#include <cstdio>
#include <string>
#include <vector>
#include "gromacs/gpu_utils/gpu_macros.h"
#include "gromacs/utility/basedefinitions.h"

Include dependency graph for gpu_utils.h:

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

Description

Declare functions for detection and initialization for GPU devices.

Author: Szilard Pall pall..nosp@m.szil.nosp@m.ard@g.nosp@m.mail.nosp@m..com; Mark Abraham mark..nosp@m.j.ab.nosp@m.raham.nosp@m.@gma.nosp@m.il.co.nosp@m.m

Typedefs
typedef void	gmx_host_alloc_t (void **ptr, size_t nbytes)
	Returns a pointer *ptr to page-locked memory of size nbytes. More...

typedef void	gmx_host_free_t (void *ptr)
	Frees page-locked memory pointed to by *ptr. More...

Enumerations
enum	GpuApiCallBehavior { Sync, Async }
	Enum which is only used to describe transfer calls at the moment.

enum	GpuTaskCompletion { Wait, Check }
	Types of actions associated to waiting or checking the completion of GPU tasks.

Functions
bool	canDetectGpus (std::string *errorMessage)
	Return whether GPUs can be detected. More...

void	findGpus (gmx_gpu_info_t *gpu_info)
	Find all GPUs in the system. More...

std::vector< int >	getCompatibleGpus (const gmx_gpu_info_t &gpu_info)
	Return a container of the detected GPUs that are compatible. More...

const char *	getGpuCompatibilityDescription (const gmx_gpu_info_t &gpu_info, int index)
	Return a string describing how compatible the GPU with given `index` is. More...

void	free_gpu_info (const gmx_gpu_info_t *gpu_info)
	Frees the gpu_dev and dev_use array fields of `gpu_info`. More...

void	init_gpu (const gmx_device_info_t *deviceInfo)
	Initializes the GPU described by `deviceInfo`. More...

void	free_gpu (const gmx_device_info_t *deviceInfo)
	Frees up the CUDA GPU used by the active context at the time of calling. More...

gmx_device_info_t *	getDeviceInfo (const gmx_gpu_info_t &gpu_info, int deviceId)
	Return a pointer to the device info for `deviceId`. More...

int	get_current_cuda_gpu_device_id (void)
	Returns the device ID of the CUDA GPU currently in use. More...

void	get_gpu_device_info_string (char *s, const gmx_gpu_info_t &gpu_info, int index)
	Formats and returns a device information string for a given GPU. More...

bool	areAllGpuDevicesFromAmd (const gmx_gpu_info_t &gpuInfo)
	Returns whether all compatible OpenCL devices are from AMD. More...

size_t	sizeof_gpu_dev_info ()
	Returns the size of the gpu_dev_info struct. More...

void	gpu_set_host_malloc_and_free (bool bUseGpuKernels, gmx_host_alloc_t nb_alloc, gmx_host_free_t nb_free)
	Set page-locked memory allocation functions used by the GPU host. More...

int	gpu_info_get_stat (const gmx_gpu_info_t &info, int index)
	Get status of device with specified index.

bool	buildSupportsNonbondedOnGpu (std::string *error)
	Check if GROMACS has been built with GPU support. More...

void	startGpuProfiler (void)
	Starts the GPU profiler if mdrun is being profiled. More...

void	resetGpuProfiler (void)
	Resets the GPU profiler if mdrun is being profiled. More...

void	stopGpuProfiler (void)
	Stops the CUDA profiler if mdrun is being profiled. More...

bool	isHostMemoryPinned (const void *h_ptr)
	Tells whether the host buffer was pinned for non-blocking transfers. Only implemented for CUDA.

Typedef Documentation

typedef void gmx_host_alloc_t(void **ptr, size_t nbytes)

Returns a pointer *ptr to page-locked memory of size nbytes.

The allocated memory is suitable to be used for data transfers between host and GPU. Error handling should be done within this function.

typedef void gmx_host_free_t(void *ptr)

Frees page-locked memory pointed to by *ptr.

NULL should not be passed to this function.

Function Documentation

bool areAllGpuDevicesFromAmd ( const gmx_gpu_info_t & gpuInfo )

Returns whether all compatible OpenCL devices are from AMD.

This is currently the most useful and best tested platform for supported OpenCL devices, so some modules may need to check what degree of support they should offer.

Todo:: An enumeration visible in the hardware module would make such checks more configurable, if we discover other needs in future.

Returns: whether all detected compatible devices have AMD for the vendor.

bool buildSupportsNonbondedOnGpu ( std::string * error )

Check if GROMACS has been built with GPU support.

Parameters

[in] error Pointer to error string or nullptr.

Todo:: Move this to NB module once it exists.

bool canDetectGpus ( std::string * errorMessage )

Return whether GPUs can be detected.

Returns true when this is a build of GROMACS configured to support GPU usage, and a valid device driver, ICD, and/or runtime was detected.

This function is not intended to be called from build configurations that do not support GPUs, and there will be no descriptive message in that case.

Parameters

[out] errorMessage When returning false on a build configured with GPU support and non-nullptr was passed, the string contains a descriptive message about why GPUs cannot be detected.

Does not throw.

Return whether GPUs can be detected.

void findGpus ( gmx_gpu_info_t * gpu_info )

Find all GPUs in the system.

