|
0
|
1 #include <stdio.h>
|
|
|
2 #include <stdlib.h>
|
|
|
3 #include <string.h>
|
|
|
4 #include <sys/stat.h>
|
|
|
5 #include <sys/mman.h>
|
|
|
6 #include <sys/types.h>
|
|
|
7 #include <fcntl.h>
|
|
|
8 #include <unistd.h>
|
|
|
9 #include <OpenCL/opencl.h>
|
|
|
10 #include <oclUtils.h>
|
|
|
11
|
|
1
|
12 #define OUT_PARAM_NUM 4
|
|
|
13 #define PRINT_PARAM_NUM 2
|
|
0
|
14
|
|
|
15 typedef struct {
|
|
|
16 caddr_t file_mmap;
|
|
|
17 off_t size;
|
|
|
18 } st_mmap_t;
|
|
|
19
|
|
1
|
20 typedef struct {
|
|
|
21 int work_num;
|
|
|
22 int size;
|
|
|
23 int remain_size;
|
|
|
24 } param_t;
|
|
0
|
25
|
|
|
26
|
|
|
27 static st_mmap_t
|
|
|
28 my_mmap(char *filename)
|
|
|
29 {
|
|
|
30
|
|
|
31 /*マッピングだよ!*/
|
|
|
32 int fd = -1;
|
|
|
33 int map = MAP_PRIVATE;
|
|
|
34 st_mmap_t st_mmap;
|
|
|
35 struct stat sb;
|
|
|
36
|
|
|
37 if ((fd=open(filename,O_RDONLY,0666))==0) {
|
|
|
38 fprintf(stderr,"can't open %s\n",filename);
|
|
|
39 }
|
|
|
40
|
|
|
41 if (fstat(fd,&sb)) {
|
|
|
42 fprintf(stderr,"can't fstat %s\n",filename);
|
|
|
43 }
|
|
|
44
|
|
1
|
45 st_mmap.size = sb.st_size;
|
|
0
|
46
|
|
1
|
47 printf("file size %d\n",(int)st_mmap.size);
|
|
0
|
48
|
|
|
49 st_mmap.file_mmap = (char*)mmap(NULL,st_mmap.size,PROT_READ,map,fd,(off_t)0);
|
|
|
50 if (st_mmap.file_mmap == (caddr_t)-1) {
|
|
|
51 fprintf(stderr,"Can't mmap file\n");
|
|
|
52 perror(NULL);
|
|
|
53 exit(0);
|
|
|
54 }
|
|
|
55
|
|
|
56 return st_mmap;
|
|
|
57
|
|
|
58 }
|
|
|
59
|
|
|
60
|
|
|
61
|
|
|
62 int main(int args, char *argv[])
|
|
|
63 {
|
|
|
64
|
|
|
65 char *filename = 0;
|
|
|
66
|
|
|
67 for (int i = 1; argv[i]; ++i) {
|
|
|
68 if (strcmp(argv[i], "-file") == 0) {
|
|
|
69 filename = argv[i+1];
|
|
|
70 } else if (strcmp(argv[i], "-help")) {
|
|
|
71 printf("Usage: ./word_count [-file filename]\n");
|
|
|
72 }
|
|
|
73 }
|
|
|
74
|
|
|
75 if (filename == 0) {
|
|
|
76 printf("Usage: ./word_count [-file filename]\n");
|
|
|
77 return 0;
|
|
|
78 }
|
|
|
79
|
|
|
80 //指定されたファイルをメモリにmap
|
|
|
81 st_mmap_t st_mmap = my_mmap(filename);
|
|
|
82
|
|
|
83 //kernelファイルの大きさ取得して、メモリ確保
|
|
|
84 int fd = -1;
|
|
|
85 const char *kernel_filename = "./word_count.cl";
|
|
|
86 if ((fd=open(kernel_filename,O_RDONLY,0666))==0) {
|
|
|
87 fprintf(stderr,"can't open %s\n",kernel_filename);
|
|
|
88 }
|
|
|
89
|
|
|
90 struct stat sb;
|
|
|
91 if (fstat(fd,&sb)) {
|
|
|
92 fprintf(stderr,"can't fstat %s\n",filename);
|
|
|
93 }
|
|
|
94
|
|
|
95 size_t source_size = sb.st_size;
|
|
|
