e73150124a058e00667e4543e82ed6fb936f8e56
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
[] )
113 // ***************************** //
114 // **** ADD YOUR CODE HERE ***** //
115 // ***************************** //
117 // feel free to make a separate function for MI and MSI versions.
118 int i
, j
, k
, ii
, jj
, bsize
, start
;
120 start
= bsize
*coreid
;
121 for ( jj
= start
; jj
< lda
; jj
+= bsize
*ncores
) {
123 for ( ii
= start
; ii
!=start
|| first
; ii
=(bsize
+ii
) % lda
) {
125 for ( j
= jj
; j
< lda
&& j
< jj
+ bsize
; j
+=4) {
126 for ( i
= ii
; i
< lda
&& i
< ii
+ bsize
; i
+=2) {
127 data_t c1
= C
[i
+ j
*lda
];
128 data_t c2
= C
[i
+ j
*lda
+ 1];
129 data_t c3
= C
[i
+ (j
+1)*lda
];
130 data_t c4
= C
[i
+ (j
+1)*lda
+ 1];
131 data_t c5
= C
[i
+ (j
+2)*lda
];
132 data_t c6
= C
[i
+ (j
+2)*lda
+ 1];
133 data_t c7
= C
[i
+ (j
+3)*lda
];
134 data_t c8
= C
[i
+ (j
+3)*lda
+ 1];
135 for ( k
= 0; k
< lda
; k
+=8){
136 for (int x
= 0; x
< 8; x
++) {
137 data_t a
= A
[j
*lda
+ k
+x
];
138 data_t a1
= A
[(j
+1)*lda
+k
+x
];
139 data_t a2
= A
[(j
+2)*lda
+k
+x
];
140 data_t a3
= A
[(j
+3)*lda
+k
+x
];
141 data_t b1
= B
[(k
+x
)*lda
+ i
];
142 data_t b2
= B
[(k
+x
)*lda
+ i
+ 1];
154 C
[i
+ j
*lda
+ 1] = c2
;
155 C
[i
+ (j
+1)*lda
] = c3
;
156 C
[i
+ (j
+1)*lda
+ 1] = c4
;
157 C
[i
+ (j
+2)*lda
] = c5
;
158 C
[i
+ (j
+2)*lda
+ 1] = c6
;
159 C
[i
+ (j
+3)*lda
] = c7
;
160 C
[i
+ (j
+3)*lda
+ 1] = c8
;
166 //--------------------------------------------------------------------------
169 // all threads start executing thread_entry(). Use their "coreid" to
170 // differentiate between threads (each thread is running on a separate core).
172 void thread_entry(int cid
, int nc
)
177 // static allocates data in the binary, which is visible to both threads
178 static data_t results_data
[ARRAY_SIZE
];
181 // Execute the provided, naive matmul
183 stats(matmul_naive(DIM_SIZE
, input1_data
, input2_data
, results_data
); barrier(nc
));
187 verifyMT(ARRAY_SIZE
, results_data
, verify_data
);
189 // clear results from the first trial
192 for (i
=0; i
< ARRAY_SIZE
; i
++)
197 // Execute your faster matmul
199 stats(matmul(DIM_SIZE
, input1_data
, input2_data
, results_data
); barrier(nc
));
202 printArrayMT("results:", ARRAY_SIZE
, results_data
);
203 printArrayMT("verify :", ARRAY_SIZE
, verify_data
);
207 verifyMT(ARRAY_SIZE
, results_data
, verify_data
);