Mercurial > hg > Game > Cerium
annotate example/many_task/gpu/sort_test.cc @ 2048:6796d85f3d6b draft
remove error
| author | Masataka Kohagura <kohagura@cr.ie.u-ryukyu.ac.jp> |
|---|---|
| date | Thu, 28 Jan 2016 00:05:49 +0900 |
| parents | fac06524090b |
| children |
| rev | line source |
|---|---|
| 1520 | 1 #include <stdlib.h> |
| 2 #include <OpenCL/opencl.h> | |
| 3 #include <stdio.h> | |
| 4 #include <fcntl.h> | |
| 5 #include <string.h> | |
| 6 #include <sys/time.h> | |
| 7 #include <sys/stat.h> | |
| 8 #include "sort.h" | |
| 9 #define DEFAULT 432 | |
| 10 | |
| 11 extern int data_length; | |
| 12 extern DataPtr data; | |
| 13 | |
| 14 // 計測用 | |
| 15 static double st_time; | |
| 16 static double ed_time; | |
| 17 static int length = DEFAULT; | |
| 18 | |
|
1538
fac06524090b
add gpu task wordcount. But not work print
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
1520
diff
changeset
|
19 |
| 1520 | 20 int |
| 21 init(int argc, char **argv) | |
| 22 { | |
| 23 for (int i = 1; argv[i]; ++i) { | |
| 24 if (strcmp(argv[i], "--length") == 0 || strcmp(argv[i], "-l") == 0) { | |
| 25 length = atoi(argv[++i]); | |
| 26 } | |
| 27 } | |
| 28 | |
| 29 return 0; | |
| 30 } | |
| 31 | |
| 32 int | |
| 33 get_split_num(int len, int num) | |
| 34 { | |
| 35 if (len / num < MAX_BLOCK_SIZE) { | |
| 36 return num; | |
| 37 } else { | |
| 38 // 切り上げ | |
| 39 return (len + MAX_BLOCK_SIZE - 1) / MAX_BLOCK_SIZE; | |
| 40 } | |
| 41 } | |
| 42 | |
| 43 | |
| 44 static double | |
| 45 getTime() | |
| 46 { | |
| 47 struct timeval tv; | |
| 48 gettimeofday(&tv, NULL); | |
| 49 return tv.tv_sec + (double)tv.tv_usec*1e-6; | |
| 50 } | |
| 51 | |
| 52 void | |
| 53 show( Data *data, int size ) | |
| 54 { | |
| 55 puts("-----------------------------------------------"); | |
| 56 for(int i=0; i<=size; i++) printf("data[%02d].index = %d\n", i, data[i].index); | |
| 57 puts("-----------------------------------------------"); | |
| 58 return; | |
| 59 } | |
| 60 | |
| 61 Sort sorter; | |
| 62 | |
| 63 static void | |
| 64 check_data() | |
| 65 { | |
| 66 for(int i=0; i< sorter.data_length-1;i++) { | |
| 67 if (sorter.data[i].index>sorter.data[i+1].index) { | |
| 68 printf("Data are not sorted at %d. %d > %d \n",i, sorter.data[i].index,sorter.data[i+1].index); | |
| 69 return; | |
| 70 } | |
| 71 } | |
| 72 printf("Data are sorted\n"); | |
| 73 } | |
| 74 | |
| 75 void | |
| 76 gpu_init() | |
| 77 { | |
|
1538
fac06524090b
add gpu task wordcount. But not work print
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
1520
diff
changeset
|
78 |
|
fac06524090b
add gpu task wordcount. But not work print
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
1520
diff
changeset
|
79 } |
|
fac06524090b
add gpu task wordcount. But not work print
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
1520
diff
changeset
|
80 |
|
fac06524090b
add gpu task wordcount. But not work print
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
1520
diff
changeset
|
81 void |
|
fac06524090b
add gpu task wordcount. But not work print
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
1520
diff
changeset
|
82 sort_start(Sort s){ |
|
fac06524090b
add gpu task wordcount. But not work print
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
1520
diff
changeset
|
83 |
|
fac06524090b
add gpu task wordcount. But not work print
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
1520
diff
changeset
|
84 Sort sorter = s; |
|
fac06524090b
add gpu task wordcount. But not work print
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
1520
diff
changeset
|
85 int length = sorter.data_length; |
|
fac06524090b
add gpu task wordcount. But not work print
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
1520
diff
changeset
|
86 cl_platform_id platform_id = NULL; |
|
fac06524090b
add gpu task wordcount. But not work print
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
1520
diff
changeset
|
87 cl_uint ret_num_platforms = NULL; |
|
fac06524090b
add gpu task wordcount. But not work print
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
1520
diff
changeset
|
88 cl_device_id device_id = NULL; |
|
fac06524090b
add gpu task wordcount. But not work print
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
1520
diff
changeset
|
89 cl_uint ret_num_devices = NULL; |
|
fac06524090b
add gpu task wordcount. But not work print
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
1520
diff
changeset
|
90 cl_int ret = NULL; |
