批改娘 10105. Multiple Device (OpenCL)

contents

  1. 1. 題目描述
  2. 2. 輸入格式
  3. 3. 輸出格式
  4. 4. Sample Input
  5. 5. Sample Output
  6. 6. Solution
    1. 6.1. main.c
    2. 6.2. matrix-lib.cl

題目描述

小明的數學作業要計算方陣,現在請你幫幫他!

題目給定數個 $N \times N$ 的矩陣和 $2$ 小題。

  • $X = AB+CD$
  • $Y = ABE+CDF$

輸入格式

多組測資,每組第一行會有一個整數 $N$,表示題目給定 $N \times N$ 矩陣,第二行上會有 $6$ 個整數,分別為矩陣 $A, B, C, D, E, F$ 的生成種子。

  • $1 \le N \le 1024$
  • $0 \le S_i \le 2^{31}$

輸出格式

輸出兩行 $X$ 和 $Y$ 的雜湊值,可參考 sequence.c 的流程。

Sample Input

1
2
3
4
2
0 1 2 3 4 5
10
0 1 2 3 4 5

Sample Output

1
2
3
4
2385860290
1374821695
617438354
1897844131

Solution

這一題要充分實作使用 real-time 分配工作到沒有運行的 GPU 上,利用在 OpenMP 學到的平行技巧,讓多個 thread 等待工作,一抓到工作立即運行。

main.c

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
#include <stdio.h>
#include <assert.h>
#include <inttypes.h>
#include <string.h>
#include <signal.h>
#include <unistd.h>
#include <CL/cl.h>
#include <omp.h>
#define MAXGPU 3
#define MAXN 1024
uint32_t hostMtx[MAXGPU][6][MAXN*MAXN];
uint32_t hostMid[MAXGPU][2][MAXN*MAXN];
char clSrcFormat[32767] = "";
char clSrc[32767] = "";
// -- start working with OpenCL
const int clNeedDevCnt = 3;
cl_context clCtx[MAXGPU];
cl_program clPrg[MAXGPU];
cl_kernel clKrnAdd[MAXGPU], clKrnMul[MAXGPU];
cl_command_queue clQue[MAXGPU];
cl_mem clMtx[MAXGPU][6], clMtxTmp[MAXGPU][6];

#define CheckFailAndExit(status) \
if (status != CL_SUCCESS) { \
fprintf(stderr, "Error %d: Line %u in file %s\n\n", status, __LINE__, __FILE__), \
destroyGPU(clCtx, clPrg, clKrnAdd, clKrnMul, clQue, clMtx, clMtxTmp); \
}

#define clFuncArgs cl_context clCtx[], cl_program clPrg[], cl_kernel clKrnAdd[], \
cl_kernel clKrnMul[], cl_command_queue clQue[], cl_mem clMtx[][6], cl_mem clMtxTmp[][6]

#define clCallFunc clCtx, clPrg, clKrnAdd, clKrnMul, clQue, clMtx, clMtxTmp
#define clCallFuncOuter clCtx, clPrg, clKrnAdd, clKrnMul, clQue, clMtx, clMtxTmp

uint32_t writeOut(uint32_t *hostC, int N) {
uint32_t h = 0;
uint32_t *Cend = hostC + N*N, *C = hostC;
for (; C != Cend; C++)
h = (h + *C) * 2654435761LU;
return h;
}
void destroyGPU(clFuncArgs) {
fprintf(stderr, "Starting Cleanup ...\n\n");
for (int i = 0; i < clNeedDevCnt; i++) {
for (int j = 0; j < 6; j++) {
if (clMtx[i][j])
clReleaseMemObject(clMtx[i][j]);
if (clMtxTmp[i][j])
clReleaseMemObject(clMtxTmp[i][j]);

}

if (clKrnAdd[i])
clReleaseKernel(clKrnAdd[i]);
if (clKrnMul[i])
clReleaseKernel(clKrnMul[i]);
if (clPrg[i])
clReleaseProgram(clPrg[i]);
if (clQue[i])
clReleaseCommandQueue(clQue[i]);
if (clCtx[i])
clReleaseContext(clCtx[i]);
}
exit(0);
}
int initAllGPU(char fileName[], clFuncArgs) {
// -- generate kernel code
FILE *codefin = fopen(fileName, "r");
assert(codefin != NULL);
assert(fread(clSrcFormat, 1, 32767, codefin) < 32767);
sprintf(clSrc, clSrcFormat);
size_t clSrcLen = strlen(clSrc);
fclose(codefin);

cl_int clStat;
cl_uint clPlatN, clGPUN, clDevN;
cl_platform_id clPlatID;
cl_device_id clGPUID[MAXGPU];
const char *clSrcPtr = clSrc;

