5ca1dbe427c11ec7ec99df9057d5fa865be504f9
1 //**************************************************************************
2 // Multi-threaded Matrix Multiply benchmark
3 //--------------------------------------------------------------------------
4 // TA : Christopher Celio
5 // Student: Benjamin Han
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 // ***************************** //
114 // **** ADD YOUR CODE HERE ***** //
115 // ***************************** //
117 // feel free to make a separate function for MI and MSI versions.
118 int j2
, i2
, k2
, j
, i
, k
;
119 int tmpC00
, tmpC01
, tmpC02
, tmpC03
, tmpC04
, tmpC05
, tmpC06
, tmpC07
;
120 int tmpC10
, tmpC11
, tmpC12
, tmpC13
, tmpC14
, tmpC15
, tmpC16
, tmpC17
;
124 static __thread
int tB
[4096]; //__thread
126 int endInd
= lda
>> 1;
132 //tranpose B (block?)
133 for (i
= 0; i
< lda
; i
+= 2) {
134 for (j
= startInd
; j
< endInd
; j
+= 2) {
135 tB
[j
*lda
+ i
] = B
[i
*lda
+ j
];
136 tB
[(j
+ 1)*lda
+ i
] = B
[i
*lda
+ j
+ 1];
137 tB
[j
*lda
+ i
+ 1] = B
[(i
+ 1)*lda
+ j
];
138 tB
[(j
+ 1)*lda
+ i
+ 1] = B
[(i
+ 1)*lda
+ j
+ 1];
143 // compute C[j*n + i] += A[j*n + k] + Btranspose[i*n + k]
144 for ( j2
= 0; j2
< lda
; j2
+= jBLOCK
)
145 for ( i2
= startInd
; i2
< endInd
; i2
+= iBLOCK
)
146 for ( j
= j2
; j
< j2
+ jBLOCK
; j
+= 2 )
147 for ( k2
= 0; k2
< lda
; k2
+= kBLOCK
)
148 for ( i
= i2
; i
< i2
+ iBLOCK
; i
+= 4) {
149 tmpC00
= C
[j
*lda
+ i
+ 0]; tmpC10
= C
[(j
+ 1)*lda
+ i
+ 0];
150 tmpC01
= C
[j
*lda
+ i
+ 1]; tmpC11
= C
[(j
+ 1)*lda
+ i
+ 1];
151 tmpC02
= C
[j
*lda
+ i
+ 2]; tmpC12
= C
[(j
+ 1)*lda
+ i
+ 2];
152 tmpC03
= C
[j
*lda
+ i
+ 3]; tmpC13
= C
[(j
+ 1)*lda
+ i
+ 3];
153 //tmpC04 = C[j*lda + i + 4]; tmpC14 = C[(j + 1)*lda + i + 4];
154 //tmpC05 = C[j*lda + i + 5]; tmpC15 = C[(j + 1)*lda + i + 5];
155 //tmpC06 = C[j*lda + i + 6]; tmpC16 = C[(j + 1)*lda + i + 6];
156 //tmpC07 = C[j*lda + i + 7]; tmpC17 = C[(j + 1)*lda + i + 7];
157 for ( k
= k2
; k
< k2
+ kBLOCK
; k
+= 4) {
158 tmpC00
+= A
[j
*lda
+ k
] * tB
[(i
+ 0)*lda
+ k
];
159 tmpC01
+= A
[j
*lda
+ k
] * tB
[(i
+ 1)*lda
+ k
];
160 tmpC02
+= A
[j
*lda
+ k
] * tB
[(i
+ 2)*lda
+ k
];
161 tmpC03
+= A
[j
*lda
+ k
] * tB
[(i
+ 3)*lda
+ k
];
162 //tmpC04 += A[j*lda + k] * tB[(i + 4)*lda + k];
163 //tmpC05 += A[j*lda + k] * tB[(i + 5)*lda + k];
164 //tmpC06 += A[j*lda + k] * tB[(i + 6)*lda + k];
165 //tmpC07 += A[j*lda + k] * tB[(i + 7)*lda + k];
166 tmpC10
+= A
[(j
+ 1)*lda
+ k
] * tB
[(i
+ 0)*lda
+ k
];
167 tmpC11
