0123, vget_low_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
494         vo1c4567 = vfmaq_lane_f32(vo1c4567, vk01c1x4567, vget_low_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
507         vo0c0123 = vfmaq_lane_f32(vo0c0123, vk21c1x0123, vget_low_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
509         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk21c1x4567, vget_low_f32(vi2x1), 1);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
516         vo1c0123 = vfmaq_lane_f32(vo1c0123, vk01c2x0123, vget_high_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
518         vo1c4567 = vfmaq_lane_f32(vo1c4567, vk01c2x4567, vget_high_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
531         vo0c0123 = vfmaq_lane_f32(vo0c0123, vk21c2x0123, vget_high_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
533         vo0c4567 = vfmaq_lane_f32(vo0c4567, vk21c2x4567, vget_high_f32(vi2x1), 0);  in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()