65     float16x8_t vacc2x01234567 = vacc0x01234567;  in xnn_f16_gemm_minmax_ukernel_4x8__neonfp16arith_ld64()  local
80         vacc2x01234567 = vfmaq_lane_f16(vacc2x01234567, vb01234567c0, va2, 0);  in xnn_f16_gemm_minmax_ukernel_4x8__neonfp16arith_ld64()
90         vacc2x01234567 = vfmaq_f16(vacc2x01234567, va2c0, vb01234567c0);  in xnn_f16_gemm_minmax_ukernel_4x8__neonfp16arith_ld64()
98         vacc2x01234567 = vfmaq_lane_f16(vacc2x01234567, vb01234567c1, va2, 1);  in xnn_f16_gemm_minmax_ukernel_4x8__neonfp16arith_ld64()
108         vacc2x01234567 = vfmaq_f16(vacc2x01234567, va2c1, vb01234567c1);  in xnn_f16_gemm_minmax_ukernel_4x8__neonfp16arith_ld64()
116         vacc2x01234567 = vfmaq_lane_f16(vacc2x01234567, vb01234567c2, va2, 2);  in xnn_f16_gemm_minmax_ukernel_4x8__neonfp16arith_ld64()
126         vacc2x01234567 = vfmaq_f16(vacc2x01234567, va2c2, vb01234567c2);  in xnn_f16_gemm_minmax_ukernel_4x8__neonfp16arith_ld64()
134         vacc2x01234567 = vfmaq_lane_f16(vacc2x01234567, vb01234567c3, va2, 3);  in xnn_f16_gemm_minmax_ukernel_4x8__neonfp16arith_ld64()
144         vacc2x01234567 = vfmaq_f16(vacc2x01234567, va2c3, vb01234567c3);  in xnn_f16_gemm_minmax_ukernel_4x8__neonfp16arith_ld64()
161         vacc2x01234567 = vfmaq_f16(vacc2x01234567, va2, vb01234567);  in xnn_f16_gemm_minmax_ukernel_4x8__neonfp16arith_ld64()
171     vacc2x01234567 = vmulq_f16(vacc2x01234567, vscale);  in xnn_f16_gemm_minmax_ukernel_4x8__neonfp16arith_ld64()
177     vacc2x01234567 = vminq_f16(vacc2x01234567, vmax);  in xnn_f16_gemm_minmax_ukernel_4x8__neonfp16arith_ld64()
183     vacc2x01234567 = vmaxq_f16(vacc2x01234567, vmin);  in xnn_f16_gemm_minmax_ukernel_4x8__neonfp16arith_ld64()
191       vst1q_f16(c2, vacc2x01234567);  in xnn_f16_gemm_minmax_ukernel_4x8__neonfp16arith_ld64()
205       float16x4_t vacc2x0123 = vget_low_f16(vacc2x01234567);  in xnn_f16_gemm_minmax_ukernel_4x8__neonfp16arith_ld64()
215         vacc2x0123 = vget_high_f16(vacc2x01234567);  in xnn_f16_gemm_minmax_ukernel_4x8__neonfp16arith_ld64()