kernels.cu File Reference

#include <cuda.h>
#include <cuda_runtime.h>

Functions
__global__ void	forwardPassV2Kernel (int step, int sequenceOffset, float *activity, float *input, float *weight, int numNeurons, int numIONeurons, float *buffer)
	Forward pass.
__global__ void	forwardPassV21Kernel (int step, int sequenceOffset, float *activity, float *input, float *buffer, float *potential, float *weight, float *previousPotential, float *error, int *deltaT, int numNeurons, int numIONeurons)
	Forward pass.
__global__ void	forwardPassV1Kernel (int step, int sequenceOffset, float *activity, float *input, float *weight, float *previousPotential, float *error, float *potential, int *deltaT, int numNeurons, int numIONeurons)
	Forward pass.
__global__ void	backwardPassV1Kernel (int step, int sequenceOffset, int numNeurons, int numIONeurons, float *input, float *activity, float *delta, float *deltaWeight, float *previousDelta, float *error, float *individualError, int *deltaT, float *weight)
	Calculates deltas, deltas on weights and errors parts.
__global__ void	backwardPassV11Kernel (int step, int numNeurons, int numIONeurons, float *activity, float *delta, float *previousDelta, int *deltaT, float *weight)
	Calculates deltas on hidden neurons.
__global__ void	backwardPassV2Kernel (int step, int sequenceOffset, int numNeurons, int numIONeurons, float *input, float *activity, float *delta, float *deltaWeight, float *previousDelta, float *error, float *individualError, int *deltaT, float *weight, float *buffer)
	Loads the buffer with deltas on weights fractions calculate deltas, deltas on weights and errors parts.
__global__ void	backwardPassV21Kernel (float *input, float *output, int numNeurons, int numIONeurons)
	Calculates deltas on hidden neurons.
__global__ void	backwardPassV3Kernel (int step, int numNeurons, int numIONeurons, float *activity, float *delta, float *previousDelta, float *deltaWeight, int *deltaT, float *weight)
	Calculates deltas on weights on hidden neurons and biases.
__global__ void	updateWeightsKernel (float learningRate, float momentum, float *weight, float *deltaWeight, float *previousDeltaWeight, int numWeights)
	Updates weights.
__global__ void	sumDeltaWeightsP2PKernel (int numWeights, float *masterDeltaWeight, float *peerDeltaWeight)
	Sums delta weights on the master device.
__global__ void	updateWeightsP2PKernel (int numWeights, float learningRate, float momentum, float *masterWeight, float *peerWeight, float *deltaWeight, float *previousDeltaWeight)
	Modifies weights on the master device and copies to the peer device.
__global__ void	sumErrorP2PKernel (float *masterError, float *peerError)
	Modifies weights on the master device and copies to the peer device.
__global__ void	setInitStatesKernel (float initState, float *activity, int numNeurons, int numIONeurons, int numFastNeurons)
	Sets the initial states for all the units on device.
__global__ void	resetParametersKernel (int numNeurons, int maxSequenceSteps, float *delta, float *previousDelta, float *potential, float *previousPotential, float *error)
	Resets delta and error parameters.
__global__ void	resetDeltaWeightsKernel (int numWeights, int numIONeurons, float *deltaWeight, float *individualError)
	Resets delta weights and errors.
template<unsigned int blockSize>
__global__ void	reduceKernel (float *input, float *output, unsigned int n, bool nIsPow2)
	Parallel reduction sum modified from NVIDIA SDK.
void	resetDeltaWeightsOnDevice (dim3 grid, dim3 block, cudaStream_t stream, int numWeights, int numIONeurons, float *deltaWeight, float *individualError)
