Base class for Single Instruction Multiple Thread style backends. More...

#include <backendSIMT.h>

Inheritance diagram for CodeGenerator::BackendSIMT:

Public Types
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...

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

virtual bool	areSharedMemAtomicsSlow () const =0
	On some older devices, shared memory atomics are actually slower than global memory atomics so should be avoided. More...

virtual std::string	getSharedPrefix () const =0
	Get the prefix to use for shared memory variables. More...

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

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

virtual std::string	getCLZ () const =0
	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 =0
	Get name of atomic operation. More...

virtual void	genSharedMemBarrier (CodeStream &os) const =0
	Generate a shared memory barrier. More...

virtual void	genPopulationRNGInit (CodeStream &os, const std::string &globalRNG, const std::string &seed, const std::string &sequence) const =0
	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 =0
	Generate a preamble to add substitution name for population RNG. More...

virtual void	genPopulationRNGPostamble (CodeStream &os, const std::string &globalRNG) const =0
	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 =0
	Generate code to skip ahead local copy of global RNG. More...

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 void	genNeuronUpdate (CodeStream &os, const ModelSpecMerged &modelMerged, HostHandler preambleHandler, HostHandler pushEGPHandler) const =0
	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 =0
	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 =0
	Generate platform-specific functions to perform custom updates. More...

virtual void	genInit (CodeStream &os, const ModelSpecMerged &modelMerged, HostHandler preambleHandler, HostHandler initPushEGPHandler, HostHandler initSparsePushEGPHandler) const =0
	Generate platform-specific function to initialise model. More...

virtual void	genDefinitionsPreamble (CodeStream &os, const ModelSpecMerged &modelMerged) const =0
	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 =0
	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 =0

virtual void	genAllocateMemPreamble (CodeStream &os, const ModelSpecMerged &modelMerged, const MemAlloc &memAlloc) const =0

virtual void	genFreeMemPreamble (CodeStream &os, const ModelSpecMerged &modelMerged) const =0

virtual void	genStepTimeFinalisePreamble (CodeStream &os, const ModelSpecMerged &modelMerged) const =0
	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 =0

virtual void	genVariableImplementation (CodeStream &os, const std::string &type, const std::string &name, VarLocation loc) const =0

virtual void	genVariableAllocation (CodeStream &os, const std::string &type, const std::string &name, VarLocation loc, size_t count, MemAlloc &memAlloc) const =0

virtual void	genVariableFree (CodeStream &os, const std::string &name, VarLocation loc) const =0

virtual void	genExtraGlobalParamDefinition (CodeStream &definitions, CodeStream &definitionsInternal, const std::string &type, const std::string &name, VarLocation loc) const =0

virtual void	genExtraGlobalParamImplementation (CodeStream &os, const std::string &type, const std::string &name, VarLocation loc) const =0

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 =0

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 =0

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 =0

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 =0
	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 =0
	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 =0
	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 =0
	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 =0
	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 =0
	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 =0
	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 =0
	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 =0
	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 =0
	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 =0
	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 =0
	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 =0
	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 =0

virtual void	genReturnFreeDeviceMemoryBytes (CodeStream &os) const =0
	Generate code to return amount of free 'device' memory in bytes. More...

virtual void	genMakefilePreamble (std::ostream &os) const =0
	This function can be used to generate a preamble for the GNU makefile used to build. More...

virtual void	genMakefileLinkRule (std::ostream &os) const =0

virtual void	genMakefileCompileRule (std::ostream &os) const =0

virtual void	genMSBuildConfigProperties (std::ostream &os) const =0

virtual void	genMSBuildImportProps (std::ostream &os) const =0

virtual void	genMSBuildItemDefinitions (std::ostream &os) const =0

virtual void	genMSBuildCompileModule (const std::string &moduleName, std::ostream &os) const =0

virtual void	genMSBuildImportTarget (std::ostream &os) const =0

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

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...

virtual bool	isPopulationRNGInitialisedOnDevice () const =0
	Different backends seed RNGs in different ways. Does this one initialise population RNGS on device? More...

virtual bool	isHostReductionRequired () const =0
	Backends which support batch-parallelism might require an additional host reduction phase after reduction kernels. More...

virtual size_t	getDeviceMemoryBytes () const =0
	How many bytes of memory does 'device' have. More...

virtual MemorySpaces	getMergedGroupMemorySpaces (const ModelSpecMerged &modelMerged) const =0

virtual bool	supportsNamespace () const =0
	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 =0
	Get hash digest of this backends identification and the preferences it has been configured with. 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

Static Public Member Functions
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
static const char *	KernelNames [KernelMax]

Protected Member Functions
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

Detailed Description

Base class for Single Instruction Multiple Thread style backends.