// -- basic OpenCL setup
clGetPlatformIDs(1, &clPlatID, &clPlatN);
clGetDeviceIDs(clPlatID, CL_DEVICE_TYPE_GPU, MAXGPU, clGPUID, &clDevN);
assert(clDevN >= clNeedDevCnt);
for (int i = 0; i < clNeedDevCnt; i++) {
clCtx[i] = clCreateContext(NULL, 1, clGPUID+i, NULL, NULL, &clStat);
CheckFailAndExit(clStat);
clQue[i] = clCreateCommandQueue(clCtx[i], clGPUID[i], 0, &clStat);
CheckFailAndExit(clStat);
clPrg[i] = clCreateProgramWithSource(clCtx[i], 1, &clSrcPtr, &clSrcLen, &clStat);
CheckFailAndExit(clStat);
clStat = clBuildProgram(clPrg[i], 1, clGPUID+i, NULL, NULL, NULL);
if (clStat != CL_SUCCESS) {
fprintf(stderr, "Error: Line %u in file %s\n\n", __LINE__, __FILE__);
size_t log_size;
clGetProgramBuildInfo(*clPrg, clGPUID[0],
CL_PROGRAM_BUILD_LOG, 0, NULL, &log_size);
char *program_log = (char *) calloc(log_size+1, sizeof(char));
clGetProgramBuildInfo(*clPrg, clGPUID[0],
CL_PROGRAM_BUILD_LOG, log_size+1, program_log, NULL);
printf("%s", program_log);
free(program_log);
CheckFailAndExit(CL_BUILD_PROGRAM_FAILURE);
}
clKrnAdd[i] = clCreateKernel(clPrg[i], "matrixAdd", &clStat);
CheckFailAndExit(clStat);
clKrnMul[i] = clCreateKernel(clPrg[i], "matrixMul", &clStat);
CheckFailAndExit(clStat);

for (int j = 0; j < 6; j++) {
clMtx[i][j] = clCreateBuffer(clCtx[i], CL_MEM_READ_WRITE,
sizeof(uint32_t)*MAXN*MAXN, NULL, &clStat);
CheckFailAndExit(clStat);
clMtxTmp[i][j] = clCreateBuffer(clCtx[i], CL_MEM_READ_WRITE,
sizeof(uint32_t)*MAXN*MAXN, NULL, &clStat);
CheckFailAndExit(clStat);
}
}

return 1;
}
void matrix_mul(int N, int devIdx, cl_mem *LIN, cl_mem *RIN, cl_mem *OUT, clFuncArgs) {
cl_int clStat;
size_t globalOffset[] = {0};
size_t globalSize[] = {N*N};
size_t localSize[] = {0};
for (int i = 1; i <= N; i++) {
if (N%i == 0 && i*N <= 32768/2)
localSize[0] = i;
}
// -- set argument to kernel
clStat = clSetKernelArg(clKrnMul[devIdx], 0, sizeof(cl_mem), LIN);
CheckFailAndExit(clStat);
clStat = clSetKernelArg(clKrnMul[devIdx], 1, sizeof(cl_mem), RIN);
CheckFailAndExit(clStat);
clStat = clSetKernelArg(clKrnMul[devIdx], 2, sizeof(cl_mem), OUT);
CheckFailAndExit(clStat);
clStat = clSetKernelArg(clKrnMul[devIdx], 3, sizeof(cl_int), &N);

CheckFailAndExit(clStat);

// -- execute
clStat = clEnqueueNDRangeKernel(clQue[devIdx], clKrnMul[devIdx], 1, globalOffset,
globalSize, NULL, 0, NULL, NULL);
CheckFailAndExit(clStat);
}
void matrix_add(int N, int devIdx, cl_mem *LIN, cl_mem *RIN, cl_mem *OUT, clFuncArgs) {
cl_int clStat;
size_t globalOffset[] = {0};
size_t globalSize[] = {N*N};
// -- set argument to kernel
clStat = clSetKernelArg(clKrnAdd[devIdx], 0, sizeof(cl_mem), LIN);
CheckFailAndExit(clStat);
clStat = clSetKernelArg(clKrnAdd[devIdx], 1, sizeof(cl_mem), RIN);
CheckFailAndExit(clStat);
clStat = clSetKernelArg(clKrnAdd[devIdx], 2, sizeof(cl_mem), OUT);
CheckFailAndExit(clStat);

