starpu-max/doc/doxygen/chapters/310_data_management.doxy

/* StarPU --- Runtime system for heterogeneous multicore architectures.
 *
 * Copyright (C) 2009-2021  Université de Bordeaux, CNRS (LaBRI UMR 5800), Inria
 *
 * StarPU is free software; you can redistribute it and/or modify
 * it under the terms of the GNU Lesser General Public License as published by
 * the Free Software Foundation; either version 2.1 of the License, or (at
 * your option) any later version.
 *
 * StarPU is distributed in the hope that it will be useful, but
 * WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
 *
 * See the GNU Lesser General Public License in COPYING.LGPL for more details.
 */

/*! \page DataManagement Data Management

TODO: intro which mentions consistency among other things

\section DataInterface Data Interface

StarPU provides several data interfaces for programmers to describe
the data layout of their application. There are predefined interfaces
already available in StarPU. Users can define new data interfaces as
explained in \ref DefiningANewDataInterface. All functions provided by
StarPU are documented in \ref API_Data_Interfaces. You will find a
short list below.

\subsection VariableDataInterface Variable Data Interface

A variable is a given-size byte element, typically a scalar. Here an
example of how to register a variable data to StarPU by using
starpu_variable_data_register().


\code{.c}
float var = 42.0;
starpu_data_handle_t var_handle;
starpu_variable_data_register(&var_handle, STARPU_MAIN_RAM, (uintptr_t)&var, sizeof(var));
\endcode

\subsection VectorDataInterface Vector Data Interface

A vector is a fixed number of elements of a given size. Here an
example of how to register a vector data to StarPU by using
starpu_vector_data_register().

\code{.c}
float vector[NX];
starpu_data_handle_t vector_handle;
starpu_vector_data_register(&vector_handle, STARPU_MAIN_RAM, (uintptr_t)vector, NX, sizeof(vector[0]));
\endcode

Vectors can be partitioned into pieces by using
starpu_vector_filter_block(). They can also be partitioned with some overlapping
by using starpu_vector_filter_block_shadow(). By default StarPU
uses the same size for each piece. If different sizes are desired,
starpu_vector_filter_list() or starpu_vector_filter_list_long() can be used
instead. To just divide in two pieces, starpu_vector_filter_divide_in_2() can be used.

\subsection MatrixDataInterface Matrix Data Interface

To register 2-D matrices with a potential padding, one can use the
matrix data interface. Here an example of how to register a matrix
data to StarPU by using starpu_matrix_data_register().

\code{.c}
float *matrix;
starpu_data_handle_t matrix_handle;
matrix = (float*)malloc(width * height * sizeof(float));
starpu_matrix_data_register(&matrix_handle, STARPU_MAIN_RAM, (uintptr_t)matrix, width, width, height, sizeof(float));
\endcode

2D matrices can be partitioned into 2D matrices along the x dimension by
using starpu_matrix_filter_block(), and along the y dimension by using
starpu_matrix_filter_vertical_block(). They can also be partitioned
with some overlapping by using starpu_matrix_filter_block_shadow() and
starpu_matrix_filter_vertical_block_shadow().

\subsection BlockDataInterface Block Data Interface

To register 3-D matrices with potential paddings on Y and Z dimensions,
one can use the block data interface. Here an example of how to
register a block data to StarPU by using starpu_block_data_register().

\code{.c}
float *block;
starpu_data_handle_t block_handle;
block = (float*)malloc(nx*ny*nz*sizeof(float));
starpu_block_data_register(&block_handle, STARPU_MAIN_RAM, (uintptr_t)block, nx, nx*ny, nx, ny, nz, sizeof(float));
\endcode

3D matrices can be partitioned along the x dimension by
using starpu_block_filter_block(), or along the y dimension
by using starpu_block_filter_vertical_block(), or along the
z dimension by using starpu_block_filter_depth_block(). They
can also be partitioned with some overlapping by using
starpu_block_filter_block_shadow(), starpu_block_filter_vertical_block_shadow(),
or starpu_block_filter_depth_block_shadow().

