gpu_opencl.c File Reference

#include "include/gpu_opencl.h"

Include dependency graph for gpu_opencl.c:

Functions
void	init_opencl_devices (TREE_RECORD *tree_record, TREE_PARAMETERS *params)
void	free_opencl_devices (TREE_RECORD *tree_record, TREE_PARAMETERS *params)
void	free_train_buffers_gpu (TREE_RECORD *tree_record, TREE_PARAMETERS *params)
void	free_query_buffers_gpu (TREE_RECORD *tree_record, TREE_PARAMETERS *params)
void	allocate_memory_opencl_devices (TREE_RECORD *tree_record, TREE_PARAMETERS *params)
void	process_all_buffers_gpu (TREE_RECORD *tree_record, TREE_PARAMETERS *params)
void	process_buffers_brute_force_gpu (TREE_RECORD *tree_record, TREE_PARAMETERS *params, INT_TYPE all_brute)
INT_TYPE	retrieve_indices_from_buffers_gpu (TREE_RECORD *tree_record, TREE_PARAMETERS *params, INT_TYPE all_brute, INT_TYPE *tindices_removed, INT_TYPE *fr_indices, INT_TYPE *to_indices)
void	process_buffers_brute_force_in_chunks_gpu (TREE_RECORD *tree_record, TREE_PARAMETERS *params, INT_TYPE all_brute, INT_TYPE *tindices_removed, INT_TYPE n_tindices_removed, INT_TYPE *fr_indices, INT_TYPE *to_indices)
void	do_brute_force_all_leaves_FIRST_gpu (INT_TYPE *test_indices, INT_TYPE n_test_indices, INT_TYPE *fr_indices, INT_TYPE *to_indices, TREE_RECORD *tree_record, TREE_PARAMETERS *params, INT_TYPE n_device_train_patterns, cl_mem device_train_patterns, INT_TYPE chunk_offset, INT_TYPE all_brute, INT_TYPE current_chunk)
void	do_brute_force_all_leaves_SECOND_gpu (INT_TYPE *test_indices, INT_TYPE n_test_indices, INT_TYPE *fr_indices, INT_TYPE *to_indices, TREE_RECORD *tree_record, TREE_PARAMETERS *params, INT_TYPE n_device_train_patterns, cl_mem device_train_patterns, INT_TYPE chunk_offset, INT_TYPE all_brute, INT_TYPE current_chunk)
void	find_leaf_idx_batch_gpu (INT_TYPE *all_next_indices, INT_TYPE num_all_next_indices, INT_TYPE *ret_vals, TREE_RECORD *tree_record, TREE_PARAMETERS *params)
void	get_distances_and_indices_gpu (TREE_RECORD *tree_record, TREE_PARAMETERS *params)
void	write_sorted_training_patterns_gpu (TREE_RECORD *tree_record, TREE_PARAMETERS *params)
void	init_train_patterns_buffers (TREE_RECORD *tree_record, TREE_PARAMETERS *params, INT_TYPE chunk, INT_TYPE n_indices)
void	copy_train_patterns_to_device (TREE_RECORD *tree_record, TREE_PARAMETERS *params, INT_TYPE chunk, INT_TYPE start_idx, INT_TYPE end_idx)
void	free_train_patterns_device (TREE_RECORD *tree_record, TREE_PARAMETERS *params, INT_TYPE chunk)
int	training_chunks_inactive (TREE_RECORD *tree_record, TREE_PARAMETERS *params)

Function Documentation

void allocate_memory_opencl_devices	(	TREE_RECORD *	tree_record,
		TREE_PARAMETERS *	params
	)

Allocates memory for testing phase.

