Cublas batched lu decomposition get segmentation fault

Accelerated Computing GPU-Accelerated Libraries
1

I need to perform lu decomposition to several small matrices in the same time. However I get segmentation fault. I attach my source code as following and my Gpu card is C2075 with cuda 5.5 and the driver is 319.37. I appreciate any suggestion.

//matrix_b include batchsize number of (dim*dim) matrices
int matrix_gpu_lu_batch(double *matrix_b, int dim, int batchsize )
{
	int m = dim;
	cublasStatus_t status_lu;
	cublasStatus_t status;
	cudaError_t cudaStatus;
	cudaError_t err1,err2,err3;
	
	cublasHandle_t handle;
	if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS)
	{
		fprintf(stdout, "CUBLAS initialization failed!\n");
		cudaDeviceReset();
		exit(1);
	}
	cudaStream_t streamArray; 
	cublasSetStream(handle, streamArray);
	
	
	//matrix buffers for LU decomposition, I need to keep the original matrices 
	double *d_lubuffer;
	err1 = cudaMalloc((void **)&d_lubuffer, dim*dim*batchsize*sizeof(double));
	
	//pointer array hold all matrix pointers
	double **h_ptrlubuffer = new double *[batchsize];

	for(int i=0;i<batchsize;i++){
		err1 = cudaMemcpy(&d_lubuffer[i*m*m],&matrix_b[i*m*m],dim*dim*sizeof(double),cudaMemcpyDeviceToDevice);	//matrix_b locates in device memory		
		h_ptrlubuffer[i] = &d_lubuffer[i*m*m];
	}

	//pointer array in device memory hold all matrix pointers
	double **d_ptrlubuffer;
	err1 = cudaMalloc((void **)&d_ptrlubuffer,batchsize*sizeof(double *));
	err2 = cudaMemcpy(d_ptrlubuffer,h_ptrlubuffer,batchsize*sizeof(double *),cudaMemcpyHostToDevice);
 
	int *d_ipiv;
	err1 = cudaMalloc((void **)&d_ipiv, dim*batchsize*sizeof(int));	
	int *d_info;
	err1 = cudaMalloc((void **)&d_info, batchsize*sizeof(int));

	status_lu = cublasDgetrfBatched(handle,dim,d_ptrlubuffer,dim,d_ipiv,d_info,batchsize);
	
	
	cublasDestroy(handle);
	cudaDeviceReset();
}
2

(1) Do any of the status returns (err1, err2) indicate an error ? The code as posted doesn’t seem to check the status returns.
(2) Have you been able to narrow down which line of the code gives rise to the segfault?

The fact that there is a segmentation fault would indicate that host code is trying to de-reference an invalid pointer, possibly by use of a device pointer where a host pointer was intended, or by an out of bounds reference in the host code.

In a quick perusal of the code I have not been able to spot anything that looks like out of whack. You may want to consider unifying the code to use of either “dim” or “m”. The interchangeable use of both (sometimes in the same line) is a bit confusing.

3

There is no status error. I did not paste the status check part to save space.

status_lu = cublasDgetrfBatched(handle,dim,d_ptrlubuffer,dim,d_ipiv,d_info,batchsize);

This line give me the segmentation fault.

I will try to unify the use of “dim” and “m”

4

This is not the correct way to set a stream:

cudaStream_t streamArray; 	
cublasSetStream(handle, streamArray);

You need to actually create the stream, before passing it to the cublas function:

cudaStream_t streamArray; 	
cudaStreamCreate(&streamArray);
cublasSetStream(handle, streamArray);