113       const float16x4_t va7 = vld1_f16(a7); a7 += 4;  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()
134         vacc7x89ABCDEF = vfmaq_lane_f16(vacc7x89ABCDEF, vb89ABCDEFc0, va7, 0);  in xnn_f16_gemm_minmax_ukernel_8x16__neonfp16arith_ld64()
143         const float16x8_t va7c0 = vdupq_lane_f16(va7, 0);  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()
181         vacc7x89ABCDEF = vfmaq_lane_f16(vacc7x89ABCDEF, vb89ABCDEFc1, va7, 1);  in xnn_f16_gemm_minmax_ukernel_8x16__neonfp16arith_ld64()
190         const float16x8_t va7c1 = vdupq_lane_f16(va7, 1);  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()
228         vacc7x89ABCDEF = vfmaq_lane_f16(vacc7x89ABCDEF, vb89ABCDEFc2, va7, 2);  in xnn_f16_gemm_minmax_ukernel_8x16__neonfp16arith_ld64()
237         const float16x8_t va7c2 = vdupq_lane_f16(va7, 2);  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()
275         vacc7x89ABCDEF = vfmaq_lane_f16(vacc7x89ABCDEF, vb89ABCDEFc3, va7, 3);  in xnn_f16_gemm_minmax_ukernel_8x16__neonfp16arith_ld64()
284         const float16x8_t va7c3 = vdupq_lane_f16(va7, 3);  in xnn_f16_gemm_minmax_ukernel_8x16__neonfp16arith_ld64()
315         const float16x8_t va7 = vld1q_dup_f16(a7); a7 += 1;  in xnn_f16_gemm_minmax_ukernel_8x16__neonfp16arith_ld64()  local
327         vacc7x01234567 = vfmaq_f16(vacc7x01234567, va7, vb01234567);  in xnn_f16_gemm_minmax_ukernel_8x16__neonfp16arith_ld64()
335         vacc7x89ABCDEF = vfmaq_f16(vacc7x89ABCDEF, va7, vb89ABCDEF);  in xnn_f16_gemm_minmax_ukernel_8x16__neonfp16arith_ld64()