StarPU Handbook
|
Data Structures | |
struct | starpu_data_filter |
Basic API | |
void | starpu_data_partition (starpu_data_handle_t initial_handle, struct starpu_data_filter *f) |
void | starpu_data_unpartition (starpu_data_handle_t root_data, unsigned gathering_node) |
starpu_data_handle_t | starpu_data_get_child (starpu_data_handle_t handle, unsigned i) |
int | starpu_data_get_nb_children (starpu_data_handle_t handle) |
starpu_data_handle_t | starpu_data_get_sub_data (starpu_data_handle_t root_data, unsigned depth,...) |
starpu_data_handle_t | starpu_data_vget_sub_data (starpu_data_handle_t root_data, unsigned depth, va_list pa) |
void | starpu_data_map_filters (starpu_data_handle_t root_data, unsigned nfilters,...) |
void | starpu_data_vmap_filters (starpu_data_handle_t root_data, unsigned nfilters, va_list pa) |
Predefined BCSR Filter Functions | |
Predefined partitioning functions for BCSR data. Examples on how to use them are shown in Partitioning Data. | |
void | starpu_bcsr_filter_canonical_block (void *father_interface, void *child_interface, struct starpu_data_filter *f, unsigned id, unsigned nparts) |
unsigned | starpu_bcsr_filter_canonical_block_get_nchildren (struct starpu_data_filter *f, starpu_data_handle_t handle) |
struct starpu_data_interface_ops * | starpu_bcsr_filter_canonical_block_child_ops (struct starpu_data_filter *f, unsigned child) |
void | starpu_bcsr_filter_vertical_block (void *father_interface, void *child_interface, struct starpu_data_filter *f, unsigned id, unsigned nparts) |
Predefined CSR Filter Functions | |
Predefined partitioning functions for CSR data. Examples on how to use them are shown in Partitioning Data. | |
void | starpu_csr_filter_vertical_block (void *father_interface, void *child_interface, struct starpu_data_filter *f, unsigned id, unsigned nparts) |
Predefined Matrix Filter Functions | |
Predefined partitioning functions for matrix data. Examples on how to use them are shown in Partitioning Data. Note: this is using the C element order which is row-major, i.e. elements with consecutive x coordinates are consecutive in memory. | |
void | starpu_matrix_filter_block (void *father_interface, void *child_interface, struct starpu_data_filter *f, unsigned id, unsigned nparts) |
void | starpu_matrix_filter_block_shadow (void *father_interface, void *child_interface, struct starpu_data_filter *f, unsigned id, unsigned nparts) |
void | starpu_matrix_filter_vertical_block (void *father_interface, void *child_interface, struct starpu_data_filter *f, unsigned id, unsigned nparts) |
void | starpu_matrix_filter_vertical_block_shadow (void *father_interface, void *child_interface, struct starpu_data_filter *f, unsigned id, unsigned nparts) |
Predefined Vector Filter Functions | |
Predefined partitioning functions for vector data. Examples on how to use them are shown in Partitioning Data. | |
void | starpu_vector_filter_block (void *father_interface, void *child_interface, struct starpu_data_filter *f, unsigned id, unsigned nparts) |
void | starpu_vector_filter_block_shadow (void *father_interface, void *child_interface, struct starpu_data_filter *f, unsigned id, unsigned nparts) |
void | starpu_vector_filter_list_long (void *father_interface, void *child_interface, struct starpu_data_filter *f, unsigned id, unsigned nparts) |
void | starpu_vector_filter_list (void *father_interface, void *child_interface, struct starpu_data_filter *f, unsigned id, unsigned nparts) |
void | starpu_vector_filter_divide_in_2 (void *father_interface, void *child_interface, struct starpu_data_filter *f, unsigned id, unsigned nparts) |
Predefined Block Filter Functions | |
Predefined partitioning functions for block data. Examples on how to use them are shown in Partitioning Data. An example is available in | |
void | starpu_block_filter_block (void *father_interface, void *child_interface, struct starpu_data_filter *f, unsigned id, unsigned nparts) |
void | starpu_block_filter_block_shadow (void *father_interface, void *child_interface, struct starpu_data_filter *f, unsigned id, unsigned nparts) |
