//----------------------------------------------------------
//--breadth first search on GPUs
//----------------------------------------------------------
void run_bfs_gpu(int no_of_nodes, Node *h_graph_nodes, int edge_list_size, \
int *h_graph_edges, char *h_graph_mask, char *h_updating_graph_mask, \
char *h_graph_visited, int *h_cost)
throw(std::string){
//int number_elements = height*width;
char h_over;
cl_mem d_graph_nodes, d_graph_edges, d_graph_mask, d_updating_graph_mask, \
d_graph_visited, d_cost, d_over;
try{
//--1 transfer data from host to device
_clInit();
d_graph_nodes = _clMalloc(no_of_nodes*sizeof(Node), h_graph_nodes);
d_graph_edges = _clMalloc(edge_list_size*sizeof(int), h_graph_edges);
d_graph_mask = _clMallocRW(no_of_nodes*sizeof(char), h_graph_mask);
d_updating_graph_mask = _clMallocRW(no_of_nodes*sizeof(char), h_updating_graph_mask);
d_graph_visited = _clMallocRW(no_of_nodes*sizeof(char), h_graph_visited);
d_cost = _clMallocRW(no_of_nodes*sizeof(int), h_cost);
d_over = _clMallocRW(sizeof(char), &h_over);
_clMemcpyH2D(d_graph_nodes, no_of_nodes*sizeof(Node), h_graph_nodes);
_clMemcpyH2D(d_graph_edges, edge_list_size*sizeof(int), h_graph_edges);
_clMemcpyH2D(d_graph_mask, no_of_nodes*sizeof(char), h_graph_mask);
_clMemcpyH2D(d_updating_graph_mask, no_of_nodes*sizeof(char), h_updating_graph_mask);
_clMemcpyH2D(d_graph_visited, no_of_nodes*sizeof(char), h_graph_visited);
_clMemcpyH2D(d_cost, no_of_nodes*sizeof(int), h_cost);
//--2 invoke kernel
#ifdef PROFILING
timer kernel_timer;
double kernel_time = 0.0;
kernel_timer.reset();
kernel_timer.start();
#endif
do{
h_over = false;
_clMemcpyH2D(d_over, sizeof(char), &h_over);
//--kernel 0
int kernel_id = 0;
int kernel_idx = 0;
_clSetArgs(kernel_id, kernel_idx++, d_graph_nodes);
_clSetArgs(kernel_id, kernel_idx++, d_graph_edges);
_clSetArgs(kernel_id, kernel_idx++, d_graph_mask);
_clSetArgs(kernel_id, kernel_idx++, d_updating_graph_mask);
_clSetArgs(kernel_id, kernel_idx++, d_graph_visited);
_clSetArgs(kernel_id, kernel_idx++, d_cost);
_clSetArgs(kernel_id, kernel_idx++, &no_of_nodes, sizeof(int));
//int work_items = no_of_nodes;
_clInvokeKernel(kernel_id, no_of_nodes, work_group_size);
//--kernel 1
kernel_id = 1;
kernel_idx = 0;
_clSetArgs(kernel_id, kernel_idx++, d_graph_mask);
_clSetArgs(kernel_id, kernel_idx++, d_updating_graph_mask);
_clSetArgs(kernel_id, kernel_idx++, d_graph_visited);
_clSetArgs(kernel_id, kernel_idx++, d_over);
_clSetArgs(kernel_id, kernel_idx++, &no_of_nodes, sizeof(int));
//work_items = no_of_nodes;
_clInvokeKernel(kernel_id, no_of_nodes, work_group_size);
_clMemcpyD2H(d_over,sizeof(char), &h_over);
}while(h_over);
_clFinish();
#ifdef PROFILING
kernel_timer.stop();
kernel_time = kernel_timer.getTimeInSeconds();
#endif
//--3 transfer data from device to host
_clMemcpyD2H(d_cost,no_of_nodes*sizeof(int), h_cost);
//--statistics
#ifdef PROFILING
std::cout<<"kernel time(s):"<<kernel_time<<std::endl;
#endif
//--4 release cl resources.
