"Invalid argument" error with Streams

Accelerated Computing CUDA CUDA Programming and Performance
1

Hello,

i’m beggining to “play” with streams and asynchronous execution operations.

Well, first i wrote a code where i have to copy from device memory to host memory a few bytes.

for( i = iniciot;i < fimt;i++){

		cutilSafeCall(cudaMemcpy(parada + (i*raios+inicior),d_parada+(i*raios+inicior),(fimr-inicior)*sizeof(int),cudaMemcpyDeviceToHost,));

		

		for(j = inicior;j < fimr;j++){ 

			posicao = i*(int)posicao1 + j*(int)posicao2;

				

			size = (parada[i*raios+j]-posicao)*sizeof(Slots);

			out[i][j] = (Slots*)malloc(size);

					

			cutilSafeCall(cudaMemcpyAsync( out[i][j], d_out + posicao , size ,cudaMemcpyDeviceToHost));

		}

	}

Tested and it works.

So i tried to add some stream support to do all this copy faster. I got the idea from programming guide (Section 3.2.6.5).This is the code:

cudaStream_t stream1[fimt-iniciot],stream2[fimr-inicior];

	

	for(i=0;i < (fimt-iniciot);i++)

		cudaStreamCreate(&stream1[i]);

	for(i=0;i < (fimr-inicior);i++)	

		cudaStreamCreate(&stream2[i]);

	

	for( i = iniciot;i < fimt;i++){

		cutilSafeCall(cudaMemcpyAsync(parada + (i*raios+inicior),d_parada+(i*raios+inicior),(fimr-inicior)*sizeof(int),cudaMemcpyDeviceToHost,stream1[i]));

		

		for(j = inicior;j < fimr;j++){ 

			//A variavel posicao marca o endereço aonde o raio verificado começou a ser gravado na memoria global da GPU

			posicao = i*(int)posicao1 + j*(int)posicao2;

				

			size = (parada[i*raios+j]-posicao)*sizeof(Slots);

			out[i][j] = (Slots*)malloc(size);

					

			cutilSafeCall(cudaMemcpyAsync( out[i][j], d_out + posicao , size ,cudaMemcpyDeviceToHost,stream2[j]));

		}

	}

	cudaThreadSynchronize();

Well, when i execute i got this: cudaSafeCall() Runtime API error : invalid argument.

What i’m doing wrong?

2

Sorry to bump this old thread, but I didn’t see the solution posted anywhere else. The solution is probably this line from the Reference Manual:

3

For what it’s worth, this error does not appear if the NULL stream is used, even on non-pagedlocked host memory.