50       const float16x4_t va0 = vld1_f16(a0); a0 += 4;  in xnn_f16_gemm_minmax_ukernel_1x16__neonfp16arith_ld64()  local
56         vacc0x01234567 = vfmaq_lane_f16(vacc0x01234567, vb01234567c0, va0, 0);  in xnn_f16_gemm_minmax_ukernel_1x16__neonfp16arith_ld64()
57         vacc0x89ABCDEF = vfmaq_lane_f16(vacc0x89ABCDEF, vb89ABCDEFc0, va0, 0);  in xnn_f16_gemm_minmax_ukernel_1x16__neonfp16arith_ld64()
59         const float16x8_t va0c0 = vdupq_lane_f16(va0, 0);  in xnn_f16_gemm_minmax_ukernel_1x16__neonfp16arith_ld64()
68         vacc0x01234567 = vfmaq_lane_f16(vacc0x01234567, vb01234567c1, va0, 1);  in xnn_f16_gemm_minmax_ukernel_1x16__neonfp16arith_ld64()
69         vacc0x89ABCDEF = vfmaq_lane_f16(vacc0x89ABCDEF, vb89ABCDEFc1, va0, 1);  in xnn_f16_gemm_minmax_ukernel_1x16__neonfp16arith_ld64()
71         const float16x8_t va0c1 = vdupq_lane_f16(va0, 1);  in xnn_f16_gemm_minmax_ukernel_1x16__neonfp16arith_ld64()
80         vacc0x01234567 = vfmaq_lane_f16(vacc0x01234567, vb01234567c2, va0, 2);  in xnn_f16_gemm_minmax_ukernel_1x16__neonfp16arith_ld64()
81         vacc0x89ABCDEF = vfmaq_lane_f16(vacc0x89ABCDEF, vb89ABCDEFc2, va0, 2);  in xnn_f16_gemm_minmax_ukernel_1x16__neonfp16arith_ld64()
83         const float16x8_t va0c2 = vdupq_lane_f16(va0, 2);  in xnn_f16_gemm_minmax_ukernel_1x16__neonfp16arith_ld64()
92         vacc0x01234567 = vfmaq_lane_f16(vacc0x01234567, vb01234567c3, va0, 3);  in xnn_f16_gemm_minmax_ukernel_1x16__neonfp16arith_ld64()
93         vacc0x89ABCDEF = vfmaq_lane_f16(vacc0x89ABCDEF, vb89ABCDEFc3, va0, 3);  in xnn_f16_gemm_minmax_ukernel_1x16__neonfp16arith_ld64()
95         const float16x8_t va0c3 = vdupq_lane_f16(va0, 3);  in xnn_f16_gemm_minmax_ukernel_1x16__neonfp16arith_ld64()
105         const float16x8_t va0 = vld1q_dup_f16(a0); a0 += 1;  in xnn_f16_gemm_minmax_ukernel_1x16__neonfp16arith_ld64()  local
110         vacc0x01234567 = vfmaq_f16(vacc0x01234567, va0, vb01234567);  in xnn_f16_gemm_minmax_ukernel_1x16__neonfp16arith_ld64()
111         vacc0x89ABCDEF = vfmaq_f16(vacc0x89ABCDEF, va0, vb89ABCDEF);  in xnn_f16_gemm_minmax_ukernel_1x16__neonfp16arith_ld64()