cublasStrsm_v2 returns unchanged input matrix

Accelerated Computing GPU-Accelerated Libraries
1

I am trying to execute CUBLAS TRSM (AX=B, solving for X, left-lower-nonTranspose-Unit combination)function in CUDA version 4.2. But no matter what, the output result remains the same as input B Matrix. I’d appreciate if anyone can help me figure it out. Here is my program:

#include <stdio.h>
#include <cuda.h>
#include <cuda_runtime.h>
#include <cublas_v2.h>

// Initializing Data for right hand side Matrix B
void randomInit(float* data, int size)
{
   for (int i = 0; i < size; ++i) {
     	float a = rand() / (float)RAND_MAX;
       int inti = (int) (a * 10);
       data[i] = inti;
    }
}

//Initializing Data for Lower Traingular Square Matrix A
void randomLowerTriangle(float* data, int n)
{
   for (int i = 0; i < n; i++) {
	for (int j = 0; j < n; j++) {
		if ( i == j ) {
       		data[ (i*n)+j ] = 1;
		} else if ( j > i ) {
       		data[ (i*n)+j ] = 0;
		} else { 
			float a = rand() / (float) RAND_MAX;
       		int inti = (int) (a * 10);
       		data[ (i*n)+j ] = inti;	
		}
	}
    }
}

int main(int argc , char **argv)
{
	int m = 5, n = 5, i, j;
	float alpha = 1.0;
       
      	int size_A = m * m;
       int mem_size_A = sizeof(float) * size_A;
       float* h_A = (float*)malloc(mem_size_A);

	int size_B = m * n;
       int mem_size_B = sizeof(float) * size_B;
  	float* h_B = (float*)malloc(mem_size_B);
	float* h_CUBLAS = (float*)malloc(mem_size_B);
	
	printf("Assigning random numbers to A and B\n");
	randomLowerTriangle(h_A, m);
   	randomInit(h_B, size_B);

	printf("Matrix A[%d X %d]:\n",m, n);
	for(i = 0; i < m; i++ ) {
     	   for(j = 0 ; j < m; j++ )
           	printf("%.3f\t", h_A[i*m+j]);
               	printf("\n");
       }

	printf("Matrix B[%d X %d]:\n", m, n);
       for(i = 0; i < m; i++ ){
     		for(j = 0; j < n; j++ )
          	printf("%.3f\t", h_B[ i*n+j ]);
                 printf("\n");
       }

	// allocate device memory
  	float* d_A, *d_B;
  	printf("Allocating memory on device for A and B\n");

  	cudaMalloc((void**) &d_A, mem_size_A);
  	cudaMalloc((void**) &d_B, mem_size_B);

	cudaMemcpy(d_A, h_A, mem_size_A, cudaMemcpyHostToDevice);
  	cudaMemcpy(d_B, h_B, mem_size_B, cudaMemcpyHostToDevice);
	
	cublasStatus_t ret;
  	cublasHandle_t handle;
  	ret = cublasCreate_v2(&handle);
   
     	ret = cublasStrsm_v2(handle,CUBLAS_SIDE_LEFT, CUBLAS_FILL_MODE_LOWER, CUBLAS_OP_N,
				 CUBLAS_DIAG_UNIT, m, n, &alpha, d_A, m, d_B, n);
		  	
  	cudaMemcpy(h_CUBLAS, d_B, mem_size_B, cudaMemcpyDeviceToHost);
	
	cudaDeviceSynchronize();

	printf("Matrix X (CUBLAS)[%d X %d]:\n", m, n);
       for(i = 0; i < m; i++ ){
     		for(j = 0; j < n; j++ )
          	printf("%.3f\t", h_CUBLAS[ i*n+j ]);
                 printf("\n");
       }

	free(h_A);
  	free(h_B);
	free(h_CUBLAS);  	
  	cudaFree(d_A);
  	cudaFree(d_B);
	cublasDestroy_v2(handle);  
  	cudaDeviceReset();
	return 0;
}
2

There may be more than a few things wrong here. First off you should make sure that the matrices are valid inputs. Also I think the the second and fourth parameters are wrong.

The second should be CUBLAS_FILL_MODE_UPPER, and the fourth should be CUBLAS_DIAG_NON_UNIT.

I am not sure why you have the cudaDeviceSynchronize() after the memcpy. You should check for errors at that point instead.

Also remember the whole column major thing, when entering your parameters.

Disclaimer: I did not test this but have used Strsm() before successfully, and at first glance think those are the possible problems. There may be others well, just experiment yourself until you get it right.

read this;

[url]cuBLAS :: CUDA Toolkit Documentation

3

Thank you! I keep forgetting that Cublas expects input in col-major format. That was the only issue.