Sgemm question

Accelerated Computing CUDA CUDA Programming and Performance
#1

I tried multiply 2 matrix in GPU memory with this code:

test=(float*)malloc(Ne*Ne*sizeof(test[0]));

	initmatrix(test,Ne,3.0); //initialize first matrix with 3 each element

	cudaMallocPitch((void**)&d_wmas, &pitch, Ne*Ne*sizeof(float),Nl);

	cublasAlloc(Ne*Ne,sizeof(d_test[0]),(void**)&d_test);

	cublasSetVector(Ne*Ne,sizeof(d_wmas[0]),test,1,d_wmas,1);

	initmatrix(test,Ne,2.0);

	cublasSetVector(Ne*Ne,sizeof(test[0]),test,1,*(&d_wmas+Ne*Ne),1);

	cudaThreadSynchronize();

	cublasSgemm('n','n',Ne,Ne,Ne,1.0,d_wmas,Ne,*(&d_wmas+Ne*Ne),Ne,0.0,d_test,Ne);

	status = cublasGetError();

 Â  Â if (status != CUBLAS_STATUS_SUCCESS) {

 Â  Â  Â  Â fprintf (stderr, "!!!! kernel execution error.\n");

 Â  Â  

 Â  Â }

	cudaThreadSynchronize();

	cublasGetVector(Ne*Ne,sizeof(d_test[0]),d_test,1,test,1);

	cudaThreadSynchronize();

And I got just second matrix returned in the test array. Am I doing something wrong? I checked that both first matrix writes in *d_wmas and second in (&d_wmas+NeNe) correct.

#2

I’m not completely sure, but you have to specify the size of each element, don’t you?
-> (&d_wmas+NeNe*sizeof(d_float))

Its different to pointer arithmetic I think.

float * fPoint;
you can point to the next float by
fPoint ++;

#3

As far as I know if in C add value to the pointer on array some value pointer will be increased for a value elements of array type. But I tried your solution too with the same result :( Anyway thank you for help.