+= A
[(j
+ 1)*lda
+ k
] * tB
[(i
+ 1)*lda
+ k
];
168 tmpC12
+= A
[(j
+ 1)*lda
+ k
] * tB
[(i
+ 2)*lda
+ k
];
169 tmpC13
+= A
[(j
+ 1)*lda
+ k
] * tB
[(i
+ 3)*lda
+ k
];
170 //tmpC14 += A[(j + 1)*lda + k] * tB[(i + 4)*lda + k];
171 //tmpC15 += A[(j + 1)*lda + k] * tB[(i + 5)*lda + k];
172 //tmpC16 += A[(j + 1)*lda + k] * tB[(i + 6)*lda + k];
173 //tmpC17 += A[(j + 1)*lda + k] * tB[(i + 7)*lda + k];
175 tmpC00
+= A
[j
*lda
+ k
+ 1] * tB
[(i
+ 0)*lda
+ k
+ 1];
176 tmpC01
+= A
[j
*lda
+ k
+ 1] * tB
[(i
+ 1)*lda
+ k
+ 1];
177 tmpC02
+= A
[j
*lda
+ k
+ 1] * tB
[(i
+ 2)*lda
+ k
+ 1];
178 tmpC03
+= A
[j
*lda
+ k
+ 1] * tB
[(i
+ 3)*lda
+ k
+ 1];
179 //tmpC04 += A[j*lda + k + 1] * tB[(i + 4)*lda + k + 1];
180 //tmpC05 += A[j*lda + k + 1] * tB[(i + 5)*lda + k + 1];
181 //tmpC06 += A[j*lda + k + 1] * tB[(i + 6)*lda + k + 1];
182 //tmpC07 += A[j*lda + k + 1] * tB[(i + 7)*lda + k + 1];
183 tmpC10
+= A
[(j
+ 1)*lda
+ k
+ 1] * tB
[(i
+ 0)*lda
+ k
+ 1];
184 tmpC11
+= A
[(j
+ 1)*lda
+ k
+ 1] * tB
[(i
+ 1)*lda
+ k
+ 1];
185 tmpC12
+= A
[(j
+ 1)*lda
+ k
+ 1] * tB
[(i
+ 2)*lda
+ k
+ 1];
186 tmpC13
+= A
[(j
+ 1)*lda
+ k
+ 1] * tB
[(i
+ 3)*lda
+ k
+ 1];
187 //tmpC14 += A[(j + 1)*lda + k + 1] * tB[(i + 4)*lda + k + 1];
188 //tmpC15 += A[(j + 1)*lda + k + 1] * tB[(i + 5)*lda + k + 1];
189 //tmpC16 += A[(j + 1)*lda + k + 1] * tB[(i + 6)*lda + k + 1];
190 //tmpC17 += A[(j + 1)*lda + k + 1] * tB[(i + 7)*lda + k + 1];
192 tmpC00
+= A
[j
*lda
+ k
+ 2] * tB
[(i
+ 0)*lda
+ k
+ 2];
193 tmpC01
+= A
[j
*lda
+ k
+ 2] * tB
[(i
+ 1)*lda
+ k
+ 2];
194 tmpC02
+= A
[j
*lda
+ k
+ 2] * tB
[(i
+ 2)*lda
+ k
+ 2];
195 tmpC03
+= A
[j
*lda
+ k
+ 2] * tB
[(i
+ 3)*lda
+ k
+ 2];
196 //tmpC04 += A[j*lda + k + 2] * tB[(i + 4)*lda + k + 2];
197 //tmpC05 += A[j*lda + k + 2] * tB[(i + 5)*lda + k + 2];
198 //tmpC06 += A[j*lda + k + 2] * tB[(i + 6)*lda + k + 2];
199 //tmpC07 += A[j*lda + k + 2] * tB[(i + 7)*lda + k + 2];
200 tmpC10
+= A
[(j
+ 1)*lda
+ k
+ 2] * tB
[(i
+ 0)*lda
+ k
+ 2];
201 tmpC11
+= A
[(j
+ 1)*lda
+ k
+ 2] * tB
[(i
+ 1)*lda
+ k
+ 2];
202 tmpC12
+= A
[(j
+ 1)*lda
+ k
+ 2] * tB
[(i
+ 2)*lda
+ k
+ 2];
203 tmpC13
+= A
[(j
+ 1)*lda
+ k
+ 2] * tB
[(i
+ 3)*lda
+ k
+ 2];
204 //tmpC14 += A[(j + 1)*lda + k + 2] * tB[(i + 4)*lda + k + 2];
205 //tmpC15 += A[(j + 1)*lda + k + 2] * tB[(i + 5)*lda + k + 2];
206 //tmpC16 += A[(j + 1)*lda + k + 2] * tB[(i + 6)*lda + k + 2];
207 //tmpC17 += A[(j + 1)*lda + k + 2] * tB[(i + 7)*lda + k + 2];
209 tmpC00
+= A
[j
*lda
+ k
+ 3] * tB
[(i
+ 0)*lda
