101     float16x8_t vacc7x01234567 = vacc0x01234567;  in xnn_f16_gemm_minmax_ukernel_8x16__neonfp16arith_ld64()  local
126         vacc7x01234567 = vfmaq_lane_f16(vacc7x01234567, vb01234567c0, va7, 0);  in xnn_f16_gemm_minmax_ukernel_8x16__neonfp16arith_ld64()
152         vacc7x01234567 = vfmaq_f16(vacc7x01234567, va7c0, vb01234567c0);  in xnn_f16_gemm_minmax_ukernel_8x16__neonfp16arith_ld64()
173         vacc7x01234567 = vfmaq_lane_f16(vacc7x01234567, vb01234567c1, va7, 1);  in xnn_f16_gemm_minmax_ukernel_8x16__neonfp16arith_ld64()
199         vacc7x01234567 = vfmaq_f16(vacc7x01234567, va7c1, vb01234567c1);  in xnn_f16_gemm_minmax_ukernel_8x16__neonfp16arith_ld64()
220         vacc7x01234567 = vfmaq_lane_f16(vacc7x01234567, vb01234567c2, va7, 2);  in xnn_f16_gemm_minmax_ukernel_8x16__neonfp16arith_ld64()
246         vacc7x01234567 = vfmaq_f16(vacc7x01234567, va7c2, vb01234567c2);  in xnn_f16_gemm_minmax_ukernel_8x16__neonfp16arith_ld64()
267         vacc7x01234567 = vfmaq_lane_f16(vacc7x01234567, vb01234567c3, va7, 3);  in xnn_f16_gemm_minmax_ukernel_8x16__neonfp16arith_ld64()
293         vacc7x01234567 = vfmaq_f16(vacc7x01234567, va7c3, vb01234567c3);  in xnn_f16_gemm_minmax_ukernel_8x16__neonfp16arith_ld64()
327         vacc7x01234567 = vfmaq_f16(vacc7x01234567, va7, vb01234567);  in xnn_f16_gemm_minmax_ukernel_8x16__neonfp16arith_ld64()
349     vacc7x01234567 = vmulq_f16(vacc7x01234567, vscale);  in xnn_f16_gemm_minmax_ukernel_8x16__neonfp16arith_ld64()
367     vacc7x01234567 = vminq_f16(vacc7x01234567, vmax);  in xnn_f16_gemm_minmax_ukernel_8x16__neonfp16arith_ld64()
385     vacc7x01234567 = vmaxq_f16(vacc7x01234567, vmin);  in xnn_f16_gemm_minmax_ukernel_8x16__neonfp16arith_ld64()
417       vst1q_f16(c7, vacc7x01234567);  in xnn_f16_gemm_minmax_ukernel_8x16__neonfp16arith_ld64()
440         vst1q_f16(c7, vacc7x01234567); c7 += 8;  in xnn_f16_gemm_minmax_ukernel_8x16__neonfp16arith_ld64()
449         vacc7x01234567 = vacc7x89ABCDEF;  in xnn_f16_gemm_minmax_ukernel_8x16__neonfp16arith_ld64()
458       float16x4_t vacc7x0123 = vget_low_f16(vacc7x01234567);  in xnn_f16_gemm_minmax_ukernel_8x16__neonfp16arith_ld64()
476         vacc7x0123 = vget_high_f16(vacc7x01234567);  in xnn_f16_gemm_minmax_ukernel_8x16__neonfp16arith_ld64()