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
|
#include <assert.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <unistd.h>
#include <stdint.h>
#include "../../support/common.h"
#include "../../support/graph.h"
#include "../../support/params.h"
#include "../../support/timer.h"
#include "../../support/utils.h"
__global__ void bfs_kernel(CSRGraph csrGraph, uint32_t* nodeLevel, uint32_t* prevFrontier, uint32_t* currFrontier, uint32_t numPrevFrontier, uint32_t* numCurrFrontier, uint32_t level) {
uint32_t i = blockIdx.x*blockDim.x + threadIdx.x;
if(i < numPrevFrontier) {
uint32_t node = prevFrontier[i];
for(uint32_t edge = csrGraph.nodePtrs[node]; edge < csrGraph.nodePtrs[node + 1]; ++edge) {
uint32_t neighbor = csrGraph.neighborIdxs[edge];
if(atomicCAS(&nodeLevel[neighbor], UINT32_MAX, level) == UINT32_MAX) { // Node not previously visited
uint32_t currFrontierIdx = atomicAdd(numCurrFrontier, 1);
currFrontier[currFrontierIdx] = neighbor;
}
}
}
}
int main(int argc, char** argv) {
// Process parameters
struct Params p = input_params(argc, argv);
// Initialize BFS data structures
PRINT_INFO(p.verbosity >= 1, "Reading graph %s", p.fileName);
struct COOGraph cooGraph = readCOOGraph(p.fileName);
PRINT_INFO(p.verbosity >= 1, " Graph has %d nodes and %d edges", cooGraph.numNodes, cooGraph.numEdges);
struct CSRGraph csrGraph = coo2csr(cooGraph);
uint32_t* nodeLevel_cpu = (uint32_t*) malloc(csrGraph.numNodes*sizeof(uint32_t));
uint32_t* nodeLevel_gpu = (uint32_t*) malloc(csrGraph.numNodes*sizeof(uint32_t));
for(uint32_t i = 0; i < csrGraph.numNodes; ++i) {
nodeLevel_cpu[i] = UINT32_MAX; // Unreachable
nodeLevel_gpu[i] = UINT32_MAX; // Unreachable
}
uint32_t srcNode = 0;
// Allocate GPU memory
CSRGraph csrGraph_d;
csrGraph_d.numNodes = csrGraph.numNodes;
csrGraph_d.numEdges = csrGraph.numEdges;
cudaMalloc((void**) &csrGraph_d.nodePtrs, (csrGraph_d.numNodes + 1)*sizeof(uint32_t));
cudaMalloc((void**) &csrGraph_d.neighborIdxs, csrGraph_d.numEdges*sizeof(uint32_t));
uint32_t* nodeLevel_d;
cudaMalloc((void**) &nodeLevel_d, csrGraph_d.numNodes*sizeof(uint32_t));
uint32_t* buffer1_d;
cudaMalloc((void**) &buffer1_d, csrGraph_d.numNodes*sizeof(uint32_t));
uint32_t* buffer2_d;
cudaMalloc((void**) &buffer2_d, csrGraph_d.numNodes*sizeof(uint32_t));
uint32_t* numCurrFrontier_d;
cudaMalloc((void**) &numCurrFrontier_d, sizeof(uint32_t));
uint32_t* prevFrontier_d = buffer1_d;
uint32_t* currFrontier_d = buffer2_d;
// Copy data to GPU
cudaMemcpy(csrGraph_d.nodePtrs, csrGraph.nodePtrs, (csrGraph_d.numNodes + 1)*sizeof(uint32_t), cudaMemcpyHostToDevice);
cudaMemcpy(csrGraph_d.neighborIdxs, csrGraph.neighborIdxs, csrGraph_d.numEdges*sizeof(uint32_t), cudaMemcpyHostToDevice);
nodeLevel_gpu[srcNode] = 0;
cudaMemcpy(nodeLevel_d, nodeLevel_gpu, csrGraph_d.numNodes*sizeof(uint32_t), cudaMemcpyHostToDevice);
