#include <backend.h>

Inheritance diagram for CodeGenerator::CUDA::Backend:

Public Member Functions
	Backend (const KernelBlockSize &kernelBlockSizes, const Preferences &preferences, const std::string &scalarType, int device)

virtual bool	areSharedMemAtomicsSlow () const override
	On some older devices, shared memory atomics are actually slower than global memory atomics so should be avoided. More...

virtual std::string	getSharedPrefix () const override
	Get the prefix to use for shared memory variables. More...

virtual std::string	getThreadID (unsigned int axis=0) const override
	Get the ID of the current thread within the threadblock. More...

virtual std::string	getBlockID (unsigned int axis=0) const override
	Get the ID of the current thread block. More...

virtual std::string	getCLZ () const override
	Get the name of the count-leading-zeros function. More...

virtual std::string	getAtomic (const std::string &type, AtomicOperation op=AtomicOperation::ADD, AtomicMemSpace memSpace=AtomicMemSpace::GLOBAL) const override
	Get name of atomic operation. More...

virtual void	genSharedMemBarrier (CodeStream &os) const override
	Generate a shared memory barrier. More...

virtual void	genPopulationRNGInit (CodeStream &os, const std::string &globalRNG, const std::string &seed, const std::string &sequence) const override
	For SIMT backends which initialize RNGs on device, initialize population RNG with specified seed and sequence. More...

virtual void	genPopulationRNGPreamble (CodeStream &os, Substitutions &subs, const std::string &globalRNG, const std::string &name="rng") const override
	Generate a preamble to add substitution name for population RNG. More...

virtual void	genPopulationRNGPostamble (CodeStream &os, const std::string &globalRNG) const override
	If required, generate a postamble for population RNG. More...

virtual void	genGlobalRNGSkipAhead (CodeStream &os, Substitutions &subs, const std::string &sequence, const std::string &name="rng") const override
	Generate code to skip ahead local copy of global RNG. More...

virtual void	genNeuronUpdate (CodeStream &os, const ModelSpecMerged &modelMerged, HostHandler preambleHandler, HostHandler pushEGPHandler) const override
	Generate platform-specific function to update the state of all neurons. More...

virtual void	genSynapseUpdate (CodeStream &os, const ModelSpecMerged &modelMerged, HostHandler preambleHandler, HostHandler pushEGPHandler) const override
	Generate platform-specific function to update the state of all synapses. More...

virtual void	genCustomUpdate (CodeStream &os, const ModelSpecMerged &modelMerged, HostHandler preambleHandler, HostHandler pushEGPHandler) const override
	Generate platform-specific functions to perform custom updates. More...

virtual void	genInit (CodeStream &os, const ModelSpecMerged &modelMerged, HostHandler preambleHandler, HostHandler initPushEGPHandler, HostHandler initSparsePushEGPHandler) const override
	Generate platform-specific function to initialise model. More...

virtual void	genDefinitionsPreamble (CodeStream &os, const ModelSpecMerged &modelMerged) const override
	Definitions is the usercode-facing header file for the generated code. This function generates a 'preamble' to this header file. More...

virtual void	genDefinitionsInternalPreamble (CodeStream &os, const ModelSpecMerged &modelMerged) const override
	Definitions internal is the internal header file for the generated code. This function generates a 'preamble' to this header file. More...

virtual void	genRunnerPreamble (CodeStream &os, const ModelSpecMerged &modelMerged, const MemAlloc &memAlloc) const override

virtual void	genAllocateMemPreamble (CodeStream &os, const ModelSpecMerged &modelMerged, const MemAlloc &memAlloc) const override

virtual void	genFreeMemPreamble (CodeStream &os, const ModelSpecMerged &modelMerged) const override

virtual void	genStepTimeFinalisePreamble (CodeStream &os, const ModelSpecMerged &modelMerged) const override
	After all timestep logic is complete. More...

virtual void	genVariableDefinition (CodeStream &definitions, CodeStream &definitionsInternal, const std::string &type, const std::string &name, VarLocation loc) const override

virtual void	genVariableImplementation (CodeStream &os, const std::string &type, const std::string &name, VarLocation loc) const override

virtual void	genVariableAllocation (CodeStream &os, const std::string &type, const std::string &name, VarLocation loc, size_t count, MemAlloc &memAlloc) const override

virtual void	genVariableFree (CodeStream &os, const std::string &name, VarLocation loc) const override

virtual void	genExtraGlobalParamDefinition (CodeStream &definitions, CodeStream &definitionsInternal, const std::string &type, const std::string &name, VarLocation loc) const override

virtual void	genExtraGlobalParamImplementation (CodeStream &os, const std::string &type, const std::string &name, VarLocation loc) const override

virtual void	genExtraGlobalParamAllocation (CodeStream &os, const std::string &type, const std::string &name, VarLocation loc, const std::string &countVarName="count", const std::string &prefix="") const override

virtual void	genExtraGlobalParamPush (CodeStream &os, const std::string &type, const std::string &name, VarLocation loc, const std::string &countVarName="count", const std::string &prefix="") const override

virtual void	genExtraGlobalParamPull (CodeStream &os, const std::string &type, const std::string &name, VarLocation loc, const std::string &countVarName="count", const std::string &prefix="") const override

virtual void	genMergedExtraGlobalParamPush (CodeStream &os, const std::string &suffix, size_t mergedGroupIdx, const std::string &groupIdx, const std::string &fieldName, const std::string &egpName) const override
	Generate code for pushing an updated EGP value into the merged group structure on 'device'. More...

virtual std::string	getMergedGroupFieldHostType (const std::string &type) const override
	When generating function calls to push to merged groups, backend without equivalent of Unified Virtual Addressing e.g. OpenCL 1.2 may use different types on host. More...

