73       const int8x8_t va0x1 = vld1_s8(a0); a0 += 8;  in xnn_qs8_gemm_minmax_ukernel_3x16c2__neon_mlal_padal_dup()  local
100 …l_s8(vprod0x0123c0, vb0123c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 0)));  in xnn_qs8_gemm_minmax_ukernel_3x16c2__neon_mlal_padal_dup()
110 …l_s8(vprod0x4567c0, vb4567c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 0)));  in xnn_qs8_gemm_minmax_ukernel_3x16c2__neon_mlal_padal_dup()
120 …l_s8(vprod0x89ABc0, vb89ABc0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 0)));  in xnn_qs8_gemm_minmax_ukernel_3x16c2__neon_mlal_padal_dup()
130 …l_s8(vprod0xCDEFc0, vbCDEFc0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 0)));  in xnn_qs8_gemm_minmax_ukernel_3x16c2__neon_mlal_padal_dup()
140 …l_s8(vprod0x0123c1, vb0123c1x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 1)));  in xnn_qs8_gemm_minmax_ukernel_3x16c2__neon_mlal_padal_dup()
150 …l_s8(vprod0x4567c1, vb4567c1x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 1)));  in xnn_qs8_gemm_minmax_ukernel_3x16c2__neon_mlal_padal_dup()
160 …l_s8(vprod0x89ABc1, vb89ABc1x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 1)));  in xnn_qs8_gemm_minmax_ukernel_3x16c2__neon_mlal_padal_dup()
170 …l_s8(vprod0xCDEFc1, vbCDEFc1x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 1)));  in xnn_qs8_gemm_minmax_ukernel_3x16c2__neon_mlal_padal_dup()
180 …l_s8(vprod0x0123c2, vb0123c2x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 2)));  in xnn_qs8_gemm_minmax_ukernel_3x16c2__neon_mlal_padal_dup()
190 …l_s8(vprod0x4567c2, vb4567c2x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 2)));  in xnn_qs8_gemm_minmax_ukernel_3x16c2__neon_mlal_padal_dup()
200 …l_s8(vprod0x89ABc2, vb89ABc2x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 2)));  in xnn_qs8_gemm_minmax_ukernel_3x16c2__neon_mlal_padal_dup()
210 …l_s8(vprod0xCDEFc2, vbCDEFc2x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 2)));  in xnn_qs8_gemm_minmax_ukernel_3x16c2__neon_mlal_padal_dup()
220 …l_s8(vprod0x0123c3, vb0123c3x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 3)));  in xnn_qs8_gemm_minmax_ukernel_3x16c2__neon_mlal_padal_dup()
230 …l_s8(vprod0x4567c3, vb4567c3x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 3)));  in xnn_qs8_gemm_minmax_ukernel_3x16c2__neon_mlal_padal_dup()
240 …l_s8(vprod0x89ABc3, vb89ABc3x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 3)));  in xnn_qs8_gemm_minmax_ukernel_3x16c2__neon_mlal_padal_dup()
250 …l_s8(vprod0xCDEFc3, vbCDEFc3x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 3)));  in xnn_qs8_gemm_minmax_ukernel_3x16c2__neon_mlal_padal_dup()