Parameters

*tree_record	Pointer to struct instance storing the model
*params	Pointer to struct instance storing all model parameters

void copy_train_patterns_to_device	(	TREE_RECORD *	tree_record,
		TREE_PARAMETERS *	params,
		INT_TYPE	chunk,
		INT_TYPE	start_idx,
		INT_TYPE	end_idx
	)

Parameters

*tree_record	Pointer to struct instance storing the model
*params	Pointer to struct instance storing all model parameters
chunk	Specifies the chunk 0 or 1
start_idx	The start index
end_idx	The end index

void do_brute_force_all_leaves_FIRST_gpu	(	INT_TYPE *	test_indices,
		INT_TYPE	n_test_indices,
		INT_TYPE *	fr_indices,
		INT_TYPE *	to_indices,
		TREE_RECORD *	tree_record,
		TREE_PARAMETERS *	params,
		INT_TYPE	n_device_train_patterns,
		cl_mem	device_train_patterns,
		INT_TYPE	chunk_offset,
		INT_TYPE	all_brute,
		INT_TYPE	current_chunk
	)

Apply brute-force approach for all leaves (first stage)

Parameters

*test_indices	Array of test indices to be processed
n_test_indices	Number of test indices
*fr_indices	Array of "from" indices, one for each test index
*to_indices	Array of "to" indices, one for each test index
*tree_record	Pointer to struct instance storing the model
*params	Pointer to struct instance storing all model parameters
n_device_train_patterns	Number of training patterns on device
device_train_patterns	Pointer to training patterns
chunk_offset	Offset (for indices) for the given chunk
all_brute	Flag specififying if all indices should be processed completely via brute force
current_chunk	The current chunk (id)

void do_brute_force_all_leaves_SECOND_gpu	(	INT_TYPE *	test_indices,
		INT_TYPE	n_test_indices,
		INT_TYPE *	fr_indices,
		INT_TYPE *	to_indices,
		TREE_RECORD *	tree_record,
		TREE_PARAMETERS *	params,
		INT_TYPE	n_device_train_patterns,
		cl_mem	device_train_patterns,
		INT_TYPE	chunk_offset,
		INT_TYPE	all_brute,
		INT_TYPE	current_chunk
	)

Second stage of the brute-force processing

Parameters

*test_indices	Array of test indices to be processed
n_test_indices	Number of test indices
*fr_indices	Array of "from" indices, one for each test index
*to_indices	Array of "to" indices, one for each test index
*tree_record	Pointer to struct instance storing the model
*params	Pointer to struct instance storing all model parameters
n_device_train_patterns	Number of training patterns on device
device_train_patterns	Pointer to training patterns
chunk_offset	Offset (for indices) for the given chunk
all_brute	Flag specififying if all indices should be processed completely via brute force
current_chunk	The current chunk (id)

void find_leaf_idx_batch_gpu	(	INT_TYPE *	all_next_indices,
		INT_TYPE	num_all_next_indices,
		INT_TYPE *	ret_vals,
		TREE_RECORD *	tree_record,
		TREE_PARAMETERS *	params
	)

Finds the next leaf indices for all test patterns indexed by all_next_indices.

Parameters

*all_next_indices	Array containing all indices that need to be processed next
num_all_next_indices	Number of indices
*ret_vals	Array containing the next leaf ids for each index
*tree_record	Pointer to struct instance storing the model
*params	Pointer to struct instance storing all model parameters

void free_opencl_devices	(	TREE_RECORD *	tree_record,
		TREE_PARAMETERS *	params
	)

After having performed all queries: Free memory etc.

Parameters

*tree_record	Pointer to struct instance storing the model
*params	Pointer to struct instance storing all model parameters

void free_query_buffers_gpu	(	TREE_RECORD *	tree_record,
		TREE_PARAMETERS *	params
	)

Free buffers needed for querying phase.

Parameters

*tree_record	Pointer to struct instance storing the model
*params	Pointer to struct instance storing all model parameters

void free_train_buffers_gpu	(	TREE_RECORD *	tree_record,
		TREE_PARAMETERS *	params
	)

Free buffers used during training.