_clFree(d_graph_nodes);
_clFree(d_graph_edges);
_clFree(d_graph_mask);
_clFree(d_updating_graph_mask);
_clFree(d_graph_visited);
_clFree(d_cost);
_clFree(d_over);
_clRelease();
}
catch(std::string msg){
_clFree(d_graph_nodes);
_clFree(d_graph_edges);
_clFree(d_graph_mask);
_clFree(d_updating_graph_mask);
_clFree(d_graph_visited);
_clFree(d_cost);
_clFree(d_over);
_clRelease();
std::string e_str = "in run_transpose_gpu -> ";
e_str += msg;
throw(e_str);
}
return ;
}
//----------------------------------------------------------
//--breadth first search on GPUs
//----------------------------------------------------------
void run_bfs_gpu(int no_of_nodes, Node *h_graph_nodes, int edge_list_size, \
int *h_graph_edges, bool *h_graph_mask, bool *h_updating_graph_mask, \
bool *h_graph_visited, int *h_cost)
throw(std::string){
//int number_elements = height*width;
bool h_over;
cl_mem d_graph_nodes, d_graph_edges, d_graph_mask, d_updating_graph_mask, \
d_graph_visited, d_cost, d_over;
// try{
//--1 transfer data from host to device
//printf("initializing\n");
_clInit();
//printf("allocating\n");
d_graph_nodes = _clMalloc(no_of_nodes*sizeof(Node), h_graph_nodes);
d_graph_edges = _clMalloc(edge_list_size*sizeof(int), h_graph_edges);
d_graph_mask = _clMalloc(no_of_nodes*sizeof(bool), h_graph_mask);
d_updating_graph_mask = _clMalloc(no_of_nodes*sizeof(bool), h_updating_graph_mask);
d_graph_visited = _clMalloc(no_of_nodes*sizeof(bool), h_graph_visited);
d_cost = _clMallocRW(no_of_nodes*sizeof(int), h_cost);
d_over = _clMallocRW(sizeof(bool), &h_over);
//printf("copyin\n");
_clMemcpyH2D(d_graph_nodes, no_of_nodes*sizeof(Node), h_graph_nodes);
_clMemcpyH2D(d_graph_edges, edge_list_size*sizeof(int), h_graph_edges);
_clMemcpyH2D(d_graph_mask, no_of_nodes*sizeof(bool), h_graph_mask);
_clMemcpyH2D(d_updating_graph_mask, no_of_nodes*sizeof(bool), h_updating_graph_mask);
_clMemcpyH2D(d_graph_visited, no_of_nodes*sizeof(bool), h_graph_visited);
_clMemcpyH2D(d_cost, no_of_nodes*sizeof(int), h_cost);
//--2 invoke kernel
#ifdef PROFILING
timer kernel_timer;
double kernel_time = 0.0;
kernel_timer.reset();
kernel_timer.start();
#endif
int kerId=0;
// printf("launching kernel\n");
do{
// printf("copy in\n");
h_over = false;
_clMemcpyH2D(d_over, sizeof(bool), &h_over);
//--kernel 0
int kernel_id = 0;
int kernel_idx = 0;
// printf("set arg 1\n");
_clSetArgs(kernel_id, kernel_idx++, d_graph_nodes);
_clSetArgs(kernel_id, kernel_idx++, d_graph_edges);
_clSetArgs(kernel_id, kernel_idx++, d_graph_mask);
_clSetArgs(kernel_id, kernel_idx++, d_updating_graph_mask);
_clSetArgs(kernel_id, kernel_idx++, d_graph_visited);
_clSetArgs(kernel_id, kernel_idx++, d_cost);
_clSetArgs(kernel_id, kernel_idx++, &no_of_nodes, sizeof(int));
//int work_items = no_of_nodes;
// printf("invoke 1\n");
_clInvokeKernel(kernel_id, no_of_nodes, work_group_size);
//--kernel 1
kernel_id = 1;
kernel_idx = 0;
// printf("set arg 2\n");
_clSetArgs(kernel_id, kernel_idx++, d_graph_mask);
_clSetArgs(kernel_id, kernel_idx++, d_updating_graph_mask);
_clSetArgs(kernel_id, kernel_idx++, d_graph_visited);
_clSetArgs(kernel_id, kernel_idx++, d_over);
_clSetArgs(kernel_id, kernel_idx++, &no_of_nodes, sizeof(int));
//work_items = no_of_nodes;
// printf("invoke 2\n");
_clInvokeKernel(kernel_id, no_of_nodes, work_group_size);
// printf("copy back\n");
_clMemcpyD2H(d_over,sizeof(bool), &h_over);
// printf("done\n");
// printf("K%d\n",kerId++);
}while(h_over);
// printf("done!");
_clFinish();
#ifdef PROFILING
kernel_timer.stop();
kernel_time = kernel_timer.getTimeInSeconds();
#endif
//--3 transfer data from device to host
_clMemcpyD2H(d_cost,no_of_nodes*sizeof(int), h_cost);
//--statistics
#ifdef PROFILING
std::cout<<"kernel time(s):"<<kernel_time<<std::endl;
#endif
//--4 release cl resources.
