138         const int16x8_t vxa4 = vreinterpretq_s16_u16(vmovl_u8(va4));  in xnn_q8_igemm_ukernel_8x8__neon()  local
155           vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa4), 0);  in xnn_q8_igemm_ukernel_8x8__neon()
156 …        vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa4), 0);  in xnn_q8_igemm_ukernel_8x8__neon()
177           vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa4), 1);  in xnn_q8_igemm_ukernel_8x8__neon()
178 …        vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa4), 1);  in xnn_q8_igemm_ukernel_8x8__neon()
199           vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa4), 2);  in xnn_q8_igemm_ukernel_8x8__neon()
200 …        vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa4), 2);  in xnn_q8_igemm_ukernel_8x8__neon()
221           vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa4), 3);  in xnn_q8_igemm_ukernel_8x8__neon()
222 …        vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa4), 3);  in xnn_q8_igemm_ukernel_8x8__neon()
243 …        vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa4), 0);  in xnn_q8_igemm_ukernel_8x8__neon()
244 …       vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567), vget_high_s16(vxa4), 0);  in xnn_q8_igemm_ukernel_8x8__neon()
265 …        vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa4), 1);  in xnn_q8_igemm_ukernel_8x8__neon()
266 …       vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567), vget_high_s16(vxa4), 1);  in xnn_q8_igemm_ukernel_8x8__neon()
287 …        vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa4), 2);  in xnn_q8_igemm_ukernel_8x8__neon()
288 …       vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567), vget_high_s16(vxa4), 2);  in xnn_q8_igemm_ukernel_8x8__neon()
309 …        vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa4), 3);  in xnn_q8_igemm_ukernel_8x8__neon()
310 …       vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567), vget_high_s16(vxa4), 3);  in xnn_q8_igemm_ukernel_8x8__neon()
334         const int16x8_t vxa4 = vreinterpretq_s16_u16(vmovl_u8(va4));  in xnn_q8_igemm_ukernel_8x8__neon()  local
351           vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa4), 0);  in xnn_q8_igemm_ukernel_8x8__neon()
352 …        vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa4), 0);  in xnn_q8_igemm_ukernel_8x8__neon()
373           vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa4), 1);  in xnn_q8_igemm_ukernel_8x8__neon()
374 …        vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa4), 1);  in xnn_q8_igemm_ukernel_8x8__neon()
394 …         vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa4), 2);  in xnn_q8_igemm_ukernel_8x8__neon()
395 …        vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa4), 2);  in xnn_q8_igemm_ukernel_8x8__neon()
415 …         vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa4), 3);  in xnn_q8_igemm_ukernel_8x8__neon()
416 …        vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa4), 3);  in xnn_q8_igemm_ukernel_8x8__neon()
436 …        vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa4), 0);  in xnn_q8_igemm_ukernel_8x8__neon()
437 …       vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567), vget_high_s16(vxa4), 0);  in xnn_q8_igemm_ukernel_8x8__neon()
457 …        vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa4), 1);  in xnn_q8_igemm_ukernel_8x8__neon()
458 …       vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567), vget_high_s16(vxa4), 1);  in xnn_q8_igemm_ukernel_8x8__neon()
478 …        vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa4), 2);  in xnn_q8_igemm_ukernel_8x8__neon()
479 …       vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567), vget_high_s16(vxa4), 2);  in xnn_q8_igemm_ukernel_8x8__neon()