virtual std::string	getMergedGroupSimRNGType () const override
	When generating merged structures what type to use for simulation RNGs. More...

virtual void	genVariablePush (CodeStream &os, const std::string &type, const std::string &name, VarLocation loc, bool autoInitialized, size_t count) const override
	Generate code for pushing a variable to the 'device'. More...

virtual void	genVariablePull (CodeStream &os, const std::string &type, const std::string &name, VarLocation loc, size_t count) const override
	Generate code for pulling a variable from the 'device'. More...

virtual void	genCurrentVariablePush (CodeStream &os, const NeuronGroupInternal &ng, const std::string &type, const std::string &name, VarLocation loc, unsigned int batchSize) const override
	Generate code for pushing a variable's value in the current timestep to the 'device'. More...

virtual void	genCurrentVariablePull (CodeStream &os, const NeuronGroupInternal &ng, const std::string &type, const std::string &name, VarLocation loc, unsigned int batchSize) const override
	Generate code for pulling a variable's value in the current timestep from the 'device'. More...

virtual void	genCurrentTrueSpikePush (CodeStream &os, const NeuronGroupInternal &ng, unsigned int batchSize) const override
	Generate code for pushing true spikes emitted by a neuron group in the current timestep to the 'device'. More...

virtual void	genCurrentTrueSpikePull (CodeStream &os, const NeuronGroupInternal &ng, unsigned int batchSize) const override
	Generate code for pulling true spikes emitted by a neuron group in the current timestep from the 'device'. More...

virtual void	genCurrentSpikeLikeEventPush (CodeStream &os, const NeuronGroupInternal &ng, unsigned int batchSize) const override
	Generate code for pushing spike-like events emitted by a neuron group in the current timestep to the 'device'. More...

virtual void	genCurrentSpikeLikeEventPull (CodeStream &os, const NeuronGroupInternal &ng, unsigned int batchSize) const override
	Generate code for pulling spike-like events emitted by a neuron group in the current timestep from the 'device'. More...

virtual void	genGlobalDeviceRNG (CodeStream &definitions, CodeStream &definitionsInternal, CodeStream &runner, CodeStream &allocations, CodeStream &free, MemAlloc &memAlloc) const override
	Generate a single RNG instance. More...

virtual void	genPopulationRNG (CodeStream &definitions, CodeStream &definitionsInternal, CodeStream &runner, CodeStream &allocations, CodeStream &free, const std::string &name, size_t count, MemAlloc &memAlloc) const override
	Generate an RNG with a state per population member. More...

virtual void	genTimer (CodeStream &definitions, CodeStream &definitionsInternal, CodeStream &runner, CodeStream &allocations, CodeStream &free, CodeStream &stepTimeFinalise, const std::string &name, bool updateInStepTime) const override

virtual void	genReturnFreeDeviceMemoryBytes (CodeStream &os) const override
	Generate code to return amount of free 'device' memory in bytes. More...

virtual void	genMakefilePreamble (std::ostream &os) const override
	This function can be used to generate a preamble for the GNU makefile used to build. More...

virtual void	genMakefileLinkRule (std::ostream &os) const override

virtual void	genMakefileCompileRule (std::ostream &os) const override

virtual void	genMSBuildConfigProperties (std::ostream &os) const override

virtual void	genMSBuildImportProps (std::ostream &os) const override

virtual void	genMSBuildItemDefinitions (std::ostream &os) const override

virtual void	genMSBuildCompileModule (const std::string &moduleName, std::ostream &os) const override

virtual void	genMSBuildImportTarget (std::ostream &os) const override

virtual std::string	getAllocateMemParams (const ModelSpecMerged &) const override
	Get backend-specific allocate memory parameters. More...

virtual bool	isPopulationRNGInitialisedOnDevice () const override
	Different backends seed RNGs in different ways. Does this one initialise population RNGS on device? More...

virtual bool	isHostReductionRequired () const override
	Backends which support batch-parallelism might require an additional host reduction phase after reduction kernels. More...

virtual size_t	getDeviceMemoryBytes () const override
	How many bytes of memory does 'device' have. More...

virtual MemorySpaces	getMergedGroupMemorySpaces (const ModelSpecMerged &modelMerged) const override

virtual bool	supportsNamespace () const override
	Does this backend support namespaces i.e. can C++ implementation of support functions be used. More...

virtual boost::uuids::detail::sha1::digest_type	getHashDigest () const override
	Get hash digest of this backends identification and the preferences it has been configured with. More...

const cudaDeviceProp &	getChosenCUDADevice () const

int	getChosenDeviceID () const

int	getRuntimeVersion () const

std::string	getNVCCFlags () const

Public Member Functions inherited from CodeGenerator::BackendSIMT
	BackendSIMT (const KernelBlockSize &kernelBlockSizes, const PreferencesBase &preferences, const std::string &scalarType)

virtual size_t	getSynapticMatrixRowStride (const SynapseGroupInternal &sg) const final
	Gets the stride used to access synaptic matrix rows, taking into account sparse data structure, padding etc. More...

virtual std::string	getDeviceVarPrefix () const final

virtual void	genPopVariableInit (CodeStream &os, const Substitutions &kernelSubs, Handler handler) const final

virtual void	genVariableInit (CodeStream &os, const std::string &count, const std::string &indexVarName, const Substitutions &kernelSubs, Handler handler) const final

virtual void	genSparseSynapseVariableRowInit (CodeStream &os, const Substitutions &kernelSubs, Handler handler) const final

virtual void	genDenseSynapseVariableRowInit (CodeStream &os, const Substitutions &kernelSubs, Handler handler) const final

virtual void	genKernelSynapseVariableInit (CodeStream &os, const SynapseInitGroupMerged &sg, const Substitutions &kernelSubs, Handler handler) const final

virtual void	genKernelCustomUpdateVariableInit (CodeStream &os, const CustomWUUpdateInitGroupMerged &cu, const Substitutions &kernelSubs, Handler handler) const final

virtual bool	isDeviceScalarRequired () const final
	Should 'scalar' variables be implemented on device or can host variables be used directly? More...