96 void *source_str = malloc(source_size);
|
|
|
97
|
|
|
98 int err = read(fd, source_str, source_size);
|
|
|
99 if (err == -1) {
|
|
|
100 fprintf(stderr,"can't read %s\n",filename);
|
|
|
101 }
|
|
|
102
|
|
|
103 close(fd);
|
|
|
104
|
|
|
105 cl_platform_id platform_id = NULL;
|
|
|
106 cl_uint ret_num_platforms = NULL;
|
|
|
107 cl_int ret = clGetPlatformIDs(1, &platform_id, &ret_num_platforms);
|
|
|
108
|
|
|
109 cl_device_id device_id = NULL;
|
|
|
110 cl_uint ret_num_devices = NULL;
|
|
|
111 // CL_DEVICE_TYPE_CPU, CL_DEVICE_TYPE_GPU, と指定できる
|
|
|
112 // CL_DEVICE_TYPE_DEFAULT はどうなるのか
|
|
1
|
113 ret = clGetDeviceIDs( platform_id, CL_DEVICE_TYPE_CPU,
|
|
0
|
114 1, &device_id, &ret_num_devices);
|
|
|
115
|
|
|
116 oclCheckError(ret, CL_SUCCESS);
|
|
|
117
|
|
|
118 //OpenCLコンテキストの作成
|
|
|
119 cl_context context = clCreateContext( NULL, 1, &device_id, NULL, NULL, &ret);
|
|
|
120 oclCheckError(ret, CL_SUCCESS);
|
|
|
121 cl_command_queue command_queue = clCreateCommandQueue(context, device_id, 0, &ret);
|
|
|
122 oclCheckError(ret, CL_SUCCESS);
|
|
|
123 // カーネルプログラムを読み込む
|
|
|
124 cl_program program = clCreateProgramWithSource(context, 1, (const char **)&source_str,
|
|
|
125 (const size_t *)&source_size, &ret);
|
|
1
|
126
|
|
0
|
127 ret = clBuildProgram(program, 1, &device_id, NULL, NULL, NULL);
|
|
|
128
|
|
1
|
129 if (ret != CL_SUCCESS) {
|
|
|
130 oclLogBuildInfo(program, device_id);
|
|
|
131 }
|
|
|
132
|
|
0
|
133 oclCheckError(ret, CL_SUCCESS);
|
|
|
134
|
|
|
135 //カーネルプログラムをビルド
|
|
|
136 //Task選択にあたる
|
|
|
137 cl_kernel kernel = clCreateKernel(program, "word_count", &ret);
|
|
|
138 oclCheckError(ret, CL_SUCCESS);
|
|
|
139
|
|
|
140 //カウントするテキストデータのメモリオブジェクト
|
|
|
141 cl_mem text_memobj = clCreateBuffer(context, CL_MEM_READ_ONLY,
|
|
|
142 st_mmap.size * sizeof(char), NULL, &ret);
|
|
|
143
|
|
|
144 ret = clEnqueueWriteBuffer(command_queue, text_memobj,
|
|
|
145 CL_TRUE, 0, st_mmap.size, (char*)st_mmap.file_mmap,
|
|
|
146 0, NULL, NULL);
|
|
|
147
|
|
|
148 oclCheckError(ret, CL_SUCCESS);
|
|
|
149
|
|
|
150 // 必要なパラメータのオブジェクト
|
|
|
151 cl_mem param_memobj = clCreateBuffer(context, CL_MEM_READ_ONLY,
|
|
1
|
152 sizeof(param_t), NULL, &ret);
|
|
0
|
153
|
|
|
154 ret = clEnqueueWriteBuffer(command_queue, param_memobj,
|
|
|
155 CL_TRUE, 0, sizeof(int), (int*)(&st_mmap.size),
|
|
|
156 0, NULL, NULL);
|
|
|
157
|
|
|
158 oclCheckError(ret, CL_SUCCESS);
|
|
|
159
|
|
|
160
|
|
|
161 /*
|
|
|
162 * 並列度の計算
|
|
|
163 * wcするファイルの大きさに合わせる
|
|
|
164 *
|
|
|
165 */
|
