59     float32x2_t vacc0x01 = vld1_f32(w); w += 2;  in xnn_f32_igemm_minmax_ukernel_4x2__neonfma_lane_ld64()
60     float32x2_t vacc1x01 = vacc0x01;  in xnn_f32_igemm_minmax_ukernel_4x2__neonfma_lane_ld64()
61     float32x2_t vacc2x01 = vacc0x01;  in xnn_f32_igemm_minmax_ukernel_4x2__neonfma_lane_ld64()
62     float32x2_t vacc3x01 = vacc0x01;  in xnn_f32_igemm_minmax_ukernel_4x2__neonfma_lane_ld64()
90         const float32x2_t va0 = vld1_f32(a0); a0 += 2;  in xnn_f32_igemm_minmax_ukernel_4x2__neonfma_lane_ld64()
91         const float32x2_t va1 = vld1_f32(a1); a1 += 2;  in xnn_f32_igemm_minmax_ukernel_4x2__neonfma_lane_ld64()
92         const float32x2_t va2 = vld1_f32(a2); a2 += 2;  in xnn_f32_igemm_minmax_ukernel_4x2__neonfma_lane_ld64()
93         const float32x2_t va3 = vld1_f32(a3); a3 += 2;  in xnn_f32_igemm_minmax_ukernel_4x2__neonfma_lane_ld64()
95         const float32x2_t vb01c0 = vld1_f32(w); w += 2;  in xnn_f32_igemm_minmax_ukernel_4x2__neonfma_lane_ld64()
103           const float32x2_t va0c0 = vdup_lane_f32(va0, 0);  in xnn_f32_igemm_minmax_ukernel_4x2__neonfma_lane_ld64()
104           const float32x2_t va1c0 = vdup_lane_f32(va1, 0);  in xnn_f32_igemm_minmax_ukernel_4x2__neonfma_lane_ld64()
105           const float32x2_t va2c0 = vdup_lane_f32(va2, 0);  in xnn_f32_igemm_minmax_ukernel_4x2__neonfma_lane_ld64()
106           const float32x2_t va3c0 = vdup_lane_f32(va3, 0);  in xnn_f32_igemm_minmax_ukernel_4x2__neonfma_lane_ld64()
112         const float32x2_t vb01c1 = vld1_f32(w); w += 2;  in xnn_f32_igemm_minmax_ukernel_4x2__neonfma_lane_ld64()
120           const float32x2_t va0c1 = vdup_lane_f32(va0, 1);  in xnn_f32_igemm_minmax_ukernel_4x2__neonfma_lane_ld64()
121           const float32x2_t va1c1 = vdup_lane_f32(va1, 1);  in xnn_f32_igemm_minmax_ukernel_4x2__neonfma_lane_ld64()
122           const float32x2_t va2c1 = vdup_lane_f32(va2, 1);  in xnn_f32_igemm_minmax_ukernel_4x2__neonfma_lane_ld64()
123           const float32x2_t va3c1 = vdup_lane_f32(va3, 1);  in xnn_f32_igemm_minmax_ukernel_4x2__neonfma_lane_ld64()
131         const float32x2_t va0 = vld1_dup_f32(a0);  in xnn_f32_igemm_minmax_ukernel_4x2__neonfma_lane_ld64()
132         const float32x2_t va1 = vld1_dup_f32(a1);  in xnn_f32_igemm_minmax_ukernel_4x2__neonfma_lane_ld64()
133         const float32x2_t va2 = vld1_dup_f32(a2);  in xnn_f32_igemm_minmax_ukernel_4x2__neonfma_lane_ld64()
134         const float32x2_t va3 = vld1_dup_f32(a3);  in xnn_f32_igemm_minmax_ukernel_4x2__neonfma_lane_ld64()
136         const float32x2_t vb01 = vld1_f32(w); w += 2;  in xnn_f32_igemm_minmax_ukernel_4x2__neonfma_lane_ld64()
146     const float32x2_t vmax = vld1_dup_f32(&params->scalar.max);  in xnn_f32_igemm_minmax_ukernel_4x2__neonfma_lane_ld64()
152     const float32x2_t vmin = vld1_dup_f32(&params->scalar.min);  in xnn_f32_igemm_minmax_ukernel_4x2__neonfma_lane_ld64()