virtual bool	isGlobalHostRNGRequired (const ModelSpecMerged &modelMerged) const final
	Different backends use different RNGs for different things. Does this one require a global host RNG for the specified model? More...

virtual bool	isGlobalDeviceRNGRequired (const ModelSpecMerged &modelMerged) const final
	Different backends use different RNGs for different things. Does this one require a global device RNG for the specified model? More...

virtual bool	isPopulationRNGRequired () const final
	Different backends use different RNGs for different things. Does this one require population RNGs? More...

virtual bool	isPostsynapticRemapRequired () const final
	Different backends may implement synaptic plasticity differently. Does this one require a postsynaptic remapping data structure? More...

size_t	getNumInitialisationRNGStreams (const ModelSpecMerged &modelMerged) const
	Get total number of RNG streams potentially used to initialise model. More...

size_t	getKernelBlockSize (Kernel kernel) const

size_t	getPaddedNumCustomUpdateThreads (const CustomUpdateInternal &cg, unsigned int batchSize) const

size_t	getPaddedNumCustomUpdateWUThreads (const CustomUpdateWUInternal &cg, unsigned int batchSize) const

size_t	getPaddedNumCustomUpdateTransposeWUThreads (const CustomUpdateWUInternal &cg, unsigned int batchSize) const

Public Member Functions inherited from CodeGenerator::BackendBase
	BackendBase (const std::string &scalarType, const PreferencesBase &preferences)

virtual	~BackendBase ()

virtual std::vector< filesystem::path >	getFilesToCopy (const ModelSpecMerged &) const
	Get list of files to copy into generated code. More...

virtual std::string	getHostVarPrefix () const

virtual std::string	getPointerPrefix () const
	Different backends may have different or no pointer prefix (e.g. __global for OpenCL) More...

void	genVariablePushPull (CodeStream &push, CodeStream &pull, const std::string &type, const std::string &name, VarLocation loc, bool autoInitialized, size_t count) const
	Helper function to generate matching push and pull functions for a variable. More...

void	genCurrentVariablePushPull (CodeStream &push, CodeStream &pull, const NeuronGroupInternal &ng, const std::string &type, const std::string &name, VarLocation loc, unsigned int batchSize) const
	Helper function to generate matching push and pull functions for the current state of a variable. More...

void	genArray (CodeStream &definitions, CodeStream &definitionsInternal, CodeStream &runner, CodeStream &allocations, CodeStream &free, const std::string &type, const std::string &name, VarLocation loc, size_t count, MemAlloc &memAlloc) const
	Helper function to generate matching definition, declaration, allocation and free code for an array. More...

size_t	getSize (const std::string &type) const
	Get the size of the type. More...

std::string	getLowestValue (const std::string &type) const
	Get the lowest value of a type. More...

std::string	getScalarAddressPrefix () const
	Get the prefix for accessing the address of 'scalar' variables. More...

bool	areSixtyFourBitSynapseIndicesRequired (const SynapseGroupMergedBase &sg) const

const PreferencesBase &	getPreferences () const

template<typename T >
const T &	getPreferences () const

Additional Inherited Members
Public Types inherited from CodeGenerator::BackendSIMT
enum	AtomicOperation { AtomicOperation::ADD, AtomicOperation::OR }
	What atomic operation is required. More...

enum	AtomicMemSpace { AtomicMemSpace::GLOBAL, AtomicMemSpace::SHARED }
	What memory space atomic operation is required. More...

Public Types inherited from CodeGenerator::BackendBase
typedef std::function< void(CodeStream &)>	HostHandler

typedef std::function< void(CodeStream &, Substitutions &)>	Handler

template<typename T >
using	GroupHandler = std::function< void(CodeStream &, const T &, Substitutions &)>

typedef std::vector< std::pair< std::string, size_t > >	MemorySpaces
	Vector of prefixes required to allocate in memory space and size of memory space. More...

Static Public Member Functions inherited from CodeGenerator::BackendSIMT
static size_t	getNumPresynapticUpdateThreads (const SynapseGroupInternal &sg, const PreferencesBase &preferences)

static size_t	getNumPostsynapticUpdateThreads (const SynapseGroupInternal &sg)

static size_t	getNumSynapseDynamicsThreads (const SynapseGroupInternal &sg)

static size_t	getNumConnectivityInitThreads (const SynapseGroupInternal &sg)

static size_t	getNumInitThreads (const SynapseGroupInternal &sg)

static size_t	getNumInitThreads (const CustomUpdateWUInternal &cg)

static void	addPresynapticUpdateStrategy (PresynapticUpdateStrategySIMT::Base *strategy)
	Register a new presynaptic update strategy. More...

