7fe737b10794bdc99dc36a567c7a6796f7b13fe8
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 printArray( 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
)) verify(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 int row
,row2
, column
, column2
, column3
, column4
, column5
, column6
, column7
, column8
;
114 size_t max_dim
= 32*32;
115 data_t element
, element2
, element3
, element4
, element5
, element6
, element7
, element8
;
116 data_t temp_mat
[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};
117 data_t temp_mat2
[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};
118 //for (i=coreid*max_dim/ncores; i<(max_dim/ncores+coreid*max_dim/ncores); i+=8){
119 for (l
=coreid
*32/ncores
; l
<32*(1+coreid
)/ncores
; l
+=2){
122 for (i
=0; i
<lda
; i
+=4){
124 element2
= A
[row
+i
+1];
125 element3
= A
[row
+i
+2];
126 element4
= A
[row
+i
+3];
127 element5
= A
[row2
+i
];
128 element6
= A
[row2
+i
+1];
129 element7
= A
[row2
+i
+2];
130 element8
= A
[row2
+i
+3];
135 for (j
=0; j
<32; j
+=4){
136 temp_mat
[j
]+=element
*B
[column
+j
]+element2
*B
[column2
+j
]+element3
*B
[column3
+j
]+element4
*B
[column4
+j
];
137 temp_mat
[j
+1]+=element
*B
[column
+j
+1]+element2
*B
[column2
+j
+1]+element3
*B
[column3
+j
+1]+element4
*B
[column4
+j
+1];
138 temp_mat
[j
+2]+=element
*B
[column
+j
+2]+element2
*B
[column2
+j
+2]+element3
*B
[column3
+j
+2]+element4
*B
[column4
+j
+2];
139 temp_mat
[j
+3]+=element
*B
[column
+j
+3]+element2
*B
[column2
+j
+3]+element3
*B
[column3
+j
+3]+element4
*B
[column4
+j
+3];
140 temp_mat2
[j
]+=element5
*B
[column
+j
]+element6
*B
[column2
+j
]+element7
*B
[column3
+j
]+element8
*B
[column4
+j
];
141 temp_mat2
[j
+1]+=element5
*B
[column
+j
+1]+element6
*B
[column2
+j
+1]+element7
*B
[column3
+j
+1]+element8
*B
[column4
+j
+1];
142 temp_mat2
[j
+2]+=element5
*B
[column
+j
+2]+element6
*B
[column2
+j
+2]+element7
*B
[column3
+j
+2]+element8
*B
[column4
+j
+2];
143 temp_mat2
[j
+3]+=element5
*B
[column
+j
+3]+element6
*B
[column2
+j
+3]+element7
*B
[column3
+j
+3]+element8
*B
[column4
+j
+3];
147 C[row+k]=temp_mat[k];
148 C[row2+k]=temp_mat2[k];
155 C
[row
+k
]=temp_mat
[k
];
156 C
[row2
+k
]=temp_mat2
[k
];
162 // ***************************** //
163 // **** ADD YOUR CODE HERE ***** //
164 // ***************************** //
166 // feel free to make a separate function for MI and MSI versions.
170 //--------------------------------------------------------------------------
173 // all threads start executing thread_entry(). Use their "coreid" to
174 // differentiate between threads (each thread is running on a separate core).
176 void thread_entry(int cid
, int nc
)
181 // static allocates data in the binary, which is visible to both threads
182 static data_t results_data
[ARRAY_SIZE
];
185 // // Execute the provided, naive matmul
187 // stats(matmul_naive(DIM_SIZE, input1_data, input2_data, results_data); barrier());
191 // verify(ARRAY_SIZE, results_data, verify_data);
193 // // clear results from the first trial
196 // for (i=0; i < ARRAY_SIZE; i++)
197 // results_data[i] = 0;
201 // Execute your faster matmul
203 stats(matmul(DIM_SIZE
, input1_data
, input2_data
, results_data
); barrier());
206 printArray("results:", ARRAY_SIZE
, results_data
);
207 printArray("verify :", ARRAY_SIZE
, verify_data
);
211 verify(ARRAY_SIZE
, results_data
, verify_data
);