\subsection TensorDataInterface Tensor Data Interface

To register 4-D matrices with potential paddings on Y, Z, and T dimensions,
one can use the tensor data interface. Here an example of how to
register a tensor data to StarPU by using starpu_tensor_data_register().

\code{.c}
float *block;
starpu_data_handle_t block_handle;
block = (float*)malloc(nx*ny*nz*nt*sizeof(float));
starpu_tensor_data_register(&block_handle, STARPU_MAIN_RAM, (uintptr_t)block, nx, nx*ny, nx*ny*nz, nx, ny, nz, nt, sizeof(float));
\endcode

Partitioning filters are not implemented yet.

\subsection BCSRDataInterface BCSR Data Interface

BCSR (Blocked Compressed Sparse Row Representation) sparse matrix data
can be registered to StarPU using the bcsr data interface. Here an
example on how to do so by using starpu_bcsr_data_register().

\code{.c}
/*
 * We use the following matrix:
 *
 *   +----------------+
 *   |  0   1   0   0 |
 *   |  2   3   0   0 |
 *   |  4   5   8   9 |
 *   |  6   7  10  11 |
 *   +----------------+
 *
 * nzval  = [0, 1, 2, 3] ++ [4, 5, 6, 7] ++ [8, 9, 10, 11]
 * colind = [0, 0, 1]
 * rowptr = [0, 1, 3]
 * r = c = 2
 */

/* Size of the blocks */
int R = 2;
int C = 2;

int NROWS = 2;
int NNZ_BLOCKS = 3;    /* out of 4 */
int NZVAL_SIZE = (R*C*NNZ_BLOCKS);

int nzval[NZVAL_SIZE]  =
{
	0, 1, 2, 3,    /* First block  */
	4, 5, 6, 7,    /* Second block */
	8, 9, 10, 11   /* Third block  */
};
uint32_t colind[NNZ_BLOCKS] =
{
	0, /* block-column index for first block in nzval */
	0, /* block-column index for second block in nzval */
	1  /* block-column index for third block in nzval */
};
uint32_t rowptr[NROWS+1] =
{
	0, / * block-index in nzval of the first block of the first row. */
	1, / * block-index in nzval of the first block of the second row. */
	NNZ_BLOCKS /* number of blocks, to allow an easier element's access for the kernels */
};

starpu_data_handle_t bcsr_handle;
starpu_bcsr_data_register(&bcsr_handle,
			  STARPU_MAIN_RAM,
			  NNZ_BLOCKS,
			  NROWS,
			  (uintptr_t) nzval,
			  colind,
			  rowptr,
			  0, /* firstentry */
			  R,
			  C,
			  sizeof(nzval[0]));
\endcode

StarPU provides an example on how to deal with such matrices in
<c>examples/spmv</c>.

BCSR data handles can be partitioned into its dense matrix blocks by using
starpu_bcsr_filter_canonical_block(), or split into other BCSR data handles by
using starpu_bcsr_filter_vertical_block() (but only split along the leading dimension is
supported, i.e. along adjacent nnz blocks)

\subsection CSRDataInterface CSR Data Interface

TODO

CSR data handles can be partitioned into vertical CSR matrices by using
starpu_csr_filter_vertical_block().

\subsection VariableSizeDataInterface Data Interface with Variable Size

Tasks are actually allowed to change the size of data interfaces.

The simplest case is just changing the amount of data actually used within the
allocated buffer. This is for instance implemented for the matrix interface: one
can set the new NX/NY values with STARPU_MATRIX_SET_NX(), STARPU_MATRIX_SET_NY(), and STARPU_MATRIX_SET_LD()
at the end of the task implementation. Data transfers achieved by StarPU will
then use these values instead of the whole allocated size. The values of course
need to be set within the original allocation. To reserve room for increasing
the NX/NY values, one can use starpu_matrix_data_register_allocsize() instead of
starpu_matrix_data_register(), to specify the allocation size to be used instead
of the default NX*NY*ELEMSIZE. To support this, the data interface
has to implement the functions starpu_data_interface_ops::alloc_footprint and
starpu_data_interface_ops::alloc_compare, for proper StarPU allocation
management.