|
fac06524090b
add gpu task wordcount. But not work print
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
1520
diff
changeset
|
91 |
| 1520 | 92 clGetPlatformIDs(1, &platform_id, &ret_num_platforms); |
| 93 clGetDeviceIDs(platform_id, CL_DEVICE_TYPE_DEFAULT, 1, &device_id, | |
| 94 &ret_num_devices); | |
| 95 | |
|
1538
fac06524090b
add gpu task wordcount. But not work print
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
1520
diff
changeset
|
96 cl_context context = clCreateContext( NULL, 1, &device_id, NULL, NULL, &ret); |
|
fac06524090b
add gpu task wordcount. But not work print
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
1520
diff
changeset
|
97 cl_command_queue command_queue = clCreateCommandQueue(context, device_id, 0, &ret); |
| 1520 | 98 |
| 99 //ファイルオープン | |
| 100 | |
| 101 const char* filename = "QuickSort.cl"; | |
| 102 const char* functionname = "quick_sort"; | |
| 103 | |
| 104 int fp = open(filename, O_RDONLY); | |
| 105 | |
| 106 if (!fp) { | |
| 107 fprintf(stderr, "Failed to load kernel.\n"); | |
| 108 exit(1); | |
| 109 } | |
| 110 | |
| 111 struct stat stats; | |
| 112 fstat(fp,&stats); | |
| 113 off_t size = stats.st_size; | |
| 114 | |
| 115 if (!size) { | |
| 116 fprintf(stderr, "Failed to load kernel.\n"); | |
| 117 exit(1); | |
| 118 } | |
| 119 | |
| 120 char *kernel_src_str = new char[size]; | |
| 121 size_t kernel_code_size = read(fp, kernel_src_str, size); | |
| 122 close(fp); | |
| 123 | |
|
1538
fac06524090b
add gpu task wordcount. But not work print
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
1520
diff
changeset
|
124 cl_program program = clCreateProgramWithSource(context, 1, (const char **)&kernel_src_str, |
| 1520 | 125 (const size_t *)&kernel_code_size, &ret); |
| 126 clBuildProgram(program, 1, &device_id, NULL, NULL, NULL); | |
|
1538
fac06524090b
add gpu task wordcount. But not work print
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
1520
diff
changeset
|
127 cl_kernel kernel = clCreateKernel(program,functionname, &ret); |
| 1520 | 128 |
| 129 //メモリバッファの作成 | |
| 130 cl_mem mem_count = clCreateBuffer(context, CL_MEM_READ_ONLY,sizeof(int),NULL, &ret); | |
| 131 cl_mem mem_data = clCreateBuffer(context, CL_MEM_READ_WRITE, sizeof(Data)*length, NULL, &ret); | |
| 132 | |
| 133 st_time = getTime(); | |
| 134 | |
| 135 //メモリバッファに入力データを書き込み | |
| 136 ret = clEnqueueWriteBuffer(command_queue, mem_count, CL_TRUE, 0, | |
| 137 sizeof(int), &length, 0, NULL, NULL); | |
| 138 ret = clEnqueueWriteBuffer(command_queue, mem_data, CL_TRUE, 0, | |
| 139 sizeof(Data)*length, sorter.data, 0, NULL, NULL); | |
| 140 | |
| 141 //print_data(data, count, "before"); | |
| 142 clSetKernelArg(kernel, 0, sizeof(cl_mem), (void *)&mem_count); | |
| 143 clSetKernelArg(kernel, 1, sizeof(cl_mem), (void *)&mem_data); | |
| 144 | |
|
1538
fac06524090b
add gpu task wordcount. But not work print
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
1520
diff
changeset
|
145 cl_event ev = NULL; |
|
fac06524090b
add gpu task wordcount. But not work print
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
1520
diff
changeset
|
146 |
| 1520 | 147 ret = clEnqueueTask(command_queue, kernel, 0, NULL, &ev); |
| 148 | |
| 149 //メモリバッファから結果を取得 | |
| 150 ret = clEnqueueReadBuffer(command_queue, mem_data, CL_TRUE, 0,sizeof(Data)*length, sorter.data, 1, &ev, NULL); | |
| 151 clFlush(command_queue); | |
| 152 ed_time = getTime(); | |
| 153 show(sorter.data, length-1); | |
| 154 check_data(); | |
| 155 printf("Time: %0.6f\n",ed_time-st_time); | |
| 156 | |
| 157 clReleaseKernel(kernel); | |
| 158 clReleaseProgram(program); | |
| 159 clReleaseMemObject(mem_data); | |
| 160 clReleaseEvent(ev); | |
| 161 clReleaseCommandQueue(command_queue); | |
| 162 clReleaseContext(context); | |
| 163 | |
| 164 } | |
| 165 | |
| 166 int main(int argc, char *argv[]) { | |
| 167 | |
| 168 // 無効な引数ならデフォルトの値として432を設定 | |
| 169 | |
| 170 | |
| 171 if (argc>1) { | |
| 172 if (init(argc,argv) < 0) { | |
| 173 return -1; | |
| 174 } | |
| 175 } | |
| 176 | |
| 177 gpu_init(); | |
| 178 | |
| 179 sorter.data = new Data[length]; | |
| 180 sorter.data_length = length; | |
| 181 sorter.split_num = get_split_num(sorter.data_length, 1); // (length, cpu_num) | |
| 182 | |
| 183 for (int i = 0; i < length; i++) { | |
| 184 sorter.data[i].index = rand()%10000; | |
| 185 sorter.data[i].ptr = i; | |
| 186 } | |
| 187 | |
| 188 sort_start(sorter); | |
| 189 return 0; | |
| 190 } |