+++ /dev/null
-//**************************************************************************\r
-// Multi-threaded Matrix Multiply benchmark\r
-//--------------------------------------------------------------------------\r
-// TA : Christopher Celio\r
-// Student: \r
-//\r
-//\r
-// This benchmark multiplies two 2-D arrays together and writes the results to\r
-// a third vector. The input data (and reference data) should be generated\r
-// using the matmul_gendata.pl perl script and dumped to a file named\r
-// dataset.h. \r
-\r
-\r
-// print out arrays, etc.\r
-//#define DEBUG\r
-\r
-//--------------------------------------------------------------------------\r
-// Includes \r
-\r
-#include <string.h>\r
-#include <stdlib.h>\r
-#include <stdio.h>\r
-\r
-\r
-//--------------------------------------------------------------------------\r
-// Input/Reference Data\r
-\r
-typedef float data_t;\r
-#include "dataset.h"\r
- \r
- \r
-//--------------------------------------------------------------------------\r
-// Basic Utilities and Multi-thread Support\r
-\r
-__thread unsigned long coreid;\r
-unsigned long ncores;\r
-\r
-#include "util.h"\r
- \r
-#define stringify_1(s) #s\r
-#define stringify(s) stringify_1(s)\r
-#define stats(code) do { \\r
- unsigned long _c = -rdcycle(), _i = -rdinstret(); \\r
- code; \\r
- _c += rdcycle(), _i += rdinstret(); \\r
- if (coreid == 0) \\r
- printf("%s: %ld cycles, %ld.%ld cycles/iter, %ld.%ld CPI\n", \\r
- 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); \\r
- } while(0)\r
- \r
-\r
-//--------------------------------------------------------------------------\r
-// Helper functions\r
- \r
-void printArrayMT( char name[], int n, data_t arr[] )\r
-{\r
- int i;\r
- if (coreid != 0)\r
- return;\r
- \r
- printf( " %10s :", name );\r
- for ( i = 0; i < n; i++ )\r
- printf( " %3ld ", (long) arr[i] );\r
- printf( "\n" );\r
-}\r
- \r
-void __attribute__((noinline)) verifyMT(size_t n, const data_t* test, const data_t* correct)\r
-{\r
- if (coreid != 0)\r
- return;\r
-\r
- size_t i;\r
- for (i = 0; i < n; i++)\r
- {\r
- if (test[i] != correct[i])\r
- {\r
- printf("FAILED test[%d]= %3ld, correct[%d]= %3ld\n", \r
- i, (long)test[i], i, (long)correct[i]);\r
- exit(-1);\r
- }\r
- }\r
- \r
- return;\r
-}\r
- \r
-//--------------------------------------------------------------------------\r
-// matmul function\r
- \r
-// single-thread, naive version\r
-void __attribute__((noinline)) matmul_naive(const int lda, const data_t A[], const data_t B[], data_t C[] )\r
-{\r
- int i, j, k;\r
-\r
- if (coreid > 0)\r
- return;\r
- \r
- for ( i = 0; i < lda; i++ )\r
- for ( j = 0; j < lda; j++ ) \r
- {\r
- for ( k = 0; k < lda; k++ ) \r
- {\r
- C[i + j*lda] += A[j*lda + k] * B[k*lda + i];\r
- }\r
- }\r
-\r
-}\r
- \r
-void __attribute__((noinline)) matmul(const int lda, const data_t A[], const data_t B[], data_t C[] )\r
-{\r
- static __thread int i, j, k;\r
