b1c0a39b07200b6ab75b65ad2add4cacc02304f5
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
[] )
98 for ( i
= 0; i
< lda
; i
++ )
99 for ( j
= 0; j
< lda
; j
++ )
101 for ( k
= 0; k
< lda
; k
++ )
103 C
[i
+ j
*lda
] += A
[j
*lda
+ k
] * B
[k
*lda
+ i
];
110 for ( i = 0; i < 16; i+=8 )
112 for ( j = 0; j < 32; j++ )
122 for ( kk = 0; kk < 32; kk+=8 )
123 for ( k = kk; k < kk+8; k++ )
124 // for ( k = 0; k < 32; k++ )
126 data_t tempA = A[j*32+k];
127 temp0 += tempA * B[k*32 + i];
128 temp1 += tempA * B[k*32 + i+1];
129 temp2 += tempA * B[k*32 + i+2];
130 temp3 += tempA * B[k*32 + i+3];
131 temp4 += tempA * B[k*32 + i+4];
132 temp5 += tempA * B[k*32 + i+5];
133 temp6 += tempA * B[k*32 + i+6];
134 temp7 += tempA * B[k*32 + i+7];
147 for ( i = 16; i < 32; i+=8 )
149 for ( j = 0; j < 32; j++ )
159 for ( kk = 0; kk < 32; kk+=8 )
160 for ( k = kk; k < kk+8; k++ )
162 data_t tempA = A[j*32+k];
163 temp0 += tempA * B[k*32 + i];
164 temp1 += tempA * B[k*32 + i+1];
165 temp2 += tempA * B[k*32 + i+2];
166 temp3 += tempA * B[k*32 + i+3];
167 temp4 += tempA * B[k*32 + i+4];
168 temp5 += tempA * B[k*32 + i+5];
169 temp6 += tempA * B[k*32 + i+6];
170 temp7 += tempA * B[k*32 + i+7];
191 void __attribute__((noinline
)) matmul(const int lda
, const data_t A
[], const data_t B
[], data_t C
[] )
194 // ***************************** //
195 // **** ADD YOUR CODE HERE ***** //
196 // ***************************** //
198 // feel free to make a separate function for MI and MSI versions.
199 int i
, j
, k
, ii
, jj
, kk
;
201 // for ( ii = 0; ii < 32; ii+=IC )
202 for ( jj
= 0; jj
< 16; jj
+=16 )
203 for ( kk
= 0; kk
< 32; kk
+=16 )
204 for ( j
= jj
; j
< jj
+16 && j
< 16; j
++ )
205 // for ( j = 0; j < 16; j++ )
207 for ( i
= 0; i
< 32; i
+=8 )
208 // for ( i = ii; i < ii + IC && i < 32; i+=8 )
210 data_t temp0
= C
[i
+j
*32];
211 data_t temp1
= C
[i
+j
*32+1];
212 data_t temp2
= C
[i
+j
*32+2];
213 data_t temp3
= C
[i
+j
*32+3];
214 data_t temp4
= C
[i
+j
*32+4];
215 data_t temp5
= C
[i
+j
*32+5];
216 data_t temp6
= C
[i
+j
*32+6];
217 data_t temp7
= C
[i
+j
*32+7];
218 for ( k
= kk
; k
< kk
+16 && k
< 32; k
++ )
219 // for ( k = 0; k < 32; k++ )
221 data_t tempA
= A
[j
*32+k
];
222 temp0
+= tempA
* B
[k
*32 + i
];
223 temp1
+= tempA
* B
[k
*32 + i
+1];
224 temp2
+= tempA
* B
[k
*32 + i
+2];
225 temp3
+= tempA
* B
[k
*32 + i
+3];
226 temp4
+= tempA
* B
[k
*32 + i
+4];
227 temp5
+= tempA
* B
[k
*32 + i
+5];
228 temp6
+= tempA
* B
[k
*32 + i
+6];
229 temp7
+= tempA
* B
[k
*32 + i
+7];
242 // for ( ii = 0; ii < 32; ii+=IC )
243 for ( jj
= 16; jj
< 32; jj
+= 16 ) {
244 for ( kk
= 16; kk
< 32; kk
+=16 )
245 for ( j
= jj
; j
< jj
+16 && j