|
|
166
|
|
|
167 // このdivi_size はどうやって決めるよ
|
|
|
168 int divi_size = 1024;
|
|
1
|
169 size_t global_work_size = (st_mmap.size + divi_size - 1) / divi_size;
|
|
|
170 int remain_size = st_mmap.size - global_work_size * divi_size;
|
|
0
|
171
|
|
1
|
172 /*
|
|
|
173 * 行数、単語数, 分割地点のフラグを格納する配列
|
|
|
174 * word_num, line_num, head, tail
|
|
|
175 */
|
|
0
|
176
|
|
|
177 int out_size = sizeof(int) * OUT_PARAM_NUM * global_work_size;
|
|
|
178 cl_mem out_memobj = clCreateBuffer(context, CL_MEM_WRITE_ONLY,
|
|
|
179 out_size * sizeof(char), NULL, &ret);
|
|
|
180
|
|
|
181 oclCheckError(ret, CL_SUCCESS);
|
|
|
182
|
|
|
183
|
|
|
184 // 引数のSet
|
|
|
185 // memory object にしなくてもできるsetできるかも
|
|
|
186 ret = clSetKernelArg(kernel, 0, sizeof(cl_mem), (void *)&text_memobj);
|
|
|
187 ret |= clSetKernelArg(kernel, 1, sizeof(cl_mem), (void *)¶m_memobj);
|
|
|
188 ret |= clSetKernelArg(kernel, 2, sizeof(cl_mem), (void *)&out_memobj);
|
|
|
189
|
|
|
190 oclCheckError(ret, CL_SUCCESS);
|
|
|
191
|
|
|
192 /*
|
|
|
193 * kernel実行
|
|
|
194 * 並列に処理せずに work-item ひとつで動かしたい場合は、clEnqueueNDRangeKernel の簡易版 clEnqueueTask が使える
|
|
|
195 *
|
|
1
|
196 */
|
|
0
|
197
|
|
|
198 //ret = clEnqueueTask(command_queue, kernel, 0, NULL, NULL);
|
|
|
199 // global_work_size は配列。並列動作させる時の次元数にあわせて、配列の次元数も決まるはず
|
|
|
200 ret = clEnqueueNDRangeKernel(command_queue, kernel, 1, NULL, &global_work_size, NULL, 0, NULL, NULL);
|
|
|
201
|
|
|
202 oclCheckError(ret, CL_SUCCESS);
|
|
|
203
|
|
|
204 int *out_data = (int*)malloc(out_size);
|
|
|
205
|
|
|
206 // 演算結果の読み込み
|
|
|
207 ret = clEnqueueReadBuffer(command_queue, out_memobj, CL_TRUE, 0,
|
|
1
|
208 out_size, out_data, 0, NULL, NULL);
|
|
0
|
209
|
|
|
210 oclCheckError(ret, CL_SUCCESS);
|
|
|
211
|
|
1
|
212 printf("global_work_size %d\n", (int)global_work_size);
|
|
|
213
|
|
0
|
214 for (int i = 0; i < global_work_size; i++) {
|
|
1
|
215 for (int j = 0; j < PRINT_PARAM_NUM; j++) {
|
|
0
|
216 printf("%d ", out_data[i*OUT_PARAM_NUM+j]);
|
|
|
217 }
|
|
|
218 printf("\n");
|
|
|
219 }
|
|
|
220
|
|
|
221 clFlush(command_queue);
|
|
|
222 clFinish(command_queue);
|
|
|
223 clReleaseKernel(kernel);
|
|
|
224 clReleaseProgram(program);
|
|
|
225 clReleaseMemObject(text_memobj);
|
|
|
226 clReleaseMemObject(param_memobj);
|
|
|
227 clReleaseMemObject(out_memobj);
|
|
|
228 clReleaseCommandQueue(command_queue);
|
|
|
229 clReleaseContext(context);
|
|
|
230
|
|
|
231 free(source_str);
|
|
|
232 free(out_data);
|
|
|
233
|
|
|
234 return 0;
|
|
|
235
|
|
|
236 }
|