void | starpu_block_filter_vertical_block (void *father_interface, void *child_interface, struct starpu_data_filter *f, unsigned id, unsigned nparts) |
void | starpu_block_filter_vertical_block_shadow (void *father_interface, void *child_interface, struct starpu_data_filter *f, unsigned id, unsigned nparts) |
void | starpu_block_filter_depth_block (void *father_interface, void *child_interface, struct starpu_data_filter *f, unsigned id, unsigned nparts) |
void | starpu_block_filter_depth_block_shadow (void *father_interface, void *child_interface, struct starpu_data_filter *f, unsigned id, unsigned nparts) |
void | starpu_filter_nparts_compute_chunk_size_and_offset (unsigned n, unsigned nparts, size_t elemsize, unsigned id, unsigned ld, unsigned *chunk_size, size_t *offset) |
struct starpu_data_filter |
Describe a data partitioning operation, to be given to starpu_data_partition()
Data Fields | |
void(* | filter_func )(void *father_interface, void *child_interface, struct starpu_data_filter *, unsigned id, unsigned nparts) |
unsigned | nchildren |
unsigned(* | get_nchildren )(struct starpu_data_filter *, starpu_data_handle_t initial_handle) |
struct starpu_data_interface_ops *(* | get_child_ops )(struct starpu_data_filter *, unsigned id) |
unsigned | filter_arg |
void * | filter_arg_ptr |
void(* starpu_data_filter::filter_func) (void *father_interface, void *child_interface, struct starpu_data_filter *, unsigned id, unsigned nparts) |
Fill the child_interface
structure with interface information for the i
-th child of the parent father_interface
(among nparts
). The filter
structure is provided, allowing to inspect the starpu_data_filter::filter_arg and starpu_data_filter::filter_arg_ptr parameters. The details of what needs to be filled in child_interface
vary according to the data interface, but generally speaking:
id
is usually just copied over from the father, when the sub data has the same structure as the father, e.g. a subvector is a vector, a submatrix is a matrix, etc. This is however not the case for instance when dividing a BCSR matrix into its dense blocks, which then are matrices. nx
, ny
and alike are usually divided by the number of subdata, depending how the subdivision is done (e.g. nx division vs ny division for vertical matrix division vs horizontal matrix division). ld
for matrix interfaces are usually just copied over: the leading dimension (ld) usually does not change. elemsize
is usually just copied over. ptr
, the pointer to the data, has to be computed according to i
and the father's ptr
, so as to point to the start of the sub data. This should however be done only if the father has ptr
different from NULL: in the OpenCL case notably, the dev_handle
and offset
fields are used instead. dev_handle
should be just copied over from the parent. offset
has to be computed according to i
and the father's offset
, so as to provide the offset of the start of the sub data. This is notably used for the OpenCL case. unsigned starpu_data_filter::nchildren |
Number of parts to partition the data into.
unsigned(* starpu_data_filter::get_nchildren) (struct starpu_data_filter *, starpu_data_handle_t initial_handle) |
Return the number of children. This can be used instead of starpu_data_filter::nchildren when the number of children depends on the actual data (e.g. the number of blocks in a sparse matrix).
struct starpu_data_interface_ops*(* starpu_data_filter::get_child_ops) (struct starpu_data_filter *, unsigned id) |
When children use different data interface, return which interface is used by child number id
.
unsigned starpu_data_filter::filter_arg |
Additional parameter for the filter function
void* starpu_data_filter::filter_arg_ptr |
Additional pointer parameter for the filter function, such as the sizes of the different parts.
void starpu_data_partition | ( | starpu_data_handle_t | initial_handle, |
struct starpu_data_filter * | f | ||
) |
Request the partitioning of initial_handle
into several subdata according to the filter f
.