CUDA terminology is used throughout i.e. thread blocks and shared memory

Member Enumeration Documentation

◆ AtomicMemSpace

enum CodeGenerator::BackendSIMT::AtomicMemSpace

strong

What memory space atomic operation is required.

Enumerator
GLOBAL
SHARED

◆ AtomicOperation

enum CodeGenerator::BackendSIMT::AtomicOperation

strong

What atomic operation is required.

Enumerator
ADD
OR

Constructor & Destructor Documentation

◆ BackendSIMT()

CodeGenerator::BackendSIMT::BackendSIMT	(	const KernelBlockSize &	kernelBlockSizes,
		const PreferencesBase &	preferences,
		const std::string &	scalarType
	)

inline

Member Function Documentation

◆ addDeviceType()

void CodeGenerator::BackendSIMT::addDeviceType	(	const std::string &	type,
		size_t	size,
		const std::string &	maxValue = `""`
	)

protected

Adds a type - both to backend base's list of sized types but also to device types set.

◆ addPresynapticUpdateStrategy()

void CodeGenerator::BackendSIMT::addPresynapticUpdateStrategy ( PresynapticUpdateStrategySIMT::Base * strategy )

static

Register a new presynaptic update strategy.

This function should be called with strategies in ascending order of preference

◆ areSharedMemAtomicsSlow()

virtual bool CodeGenerator::BackendSIMT::areSharedMemAtomicsSlow ( ) const

pure virtual

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

Implemented in CodeGenerator::CUDA::Backend, and CodeGenerator::OpenCL::Backend.

◆ genCustomTransposeUpdateWUKernel()

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

protected

◆ genCustomUpdateKernel()

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

protected

◆ genCustomUpdateWUKernel()

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

protected

◆ genDenseSynapseVariableRowInit()

virtual void CodeGenerator::BackendSIMT::genDenseSynapseVariableRowInit	(	CodeStream &	os,
		const Substitutions &	kernelSubs,
		Handler	handler
	)		const

inlinefinalvirtual

Implements CodeGenerator::BackendBase.

◆ genGlobalRNGSkipAhead()

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

pure virtual

Generate code to skip ahead local copy of global RNG.

Implemented in CodeGenerator::CUDA::Backend, and CodeGenerator::OpenCL::Backend.

◆ genInitializeKernel()

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

protected

◆ genInitializeSparseKernel()

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

protected

◆ genKernelCustomUpdateVariableInit()

void CodeGenerator::BackendSIMT::genKernelCustomUpdateVariableInit	(	CodeStream &	os,
		const CustomWUUpdateInitGroupMerged &	cu,
		const Substitutions &	kernelSubs,
		Handler	handler
	)		const

finalvirtual

Implements CodeGenerator::BackendBase.

◆ genKernelSynapseVariableInit()

void CodeGenerator::BackendSIMT::genKernelSynapseVariableInit	(	CodeStream &	os,
		const SynapseInitGroupMerged &	sg,
		const Substitutions &	kernelSubs,
		Handler	handler
	)		const

finalvirtual

Implements CodeGenerator::BackendBase.

◆ genNeuronPrevSpikeTimeUpdateKernel()

void CodeGenerator::BackendSIMT::genNeuronPrevSpikeTimeUpdateKernel	(	CodeStream &	os,
		const Substitutions &	kernelSubs,
		const ModelSpecMerged &	modelMerged,
		size_t &	idStart
	)		const

protected

◆ genNeuronSpikeQueueUpdateKernel()

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

protected

◆ genNeuronUpdateKernel()

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

protected

◆ genPopulationRNGInit()

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

pure virtual

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

Implemented in CodeGenerator::CUDA::Backend, and CodeGenerator::OpenCL::Backend.

◆ genPopulationRNGPostamble()

virtual void CodeGenerator::BackendSIMT::genPopulationRNGPostamble	(	CodeStream &	os,
		const std::string &	globalRNG
	)		const

pure virtual

If required, generate a postamble for population RNG.

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

Implemented in CodeGenerator::CUDA::Backend, and CodeGenerator::OpenCL::Backend.

◆ genPopulationRNGPreamble()

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

pure virtual

Generate a preamble to add substitution name for population RNG.

Implemented in CodeGenerator::CUDA::Backend, and CodeGenerator::OpenCL::Backend.

◆ genPopVariableInit()

void CodeGenerator::BackendSIMT::genPopVariableInit	(	CodeStream &	os,
		const Substitutions &	kernelSubs,
		Handler	handler
	)		const

finalvirtual

Implements CodeGenerator::BackendBase.

◆ genPostsynapticUpdateKernel()

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

protected

◆ genPresynapticUpdateKernel()

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

protected

◆ genSharedMemBarrier()

virtual void CodeGenerator::BackendSIMT::genSharedMemBarrier ( CodeStream & os ) const

pure virtual

Generate a shared memory barrier.

Implemented in CodeGenerator::CUDA::Backend, and CodeGenerator::OpenCL::Backend.