	Wrapper for resetDeltaWeightsKernel.
void	setInitStatesOnDevice (dim3 grid, dim3 block, cudaStream_t stream, float initState, float *activity, int numNeurons, int numIONeurons, int numFastNeurons)
	Wrapper for setInitStatesKernel.
void	resetParametersOnDevice (dim3 grid, dim3 block, cudaStream_t stream, int numNeurons, int maxSequenceSteps, float *delta, float *previousDelta, float *potential, float *previousPotential, float *error)
	Wrapper for resetParametersKernel.
void	updateWeightsOnDevice (dim3 grid, dim3 block, float learningRate, float momentum, float *weight, float *deltaWeight, float *previousDeltaWeight, int numWeights)
	Wrapper for updateWeightsKernel.
void	forwardPassV1onDevice (dim3 grid, dim3 block, cudaStream_t stream, int step, int sequenceOffset, float *activity, float *input, float *weight, float *previousPotential, float *error, float *potential, int *deltaT, int numNeurons, int numIONeurons)
	Wrapper for forwardPassV1Kernel.
void	forwardPassV2onDevice (dim3 grid, dim3 block, cudaStream_t stream, int step, int sequenceOffset, float *activity, float *input, float *weight, int numNeurons, int numIONeurons, float *buffer)
	Wrapper for forwardPassV2Kernel.
void	forwardPassV21onDevice (dim3 grid, dim3 block, int smemSize, cudaStream_t stream, int step, int sequenceOffset, float *activity, float *input, float *buffer, float *potential, float *weight, float *previousPotential, float *error, int *deltaT, int numNeurons, int numIONeurons)
	Wrapper for forwardPassV21Kernel.
void	backwardPassV1onDevice (dim3 grid, dim3 block, cudaStream_t stream, int step, int sequenceOffset, int numNeurons, int numIONeurons, float *input, float *activity, float *delta, float *deltaWeight, float *previousDelta, float *error, float *individualError, int *deltaT, float *weight)
	Wrapper for backwardPassV1Kernel.
void	backwardPassV11onDevice (dim3 grid, dim3 block, cudaStream_t stream, int step, int numNeurons, int numIONeurons, float *activity, float *delta, float *previousDelta, int *deltaT, float *weight)
	Wrapper for backwardPassV11Kernel.
void	backwardPassV2onDevice (dim3 grid, dim3 block, cudaStream_t stream, int step, int sequenceOffset, int numNeurons, int numIONeurons, float *input, float *activity, float *delta, float *deltaWeight, float *previousDelta, float *error, float *individualError, int *deltaT, float *weight, float *buffer)
	Wrapper for backwardPassV2Kernel.
void	backwardPassV21onDevice (dim3 grid, dim3 block, int smemSize, cudaStream_t stream, float *input, float *output, int numNeurons, int numIONeurons)
	Wrapper for backwardPassV21Kernel.
void	backwardPassV3onDevice (dim3 grid, dim3 block, cudaStream_t stream, int step, int numNeurons, int numIONeurons, float *activity, float *delta, float *previousDelta, float *deltaWeight, int *deltaT, float *weight)
	Wrapper for backwardPassV3Kernel.
void	reduceOnDevice (int size, dim3 grid, dim3 block, int smemSize, cudaStream_t stream, float *input, float *output, unsigned int n, bool nIsPow2)
	Wrapper for reduceKernel.
void	sumDeltaWeightsP2PonDevice (dim3 grid, dim3 block, int numWeights, float *masterDeltaWeight, float *peerDeltaWeight)
	Wrapper for sumDeltaWeightsP2PKernel.
void	updateWeightsP2PonDevice (dim3 grid, dim3 block, int numWeights, float learningRate, float momentum, float *masterWeight, float *peerWeight, float *deltaWeight, float *previousDeltaWeight)
	Wrapper for updateWeightsP2PKernel.
void	sumErrorP2PonDevice (dim3 grid, dim3 block, float *masterError, float *peerError)
	Wrapper for sumErrorP2PKernel.

