Unable to run Magma_dgeqrf_batched

[naveenkumar]

Hello, I wanted to run batched QR decomposition using magma. I was able to successfully run it using the cublas_dgeqrf_batched code but couldn't run it using magma_dgeqrf_batched. As I am new to magma I am not able to figure out where I am going wrong, so can anyone guide me to what correction do I have to do in my code to successfully run the magma_dgeqrf_batched. I will be posting my code below. Any help would be much appreciated, Thank you in advance

Code: Select all

int main() {

   magma_init (); // initialize Magma
    magma_queue_t queue = NULL ;
    magma_int_t dev =0;
    magma_queue_create (dev ,& queue );

	magma_int_t m = 3;
	magma_int_t  n = 2;
	magma_int_t  batchSize = 2;//2 small matrices
	
	magma_int_t lda = 3;
	
	magma_int_t *info;
	 magma_imalloc_cpu(&info,batchSize);
	 
	
//step -1 Allocate storage for batch count 
    double **A,**tau;
	
	
	A = (double**)malloc(batchSize * sizeof(double*));
	for (int i = 0; i < batchSize; i++) {
		A[i] = (double*)malloc(m*n * sizeof(double));
	}
	
	
    tau = (double**)malloc(batchSize * sizeof(double*));        
    for (int i = 0; i < batchSize; i++) 
    {
    		tau[i] = (double*)malloc(n * sizeof(double));
    }
	
     
	//info = (int*)malloc(batchSize * sizeof(int));
	
//step -2 create host pointer array to the gpu array
    double **d_A, **h_d_A,**h_d_tau,**d_tau;
	h_d_A = (double**)malloc(batchSize * sizeof(double*));
	h_d_tau = (double**)malloc(batchSize * sizeof(double*));
	
	
	for (int i = 0; i < batchSize; i++) {
		cudaMalloc((void**)&h_d_A[i], m*n * sizeof(double));
		h_d_tau[i] = (double*)malloc(n * sizeof(double));
	}
	
	
//step -3 copy host array of pointers to device
	cudaMalloc((void**)&d_A, batchSize * sizeof(double*));
	cudaMalloc((void**)&d_tau, batchSize * sizeof(double*));
	
	
	cudaMemcpy(d_A, h_d_A, batchSize * sizeof(double*), cudaMemcpyHostToDevice);
	cudaMemcpy(d_tau, h_d_tau, batchSize * sizeof(double*), cudaMemcpyHostToDevice);
	
	
	//fill up the matrix A
	for (int k = 0; k < batchSize; k++) {
		for (int j = 0; j < n; j++) {
			for (int i = 0; i < m; i++) {
				int index = j * m + i;//not tested
				if (i == j) {
					(A[k])[index] = 2.0000000;
				}
				else {
					(A[k])[index] = 1.0000000;

				}
			} // i  
		} // j
	} // k
	
  
	
	for (int i = 0; i < batchSize; i++)
	{
		magma_dsetmatrix( m, n, A[i], lda, h_d_A[i], lda ,queue);
		
		
	}
	
   
	
	magma_dgeqrf_batched( m, n, d_A, lda, d_tau, info, batchSize,queue);
	
	for (int i = 0; i < batchSize; i++)
	{
		magma_dgetmatrix(m, n, h_d_A[i], lda, A[i], lda,queue);
		//cudaMemcpy(A[i], h_d_A[i], m*n * sizeof(double), cudaMemcpyDeviceToHost);
	}
	
	
//print the A matrix
 
	for (int k = 0; k < batchSize; k++) {
		for (int j = 0; j < m; j++) {
			for (int i = 0; i < n; i++) {
				int index = j * m + i;//not tested
				//count = count + 1;
				printf("\n %d The values are %lf",k+index, A[k][index]);
			} // i
		} // j
	} // k
	
	free(tau);
	free(A);
	free(h_d_A);
	free(h_d_tau);
	cudaFree(d_A);
	cudaFree(d_tau);
	
}

[mgates3]

See magma/testing/testing_dgeqrf_batched.cpp for sample code.

I can't debug your code for you, but from a quick scan, I'm wondering why h_d_A is allocated with cudaMalloc on the GPU device, while h_d_tau is allocated with malloc on the CPU host.

-mark

[Sorry for the delay in posting. A misconfiguration left messages in the moderation queue.]