30703df8f4b07600cf46f560fae385fc971df0e9
1 //**************************************************************************
2 // Vector-vector add benchmark
3 //--------------------------------------------------------------------------
4 // Author : Andrew Waterman
5 // TA : Christopher Celio
8 // This benchmark adds two vectors and writes the results to a
9 // third vector. The input data (and reference data) should be
10 // generated using the vvadd_gendata.pl perl script and dumped
11 // to a file named dataset.h
13 // to print out arrays, etc.
16 //--------------------------------------------------------------------------
24 //--------------------------------------------------------------------------
25 // Input/Reference Data
31 //--------------------------------------------------------------------------
32 // Basic Utilities and Multi-thread Support
34 __thread
unsigned long coreid
;
39 #define stringify_1(s) #s
40 #define stringify(s) stringify_1(s)
41 #define stats(code) do { \
42 unsigned long _c = -rdcycle(), _i = -rdinstret(); \
44 _c += rdcycle(), _i += rdinstret(); \
46 printf("%s: %ld cycles, %ld.%ld cycles/iter, %ld.%ld CPI\n", \
47 stringify(code), _c, _c/DATA_SIZE, 10*_c/DATA_SIZE%10, _c/_i, 10*_c/_i%10); \
51 //--------------------------------------------------------------------------
54 void printArray( char name
[], int n
, data_t arr
[] )
60 printf( " %10s :", name
);
61 for ( i
= 0; i
< n
; i
++ )
62 printf( " %4ld ", (long) arr
[i
] );
66 void __attribute__((noinline
)) verify(size_t n
, const data_t
* test
, const data_t
* correct
)
72 for (i
= 0; i
< n
; i
++)
74 if (test
[i
] != correct
[i
])
76 printf("FAILED test[%d]= %4ld, correct[%d]= %4ld\n",
77 i
, (long) test
[i
], i
, (long)correct
[i
]);
85 //--------------------------------------------------------------------------
88 //perform in-place vvadd
89 void __attribute__((noinline
)) vvadd(size_t n
, data_t
* __restrict__ x
, const data_t
* __restrict__ y
)
93 // interleave accesses
94 for (i
= coreid
; i
< n
; i
+=ncores
)
100 void __attribute__((noinline
)) vvadd_opt(size_t n
, data_t
* __restrict__ x
, const data_t
* __restrict__ y
)
103 size_t start
= n
* coreid
/ ncores
;
104 size_t end
= (coreid
== ncores
- 1) ? n
: n
* (coreid
+1)/ ncores
;
105 for (i
= start
; i
< end
; i
++) {
110 //--------------------------------------------------------------------------
113 // all threads start executing thread_entry(). Use their "coreid" to
114 // differentiate between threads (each thread is running on a separate core).
116 void thread_entry(int cid
, int nc
)
121 // static allocates data in the binary, which is visible to both threads
122 static data_t results_data
[DATA_SIZE
];
124 // because we're going to perform an in-place vvadd (and we're going to run
125 // it a couple of times) let's copy the input data to a temporary results
131 for (i
= 0; i
< DATA_SIZE
; i
++)
132 results_data
[i
] = input1_data
[i
];
136 // Execute the provided, terrible vvadd
138 stats(vvadd(DATA_SIZE
, results_data
, input2_data
); barrier());
142 verify(DATA_SIZE
, results_data
, verify_data
);
144 // reset results from the first trial
147 for (i
=0; i
< DATA_SIZE
; i
++)
148 results_data
[i
] = input1_data
[i
];
153 // Execute your faster vvadd
155 stats(vvadd_opt(DATA_SIZE
, results_data
, input2_data
); barrier());
158 printArray("results: ", DATA_SIZE
, results_data
);
159 printArray("verify : ", DATA_SIZE
, verify_data
);
163 verify(DATA_SIZE
, results_data
, verify_data
);