--- /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 printArray( 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)) verify(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
+ \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
+\r
+\r
+void __attribute__((noinline)) matmul(const int lda, const data_t A[], const data_t B[], data_t C[] )\r
+{\r
+ \r
+ // ***************************** //\r
+ // **** ADD YOUR CODE HERE ***** //\r
+ // ***************************** //\r
+ //\r
+ // feel free to make a separate function for MI and MSI versions.\r
+ \r
+ int m, i, j, k, iB0, iB1;\r
+ data_t tempC0, tempC1, tempC2, tempC3, tempC4, tempC5, tempC6, tempC7;\r
+ data_t tempA0, tempA1;\r
+ \r
+ if (coreid == 0){\r
+ for (m = 0; m < 2; m++){\r
+ for (j = 0; j < lda/2; j++){\r
+ for (i = 0; i < lda; i+=8){\r
+ tempC0 = C[i + j*lda];\r
+ tempC1 = C[i + j*lda+1];\r
+ tempC2 = C[i + j*lda+2];\r
+ tempC3 = C[i + j*lda+3];\r
+ tempC4 = C[i + j*lda+4];\r
+ tempC5 = C[i + j*lda+5];\r
+ tempC6 = C[i + j*lda+6];\r
+ tempC7 = C[i + j*lda+7];\r
+ iB0 = m*lda*lda/2+i;\r
+ iB1 = iB0+lda;\r
+ for (k = m*lda/2; k < (m+1)*lda/2; k+=2){\r
+ tempA0 = A[j*lda+k];\r
+ tempA1 = A[j*lda+k+1];\r
+ tempC0 += tempA0*B[iB0]+tempA1*B[iB1];\r
+ tempC1 += tempA0*B[iB0+1]+tempA1*B[iB1+1];\r
+ tempC2 += tempA0*B[iB0+2]+tempA1*B[iB1+2];\r
+ tempC3 += tempA0*B[iB0+3]+tempA1*B[iB1+3];\r
+ tempC4 += tempA0*B[iB0+4]+tempA1*B[iB1+4];\r
+ tempC5 += tempA0*B[iB0+5]+tempA1*B[iB1+5];\r
+ tempC6 += tempA0*B[iB0+6]+tempA1*B[iB1+6];\r
+ tempC7 += tempA0*B[iB0+7]+tempA1*B[iB1+7];\r
+ iB0 += 2*lda;\r
+ iB1 += 2*lda;\r
+ \r
+ }\r
+ C[i + j*lda] = tempC0;\r
+ C[i + j*lda + 1] = tempC1;\r
+ C[i + j*lda + 2] = tempC2;\r
+ C[i + j*lda + 3] = tempC3;\r
+ C[i + j*lda + 4] = tempC4;\r
+ C[i + j*lda + 5] = tempC5;\r
+ C[i + j*lda + 6] = tempC6;\r
+ C[i + j*lda + 7] = tempC7;\r
+ }\r
+ }\r
+ }\r
+ } else {\r
+ for (m = 2; m > 0; m--){\r
+ for (j = lda-1; j >= lda/2; j--){\r
+ for (i = lda-1; i >= 0; i-=8){\r
+ tempC0 = C[i + j*lda];\r
+ tempC1 = C[i + j*lda - 1];\r
+ tempC2 = C[i + j*lda - 2];\r
+ tempC3 = C[i + j*lda - 3];\r
+ tempC4 = C[i + j*lda - 4];\r
+ tempC5 = C[i + j*lda - 5];\r
+ tempC6 = C[i + j*lda - 6];\r
+ tempC7 = C[i + j*lda - 7];\r
+ for (k = m*lda/2-1; k >= (m-1)*lda/2; k-=2){\r
+ tempA0 = A[j*lda+k];\r
+ tempA1 = A[j*lda+k-1];\r
+ tempC0 += tempA0*B[k*lda+i]+tempA1*B[(k-1)*lda+i];\r
+ tempC1 += tempA0*B[k*lda+i-1]+tempA1*B[(k-1)*lda+i-1];\r
+ tempC2 += tempA0*B[k*lda+i-2]+tempA1*B[(k-1)*lda+i-2];\r
+ tempC3 += tempA0*B[k*lda+i-3]+tempA1*B[(k-1)*lda+i-3];\r
+ tempC4 += tempA0*B[k*lda+i-4]+tempA1*B[(k-1)*lda+i-4];\r
+ tempC5 += tempA0*B[k*lda+i-5]+tempA1*B[(k-1)*lda+i-5];\r
+ tempC6 += tempA0*B[k*lda+i-6]+tempA1*B[(k-1)*lda+i-6];\r
+ tempC7 += tempA0*B[k*lda+i-7]+tempA1*B[(k-1)*lda+i-7];\r
+ }\r
+ C[i + j*lda] = tempC0;\r
+ C[i + j*lda - 1] = tempC1;\r
+ C[i + j*lda - 2] = tempC2;\r
+ C[i + j*lda - 3] = tempC3;\r
+ C[i + j*lda - 4] = tempC4;\r
+ C[i + j*lda - 5] = tempC5;\r
+ C[i + j*lda - 6] = tempC6;\r
+ C[i + j*lda - 7] = tempC7;\r
+ }\r
+ }\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();\r
+// stats(matmul_naive(DIM_SIZE, input1_data, input2_data, results_data); barrier());\r
+// \r
+// \r
+// // verify\r
+// verify(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();\r
+ \r
+ \r
+ // Execute your faster matmul\r
+ barrier();\r
+ stats(matmul(DIM_SIZE, input1_data, input2_data, results_data); barrier());\r
+ \r
+#ifdef DEBUG\r
+ printArray("results:", ARRAY_SIZE, results_data);\r
+ printArray("verify :", ARRAY_SIZE, verify_data);\r
+#endif\r
+ \r
+ // verify\r
+ verify(ARRAY_SIZE, results_data, verify_data);\r
+ barrier();\r
+ \r
+ exit(0);\r
+}\r
+\r