Function Documentation

__global__ void backwardPassV11Kernel	(	int	step,
		int	numNeurons,
		int	numIONeurons,
		float *	activity,
		float *	delta,
		float *	previousDelta,
		int *	deltaT,
		float *	weight
	)

Calculates deltas on hidden neurons.

Note

Slower version for larger networks over 1024 neurons.

Parameters

[in]	step	- current step
[in]	numNeurons	- number of neurons
[in]	numIONeurons	- number of input-ouput neurons
[in]	activity	- activities
[in]	deltaT	- delta-t values
[in]	weight	- weights
[in]	previousDelta	- previous deltas
[out]	delta	- deltas

void backwardPassV11onDevice	(	dim3	grid,
		dim3	block,
		cudaStream_t	stream,
		int	step,
		int	numNeurons,
		int	numIONeurons,
		float *	activity,
		float *	delta,
		float *	previousDelta,
		int *	deltaT,
		float *	weight
	)

Wrapper for backwardPassV11Kernel.

Parameters

[in]	grid	- CUDA grid size
[in]	block	- CUDA block size
[in]	stream	- CUDA stream
[in]	step	- current step
[in]	numNeurons	- number of neurons
[in]	numIONeurons	- number of input-ouput neurons
[in]	activity	- activities
[in]	deltaT	- delta-t values
[in]	weight	- weights
[in]	previousDelta	- previous deltas
[out]	delta	- deltas

__global__ void backwardPassV1Kernel	(	int	step,
		int	sequenceOffset,
		int	numNeurons,
		int	numIONeurons,
		float *	input,
		float *	activity,
		float *	delta,
		float *	deltaWeight,
		float *	previousDelta,
		float *	error,
		float *	individualError,
		int *	deltaT,
		float *	weight
	)

Calculates deltas, deltas on weights and errors parts.

Note

Slower version for larger networks over 1024 neurons.

Parameters

[in]	step	- current step
[in]	sequenceOffset	- sequence offsets
[in]	numNeurons	- number of neurons
[in]	numIONeurons	- number of input-output neurons
[in]	input	- input
[in]	activity	- activities
[in]	error	- errors
[in]	individualError	- error buffer
[in]	deltaT	- delta-t values
[in]	weight	- weights
[in]	previousDelta	- previous deltas
[out]	delta	- deltas
[out]	deltaWeight	- delta weights

void backwardPassV1onDevice	(	dim3	grid,
		dim3	block,
		cudaStream_t	stream,
		int	step,
		int	sequenceOffset,
		int	numNeurons,
		int	numIONeurons,
		float *	input,
		float *	activity,
		float *	delta,
		float *	deltaWeight,
		float *	previousDelta,
		float *	error,
		float *	individualError,
		int *	deltaT,
		float *	weight
	)

Wrapper for backwardPassV1Kernel.

Parameters

[in]	grid	- CUDA grid size
[in]	block	- CUDA block size
[in]	stream	- CUDA stream
[in]	step	- current step
[in]	sequenceOffset	- sequence offsets
[in]	numNeurons	- number of neurons
[in]	numIONeurons	- number of input-output neurons
[in]	input	- input
[in]	activity	- activities
[in]	error	- errors
[in]	individualError	- error buffer
[in]	deltaT	- delta-t values
[in]	weight	- weights
[in]	previousDelta	- previous deltas
[out]	delta	- deltas
[out]	deltaWeight	- delta weights

__global__ void backwardPassV21Kernel	(	float *	input,
		float *	output,
		int	numNeurons,
		int	numIONeurons
	)

Calculates deltas on hidden neurons.

Note

Used together with backwardPassV2Kerne. Faster version for networks of up to 1024 neurons.

Parameters

[in]	input	- input
[in]	numNeurons	- number of neurons
[in]	numIONeurons	- number of input-output neurons
[out]	output	- output

void backwardPassV21onDevice	(	dim3	grid,
		dim3	block,
		int	smemSize,
		cudaStream_t	stream,
		float *	input,
		float *	output,
		int	numNeurons,
		int	numIONeurons
	)

Wrapper for backwardPassV21Kernel.

Parameters

[in]	grid	- CUDA grid size
[in]	block	- CUDA block size
[in]	smemSize	- CUDA shared memory size
[in]	stream	- CUDA stream
[in]	input	- input
[in]	numNeurons	- number of neurons
[in]	numIONeurons	- number of input-output neurons
[out]	output	- output

__global__ void backwardPassV2Kernel	(	int	step,
		int	sequenceOffset,
		int	numNeurons,
		int	numIONeurons,
		float *	input,
		float *	activity,
		float *	delta,
		float *	deltaWeight,
		float *	previousDelta,
		float *	error,
		float *	individualError,
		int *	deltaT,
		float *	weight,
		float *	buffer
	)

Loads the buffer with deltas on weights fractions calculate deltas, deltas on weights and errors parts.