Here an example of how to use the function.
void starpu_data_unpartition | ( | starpu_data_handle_t | root_data, |
unsigned | gathering_node | ||
) |
Unapply the filter which has been applied to root_data
, thus unpartitioning the data. The pieces of data are collected back into one big piece in the gathering_node
(usually STARPU_MAIN_RAM). Tasks working on the partitioned data will be waited for by starpu_data_unpartition().
Here an example of how to use the function.
starpu_data_handle_t starpu_data_get_child | ( | starpu_data_handle_t | handle, |
unsigned | i | ||
) |
Return the i
-th child of the given handle
, which must have been partitionned beforehand.
int starpu_data_get_nb_children | ( | starpu_data_handle_t | handle | ) |
Return the number of children handle
has been partitioned into.
starpu_data_handle_t starpu_data_get_sub_data | ( | starpu_data_handle_t | root_data, |
unsigned | depth, | ||
... | |||
) |
After partitioning a StarPU data by applying a filter, starpu_data_get_sub_data() can be used to get handles for each of the data portions. root_data
is the parent data that was partitioned. depth
is the number of filters to traverse (in case several filters have been applied, to e.g. partition in row blocks, and then in column blocks), and the subsequent parameters are the indexes. The function returns a handle to the subdata.
Here an example of how to use the function.
starpu_data_handle_t starpu_data_vget_sub_data | ( | starpu_data_handle_t | root_data, |
unsigned | depth, | ||
va_list | pa | ||
) |
Similar to starpu_data_get_sub_data() but use a va_list
for the parameter list.
void starpu_data_map_filters | ( | starpu_data_handle_t | root_data, |
unsigned | nfilters, | ||
... | |||
) |
Apply nfilters
filters to the handle designated by root_handle
recursively. nfilters
pointers to variables of the type starpu_data_filter should be given.
void starpu_data_vmap_filters | ( | starpu_data_handle_t | root_data, |
unsigned | nfilters, | ||
va_list | pa | ||
) |
Apply nfilters
filters to the handle designated by root_handle
recursively. Use a va_list
of pointers to variables of the type starpu_data_filter.
void starpu_data_partition_plan | ( | starpu_data_handle_t | initial_handle, |
struct starpu_data_filter * | f, | ||
starpu_data_handle_t * | children | ||
) |
Plan to partition initial_handle
into several subdata according to the filter f
. The handles are returned into the children
array, which has to be the same size as the number of parts described in f
. These handles are not immediately usable, starpu_data_partition_submit() has to be called to submit the actual partitioning.
Here is an example of how to use the function:
void starpu_data_partition_submit | ( | starpu_data_handle_t | initial_handle, |
unsigned | nparts, | ||
starpu_data_handle_t * | children | ||
) |
Submit the actual partitioning of initial_handle
into the nparts
children
handles. This call is asynchronous, it only submits that the partitioning should be done, so that the children
handles can now be used to submit tasks, and initial_handle
can not be used to submit tasks any more (to guarantee coherency). For instance,
void starpu_data_partition_readonly_submit | ( | starpu_data_handle_t | initial_handle, |
unsigned | nparts, | ||
starpu_data_handle_t * | children | ||
) |
Similar to starpu_data_partition_submit(), but do not invalidate initial_handle
. This allows to continue using it, but the application has to be careful not to write to initial_handle
or children
handles, only read from them, since the coherency is otherwise not guaranteed. This thus allows to submit various tasks which concurrently read from various partitions of the data.
When the application wants to write to initial_handle
again, it should call starpu_data_unpartition_submit(), which will properly add dependencies between the reads on the children
and the writes to be submitted.
