239       const int16x8_t vxb01234567c6 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c6, vb_zero_point));  in xnn_q8_gemm_ukernel_8x8__neon()  local
241       vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);  in xnn_q8_gemm_ukernel_8x8__neon()
242       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);  in xnn_q8_gemm_ukernel_8x8__neon()
243       vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa1), 2);  in xnn_q8_gemm_ukernel_8x8__neon()
244       vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa1), 2);  in xnn_q8_gemm_ukernel_8x8__neon()
245       vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa2), 2);  in xnn_q8_gemm_ukernel_8x8__neon()
246       vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa2), 2);  in xnn_q8_gemm_ukernel_8x8__neon()
247       vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa3), 2);  in xnn_q8_gemm_ukernel_8x8__neon()
248       vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa3), 2);  in xnn_q8_gemm_ukernel_8x8__neon()
249       vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa4), 2);  in xnn_q8_gemm_ukernel_8x8__neon()
250       vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa4), 2);  in xnn_q8_gemm_ukernel_8x8__neon()
251       vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa5), 2);  in xnn_q8_gemm_ukernel_8x8__neon()
252       vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa5), 2);  in xnn_q8_gemm_ukernel_8x8__neon()
253       vacc6x0123 = vmlal_lane_s16(vacc6x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa6), 2);  in xnn_q8_gemm_ukernel_8x8__neon()
254       vacc6x4567 = vmlal_lane_s16(vacc6x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa6), 2);  in xnn_q8_gemm_ukernel_8x8__neon()
255       vacc7x0123 = vmlal_lane_s16(vacc7x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa7), 2);  in xnn_q8_gemm_ukernel_8x8__neon()
256       vacc7x4567 = vmlal_lane_s16(vacc7x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa7), 2);  in xnn_q8_gemm_ukernel_8x8__neon()
425 …      const int16x8_t vxb01234567c6 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c6, vb_zero_point));  in xnn_q8_gemm_ukernel_8x8__neon()  local
427 …      vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);  in xnn_q8_gemm_ukernel_8x8__neon()
428 …     vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);  in xnn_q8_gemm_ukernel_8x8__neon()
429 …      vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa1), 2);  in xnn_q8_gemm_ukernel_8x8__neon()
430 …     vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa1), 2);  in xnn_q8_gemm_ukernel_8x8__neon()
431 …      vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa2), 2);  in xnn_q8_gemm_ukernel_8x8__neon()
432 …     vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa2), 2);  in xnn_q8_gemm_ukernel_8x8__neon()
433 …      vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa3), 2);  in xnn_q8_gemm_ukernel_8x8__neon()
434 …     vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa3), 2);  in xnn_q8_gemm_ukernel_8x8__neon()
435 …      vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa4), 2);  in xnn_q8_gemm_ukernel_8x8__neon()
436 …     vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa4), 2);  in xnn_q8_gemm_ukernel_8x8__neon()
437 …      vacc5x0123 = vmlal_lane_s16(vacc5x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa5), 2);  in xnn_q8_gemm_ukernel_8x8__neon()
438 …     vacc5x4567 = vmlal_lane_s16(vacc5x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa5), 2);  in xnn_q8_gemm_ukernel_8x8__neon()
439 …      vacc6x0123 = vmlal_lane_s16(vacc6x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa6), 2);  in xnn_q8_gemm_ukernel_8x8__neon()
440 …     vacc6x4567 = vmlal_lane_s16(vacc6x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa6), 2);  in xnn_q8_gemm_ukernel_8x8__neon()
441 …      vacc7x0123 = vmlal_lane_s16(vacc7x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa7), 2);  in xnn_q8_gemm_ukernel_8x8__neon()
442 …     vacc7x4567 = vmlal_lane_s16(vacc7x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa7), 2);  in xnn_q8_gemm_ukernel_8x8__neon()