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 printArrayMT( 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
)) verifyMT(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
];
110 void __attribute__((noinline
)) matmul(const int lda
, const data_t A
[], const data_t B
[], data_t C
[] )
118 size_t max_dim
= lda
*lda
;
119 size_t block_size
= lda
/2;
121 data_t temp_mat1
[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};
123 for (k
= 0; k
< lda
/2; k
++) {
124 int columnIndex
= 32*k
;
126 //temp_mat1 will store the kth column of B
127 for (i
= 0; i
< lda
; i
++) {
128 temp_mat1
[i
] = B
[32*i
+ k
];
131 for (j
=0; j
< lda
; j
++) {
133 //iterate through each element of A in row J and accumulate result
134 for (i2
= 0; i2
<lda
; i2
+= 4) {
135 int elementA
= A
[rowIndex
+i2
];
136 int elementA2
= A
[rowIndex
+i2
+1];
137 int elementA3
= A
[rowIndex
+i2
+2];
138 int elementA4
= A
[rowIndex
+i2
+3];
139 result
+= elementA
*temp_mat1
[i2
] + elementA2
*temp_mat1
[i2
+1] + elementA3
*temp_mat1
[i2
+2] + elementA4
*temp_mat1
[i2
+3] ;
141 C
[k
+rowIndex
] = result
;
147 for (k
= lda
/2; k
< lda
; k
++) {
148 int columnIndex
= 32*k
;
150 //temp_mat1 will store the kth column of B
151 for (i
= 0; i
< lda
; i
++) {
152 temp_mat1
[i
] = B
[32*i
+ k
];
155 for (j
=0; j
< lda
; j
++) {
157 //iterate through each element of A in row J and accumulate result
158 for (i2
= 0; i2
<lda
; i2
+= 4) {
159 int elementA
= A
[rowIndex
+i2
];
160 int elementA2
= A
[rowIndex
+i2
+1];
161 int elementA3
= A
[rowIndex
+i2
+2];
162 int elementA4
= A
[rowIndex
+i2
+3];
163 result
+= elementA
*temp_mat1
[i2
] + elementA2
*temp_mat1
[i2
+1] + elementA3
*temp_mat1
[i2
+2] + elementA4
*temp_mat1
[i2
+3] ;
165 C
[k
+rowIndex
] = result
;
176 // ***************************** //
177 // **** ADD YOUR CODE HERE ***** //
178 // ***************************** //
180 // feel free to make a separate function for MI and MSI versions.
184 //--------------------------------------------------------------------------
187 // all threads start executing thread_entry(). Use their "coreid" to
188 // differentiate between threads (each thread is running on a separate core).
190 void thread_entry(int cid
, int nc
)
195 // static allocates data in the binary, which is visible to both threads
196 static data_t results_data
[ARRAY_SIZE
];
199 // Execute the provided, naive matmul
201 stats(matmul_naive(DIM_SIZE
, input1_data
, input2_data
, results_data
); barrier(nc
));
205 verifyMT(ARRAY_SIZE
, results_data
, verify_data
);
207 // clear results from the first trial
210 for (i
=0; i
< ARRAY_SIZE
; i
++)
215 // Execute your faster matmul
217 stats(matmul(DIM_SIZE
, input1_data
, input2_data
, results_data
); barrier(nc
));
220 printArrayMT("results:", ARRAY_SIZE
, results_data
);
221 printArrayMT("verify :", ARRAY_SIZE
, verify_data
);
225 verifyMT(ARRAY_SIZE
, results_data
, verify_data
);