If instead the application wants to write to children
handles, it should call starpu_data_partition_readwrite_upgrade_submit(), which will correctly add dependencies between the reads on the initial_handle
and the writes to be submitted.
void starpu_data_partition_readwrite_upgrade_submit | ( | starpu_data_handle_t | initial_handle, |
unsigned | nparts, | ||
starpu_data_handle_t * | children | ||
) |
Assume that a partitioning of initial_handle
has already been submited in readonly mode through starpu_data_partition_readonly_submit(), and will upgrade that partitioning into read-write mode for the children
, by invalidating initial_handle
, and adding the necessary dependencies.
void starpu_data_unpartition_submit | ( | starpu_data_handle_t | initial_handle, |
unsigned | nparts, | ||
starpu_data_handle_t * | children, | ||
int | gathering_node | ||
) |
Assuming that initial_handle
is partitioned into children
, submit an unpartitionning of initial_handle
, i.e. submit a gathering of the pieces on the requested gathering_node
memory node, and submit an invalidation of the children.
void starpu_data_unpartition_readonly_submit | ( | starpu_data_handle_t | initial_handle, |
unsigned | nparts, | ||
starpu_data_handle_t * | children, | ||
int | gathering_node | ||
) |
Similar to starpu_data_partition_submit(), but do not invalidate initial_handle
. This allows to continue using it, but the application has to be careful not to write to initial_handle
or children
handles, only read from them, since the coherency is otherwise not guaranteed. This thus allows to submit various tasks which concurrently read from various partitions of the data.
void starpu_data_partition_clean | ( | starpu_data_handle_t | root_data, |
unsigned | nparts, | ||
starpu_data_handle_t * | children | ||
) |
Clear the partition planning established between root_data
and children
with starpu_data_partition_plan(). This will notably submit an unregister all the children
, which can thus not be used any more afterwards.
void starpu_data_unpartition_submit_sequential_consistency_cb | ( | starpu_data_handle_t | initial_handle, |
unsigned | nparts, | ||
starpu_data_handle_t * | children, | ||
int | gather_node, | ||
int | sequential_consistency, | ||
void(*)(void *) | callback_func, | ||
void * | callback_arg | ||
) |
Similar to starpu_data_unpartition_submit_sequential_consistency() but allow to specify a callback function for the unpartitiong task
void starpu_data_partition_submit_sequential_consistency | ( | starpu_data_handle_t | initial_handle, |
unsigned | nparts, | ||
starpu_data_handle_t * | children, | ||
int | sequential_consistency | ||
) |
Similar to starpu_data_partition_submit() but also allow to specify the coherency to be used for the main data initial_handle
through the parameter sequential_consistency
.
void starpu_data_unpartition_submit_sequential_consistency | ( | starpu_data_handle_t | initial_handle, |
unsigned | nparts, | ||
starpu_data_handle_t * | children, | ||
int | gathering_node, | ||
int | sequential_consistency | ||
) |
Similar to starpu_data_unpartition_submit() but also allow to specify the coherency to be used for the main data initial_handle
through the parameter sequential_consistency
.
void starpu_bcsr_filter_canonical_block | ( | void * | father_interface, |
void * | child_interface, | ||
struct starpu_data_filter * | f, | ||
unsigned | id, | ||
unsigned | nparts | ||
) |
Partition a block-sparse matrix into dense matrices. starpu_data_filter::get_child_ops needs to be set to starpu_bcsr_filter_canonical_block_child_ops() and starpu_data_filter::get_nchildren set to starpu_bcsr_filter_canonical_block_get_nchildren().
unsigned starpu_bcsr_filter_canonical_block_get_nchildren | ( | struct starpu_data_filter * | f, |
starpu_data_handle_t | handle | ||
) |
Return the number of children obtained with starpu_bcsr_filter_canonical_block().
struct starpu_data_interface_ops* starpu_bcsr_filter_canonical_block_child_ops | ( | struct starpu_data_filter * | f, |
unsigned | child | ||
) |
Return the child_ops of the partition obtained with starpu_bcsr_filter_canonical_block().
void starpu_bcsr_filter_vertical_block | ( | void * | father_interface, |
void * | child_interface, | ||
struct starpu_data_filter * | f, | ||
unsigned | id, | ||
unsigned | nparts | ||
) |
Partition a block-sparse matrix into block-sparse matrices.