Static Public Attributes inherited from CodeGenerator::BackendSIMT
static const char *	KernelNames [KernelMax]

Protected Member Functions inherited from CodeGenerator::BackendSIMT
void	genNeuronPrevSpikeTimeUpdateKernel (CodeStream &os, const Substitutions &kernelSubs, const ModelSpecMerged &modelMerged, size_t &idStart) const

void	genNeuronSpikeQueueUpdateKernel (CodeStream &os, const ModelSpecMerged &modelMerged, size_t &idStart) const

void	genNeuronUpdateKernel (CodeStream &os, const Substitutions &kernelSubs, const ModelSpecMerged &modelMerged, size_t &idStart) const

void	genSynapseDendriticDelayUpdateKernel (CodeStream &os, const ModelSpecMerged &modelMerged, size_t &idStart) const

void	genPresynapticUpdateKernel (CodeStream &os, const Substitutions &kernelSubs, const ModelSpecMerged &modelMerged, size_t &idStart) const

void	genPostsynapticUpdateKernel (CodeStream &os, const Substitutions &kernelSubs, const ModelSpecMerged &modelMerged, size_t &idStart) const

void	genSynapseDynamicsKernel (CodeStream &os, const Substitutions &kernelSubs, const ModelSpecMerged &modelMerged, size_t &idStart) const

void	genCustomUpdateKernel (CodeStream &os, const Substitutions &kernelSubs, const ModelSpecMerged &modelMerged, const std::string &updateGroup, size_t &idStart) const

void	genCustomUpdateWUKernel (CodeStream &os, const Substitutions &kernelSubs, const ModelSpecMerged &modelMerged, const std::string &updateGroup, size_t &idStart) const

void	genCustomTransposeUpdateWUKernel (CodeStream &os, const Substitutions &kernelSubs, const ModelSpecMerged &modelMerged, const std::string &updateGroup, size_t &idStart) const

void	genInitializeKernel (CodeStream &os, const Substitutions &kernelSubs, const ModelSpecMerged &modelMerged, size_t &idStart) const

void	genInitializeSparseKernel (CodeStream &os, const Substitutions &kernelSubs, const ModelSpecMerged &modelMerged, size_t numInitializeThreads, size_t &idStart) const

void	addDeviceType (const std::string &type, size_t size, const std::string &maxValue="")
	Adds a type - both to backend base's list of sized types but also to device types set. More...

bool	isDeviceType (const std::string &type) const
	Is type a a device only type? More...

size_t	padKernelSize (size_t size, Kernel kernel) const
	Helper wrapper around padSize to pad size to a kernel size. More...

const KernelBlockSize &	getKernelBlockSize () const
	Get kernel block size. More...

Protected Member Functions inherited from CodeGenerator::BackendBase
void	addType (const std::string &type, size_t size, const std::string &lowestValue="")

void	setPointerBytes (size_t pointerBytes)

void	genNeuronIndexCalculation (CodeStream &os, const NeuronUpdateGroupMerged &ng, unsigned int batchSize) const

void	genSynapseIndexCalculation (CodeStream &os, const SynapseGroupMergedBase &sg, unsigned int batchSize) const

void	genCustomUpdateIndexCalculation (CodeStream &os, const CustomUpdateGroupMerged &cu) const

std::vector< ReductionTarget >	genInitReductionTargets (CodeStream &os, const CustomUpdateGroupMerged &cg, const std::string &idx="") const

std::vector< ReductionTarget >	genInitReductionTargets (CodeStream &os, const CustomUpdateWUGroupMerged &cg, const std::string &idx="") const

Constructor & Destructor Documentation

◆ Backend()

CodeGenerator::CUDA::Backend::Backend	(	const KernelBlockSize &	kernelBlockSizes,
		const Preferences &	preferences,
		const std::string &	scalarType,
		int	device
	)

Member Function Documentation

◆ areSharedMemAtomicsSlow()

bool CodeGenerator::CUDA::Backend::areSharedMemAtomicsSlow ( ) const

overridevirtual

On some older devices, shared memory atomics are actually slower than global memory atomics so should be avoided.

Implements CodeGenerator::BackendSIMT.

◆ genAllocateMemPreamble()

void CodeGenerator::CUDA::Backend::genAllocateMemPreamble	(	CodeStream &	os,
		const ModelSpecMerged &	modelMerged,
		const MemAlloc &	memAlloc
	)		const

overridevirtual

Allocate memory is the first function in GeNN generated code called by usercode and it should only ever be called once. Therefore it's a good place for any global initialisation. This function generates a 'preamble' to this function.

Implements CodeGenerator::BackendBase.

◆ genCurrentSpikeLikeEventPull()

virtual void CodeGenerator::CUDA::Backend::genCurrentSpikeLikeEventPull	(	CodeStream &	os,
		const NeuronGroupInternal &	ng,
		unsigned int	batchSize
	)		const

inlineoverridevirtual

Generate code for pulling spike-like events emitted by a neuron group in the current timestep from the 'device'.

Implements CodeGenerator::BackendBase.

◆ genCurrentSpikeLikeEventPush()

virtual void CodeGenerator::CUDA::Backend::genCurrentSpikeLikeEventPush	(	CodeStream &	os,
		const NeuronGroupInternal &	ng,
		unsigned int	batchSize
	)		const

inlineoverridevirtual

Generate code for pushing spike-like events emitted by a neuron group in the current timestep to the 'device'.

Implements CodeGenerator::BackendBase.

◆ genCurrentTrueSpikePull()

virtual void CodeGenerator::CUDA::Backend::genCurrentTrueSpikePull	(	CodeStream &	os,
		const NeuronGroupInternal &	ng,
		unsigned int	batchSize
	)		const

inlineoverridevirtual

Generate code for pulling true spikes emitted by a neuron group in the current timestep from the 'device'.

Implements CodeGenerator::BackendBase.

◆ genCurrentTrueSpikePush()

virtual void CodeGenerator::CUDA::Backend::genCurrentTrueSpikePush	(	CodeStream &	os,
		const NeuronGroupInternal &	ng,
		unsigned int	batchSize
	)		const

inlineoverridevirtual

Generate code for pushing true spikes emitted by a neuron group in the current timestep to the 'device'.

Implements CodeGenerator::BackendBase.

