Alien-XGBoost
view release on metacpan or search on metacpan
xgboost/cub/cub/iterator/tex_ref_input_iterator.cuh view on Meta::CPAN
*/
/**
* \brief A random-access input wrapper for dereferencing array values through texture cache. Uses older Tesla/Fermi-style texture references.
*
* \par Overview
* - TexRefInputIteratorTwraps a native device pointer of type <tt>ValueType*</tt>. References
* to elements are to be loaded through texture cache.
* - Can be used to load any data type from memory through texture cache.
* - Can be manipulated and exchanged within and between host and device
* functions, can only be constructed within host functions, and can only be
* dereferenced within device functions.
* - The \p UNIQUE_ID template parameter is used to statically name the underlying texture
* reference. Only one TexRefInputIteratorTinstance can be bound at any given time for a
* specific combination of (1) data type \p T, (2) \p UNIQUE_ID, (3) host
* thread, and (4) compilation .o unit.
* - With regard to nested/dynamic parallelism, TexRefInputIteratorTiterators may only be
* created by the host thread and used by a top-level kernel (i.e. the one which is launched
* from the host).
* - Compatible with Thrust API v1.7 or newer.
* - Compatible with CUDA toolkit v5.5 or newer.
*
* \par Snippet
* The code snippet below illustrates the use of \p TexRefInputIteratorTto
* dereference a device array of doubles through texture cache.
* \par
* \code
* #include <cub/cub.cuh> // or equivalently <cub/iterator/tex_ref_input_iterator.cuh>
*
* // Declare, allocate, and initialize a device array
* int num_items; // e.g., 7
* double *d_in; // e.g., [8.0, 6.0, 7.0, 5.0, 3.0, 0.0, 9.0]
*
* // Create an iterator wrapper
* cub::TexRefInputIterator<double, __LINE__> itr;
* itr.BindTexture(d_in, sizeof(double) * num_items);
* ...
*
* // Within device code:
* printf("%f\n", itr[0]); // 8.0
* printf("%f\n", itr[1]); // 6.0
* printf("%f\n", itr[6]); // 9.0
*
* ...
* itr.UnbindTexture();
*
* \endcode
*
* \tparam T The value type of this iterator
* \tparam UNIQUE_ID A globally-unique identifier (within the compilation unit) to name the underlying texture reference
* \tparam OffsetT The difference type of this iterator (Default: \p ptrdiff_t)
*/
template <
typename T,
int UNIQUE_ID,
typename OffsetT = ptrdiff_t>
class TexRefInputIterator
{
public:
// Required iterator traits
typedef TexRefInputIterator self_type; ///< My own type
typedef OffsetT difference_type; ///< Type to express the result of subtracting one iterator from another
typedef T value_type; ///< The type of the element the iterator can point to
typedef T* pointer; ///< The type of a pointer to an element the iterator can point to
typedef T reference; ///< The type of a reference to an element the iterator can point to
#if (THRUST_VERSION >= 100700)
// Use Thrust's iterator categories so we can use these iterators in Thrust 1.7 (or newer) methods
typedef typename thrust::detail::iterator_facade_category<
thrust::device_system_tag,
thrust::random_access_traversal_tag,
value_type,
reference
>::type iterator_category; ///< The iterator category
#else
typedef std::random_access_iterator_tag iterator_category; ///< The iterator category
#endif // THRUST_VERSION
private:
T* ptr;
difference_type tex_offset;
// Texture reference wrapper (old Tesla/Fermi-style textures)
typedef typename IteratorTexRef<T>::template TexId<UNIQUE_ID> TexId;
public:
/*
/// Constructor
__host__ __device__ __forceinline__ TexRefInputIterator()
:
ptr(NULL),
tex_offset(0)
{}
*/
/// Use this iterator to bind \p ptr with a texture reference
template <typename QualifiedT>
cudaError_t BindTexture(
QualifiedT *ptr, ///< Native pointer to wrap that is aligned to cudaDeviceProp::textureAlignment
size_t bytes = size_t(-1), ///< Number of bytes in the range
size_t tex_offset = 0) ///< OffsetT (in items) from \p ptr denoting the position of the iterator
{
this->ptr = const_cast<typename RemoveQualifiers<QualifiedT>::Type *>(ptr);
size_t offset;
cudaError_t retval = TexId::BindTexture(this->ptr + tex_offset, offset);
this->tex_offset = (difference_type) (offset / sizeof(QualifiedT));
return retval;
}
/// Unbind this iterator from its texture reference
cudaError_t UnbindTexture()
{
return TexId::UnbindTexture();
}
/// Postfix increment
__host__ __device__ __forceinline__ self_type operator++(int)
{
self_type retval = *this;
tex_offset++;
return retval;
}
/// Prefix increment
__host__ __device__ __forceinline__ self_type operator++()
{
tex_offset++;
return *this;
}
/// Indirection
__host__ __device__ __forceinline__ reference operator*() const
{
#if (CUB_PTX_ARCH == 0)
// Simply dereference the pointer on the host
return ptr[tex_offset];
#else
// Use the texture reference
return TexId::Fetch(tex_offset);
#endif
}
/// Addition
template <typename Distance>
__host__ __device__ __forceinline__ self_type operator+(Distance n) const
{
self_type retval;
( run in 1.112 second using v1.01-cache-2.11-cpan-d7f47b0818f )