◆ genSparseSynapseVariableRowInit()

virtual void CodeGenerator::BackendSIMT::genSparseSynapseVariableRowInit	(	CodeStream &	os,
		const Substitutions &	kernelSubs,
		Handler	handler
	)		const

inlinefinalvirtual

Implements CodeGenerator::BackendBase.

◆ genSynapseDendriticDelayUpdateKernel()

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

protected

◆ genSynapseDynamicsKernel()

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

protected

◆ genVariableInit()

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

finalvirtual

Implements CodeGenerator::BackendBase.

◆ getAtomic()

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

pure virtual

Get name of atomic operation.

Implemented in CodeGenerator::CUDA::Backend, and CodeGenerator::OpenCL::Backend.

◆ getBlockID()

virtual std::string CodeGenerator::BackendSIMT::getBlockID ( unsigned int axis = 0 ) const

pure virtual

Get the ID of the current thread block.

Implemented in CodeGenerator::CUDA::Backend, and CodeGenerator::OpenCL::Backend.

◆ getCLZ()

virtual std::string CodeGenerator::BackendSIMT::getCLZ ( ) const

pure virtual

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

Implemented in CodeGenerator::CUDA::Backend, and CodeGenerator::OpenCL::Backend.

◆ getDeviceVarPrefix()

virtual std::string CodeGenerator::BackendSIMT::getDeviceVarPrefix ( ) const

inlinefinalvirtual

When backends require separate 'device' and 'host' versions of variables, they are identified with a prefix. This function returns the device prefix so it can be used in otherwise platform-independent code.

Reimplemented from CodeGenerator::BackendBase.

◆ getKernelBlockSize() [1/2]

size_t CodeGenerator::BackendSIMT::getKernelBlockSize ( Kernel kernel ) const

inline

◆ getKernelBlockSize() [2/2]

const KernelBlockSize& CodeGenerator::BackendSIMT::getKernelBlockSize ( ) const

inlineprotected

Get kernel block size.

◆ getNumConnectivityInitThreads()

size_t CodeGenerator::BackendSIMT::getNumConnectivityInitThreads ( const SynapseGroupInternal & sg )

static

◆ getNumInitialisationRNGStreams()

size_t CodeGenerator::BackendSIMT::getNumInitialisationRNGStreams ( const ModelSpecMerged & modelMerged ) const

Get total number of RNG streams potentially used to initialise model.

NOTE because RNG supports 2^64 streams, we are overly conservative

◆ getNumInitThreads() [1/2]

size_t CodeGenerator::BackendSIMT::getNumInitThreads ( const SynapseGroupInternal & sg )

static

◆ getNumInitThreads() [2/2]

size_t CodeGenerator::BackendSIMT::getNumInitThreads ( const CustomUpdateWUInternal & cg )

static

◆ getNumPostsynapticUpdateThreads()

size_t CodeGenerator::BackendSIMT::getNumPostsynapticUpdateThreads ( const SynapseGroupInternal & sg )

static

◆ getNumPresynapticUpdateThreads()

size_t CodeGenerator::BackendSIMT::getNumPresynapticUpdateThreads	(	const SynapseGroupInternal &	sg,
		const PreferencesBase &	preferences
	)

static

◆ getNumSynapseDynamicsThreads()

size_t CodeGenerator::BackendSIMT::getNumSynapseDynamicsThreads ( const SynapseGroupInternal & sg )

static

◆ getPaddedNumCustomUpdateThreads()

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

◆ getPaddedNumCustomUpdateTransposeWUThreads()

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

◆ getPaddedNumCustomUpdateWUThreads()

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

◆ getSharedPrefix()

virtual std::string CodeGenerator::BackendSIMT::getSharedPrefix ( ) const

pure virtual

Get the prefix to use for shared memory variables.

Implemented in CodeGenerator::CUDA::Backend, and CodeGenerator::OpenCL::Backend.

◆ getSynapticMatrixRowStride()

size_t CodeGenerator::BackendSIMT::getSynapticMatrixRowStride ( const SynapseGroupInternal & sg ) const

finalvirtual

Gets the stride used to access synaptic matrix rows, taking into account sparse data structure, padding etc.

Implements CodeGenerator::BackendBase.

◆ getThreadID()

virtual std::string CodeGenerator::BackendSIMT::getThreadID ( unsigned int axis = 0 ) const

pure virtual

Get the ID of the current thread within the threadblock.

Implemented in CodeGenerator::CUDA::Backend, and CodeGenerator::OpenCL::Backend.

◆ isDeviceScalarRequired()

virtual bool CodeGenerator::BackendSIMT::isDeviceScalarRequired ( ) const

inlinefinalvirtual

Should 'scalar' variables be implemented on device or can host variables be used directly?

Implements CodeGenerator::BackendBase.