◆ genCurrentVariablePull()

void CodeGenerator::CUDA::Backend::genCurrentVariablePull	(	CodeStream &	os,
		const NeuronGroupInternal &	ng,
		const std::string &	type,
		const std::string &	name,
		VarLocation	loc,
		unsigned int	batchSize
	)		const

overridevirtual

Generate code for pulling a variable's value in the current timestep from the 'device'.

Implements CodeGenerator::BackendBase.

◆ genCurrentVariablePush()

void CodeGenerator::CUDA::Backend::genCurrentVariablePush	(	CodeStream &	os,
		const NeuronGroupInternal &	ng,
		const std::string &	type,
		const std::string &	name,
		VarLocation	loc,
		unsigned int	batchSize
	)		const

overridevirtual

Generate code for pushing a variable's value in the current timestep to the 'device'.

Implements CodeGenerator::BackendBase.

◆ genCustomUpdate()

void CodeGenerator::CUDA::Backend::genCustomUpdate	(	CodeStream &	os,
		const ModelSpecMerged &	modelMerged,
		HostHandler	preambleHandler,
		HostHandler	pushEGPHandler
	)		const

overridevirtual

Generate platform-specific functions to perform custom updates.

Parameters

os	CodeStream to write function to
modelMerged	merged model to generate code for
preambleHandler	callback to write functions for pushing extra-global parameters
pushEGPHandler	callback to write required extra-global parameter pushing code to start of customUpdate function

Implements CodeGenerator::BackendBase.

◆ genDefinitionsInternalPreamble()

void CodeGenerator::CUDA::Backend::genDefinitionsInternalPreamble	(	CodeStream &	os,
		const ModelSpecMerged &	modelMerged
	)		const

overridevirtual

Definitions internal is the internal header file for the generated code. This function generates a 'preamble' to this header file.

This will only be included by the platform-specific compiler used to build this backend so can include platform-specific types or headers

Implements CodeGenerator::BackendBase.

◆ genDefinitionsPreamble()

void CodeGenerator::CUDA::Backend::genDefinitionsPreamble	(	CodeStream &	os,
		const ModelSpecMerged &	modelMerged
	)		const

overridevirtual

Definitions is the usercode-facing header file for the generated code. This function generates a 'preamble' to this header file.

This will be included from a standard C++ compiler so shouldn't include any platform-specific types or headers

Implements CodeGenerator::BackendBase.

◆ genExtraGlobalParamAllocation()

void CodeGenerator::CUDA::Backend::genExtraGlobalParamAllocation	(	CodeStream &	os,
		const std::string &	type,
		const std::string &	name,
		VarLocation	loc,
		const std::string &	countVarName = `"count"`,
		const std::string &	prefix = `""`
	)		const

overridevirtual

Implements CodeGenerator::BackendBase.

◆ genExtraGlobalParamDefinition()

void CodeGenerator::CUDA::Backend::genExtraGlobalParamDefinition	(	CodeStream &	definitions,
		CodeStream &	definitionsInternal,
		const std::string &	type,
		const std::string &	name,
		VarLocation	loc
	)		const

overridevirtual

Implements CodeGenerator::BackendBase.

◆ genExtraGlobalParamImplementation()

void CodeGenerator::CUDA::Backend::genExtraGlobalParamImplementation	(	CodeStream &	os,
		const std::string &	type,
		const std::string &	name,
		VarLocation	loc
	)		const

overridevirtual

Implements CodeGenerator::BackendBase.

◆ genExtraGlobalParamPull()

void CodeGenerator::CUDA::Backend::genExtraGlobalParamPull	(	CodeStream &	os,
		const std::string &	type,
		const std::string &	name,
		VarLocation	loc,
		const std::string &	countVarName = `"count"`,
		const std::string &	prefix = `""`
	)		const

overridevirtual

Implements CodeGenerator::BackendBase.

◆ genExtraGlobalParamPush()

void CodeGenerator::CUDA::Backend::genExtraGlobalParamPush	(	CodeStream &	os,
		const std::string &	type,
		const std::string &	name,
		VarLocation	loc,
		const std::string &	countVarName = `"count"`,
		const std::string &	prefix = `""`
	)		const

overridevirtual

Implements CodeGenerator::BackendBase.

◆ genFreeMemPreamble()

void CodeGenerator::CUDA::Backend::genFreeMemPreamble	(	CodeStream &	os,
		const ModelSpecMerged &	modelMerged
	)		const

overridevirtual

Free memory is called by usercode to free all memory allocatd by GeNN and should only ever be called once. This function generates a 'preamble' to this function, for example to free backend-specific objects

Implements CodeGenerator::BackendBase.

◆ genGlobalDeviceRNG()

void CodeGenerator::CUDA::Backend::genGlobalDeviceRNG	(	CodeStream &	definitions,
		CodeStream &	definitionsInternal,
		CodeStream &	runner,
		CodeStream &	allocations,
		CodeStream &	free,
		MemAlloc &	memAlloc
	)		const

overridevirtual

Generate a single RNG instance.

On single-threaded platforms this can be a standard RNG like M.T. but, on parallel platforms, it is likely to be a counter-based RNG

Implements CodeGenerator::BackendBase.

◆ genGlobalRNGSkipAhead()

void CodeGenerator::CUDA::Backend::genGlobalRNGSkipAhead	(	CodeStream &	os,
		Substitutions &	subs,
		const std::string &	sequence,
		const std::string &	name = `"rng"`
	)		const

overridevirtual

Generate code to skip ahead local copy of global RNG.

Implements CodeGenerator::BackendSIMT.

◆ genInit()

void CodeGenerator::CUDA::Backend::genInit	(	CodeStream &	os,
		const ModelSpecMerged &	modelMerged,
		HostHandler	preambleHandler,
		HostHandler	initPushEGPHandler,
		HostHandler	initSparsePushEGPHandler
	)		const

overridevirtual

Generate platform-specific function to initialise model.