The split is done along the leading dimension, i.e. along adjacent nnz blocks.
void starpu_csr_filter_vertical_block | ( | void * | father_interface, |
void * | child_interface, | ||
struct starpu_data_filter * | f, | ||
unsigned | id, | ||
unsigned | nparts | ||
) |
Partition a block-sparse matrix into vertical block-sparse matrices.
void starpu_matrix_filter_block | ( | void * | father_interface, |
void * | child_interface, | ||
struct starpu_data_filter * | f, | ||
unsigned | id, | ||
unsigned | nparts | ||
) |
Partition a dense Matrix along the x dimension, thus getting (x/nparts
,y) matrices. If nparts
does not divide x, the last submatrix contains the remainder.
void starpu_matrix_filter_block_shadow | ( | void * | father_interface, |
void * | child_interface, | ||
struct starpu_data_filter * | f, | ||
unsigned | id, | ||
unsigned | nparts | ||
) |
Partition a dense Matrix along the x dimension, with a shadow border filter_arg_ptr
, thus getting ((x-2*shadow)/nparts
+2*shadow,y) matrices. If nparts
does not divide x-2*shadow, the last submatrix contains the remainder.
IMPORTANT: This can only be used for read-only access, as no coherency is enforced for the shadowed parts. A usage example is available in examples/filters/shadow2d.c
void starpu_matrix_filter_vertical_block | ( | void * | father_interface, |
void * | child_interface, | ||
struct starpu_data_filter * | f, | ||
unsigned | id, | ||
unsigned | nparts | ||
) |
Partition a dense Matrix along the y dimension, thus getting (x,y/nparts
) matrices. If nparts
does not divide y, the last submatrix contains the remainder.
void starpu_matrix_filter_vertical_block_shadow | ( | void * | father_interface, |
void * | child_interface, | ||
struct starpu_data_filter * | f, | ||
unsigned | id, | ||
unsigned | nparts | ||
) |
Partition a dense Matrix along the y dimension, with a shadow border filter_arg_ptr
, thus getting (x,(y-2*shadow)/nparts
+2*shadow) matrices. If nparts
does not divide y-2*shadow, the last submatrix contains the remainder.
IMPORTANT: This can only be used for read-only access, as no coherency is enforced for the shadowed parts. A usage example is available in examples/filters/shadow2d.c
void starpu_vector_filter_block | ( | void * | father_interface, |
void * | child_interface, | ||
struct starpu_data_filter * | f, | ||
unsigned | id, | ||
unsigned | nparts | ||
) |
Return in child_interface
the id
th element of the vector represented by father_interface
once partitioned in nparts
chunks of equal size.
void starpu_vector_filter_block_shadow | ( | void * | father_interface, |
void * | child_interface, | ||
struct starpu_data_filter * | f, | ||
unsigned | id, | ||
unsigned | nparts | ||
) |
Return in child_interface
the id
th element of the vector represented by father_interface
once partitioned in nparts
chunks of equal size with a shadow border filter_arg_ptr
, thus getting a vector of size (n-2*shadow)/nparts+2*shadow
. The filter_arg_ptr
field of f
must be the shadow size casted into void*
.
IMPORTANT: This can only be used for read-only access, as no coherency is enforced for the shadowed parts. An usage example is available in examples/filters/shadow.c
void starpu_vector_filter_list_long | ( | void * | father_interface, |
void * | child_interface, | ||
struct starpu_data_filter * | f, | ||
unsigned | id, | ||
unsigned | nparts | ||
) |
Return in child_interface
the id
th element of the vector represented by father_interface
once partitioned into nparts
chunks according to the filter_arg_ptr
field of f
. The filter_arg_ptr
field must point to an array of nparts
long elements, each of which specifies the number of elements in each chunk of the partition.
