CODEWITHSUNDEEP
c++c++11
briantilley
briantilley

Reputation: 31

Why won't this value assign?

I have a member variable, m_width, that won't hold any value assigned to it. The relevant part of the class looks as follows:

class GPUFrame
{
private:
    std::shared_ptr<void> m_deviceData;
    unsigned m_pitch = 0;
    unsigned m_width = 2; // testing if m_width will take any value
    unsigned m_height = 0;
    unsigned m_timestamp = 0; // time value in microseconds (absolute value is arbitrary)
    bool m_endOfStream = false; // signifies last frame in the stream

public:

    // make an entirely new allocation
    GPUFrame(unsigned imageWidth, unsigned imageHeight, unsigned allocationCols, unsigned allocationRows,
         unsigned timestamp, bool eos=false)
    {
        // initializer list was causing headaches
        m_pitch = 0;
        m_width = imageWidth;
        m_height = imageHeight;
        m_timestamp = timestamp;
        m_endOfStream = eos;

        // get space from CUDA
        void* newAllocation;
        cudaErr(cudaMallocPitch(&newAllocation, reinterpret_cast<size_t*>(&m_pitch), static_cast<size_t>(allocationCols), static_cast<size_t>(allocationRows)));

        // track allocation with the shared_ptr
        m_deviceData = std::shared_ptr<void>(newAllocation, [=](void* p){ cudaErrNE(cudaFree(p)); });

        std::cout << "imageWidth = " << imageWidth << ", m_width = " << m_width << std::endl;
    }

    // copy from given location
    GPUFrame(CUdeviceptr devPtr, unsigned pitch,
         unsigned imageWidth, unsigned imageHeight, unsigned allocationCols, unsigned allocationRows,
         unsigned timestamp, bool eos=false): GPUFrame(imageWidth, imageHeight, allocationCols, allocationRows, timestamp)
    {
        // copy into a more permanent chunk of memory allocated by above ctor
        cudaErr(cudaMemcpy2D(data(), m_pitch, reinterpret_cast<void*>(devPtr), pitch, allocationCols, allocationRows, cudaMemcpyDeviceToDevice));
    }
}

The output I keep getting:

imageWidth = 1920, m_width = 0

I'm confused why m_width would even be 0, that doesn't even seem like an option. Does anyone have any clue as to what I'm doing wrong?

FWIW, I'm compiling with g++-5 using the --std=gnu++11 option. The full code is available at https://github.com/briantilley/computer-vision.

Upvotes: 1

Views: 112

Answers (2)

Damyan
Damyan

Reputation: 1613

This:

cudaMallocPitch(&newAllocation, reinterpret_cast<size_t*>(&m_pitch), static_cast<size_t>(allocationCols), static_cast<size_t>(allocationRows));

Will cause the pitch to be written to m_pitch. This expects to write a size_t's worth of data to &m_pitch.

m_pitch is declared as "unsigned". This is not necessarily the same size as a size_t. If you output:

std::cout << sizeof(unsigned) << "\n" << sizeof(size_t)

Then I expect that you'll see "4" and "8".

So cudaMallocPitch will write 8 bytes, starting at &m_pitch. This will overwrite the next field, which is m_width.

Changing the type of m_pitch to size_t should solve this. You should also be able to remove the reinterpret_cast as well.

Upvotes: 4

T.C.
T.C.

Reputation: 137330

unsigned is typically 4 bytes. std::size_t, on a 64-bit system, is typically 8 bytes.1

cudaMallocPitch writes to the thing pointed to by its second argument assuming that it's a size_t, but m_pitch is an unsigned, so it ends up clobbering m_width.

1 Usual disclaimers about unicorn systems apply.

Upvotes: 2

Related Questions