Parameters

os	CodeStream to write function to
modelMerged	merged model to generate code for
preambleHandler	callback to write functions for pushing extra-global parameters
initPushEGPHandler	callback to write required extra-global parameter pushing code to start of initialize function
initSparsePushEGPHandler	callback to write required extra-global parameter pushing code to start of initializeSparse function

Implements CodeGenerator::BackendBase.

◆ genMakefileCompileRule()

void CodeGenerator::CUDA::Backend::genMakefileCompileRule ( std::ostream & os ) const

overridevirtual

The GNU make build system uses 'pattern rules' (https://www.gnu.org/software/make/manual/html_node/Pattern-Intro.html) to build backend modules into objects. This function should generate a GNU make pattern rule capable of building each module (i.e. compiling .cc file $< into .o file $@).

Implements CodeGenerator::BackendBase.

◆ genMakefileLinkRule()

void CodeGenerator::CUDA::Backend::genMakefileLinkRule ( std::ostream & os ) const

overridevirtual

The GNU make build system will populate a variable called with a list of objects to link. This function should generate a GNU make rule to build these objects into a shared library.

Implements CodeGenerator::BackendBase.

◆ genMakefilePreamble()

void CodeGenerator::CUDA::Backend::genMakefilePreamble ( std::ostream & os ) const

overridevirtual

This function can be used to generate a preamble for the GNU makefile used to build.

Implements CodeGenerator::BackendBase.

◆ genMergedExtraGlobalParamPush()

void CodeGenerator::CUDA::Backend::genMergedExtraGlobalParamPush	(	CodeStream &	os,
		const std::string &	suffix,
		size_t	mergedGroupIdx,
		const std::string &	groupIdx,
		const std::string &	fieldName,
		const std::string &	egpName
	)		const

overridevirtual

Generate code for pushing an updated EGP value into the merged group structure on 'device'.

Implements CodeGenerator::BackendBase.

◆ genMSBuildCompileModule()

void CodeGenerator::CUDA::Backend::genMSBuildCompileModule	(	const std::string &	moduleName,
		std::ostream &	os
	)		const

overridevirtual

Implements CodeGenerator::BackendBase.

◆ genMSBuildConfigProperties()

void CodeGenerator::CUDA::Backend::genMSBuildConfigProperties ( std::ostream & os ) const

overridevirtual

In MSBuild, 'properties' are used to configure global project settings e.g. whether the MSBuild project builds a static or dynamic library This function can be used to add additional XML properties to this section.

see https://docs.microsoft.com/en-us/visualstudio/msbuild/msbuild-properties for more information.

Implements CodeGenerator::BackendBase.

◆ genMSBuildImportProps()

void CodeGenerator::CUDA::Backend::genMSBuildImportProps ( std::ostream & os ) const

overridevirtual

Implements CodeGenerator::BackendBase.

◆ genMSBuildImportTarget()

void CodeGenerator::CUDA::Backend::genMSBuildImportTarget ( std::ostream & os ) const

overridevirtual

Implements CodeGenerator::BackendBase.

◆ genMSBuildItemDefinitions()

void CodeGenerator::CUDA::Backend::genMSBuildItemDefinitions ( std::ostream & os ) const

overridevirtual

In MSBuild, the 'item definitions' are used to override the default properties of 'items' such as <ClCompile> or <Link>. This function should generate XML to correctly configure the 'items' required to build the generated code, taking into account etc.

see https://docs.microsoft.com/en-us/visualstudio/msbuild/msbuild-items#item-definitions for more information.

Implements CodeGenerator::BackendBase.

◆ genNeuronUpdate()

void CodeGenerator::CUDA::Backend::genNeuronUpdate	(	CodeStream &	os,
		const ModelSpecMerged &	modelMerged,
		HostHandler	preambleHandler,
		HostHandler	pushEGPHandler
	)		const

overridevirtual

Generate platform-specific function to update the state of all neurons.

Parameters

os	CodeStream to write function to
modelMerged	merged model to generate code for
preambleHandler	callback to write functions for pushing extra-global parameters
pushEGPHandler	callback to write required extra-global parameter pushing code to start of neuronUpdate function

Implements CodeGenerator::BackendBase.

◆ genPopulationRNG()

void CodeGenerator::CUDA::Backend::genPopulationRNG	(	CodeStream &	definitions,
		CodeStream &	definitionsInternal,
		CodeStream &	runner,
		CodeStream &	allocations,
		CodeStream &	free,
		const std::string &	name,
		size_t	count,
		MemAlloc &	memAlloc
	)		const

overridevirtual

Generate an RNG with a state per population member.

Implements CodeGenerator::BackendBase.

◆ genPopulationRNGInit()

void CodeGenerator::CUDA::Backend::genPopulationRNGInit	(	CodeStream &	os,
		const std::string &	globalRNG,
		const std::string &	seed,
		const std::string &	sequence
	)		const

overridevirtual

For SIMT backends which initialize RNGs on device, initialize population RNG with specified seed and sequence.

Implements CodeGenerator::BackendSIMT.

◆ genPopulationRNGPostamble()

void CodeGenerator::CUDA::Backend::genPopulationRNGPostamble	(	CodeStream &	os,
		const std::string &	globalRNG
	)		const

overridevirtual

If required, generate a postamble for population RNG.

For example, in OpenCL, this is used to write local RNG state back to global memory

Implements CodeGenerator::BackendSIMT.

◆ genPopulationRNGPreamble()

void CodeGenerator::CUDA::Backend::genPopulationRNGPreamble	(	CodeStream &	os,
		Substitutions &	subs,
		const std::string &	globalRNG,
		const std::string &	name = `"rng"`
	)		const

overridevirtual

Generate a preamble to add substitution name for population RNG.

