//************************************************************************** // 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[] ) { size_t c_start = lda / ncores * coreid; size_t c_row; size_t c_col; size_t colSplit = 0; size_t i; size_t useSplit = 0; data_t a1; data_t a2; data_t a3; data_t a4; data_t a5; data_t a6; data_t a7; data_t a8; data_t c1; data_t c2; data_t c3; data_t c4; data_t c5; data_t c6; data_t c7; data_t c8; size_t block; for (block = 0; block < 2; block++) { for (colSplit = 0; colSplit < 4; colSplit++) { useSplit = (coreid == 0) ? colSplit : (colSplit + 2 ) % 4; for (c_row = c_start + block * 8; c_row < c_start + block * 8 + 8; c_row += 2) { for (c_col = 0; c_col < lda; c_col+=4) { c1 = C[c_row*lda+c_col]; c2 = C[(c_row+1)*lda+c_col]; c3 = C[c_row*lda+c_col+1]; c4 = C[(c_row+1)*lda+c_col+1]; c5 = C[c_row*lda+c_col+2]; c6 = C[(c_row+1)*lda+c_col+2]; c7 = C[c_row*lda+c_col+3]; c8 = C[(c_row+1)*lda+c_col+3]; for (i = useSplit * lda / 4; i < (useSplit + 1) * lda / 4; i+=4) { a1 = A[c_row*lda+i]; a2 = A[(c_row+1)*lda+i]; a3 = A[c_row*lda+i+1]; a4 = A[(c_row+1)*lda+i+1]; a5 = A[c_row*lda+i+2]; a6 = A[(c_row+1)*lda+i+2]; a7 = A[c_row*lda+i+3]; a8 = A[(c_row+1)*lda+i+3]; c1 += a1 * B[i*lda+c_col]; c2 += a2 * B[i*lda+c_col]; c1 += a3 * B[(i+1)*lda+c_col]; c2 += a4 * B[(i+1)*lda+c_col]; c1 += a5 * B[(i+2)*lda+c_col]; c2 += a6 * B[(i+2)*lda+c_col]; c1 += a7 * B[(i+3)*lda+c_col]; c2 += a8 * B[(i+3)*lda+c_col]; c3 += a1 * B[i*lda+c_col+1]; c4 += a2 * B[i*lda+c_col+1]; c3 += a3 * B[(i+1)*lda+c_col+1]; c4 += a4 * B[(i+1)*lda+c_col+1]; c3 += a5 * B[(i+2)*lda+c_col+1]; c4 += a6 * B[(i+2)*lda+c_col+1]; c3 += a7 * B[(i+3)*lda+c_col+1]; c4 += a8 * B[(i+3)*lda+c_col+1]; c5 += a1 * B[i*lda+c_col+2]; c6 += a2 * B[i*lda+c_col+2]; c5 += a3 * B[(i+1)*lda+c_col+2]; c6 += a4 * B[(i+1)*lda+c_col+2]; c5 += a5 * B[(i+2)*lda+c_col+2]; c6 += a6 * B[(i+2)*lda+c_col+2]; c5 += a7 * B[(i+3)*lda+c_col+2]; c6 += a8 * B[(i+3)*lda+c_col+2]; c7 += a1 * B[i*lda+c_col+3]; c8 += a2 * B[i*lda+c_col+3]; c7 += a3 * B[(i+1)*lda+c_col+3]; c8 += a4 * B[(i+1)*lda+c_col+3]; c7 += a5 * B[(i+2)*lda+c_col+3]; c8 += a6 * B[(i+2)*lda+c_col+3]; c7 += a7 * B[(i+3)*lda+c_col+3]; c8 += a8 * B[(i+3)*lda+c_col+3]; } C[c_row*lda+c_col] = c1; C[(c_row+1)*lda+c_col] = c2; C[c_row*lda+c_col+1] = c3; C[(c_row+1)*lda+c_col+1] = c4; C[c_row*lda+c_col+2] = c5; C[(c_row+1)*lda+c_col+2] = c6; C[c_row*lda+c_col+3] = c7; C[(c_row+1)*lda+c_col+3] = c8; } } } } } //-------------------------------------------------------------------------- // 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); }