64         const int8x8_t va0x1 = vld1_s8(a0); a0 += 8;  in xnn_qs8_igemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()  local
85 …l_s8(vprod0x0123c0, vb0123c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 0)));  in xnn_qs8_igemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
89 …l_s8(vprod0x4567c0, vb4567c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 0)));  in xnn_qs8_igemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
93 …l_s8(vprod0x89ABc0, vb89ABc0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 0)));  in xnn_qs8_igemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
97 …l_s8(vprod0xCDEFc0, vbCDEFc0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 0)));  in xnn_qs8_igemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
101 …l_s8(vprod0x0123c1, vb0123c1x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 1)));  in xnn_qs8_igemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
105 …l_s8(vprod0x4567c1, vb4567c1x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 1)));  in xnn_qs8_igemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
109 …l_s8(vprod0x89ABc1, vb89ABc1x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 1)));  in xnn_qs8_igemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
113 …l_s8(vprod0xCDEFc1, vbCDEFc1x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 1)));  in xnn_qs8_igemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
117 …l_s8(vprod0x0123c2, vb0123c2x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 2)));  in xnn_qs8_igemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
121 …l_s8(vprod0x4567c2, vb4567c2x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 2)));  in xnn_qs8_igemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
125 …l_s8(vprod0x89ABc2, vb89ABc2x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 2)));  in xnn_qs8_igemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
129 …l_s8(vprod0xCDEFc2, vbCDEFc2x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 2)));  in xnn_qs8_igemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
133 …l_s8(vprod0x0123c3, vb0123c3x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 3)));  in xnn_qs8_igemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
137 …l_s8(vprod0x4567c3, vb4567c3x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 3)));  in xnn_qs8_igemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
141 …l_s8(vprod0x89ABc3, vb89ABc3x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 3)));  in xnn_qs8_igemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()
145 …l_s8(vprod0xCDEFc3, vbCDEFc3x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 3)));  in xnn_qs8_igemm_minmax_ukernel_1x16c2__neon_mlal_padal_dup()