////////////////////////////////////////////////////////////////////////////////
// Program main
////////////////////////////////////////////////////////////////////////////////
int main(int argc, char **argv)
{
bool doCPU = false;
bool doGPU = false;
bool doMultiGPU = false;
bool doCPUGPU = false;
bool doRef = false;
int numSources = 100;
int generateVerts = 100000;
int generateEdgesPerVert = 10;
parseCommandLineArgs(argc, argv, doCPU, doGPU,
doMultiGPU, doCPUGPU, doRef,
&numSources, &generateVerts, &generateEdgesPerVert);
cl_platform_id platform;
cl_context gpuContext;
cl_context cpuContext;
cl_int errNum;
// First, select an OpenCL platform to run on. For this example, we
// simply choose the first available platform. Normally, you would
// query for all available platforms and select the most appropriate one.
cl_uint numPlatforms;
errNum = clGetPlatformIDs(1, &platform, &numPlatforms);
printf("Number of OpenCL Platforms: %d\n", numPlatforms);
if (errNum != CL_SUCCESS || numPlatforms <= 0)
{
printf("Failed to find any OpenCL platforms.\n");
return 1;
}
// create the OpenCL context on available GPU devices
gpuContext = clCreateContextFromType(0, CL_DEVICE_TYPE_GPU, NULL, NULL, &errNum);
if (errNum != CL_SUCCESS)
{
printf("No GPU devices found.\n");
}
// Create an OpenCL context on available CPU devices
cpuContext = clCreateContextFromType(0, CL_DEVICE_TYPE_CPU, NULL, NULL, &errNum);
if (errNum != CL_SUCCESS)
{
printf("No CPU devices found.\n");
}
// Allocate memory for arrays
GraphData graph;
generateRandomGraph(&graph, generateVerts, generateEdgesPerVert);
printf("Vertex Count: %d\n", graph.vertexCount);
printf("Edge Count: %d\n", graph.edgeCount);
std::vector<int> sourceVertices;
for(int source = 0; source < numSources; source++)
{
sourceVertices.push_back(source % graph.vertexCount);
}
int *sourceVertArray = (int*) malloc(sizeof(int) * sourceVertices.size());
std::copy(sourceVertices.begin(), sourceVertices.end(), sourceVertArray);
float *results = (float*) malloc(sizeof(float) * sourceVertices.size() * graph.vertexCount);
// Run Dijkstra's algorithm
pt::ptime startTimeCPU = pt::microsec_clock::local_time();
if (doCPU)
{
runDijkstra(cpuContext, getMaxFlopsDev(cpuContext), &graph, sourceVertArray,
results, sourceVertices.size() );
}
pt::time_duration timeCPU = pt::microsec_clock::local_time() - startTimeCPU;
pt::ptime startTimeGPU = pt::microsec_clock::local_time();
if (doGPU)
{
runDijkstra(gpuContext, getMaxFlopsDev(gpuContext), &graph, sourceVertArray,
results, sourceVertices.size() );
}
pt::time_duration timeGPU = pt::microsec_clock::local_time() - startTimeGPU;
pt::ptime startTimeMultiGPU = pt::microsec_clock::local_time();
if (doMultiGPU)
{
runDijkstraMultiGPU(gpuContext, &graph, sourceVertArray,
results, sourceVertices.size() );
}
pt::time_duration timeMultiGPU = pt::microsec_clock::local_time() - startTimeMultiGPU;
pt::ptime startTimeGPUCPU = pt::microsec_clock::local_time();
if (doCPUGPU)
{
runDijkstraMultiGPUandCPU(gpuContext, cpuContext, &graph, sourceVertArray,
results, sourceVertices.size() );
}
pt::time_duration timeGPUCPU = pt::microsec_clock::local_time() - startTimeGPUCPU;
pt::ptime startTimeRef = pt::microsec_clock::local_time();
if (doRef)
{
runDijkstraRef( &graph, sourceVertArray,
results, sourceVertices.size() );
}
pt::time_duration timeRef = pt::microsec_clock::local_time() - startTimeRef;
if (doCPU)
{
printf("\nrunDijkstra - CPU Time: %f s\n", (float)timeCPU.total_milliseconds() / 1000.0f);
}
if (doGPU)
{
printf("\nrunDijkstra - Single GPU Time: %f s\n", (float)timeGPU.total_milliseconds() / 1000.0f);
}
if (doMultiGPU)
{
printf("\nrunDijkstra - Multi GPU Time: %f s\n", (float)timeMultiGPU.total_milliseconds() / 1000.0f);
}
if (doCPUGPU)
{
printf("\nrunDijkstra - Multi GPU and CPU Time: %f s\n", (float)timeGPUCPU.total_milliseconds() / 1000.0f);
}
if (doRef)
{
printf("\nrunDijkstra - Reference (CPU): %f s\n", (float)timeRef.total_milliseconds() / 1000.0f);
}
free(sourceVertArray);
free(results);
clReleaseContext(gpuContext);
// finish
//shrEXIT(argc, argv);
}
////////////////////////////////////////////////////////////////////////////////
// Program main
////////////////////////////////////////////////////////////////////////////////
int
main( int argc, char** argv)
{
//@TEMP - why do I need this for link to work on Linux?