< 32; j
++ )
246 // for ( j = 16; j < 32; j++ )
248 for ( i
= 0; i
< 32; i
+=8 )
249 // for ( i = ii; i < ii + IC && i < 32; i+=8 )
251 data_t temp0
= C
[i
+j
*32];
252 data_t temp1
= C
[i
+j
*32+1];
253 data_t temp2
= C
[i
+j
*32+2];
254 data_t temp3
= C
[i
+j
*32+3];
255 data_t temp4
= C
[i
+j
*32+4];
256 data_t temp5
= C
[i
+j
*32+5];
257 data_t temp6
= C
[i
+j
*32+6];
258 data_t temp7
= C
[i
+j
*32+7];
259 for ( k
= kk
; k
< kk
+16 && k
< 32; k
++ )
261 data_t tempA
= A
[j
*32+k
];
262 temp0
+= tempA
* B
[k
*32 + i
];
263 temp1
+= tempA
* B
[k
*32 + i
+1];
264 temp2
+= tempA
* B
[k
*32 + i
+2];
265 temp3
+= tempA
* B
[k
*32 + i
+3];
266 temp4
+= tempA
* B
[k
*32 + i
+4];
267 temp5
+= tempA
* B
[k
*32 + i
+5];
268 temp6
+= tempA
* B
[k
*32 + i
+6];
269 temp7
+= tempA
* B
[k
*32 + i
+7];
282 for ( kk
= 0; kk
< 16; kk
+=16 )
283 for ( j
= jj
; j
< jj
+16 && j
< 32; j
++ )
284 // for ( j = 16; j < 32; j++ )
286 for ( i
= 0; i
< 32; i
+=8 )
287 // for ( i = ii; i < ii + IC && i < 32; i+=8 )
289 data_t temp0
= C
[i
+j
*32];
290 data_t temp1
= C
[i
+j
*32+1];
291 data_t temp2
= C
[i
+j
*32+2];
292 data_t temp3
= C
[i
+j
*32+3];
293 data_t temp4
= C
[i
+j
*32+4];
294 data_t temp5
= C
[i
+j
*32+5];
295 data_t temp6
= C
[i
+j
*32+6];
296 data_t temp7
= C
[i
+j
*32+7];
297 for ( k
= kk
; k
< kk
+16 && k
< 32; k
++ )
299 data_t tempA
= A
[j
*32+k
];
300 temp0
+= tempA
* B
[k
*32 + i
];
301 temp1
+= tempA
* B
[k
*32 + i
+1];
302 temp2
+= tempA
* B
[k
*32 + i
+2];
303 temp3
+= tempA
* B
[k
*32 + i
+3];
304 temp4
+= tempA
* B
[k
*32 + i
+4];
305 temp5
+= tempA
* B
[k
*32 + i
+5];
306 temp6
+= tempA
* B
[k
*32 + i
+6];
307 temp7
+= tempA
* B
[k
*32 + i
+7];
324 //--------------------------------------------------------------------------
327 // all threads start executing thread_entry(). Use their "coreid" to
328 // differentiate between threads (each thread is running on a separate core).
330 void thread_entry(int cid
, int nc
)
335 // static allocates data in the binary, which is visible to both threads
336 static data_t results_data
[ARRAY_SIZE
];
339 // // Execute the provided, naive matmul
341 // stats(matmul_naive(DIM_SIZE, input1_data, input2_data, results_data); barrier());
345 // verify(ARRAY_SIZE, results_data, verify_data);
347 // // clear results from the first trial
350 // for (i=0; i < ARRAY_SIZE; i++)
351 // results_data[i] = 0;
355 // Execute your faster matmul
357 stats(matmul(DIM_SIZE
, input1_data
, input2_data
, results_data
); barrier());
360 printArray("results:", ARRAY_SIZE
, results_data
);
361 printArray("verify :", ARRAY_SIZE
, verify_data
);
365 verify(ARRAY_SIZE
, results_data
, verify_data
);