+ k
+ 3];
210 tmpC01
+= A
[j
*lda
+ k
+ 3] * tB
[(i
+ 1)*lda
+ k
+ 3];
211 tmpC02
+= A
[j
*lda
+ k
+ 3] * tB
[(i
+ 2)*lda
+ k
+ 3];
212 tmpC03
+= A
[j
*lda
+ k
+ 3] * tB
[(i
+ 3)*lda
+ k
+ 3];
213 //tmpC04 += A[j*lda + k + 3] * tB[(i + 4)*lda + k + 3];
214 //tmpC05 += A[j*lda + k + 3] * tB[(i + 5)*lda + k + 3];
215 //tmpC06 += A[j*lda + k + 3] * tB[(i + 6)*lda + k + 3];
216 //tmpC07 += A[j*lda + k + 3] * tB[(i + 7)*lda + k + 3];
217 tmpC10
+= A
[(j
+ 1)*lda
+ k
+ 3] * tB
[(i
+ 0)*lda
+ k
+ 3];
218 tmpC11
+= A
[(j
+ 1)*lda
+ k
+ 3] * tB
[(i
+ 1)*lda
+ k
+ 3];
219 tmpC12
+= A
[(j
+ 1)*lda
+ k
+ 3] * tB
[(i
+ 2)*lda
+ k
+ 3];
220 tmpC13
+= A
[(j
+ 1)*lda
+ k
+ 3] * tB
[(i
+ 3)*lda
+ k
+ 3];
221 //tmpC14 += A[(j + 1)*lda + k + 3] * tB[(i + 4)*lda + k + 3];
222 //tmpC15 += A[(j + 1)*lda + k + 3] * tB[(i + 5)*lda + k + 3];
223 //tmpC16 += A[(j + 1)*lda + k + 3] * tB[(i + 6)*lda + k + 3];
224 //tmpC17 += A[(j + 1)*lda + k + 3] * tB[(i + 7)*lda + k + 3];
226 C
[j
*lda
+ i
+ 0] = tmpC00
; C
[(j
+ 1)*lda
+ i
+ 0] = tmpC10
;
227 C
[j
*lda
+ i
+ 1] = tmpC01
; C
[(j
+ 1)*lda
+ i
+ 1] = tmpC11
;
228 C
[j
*lda
+ i
+ 2] = tmpC02
; C
[(j
+ 1)*lda
+ i
+ 2] = tmpC12
;
229 C
[j
*lda
+ i
+ 3] = tmpC03
; C
[(j
+ 1)*lda
+ i
+ 3] = tmpC13
;
230 //C[j*lda + i + 4] = tmpC04; C[(j + 1)*lda + i + 4] = tmpC14;
231 //C[j*lda + i + 5] = tmpC05; C[(j + 1)*lda + i + 5] = tmpC15;
232 //C[j*lda + i + 6] = tmpC06; C[(j + 1)*lda + i + 6] = tmpC16;
233 //C[j*lda + i + 7] = tmpC07; C[(j + 1)*lda + i + 7] = tmpC17;
237 //--------------------------------------------------------------------------
240 // all threads start executing thread_entry(). Use their "coreid" to
241 // differentiate between threads (each thread is running on a separate core).
243 void thread_entry(int cid
, int nc
)
248 // static allocates data in the binary, which is visible to both threads
249 static data_t results_data
[ARRAY_SIZE
];
252 // // Execute the provided, naive matmul
254 // stats(matmul_naive(DIM_SIZE, input1_data, input2_data, results_data); barrier());
258 // verify(ARRAY_SIZE, results_data, verify_data);
260 // // clear results from the first trial
263 // for (i=0; i < ARRAY_SIZE; i++)
264 // results_data[i] = 0;
268 // Execute your faster matmul
270 stats(matmul(DIM_SIZE
, input1_data
, input2_data
, results_data
); barrier());
273 printArray("results:", ARRAY_SIZE
, results_data
);
274 printArray("verify :", ARRAY_SIZE
, verify_data
);
278 verify(ARRAY_SIZE
, results_data
, verify_data
);