//cutWaitForThreads(NULL,0);
//@TEMP
// use command-line specified CUDA device, otherwise use device with highest Gflops/s
//if( cutCheckCmdLineFlag(argc, (const char**)argv, "device") )
// cutilDeviceInit(argc, argv);
//else
// cudaSetDevice( cutGetMaxGflopsDeviceId() );
bool doGPU;
bool doRef;
bool doMultiGPU;
int generateVerts;
int generateEdgesPerVert;
int numSources;
int sourceId;
//parseCommandLineArgs(argc, (const char**)argv, doGPU, doMultiGPU, doRef, &numSources, &generateVerts, &generateEdgesPerVert);
doGPU = 1;
doMultiGPU = 0;
doRef = 0;
// Allocate memory for arrays
GraphData graph;
//generateRandomGraph(&graph, generateVerts, generateEdgesPerVert);
generateLVAGraph(&graph, generateVerts, generateEdgesPerVert);
//printf("Vertex Count: %d\n", graph.vertexCount);
//printf("Edge Count: %d\n", graph.edgeCount);
std::vector<int> sourceVertices;
//for(int source = 0; source < numSources; source++)
//{
// sourceVertices.push_back(source % graph.vertexCount);
//}
FILE *fp;
fp = fopen("nodes2cal.out", "r");
if(fp){
//printf("Reading file nodes2cal.out...\n");
if(fscanf(fp, "%d", &numSources)!=EOF){
if(numSources<1){
printf("nodes2cal.out must have numSources>0.\n");
return 1;
}
}
else{
printf("nodes2cal.out must have as first line 'numSources'.\n");
return 1;
}
while(fscanf(fp, "%d", &sourceId)!=EOF){
sourceVertices.push_back(sourceId-1);
}
}
else{
printf("nodes2cal.out not exists!.\n");
return 1;
}
int *sourceVertArray = (int*) malloc(sizeof(int) * sourceVertices.size());
std::copy(sourceVertices.begin(), sourceVertices.end(), sourceVertArray);
float *results = (float*) malloc(sizeof(float) * sourceVertices.size() * graph.vertexCount);
unsigned int gpuTimer = 0;
//cutilCheckError(cutCreateTimer(&gpuTimer));
//cutilCheckError(cutStartTimer(gpuTimer));
// Run Dijkstra's algorithm
if ( doGPU )
{
runDijkstra(&graph, sourceVertArray, results, sourceVertices.size() );
}
//cutilCheckError(cutStopTimer(gpuTimer));
unsigned int multiGPUTimer = 0;
//cutilCheckError(cutCreateTimer(&multiGPUTimer));
//cutilCheckError(cutStartTimer(multiGPUTimer));
if ( doMultiGPU )
{
runDijkstraMultiGPU(&graph, sourceVertArray, results, sourceVertices.size() );
}
//cutilCheckError(cutStopTimer(multiGPUTimer));
unsigned int refTimer = 0;
//cutilCheckError(cutCreateTimer(&refTimer));
//cutilCheckError(cutStartTimer(refTimer));
if ( doRef )
{
runDijkstraRef(&graph, sourceVertArray, results, sourceVertices.size() );
}
//cutilCheckError(cutStopTimer(refTimer));
for (unsigned int i = 0; i < sourceVertices.size(); i++)
{
for (int j = 0; j < graph.vertexCount; j++)
{
//if (i != j)
//{
//printf("%d --> %d: %f\n", sourceVertArray[i], j, results[i * graph.vertexCount + j] );
printf("%f\n", results[i * graph.vertexCount + j] );
//}
}
}
free(sourceVertArray);
free(results);
//cudaThreadExit();
return 0;
}