Runtime Error (17) in Win7 (64-bit)

[zhangft]

When I use culaDeviceSgesv and culaDeviceSgetrf in Windows (32-bit), everything's ok. But when I change the platform and culatools to 64-bit,
Runtime Error (17) occurs with the same code. I don't know why.

[zhangft]

part of the code is as below:

Code: Select all

float *d_A;
   cutilSafeCall(cudaMalloc((void**) &d_A, sizeof(float)*size*size ));   
    cutilSafeCall(cudaMemcpy(d_A, matrixA, sizeof(float)*size*size,
                              cudaMemcpyHostToDevice));
   int* d_ipiv;
   cutilSafeCall(cudaMalloc((void**) &d_ipiv, sizeof(int)*size ));   
   cutilSafeCall(cudaMemset ((void *)d_ipiv, 0, sizeof(int)*size));

   float *d_B;
   cutilSafeCall(cudaMalloc((void**) &d_B, sizeof(float)*size*size ));   
   cutilSafeCall(cudaMemset ((void *)d_B, 0, sizeof(int)*size*size ));
   initIdentityGPU_old <<< (size+255)/256, 256 >>> (d_B, size);

   culaStatus status;   
    status = culaInitialize();
    checkStatus(status);   
   // get inverse of d_A, and store in d_B   
   status = culaDeviceSgesv(size, size, d_A, size, d_ipiv, d_B, size);
    checkStatus(status);   
    culaShutdown();

[zhangft]

culaSgesv runs well, however culaDeviceSgesv desn't. culaGetDeviceCount and culaGetExecutingDevice are fine. I wonder wether there are some bugs in culapackdevice for x64.

[zhangft]

I tested your examples ---> device, the same error happened. The CUDA version I use is CUDA3.0 for x64.

[zhangft]

I found the reason! culaDevice doesn't support CUDA3.0 yet!
I changed back to CUDA2.3 and the issue was sovled!
Hope cula2.0 release soon!

[john]

Indeed, soon. In the meantime you'll probably find that there is very little speed difference between using the host and device interfaces.

[zhangft]

Yes, I found it. But I would do some other caculations with the result of Sgesv in GPU. To reduce the I/O time I would not transfer the result out and in, so I choose the device interface.

[john]

I would recommend always benchmarking the times to prove to yourself that the use of the Device interface will reduce your overall time. When you allow CULA to do the allocation and transfers, we have more flexibility. Often the performance between our Host and Device interfaces is negligible and occasionally the Host interface outperforms a poorly allocated Device call.