Will detect every GPU supported by the device driver in use. Must only be called if canDetectGpus() has returned true. This routine also checks for the compatibility of each and fill the gpu_info->gpu_dev array with the required information on each the device: ID, device properties, status.

Note that this function leaves the GPU runtime API error state clean; this is implemented ATM in the CUDA flavor. TODO: check if errors do propagate in OpenCL as they do in CUDA and whether there is a mechanism to "clear" them.

Parameters

[in] gpu_info pointer to structure holding GPU information.

Exceptions

InternalError if a GPU API returns an unexpected failure (because the call to canDetectGpus() should always prevent this occuring)

Find all GPUs in the system.

void free_gpu ( const gmx_device_info_t * deviceInfo )

Frees up the CUDA GPU used by the active context at the time of calling.

If deviceInfo is nullptr, then it is understood that no device was selected so no context is active to be freed. Otherwise, the context is explicitly destroyed and therefore all data uploaded to the GPU is lost. This must only be called when none of this data is required anymore, because subsequent attempts to free memory associated with the context will otherwise fail.

Calls gmx_warning upon errors.

Parameters

[in] deviceInfo device info of the GPU to clean up for

Returns: true if no error occurs during the freeing.

void free_gpu_info ( const gmx_gpu_info_t * gpu_info )

Frees the gpu_dev and dev_use array fields of gpu_info.

Parameters

[in] gpu_info pointer to structure holding GPU information

int get_current_cuda_gpu_device_id ( void )

Returns the device ID of the CUDA GPU currently in use.

The GPU used is the one that is active at the time of the call in the active context.

Returns: device ID of the GPU in use at the time of the call

void get_gpu_device_info_string	(	char *	s,
		const gmx_gpu_info_t &	gpu_info,
		int	index
	)

Formats and returns a device information string for a given GPU.

Given an index directly into the array of available GPUs (gpu_dev) returns a formatted info string for the respective GPU which includes ID, name, compute capability, and detection status.

Parameters

[out]	s	pointer to output string (has to be allocated externally)
[in]	gpu_info	Information about detected GPUs
[in]	index	an index directly into the array of available GPUs

Formats and returns a device information string for a given GPU.

std::vector<int> getCompatibleGpus ( const gmx_gpu_info_t & gpu_info )

Return a container of the detected GPUs that are compatible.

This function filters the result of the detection for compatible GPUs, based on the previously run compatibility tests.

Parameters

[in] gpu_info Information detected about GPUs, including compatibility.

Returns: vector of IDs of GPUs already recorded as compatible

gmx_device_info_t* getDeviceInfo	(	const gmx_gpu_info_t &	gpu_info,
		int	deviceId
	)

Return a pointer to the device info for deviceId.

Parameters

[in]	gpu_info	GPU info of all detected devices in the system.
[in]	deviceId	ID for the GPU device requested.

Returns: Pointer to the device info for deviceId.

Return a pointer to the device info for deviceId.

const char* getGpuCompatibilityDescription	(	const gmx_gpu_info_t &	gpu_info,
		int	index
	)

Return a string describing how compatible the GPU with given index is.

Parameters

[in]	gpu_info	Information about detected GPUs
[in]	index	index of GPU to ask about

Returns: A null-terminated C string describing the compatibility status, useful for error messages.

void gpu_set_host_malloc_and_free	(	bool	,
		gmx_host_alloc_t **	nb_alloc,
		gmx_host_free_t **	nb_free
	)

Set page-locked memory allocation functions used by the GPU host.

Since GPU support is not configured, there is no host memory to allocate.

void init_gpu ( const gmx_device_info_t * deviceInfo )

Initializes the GPU described by deviceInfo.

TODO Doxygen complains about these - probably a Doxygen bug, since the patterns here are the same as elsewhere in this header.

Parameters

[in] deviceInfo device info of the GPU to initialize

Issues a fatal error for any critical errors that occur during initialization.

Initializes the GPU described by deviceInfo.

void resetGpuProfiler ( void )

Resets the GPU profiler if mdrun is being profiled.

When a profiler run is in progress (based on the presence of the NVPROF_ID env. var.), the profiler data is restet in order to eliminate the data collected from the preceding part fo the run.

This function should typically be called at the mdrun counter reset time.

Note that this is implemented only for the CUDA API.

size_t sizeof_gpu_dev_info ( )

Returns the size of the gpu_dev_info struct.

The size of gpu_dev_info can be used for allocation and communication.

Returns: size in bytes of gpu_dev_info

Returns the size of the gpu_dev_info struct.

void startGpuProfiler ( void )

Starts the GPU profiler if mdrun is being profiled.

When a profiler run is in progress (based on the presence of the NVPROF_ID env. var.), the profiler is started to begin collecting data during the rest of the run (or until stopGpuProfiler is called).

Note that this is implemented only for the CUDA API.

void stopGpuProfiler ( void )

Stops the CUDA profiler if mdrun is being profiled.

This function can be called at cleanup when skipping recording recording subsequent API calls from being traces/profiled is desired, e.g. before uninitialization.

Note that this is implemented only for the CUDA API.

Description

Typedefs

Enumerations

Functions

Typedef Documentation

Function Documentation