63     float16x8_t vacc2x01234567 = vacc0x01234567;  in xnn_f16_igemm_minmax_ukernel_4x16__neonfp16arith_ld64()  local
105           vacc2x01234567 = vfmaq_lane_f16(vacc2x01234567, vb01234567c0, va2, 0);  in xnn_f16_igemm_minmax_ukernel_4x16__neonfp16arith_ld64()
119           vacc2x01234567 = vfmaq_f16(vacc2x01234567, va2c0, vb01234567c0);  in xnn_f16_igemm_minmax_ukernel_4x16__neonfp16arith_ld64()
132           vacc2x01234567 = vfmaq_lane_f16(vacc2x01234567, vb01234567c1, va2, 1);  in xnn_f16_igemm_minmax_ukernel_4x16__neonfp16arith_ld64()
146           vacc2x01234567 = vfmaq_f16(vacc2x01234567, va2c1, vb01234567c1);  in xnn_f16_igemm_minmax_ukernel_4x16__neonfp16arith_ld64()
159           vacc2x01234567 = vfmaq_lane_f16(vacc2x01234567, vb01234567c2, va2, 2);  in xnn_f16_igemm_minmax_ukernel_4x16__neonfp16arith_ld64()
173           vacc2x01234567 = vfmaq_f16(vacc2x01234567, va2c2, vb01234567c2);  in xnn_f16_igemm_minmax_ukernel_4x16__neonfp16arith_ld64()
186           vacc2x01234567 = vfmaq_lane_f16(vacc2x01234567, vb01234567c3, va2, 3);  in xnn_f16_igemm_minmax_ukernel_4x16__neonfp16arith_ld64()
200           vacc2x01234567 = vfmaq_f16(vacc2x01234567, va2c3, vb01234567c3);  in xnn_f16_igemm_minmax_ukernel_4x16__neonfp16arith_ld64()
220           vacc2x01234567 = vfmaq_f16(vacc2x01234567, va2, vb01234567);  in xnn_f16_igemm_minmax_ukernel_4x16__neonfp16arith_ld64()
236     vacc2x01234567 = vmulq_f16(vacc2x01234567, vscale);  in xnn_f16_igemm_minmax_ukernel_4x16__neonfp16arith_ld64()
246     vacc2x01234567 = vminq_f16(vacc2x01234567, vmax);  in xnn_f16_igemm_minmax_ukernel_4x16__neonfp16arith_ld64()
256     vacc2x01234567 = vmaxq_f16(vacc2x01234567, vmin);  in xnn_f16_igemm_minmax_ukernel_4x16__neonfp16arith_ld64()
267       vst1q_f16(c2, vacc2x01234567);  in xnn_f16_igemm_minmax_ukernel_4x16__neonfp16arith_ld64()
282         vst1q_f16(c2, vacc2x01234567); c2 += 8;  in xnn_f16_igemm_minmax_ukernel_4x16__neonfp16arith_ld64()
287         vacc2x01234567 = vacc2x89ABCDEF;  in xnn_f16_igemm_minmax_ukernel_4x16__neonfp16arith_ld64()
292       float16x4_t vacc2x0123 = vget_low_f16(vacc2x01234567);  in xnn_f16_igemm_minmax_ukernel_4x16__neonfp16arith_ld64()
302         vacc2x0123 = vget_high_f16(vacc2x01234567);  in xnn_f16_igemm_minmax_ukernel_4x16__neonfp16arith_ld64()