Parameters

*tree_record	Pointer to struct instance storing the model
*params	Pointer to struct instance storing all model parameters

void free_train_patterns_device	(	TREE_RECORD *	tree_record,
		TREE_PARAMETERS *	params,
		INT_TYPE	chunk
	)

Releases training patterns on device

Parameters

*tree_record	Pointer to struct instance storing the model
*params	Pointer to struct instance storing all model parameters
chunk	Specifies the chunk 0 or 1

void get_distances_and_indices_gpu	(	TREE_RECORD *	tree_record,
		TREE_PARAMETERS *	params
	)

Copies the arrays dist_min_global and idx_min_global from GPU to CPU Updates the distances and indices (w.r.t the original indices)

Parameters

*tree_record	Pointer to struct instance storing the model
*params	Pointer to struct instance storing all model parameters

void init_opencl_devices	(	TREE_RECORD *	tree_record,
		TREE_PARAMETERS *	params
	)

Initializes all devices at the beginning of the querying process.

Parameters

*tree_record	Pointer to struct instance storing the model
*params	Pointer to struct instance storing all model parameters

void init_train_patterns_buffers	(	TREE_RECORD *	tree_record,
		TREE_PARAMETERS *	params,
		INT_TYPE	chunk,
		INT_TYPE	n_indices
	)

Parameters

*tree_record	Pointer to struct instance storing the model
*params	Pointer to struct instance storing all model parameters
chunk	Specifies the chunk 0 or 1
n_indices	Number of indices for which space shall be allocated

void process_all_buffers_gpu	(	TREE_RECORD *	tree_record,
		TREE_PARAMETERS *	params
	)

Processes all buffers on the GPU. It is important that only INDICES are moved from the CPU to the GPU (only CPU->GPU; GPU->CPU is not necessary). Further, the global distances and the indices are updated ON THE GPU.

Parameters

*tree_record	Pointer to struct instance storing the model
*params	Pointer to struct instance storing all model parameters

void process_buffers_brute_force_gpu	(	TREE_RECORD *	tree_record,
		TREE_PARAMETERS *	params,
		INT_TYPE	all_brute
	)

Helper function that processed all buffers via different methods, specified via the flag 'all_brute'.

Parameters

*tree_record	Pointer to struct instance storing the model
*params	Pointer to struct instance storing all model parameters
all_brute

void process_buffers_brute_force_in_chunks_gpu	(	TREE_RECORD *	tree_record,
		TREE_PARAMETERS *	params,
		INT_TYPE	all_brute,
		INT_TYPE *	tindices_removed,
		INT_TYPE	n_tindices_removed,
		INT_TYPE *	fr_indices,
		INT_TYPE *	to_indices
	)

Processes all indices stored in all buffers

Parameters

*tree_record	Pointer to struct instance storing the model
*params	Pointer to struct instance storing all model parameters
all_brute	Flag specififying if all indices should be processed completely via brute force
*tindices_removed	Array containing the test indices afterwards
n_tindices_removed	Number of indices
*fr_indices	Array of "from" indices
*to_indices	Array of "to" indices

INT_TYPE retrieve_indices_from_buffers_gpu	(	TREE_RECORD *	tree_record,
		TREE_PARAMETERS *	params,
		INT_TYPE	all_brute,
		INT_TYPE *	tindices_removed,
		INT_TYPE *	fr_indices,
		INT_TYPE *	to_indices
	)

Retrieves all indices from all buffers

Parameters

*tree_record	Pointer to struct instance storing the model
*params	Pointer to struct instance storing all model parameters
all_brute	Flag specififying if all indices should be processed completely via brute force
*tindices_removed	Array containing the test indices afterwards
*fr_indices	Array of "from" indices
*to_indices	Array of "to" indices

int training_chunks_inactive	(	TREE_RECORD *	tree_record,
		TREE_PARAMETERS *	params
	)

Helper method that checks of training chunks are used or not

Parameters

*tree_record	Pointer to struct instance storing the model
*params	Pointer to struct instance storing all model parameters

void write_sorted_training_patterns_gpu	(	TREE_RECORD *	tree_record,
		TREE_PARAMETERS *	params
	)

Writes the training patterns in a specific ordering (needed for coalesced data access on the GPU etc.

Parameters

*tree_record	Pointer to struct instance storing the model
*params	Pointer to struct instance storing all model parameters