1 //**************************************************************************
2 // Multi-threaded Matrix Multiply benchmark
3 //--------------------------------------------------------------------------
4 // TA : Christopher Celio
5 // Student: Felix Li $ Ronald Lee
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 data_t element1
, element2
, element3
, element4
, element5
, element6
, element7
, element8
;
115 int column1
, column2
, column3
, column4
, column5
, column6
, column7
, column8
;
116 data_t temp
[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 temp2
[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};
119 for (i
=0; i
<32; i
+=2){
122 for (j
=0; j
<16; j
+=4){
124 element2
= A
[row
+j
+1];
125 element3
= A
[row
+j
+2];
126 element4
= A
[row
+j
+3];
131 element5
= A
[row2
+j
];
132 element6
= A
[row2
+j
+1];
133 element7
= A
[row2
+j
+2];
134 element8
= A
[row2
+j
+3];
136 for (k
=0; k
<32; k
+=4){
137 temp
[k
]+=element1
*B
[column1
+k
]+element2
*B
[column2
+k
]+element3
*B
[column3
+k
]+element4
*B
[column4
+k
];
138 temp
[k
+1]+=element1
*B
[column1
+k
+1]+element2
*B
[column2
+k
+1]+element3
*B
[column3
+k
+1]+element4
*B
[column4
+k
+1];
139 temp
[k
+2]+=element1
*B
[column1
+k
+2]+element2
*B
[column2
+k
+2]+element3
*B
[column3
+k
+2]+element4
*B
[column4
+k
+2];
140 temp
[k
+3]+=element1
*B
[column1
+k
+3]+element2
*B
[column2
+k
+3]+element3
*B
[column3
+k
+3]+element4
*B
[column4
+k
+3];
141 temp2
[k
]+=element5
*B
[column1
+k
]+element6
*B
[column2
+k
]+element7
*B
[column3
+k
]+element8
*B
[column4
+k
];
142 temp2
[k
+1]+=element5
*B
[column1
+k
+1]+element6
*B
[column2
+k
+1]+element7
*B
[column3
+k
+1]+element8
*B
[column4
+k
+1];
143 temp2
[k
+2]+=element5
*B
[column1
+k
+2]+element6
*B
[column2
+k
+2]+element7
*B
[column3
+k
+2]+element8
*B
[column4
+k
+2];
144 temp2
[k
+3]+=element5
*B
[column1
+k
+3]+element6
*B
[column2
+k
+3]+element7
*B
[column3
+k
+3]+element8
*B
[column4
+k
+3];
149 for (l
=0; l
<32; l
++){
159 for (i
=0; i
<32; i
+=2){
162 for (j
=16; j
<32; j
+=4){
164 element2
= A
[row
+j
+1];
165 element3
= A
[row
+j
+2];
166 element4
= A
[row
+j
+3];
167 element5
= A
[row2
+j
];
168 element6
= A
[row2
+j
+1];
169 element7
= A
[row2
+j
+2];
170 element8
= A
[row2
+j
+3];
175 for (k
=0; k
<32; k
+=4){
176 temp
[k
]+=element1
*B
[column1
+k
]+element2
*B
[column2
+k
]+element3
*B
[column3
+k
]+element4
*B
[column4
+k
];
177 temp
[k
+1]+=element1
*B
[column1
+k
+1]+element2
*B
[column2
+k
+1]+element3
*B
[column3
+k
+1]+element4
*B
[column4
+k
+1];
178 temp
[k
+2]+=element1
*B
[column1
+k
+2]+element2
*B
[column2
+k
+2]+element3
*B
[column3
+k
+2]+element4
*B
[column4
+k
+2];
179 temp
[k
+3]+=element1
*B
[column1
+k
+3]+element2
*B
[column2
+k
+3]+element3
*B
[column3
+k
+3]+element4
*B
[column4
+k
+3];
180 temp2
[k
]+=element5
*B
[column1
+k
]+element6
*B
[column2
+k
]+element7
*B
[column3
+k
]+element8
*B
[column4
+k
];
181 temp2
[k
+1]+=element5
*B
[column1
+k
+1]+element6
*B
[column2
+k
+1]+element7
*B
[column3
+k
+1]+element8
*B
[column4
+k
+1];
182 temp2
[k
+2]+=element5
*B
[column1
+k
+2]+element6
*B
[column2
+k
+2]+element7
*B
[column3
+k
+2]+element8
*B
[column4
+k
+2];
183 temp2
[k
+3]+=element5
*B
[column1
+k
+3]+element6
*B
[column2
+k
+3]+element7
*B
[column3
+k
+3]+element8
*B
[column4
+k
+3];
189 for (l
=0; l
<32; l
++){
197 // ***************************** //
198 // **** ADD YOUR CODE HERE ***** //
199 // ***************************** //
201 // feel free to make a separate function for MI and MSI versions.
205 //--------------------------------------------------------------------------
208 // all threads start executing thread_entry(). Use their "coreid" to
209 // differentiate between threads (each thread is running on a separate core).
211 void thread_entry(int cid
, int nc
)
216 // static allocates data in the binary, which is visible to both threads
217 static data_t results_data
[ARRAY_SIZE
];
220 // // Execute the provided, naive matmul
222 // stats(matmul_naive(DIM_SIZE, input1_data, input2_data, results_data); barrier());
226 // verify(ARRAY_SIZE, results_data, verify_data);
228 // // clear results from the first trial
231 // for (i=0; i < ARRAY_SIZE; i++)
232 // results_data[i] = 0;
236 // Execute your faster matmul
238 stats(matmul(DIM_SIZE
, input1_data
, input2_data
, results_data
); barrier());
241 printArray("results:", ARRAY_SIZE
, results_data
);
242 printArray("verify :", ARRAY_SIZE
, verify_data
);
246 verify(ARRAY_SIZE
, results_data
, verify_data
);