CUDA functions in DLL, __declspec(dllexport) works but strange results? - qt

Question

I've come across a weird problem with a piece of CUDA code. It's compiled into a DLL using msvc community 2015 and nvcc in Windows 10. I'm using CUDA 8. The application calling the dll is being developed with Qt5.
The application is fairly large and complicated: using Qt, CUDA, VTK, HDF5. It all seems to work, the app runs and does what it's supposed to, but fails in a reproducible manner that doesn't seem to make any sense. The example function below seems to reproduce a similar error.
I'm compiling the dll with:
nvcc -m64 -arch=sm_20 -o fdm1_cuda.dll -Xcompiler "/LD /D_USRDLL /D_WINDLL" fdm1_cuda.cu
This function seems to exhibit the same problem as the main code:
extern "C" __declspec(dllexport) void fdm1_funnyproblemchecker(){
cudaError_t errorcode;
float *a_host;
float *b_host;
float *a_device;
int num, i;
num=10;
a_host = (float *)malloc(sizeof(float)*num);
if( a_host) printf("Result check, allocate host memory a: success\n");
if(!a_host) printf("Result check, allocate host memory a: failed!\n");
for(i=0;i<num;i++) a_host[i] = (float)i;
for(i=0;i<num;i++) printf("%6.3f ", a_host[i]);
printf("\n");
b_host = (float *)malloc(sizeof(float)*num);
if( b_host) printf("Result check, allocate host memory b: success\n");
if(!b_host) printf("Result check, allocate host memory b: failed!\n");
errorcode = cudaSuccess;
cudaMalloc((void **) &a_device, sizeof(float)*num);
errorcode = cudaGetLastError();
printf("Result check, allocate device memory: %s\n", cudaGetErrorString(errorcode));
errorcode = cudaSuccess;
cudaMemcpy(a_device, a_host, num*sizeof(float), cudaMemcpyHostToDevice);
errorcode = cudaGetLastError();
printf("Result check, copy host to device : %s\n", cudaGetErrorString(errorcode));
errorcode = cudaSuccess;
cudaMemcpy(b_host, a_device, num*sizeof(float), cudaMemcpyDeviceToHost);
errorcode = cudaGetLastError();
printf("Result check, copy device to host : %s\n", cudaGetErrorString(errorcode));
for(i=0;i<num;i++) printf("%6.3f ", b_host[i]);
printf("\n");
fflush(stdout);
cudaFree(a_device);
free(a_host);
free(b_host);
}
Sometimes the output from this is:
Result check, allocate host memory a: success
0.000 1.000 2.000 3.000 4.000 5.000 6.000 7.000 8.000 9.000
Result check, allocate host memory b: success
Result check, allocate device memory: no error
Result check, copy host to device : no error
Result check, copy device to host : no error
0.000 1.000 2.000 3.000 4.000 5.000 6.000 7.000 8.000 9.000
If I change something which I don't think is related elsewhere in the application (changing the size of a model during runtime), I get this:
Result check, allocate host memory a: success
0.000 1.000 2.000 3.000 4.000 5.000 6.000 7.000 8.000 9.000
Result check, allocate host memory b: success
Result check, allocate device memory: no error
Result check, copy host to device : an illegal memory access was encountered
Result check, copy device to host : an illegal memory access was encountered
0.000 0.000 0.000 0.000 0.000 0.000 0.000 270355481144287188484096.000 74936693461279934656588472647680.000 0.000
So, there is a cudaMemcpy failure.
I can't tell if it is a host malloc issue, a cudaMalloc issue, or something related to it running from a dll.
Can anyone see what I'm missing here?
I've had this application running in Linux and Mac, using dynamic libraries, without any major problems. I'm now trying to get it going under windows.

Answer 1

Problem solved.
It was a kernel elsewher in the code accessing an array[-1] out of bounds.
My error checking with cudaGetLastError was incorrect. If I do cudaDeviceSynchronize before each cudaGetLastError it reports errors I was missing.
Thanks.