void starpu_vector_filter_list | ( | void * | father_interface, |
void * | child_interface, | ||
struct starpu_data_filter * | f, | ||
unsigned | id, | ||
unsigned | nparts | ||
) |
Return in child_interface
the id
th element of the vector represented by father_interface
once partitioned into nparts
chunks according to the filter_arg_ptr
field of f
. The filter_arg_ptr
field must point to an array of nparts
uint32_t elements, each of which specifies the number of elements in each chunk of the partition.
void starpu_vector_filter_divide_in_2 | ( | void * | father_interface, |
void * | child_interface, | ||
struct starpu_data_filter * | f, | ||
unsigned | id, | ||
unsigned | nparts | ||
) |
Return in child_interface
the id
th element of the vector represented by father_interface
once partitioned in 2
chunks of equal size, ignoring nparts. Thus, id
must be 0
or 1
.
void starpu_block_filter_block | ( | void * | father_interface, |
void * | child_interface, | ||
struct starpu_data_filter * | f, | ||
unsigned | id, | ||
unsigned | nparts | ||
) |
Partition a block along the X dimension, thus getting (x/nparts
,y,z) 3D matrices. If nparts
does not divide x, the last submatrix contains the remainder.
void starpu_block_filter_block_shadow | ( | void * | father_interface, |
void * | child_interface, | ||
struct starpu_data_filter * | f, | ||
unsigned | id, | ||
unsigned | nparts | ||
) |
Partition a block along the X dimension, with a shadow border filter_arg_ptr
, thus getting ((x-2*shadow)/nparts
+2*shadow,y,z) blocks. If nparts
does not divide x, the last submatrix contains the remainder.
IMPORTANT: This can only be used for read-only access, as no coherency is enforced for the shadowed parts.
void starpu_block_filter_vertical_block | ( | void * | father_interface, |
void * | child_interface, | ||
struct starpu_data_filter * | f, | ||
unsigned | id, | ||
unsigned | nparts | ||
) |
Partition a block along the Y dimension, thus getting (x,y/nparts
,z) blocks. If nparts
does not divide y, the last submatrix contains the remainder.
void starpu_block_filter_vertical_block_shadow | ( | void * | father_interface, |
void * | child_interface, | ||
struct starpu_data_filter * | f, | ||
unsigned | id, | ||
unsigned | nparts | ||
) |
Partition a block along the Y dimension, with a shadow border filter_arg_ptr
, thus getting (x,(y-2*shadow)/nparts
+2*shadow,z) 3D matrices. If nparts
does not divide y, the last submatrix contains the remainder.
IMPORTANT: This can only be used for read-only access, as no coherency is enforced for the shadowed parts.
void starpu_block_filter_depth_block | ( | void * | father_interface, |
void * | child_interface, | ||
struct starpu_data_filter * | f, | ||
unsigned | id, | ||
unsigned | nparts | ||
) |
Partition a block along the Z dimension, thus getting (x,y,z/nparts
) blocks. If nparts
does not divide z, the last submatrix contains the remainder.
void starpu_block_filter_depth_block_shadow | ( | void * | father_interface, |
void * | child_interface, | ||
struct starpu_data_filter * | f, | ||
unsigned | id, | ||
unsigned | nparts | ||
) |
Partition a block along the Z dimension, with a shadow border filter_arg_ptr
, thus getting (x,y,(z-2*shadow)/nparts
+2*shadow) blocks. If nparts
does not divide z, the last submatrix contains the remainder.
IMPORTANT: This can only be used for read-only access, as no coherency is enforced for the shadowed parts.
void starpu_filter_nparts_compute_chunk_size_and_offset | ( | unsigned | n, |
unsigned | nparts, | ||
size_t | elemsize, | ||
unsigned | id, | ||
unsigned | ld, | ||
unsigned * | chunk_size, | ||
size_t * | offset | ||
) |
Given an integer n
, n
the number of parts it must be divided in, id
the part currently considered, determines the chunk_size
and the offset
, taking into account the size of the elements stored in the data structure elemsize
and ld
, the leading dimension, which is most often 1.