- static __thread data_t tempA0, tempA1, tempA2, tempA3, tempA4, tempA5, tempA6, tempA7;\r
- static __thread data_t tempC0, tempC1, tempC2, tempC3, tempC4, tempC5, tempC6, tempC7, tempC8, tempC9, tempC10, tempC11, tempC12, tempC13, tempC14, tempC15;\r
-\r
- static __thread int start, end, jStride, jToRow, jToCol;\r
- \r
- start = coreid << 9;\r
- end = (coreid+1) << 9;\r
- jStride = 8;\r
-\r
- for (j=start; j < end; j+=jStride) {\r
- jToRow = (j>>5)<<5;\r
- jToCol = j%32;\r
- tempC0 = 0;\r
- tempC1 = 0;\r
- tempC2 = 0;\r
- tempC3 = 0;\r
- tempC4 = 0;\r
- tempC5 = 0;\r
- tempC6 = 0;\r
- tempC7 = 0;\r
- for ( i=0; i < lda; i+=2 ) {\r
- tempA0 = A[i + jToRow];\r
- tempA1 = A[i+1 + jToRow];\r
- tempC0 += tempA0 * B[(jToCol ) + (i<<5)];\r
- tempC1 += tempA0 * B[(jToCol+1 ) + (i<<5)];\r
- tempC2 += tempA0 * B[(jToCol+2 ) + (i<<5)];\r
- tempC3 += tempA0 * B[(jToCol+3 ) + (i<<5)];\r
- tempC4 += tempA0 * B[(jToCol+4 ) + (i<<5)];\r
- tempC5 += tempA0 * B[(jToCol+5 ) + (i<<5)];\r
- tempC6 += tempA0 * B[(jToCol+6 ) + (i<<5)];\r
- tempC7 += tempA0 * B[(jToCol+7 ) + (i<<5)];\r
- tempC0 += tempA1 * B[(jToCol ) + ((i+1)<<5)];\r
- tempC1 += tempA1 * B[(jToCol+1 ) + ((i+1)<<5)];\r
- tempC2 += tempA1 * B[(jToCol+2 ) + ((i+1)<<5)];\r
- tempC3 += tempA1 * B[(jToCol+3 ) + ((i+1)<<5)];\r
- tempC4 += tempA1 * B[(jToCol+4 ) + ((i+1)<<5)];\r
- tempC5 += tempA1 * B[(jToCol+5 ) + ((i+1)<<5)];\r
- tempC6 += tempA1 * B[(jToCol+6 ) + ((i+1)<<5)];\r
- tempC7 += tempA1 * B[(jToCol+7 ) + ((i+1)<<5)];\r
- }\r
- C[j] =tempC0;\r
- C[j + 1 ]=tempC1;\r
- C[j + 2 ]=tempC2;\r
- C[j + 3 ]=tempC3;\r
- C[j + 4 ]=tempC4;\r
- C[j + 5 ]=tempC5;\r
- C[j + 6 ]=tempC6;\r
- C[j + 7 ]=tempC7;\r
- }\r
- \r
-}\r
-\r
-//--------------------------------------------------------------------------\r
-// Main\r
-//\r
-// all threads start executing thread_entry(). Use their "coreid" to\r
-// differentiate between threads (each thread is running on a separate core).\r
- \r
-void thread_entry(int cid, int nc)\r
-{\r
- coreid = cid;\r
- ncores = nc;\r
-\r
- // static allocates data in the binary, which is visible to both threads\r
- static data_t results_data[ARRAY_SIZE];\r
-\r
-\r
- //// Execute the provided, naive matmul\r
- //barrier(nc);\r
- //stats(matmul_naive(DIM_SIZE, input1_data, input2_data, results_data); barrier(nc));\r
- \r
- //\r
- //// verify\r
- //verifyMT(ARRAY_SIZE, results_data, verify_data);\r
- //\r
- //// clear results from the first trial\r
- //size_t i;\r
- //if (coreid == 0) \r
- // for (i=0; i < ARRAY_SIZE; i++)\r
- // results_data[i] = 0;\r
- //barrier(nc);\r
-\r
- \r
- // Execute your faster matmul\r
- barrier(nc);\r
- stats(matmul(DIM_SIZE, input1_data, input2_data, results_data); barrier(nc));\r
- \r
-#ifdef DEBUG\r
- printArrayMT("results:", ARRAY_SIZE, results_data);\r
- printArrayMT("verify :", ARRAY_SIZE, verify_data);\r
-#endif\r
- \r
- // verify\r
- verifyMT(ARRAY_SIZE, results_data, verify_data);\r
- barrier(nc);\r
-\r
- exit(0);\r
-}\r
-\r
-\r