79       int8x8_t va3x0 = vld1_s8(a3); a3 += 8;  in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal()  local
94       int16x8_t vprod3x0123c0 = vmull_s8(vb0123c0x0, va3x0);  in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal()
107       int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0x0, va3x0);  in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal()
123       va3x0 = vext_s8(va3x0, va3x0, 2);  in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal()
128       int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1x0, va3x0);  in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal()
141       int16x8_t vprod3x4567c1 = vmull_s8(vb4567c1x0, va3x0);  in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal()
157       va3x0 = vext_s8(va3x0, va3x0, 2);  in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal()
162       int16x8_t vprod3x0123c2 = vmull_s8(vb0123c2x0, va3x0);  in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal()
175       int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2x0, va3x0);  in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal()
191       va3x0 = vext_s8(va3x0, va3x0, 2);  in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal()
196       int16x8_t vprod3x0123c3 = vmull_s8(vb0123c3x0, va3x0);  in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal()
209       int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3x0, va3x0);  in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal()
226       int8x8_t va3x0 = vld1_s8(a3); a3 += 8;  in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal()  local
240       int16x8_t vprod3x0123c0 = vmull_s8(vb0123c0x0, va3x0);  in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal()
248       int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0x0, va3x0);  in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal()
256       va3x0 = vext_s8(va3x0, va3x0, 2);  in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal()
260       int16x8_t vprod3x0123c1 = vmull_s8(vb0123c1x0, va3x0);  in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal()
268       int16x8_t vprod3x4567c1 = vmull_s8(vb4567c1x0, va3x0);  in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal()
276       va3x0 = vext_s8(va3x0, va3x0, 2);  in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal()
280       int16x8_t vprod3x0123c2 = vmull_s8(vb0123c2x0, va3x0);  in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal()
288       int16x8_t vprod3x4567c2 = vmull_s8(vb4567c2x0, va3x0);  in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal()
296       va3x0 = vext_s8(va3x0, va3x0, 2);  in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal()
300       int16x8_t vprod3x0123c3 = vmull_s8(vb0123c3x0, va3x0);  in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal()
308       int16x8_t vprod3x4567c3 = vmull_s8(vb4567c3x0, va3x0);  in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal()