Namespace providing routines for handling the different memory domains. More...

Namespaces
	cpu_ram
	Provides implementations for handling memory buffers in CPU RAM.

	cuda
	Provides implementations for handling CUDA memory buffers.

	detail
	Implementation details for the generic memory backend interface.

	opencl
	Provides implementations for handling OpenCL memory buffers.

Classes
class	mem_handle
	Main abstraction class for multiple memory domains. Represents a buffer in either main RAM, an OpenCL context, or a CUDA device. More...

class	typesafe_host_array
	Helper class implementing an array on the host. Default case: No conversion necessary. More...

class	typesafe_host_array< T, true >
	Special host array type for conversion between OpenCL types and pure CPU types. More...

Functions
memory_types	default_memory_type ()
	Returns the default memory type for the given configuration. More...

memory_types	default_memory_type (memory_types new_memory_type)
	Sets the default memory type for the given configuration. More...

void	finish ()
	Synchronizes the execution. finish() will only return after all compute kernels (CUDA, OpenCL) have completed. More...

void	memory_create (mem_handle &handle, vcl_size_t size_in_bytes, viennacl::context const &ctx, const void *host_ptr=NULL)
	Creates an array of the specified size. If the second argument is provided, the buffer is initialized with data from that pointer. More...

void	memory_copy (mem_handle const &src_buffer, mem_handle &dst_buffer, vcl_size_t src_offset, vcl_size_t dst_offset, vcl_size_t bytes_to_copy)
	Copies 'bytes_to_copy' bytes from address 'src_buffer + src_offset' to memory starting at address 'dst_buffer + dst_offset'. More...

void	memory_shallow_copy (mem_handle const &src_buffer, mem_handle &dst_buffer)
	A 'shallow' copy operation from an initialized buffer to an uninitialized buffer. The uninitialized buffer just copies the raw handle. More...

void	memory_write (mem_handle &dst_buffer, vcl_size_t dst_offset, vcl_size_t bytes_to_write, const void *ptr, bool async=false)
	Writes data from main RAM identified by 'ptr' to the buffer identified by 'dst_buffer'. More...

void	memory_read (mem_handle const &src_buffer, vcl_size_t src_offset, vcl_size_t bytes_to_read, void *ptr, bool async=false)
	Reads data from a buffer back to main RAM. More...

template<typename DataType >
void	switch_memory_context (mem_handle &handle, viennacl::context new_ctx)
	Switches the active memory domain within a memory handle. Data is copied if the new active domain differs from the old one. Memory in the source handle is not free'd. More...

template<typename DataType >
void	typesafe_memory_copy (mem_handle const &handle_src, mem_handle &handle_dst)
	Copies data of the provided 'DataType' from 'handle_src' to 'handle_dst' and converts the data if the binary representation of 'DataType' among the memory domains differs. More...

Detailed Description

Namespace providing routines for handling the different memory domains.

Function Documentation

memory_types viennacl::backend::default_memory_type ( )

inline

Returns the default memory type for the given configuration.

CUDA has precedence over OpenCL, which has precedence over main memory. Depends on which VIENNACL_WITH_{CUDA/OPENCL/OPENMP} macros are defined.

Definition at line 73 of file mem_handle.hpp.

memory_types viennacl::backend::default_memory_type ( memory_types new_memory_type )

inline

Sets the default memory type for the given configuration.

Make sure the respective new memory type is enabled. For example, passing CUDA_MEMORY if no CUDA backend is selected will result in exceptions being thrown as soon as you try to allocate buffers.

Definition at line 80 of file mem_handle.hpp.

void viennacl::backend::finish ( )

inline

Synchronizes the execution. finish() will only return after all compute kernels (CUDA, OpenCL) have completed.

Examples:: amg.cpp, and blas3.cpp.

Definition at line 54 of file memory.hpp.

void viennacl::backend::memory_copy	(	mem_handle const &	src_buffer,
		mem_handle &	dst_buffer,
		vcl_size_t	src_offset,
		vcl_size_t	dst_offset,
		vcl_size_t	bytes_to_copy
	)

inline

Copies 'bytes_to_copy' bytes from address 'src_buffer + src_offset' to memory starting at address 'dst_buffer + dst_offset'.

