70     float16x8_t vacc3x89ABCDEF = vacc0x89ABCDEF;  in xnn_f16_gemm_minmax_ukernel_4x16__neonfp16arith_ld64()  local
90         vacc3x89ABCDEF = vfmaq_lane_f16(vacc3x89ABCDEF, vb89ABCDEFc0, va3, 0);  in xnn_f16_gemm_minmax_ukernel_4x16__neonfp16arith_ld64()
104         vacc3x89ABCDEF = vfmaq_f16(vacc3x89ABCDEF, va3c0, vb89ABCDEFc0);  in xnn_f16_gemm_minmax_ukernel_4x16__neonfp16arith_ld64()
117         vacc3x89ABCDEF = vfmaq_lane_f16(vacc3x89ABCDEF, vb89ABCDEFc1, va3, 1);  in xnn_f16_gemm_minmax_ukernel_4x16__neonfp16arith_ld64()
131         vacc3x89ABCDEF = vfmaq_f16(vacc3x89ABCDEF, va3c1, vb89ABCDEFc1);  in xnn_f16_gemm_minmax_ukernel_4x16__neonfp16arith_ld64()
144         vacc3x89ABCDEF = vfmaq_lane_f16(vacc3x89ABCDEF, vb89ABCDEFc2, va3, 2);  in xnn_f16_gemm_minmax_ukernel_4x16__neonfp16arith_ld64()
158         vacc3x89ABCDEF = vfmaq_f16(vacc3x89ABCDEF, va3c2, vb89ABCDEFc2);  in xnn_f16_gemm_minmax_ukernel_4x16__neonfp16arith_ld64()
171         vacc3x89ABCDEF = vfmaq_lane_f16(vacc3x89ABCDEF, vb89ABCDEFc3, va3, 3);  in xnn_f16_gemm_minmax_ukernel_4x16__neonfp16arith_ld64()
185         vacc3x89ABCDEF = vfmaq_f16(vacc3x89ABCDEF, va3c3, vb89ABCDEFc3);  in xnn_f16_gemm_minmax_ukernel_4x16__neonfp16arith_ld64()
207         vacc3x89ABCDEF = vfmaq_f16(vacc3x89ABCDEF, va3, vb89ABCDEF);  in xnn_f16_gemm_minmax_ukernel_4x16__neonfp16arith_ld64()
221     vacc3x89ABCDEF = vmulq_f16(vacc3x89ABCDEF, vscale);  in xnn_f16_gemm_minmax_ukernel_4x16__neonfp16arith_ld64()
231     vacc3x89ABCDEF = vminq_f16(vacc3x89ABCDEF, vmax);  in xnn_f16_gemm_minmax_ukernel_4x16__neonfp16arith_ld64()
241     vacc3x89ABCDEF = vmaxq_f16(vacc3x89ABCDEF, vmin);  in xnn_f16_gemm_minmax_ukernel_4x16__neonfp16arith_ld64()
254       vst1q_f16(c3 + 8, vacc3x89ABCDEF);  in xnn_f16_gemm_minmax_ukernel_4x16__neonfp16arith_ld64()
273         vacc3x01234567 = vacc3x89ABCDEF;  in xnn_f16_gemm_minmax_ukernel_4x16__neonfp16arith_ld64()