#include <string>
#include "gromacs/gpu_utils/gmxopencl.h"
#include "gromacs/gpu_utils/gputraits_ocl.h"
#include "gromacs/utility/exceptions.h"
#include "gromacs/utility/gmxassert.h"

Include dependency graph for oclutils.h:

This graph shows which files directly or indirectly include this file:

Description

Declare utility routines for OpenCL.

Author: Anca Hamuraru anca@.nosp@m.stre.nosp@m.amcom.nosp@m.puti.nosp@m.ng.eu

Enumerations
enum	ocl_vendor_id_t { OCL_VENDOR_NVIDIA = 0, OCL_VENDOR_AMD, OCL_VENDOR_INTEL, OCL_VENDOR_UNKNOWN }
	OpenCL vendor IDs.

Functions
int	ocl_copy_D2H (void h_dest, cl_mem d_src, size_t offset, size_t bytes, GpuApiCallBehavior transferKind, cl_command_queue command_queue, cl_event copy_event)
	Launches synchronous or asynchronous device to host memory copy. More...

int	ocl_copy_D2H_async (void h_dest, cl_mem d_src, size_t offset, size_t bytes, cl_command_queue command_queue, cl_event copy_event)
	Launches asynchronous device to host memory copy. More...

int	ocl_copy_H2D (cl_mem d_dest, void h_src, size_t offset, size_t bytes, GpuApiCallBehavior transferKind, cl_command_queue command_queue, cl_event copy_event)
	Launches synchronous or asynchronous host to device memory copy. More...

int	ocl_copy_H2D_async (cl_mem d_dest, void h_src, size_t offset, size_t bytes, cl_command_queue command_queue, cl_event copy_event)
	Launches asynchronous host to device memory copy. More...

int	ocl_copy_H2D_sync (cl_mem d_dest, void *h_src, size_t offset, size_t bytes, cl_command_queue command_queue)
	Launches synchronous host to device memory copy.

void	pmalloc (void **h_ptr, size_t nbytes)
	Allocate host memory in malloc style. More...

void	pfree (void *h_ptr)
	Free host memory in malloc style. More...

std::string	ocl_get_error_string (cl_int error)
	Convert error code to diagnostic string.

static void	gpuStreamSynchronize (cl_command_queue s)
	Calls clFinish() in the stream `s`. More...

void	ensureReferenceCount (const cl_event &event, unsigned int refCount)
	A debug checker to track cl_events being released correctly.

static bool	haveStreamTasksCompleted (cl_command_queue s)
	Pretend to synchronize an OpenCL stream (dummy implementation). More...

void	prepareGpuKernelArgument (cl_kernel kernel, const KernelLaunchConfig &config, size_t argIndex)
	A function for setting up a single OpenCL kernel argument. This is the tail of the compile-time recursive function below. It has to be seen by the compiler first. As NB kernels might be using dynamic local memory as the last argument, this function also manages that, using sharedMemorySize from `config`. More...

template<typename CurrentArg , typename... RemainingArgs>
void	prepareGpuKernelArgument (cl_kernel kernel, const KernelLaunchConfig &config, size_t argIndex, const CurrentArg argPtr, const RemainingArgs ...otherArgsPtrs)
	Compile-time recursive function for setting up a single OpenCL kernel argument. This function uses one kernel argument pointer `argPtr` to call clSetKernelArg(), and calls itself on the next argument, eventually calling the tail function above. More...

template<typename... Args>
void *	prepareGpuKernelArguments (cl_kernel kernel, const KernelLaunchConfig &config, const Args *...argsPtrs)
	A wrapper function for setting up all the OpenCL kernel arguments. Calls the recursive functions above. More...

void	launchGpuKernel (cl_kernel kernel, const KernelLaunchConfig &config, CommandEvent timingEvent, const char kernelName, const void *)
	Launches the OpenCL kernel and handles the errors. More...

Function Documentation

static void gpuStreamSynchronize ( cl_command_queue s )

inlinestatic

Calls clFinish() in the stream s.

Parameters

[in] s stream to synchronize with

static bool haveStreamTasksCompleted ( cl_command_queue s )

inlinestatic

Pretend to synchronize an OpenCL stream (dummy implementation).

Parameters

[in] s queue to check

Returns: True if all tasks enqueued in the stream s (at the time of this call) have completed.

void launchGpuKernel	(	cl_kernel	kernel,
		const KernelLaunchConfig &	config,
		CommandEvent *	timingEvent,
		const char *	kernelName,
		const void *
	)

inline

Launches the OpenCL kernel and handles the errors.