Implements CodeGenerator::BackendSIMT.

◆ genReturnFreeDeviceMemoryBytes()

void CodeGenerator::CUDA::Backend::genReturnFreeDeviceMemoryBytes ( CodeStream & os ) const

overridevirtual

Generate code to return amount of free 'device' memory in bytes.

Implements CodeGenerator::BackendBase.

◆ genRunnerPreamble()

void CodeGenerator::CUDA::Backend::genRunnerPreamble	(	CodeStream &	os,
		const ModelSpecMerged &	modelMerged,
		const MemAlloc &	memAlloc
	)		const

overridevirtual

Implements CodeGenerator::BackendBase.

◆ genSharedMemBarrier()

void CodeGenerator::CUDA::Backend::genSharedMemBarrier ( CodeStream & os ) const

overridevirtual

Generate a shared memory barrier.

Implements CodeGenerator::BackendSIMT.

◆ genStepTimeFinalisePreamble()

void CodeGenerator::CUDA::Backend::genStepTimeFinalisePreamble	(	CodeStream &	os,
		const ModelSpecMerged &	modelMerged
	)		const

overridevirtual

After all timestep logic is complete.

Implements CodeGenerator::BackendBase.

◆ genSynapseUpdate()

void CodeGenerator::CUDA::Backend::genSynapseUpdate	(	CodeStream &	os,
		const ModelSpecMerged &	modelMerged,
		HostHandler	preambleHandler,
		HostHandler	pushEGPHandler
	)		const

overridevirtual

Generate platform-specific function to update the state of all synapses.

Parameters

os	CodeStream to write function to
modelMerged	merged model to generate code for
preambleHandler	callback to write functions for pushing extra-global parameters
pushEGPHandler	callback to write required extra-global parameter pushing code to start of synapseUpdate function

Implements CodeGenerator::BackendBase.

◆ genTimer()

void CodeGenerator::CUDA::Backend::genTimer	(	CodeStream &	definitions,
		CodeStream &	definitionsInternal,
		CodeStream &	runner,
		CodeStream &	allocations,
		CodeStream &	free,
		CodeStream &	stepTimeFinalise,
		const std::string &	name,
		bool	updateInStepTime
	)		const

overridevirtual

Implements CodeGenerator::BackendBase.

◆ genVariableAllocation()

void CodeGenerator::CUDA::Backend::genVariableAllocation	(	CodeStream &	os,
		const std::string &	type,
		const std::string &	name,
		VarLocation	loc,
		size_t	count,
		MemAlloc &	memAlloc
	)		const

overridevirtual

Implements CodeGenerator::BackendBase.

◆ genVariableDefinition()

void CodeGenerator::CUDA::Backend::genVariableDefinition	(	CodeStream &	definitions,
		CodeStream &	definitionsInternal,
		const std::string &	type,
		const std::string &	name,
		VarLocation	loc
	)		const

overridevirtual

Implements CodeGenerator::BackendBase.

◆ genVariableFree()

void CodeGenerator::CUDA::Backend::genVariableFree	(	CodeStream &	os,
		const std::string &	name,
		VarLocation	loc
	)		const

overridevirtual

Implements CodeGenerator::BackendBase.

◆ genVariableImplementation()

void CodeGenerator::CUDA::Backend::genVariableImplementation	(	CodeStream &	os,
		const std::string &	type,
		const std::string &	name,
		VarLocation	loc
	)		const

overridevirtual

Implements CodeGenerator::BackendBase.

◆ genVariablePull()

void CodeGenerator::CUDA::Backend::genVariablePull	(	CodeStream &	os,
		const std::string &	type,
		const std::string &	name,
		VarLocation	loc,
		size_t	count
	)		const

overridevirtual

Generate code for pulling a variable from the 'device'.

Implements CodeGenerator::BackendBase.

◆ genVariablePush()

void CodeGenerator::CUDA::Backend::genVariablePush	(	CodeStream &	os,
		const std::string &	type,
		const std::string &	name,
		VarLocation	loc,
		bool	autoInitialized,
		size_t	count
	)		const

overridevirtual

Generate code for pushing a variable to the 'device'.

Implements CodeGenerator::BackendBase.

◆ getAllocateMemParams()

std::string CodeGenerator::CUDA::Backend::getAllocateMemParams ( const ModelSpecMerged & ) const

overridevirtual

Get backend-specific allocate memory parameters.

Reimplemented from CodeGenerator::BackendBase.

◆ getAtomic()

std::string CodeGenerator::CUDA::Backend::getAtomic	(	const std::string &	type,
		AtomicOperation	op = `AtomicOperation::ADD`,
		AtomicMemSpace	memSpace = `AtomicMemSpace::GLOBAL`
	)		const

overridevirtual

Get name of atomic operation.

Implements CodeGenerator::BackendSIMT.

◆ getBlockID()

std::string CodeGenerator::CUDA::Backend::getBlockID ( unsigned int axis = 0 ) const

overridevirtual

Get the ID of the current thread block.

Implements CodeGenerator::BackendSIMT.

◆ getChosenCUDADevice()

const cudaDeviceProp& CodeGenerator::CUDA::Backend::getChosenCUDADevice ( ) const

inline

◆ getChosenDeviceID()

int CodeGenerator::CUDA::Backend::getChosenDeviceID ( ) const

inline

◆ getCLZ()

virtual std::string CodeGenerator::CUDA::Backend::getCLZ ( ) const

inlineoverridevirtual

Get the name of the count-leading-zeros function.

Implements CodeGenerator::BackendSIMT.