◆ isDeviceType()

bool CodeGenerator::BackendSIMT::isDeviceType ( const std::string & type ) const

protected

Is type a a device only type?

◆ isGlobalDeviceRNGRequired()

bool CodeGenerator::BackendSIMT::isGlobalDeviceRNGRequired ( const ModelSpecMerged & modelMerged ) const

finalvirtual

Different backends use different RNGs for different things. Does this one require a global device RNG for the specified model?

Implements CodeGenerator::BackendBase.

◆ isGlobalHostRNGRequired()

bool CodeGenerator::BackendSIMT::isGlobalHostRNGRequired ( const ModelSpecMerged & modelMerged ) const

finalvirtual

Different backends use different RNGs for different things. Does this one require a global host RNG for the specified model?

Implements CodeGenerator::BackendBase.

◆ isPopulationRNGRequired()

virtual bool CodeGenerator::BackendSIMT::isPopulationRNGRequired ( ) const

inlinefinalvirtual

Different backends use different RNGs for different things. Does this one require population RNGs?

Implements CodeGenerator::BackendBase.

◆ isPostsynapticRemapRequired()

virtual bool CodeGenerator::BackendSIMT::isPostsynapticRemapRequired ( ) const

inlinefinalvirtual

Different backends may implement synaptic plasticity differently. Does this one require a postsynaptic remapping data structure?

Implements CodeGenerator::BackendBase.

◆ padKernelSize()

size_t CodeGenerator::BackendSIMT::padKernelSize	(	size_t	size,
		Kernel	kernel
	)		const

protected

Helper wrapper around padSize to pad size to a kernel size.

Member Data Documentation

◆ KernelNames

const char * CodeGenerator::BackendSIMT::KernelNames

static

Initial value:

= {
    "updateNeuronsKernel",
    "updatePresynapticKernel",
    "updatePostsynapticKernel",
    "updateSynapseDynamicsKernel",
    "initializeKernel",
    "initializeSparseKernel",
    "neuronSpikeQueueUpdateKernel",
    "neuronPrevSpikeTimeUpdateKernel",
    "synapseDendriticDelayUpdateKernel",
    "customUpdate",
    "customTransposeUpdate"}

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

Public Types

Public Member Functions

Static Public Member Functions

Static Public Attributes

Protected Member Functions

Detailed Description

Member Enumeration Documentation

◆ AtomicMemSpace

◆ AtomicOperation

Constructor & Destructor Documentation

◆ BackendSIMT()

Member Function Documentation

◆ addDeviceType()

◆ addPresynapticUpdateStrategy()

◆ areSharedMemAtomicsSlow()

◆ genCustomTransposeUpdateWUKernel()

◆ genCustomUpdateKernel()

◆ genCustomUpdateWUKernel()

◆ genDenseSynapseVariableRowInit()

◆ genGlobalRNGSkipAhead()

◆ genInitializeKernel()

◆ genInitializeSparseKernel()

◆ genKernelCustomUpdateVariableInit()

◆ genKernelSynapseVariableInit()

◆ genNeuronPrevSpikeTimeUpdateKernel()

◆ genNeuronSpikeQueueUpdateKernel()

◆ genNeuronUpdateKernel()

◆ genPopulationRNGInit()

◆ genPopulationRNGPostamble()

◆ genPopulationRNGPreamble()

◆ genPopVariableInit()

◆ genPostsynapticUpdateKernel()

◆ genPresynapticUpdateKernel()

◆ genSharedMemBarrier()

◆ genSparseSynapseVariableRowInit()

◆ genSynapseDendriticDelayUpdateKernel()

◆ genSynapseDynamicsKernel()

◆ genVariableInit()

◆ getAtomic()

◆ getBlockID()

◆ getCLZ()

◆ getDeviceVarPrefix()

◆ getKernelBlockSize() [1/2]

◆ getKernelBlockSize() [2/2]

◆ getNumConnectivityInitThreads()

◆ getNumInitialisationRNGStreams()

◆ getNumInitThreads() [1/2]

◆ getNumInitThreads() [2/2]

◆ getNumPostsynapticUpdateThreads()

◆ getNumPresynapticUpdateThreads()

◆ getNumSynapseDynamicsThreads()

◆ getPaddedNumCustomUpdateThreads()

◆ getPaddedNumCustomUpdateTransposeWUThreads()

◆ getPaddedNumCustomUpdateWUThreads()

◆ getSharedPrefix()

◆ getSynapticMatrixRowStride()

◆ getThreadID()

◆ isDeviceScalarRequired()

◆ isDeviceType()

◆ isGlobalDeviceRNGRequired()

◆ isGlobalHostRNGRequired()

◆ isPopulationRNGRequired()

◆ isPostsynapticRemapRequired()

◆ padKernelSize()

Member Data Documentation

◆ KernelNames