Note

Deltas on weights and errors are later summed by parallel reduction.

Parameters

[in]	step	- current step
[in]	sequenceOffset	- sequence offsets
[in]	numNeurons	- number of neurons
[in]	numIONeurons	- number of input-output neurons
[in]	input	- input
[in]	activity	- activities
[in]	previousDelta	- previous deltas
[in]	error	- errors
[in]	individualError	- error buffer
[in]	deltaT	- delta-t values
[in]	weight	- weights
[out]	delta	- deltas
[out]	deltaWeight	- delta weights
[out]	buffer	- buffer used for storing delta weights

void backwardPassV2onDevice	(	dim3	grid,
		dim3	block,
		cudaStream_t	stream,
		int	step,
		int	sequenceOffset,
		int	numNeurons,
		int	numIONeurons,
		float *	input,
		float *	activity,
		float *	delta,
		float *	deltaWeight,
		float *	previousDelta,
		float *	error,
		float *	individualError,
		int *	deltaT,
		float *	weight,
		float *	buffer
	)

Wrapper for backwardPassV2Kernel.

Parameters

[in]	grid	- CUDA grid size
[in]	block	- CUDA block size
[in]	stream	- CUDA stream
[in]	step	- current step
[in]	sequenceOffset	- sequence offsets
[in]	numNeurons	- number of neurons
[in]	numIONeurons	- number of input-output neurons
[in]	input	- input
[in]	activity	- activities
[in]	previousDelta	- previous deltas
[in]	error	- errors
[in]	individualError	- error buffer
[in]	deltaT	- delta-t values
[in]	weight	- weights
[out]	delta	- deltas
[out]	deltaWeight	- delta weights
[out]	buffer	- buffer used for storing delta weights

__global__ void backwardPassV3Kernel	(	int	step,
		int	numNeurons,
		int	numIONeurons,
		float *	activity,
		float *	delta,
		float *	previousDelta,
		float *	deltaWeight,
		int *	deltaT,
		float *	weight
	)

Calculates deltas on weights on hidden neurons and biases.

Parameters

[in]	step	- current step
[in]	numNeurons	- number of neurons
[in]	numIONeurons	- number of input-output neurons
[in]	activity	- activities
[in]	deltaT	- delta-t values
[in]	weight	- weights
[in]	delta	- deltas
[out]	previousDelta	- previous deltas
[out]	deltaWeight	- delta weights

void backwardPassV3onDevice	(	dim3	grid,
		dim3	block,
		cudaStream_t	stream,
		int	step,
		int	numNeurons,
		int	numIONeurons,
		float *	activity,
		float *	delta,
		float *	previousDelta,
		float *	deltaWeight,
		int *	deltaT,
		float *	weight
	)

Wrapper for backwardPassV3Kernel.

Parameters

[in]	grid	- CUDA grid size
[in]	block	- CUDA block size
[in]	stream	- CUDA stream
[in]	step	- current step
[in]	numNeurons	- number of neurons
[in]	numIONeurons	- number of input-output neurons
[in]	activity	- activities
[in]	deltaT	- delta-t values
[in]	weight	- weights
[in]	delta	- deltas
[out]	previousDelta	- previous deltas
[out]	deltaWeight	- delta weights

__global__ void forwardPassV1Kernel	(	int	step,
		int	sequenceOffset,
		float *	activity,
		float *	input,
		float *	weight,
		float *	previousPotential,
		float *	error,
		float *	potential,
		int *	deltaT,
		int	numNeurons,
		int	numIONeurons
	)

Forward pass.

Note

Slower version for larger networks.

