###University of Pennsylvania, CIS 565: GPU Programming and Architecture, Project 2

Ju Yang

global void naive_sum(int n,int* odata, int* idata);

void scan(int n, int *odata, const int *idata);

global void prescan(int *g_odata, int g_idata, int n, inttemp);

void scan(int n, int *odata, const int *idata);

int compact(int n, int *odata, const int *idata);

void scan(int n, int *odata, const int *idata);

void scan(int n, int *odata, const int *idata);

int compactWithoutScan(int n, int *odata, const int *idata) ;

int compactWithScan(int n, int *odata, const int *idata);

global void kernMapToBoolean(int n, int *bools, const int *idata);

global void kernScatter(int n, int *odata, const int *idata, const int *bools, const int *indices);

As we can see, the thrust::exclusive_scan is rather time-costing compared with other methods. Even if I used device_vector to store the data, it is still the slowest. But since I did not free the device_vectors, the non-pow2 as second round's speed is much faster. I think the reason is, when calling thrust functions, it will apply for some blocks/threads inside the GPU, and will release later on. Although I tried my best to avoid any read/write from CPU to GPU, the scan function still cost some time to arrange for some place.

       I think this is because the SIZE is still not large enough?