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 //--------------------------------------------------------------------------
89 // single-thread, naive version
90 void __attribute__((noinline)) matmul_naive(const int lda, const data_t A[], const data_t B[], data_t C[] )
97 for ( i = 0; i < lda; i++ )
98 for ( j = 0; j < lda; j++ )
100 for ( k = 0; k < lda; k++ )
102 C[i + j*lda] += A[j*lda + k] * B[k*lda + i];
109 void __attribute__((noinline)) matmul(const int lda, const data_t A[], const data_t B[], data_t C[] )
112 data_t element1, element2, element3, element4, element5, element6, element7, element8;
114 int column1, column2, column3, column4, column5, column6, column7, column8;
115 data_t temp[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};
116 data_t temp2[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};
118 for (i=0; i<32; i+=2){
121 for (j=0; j<16; j+=4){
123 element2 = A[row+j+1];
124 element3 = A[row+j+2];
125 element4 = A[row+j+3];
130 element5 = A[row2+j];
131 element6 = A[row2+j+1];
132 element7 = A[row2+j+2];
133 element8 = A[row2+j+3];
135 for (k=0; k<32; k+=4){
136 temp[k]+=element1*B[column1+k]+element2*B[column2+k]+element3*B[column3+k]+element4*B[column4+k];
137 temp[k+1]+=element1*B[column1+k+1]+element2*B[column2+k+1]+element3*B[column3+k+1]+element4*B[column4+k+1];
138 temp[k+2]+=element1*B[column1+k+2]+element2*B[column2+k+2]+element3*B[column3+k+2]+element4*B[column4+k+2];
139 temp[k+3]+=element1*B[column1+k+3]+element2*B[column2+k+3]+element3*B[column3+k+3]+element4*B[column4+k+3];
140 temp2[k]+=element5*B[column1+k]+element6*B[column2+k]+element7*B[column3+k]+element8*B[column4+k];
141 temp2[k+1]+=element5*B[column1+k+1]+element6*B[column2+k+1]+element7*B[column3+k+1]+element8*B[column4+k+1];
142 temp2[k+2]+=element5*B[column1+k+2]+element6*B[column2+k+2]+element7*B[column3+k+2]+element8*B[column4+k+2];
143 temp2[k+3]+=element5*B[column1+k+3]+element6*B[column2+k+3]+element7*B[column3+k+3]+element8*B[column4+k+3];
146 for (l=0; l<32; l++){
157 for (i=0; i<32; i+=2){
160 for (j=16; j<32; j+=4){
162 element2 = A[row+j+1];
163 element3 = A[row+j+2];
164 element4 = A[row+j+3];
165 element5 = A[row2+j];
166 element6 = A[row2+j+1];
167 element7 = A[row2+j+2];
168 element8 = A[row2+j+3];
173 for (k=0; k<32; k+=4){
174 temp[k]+=element1*B[column1+k]+element2*B[column2+k]+element3*B[column3+k]+element4*B[column4+k];
175 temp[k+1]+=element1*B[column1+k+1]+element2*B[column2+k+1]+element3*B[column3+k+1]+element4*B[column4+k+1];
176 temp[k+2]+=element1*B[column1+k+2]+element2*B[column2+k+2]+element3*B[column3+k+2]+element4*B[column4+k+2];
177 temp[k+3]+=element1*B[column1+k+3]+element2*B[column2+k+3]+element3*B[column3+k+3]+element4*B[column4+k+3];
178 temp2[k]+=element5*B[column1+k]+element6*B[column2+k]+element7*B[column3+k]+element8*B[column4+k];
179 temp2[k+1]+=element5*B[column1+k+1]+element6*B[column2+k+1]+element7*B[column3+k+1]+element8*B[column4+k+1];
180 temp2[k+2]+=element5*B[column1+k+2]+element6*B[column2+k+2]+element7*B[column3+k+2]+element8*B[column4+k+2];
181 temp2[k+3]+=element5*B[column1+k+3]+element6*B[column2+k+3]+element7*B[column3+k+3]+element8*B[column4+k+3];
184 for (l=0; l<32; l++){
194 // ***************************** //
195 // **** ADD YOUR CODE HERE ***** //
196 // ***************************** //
198 // feel free to make a separate function for MI and MSI versions.
201 //--------------------------------------------------------------------------
204 // all threads start executing thread_entry(). Use their "coreid" to
205 // differentiate between threads (each thread is running on a separate core).
207 void thread_entry(int cid, int nc)
212 // static allocates data in the binary, which is visible to both threads
213 static data_t results_data[ARRAY_SIZE];
216 // Execute the provided, naive matmul
218 stats(matmul_naive(DIM_SIZE, input1_data, input2_data, results_data); barrier());
222 verify(ARRAY_SIZE, results_data, verify_data);
224 // clear results from the first trial
227 for (i=0; i < ARRAY_SIZE; i++)
232 // Execute your faster matmul
234 stats(matmul(DIM_SIZE, input1_data, input2_data, results_data); barrier());
237 printArray("results:", ARRAY_SIZE, results_data);
238 printArray("verify :", ARRAY_SIZE, verify_data);
242 verify(ARRAY_SIZE, results_data, verify_data);