Parameters

[in]	step	- current step
[in]	sequenceOffset	- sequence offsets
[in]	input	- input
[in]	weight	- weights
[in]	deltaT	- delta-t values
[in]	numNeurons	- number of neurons
[in]	numIONeurons	- number of input-output neurons
[out]	potential	- potentials
[out]	previousPotential	- previous potentials
[out]	activity	- activities
[out]	error	- errors

void forwardPassV1onDevice	(	dim3	grid,
		dim3	block,
		cudaStream_t	stream,
		int	step,
		int	sequenceOffset,
		float *	activity,
		float *	input,
		float *	weight,
		float *	previousPotential,
		float *	error,
		float *	potential,
		int *	deltaT,
		int	numNeurons,
		int	numIONeurons
	)

Wrapper for forwardPassV1Kernel.

Parameters

[in]	grid	- CUDA grid size
[in]	block	- CUDA block size
[in]	stream	- CUDA stream
[in]	step	- current step
[in]	sequenceOffset	- sequence offsets
[in]	input	- input
[in]	weight	- weights
[in]	deltaT	- delta-t values
[in]	numNeurons	- number of neurons
[in]	numIONeurons	- number of input-output neurons
[out]	potential	- potentials
[out]	previousPotential	- previous potentials
[out]	activity	- activities
[out]	error	- errors

__global__ void forwardPassV21Kernel	(	int	step,
		int	sequenceOffset,
		float *	activity,
		float *	input,
		float *	buffer,
		float *	potential,
		float *	weight,
		float *	previousPotential,
		float *	error,
		int *	deltaT,
		int	numNeurons,
		int	numIONeurons
	)

Forward pass.

Note

Used with forwardPassV2Kernel. Faster version for networks of up to 1024 neurons.

Parameters

[in]	step	- current step
[in]	sequenceOffset	- sequence offsets
[in]	input	- input
[in]	weight	- weights
[in]	deltaT	- delta-t values
[in]	numNeurons	- number of neurons
[in]	numIONeurons	- number of input-output neurons
[out]	potential	- potentials
[out]	previousPotential	- previous potentials
[out]	activity	- activities
[out]	error	- errors

void forwardPassV21onDevice	(	dim3	grid,
		dim3	block,
		int	smemSize,
		cudaStream_t	stream,
		int	step,
		int	sequenceOffset,
		float *	activity,
		float *	input,
		float *	buffer,
		float *	potential,
		float *	weight,
		float *	previousPotential,
		float *	error,
		int *	deltaT,
		int	numNeurons,
		int	numIONeurons
	)

Wrapper for forwardPassV21Kernel.

Parameters

[in]	grid	- CUDA grid size
[in]	block	- CUDA block size
[in]	smemSize	- CUDA shared memory size
[in]	stream	- CUDA stream
[in]	step	- current step
[in]	sequenceOffset	- sequence offsets
[in]	activity	- activations
[in]	input	- input
[in]	weight	- weights
[in]	numNeurons	- number of neurons
[in]	numIONeurons	- number of input-output neurons
[out]	buffer	- buffer used for storing new activations

__global__ void forwardPassV2Kernel	(	int	step,
		int	sequenceOffset,
		float *	activity,
		float *	input,
		float *	weight,
		int	numNeurons,
		int	numIONeurons,
		float *	buffer
	)

Forward pass.

Note

Used with forwardPassV21Kernel - faster version for networks of up to 1024 neurons.

Parameters

[in]	step	- current step
[in]	sequenceOffset	- sequence offsets
[in]	activity	- activations
[in]	input	- input
[in]	weight	- weights
[in]	numNeurons	- number of neurons
[in]	numIONeurons	- number of input-output neurons
[out]	buffer	- buffer used for storing new activations

void forwardPassV2onDevice	(	dim3	grid,
		dim3	block,
		cudaStream_t	stream,
		int	step,
		int	sequenceOffset,
		float *	activity,
		float *	input,
		float *	weight,
		int	numNeurons,
		int	numIONeurons,
		float *	buffer
	)

Wrapper for forwardPassV2Kernel.

Parameters

[in]	grid	- CUDA grid size
[in]	block	- CUDA block size
[in]	stream	- CUDA stream
[in]	step	- current step
[in]	sequenceOffset	- sequence offsets
[in]	activity	- activations
[in]	input	- input
[in]	weight	- weights
[in]	numNeurons	- number of neurons
[in]	numIONeurons	- number of input-output neurons
[out]	buffer	- buffer used for storing new activations

template<unsigned int blockSize>

__global__ void reduceKernel	(	float *	input,
		float *	output,
		unsigned int	n,
		bool	nIsPow2
	)

Parallel reduction sum modified from NVIDIA SDK.

