API Support

Tools to help implement irritating C-style APIs. More...

Modules
	Enum Validation
	A general mechanism for checking if an enum value is in range.

Classes
class	sp::Context
	A class to hold library-global state. More...
struct	sp::PtrScalar< T, Location >
	Represents a scalar value that has been passed as a pointer. More...
struct	sp::is_ptr_scalar< typename >
	Type-trait for detecting if a given type is a PtrScalar of some sort. More...
class	sp::IndirectScalar< T >
	A self-dereferencing pointer. More...
struct	sp::is_indirect_scalar< typename >
	Type-trait for detecting if a given type is a PtrScalar of some sort. More...
struct	sp::CudaVariantScalar< Location >
	Similar to a PtrScalar, but with the added quirk that the type of the scalar is not known at compile-time. More...
struct	sp::is_variant_scalar< typename >
	Trait to detect if something is a CudaVariantScalar. More...
struct	sp::VariantOutputPtr
	An output pointer represented as a void* and a cudaDataType_t. More...
struct	sp::CudaVec< T, Size >
	Provides a type mapping between CUDA vector types and sp::Vec. More...
struct	sp::TensorDescriptor< Type, Rank >
	A fully dynamic entity that can be converted into a proper Tensor at runtime. More...

Macros
#define	CUDA_VEC_E(TYPE, SIZE, CUDATYPE)
#define	CUDA_VEC(TYPE, SIZE)   CUDA_VEC_E(TYPE, SIZE, TYPE ## SIZE)

Typedefs
template<typename T , int Size>
using	sp::CudaVec_t = typename CudaVec< std::remove_cv_t< T >, Size >::type
	The CUDA vector type representing a vector of type T and length Size. Such as float4 or int2. More...
template<typename T , int Size>
using	sp::ThinCudaVec_t = typename CudaVec< std::remove_cv_t< T >, Size >::thinType
	Like CudaVec_t, except that if Size is 1 this is equal to T. More...

Enumerations
enum class	sp::PtrLocation { HOST , DEVICE , UNKNOWN }
	Represents which device a pointer resides on, if known. More...

Functions
template<typename T >
void	sp::outputToAddress (sp::Stream &s, __flat T *outputAddress, __device const T *stagingAddress)
	Conditionally copy a result back to the host. More...
template<typename T >
	sp::PtrScalar (__device const T *) -> PtrScalar< T, PtrLocation::DEVICE >
template<typename T >
cudaDataType_t	sp::asCudaType ()
	Convert a C++ type to a CUDA type enumeration value. More...

Detailed Description

Tools to help implement irritating C-style APIs.

This is helpful when:

• Writing bindings to other programming or scripting languages.
• Creating drop-in replacements for older CUDA libraries.
• Integrating a thirdparty library with an annoying API while wrapping it with modern features like RAII.

Macro Definition Documentation

CUDA_VEC_E

#define CUDA_VEC_E	(		TYPE,
			SIZE,
			CUDATYPE
	)

Value:

template<> \
struct CudaVec<TYPE, SIZE> { \
    using type = CUDATYPE; \
    using thinType = std::conditional_t<SIZE == 1, TYPE, CUDATYPE>; \
};

Typedef Documentation

CudaVec_t

template<typename T , int Size>

using sp::CudaVec_t = typedef typename CudaVec<std::remove_cv_t<T>, Size>::type

The CUDA vector type representing a vector of type T and length Size. Such as float4 or int2.

ThinCudaVec_t

template<typename T , int Size>

using sp::ThinCudaVec_t = typedef typename CudaVec<std::remove_cv_t<T>, Size>::thinType

Like CudaVec_t, except that if Size is 1 this is equal to T.

Enumeration Type Documentation

PtrLocation

enum class sp::PtrLocation

strong

Represents which device a pointer resides on, if known.

Function Documentation

asCudaType()

template<typename T >

cudaDataType_t sp::asCudaType

(

)

Convert a C++ type to a CUDA type enumeration value.

outputToAddress()

template<typename T >

void sp::outputToAddress	(	sp::Stream &	s,
		__flat T *	outputAddress,
		__device const T *	stagingAddress
	)

Conditionally copy a result back to the host.

Some APIs allow the output address to be either a host or device pointer. Kernels always produce results on the GPU, so we need to conditionally copy the result back to the device.

If they are the same address, nothing happens. Otherwise a copy (potentially between devices) occurs. This allows values to be returned to the host, or to other devices, after a library call.

Parameters

s	The stream to enqueue any copies on.
outputAddress	The address the caller wants the result to end up at.
stagingAddress	The GPU address where the result is due to be written.