gcc/testsuite/gcc.target/powerpc/mma-double-test.c

   1 /* { dg-do run { target { power10_hw } } } */
   2 /* { dg-do link { target { ! power10_hw } } } */
   3 /* { dg-require-effective-target power10_ok } */
   4 /* { dg-require-effective-target ppc_mma_hw } */
   5 /* { dg-options "-mdejagnu-cpu=power10 -O2" } */
   6
   7 #include <stdio.h>
   8 #include <stdlib.h>
   9 #include <altivec.h>
  10
  11 typedef unsigned char vec_t __attribute__ ((vector_size (16)));
  12 typedef double v4sf_t __attribute__ ((vector_size (16)));
  13 #define SAVE_ACC(ACC, ldc, J)  \
  14           __builtin_mma_disassemble_acc (result, ACC); \
  15           rowC = (v4sf_t *) &CO[0*ldc+J]; \
  16           rowC[0] += result[0]; \
  17           rowC = (v4sf_t *) &CO[1*ldc+J]; \
  18           rowC[0] += result[1]; \
  19           rowC = (v4sf_t *) &CO[2*ldc+J]; \
  20           rowC[0] += result[2]; \
  21           rowC = (v4sf_t *) &CO[3*ldc+J]; \
  22           rowC[0] += result[3];
  23
  24 void
  25 MMA (int m, int n, int k, double *A, double *B, double *C)
  26 {
  27   __vector_quad acc0, acc1, acc2, acc3, acc4, acc5, acc6, acc7;
  28   v4sf_t result[4];
  29   v4sf_t *rowC;
  30   for (int l = 0; l < n; l += 4)
  31     {
  32       double *CO;
  33       double *AO;
  34       AO = A;
  35       CO = C;
  36       C += m * 4;
  37       for (int j = 0; j < m; j += 16)
  38         {
  39           double *BO = B;
  40           __builtin_mma_xxsetaccz (&acc0);
  41           __builtin_mma_xxsetaccz (&acc1);
  42           __builtin_mma_xxsetaccz (&acc2);
  43           __builtin_mma_xxsetaccz (&acc3);
  44           __builtin_mma_xxsetaccz (&acc4);
  45           __builtin_mma_xxsetaccz (&acc5);
  46           __builtin_mma_xxsetaccz (&acc6);
  47           __builtin_mma_xxsetaccz (&acc7);
  48           unsigned long i;
  49
  50           for (i = 0; i < k; i++)
  51             {
  52               vec_t *rowA = (vec_t *) & AO[i * 16];
  53               __vector_pair rowB;
  54               vec_t *rb = (vec_t *) & BO[i * 4];
  55               __builtin_mma_assemble_pair (&rowB, rb[1], rb[0]);
  56               __builtin_mma_xvf64gerpp (&acc0, rowB, rowA[0]);
  57               __builtin_mma_xvf64gerpp (&acc1, rowB, rowA[1]);
  58               __builtin_mma_xvf64gerpp (&acc2, rowB, rowA[2]);
  59               __builtin_mma_xvf64gerpp (&acc3, rowB, rowA[3]);
  60               __builtin_mma_xvf64gerpp (&acc4, rowB, rowA[4]);
  61               __builtin_mma_xvf64gerpp (&acc5, rowB, rowA[5]);
  62               __builtin_mma_xvf64gerpp (&acc6, rowB, rowA[6]);
  63               __builtin_mma_xvf64gerpp (&acc7, rowB, rowA[7]);
  64             }
  65           SAVE_ACC (&acc0, m, 0);
  66           SAVE_ACC (&acc2, m, 4);
  67           SAVE_ACC (&acc1, m, 2);
  68           SAVE_ACC (&acc3, m, 6);
  69           SAVE_ACC (&acc4, m, 8);
  70           SAVE_ACC (&acc6, m, 12);
  71           SAVE_ACC (&acc5, m, 10);
  72           SAVE_ACC (&acc7, m, 14);
  73           AO += k * 16;
  74           BO += k * 4;
  75           CO += 16;
  76         }
  77       B += k * 4;
  78     }
  79 }
  80
  81 void
  82 init (double *matrix, int row, int column)
  83 {
  84   for (int j = 0; j < column; j++)
  85     {
  86       for (int i = 0; i < row; i++)
  87         {
  88           matrix[j * row + i] = (i * 16 + 2 + j) / 0.123;
  89         }
  90     }
  91 }
  92
  93 void
  94 init0 (double *matrix, double *matrix1, int row, int column)
  95 {
  96   for (int j = 0; j < column; j++)
  97     for (int i = 0; i < row; i++)
  98       matrix[j * row + i] = matrix1[j * row + i] = 0;
  99 }
 100
 101
 102 void
 103 print (const char *name, const double *matrix, int row, int column)
 104 {
 105   printf ("Matrix %s has %d rows and %d columns:\n", name, row, column);
 106   for (int i = 0; i < row; i++)
 107     {
 108       for (int j = 0; j < column; j++)
 109         {
 110           printf ("%f ", matrix[j * row + i]);
 111         }
 112       printf ("\n");
 113     }
 114   printf ("\n");
 115 }
 116
 117 int
 118 main (int argc, char *argv[])
 119 {
 120   int rowsA, colsB, common;
 121   int i, j, k;
 122   int ret = 0;
 123
 124   for (int t = 16; t <= 128; t += 16)
 125     {
 126       for (int t1 = 4; t1 <= 16; t1 += 4)
 127         {
 128           rowsA = t;
 129           colsB = t1;
 130           common = 1;
 131           /* printf ("Running test for rows = %d,cols = %d\n", t, t1); */
 132           double A[rowsA * common];
 133           double B[common * colsB];
 134           double C[rowsA * colsB];
 135           double D[rowsA * colsB];
 136
 137
 138           init (A, rowsA, common);
 139           init (B, common, colsB);
 140           init0 (C, D, rowsA, colsB);
 141           MMA (rowsA, colsB, common, A, B, C);
 142
 143           for (i = 0; i < colsB; i++)
 144             {
 145               for (j = 0; j < rowsA; j++)
 146                 {
 147                   D[i * rowsA + j] = 0;
 148                   for (k = 0; k < common; k++)
 149                     {
 150                       D[i * rowsA + j] +=
 151                         A[k * rowsA + j] * B[k + common * i];
 152                     }
 153                 }
 154             }
 155           for (i = 0; i < colsB; i++)
 156             {
 157               for (j = 0; j < rowsA; j++)
 158                 {
 159                   for (k = 0; k < common; k++)
 160                     {
 161                       if (D[i * rowsA + j] != C[i * rowsA + j])
 162                         {
 163                           printf ("Error %d,%d,%d\n",i,j,k);
 164                           ret++;
 165                         }
 166                     }
 167                 }
 168             }
 169           if (ret)
 170             {
 171               print ("A", A, rowsA, common);
 172               print ("B", B, common, colsB);
 173               print ("C", C, rowsA, colsB);
 174               print ("D", D, rowsA, colsB);
 175             }
 176         }
 177     }
 178
 179 #ifdef VERBOSE
 180   if (ret)
 181     printf ("MMA double test fail: %d errors\n",ret);
 182   else
 183     printf ("MMA single test success: 0 MMA errors\n");
 184 #else
 185   if (ret)
 186     abort();
 187 #endif
 188
 189   return ret;
 190 }