cudaMemcpy(prevFrontier_d, &srcNode, sizeof(uint32_t), cudaMemcpyHostToDevice);
cudaDeviceSynchronize();
// Calculating result on GPU
PRINT_INFO(p.verbosity >= 1, "Calculating result on GPU");
Timer timer;
startTimer(&timer);
uint32_t numPrevFrontier = 1;
uint32_t numThreadsPerBlock = 256;
for(uint32_t level = 1; numPrevFrontier > 0; ++level) {
// Visit nodes in previous frontier
cudaMemset(numCurrFrontier_d, 0, sizeof(uint32_t));
uint32_t numBlocks = (numPrevFrontier + numThreadsPerBlock - 1)/numThreadsPerBlock;
bfs_kernel <<< numBlocks, numThreadsPerBlock >>> (csrGraph_d, nodeLevel_d, prevFrontier_d, currFrontier_d, numPrevFrontier, numCurrFrontier_d, level);
// Swap buffers
uint32_t* tmp = prevFrontier_d;
prevFrontier_d = currFrontier_d;
currFrontier_d = tmp;
cudaMemcpy(&numPrevFrontier, numCurrFrontier_d, sizeof(uint32_t), cudaMemcpyDeviceToHost);
}
cudaDeviceSynchronize();
stopTimer(&timer);
if(p.verbosity == 0) PRINT("%f", getElapsedTime(timer)*1e3);
PRINT_INFO(p.verbosity >= 1, "Elapsed time: %f ms", getElapsedTime(timer)*1e3);
// Copy data from GPU
cudaMemcpy(nodeLevel_gpu, nodeLevel_d, csrGraph_d.numNodes*sizeof(uint32_t), cudaMemcpyDeviceToHost);
cudaDeviceSynchronize();
// Initialize frontier double buffers for CPU
uint32_t* buffer1 = (uint32_t*) malloc(csrGraph.numNodes*sizeof(uint32_t));
uint32_t* buffer2 = (uint32_t*) malloc(csrGraph.numNodes*sizeof(uint32_t));
uint32_t* prevFrontier = buffer1;
uint32_t* currFrontier = buffer2;
// Calculating result on CPU
PRINT_INFO(p.verbosity >= 1, "Calculating result on CPU");
nodeLevel_cpu[srcNode] = 0;
prevFrontier[0] = srcNode;
numPrevFrontier = 1;
for(uint32_t level = 1; numPrevFrontier > 0; ++level) {
uint32_t numCurrFrontier = 0;
// Visit nodes in the previous frontier
for(uint32_t i = 0; i < numPrevFrontier; ++i) {
uint32_t node = prevFrontier[i];
for(uint32_t edge = csrGraph.nodePtrs[node]; edge < csrGraph.nodePtrs[node + 1]; ++edge) {
uint32_t neighbor = csrGraph.neighborIdxs[edge];
if(nodeLevel_cpu[neighbor] == UINT32_MAX) { // Node not previously visited
nodeLevel_cpu[neighbor] = level;
currFrontier[numCurrFrontier] = neighbor;
++numCurrFrontier;
}
}
}
// Swap buffers
uint32_t* tmp = prevFrontier;
prevFrontier = currFrontier;
currFrontier = tmp;
numPrevFrontier = numCurrFrontier;
}
// Verify result
PRINT_INFO(p.verbosity >= 1, "Verifying the result");
for(uint32_t i = 0; i < csrGraph.numNodes; ++i) {
if(nodeLevel_cpu[i] != nodeLevel_gpu[i]) {
printf("Mismatch detected at node %u (CPU result = %u, GPU result = %u)\n", i, nodeLevel_cpu[i], nodeLevel_gpu[i]);
exit(0);
}
}
// Deallocate data structures
freeCOOGraph(cooGraph);
freeCSRGraph(csrGraph);
free(nodeLevel_cpu);
free(nodeLevel_gpu);
free(buffer1);
free(buffer2);
return 0;
}
|