// -- execute
clStat = clEnqueueNDRangeKernel(clQue[devIdx], clKrnAdd[devIdx], 1, globalOffset,
globalSize, NULL, 0, NULL, NULL);
CheckFailAndExit(clStat);
}
int solver(int N, int devId, uint32_t ret[], clFuncArgs) {
uint32_t memSz = N*N*sizeof(uint32_t);
cl_int clStat;
for (int i = 0; i < 6; i++) {
clStat = clEnqueueWriteBuffer(clQue[devId],
clMtx[devId][i], 0, 0, memSz,
hostMtx[devId][i], 0, NULL, NULL);
CheckFailAndExit(clStat);
}
// cuMtxTmp[0] = AB
matrix_mul(N, devId, &clMtx[devId][0], &clMtx[devId][1], &clMtxTmp[devId][0], clCallFunc);
// cuMtxTmp[1] = CD
matrix_mul(N, devId, &clMtx[devId][2], &clMtx[devId][3], &clMtxTmp[devId][1], clCallFunc);
// cuMtxTmp[2] = ABE
matrix_mul(N, devId, &clMtxTmp[devId][0], &clMtx[devId][4], &clMtxTmp[devId][2], clCallFunc);
// cuMtxTmp[3] = CDF
matrix_mul(N, devId, &clMtxTmp[devId][1], &clMtx[devId][5], &clMtxTmp[devId][3], clCallFunc);
// cuMtxTmp[4] = AB + CD
matrix_add(N, devId, &clMtxTmp[devId][0], &clMtxTmp[devId][1], &clMtxTmp[devId][4], clCallFunc);
// cuMtxTmp[5] = ABE+CDF
matrix_add(N, devId, &clMtxTmp[devId][2], &clMtxTmp[devId][3], &clMtxTmp[devId][5], clCallFunc);

clStat = clEnqueueReadBuffer(clQue[devId], clMtxTmp[devId][4], CL_TRUE, 0,
sizeof(uint32_t)*N*N, hostMid[devId][0], 0, NULL, NULL);
CheckFailAndExit(clStat);
clStat = clEnqueueReadBuffer(clQue[devId], clMtxTmp[devId][5], CL_TRUE, 0,
sizeof(uint32_t)*N*N, hostMid[devId][1], 0, NULL, NULL);
CheckFailAndExit(clStat);

for (int i = 0; i < 2; i++)
#pragma omp task
{
ret[i] = writeOut(hostMid[devId][i], N);
}
#pragma omp taskwait
return 1;
}
int readIn(uint32_t S[], int *n, int devId) {
int N, M;
if (scanf("%d", &N) != 1)
return 0;
M = 6;
for (int i = 0; i < M; i++)
assert(scanf("%d", &S[i]) == 1);

for (int p = 0; p < M; p++)
#pragma omp task
{
uint32_t x = 2, n = N*N, c = S[p];
x = 2;
for (int i = 0; i < N; i++) {
for (int j = 0; j < N; j++) {
x = (x * x + c + i + j)%n;
hostMtx[devId][p][i*N+j] = x;
}
}
}
#pragma omp taskwait
*n = N;
return 1;
}
void onStart(clFuncArgs) {
initAllGPU("matrix-lib.cl", clCallFunc);
int inN = 0;
static uint32_t ansQue[32767][2];
#pragma omp parallel sections
{
#pragma omp section
{
while (1) {
int f = 0, N, pid = 0;
uint32_t S[32];
#pragma omp critical
{
f = readIn(S, &N, 0);
pid = inN;
inN += f;
}
if (f == 0)
break;
solver(N, 0, ansQue[pid], clCallFunc);
}
}

#pragma omp section
{
while (1) {
int f = 0, N, pid = 0;
uint32_t S[32];
#pragma omp critical
{
f = readIn(S, &N, 1);
pid = inN;
inN += f;
}
if (f == 0)
break;
solver(N, 1, ansQue[pid], clCallFunc);
}
}
#pragma omp section
{
while (1) {
int f = 0, N, pid = 0;
uint32_t S[32];
#pragma omp critical
{
f = readIn(S, &N, 2);
pid = inN;
inN += f;
}
if (f == 0)
break;
solver(N, 2, ansQue[pid], clCallFunc);
}
}

}
for (int i = 0; i < inN; i++)
printf("%u\n%u\n", ansQue[i][0], ansQue[i][1]);
destroyGPU(clCallFunc);
}

void sigHandler(int signo) {
printf("God Bless Me\n");
destroyGPU(clCallFuncOuter);
exit(0);
}
int main(int argc, char *argv[]) {
const char sigErr[] = "I can't catch signal.\n";
if (signal(SIGTRAP, sigHandler) == SIG_ERR)
fprintf(stderr, sigErr);
if (signal(SIGSEGV, sigHandler) == SIG_ERR)
fprintf(stderr, sigErr);
if (signal(SIGILL, sigHandler) == SIG_ERR)
fprintf(stderr, sigErr);
if (signal(SIGFPE, sigHandler) == SIG_ERR)
fprintf(stderr, sigErr);
if (signal(SIGINT, sigHandler) == SIG_ERR)
fprintf(stderr, sigErr);

onStart(clCallFuncOuter);
return 0;
}

matrix-lib.cl

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
#define CTYPE unsigned int

__kernel void matrixAdd(__global CTYPE *in1,
__global CTYPE *in2,
__global CTYPE *out)
{

int x = get_global_id(0);
out[x] = in1[x] + in2[x];
}

__kernel void matrixMul(__global CTYPE *in1,
__global CTYPE *in2,
__global CTYPE *out, int N)
{

int id = get_global_id(0);
int x = id / N, y = id % N;
CTYPE sum = 0;
for (int i = 0; i < N; i++)
sum += in1[x*N + i] * in2[i*N + y];
out[x * N + y] = sum;
}