This is the generic version for CPU RAM, CUDA, and OpenCL. Copies the memory in the currently active memory domain.

Parameters

src_buffer	A smart pointer to the begin of an allocated buffer
dst_buffer	A smart pointer to the end of an allocated buffer
src_offset	Offset of the first byte to be written from the address given by 'src_buffer' (in bytes)
dst_offset	Offset of the first byte to be written to the address given by 'dst_buffer' (in bytes)
bytes_to_copy	Number of bytes to be copied

Definition at line 140 of file memory.hpp.

void viennacl::backend::memory_create	(	mem_handle &	handle,
		vcl_size_t	size_in_bytes,
		viennacl::context const &	ctx,
		const void *	host_ptr = `NULL`
	)

inline

Creates an array of the specified size. If the second argument is provided, the buffer is initialized with data from that pointer.

This is the generic version for CPU RAM, CUDA, and OpenCL. Creates the memory in the currently active memory domain.

Parameters

handle	The generic wrapper handle for multiple memory domains which will hold the new buffer.
size_in_bytes	Number of bytes to allocate
ctx	Optional context in which the matrix is created (one out of multiple OpenCL contexts, CUDA, host)
host_ptr	Pointer to data which will be copied to the new array. Must point to at least 'size_in_bytes' bytes of data.

Definition at line 87 of file memory.hpp.

void viennacl::backend::memory_read	(	mem_handle const &	src_buffer,
		vcl_size_t	src_offset,
		vcl_size_t	bytes_to_read,
		void *	ptr,
		bool	async = `false`
	)

inline

Reads data from a buffer back to main RAM.

This is the generic version for CPU RAM, CUDA, and OpenCL. Reads the memory from the currently active memory domain.

Parameters

src_buffer	A smart pointer to the beginning of an allocated source buffer
src_offset	Offset of the first byte to be read from the beginning of src_buffer (in bytes_
bytes_to_read	Number of bytes to be read
ptr	Location in main RAM where to read data should be written to
async	Whether the operation should be asynchronous

Definition at line 261 of file memory.hpp.

void viennacl::backend::memory_shallow_copy	(	mem_handle const &	src_buffer,
		mem_handle &	dst_buffer
	)

inline

A 'shallow' copy operation from an initialized buffer to an uninitialized buffer. The uninitialized buffer just copies the raw handle.

Definition at line 177 of file memory.hpp.

void viennacl::backend::memory_write	(	mem_handle &	dst_buffer,
		vcl_size_t	dst_offset,
		vcl_size_t	bytes_to_write,
		const void *	ptr,
		bool	async = `false`
	)

inline

Writes data from main RAM identified by 'ptr' to the buffer identified by 'dst_buffer'.

This is the generic version for CPU RAM, CUDA, and OpenCL. Writes the memory in the currently active memory domain.

Parameters

dst_buffer	A smart pointer to the beginning of an allocated buffer
dst_offset	Offset of the first written byte from the beginning of 'dst_buffer' (in bytes)
bytes_to_write	Number of bytes to be written
ptr	Pointer to the first byte to be written
async	Whether the operation should be asynchronous

Definition at line 220 of file memory.hpp.

template<typename DataType >

void viennacl::backend::switch_memory_context	(	mem_handle &	handle,
		viennacl::context	new_ctx
	)

Switches the active memory domain within a memory handle. Data is copied if the new active domain differs from the old one. Memory in the source handle is not free'd.

Definition at line 368 of file memory.hpp.

template<typename DataType >

void viennacl::backend::typesafe_memory_copy	(	mem_handle const &	handle_src,
		mem_handle &	handle_dst
	)

Copies data of the provided 'DataType' from 'handle_src' to 'handle_dst' and converts the data if the binary representation of 'DataType' among the memory domains differs.

Definition at line 470 of file memory.hpp.

Namespaces

Classes

Functions

Detailed Description

Function Documentation