60aa2e7c4efb5e4fb396290fd5f5793dbcfe02bb
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
27 typedef double data_t
;
31 //--------------------------------------------------------------------------
32 // Basic Utilities and Multi-thread Support
34 __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/DATA_SIZE, 10*_c/DATA_SIZE%10, _c/_i, 10*_c/_i%10); \
51 //--------------------------------------------------------------------------
54 //perform in-place vvadd
55 void __attribute__((noinline
)) vvadd(size_t n
, data_t
* __restrict__ x
, const data_t
* __restrict__ y
)
59 // interleave accesses
60 for (i
= coreid
; i
< n
; i
+=ncores
)
66 void __attribute__((noinline
)) vvadd_opt(size_t n
, data_t
* __restrict__ x
, const data_t
* __restrict__ y
)
68 // ***************************** //
69 // **** ADD YOUR CODE HERE ***** //
70 // ***************************** //
73 //--------------------------------------------------------------------------
76 // all threads start executing thread_entry(). Use their "coreid" to
77 // differentiate between threads (each thread is running on a separate core).
79 void thread_entry(int cid
, int nc
)
84 // static allocates data in the binary, which is visible to both threads
85 static data_t results_data
[DATA_SIZE
];
87 // because we're going to perform an in-place vvadd (and we're going to run
88 // it a couple of times) let's copy the input data to a temporary results
94 for (i
= 0; i
< DATA_SIZE
; i
++)
95 results_data
[i
] = input1_data
[i
];
99 // Execute the provided, terrible vvadd
101 stats(vvadd(DATA_SIZE
, results_data
, input2_data
); barrier());
105 int res
= verifyDouble(DATA_SIZE
, results_data
, verify_data
);
110 // reset results from the first trial
113 for (i
=0; i
< DATA_SIZE
; i
++)
114 results_data
[i
] = input1_data
[i
];
118 // Execute your faster vvadd
120 stats(vvadd_opt(DATA_SIZE
, results_data
, input2_data
); barrier());
123 printDoubleArray("results: ", DATA_SIZE
, results_data
);
124 printDoubleArray("verify : ", DATA_SIZE
, verify_data
);
128 res
= verifyDouble(DATA_SIZE
, results_data
, verify_data
);