1  /* { dg-do run { target { ! ia32 } } } */
       2  /* { dg-require-effective-target amx_int8 } */
       3  /* { dg-options "-O2 -mamx-int8" } */
       4  #include <immintrin.h>
       5  
       6  #define AMX_INT8
       7  #define DO_TEST test_amx_int8_dpbsud
       8  void test_amx_int8_dpbsud ();
       9  #include "amx-check.h"
      10  
      11  /* Init tile buffer with int32 value*/
      12  void init_i32_max_tile_buffer (uint8_t *buf)
      13  {
      14    int i, j;
      15    int *ptr = (int *)buf;
      16    for (i = 0; i < 16; i++)
      17      for (j = 0; j < 16; j++)
      18        ptr[i * 16 + j] = 2 * i - (16 - j);
      19  }
      20  
      21  void calc_matrix_dpbsud (__tile *dst, __tile *src1, __tile *src2)
      22  {
      23    int8_t *src1_buf = (int8_t *)src1->buf;
      24    uint8_t *src2_buf = (uint8_t *)src2->buf;
      25    int *dst_buf = (int *)dst->buf;
      26  
      27    int M = src1->rows;
      28    int N = src1->colsb / 4;
      29    int K = src2->colsb / 4;
      30    int i, j, k, t;
      31  
      32    for (i = 0; i < M; i++)
      33      for (j = 0; j < N; j++)
      34        for (k = 0; k < K; k++)
      35  	for (t = 0; t < 4; t++)
      36  	  {
      37  	    dst_buf[i * N + k] += 
      38  	      ((int) src1_buf[i * 4 * N + 4 * j + t]) *
      39  	      ((unsigned) src2_buf[j * 4 * K + 4 * k + t]);
      40  	  }
      41  }
      42  
      43  void test_amx_int8_dpbsud ()
      44  {
      45    __tilecfg_u cfg;
      46    __tile dst, dst_ref, src1, src2;
      47    uint8_t tmp_dst_buf[1024];
      48    
      49    init_i32_max_tile_buffer (tmp_dst_buf);
      50  
      51    init_tile_config (&cfg);
      52    init_tile_reg_and_src_with_buffer (1, dst, tmp_dst_buf);
      53    init_tile_reg_and_src (2, src1);
      54    init_tile_reg_and_src (3, src2);
      55  
      56    calc_matrix_dpbsud (&dst, &src1, &src2);
      57    _tile_dpbsud (1, 2, 3);
      58    _tile_stored (1, dst_ref.buf, _STRIDE);
      59    
      60    if (!check_tile_register (&dst_ref, &dst))
      61        abort();
      62  }