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
[] )
114 for ( i
= 0; i
< lda
; i
+=2 )
116 for (k
= 0; k
< lda
; k
+=4)
118 int d0
= B
[k
*lda
+ i
];
119 int c0
= B
[k
*lda
+ i
+ 1];
120 int d1
= B
[(k
+1)*lda
+ i
];
121 int c1
= B
[(k
+1)*lda
+ i
+ 1];
122 int d2
= B
[(k
+2)*lda
+ i
];
123 int c2
= B
[(k
+2)*lda
+ i
+ 1];
124 int d3
= B
[(k
+3)*lda
+ i
];
125 int c3
= B
[(k
+3)*lda
+ i
+ 1];
127 for ( j
= coreid
*(lda
/ncores
); j
< (coreid
+1)*(lda
/ncores
); j
+=4)
130 int sum
= A
[j
*lda
+ k
] * d0
;
131 sum
+= A
[j
*lda
+ k
+ 1] * d1
;
132 sum
+= A
[j
*lda
+ k
+ 2] * d2
;
133 sum
+= A
[j
*lda
+ k
+ 3] * d3
;
136 sum
= A
[j
*lda
+ k
] * c0
;
137 sum
+= A
[j
*lda
+ k
+ 1] * c1
;
138 sum
+= A
[j
*lda
+ k
+ 2] * c2
;
139 sum
+= A
[j
*lda
+ k
+ 3] * c3
;
140 C
[j
*lda
+ i
+ 1] += sum
;
142 sum
= A
[(j
+1)*lda
+ k
] * d0
;
143 sum
+= A
[(j
+1)*lda
+ k
+ 1] * d1
;
144 sum
+= A
[(j
+1)*lda
+ k
+ 2] * d2
;
145 sum
+= A
[(j
+1)*lda
+ k
+ 3] * d3
;
146 C
[(j
+1)*lda
+i
] += sum
;
148 sum
= A
[(j
+1)*lda
+ k
] * c0
;
149 sum
+= A
[(j
+1)*lda
+ k
+ 1] * c1
;
150 sum
+= A
[(j
+1)*lda
+ k
+ 2] * c2
;
151 sum
+= A
[(j
+1)*lda
+ k
+ 3] * c3
;
152 C
[(j
+1)*lda
+ i
+ 1] += sum
;
154 sum
= A
[(j
+2)*lda
+ k
] * d0
;
155 sum
+= A
[(j
+2)*lda
+ k
+ 1] * d1
;
156 sum
+= A
[(j
+2)*lda
+ k
+ 2] * d2
;
157 sum
+= A
[(j
+2)*lda
+ k
+ 3] * d3
;
158 C
[(j
+2)*lda
+i
] += sum
;
160 sum
= A
[(j
+2)*lda
+ k
] * c0
;
161 sum
+= A
[(j
+2)*lda
+ k
+ 1] * c1
;
162 sum
+= A
[(j
+2)*lda
+ k
+ 2] * c2
;
163 sum
+= A
[(j
+2)*lda
+ k
+ 3] * c3
;
164 C
[(j
+2)*lda
+ i
+ 1] += sum
;
166 sum
= A
[(j
+3)*lda
+ k
] * d0
;
167 sum
+= A
[(j
+3)*lda
+ k
+ 1] * d1
;
168 sum
+= A
[(j
+3)*lda
+ k
+ 2] * d2
;
169 sum
+= A
[(j
+3)*lda
+ k
+ 3] * d3
;
170 C
[(j
+3)*lda
+i
] += sum
;
172 sum
= A
[(j
+3)*lda
+ k
] * c0
;
173 sum
+= A
[(j
+3)*lda
+ k
+ 1] * c1
;
174 sum
+= A
[(j
+3)*lda
+ k
+ 2] * c2
;
175 sum
+= A
[(j
+3)*lda
+ k
+ 3] * c3
;
176 C
[(j
+3)*lda
+ i
+ 1] += sum
;
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++)
211 // results_data[i] = 0;
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
);