227
|
1 #include "TaskManager.h"
|
|
2 #include "sort.h"
|
|
3 #include "Func.h"
|
|
4
|
|
5 DataPtr data; // sort array
|
|
6 int data_length;
|
|
7 static int sort_count; // sort 完了に必要な回数
|
|
8 static int split_num; // data の分割数
|
|
9 static int half_num;
|
|
10 static int block_num; // 一つのタスクで sort する data 数
|
|
11 static int last_block_num;
|
|
12 static int half_block_num;
|
|
13 static int last_half_block_num;
|
|
14
|
|
15 static void sort_restart(void *);
|
|
16 static void sort_start(void);
|
|
17
|
|
18 /**
|
|
19 * 一つの block にある data の数が MAX_BLOCK_SIZE 超えないような
|
|
20 * len の分割数を返す
|
|
21 *
|
|
22 * @param len sort する data の総数
|
|
23 * @param num 使用する SPE の数
|
|
24 *
|
|
25 * @return data の分割数
|
|
26 *
|
|
27 * TODO:
|
|
28 * len が num 以下とか考えてません
|
|
29 */
|
|
30 static int
|
|
31 get_split_num(int len, int num)
|
|
32 {
|
|
33 if (len / num < MAX_BLOCK_SIZE) {
|
|
34 return num;
|
|
35 } else {
|
|
36 // 切り上げ
|
|
37 return (len + MAX_BLOCK_SIZE - 1) / MAX_BLOCK_SIZE;
|
|
38 }
|
|
39 }
|
|
40
|
|
41 /**
|
|
42 * btask が全て終了したら、再び sort_start を実行する
|
|
43 * @param d 生成された btask の数
|
|
44 */
|
|
45 static void
|
|
46 sort_restart(void *d)
|
|
47 {
|
|
48 static int cnt = 0;
|
|
49 int max = (int)d;
|
|
50
|
|
51 if (++cnt == max) {
|
|
52 cnt = 0;
|
|
53 sort_start();
|
|
54 }
|
|
55 }
|
|
56
|
|
57 static void
|
|
58 sort_start(void)
|
|
59 {
|
|
60 if (--sort_count < 0) {
|
|
61 return;
|
|
62 }
|
|
63
|
|
64 HTaskPtr fsort[split_num];
|
|
65
|
|
66 for (int i = 0; i < split_num-1; i++) {
|
|
67 fsort[i] = manager->create_task(QUICK_SORT);
|
|
68 fsort[i]->add_inData(&data[i*block_num], sizeof(Data)*block_num);
|
|
69 fsort[i]->add_outData(&data[i*block_num], sizeof(Data)*block_num);
|
|
70 fsort[i]->add_param(block_num);
|
|
71 fsort[i]->set_cpu(SPE_ANY);
|
|
72 }
|
|
73
|
|
74 // 最後の block は端数なので last_block_num を使う
|
|
75 {
|
|
76 int i = split_num-1;
|
|
77
|
|
78 fsort[i] = manager->create_task(QUICK_SORT);
|
|
79 fsort[i]->add_inData(&data[i*block_num], sizeof(Data)*last_block_num);
|
|
80 fsort[i]->add_outData(&data[i*block_num], sizeof(Data)*last_block_num);
|
|
81 fsort[i]->add_param(last_block_num);
|
|
82 fsort[i]->set_cpu(SPE_ANY);
|
|
83 }
|
|
84
|
|
85 if (split_num > 1) {
|
|
86 HTaskPtr bsort[half_num];
|
|
87
|
|
88 for (int i = 0; i < half_num-1; i++) {
|
|
89 bsort[i] = manager->create_task(QUICK_SORT);
|
|
90 bsort[i]->add_inData(&data[i*block_num+half_block_num],
|
|
91 sizeof(Data)*block_num);
|
|
92 bsort[i]->add_outData(&data[i*block_num+half_block_num],
|
|
93 sizeof(Data)*block_num);
|
|
94 bsort[i]->add_param(block_num);
|
|
95 bsort[i]->set_cpu(SPE_ANY);
|
|
96 }
|
|
97
|
|
98 {
|
|
99 int i = half_num-1;
|
|
100
|
|
101 bsort[i] = manager->create_task(QUICK_SORT);
|
|
102 bsort[i]->add_inData(&data[i*block_num+half_block_num],
|
|
103 sizeof(Data)*last_half_block_num);
|
|
104 bsort[i]->add_outData(&data[i*block_num+half_block_num],
|
|
105 sizeof(Data)*last_half_block_num);
|
|
106 bsort[i]->add_param(last_half_block_num);
|
|
107 bsort[i]->set_cpu(SPE_ANY);
|
|
108 }
|
|
109
|
|
110 for (int i = 0; i < half_num; i++) {
|
|
111 bsort[i]->wait_for(fsort[i]);
|
|
112 bsort[i]->wait_for(fsort[i+1]);
|
|
113 bsort[i]->set_post(sort_restart, (void*)(half_num));
|
|
114 bsort[i]->spawn();
|
|
115 }
|
|
116 }
|
|
117
|
|
118 for (int i = 0; i < split_num; i++) {
|
|
119 fsort[i]->spawn();
|
|
120 }
|
|
121 }
|
|
122
|
|
123 void
|
|
124 sort_init(int cpuNum, int length)
|
|
125 {
|
|
126 data = (DataPtr)manager->allocate(sizeof(Data)*length);
|
|
127 data_length = length;
|
|
128
|
|
129 for (int i = 0; i < length; i++) {
|
230
|
130 data[i].index = manager->get_random()%10000;
|
227
|
131 data[i].ptr = 0;
|
|
132 }
|
|
133
|
|
134 split_num = get_split_num(length, cpuNum);
|
|
135 half_num = split_num-1;
|
|
136 sort_count = split_num;
|
|
137
|
|
138 block_num = (length + split_num -1)/split_num;
|
|
139 half_block_num = block_num/2;
|
|
140
|
|
141 last_block_num = length - (split_num-1)*block_num;
|
|
142 last_half_block_num = half_block_num+(last_block_num/2);
|
|
143
|
|
144 sort_start();
|
|
145 }
|