mt/bb_matmul/matmul_mi.c

   1 //**************************************************************************
   2 // Multi-threaded Matrix Multiply benchmark
   3 //--------------------------------------------------------------------------
   4 // TA     : Christopher Celio
   5 // Student:
   6 //
   7 //
   8 // This benchmark multiplies two 2-D arrays together and writes the results to
   9 // a third vector. The input data (and reference data) should be generated
  10 // using the matmul_gendata.pl perl script and dumped to a file named
  11 // dataset.h.
  12
  13
  14 // print out arrays, etc.
  15 //#define DEBUG
  16
  17 //--------------------------------------------------------------------------
  18 // Includes
  19
  20 #include <string.h>
  21 #include <stdlib.h>
  22 #include <stdio.h>
  23
  24
  25 //--------------------------------------------------------------------------
  26 // Input/Reference Data
  27
  28 typedef float data_t;
  29 #include "dataset.h"
  30
  31
  32 //--------------------------------------------------------------------------
  33 // Basic Utilities and Multi-thread Support
  34
  35 __thread unsigned long coreid;
  36 unsigned long ncores;
  37
  38 #include "util.h"
  39
  40 #define stringify_1(s) #s
  41 #define stringify(s) stringify_1(s)
  42 #define stats(code) do { \
  43     unsigned long _c = -rdcycle(), _i = -rdinstret(); \
  44     code; \
  45     _c += rdcycle(), _i += rdinstret(); \
  46     if (coreid == 0) \
  47       printf("%s: %ld cycles, %ld.%ld cycles/iter, %ld.%ld CPI\n", \
  48              stringify(code), _c, _c/DIM_SIZE/DIM_SIZE/DIM_SIZE, 10*_c/DIM_SIZE/DIM_SIZE/DIM_SIZE%10, _c/_i, 10*_c/_i%10); \
  49   } while(0)
  50
  51
  52 //--------------------------------------------------------------------------
  53 // Helper functions
  54
  55 void printArrayMT( char name[], int n, data_t arr[] )
  56 {
  57    int i;
  58    if (coreid != 0)
  59       return;
  60
  61    printf( " %10s :", name );
  62    for ( i = 0; i < n; i++ )
  63       printf( " %3ld ", (long) arr[i] );
  64    printf( "\n" );
  65 }
  66
  67 void __attribute__((noinline)) verifyMT(size_t n, const data_t* test, const data_t* correct)
  68 {
  69    if (coreid != 0)
  70       return;
  71
  72    size_t i;
  73    for (i = 0; i < n; i++)
  74    {
  75       if (test[i] != correct[i])
  76       {
  77          printf("FAILED test[%d]= %3ld, correct[%d]= %3ld\n",
  78             i, (long)test[i], i, (long)correct[i]);
  79          exit(-1);
  80       }
  81    }
  82
  83    return;
  84 }
  85
  86 //--------------------------------------------------------------------------
  87 // matmul function
  88
  89 // single-thread, naive version
  90 void __attribute__((noinline)) matmul_naive(const int lda,  const data_t A[], const data_t B[], data_t C[] )
  91 {
  92    int i, j, k;
  93
  94    if (coreid > 0)
  95       return;
  96
  97    for ( i = 0; i < lda; i++ )
  98       for ( j = 0; j < lda; j++ )
  99       {
 100          for ( k = 0; k < lda; k++ )
 101          {
 102             C[i + j*lda] += A[j*lda + k] * B[k*lda + i];
