CNTK/Source/Math/cudabasetypes.h

// cudabasetypes.h -- basic types used on both CUDA and PC side
//
// F. Seide, V-hansu

#pragma once

#ifdef __CUDA_ARCH__ // we are compiling under CUDA
#define ON_CUDA 1
#ifdef __device__
#define cudacode __device__
#define cudasharedcode __device__ __host__
#else
#define cudacode
#define cudasharedcode
#endif
#else
#define ON_CUDA 0 // TODO: this does not work for some combination--fix this
#ifdef __device__
#define cudacode __device__
#define cudasharedcode __device__ __host__
#else
#define cudacode
#define cudasharedcode
#endif
#endif

#include <assert.h>

namespace msra { namespace cuda {

typedef size_t cuda_size_t; // TODO: verify if this is consistent across CPU/CUDA, or use uint32 or so

// we wrap CUDA pointers so that we don't accidentally use them in CPU code
template <typename T>
class cuda_ptr
{
    T* p; // CUDA pointers are the same as host (e.g. Win32 is restricted to 32-bit CUDA pointers)
public:
    void swap(cuda_ptr& other)
    {
        T* tmp = p;
        p = other.p;
        other.p = tmp;
    }
    cudacode T& operator[](size_t index)
    {
        return p[index];
    }
    cudacode const T& operator[](size_t index) const
    {
        return p[index];
    }
    cudasharedcode cuda_ptr operator+(size_t index) const
    {
        return cuda_ptr(p + index);
    }
    cudasharedcode cuda_ptr operator-(size_t index) const
    {
        return cuda_ptr(p - index);
    }
    cuda_ptr(T* pp)
        : p(pp)
    {
    }
    T* get() const
    {
        return p;
    }
};

// reference to a vector (without allocation) that lives in CUDA RAM
// This can be directly passed by value to CUDA functions.
template <typename T>
class vectorref
{
    cuda_ptr<T> p; // pointer in CUDA space of this device
    cuda_size_t n; // number of elements
public:
    cudasharedcode size_t size() const throw()
    {
        return n;
    }
    cudacode T& operator[](size_t i)
    {
        return p[i];
    }
    cudacode const T& operator[](size_t i) const
    {
        return p[i];
    }
    cuda_ptr<T> get() const throw()
    {
        return p;
    }
    cuda_ptr<T> reset(cuda_ptr<T> pp, size_t nn) throw()
    {
        p.swap(pp);
        n = nn;
        return pp;
    }
    vectorref(cuda_ptr<T> pp, size_t nn)
        : p(pp), n(nn)
    {
    }
    vectorref()
        : p(0), n(0)
    {
    }
};

// reference to a matrix
template <typename T>
class matrixref
{
protected:
    cuda_ptr<T> p;    // pointer in CUDA space of this device
    size_t numrows;   // rows()
    size_t numcols;   // cols()
    size_t colstride; // height of column = rows() rounded to multiples of 4
    cudasharedcode size_t locate(size_t i, size_t j) const
    {
        return j * colstride + i;
    } // matrix in column-wise storage
    matrixref()
        : p(0), numrows(0), numcols(0), colstride(0)
    {
    }

public:
    matrixref(T* p, size_t numRows, size_t numCols, size_t colStride)
        : p(p), numrows(numRows), numcols(numCols), colstride(colStride)
    {
    }
    cuda_ptr<T> get() const throw()
    {
        return p;
    }
    cudasharedcode size_t rows() const throw()
    {
        return numrows;
    }
    cudasharedcode size_t cols() const throw()
    {
        return numcols;
    }
    cudasharedcode void reshape(const size_t newrows, const size_t newcols)
    {
        assert(rows() * cols() == newrows * newcols);
        numrows = newrows;
        numcols = newcols;
    };
    cudasharedcode size_t getcolstride() const throw()
    {
        return colstride;
    }
    cudacode T& operator()(size_t i, size_t j)
    {
        return p[locate(i, j)];
    }
    cudacode const T& operator()(size_t i, size_t j) const
    {
        return p[locate(i, j)];
    }
};
} }