1 //**************************************************************************
2 // Multi-threaded Matrix Multiply benchmark
3 //--------------------------------------------------------------------------
4 // TA : Christopher Celio
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
14 // print out arrays, etc.
17 //--------------------------------------------------------------------------
25 //--------------------------------------------------------------------------
26 // Input/Reference Data
32 //--------------------------------------------------------------------------
33 // Basic Utilities and Multi-thread Support
35 __thread unsigned long coreid;
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(); \
45 _c += rdcycle(), _i += rdinstret(); \
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); \
52 //--------------------------------------------------------------------------
55 void printArray( char name[], int n, data_t arr[] )
61 printf( " %10s :", name );
62 for ( i = 0; i < n; i++ )
63 printf( " %3ld ", (long) arr[i] );
67 void __attribute__((noinline)) verify(size_t n, const data_t* test, const data_t* correct)
73 for (i = 0; i < n; i++)
75 if (test[i] != correct[i])
77 printf("FAILED test[%d]= %3ld, correct[%d]= %3ld\n",
78 i, (long)test[i], i, (long)correct[i]);
86 data_t mult(data_t x, data_t y)
89 for (i=0; i < x; i++) {
94 //--------------------------------------------------------------------------
97 // single-thread, naive version
98 void __attribute__((noinline)) matmul_naive(const int lda, const data_t A[], const data_t B[], data_t C[] )
105 for ( i = 0; i < lda; i++ )
106 for ( j = 0; j < lda; j++ )
108 for ( k = 0; k < lda; k++ )
110 C[i + j*lda] += A[j*lda + k] * B[k*lda + i];
117 void __attribute__((noinline)) matmul(const int lda, const data_t A[], const data_t B[], data_t C[] )
120 int row,row2, column, column2, column3, column4, column5, column6, column7, column8;
121 data_t element, element2, element3, element4, element5, element6, element7, element8;
122 data_t B1, B2, B3, B4;
123 data_t temp_mat[32]={0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0};
124 data_t temp_mat2[32]={0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0};
126 //for (i=coreid*max_dim/ncores; i<(max_dim/ncores+coreid*max_dim/ncores); i+=8){
127 for (l=coreid*local_lda/ncores; l<local_lda*(1+coreid)/ncores; l+=2){
131 //element5 = A[row2];
132 for (i=0; i<local_lda; i+=4){
134 element2 = A[row+i+1];
135 element3 = A[row+i+2];
136 element4 = A[row+i+3];
138 element5 = A[row2+i];
139 element6 = A[row2+i+1];
140 element7 = A[row2+i+2];
141 element8 = A[row2+i+3];
144 column2=(i+1)*local_lda;
145 column3=(i+2)*local_lda;
146 column4=(i+3)*local_lda;
153 for (j=0; j<lda; j+=4){
154 temp_mat[j]+=element*B1+element2*B2+element3*B3+element4*B4;
155 temp_mat[j+1]+=element*B[column+j+1]+element2*B[column2+j+1]+element3*B[column3+j+1]+element4*B[column4+j+1];
156 temp_mat[j+2]+=element*B[column+j+2]+element2*B[column2+j+2]+element3*B[column3+j+2]+element4*B[column4+j+2];
157 temp_mat[j+3]+=element*B[column+j+3]+element2*B[column2+j+3]+element3*B[column3+j+3]+element4*B[column4+j+3];
159 temp_mat2[j]+=element5*B1+element6*B2+element7*B3+element8*B4;
160 temp_mat2[j+1]+=element5*B[column+j+1]+element6*B[column2+j+1]+element7*B[column3+j+1]+element8*B[column4+j+1];
161 temp_mat2[j+2]+=element5*B[column+j+2]+element6*B[column2+j+2]+element7*B[column3+j+2]+element8*B[column4+j+2];
162 temp_mat2[j+3]+=element5*B[column+j+3]+element6*B[column2+j+3]+element7*B[column3+j+3]+element8*B[column4+j+3];
170 //element = A[row+i+4];
171 //element5 = A[row2+i+4];
174 for(k=0; k<local_lda; k++){
175 C[row+k]=temp_mat[k];
177 C[row2+k]=temp_mat2[k];
184 // ***************************** //
185 // **** ADD YOUR CODE HERE ***** //
186 // ***************************** //
188 // feel free to make a separate function for MI and MSI versions.
192 //--------------------------------------------------------------------------
195 // all threads start executing thread_entry(). Use their "coreid" to
196 // differentiate between threads (each thread is running on a separate core).
198 void thread_entry(int cid, int nc)
203 // static allocates data in the binary, which is visible to both threads
204 static data_t results_data[ARRAY_SIZE];
207 // Execute the provided, naive matmul
209 stats(matmul_naive(DIM_SIZE, input1_data, input2_data, results_data); barrier());
213 verify(ARRAY_SIZE, results_data, verify_data);
215 // clear results from the first trial
218 for (i=0; i < ARRAY_SIZE; i++)
223 // Execute your faster matmul
225 stats(matmul(DIM_SIZE, input1_data, input2_data, results_data); barrier());
228 printArray("results:", ARRAY_SIZE, results_data);
229 printArray("verify :", ARRAY_SIZE, verify_data);
233 verify(ARRAY_SIZE, results_data, verify_data);