293   cuda(Malloc(&random_d,size_hi_in));  in hs_bench()
294   cuda(Malloc(&vin_d,   size_hi_in));  in hs_bench()
295   cuda(Malloc(&vout_d,  size_hi_out));  in hs_bench()
306   cuda(Memcpy(random_d,random_h,size_hi_in,cudaMemcpyHostToDevice));  in hs_bench()
342   cuda(EventCreate(&start));  in hs_bench()
343   cuda(EventCreate(&end));  in hs_bench()
352       cuda(Memcpy(vin_d,random_d,count*key_size,cudaMemcpyDeviceToDevice));  in hs_bench()
370           cuda(EventRecord(start,stream0));  in hs_bench()
371           cuda(StreamWaitEvent(stream1,start,0));  in hs_bench()
372           cuda(StreamWaitEvent(stream2,start,0));  in hs_bench()
384           cuda(EventRecord(end,stream0));  in hs_bench()
386           cuda(EventSynchronize(end));  in hs_bench()
390           cuda(EventElapsedTime(&elapsed,start,end));  in hs_bench()
411 	  cuda(Memcpy(sorted_h,vin_d, size_padded_in,cudaMemcpyDeviceToHost));  in hs_bench()
412 	  cuda(Memcpy(vout_h,  vout_d,size_padded_in,cudaMemcpyDeviceToHost));  in hs_bench()
472   cuda(EventDestroy(start));  in hs_bench()
473   cuda(EventDestroy(end));  in hs_bench()
478   cuda(Free(random_d));  in hs_bench()
479   cuda(Free(vin_d));  in hs_bench()
480   cuda(Free(vout_d));  in hs_bench()
496   cuda(GetDeviceProperties(&props,device));  in main()
498   cuda(SetDeviceFlags(cudaDeviceScheduleBlockingSync));  in main()
499   cuda(SetDevice(device));  in main()
503   cuda(DriverGetVersion(&driver_version));  in main()
517   cuda(StreamCreate(&stream0));  in main()
518   cuda(StreamCreate(&stream1));  in main()
519   cuda(StreamCreate(&stream2));  in main()
597   cuda(StreamDestroy(stream0));  in main()
598   cuda(StreamDestroy(stream1));  in main()
599   cuda(StreamDestroy(stream2));  in main()
601   cuda(DeviceReset());  in main()