A more involved case is changing the amount of allocated data.
The task implementation can just reallocate the buffer during its execution, and
set the proper new values in the interface structure, e.g. nx, ny, ld, etc. so
that the StarPU core knows the new data layout. The structure starpu_data_interface_ops
however then needs to have the field starpu_data_interface_ops::dontcache
set to 1, to prevent StarPU from trying to perform any cached allocation,
since the allocated size will vary. An example is available in
<c>tests/datawizard/variable_size.c</c>. The example uses its own data
interface so as to contain some simulation information for data growth, but the
principle can be applied for any data interface.

The principle is to use starpu_malloc_on_node_flags() to make the new
allocation, and use starpu_free_on_node_flags() to release any previous
allocation. The flags have to be precisely like in the example:

\code{.c}
unsigned workerid = starpu_worker_get_id_check();
unsigned dst_node = starpu_worker_get_memory_node(workerid);
interface->ptr = starpu_malloc_on_node_flags(dst_node, size + increase, STARPU_MALLOC_PINNED | STARPU_MALLOC_COUNT | STARPU_MEMORY_OVERFLOW);
starpu_free_on_node_flags(dst_node, old, size, STARPU_MALLOC_PINNED | STARPU_MALLOC_COUNT | STARPU_MEMORY_OVERFLOW);
interface->size += increase;
\endcode

so that the allocated area has the expected properties and the allocation is accounted for properly.

Depending on the interface (vector, CSR, etc.) you may have to fix several
members of the data interface: e.g. both <c>nx</c> and <c>allocsize</c> for
vectors, and store the pointer both in <c>ptr</c> and <c>dev_handle</c>.

Some interfaces make a distinction between the actual number of elements
stored in the data and the actually allocated buffer. For instance, the vector
interface uses the <c>nx</c> field for the former, and the <c>allocsize</c> for
the latter. This allows for lazy reallocation to avoid reallocating the buffer
everytime to exactly match the actual number of elements. Computations and data
transfers will use the field <c>nx</c>, while allocation functions will use the field
<c>allocsize</c>. One just has to make sure that <c>allocsize</c> is always
bigger or equal to <c>nx</c>.

Important note: one can not change the size of a partitioned data.


\section DataManagement Data Management

When the application allocates data, whenever possible it should use
the function starpu_malloc(), which will ask CUDA or OpenCL to make
the allocation itself and pin the corresponding allocated memory, or to use the function
starpu_memory_pin() to pin memory allocated by other ways, such as local arrays. This
is needed to permit asynchronous data transfer, i.e. permit data
transfer to overlap with computations. Otherwise, the trace will show
that the state <c>DriverCopyAsync</c> takes a lot of time, this is
because CUDA or OpenCL then reverts to synchronous transfers.

The application can provide its own allocation function by calling
starpu_malloc_set_hooks(). StarPU will then use them for all data handle
allocations in the main memory.

By default, StarPU leaves replicates of data wherever they were used, in case they
will be re-used by other tasks, thus saving the data transfer time. When some
task modifies some data, all the other replicates are invalidated, and only the
processing unit which ran this task will have a valid replicate of the data. If the application knows
that this data will not be re-used by further tasks, it should advise StarPU to
immediately replicate it to a desired list of memory nodes (given through a
bitmask). This can be understood like the write-through mode of CPU caches.

\code{.c}
starpu_data_set_wt_mask(img_handle, 1<<0);
\endcode

will for instance request to always automatically transfer a replicate into the
main memory (node <c>0</c>), as bit <c>0</c> of the write-through bitmask is being set.

\code{.c}
starpu_data_set_wt_mask(img_handle, ~0U);
\endcode

will request to always automatically broadcast the updated data to all memory
nodes.