 103          }
 104       }
 105
 106 }
 107
 108
 109
 110 void __attribute__((noinline)) matmul_msi(const int lda,  const data_t A[], const data_t B[], data_t C[] ) {
 111   int i, j, k;
 112
 113   for (i = 0; i < lda; i += 2) {
 114     for (j = coreid * (lda / ncores); j < (coreid + 1) * (lda / ncores); j += 4) {
 115     //for (j = 0; j < lda; j += 4) {
 116       register data_t c00 = 0, c01 = 0;
 117       register data_t c10 = 0, c11 = 0;
 118       register data_t c20 = 0, c21 = 0;
 119       register data_t c30 = 0, c31 = 0;
 120
 121       register data_t a0, a1, a2, a3, b0, b1;
 122       for (k = 0; k < lda; k++) {
 123         a0 = A[j*lda + k + 0*lda];
 124         a1 = A[j*lda + k + 1*lda];
 125         a2 = A[j*lda + k + 2*lda];
 126         a3 = A[j*lda + k + 3*lda];
 127
 128         b0 = B[k*lda + i + 0];
 129         b1 = B[k*lda + i + 1];
 130         /*if (coreid == 0) {
 131           printf("i = %d; j = %d; k = %d\n", i, j, k);
 132           printf("%d += %d * %d; %d += %d * %d\n", (int)c00, (int)a0, (int)b0, (int)c01, (int)a0, (int)b1);
 133           printf("%d += %d * %d; %d += %d * %d\n", (int)c10, (int)a1, (int)b0, (int)c11, (int)a1, (int)b1);
 134           printf("%d += %d * %d; %d += %d * %d\n", (int)c20, (int)a2, (int)b0, (int)c21, (int)a2, (int)b1);
 135           printf("%d += %d * %d; %d += %d * %d\n", (int)c30, (int)a3, (int)b0, (int)c31, (int)a3, (int)b1);
 136           printf("\n");
 137         }*/
 138
 139         c00 += a0 * b0; c01 += a0 * b1;
 140         c10 += a1 * b0; c11 += a1 * b1;
 141         c20 += a2 * b0; c21 += a2 * b1;
 142         c30 += a3 * b0; c31 += a3 * b1;
 143       }
 144
 145       C[i + j*lda + 0 + 0*lda] = c00; C[i + j*lda + 1 + 0*lda] = c01;
 146       C[i + j*lda + 0 + 1*lda] = c10; C[i + j*lda + 1 + 1*lda] = c11;
 147       C[i + j*lda + 0 + 2*lda] = c20; C[i + j*lda + 1 + 2*lda] = c21;
 148       C[i + j*lda + 0 + 3*lda] = c30; C[i + j*lda + 1 + 3*lda] = c31;
 149     }
 150   }
 151 }
 152
 153 void __attribute__((noinline)) matmul_mi(const int lda,  const data_t A[], const data_t B[], data_t C[] ) {
 154   int i, j, k;
 155
 156   int curhalf = coreid;
 157   for (i = 0; i < lda; i += 2) {
 158     for (j = coreid * (lda / ncores); j < (coreid + 1) * (lda / ncores); j += 4) {
 159       register float c00 = 0, c01 = 0;
 160       register float c10 = 0, c11 = 0;
 161       register float c20 = 0, c21 = 0;
 162       register float c30 = 0, c31 = 0;
 163
 164       register float a0, a1, a2, a3, b0, b1;
 165       for (k = curhalf * (lda/2); k < curhalf * (lda/2)  + (lda/2); k++) {
 166         a0 = A[j*lda + k + 0*lda];
 167         a1 = A[j*lda + k + 1*lda];
 168         a2 = A[j*lda + k + 2*lda];
 169         a3 = A[j*lda + k + 3*lda];
 170
 171         b0 = B[k*lda + i + 0];
 172         b1 = B[k*lda + i + 1];
 173
 174         c00 += a0 * b0; c01 += a0 * b1;
 175         c10 += a1 * b0; c11 += a1 * b1;
 176         c20 += a2 * b0; c21 += a2 * b1;
 177         c30 += a3 * b0; c31 += a3 * b1;
 178       }
 179
 180       C[i + j*lda + 0 + 0*lda] += c00; C[i + j*lda + 1 + 0*lda] += c01;