_clFree(d_graph_nodes);
_clFree(d_graph_edges);
_clFree(d_graph_mask);
_clFree(d_updating_graph_mask);
_clFree(d_graph_visited);
_clFree(d_cost);
_clFree(d_over);
_clRelease();
// }
// catch(std::string msg){
// _clFree(d_graph_nodes);
// _clFree(d_graph_edges);
// _clFree(d_graph_mask);
// _clFree(d_updating_graph_mask);
// _clFree(d_graph_visited);
// _clFree(d_cost);
// _clFree(d_over);
// _clRelease();
// std::string e_str = "in run_transpose_gpu -> ";
// e_str += msg;
// throw(e_str);
// }
return ;
}
//----------------------------------------------------------
//--breadth first search on GPUs
//----------------------------------------------------------
void run_bfs_gpu(int no_of_nodes, Node *h_graph_nodes, int edge_list_size, \
int *h_graph_edges, int *h_graph_mask, int *h_updating_graph_mask, \
int *h_graph_visited, int *h_cost)
throw(std::string){
//int number_elements = height*width;
int h_over;
cl_mem d_graph_nodes, d_graph_edges, d_graph_mask, d_updating_graph_mask, \
d_graph_visited, d_cost, d_over;
try{
//--1 transfer data from host to device
_clInit();
d_graph_nodes = _clMalloc(no_of_nodes*sizeof(Node), h_graph_nodes);
d_graph_edges = _clMalloc(edge_list_size*sizeof(int), h_graph_edges);
d_graph_mask = _clMallocRW(no_of_nodes*sizeof(int), h_graph_mask);
d_updating_graph_mask = _clMallocRW(no_of_nodes*sizeof(int), h_updating_graph_mask);
d_graph_visited = _clMallocRW(no_of_nodes*sizeof(int), h_graph_visited);
d_cost = _clMallocRW(no_of_nodes*sizeof(int), h_cost);
d_over = _clMallocRW(sizeof(int), &h_over);
_clMemcpyH2D(d_graph_nodes, no_of_nodes*sizeof(Node), h_graph_nodes);
_clMemcpyH2D(d_graph_edges, edge_list_size*sizeof(int), h_graph_edges);
_clMemcpyH2D(d_graph_mask, no_of_nodes*sizeof(int), h_graph_mask);
_clMemcpyH2D(d_updating_graph_mask, no_of_nodes*sizeof(int), h_updating_graph_mask);
_clMemcpyH2D(d_graph_visited, no_of_nodes*sizeof(int), h_graph_visited);
_clMemcpyH2D(d_cost, no_of_nodes*sizeof(int), h_cost);
//--2 invoke kernel
#ifdef PROFILING
timer kernel_timer;
double kernel_time = 0.0;
kernel_timer.reset();
kernel_timer.start();
#endif
struct timespec startT, endT;
clock_gettime(CLOCK_MONOTONIC, &startT);
do{
h_over = false;
_clMemcpyH2D(d_over, sizeof(int), &h_over);
//--kernel 0
int kernel_id = 0;
int kernel_idx = 0;
_clSetArgs(kernel_id, kernel_idx++, d_graph_nodes);
_clSetArgs(kernel_id, kernel_idx++, d_graph_edges);
_clSetArgs(kernel_id, kernel_idx++, d_graph_mask);
_clSetArgs(kernel_id, kernel_idx++, d_updating_graph_mask);
_clSetArgs(kernel_id, kernel_idx++, d_graph_visited);
_clSetArgs(kernel_id, kernel_idx++, d_cost);
_clSetArgs(kernel_id, kernel_idx++, &no_of_nodes, sizeof(int));
//int work_items = no_of_nodes;
_clInvokeKernel(kernel_id, no_of_nodes, work_group_size);
//--kernel 1
kernel_id = 1;
kernel_idx = 0;
_clSetArgs(kernel_id, kernel_idx++, d_graph_mask);
_clSetArgs(kernel_id, kernel_idx++, d_updating_graph_mask);
_clSetArgs(kernel_id, kernel_idx++, d_graph_visited);
_clSetArgs(kernel_id, kernel_idx++, d_over);
_clSetArgs(kernel_id, kernel_idx++, &no_of_nodes, sizeof(int));
//work_items = no_of_nodes;
_clInvokeKernel(kernel_id, no_of_nodes, work_group_size);
_clMemcpyD2H(d_over,sizeof(int), &h_over);
}while(h_over);
_clFinish();
clock_gettime(CLOCK_MONOTONIC, &endT);
uint64_t diff = 1000000000 * (endT.tv_sec - startT.tv_sec);
uint64_t nanodiff = endT.tv_nsec - startT.tv_nsec;
//printf("elapsed accelerator time = %llu nanoseconds\n", (long long unsigned int) diff);
//printf("start time seconds%u \n", startT.tv_sec);
//printf("end time seconds %u \n", endT.tv_sec);
//printf("difference %u \n", diff);
//printf("start time nanoseconds %u \n", startT.tv_nsec);
//printf("end time nanoseconds %u \n", endT.tv_nsec);
printf(" accelerator time %u \n", nanodiff + diff);
#ifdef PROFILING
kernel_timer.stop();
kernel_time = kernel_timer.getTimeInSeconds();
#endif
//--3 transfer data from device to host
_clMemcpyD2H(d_cost,no_of_nodes*sizeof(int), h_cost);
//--statistics
#ifdef PROFILING
std::cout<<"kernel time(s):"<<kernel_time<<std::endl;
#endif
//--4 release cl resources.
_clFree(d_graph_nodes);
_clFree(d_graph_edges);
_clFree(d_graph_mask);
_clFree(d_updating_graph_mask);
_clFree(d_graph_visited);
_clFree(d_cost);
_clFree(d_over);
_clRelease();
}
catch(std::string msg){
_clFree(d_graph_nodes);
_clFree(d_graph_edges);
_clFree(d_graph_mask);
_clFree(d_updating_graph_mask);
_clFree(d_graph_visited);
_clFree(d_cost);
_clFree(d_over);
_clRelease();
std::string e_str = "in run_transpose_gpu -> ";
e_str += msg;
throw(e_str);
}
return ;
}