Parameters

[in]	kernel	Kernel function handle
[in]	config	Kernel configuration for launching
[in]	timingEvent	Timing event, fetched from GpuRegionTimer
[in]	kernelName	Human readable kernel description, for error handling only

Exceptions

gmx::InternalError on kernel launch failure

int ocl_copy_D2H	(	void *	h_dest,
		cl_mem	d_src,
		size_t	offset,
		size_t	bytes,
		GpuApiCallBehavior	transferKind,
		cl_command_queue	command_queue,
		cl_event *	copy_event
	)

Launches synchronous or asynchronous device to host memory copy.

If copy_event is not NULL, on return it will contain an event object identifying this particular device to host operation. The event can further be used to queue a wait for this operation or to query profiling information.

int ocl_copy_D2H_async	(	void *	h_dest,
		cl_mem	d_src,
		size_t	offset,
		size_t	bytes,
		cl_command_queue	command_queue,
		cl_event *	copy_event
	)

Launches asynchronous device to host memory copy.

If copy_event is not nullptr, on return it will contain an event object identifying this particular host to device operation. The event can further be used to queue a wait for this operation or to query profiling information.

int ocl_copy_H2D	(	cl_mem	d_dest,
		void *	h_src,
		size_t	offset,
		size_t	bytes,
		GpuApiCallBehavior	transferKind,
		cl_command_queue	command_queue,
		cl_event *	copy_event
	)

Launches synchronous or asynchronous host to device memory copy.

If copy_event is not NULL, on return it will contain an event object identifying this particular host to device operation. The event can further be used to queue a wait for this operation or to query profiling information.

int ocl_copy_H2D_async	(	cl_mem	d_dest,
		void *	h_src,
		size_t	offset,
		size_t	bytes,
		cl_command_queue	command_queue,
		cl_event *	copy_event
	)

Launches asynchronous host to device memory copy.

If copy_event is not nullptr, on return it will contain an event object identifying this particular host to device operation. The event can further be used to queue a wait for this operation or to query profiling information.

void pfree ( void * h_ptr )

Free host memory in malloc style.

Parameters

[in] h_ptr Buffer allocated with pmalloc that needs to be freed.

void pmalloc	(	void **	h_ptr,
		size_t	nbytes
	)

Allocate host memory in malloc style.

Todo:: This function should allocate page-locked memory to help reduce D2H and H2D transfer times, similar with pmalloc from pmalloc_cuda.cu.

Parameters

[in,out]	h_ptr	Pointer where to store the address of the newly allocated buffer.
[in]	nbytes	Size in bytes of the buffer to be allocated.

void prepareGpuKernelArgument	(	cl_kernel	kernel,
		const KernelLaunchConfig &	config,
		size_t	argIndex
	)

inline

A function for setting up a single OpenCL kernel argument. This is the tail of the compile-time recursive function below. It has to be seen by the compiler first. As NB kernels might be using dynamic local memory as the last argument, this function also manages that, using sharedMemorySize from config.

Parameters

[in]	kernel	Kernel function handle
[in]	config	Kernel configuration for launching
[in]	argIndex	Index of the current argument

template<typename CurrentArg , typename... RemainingArgs>

void prepareGpuKernelArgument	(	cl_kernel	kernel,
		const KernelLaunchConfig &	config,
		size_t	argIndex,
		const CurrentArg *	argPtr,
		const RemainingArgs *...	otherArgsPtrs
	)

Compile-time recursive function for setting up a single OpenCL kernel argument. This function uses one kernel argument pointer argPtr to call clSetKernelArg(), and calls itself on the next argument, eventually calling the tail function above.

Template Parameters

CurrentArg	Type of the current argument
RemainingArgs	Types of remaining arguments after the current one

Parameters

[in]	kernel	Kernel function handle
[in]	config	Kernel configuration for launching
[in]	argIndex	Index of the current argument
[in]	argPtr	Pointer to the current argument
[in]	otherArgsPtrs	Pack of pointers to arguments remaining to process after the current one

template<typename... Args>

void* prepareGpuKernelArguments	(	cl_kernel	kernel,
		const KernelLaunchConfig &	config,
		const Args *...	argsPtrs
	)

A wrapper function for setting up all the OpenCL kernel arguments. Calls the recursive functions above.

Template Parameters

Args	Types of all the kernel arguments

Parameters

[in]	kernel	Kernel function handle
[in]	config	Kernel configuration for launching
[in]	argsPtrs	Pointers to all the kernel arguments

Returns

A handle for the prepared parameter pack to be used with launchGpuKernel() as the last argument

currently always nullptr for OpenCL, as it manages kernel/arguments association by itself.

Description

Enumerations

Functions

Function Documentation