99     float16x8_t vacc6x01234567 = vacc0x01234567;  in xnn_f16_gemm_minmax_ukernel_8x16__neonfp16arith_ld64()  local
125         vacc6x01234567 = vfmaq_lane_f16(vacc6x01234567, vb01234567c0, va6, 0);  in xnn_f16_gemm_minmax_ukernel_8x16__neonfp16arith_ld64()
151         vacc6x01234567 = vfmaq_f16(vacc6x01234567, va6c0, vb01234567c0);  in xnn_f16_gemm_minmax_ukernel_8x16__neonfp16arith_ld64()
172         vacc6x01234567 = vfmaq_lane_f16(vacc6x01234567, vb01234567c1, va6, 1);  in xnn_f16_gemm_minmax_ukernel_8x16__neonfp16arith_ld64()
198         vacc6x01234567 = vfmaq_f16(vacc6x01234567, va6c1, vb01234567c1);  in xnn_f16_gemm_minmax_ukernel_8x16__neonfp16arith_ld64()
219         vacc6x01234567 = vfmaq_lane_f16(vacc6x01234567, vb01234567c2, va6, 2);  in xnn_f16_gemm_minmax_ukernel_8x16__neonfp16arith_ld64()
245         vacc6x01234567 = vfmaq_f16(vacc6x01234567, va6c2, vb01234567c2);  in xnn_f16_gemm_minmax_ukernel_8x16__neonfp16arith_ld64()
266         vacc6x01234567 = vfmaq_lane_f16(vacc6x01234567, vb01234567c3, va6, 3);  in xnn_f16_gemm_minmax_ukernel_8x16__neonfp16arith_ld64()
292         vacc6x01234567 = vfmaq_f16(vacc6x01234567, va6c3, vb01234567c3);  in xnn_f16_gemm_minmax_ukernel_8x16__neonfp16arith_ld64()
326         vacc6x01234567 = vfmaq_f16(vacc6x01234567, va6, vb01234567);  in xnn_f16_gemm_minmax_ukernel_8x16__neonfp16arith_ld64()
348     vacc6x01234567 = vmulq_f16(vacc6x01234567, vscale);  in xnn_f16_gemm_minmax_ukernel_8x16__neonfp16arith_ld64()
366     vacc6x01234567 = vminq_f16(vacc6x01234567, vmax);  in xnn_f16_gemm_minmax_ukernel_8x16__neonfp16arith_ld64()
384     vacc6x01234567 = vmaxq_f16(vacc6x01234567, vmin);  in xnn_f16_gemm_minmax_ukernel_8x16__neonfp16arith_ld64()
414       vst1q_f16(c6, vacc6x01234567);  in xnn_f16_gemm_minmax_ukernel_8x16__neonfp16arith_ld64()
439         vst1q_f16(c6, vacc6x01234567); c6 += 8;  in xnn_f16_gemm_minmax_ukernel_8x16__neonfp16arith_ld64()
448         vacc6x01234567 = vacc6x89ABCDEF;  in xnn_f16_gemm_minmax_ukernel_8x16__neonfp16arith_ld64()
457       float16x4_t vacc6x0123 = vget_low_f16(vacc6x01234567);  in xnn_f16_gemm_minmax_ukernel_8x16__neonfp16arith_ld64()
475         vacc6x0123 = vget_high_f16(vacc6x01234567);  in xnn_f16_gemm_minmax_ukernel_8x16__neonfp16arith_ld64()