◆ getDeviceMemoryBytes()

virtual size_t CodeGenerator::CUDA::Backend::getDeviceMemoryBytes ( ) const

inlineoverridevirtual

How many bytes of memory does 'device' have.

Implements CodeGenerator::BackendBase.

◆ getHashDigest()

boost::uuids::detail::sha1::digest_type CodeGenerator::CUDA::Backend::getHashDigest ( ) const

overridevirtual

Get hash digest of this backends identification and the preferences it has been configured with.

Implements CodeGenerator::BackendBase.

◆ getMergedGroupFieldHostType()

std::string CodeGenerator::CUDA::Backend::getMergedGroupFieldHostType ( const std::string & type ) const

overridevirtual

When generating function calls to push to merged groups, backend without equivalent of Unified Virtual Addressing e.g. OpenCL 1.2 may use different types on host.

Implements CodeGenerator::BackendBase.

◆ getMergedGroupMemorySpaces()

Backend::MemorySpaces CodeGenerator::CUDA::Backend::getMergedGroupMemorySpaces ( const ModelSpecMerged & modelMerged ) const

overridevirtual

Some backends will have additional small, fast, memory spaces for read-only data which might Be well-suited to storing merged group structs. This method returns the prefix required to Place arrays in these and their size in preferential order

Implements CodeGenerator::BackendBase.

◆ getMergedGroupSimRNGType()

virtual std::string CodeGenerator::CUDA::Backend::getMergedGroupSimRNGType ( ) const

inlineoverridevirtual

When generating merged structures what type to use for simulation RNGs.

Implements CodeGenerator::BackendBase.

◆ getNVCCFlags()

std::string CodeGenerator::CUDA::Backend::getNVCCFlags ( ) const

◆ getRuntimeVersion()

int CodeGenerator::CUDA::Backend::getRuntimeVersion ( ) const

inline

◆ getSharedPrefix()

virtual std::string CodeGenerator::CUDA::Backend::getSharedPrefix ( ) const

inlineoverridevirtual

Get the prefix to use for shared memory variables.

Implements CodeGenerator::BackendSIMT.

◆ getThreadID()

std::string CodeGenerator::CUDA::Backend::getThreadID ( unsigned int axis = 0 ) const

overridevirtual

Get the ID of the current thread within the threadblock.

Implements CodeGenerator::BackendSIMT.

◆ isHostReductionRequired()

virtual bool CodeGenerator::CUDA::Backend::isHostReductionRequired ( ) const

inlineoverridevirtual

Backends which support batch-parallelism might require an additional host reduction phase after reduction kernels.

Implements CodeGenerator::BackendBase.

◆ isPopulationRNGInitialisedOnDevice()

virtual bool CodeGenerator::CUDA::Backend::isPopulationRNGInitialisedOnDevice ( ) const

inlineoverridevirtual

Different backends seed RNGs in different ways. Does this one initialise population RNGS on device?

Implements CodeGenerator::BackendBase.

◆ supportsNamespace()

virtual bool CodeGenerator::CUDA::Backend::supportsNamespace ( ) const

inlineoverridevirtual

Does this backend support namespaces i.e. can C++ implementation of support functions be used.

Implements CodeGenerator::BackendBase.

The documentation for this class was generated from the following files:

Public Member Functions

Additional Inherited Members

Constructor & Destructor Documentation

◆ Backend()

Member Function Documentation

◆ areSharedMemAtomicsSlow()

◆ genAllocateMemPreamble()

◆ genCurrentSpikeLikeEventPull()

◆ genCurrentSpikeLikeEventPush()

◆ genCurrentTrueSpikePull()

◆ genCurrentTrueSpikePush()

◆ genCurrentVariablePull()

◆ genCurrentVariablePush()

◆ genCustomUpdate()

◆ genDefinitionsInternalPreamble()

◆ genDefinitionsPreamble()

◆ genExtraGlobalParamAllocation()

◆ genExtraGlobalParamDefinition()

◆ genExtraGlobalParamImplementation()

◆ genExtraGlobalParamPull()

◆ genExtraGlobalParamPush()

◆ genFreeMemPreamble()

◆ genGlobalDeviceRNG()

◆ genGlobalRNGSkipAhead()

◆ genInit()

◆ genMakefileCompileRule()

◆ genMakefileLinkRule()

◆ genMakefilePreamble()

◆ genMergedExtraGlobalParamPush()

◆ genMSBuildCompileModule()

◆ genMSBuildConfigProperties()

◆ genMSBuildImportProps()

◆ genMSBuildImportTarget()

◆ genMSBuildItemDefinitions()

◆ genNeuronUpdate()

◆ genPopulationRNG()

◆ genPopulationRNGInit()

◆ genPopulationRNGPostamble()

◆ genPopulationRNGPreamble()

◆ genReturnFreeDeviceMemoryBytes()

◆ genRunnerPreamble()

◆ genSharedMemBarrier()

◆ genStepTimeFinalisePreamble()

◆ genSynapseUpdate()

◆ genTimer()

◆ genVariableAllocation()

◆ genVariableDefinition()

◆ genVariableFree()

◆ genVariableImplementation()

◆ genVariablePull()

◆ genVariablePush()

◆ getAllocateMemParams()

◆ getAtomic()

◆ getBlockID()

◆ getChosenCUDADevice()

◆ getChosenDeviceID()

◆ getCLZ()

◆ getDeviceMemoryBytes()

◆ getHashDigest()

◆ getMergedGroupFieldHostType()

◆ getMergedGroupMemorySpaces()

◆ getMergedGroupSimRNGType()

◆ getNVCCFlags()

◆ getRuntimeVersion()

◆ getSharedPrefix()

◆ getThreadID()

◆ isHostReductionRequired()

◆ isPopulationRNGInitialisedOnDevice()

◆ supportsNamespace()