#include <backend.h>

Inheritance diagram for CodeGenerator::CUDA::Preferences:

Public Member Functions
	Preferences ()

void	updateHash (boost::uuids::detail::sha1 &hash) const

Public Member Functions inherited from CodeGenerator::PreferencesBase
void	updateHash (boost::uuids::detail::sha1 &hash) const

Public Attributes
bool	showPtxInfo = false
	Should PTX assembler information be displayed for each CUDA kernel during compilation? More...

bool	generateLineInfo = false
	Should line info be included in resultant executable for debugging/profiling purposes? More...

bool	selectGPUByDeviceID = false

bool	enableNCCLReductions = false
	Generate corresponding NCCL batch reductions. More...

bool	generateSimpleErrorHandling = false

DeviceSelect	deviceSelectMethod = DeviceSelect::OPTIMAL
	How to select GPU device. More...

unsigned int	manualDeviceID = 0
	If device select method is set to DeviceSelect::MANUAL, id of device to use. More...

BlockSizeSelect	blockSizeSelectMethod = BlockSizeSelect::OCCUPANCY
	How to select CUDA blocksize. More...

KernelBlockSize	manualBlockSizes
	If block size select method is set to BlockSizeSelect::MANUAL, block size to use for each kernel. More...

size_t	constantCacheOverhead = 72 * 5
	How much constant cache is already used and therefore can't be used by GeNN? More...

std::string	userNvccFlags = ""
	NVCC compiler options for all GPU code. More...

Public Attributes inherited from CodeGenerator::PreferencesBase
bool	optimizeCode = false
	Generate speed-optimized code, potentially at the expense of floating-point accuracy. More...

bool	debugCode = false
	Generate code with debug symbols. More...

bool	enableBitmaskOptimisations = false

bool	automaticCopy = false
	If backend/device supports it, copy data automatically when required rather than requiring push and pull. More...

bool	generateEmptyStatePushPull = true
	Should GeNN generate empty state push and pull functions. More...

bool	generateExtraGlobalParamPull = true
	Should GeNN generate pull functions for extra global parameters? These are very rarely used. More...

bool	includeModelNameInDLL = false

std::string	userCxxFlagsGNU = ""
	C++ compiler options to be used for building all host side code (used for unix based platforms) More...

std::string	userNvccFlagsGNU = ""
	NVCC compiler options they may want to use for all GPU code (used for unix based platforms) More...

plog::Severity	logLevel = plog::info
	Logging level to use for code generation. More...

Detailed Description

Preferences for CUDA backend.

Constructor & Destructor Documentation

◆ Preferences()

CodeGenerator::CUDA::Preferences::Preferences ( )

inline

Member Function Documentation

◆ updateHash()

void CodeGenerator::CUDA::Preferences::updateHash ( boost::uuids::detail::sha1 & hash ) const

inline

Update hash with preferences

Member Data Documentation

◆ blockSizeSelectMethod

BlockSizeSelect CodeGenerator::CUDA::Preferences::blockSizeSelectMethod = BlockSizeSelect::OCCUPANCY

How to select CUDA blocksize.

◆ constantCacheOverhead

size_t CodeGenerator::CUDA::Preferences::constantCacheOverhead = 72 * 5

How much constant cache is already used and therefore can't be used by GeNN?

Each of the four modules which includes CUDA headers(neuronUpdate, synapseUpdate, custom update, init and runner) Takes 72 bytes of constant memory for a lookup table used by cuRAND. If your application requires additional constant cache, increase this

◆ deviceSelectMethod

DeviceSelect CodeGenerator::CUDA::Preferences::deviceSelectMethod = DeviceSelect::OPTIMAL

How to select GPU device.

◆ enableNCCLReductions

bool CodeGenerator::CUDA::Preferences::enableNCCLReductions = false

Generate corresponding NCCL batch reductions.

◆ generateLineInfo

bool CodeGenerator::CUDA::Preferences::generateLineInfo = false

Should line info be included in resultant executable for debugging/profiling purposes?

◆ generateSimpleErrorHandling

bool CodeGenerator::CUDA::Preferences::generateSimpleErrorHandling = false

By default, GeNN generates CUDA error-handling code that generates exceptions with line numbers etc. This can make compilation of large models very slow and require a large amount of memory. Turning on this option generates much simpler error-handling code that simply raises an abort signal.

◆ manualBlockSizes

KernelBlockSize CodeGenerator::CUDA::Preferences::manualBlockSizes

If block size select method is set to BlockSizeSelect::MANUAL, block size to use for each kernel.

◆ manualDeviceID

unsigned int CodeGenerator::CUDA::Preferences::manualDeviceID = 0

If device select method is set to DeviceSelect::MANUAL, id of device to use.

◆ selectGPUByDeviceID

bool CodeGenerator::CUDA::Preferences::selectGPUByDeviceID = false

GeNN normally identifies devices by PCI bus ID to ensure that the model is run on the same device it was optimized for. However if, for example, you are running on a cluser with NVML this is not desired behaviour.

◆ showPtxInfo

bool CodeGenerator::CUDA::Preferences::showPtxInfo = false

Should PTX assembler information be displayed for each CUDA kernel during compilation?

◆ userNvccFlags

std::string CodeGenerator::CUDA::Preferences::userNvccFlags = ""

NVCC compiler options for all GPU code.

The documentation for this struct was generated from the following file:

cuda/backend.h

Public Member Functions

Public Attributes