Debugging CUDA - CudaUnknownError - c++

Question

I am trying to use CUDA to create a bitmap image of the mandlebrot set. I have looked at a few tutorials and already got some help here for the process of integrating the unmanaged CUDA dll with the managed C# gui. The problem I am having now is that my CUDA dll is not forming the bitmap correctly - and when I use an error checking macro on cudaDeviceSynchronize() after the kernel launch, I get cudaUnknownError.
Here is the relevant code:
#define gpuErrchk(ans) { gpuAssert((ans), __FILE__, __LINE__); }
inline void gpuAssert(cudaError_t code, char *file, int line, bool abort=true)
{
if (code != cudaSuccess)
{
fprintf(stderr,"GPUassert: %s %s %d\n", cudaGetErrorString(code), file, line);
if (abort) exit(code);
}
}
struct complex
{
float r, i;
__device__ complex(float _r, float _i) : r(_r), i(_i) {}
__device__ float magnitudeSquared(){ return (r*r + i*i) ; }
__device__ complex& operator*=(const complex& rhs)
{
r = (r * rhs.r - i * rhs.i);
i = (r * rhs.i + i * rhs.r);
return *this;
}
__device__ complex& operator+=(const complex& rhs)
{
r = (r + rhs.r);
i = (i + rhs.i);
return *this;
}
};
__device__ int mandlebrotDiverge(complex *z)
{
complex c(*z);
int i = 0;
for(i = 0; i < MAX_ITERATIONS; i++)
{
*z *= *z;
*z += c;
if(z->magnitudeSquared() >= 2)
{
return 1;
}
}
return 0;
}
__global__ void kernel(int *ptr, int width, int height)
{
int x = threadIdx.x + blockIdx.x * blockDim.x;
int y = threadIdx.y + blockIdx.y * blockDim.y;
int offset = x + y * blockDim.x * gridDim.x;
float scale = 1.5f;
complex z(scale*(float)(width/2 - x)/(width/2), scale*(float)(height/2 - y)/(height/2));
if(offset < (1920*1080))
{
int mValue = mandlebrotDiverge(&z);
ptr[offset*3 + (uint8_t)0] = (uint8_t)(mValue*255);
ptr[offset*3 + (uint8_t)1] = (uint8_t)(mValue*255);
ptr[offset*3 + (uint8_t)2] = (uint8_t)(mValue*255);
}
}
extern "C" __declspec(dllexport) void __cdecl generateBitmap(void *bitmap)
{
int width = 1920;
int height = 1080;
int *dev_bmp;
dim3 blocks(width/16, height/16);
dim3 threads(16, 16);
gpuErrchk(cudaMalloc((void**)&dev_bmp, (3*width*height)));
kernel<<<blocks, threads>>>(dev_bmp, width, height);
gpuErrchk(cudaPeekAtLastError());
gpuErrchk(cudaDeviceSynchronize());
gpuErrchk(cudaMemcpy(bitmap, dev_bmp, (width*height*3), cudaMemcpyDeviceToHost));
cudaFree(dev_bmp);
}
When I step through the code, everything appears to be working correctly until I get to gpuErrchk(cudaDeviceSynchronize()); - when I step into that, the error code simply says 'cudaUnknownError'. I have no clue what I am doing wrong at this point. Any help or advice on how to improve this solution would be appreciated.
EDIT:
Okay, looked at CUDA_memcheck, and I get this error (hundreds of times):
========= CUDA-MEMCHECK
========= Invalid __global__ write of size 4
========= at 0x00000a10 in C:/.../kernel.cu:77:kernel(int*, int, int)
========= by thread (15,11,0) in block (1,17,0)
========= Address 0x05a37f74 is out of bounds
========= Saved host backtrace up to driver entry point at kernel launch time
So I changed from *int to *unsigned char because I am trying to allocate arrays of individual bytes, not ints. Lots of errors cleared up, but now I get this:
========= CUDA-MEMCHECK
========= Program hit error 6 on CUDA API call to cudaDeviceSynchronize
========= Saved host backtrace up to driver entry point at error
========= Host Frame:C:\Windows\system32\nvcuda.dll (cuD3D11CtxCreate + 0x102459) [0x11e4b9]
========= Host Frame:C:\...\cudart32_55.dll (cudaDeviceSynchronize + 0xdd) [0x1149d]
========= Host Frame:C:\...\FractalMaxUnmanaged.dll (generateBitmap + 0xf0) [0x97c0]
=========
========= ERROR SUMMARY: 1 error
Okay, I'm making progress, but now when I step through the c# application, the byte buffer has 255 as the value for every byte, which doesn't make sense. Here is the c# code:
public unsafe class NativeMethods
{
[DllImport(#"C:\Users\Bill\Documents\Visual Studio 2012\Projects\FractalMaxUnmanaged\Debug\FractalMaxUnmanaged.dll", CallingConvention=CallingConvention.Cdecl)]
public static extern void generateBitmap(void *bitmap);
public static Bitmap create()
{
byte[] buf = new byte[1920 * 1080 * 3];
fixed (void* pBuffer = buf)
{
generateBitmap(pBuffer);
}
IntPtr unmanagedPtr = Marshal.AllocHGlobal(buf.Length);
Marshal.Copy(buf, 0, unmanagedPtr, buf.Length);
Bitmap img = new Bitmap(1920, 1080, 1920*3, PixelFormat.Format24bppRgb, unmanagedPtr);
Marshal.FreeHGlobal(unmanagedPtr);
return img;
}
}

Answer 1

Your problem here is that your memory allocations and copies are wrong, you are forgetting that cudaMalloc/cudaMemcpy expect the size in bytes. Since int uses 4 bytes, you are actually allocating less memory than required by your kernel. Use this instead (or use unsigned char which only needs 1 byte):
cudaMalloc((void**)&dev_bmp, (3*width*height)*sizeof(int));
cudaMemcpy(bitmap, dev_bmp, (3*width*height)*sizeof(int), cudaMemcpyDeviceToHost);
Also make sure that bitmap was allocated correctly. As #Eugene said, using cuda-memcheck is a good way to find the source of that kind of errors.