//************************************************************************** // Multi-threaded Matrix Multiply benchmark //-------------------------------------------------------------------------- // TA : Christopher Celio // Student: // // // This benchmark multiplies two 2-D arrays together and writes the results to // a third vector. The input data (and reference data) should be generated // using the matmul_gendata.pl perl script and dumped to a file named // dataset.h. // print out arrays, etc. //#define DEBUG //-------------------------------------------------------------------------- // Includes #include #include #include //-------------------------------------------------------------------------- // Input/Reference Data typedef float data_t; #include "dataset.h" //-------------------------------------------------------------------------- // Basic Utilities and Multi-thread Support __thread unsigned long coreid; unsigned long ncores; #include "util.h" #define stringify_1(s) #s #define stringify(s) stringify_1(s) #define stats(code) do { \ unsigned long _c = -rdcycle(), _i = -rdinstret(); \ code; \ _c += rdcycle(), _i += rdinstret(); \ if (coreid == 0) \ printf("%s: %ld cycles, %ld.%ld cycles/iter, %ld.%ld CPI\n", \ 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); \ } while(0) //-------------------------------------------------------------------------- // Helper functions void printArray( char name[], int n, data_t arr[] ) { int i; if (coreid != 0) return; printf( " %10s :", name ); for ( i = 0; i < n; i++ ) printf( " %3ld ", (long) arr[i] ); printf( "\n" ); } void __attribute__((noinline)) verify(size_t n, const data_t* test, const data_t* correct) { if (coreid != 0) return; size_t i; for (i = 0; i < n; i++) { if (test[i] != correct[i]) { printf("FAILED test[%d]= %3ld, correct[%d]= %3ld\n", i, (long)test[i], i, (long)correct[i]); exit(-1); } } return; } //-------------------------------------------------------------------------- // matmul function // single-thread, naive version void __attribute__((noinline)) matmul_naive(const int lda, const data_t A[], const data_t B[], data_t C[] ) { int i, j, k; if (coreid > 0) return; for ( i = 0; i < lda; i++ ) for ( j = 0; j < lda; j++ ) { for ( k = 0; k < lda; k++ ) { C[i + j*lda] += A[j*lda + k] * B[k*lda + i]; } } } void __attribute__((noinline)) matmul(const int lda, const data_t A[], const data_t B[], data_t C[] ) { // ***************************** // // **** ADD YOUR CODE HERE ***** // // ***************************** // // // feel free to make a separate function for MI and MSI versions. int i, j, k; int temp0, temp1,temp2,temp3,temp4,temp5,temp6,temp7; int start = coreid*lda/2; int end = start + lda/2; int j_lda; int temp_i; int temp_A0, temp_A1, temp_A2, temp_A3 ; for ( i = start; i < end; i+=8){ for ( j = 0; j < lda; j++) { j_lda = j*lda; temp0 = C[(i+0) + j_lda]; temp1 = C[(i+1) + j_lda]; temp2 = C[(i+2) + j_lda]; temp3 = C[(i+3) + j_lda]; temp4 = C[(i+4) + j_lda]; temp5 = C[(i+5) + j_lda]; temp6 = C[(i+6) + j_lda]; temp7 = C[(i+7) + j_lda]; for ( k = 0; k < lda; k+=4) { temp_i = i; temp_A0 = A[j_lda + (k+0)] ; temp_A1 = A[j_lda + (k+1)] ; temp_A2 = A[j_lda + (k+2)] ; temp_A3 = A[j_lda + (k+3)] ; temp0 += temp_A0 * B[(k+0)*lda + temp_i]; temp0 += temp_A1 * B[(k+1)*lda + temp_i]; temp0 += temp_A2 * B[(k+2)*lda + temp_i]; temp0 += temp_A3 * B[(k+3)*lda + temp_i]; temp_i++; temp1 += temp_A0 * B[(k+0)*lda + temp_i]; temp1 += temp_A1 * B[(k+1)*lda + temp_i]; temp1 += temp_A2 * B[(k+2)*lda + temp_i]; temp1 += temp_A3 * B[(k+3)*lda + temp_i]; temp_i++; temp2 += temp_A0 * B[(k+0)*lda + temp_i]; temp2 += temp_A1 * B[(k+1)*lda + temp_i]; temp2 += temp_A2 * B[(k+2)*lda + temp_i]; temp2 += temp_A3 * B[(k+3)*lda + temp_i]; temp_i++; temp3 += temp_A0 * B[(k+0)*lda + temp_i]; temp3 += temp_A1 * B[(k+1)*lda + temp_i]; temp3 += temp_A2 * B[(k+2)*lda + temp_i]; temp3 += temp_A3 * B[(k+3)*lda + temp_i]; temp_i++; temp4 += temp_A0 * B[(k+0)*lda + temp_i]; temp4 += temp_A1 * B[(k+1)*lda + temp_i]; temp4 += temp_A2 * B[(k+2)*lda + temp_i]; temp4 += temp_A3 * B[(k+3)*lda + temp_i]; temp_i++; temp5 += temp_A0 * B[(k+0)*lda + temp_i]; temp5 += temp_A1 * B[(k+1)*lda + temp_i]; temp5 += temp_A2 * B[(k+2)*lda + temp_i]; temp5 += temp_A3 * B[(k+3)*lda + temp_i]; temp_i++; temp6 += temp_A0 * B[(k+0)*lda + temp_i]; temp6 += temp_A1 * B[(k+1)*lda + temp_i]; temp6 += temp_A2 * B[(k+2)*lda + temp_i]; temp6 += temp_A3 * B[(k+3)*lda + temp_i]; temp_i++; temp7 += temp_A0 * B[(k+0)*lda + temp_i]; temp7 += temp_A1 * B[(k+1)*lda + temp_i]; temp7 += temp_A2 * B[(k+2)*lda + temp_i]; temp7 += temp_A3 * B[(k+3)*lda + temp_i]; temp_i++; } C[i + j*lda] = temp0; C[(i+1) + j*lda] = temp1; C[(i+2) + j*lda] = temp2; C[(i+3) + j*lda] = temp3; C[(i+4) + j*lda] = temp4; C[(i+5) + j*lda] = temp5; C[(i+6) + j*lda] = temp6; C[(i+7) + j*lda] = temp7; } } } //-------------------------------------------------------------------------- // Main // // all threads start executing thread_entry(). Use their "coreid" to // differentiate between threads (each thread is running on a separate core). void thread_entry(int cid, int nc) { coreid = cid; ncores = nc; // static allocates data in the binary, which is visible to both threads static data_t results_data[ARRAY_SIZE]; /* // Execute the provided, naive matmul barrier(); stats(matmul_naive(DIM_SIZE, input1_data, input2_data, results_data); barrier()); // verify verify(ARRAY_SIZE, results_data, verify_data); // clear results from the first trial size_t i; if (coreid == 0) for (i=0; i < ARRAY_SIZE; i++) results_data[i] = 0; barrier(); */ // Execute your faster matmul barrier(); stats(matmul(DIM_SIZE, input1_data, input2_data, results_data); barrier()); #ifdef DEBUG printArray("results:", ARRAY_SIZE, results_data); printArray("verify :", ARRAY_SIZE, verify_data); #endif // verify verify(ARRAY_SIZE, results_data, verify_data); barrier(); exit(0); }