133       const int8x8_t va1x1 = vld1_s8(a1); a1 += 8;  in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()  local
162       vprod1x0 = vmlal_s8(vprod1x0, vb0x1, va1x1);  in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
175       vprod1x1 = vmlal_s8(vprod1x1, vb1x1, va1x1);  in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
188       vprod1x2 = vmlal_s8(vprod1x2, vb2x1, va1x1);  in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
201       vprod1x3 = vmlal_s8(vprod1x3, vb3x1, va1x1);  in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
214       vprod1x4 = vmlal_s8(vprod1x4, vb4x1, va1x1);  in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
227       vprod1x5 = vmlal_s8(vprod1x5, vb5x1, va1x1);  in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
240       vprod1x6 = vmlal_s8(vprod1x6, vb6x1, va1x1);  in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
253       vprod1x7 = vmlal_s8(vprod1x7, vb7x1, va1x1);  in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
266       vprod1x8 = vmlal_s8(vprod1x8, vb8x1, va1x1);  in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
279       vprod1x9 = vmlal_s8(vprod1x9, vb9x1, va1x1);  in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
292       vprod1x10 = vmlal_s8(vprod1x10, vb10x1, va1x1);  in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
305       vprod1x11 = vmlal_s8(vprod1x11, vb11x1, va1x1);  in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
318       vprod1x12 = vmlal_s8(vprod1x12, vb12x1, va1x1);  in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
331       vprod1x13 = vmlal_s8(vprod1x13, vb13x1, va1x1);  in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
344       vprod1x14 = vmlal_s8(vprod1x14, vb14x1, va1x1);  in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
357       vprod1x15 = vmlal_s8(vprod1x15, vb15x1, va1x1);  in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()