Note

Number of threads are not known at a compile time, however, we always stick to power of 2 sizes

so here we are using templates to allow compilation for all known size, which results in higher throughput.

Parameters

[in]	input	- input
[in]	n	- number of elements to sum
[in]	nIsPow2	- determines if the number is of power of two
[out]	output	- output

void reduceOnDevice	(	int	size,
		dim3	grid,
		dim3	block,
		int	smemSize,
		cudaStream_t	stream,
		float *	input,
		float *	output,
		unsigned int	n,
		bool	nIsPow2
	)

Wrapper for reduceKernel.

Parameters

[in]	size	- number of elements to sum
[in]	grid	- CUDA grid size
[in]	block	- CUDA block size
[in]	smemSize	- CUDA shared memory size
[in]	stream	- CUDA stream
[in]	input	- input
[in]	n	- number of elements to sum
[in]	nIsPow2	- determines if the number is of power of two
[out]	output	- output

__global__ void resetDeltaWeightsKernel	(	int	numWeights,
		int	numIONeurons,
		float *	deltaWeight,
		float *	individualError
	)

Resets delta weights and errors.

Parameters

[in]	numWeights	- number of weights
[in]	numIONeurons	- number of input-output neurons
[out]	deltaWeight	- delta weights
[out]	individualError	- error buffer

void resetDeltaWeightsOnDevice	(	dim3	grid,
		dim3	block,
		cudaStream_t	stream,
		int	numWeights,
		int	numIONeurons,
		float *	deltaWeight,
		float *	individualError
	)

Wrapper for resetDeltaWeightsKernel.

Parameters

[in]	grid	- CUDA grid size
[in]	block	- CUDA block size
[in]	stream	- CUDA stream
[in]	numWeights	- number of weights
[in]	numIONeurons	- number of input-output neurons
[out]	deltaWeight	- delta weights
[out]	individualError	- error buffer

__global__ void resetParametersKernel	(	int	numNeurons,
		int	maxSequenceSteps,
		float *	delta,
		float *	previousDelta,
		float *	potential,
		float *	previousPotential,
		float *	error
	)

Resets delta and error parameters.

Parameters

[in]	numNeurons	- number of neurons
[in]	maxSequenceSteps	- maximum number of sequence steps
[out]	delta	- deltas
[out]	previousDelta	- previous deltas
[out]	potential	- potentials
[out]	previousPotential	- previous potentials
[out]	error	- errors

void resetParametersOnDevice	(	dim3	grid,
		dim3	block,
		cudaStream_t	stream,
		int	numNeurons,
		int	maxSequenceSteps,
		float *	delta,
		float *	previousDelta,
		float *	potential,
		float *	previousPotential,
		float *	error
	)

Wrapper for resetParametersKernel.

Parameters

[in]	grid	- CUDA grid size
[in]	block	- CUDA block size
[in]	stream	- CUDA stream
[in]	numNeurons	- number of neurons
[in]	maxSequenceSteps	- maximum number of sequence steps
[out]	delta	- deltas
[out]	previousDelta	- previous deltas
[out]	potential	- potentials
[out]	previousPotential	- previous potentials
[out]	error	- errors

__global__ void setInitStatesKernel	(	float	initState,
		float *	activity,
		int	numNeurons,
		int	numIONeurons,
		int	numFastNeurons
	)

Sets the initial states for all the units on device.

Note

Slow context units will be initialised with a specific value.

Parameters

[in]	initState	- initial state
[in]	numNeurons	- number of neurons
[in]	numIONeurons	- number of input-output neurons
[in]	numFastNeurons	- number of fast neurons
[out]	activity	- activities

void setInitStatesOnDevice	(	dim3	grid,
		dim3	block,
		cudaStream_t	stream,
		float	initState,
		float *	activity,
		int	numNeurons,
		int	numIONeurons,
		int	numFastNeurons
	)

Wrapper for setInitStatesKernel.