Setting the write-through mask to <c>~0U</c> can also be useful to make sure all
memory nodes always have a copy of the data, so that it is never evicted when
memory gets scarse.

Implicit data dependency computation can become expensive if a lot
of tasks access the same piece of data. If no dependency is required
on some piece of data (e.g. because it is only accessed in read-only
mode, or because write accesses are actually commutative), use the
function starpu_data_set_sequential_consistency_flag() to disable
implicit dependencies on this data.

In the same vein, accumulation of results in the same data can become a
bottleneck. The use of the mode ::STARPU_REDUX permits to optimize such
accumulation (see \ref DataReduction). To a lesser extent, the use of
the flag ::STARPU_COMMUTE keeps the bottleneck (see \ref DataCommute), but at least permits
the accumulation to happen in any order.

Applications often need a data just for temporary results.  In such a case,
registration can be made without an initial value, for instance this produces a vector data:

\code{.c}
starpu_vector_data_register(&handle, -1, 0, n, sizeof(float));
\endcode

StarPU will then allocate the actual buffer only when it is actually needed,
e.g. directly on the GPU without allocating in main memory.

In the same vein, once the temporary results are not useful any more, the
data should be thrown away. If the handle is not to be reused, it can be
unregistered:

\code{.c}
starpu_data_unregister_submit(handle);
\endcode

actual unregistration will be done after all tasks working on the handle
terminate.

If the handle is to be reused, instead of unregistering it, it can simply be invalidated:

\code{.c}
starpu_data_invalidate_submit(handle);
\endcode

the buffers containing the current value will then be freed, and reallocated
only when another task writes some value to the handle.

\section DataPrefetch Data Prefetch

The scheduling policies <c>heft</c>, <c>dmda</c> and <c>pheft</c>
perform data prefetch (see \ref STARPU_PREFETCH):
as soon as a scheduling decision is taken for a task, requests are issued to
transfer its required data to the target processing unit, if needed, so that
when the processing unit actually starts the task, its data will hopefully be
already available and it will not have to wait for the transfer to finish.

The application may want to perform some manual prefetching, for several reasons
such as excluding initial data transfers from performance measurements, or
setting up an initial statically-computed data distribution on the machine
before submitting tasks, which will thus guide StarPU toward an initial task
distribution (since StarPU will try to avoid further transfers).

This can be achieved by giving the function starpu_data_prefetch_on_node() the
handle and the desired target memory node. The variant
starpu_data_idle_prefetch_on_node() can be used to issue the transfer
only when the bus is idle.

Conversely, one can advise StarPU that some data will not be useful in the
close future by calling starpu_data_wont_use(). StarPU will then write its value
back to its home node, and evict it from GPUs when room is needed.

\section PartitioningData Partitioning Data

An existing piece of data can be partitioned in sub parts to be used by different tasks, for instance:

\code{.c}
#define NX 1048576
#define PARTS 16
int vector[NX];
starpu_data_handle_t handle;

/* Declare data to StarPU */
starpu_vector_data_register(&handle, STARPU_MAIN_RAM, (uintptr_t)vector, NX, sizeof(vector[0]));

/* Partition the vector in PARTS sub-vectors */
struct starpu_data_filter f =
{
    .filter_func = starpu_vector_filter_block,
    .nchildren = PARTS
};
starpu_data_partition(handle, &f);
\endcode

The task submission then uses the function starpu_data_get_sub_data()
to retrieve the sub-handles to be passed as tasks parameters.

\code{.c}
/* Submit a task on each sub-vector */
for (i=0; i<starpu_data_get_nb_children(handle); i++)
{
    /* Get subdata number i (there is only 1 dimension) */
    starpu_data_handle_t sub_handle = starpu_data_get_sub_data(handle, 1, i);
    struct starpu_task *task = starpu_task_create();

    task->handles[0] = sub_handle;
    task->cl = &cl;
    task->synchronous = 1;
    task->cl_arg = &factor;
    task->cl_arg_size = sizeof(factor);

    starpu_task_submit(task);
}
\endcode

Partitioning can be applied several times, see
<c>examples/basic_examples/mult.c</c> and <c>examples/filters/</c>.

