65       const int8x8_t va1x1 = vld1_s8(a1); a1 += 8;  in xnn_qs8_gemm_minmax_ukernel_2x16c2__neon_mlal_padal_dup()  local
88 …l_s8(vprod1x0123c0, vb0123c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 0)));  in xnn_qs8_gemm_minmax_ukernel_2x16c2__neon_mlal_padal_dup()
95 …l_s8(vprod1x4567c0, vb4567c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 0)));  in xnn_qs8_gemm_minmax_ukernel_2x16c2__neon_mlal_padal_dup()
102 …l_s8(vprod1x89ABc0, vb89ABc0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 0)));  in xnn_qs8_gemm_minmax_ukernel_2x16c2__neon_mlal_padal_dup()
109 …l_s8(vprod1xCDEFc0, vbCDEFc0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 0)));  in xnn_qs8_gemm_minmax_ukernel_2x16c2__neon_mlal_padal_dup()
116 …l_s8(vprod1x0123c1, vb0123c1x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 1)));  in xnn_qs8_gemm_minmax_ukernel_2x16c2__neon_mlal_padal_dup()
123 …l_s8(vprod1x4567c1, vb4567c1x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 1)));  in xnn_qs8_gemm_minmax_ukernel_2x16c2__neon_mlal_padal_dup()
130 …l_s8(vprod1x89ABc1, vb89ABc1x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 1)));  in xnn_qs8_gemm_minmax_ukernel_2x16c2__neon_mlal_padal_dup()
137 …l_s8(vprod1xCDEFc1, vbCDEFc1x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 1)));  in xnn_qs8_gemm_minmax_ukernel_2x16c2__neon_mlal_padal_dup()
144 …l_s8(vprod1x0123c2, vb0123c2x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 2)));  in xnn_qs8_gemm_minmax_ukernel_2x16c2__neon_mlal_padal_dup()
151 …l_s8(vprod1x4567c2, vb4567c2x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 2)));  in xnn_qs8_gemm_minmax_ukernel_2x16c2__neon_mlal_padal_dup()
158 …l_s8(vprod1x89ABc2, vb89ABc2x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 2)));  in xnn_qs8_gemm_minmax_ukernel_2x16c2__neon_mlal_padal_dup()
165 …l_s8(vprod1xCDEFc2, vbCDEFc2x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 2)));  in xnn_qs8_gemm_minmax_ukernel_2x16c2__neon_mlal_padal_dup()
172 …l_s8(vprod1x0123c3, vb0123c3x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 3)));  in xnn_qs8_gemm_minmax_ukernel_2x16c2__neon_mlal_padal_dup()
179 …l_s8(vprod1x4567c3, vb4567c3x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 3)));  in xnn_qs8_gemm_minmax_ukernel_2x16c2__neon_mlal_padal_dup()
186 …l_s8(vprod1x89ABc3, vb89ABc3x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 3)));  in xnn_qs8_gemm_minmax_ukernel_2x16c2__neon_mlal_padal_dup()
193 …l_s8(vprod1xCDEFc3, vbCDEFc3x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 3)));  in xnn_qs8_gemm_minmax_ukernel_2x16c2__neon_mlal_padal_dup()