Parameters

[in]	grid	- CUDA grid size
[in]	block	- CUDA block size
[in]	stream	- CUDA stream
[in]	initState	- initial state
[in]	numNeurons	- number of neurons
[in]	numIONeurons	- number of input-output neurons
[in]	numFastNeurons	- number of fast neurons
[out]	activity	- activities

__global__ void sumDeltaWeightsP2PKernel	(	int	numWeights,
		float *	masterDeltaWeight,
		float *	peerDeltaWeight
	)

Sums delta weights on the master device.

Parameters

[in]	numWeights	- number of weights
[in]	peerDeltaWeight	- delta weights from peer device
[out]	masterDeltaWeight	- delta weights from master device

void sumDeltaWeightsP2PonDevice	(	dim3	grid,
		dim3	block,
		int	numWeights,
		float *	masterDeltaWeight,
		float *	peerDeltaWeight
	)

Wrapper for sumDeltaWeightsP2PKernel.

Parameters

[in]	grid	- CUDA grid size
[in]	block	- CUDA block size
[in]	numWeights	- number of weights
[in]	peerDeltaWeight	- delta weights from peer device
[out]	masterDeltaWeight	- delta weights from master device

__global__ void sumErrorP2PKernel	(	float *	masterError,
		float *	peerError
	)

Modifies weights on the master device and copies to the peer device.

Parameters

[in]	peerError	- error from peer device
[out]	masterError	- error from master device

void sumErrorP2PonDevice	(	dim3	grid,
		dim3	block,
		float *	masterError,
		float *	peerError
	)

Wrapper for sumErrorP2PKernel.

Parameters

[in]	grid	- CUDA grid size
[in]	block	- CUDA block size
[in]	peerError	- error from peer device
[out]	masterError	- error from master device

__global__ void updateWeightsKernel	(	float	learningRate,
		float	momentum,
		float *	weight,
		float *	deltaWeight,
		float *	previousDeltaWeight,
		int	numWeights
	)

Updates weights.

Parameters

[in]	learningRate	- learning rate
[in]	momentum	- momentum
[in]	numWeights	- number of weights
[in]	deltaWeight	- delta weights
[out]	previousDeltaWeight	- previous delta weights
[out]	weight	- weights

void updateWeightsOnDevice	(	dim3	grid,
		dim3	block,
		float	learningRate,
		float	momentum,
		float *	weight,
		float *	deltaWeight,
		float *	previousDeltaWeight,
		int	numWeights
	)

Wrapper for updateWeightsKernel.

Parameters

[in]	grid	- CUDA grid size
[in]	block	- CUDA block size
[in]	learningRate	- learning rate
[in]	momentum	- momentum
[in]	numWeights	- number of weights
[in]	deltaWeight	- delta weights
[out]	previousDeltaWeight	- previous delta weights
[out]	weight	- weights

__global__ void updateWeightsP2PKernel	(	int	numWeights,
		float	learningRate,
		float	momentum,
		float *	masterWeight,
		float *	peerWeight,
		float *	deltaWeight,
		float *	previousDeltaWeight
	)

Modifies weights on the master device and copies to the peer device.

Parameters

[in]	numWeights	- number of weights
[in]	learningRate	- learning rate
[in]	momentum	- momentum
[in]	deltaWeight	- delta weights
[out]	previousDeltaWeight	- previous delta weights
[out]	masterWeight	- weigths from master device
[out]	peerWeight	- weights from peer device

void updateWeightsP2PonDevice	(	dim3	grid,
		dim3	block,
		int	numWeights,
		float	learningRate,
		float	momentum,
		float *	masterWeight,
		float *	peerWeight,
		float *	deltaWeight,
		float *	previousDeltaWeight
	)

Wrapper for updateWeightsP2PKernel.

Parameters

[in]	grid	- CUDA grid size
[in]	block	- CUDA block size
[in]	numWeights	- number of weights
[in]	learningRate	- learning rate
[in]	momentum	- momentum
[in]	deltaWeight	- delta weights
[out]	previousDeltaWeight	- previous delta weights
[out]	masterWeight	- weigths from master device
[out]	peerWeight	- weights from peer device