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
|
#pragma once
#include "./graph_base.hpp"
#include <taskflow/cuda/cudaflow.hpp>
#include <cassert>
template <typename OPT>
class GraphExecutor {
public:
GraphExecutor(Graph& graph, int dev_id = 0);
template <typename... OPT_Args>
void traversal(OPT_Args&&... args);
private:
int _dev_id;
Graph& _g;
};
template <typename OPT>
GraphExecutor<OPT>::GraphExecutor(Graph& graph, int dev_id): _g{graph}, _dev_id{dev_id} {
//TODO: why we cannot put cuda lambda function here?
}
template <typename OPT>
template <typename... OPT_Args>
void GraphExecutor<OPT>::traversal(OPT_Args&&... args) {
tf::Taskflow taskflow;
tf::Executor executor;
taskflow.emplace([this, args...]() {
tf::cudaFlowCapturer cf;
cf.make_optimizer<OPT>(args...);
std::vector<std::vector<tf::cudaTask>> tasks;
tasks.resize(_g.get_graph().size());
for(size_t l = 0; l < _g.get_graph().size(); ++l) {
tasks[l].resize((_g.get_graph())[l].size());
for(size_t i = 0; i < (_g.get_graph())[l].size(); ++i) {
bool* v = _g.at(l, i).visited;
tasks[l][i] = cf.single_task([v] __device__ () {
*v = true;
});
}
}
for(size_t l = 0; l < _g.get_graph().size() - 1; ++l) {
for(size_t i = 0; i < (_g.get_graph())[l].size(); ++i) {
for(auto&& out_node: _g.at(l, i).out_nodes) {
tasks[l][i].precede(tasks[l + 1][out_node]);
}
}
}
tf::cudaStream stream;
cf.run(stream);
stream.synchronize();
}).name("traverse");
//auto check_t = taskflow.emplace([this](){
//assert(_g.traversed());
//});
//trav_t.precede(check_t);
executor.run(taskflow).wait();
}
|
