Varying Execution time

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi,

I run the follow code and I realized that the execution time change, sometimes is 0 and sometimes is 0.001(too much for that code!). Does anyone know why this happend?

Do you thing that I can improve the performance of the code storing the scalar variable in device as well ?

//all the matrices and vectors are stored in the device.

//all the scalar are stored in the host

.

.

clock_t start = clock();

		cublasScopy(N, v, 1, v_old, 1); //v_old = v;		

		

		//v = v_hat/beta; 

		cublasScopy (N, v_hat, 1, v, 1); // v = v_hat;

		cublasSscal(N,(1/beta),v,1);

		// alpha = v'*A*v;

		cublasSsbmv('U', N, K, 1, Ab, (2*K+1), v, 1, 0, sup0, 1); // sup0= A*v;

		alpha= cublasSdot(N, sup0, 1,v, 1); // alpha = dot(sup0,sup0)=sup0'*sup0

	

				//v_hat = A*v - alpha*v - beta*v_old;

		cublasSaxpy(N, -alpha, v, 1, sup0, 1); // sup0 = -alpha*v +sup0 ----- sup0 was A*v

		cublasSaxpy(N, -beta, v_old, 1, sup0, 1); // sup0 = -beta*v_old +sup0

		cublasScopy(N, sup0, 1, v_hat, 1); //v_hat = sup0;

	

		beta_old = beta;				// beta_old = beta;

		beta = cublasSnrm2(N,v_hat,1);		 // beta = norm(v_hat);

printf("\n Time elapsed : %f \n", ((double)clock() - start) / CLOCKS_PER_SEC);

.

.

.
2

Apparently [font=“Courier New”]CLOCKS_PER_SEC[/font] is 1000 on your system. Before you try to optimize your code based on measured timings, improve the measurement itself. Run the timed code multiple times in a loop, or user a higher precision timer like CudaEvents.

3

You also need a cudaThreadSynchronize() prior to each measurement of wall clock time. gettimeofday has a much higher resolution than clock()