102         const int8x8_t va1x1 = vld1_s8(a1); a1 += 8;  in xnn_qs8_igemm_minmax_ukernel_4x16c2__neon_mlal_padal_dup()  local
131 …l_s8(vprod1x0123c0, vb0123c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 0)));  in xnn_qs8_igemm_minmax_ukernel_4x16c2__neon_mlal_padal_dup()
144 …l_s8(vprod1x4567c0, vb4567c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 0)));  in xnn_qs8_igemm_minmax_ukernel_4x16c2__neon_mlal_padal_dup()
157 …l_s8(vprod1x89ABc0, vb89ABc0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 0)));  in xnn_qs8_igemm_minmax_ukernel_4x16c2__neon_mlal_padal_dup()
170 …l_s8(vprod1xCDEFc0, vbCDEFc0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 0)));  in xnn_qs8_igemm_minmax_ukernel_4x16c2__neon_mlal_padal_dup()
183 …l_s8(vprod1x0123c1, vb0123c1x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 1)));  in xnn_qs8_igemm_minmax_ukernel_4x16c2__neon_mlal_padal_dup()
196 …l_s8(vprod1x4567c1, vb4567c1x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 1)));  in xnn_qs8_igemm_minmax_ukernel_4x16c2__neon_mlal_padal_dup()
209 …l_s8(vprod1x89ABc1, vb89ABc1x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 1)));  in xnn_qs8_igemm_minmax_ukernel_4x16c2__neon_mlal_padal_dup()
222 …l_s8(vprod1xCDEFc1, vbCDEFc1x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 1)));  in xnn_qs8_igemm_minmax_ukernel_4x16c2__neon_mlal_padal_dup()
235 …l_s8(vprod1x0123c2, vb0123c2x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 2)));  in xnn_qs8_igemm_minmax_ukernel_4x16c2__neon_mlal_padal_dup()
248 …l_s8(vprod1x4567c2, vb4567c2x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 2)));  in xnn_qs8_igemm_minmax_ukernel_4x16c2__neon_mlal_padal_dup()
261 …l_s8(vprod1x89ABc2, vb89ABc2x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 2)));  in xnn_qs8_igemm_minmax_ukernel_4x16c2__neon_mlal_padal_dup()
274 …l_s8(vprod1xCDEFc2, vbCDEFc2x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 2)));  in xnn_qs8_igemm_minmax_ukernel_4x16c2__neon_mlal_padal_dup()
287 …l_s8(vprod1x0123c3, vb0123c3x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 3)));  in xnn_qs8_igemm_minmax_ukernel_4x16c2__neon_mlal_padal_dup()
300 …l_s8(vprod1x4567c3, vb4567c3x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 3)));  in xnn_qs8_igemm_minmax_ukernel_4x16c2__neon_mlal_padal_dup()
313 …l_s8(vprod1x89ABc3, vb89ABc3x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 3)));  in xnn_qs8_igemm_minmax_ukernel_4x16c2__neon_mlal_padal_dup()
326 …l_s8(vprod1xCDEFc3, vbCDEFc3x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 3)));  in xnn_qs8_igemm_minmax_ukernel_4x16c2__neon_mlal_padal_dup()