 181       C[i + j*lda + 0 + 1*lda] += c10; C[i + j*lda + 1 + 1*lda] += c11;
 182       C[i + j*lda + 0 + 2*lda] += c20; C[i + j*lda + 1 + 2*lda] += c21;
 183       C[i + j*lda + 0 + 3*lda] += c30; C[i + j*lda + 1 + 3*lda] += c31;
 184     }
 185   }
 186
 187   barrier(nc);
 188   curhalf++;
 189   curhalf %= ncores;
 190
 191   for (i = 0; i < lda; i += 2) {
 192     for (j = coreid * (lda / ncores); j < (coreid + 1) * (lda / ncores); j += 4) {
 193       register float c00 = 0, c01 = 0;
 194       register float c10 = 0, c11 = 0;
 195       register float c20 = 0, c21 = 0;
 196       register float c30 = 0, c31 = 0;
 197
 198       register float a0, a1, a2, a3, b0, b1;
 199       for (k = curhalf * (lda/2); k < curhalf * (lda/2)  + (lda/2); k++) {
 200         a0 = A[j*lda + k + 0*lda];
 201         a1 = A[j*lda + k + 1*lda];
 202         a2 = A[j*lda + k + 2*lda];
 203         a3 = A[j*lda + k + 3*lda];
 204
 205         b0 = B[k*lda + i + 0];
 206         b1 = B[k*lda + i + 1];
 207
 208         c00 += a0 * b0; c01 += a0 * b1;
 209         c10 += a1 * b0; c11 += a1 * b1;
 210         c20 += a2 * b0; c21 += a2 * b1;
 211         c30 += a3 * b0; c31 += a3 * b1;
 212       }
 213
 214       C[i + j*lda + 0 + 0*lda] += c00; C[i + j*lda + 1 + 0*lda] += c01;
 215       C[i + j*lda + 0 + 1*lda] += c10; C[i + j*lda + 1 + 1*lda] += c11;
 216       C[i + j*lda + 0 + 2*lda] += c20; C[i + j*lda + 1 + 2*lda] += c21;
 217       C[i + j*lda + 0 + 3*lda] += c30; C[i + j*lda + 1 + 3*lda] += c31;
 218     }
 219   }
 220 }
 221
 222 void __attribute__((noinline)) matmul(const int lda,  const data_t A[], const data_t B[], data_t C[] )
 223 {
 224   matmul_mi(lda, A, B, C);
 225 }
 226
 227 //--------------------------------------------------------------------------
 228 // Main
 229 //
 230 // all threads start executing thread_entry(). Use their "coreid" to
 231 // differentiate between threads (each thread is running on a separate core).
 232
 233 void thread_entry(int cid, int nc)
 234 {
 235    coreid = cid;
 236    ncores = nc;
 237
 238    // static allocates data in the binary, which is visible to both threads
 239    static data_t results_data[ARRAY_SIZE];
 240
 241
 242 //   // Execute the provided, naive matmul
 243 //   barrier(nc);
 244 //   stats(matmul_naive(DIM_SIZE, input1_data, input2_data, results_data); barrier(nc));
 245 //
 246 //
 247 //   // verify
 248 //   verifyMT(ARRAY_SIZE, results_data, verify_data);
 249 //
 250 //   // clear results from the first trial
 251 //   size_t i;
 252 //   if (coreid == 0)
 253 //      for (i=0; i < ARRAY_SIZE; i++)
 254 //         results_data[i] = 0;
 255 //   barrier(nc);
 256
 257
 258    // Execute your faster matmul
 259    barrier(nc);
 260    stats(matmul(DIM_SIZE, input1_data, input2_data, results_data); barrier(nc));
 261
 262 #ifdef DEBUG
 263    printArrayMT("results:", ARRAY_SIZE, results_data);
 264    printArrayMT("verify :", ARRAY_SIZE, verify_data);
 265 #endif
 266
 267    // verify
 268    verifyMT(ARRAY_SIZE, results_data, verify_data);
 269    barrier(nc);
 270
 271    exit(0);
 272 }
 273