Wherever the whole piece of data is already available, the partitioning will
be done in-place, i.e. without allocating new buffers but just using pointers
inside the existing copy. This is particularly important to be aware of when
using OpenCL, where the kernel parameters are not pointers, but \c cl_mem handles. The
kernel thus needs to be also passed the offset within the OpenCL buffer:

\code{.c}
void opencl_func(void *buffers[], void *cl_arg)
{
    cl_mem vector = (cl_mem) STARPU_VECTOR_GET_DEV_HANDLE(buffers[0]);
    unsigned offset = STARPU_BLOCK_GET_OFFSET(buffers[0]);

    ...
    clSetKernelArg(kernel, 0, sizeof(vector), &vector);
    clSetKernelArg(kernel, 1, sizeof(offset), &offset);
    ...
}
\endcode

And the kernel has to shift from the pointer passed by the OpenCL driver:

\code{.c}
__kernel void opencl_kernel(__global int *vector, unsigned offset)
{
    block = (__global void *)block + offset;
    ...
}
\endcode

When the sub-data is not of the same type as the original data, the field
starpu_data_filter::get_child_ops needs to be set appropriately for StarPU
to know which type should be used.

StarPU provides various interfaces and filters for matrices, vectors, etc.,
but applications can also write their own data interfaces and filters, see
<c>examples/interface</c> and <c>examples/filters/custom_mf</c> for an example,
and see \ref DefiningANewDataInterface and \ref DefiningANewDataFilter
for documentation.

\section AsynchronousPartitioning Asynchronous Partitioning

The partitioning functions described in the previous section are synchronous:
starpu_data_partition() and starpu_data_unpartition() both wait for all the tasks
currently working on the data.  This can be a bottleneck for the application.

An asynchronous API also exists, it works only on handles with sequential
consistency. The principle is to first plan the partitioning, which returns
data handles of the partition, which are not functional yet. When submitting
tasks, one can mix using the handles of the partition, of the whole data. One
can even partition recursively and mix using handles at different levels of the
recursion. Of course, StarPU will have to introduce coherency synchronization.

<c>fmultiple_submit_implicit</c> is a complete example using this technique.
One can also look at <c>fmultiple_submit_readonly</c> which contains the
explicit coherency synchronization which are automatically introduced by StarPU
for <c>fmultiple_submit_implicit</c>.

In short, we first register a matrix and plan the partitioning:

\code{.c}
starpu_matrix_data_register(&handle, STARPU_MAIN_RAM, (uintptr_t)matrix, NX, NX, NY, sizeof(matrix[0]));
struct starpu_data_filter f_vert =
{
	.filter_func = starpu_matrix_filter_block,
	.nchildren = PARTS
};
starpu_data_partition_plan(handle, &f_vert, vert_handle);
\endcode

starpu_data_partition_plan() returns the handles for the partition in <c>vert_handle</c>.

One can then submit tasks working on the main handle, and tasks working on the handles
<c>vert_handle</c>. Between using the main handle and the handles <c>vert_handle</c>, StarPU will automatically call starpu_data_partition_submit() and
starpu_data_unpartition_submit().

All this code is asynchronous, just submitting which tasks, partitioning and
unpartitioning will be done at runtime.

Planning several partitioning of the same data is also possible, StarPU will
unpartition and repartition as needed when mixing accesses of different
partitions. If data access is done in read-only mode, StarPU will allow the
different partitioning to coexist. As soon as a data is accessed in read-write
mode, StarPU will automatically unpartition everything and activate only the
partitioning leading to the data being written to.

For instance, for a stencil application, one can split a subdomain into
its interior and halos, and then just submit a task updating the whole
subdomain, then submit MPI sends/receives to update the halos, then submit
again a task updating the whole subdomain, etc. and StarPU will automatically
partition/unpartition each time.

\section ManualPartitioning Manual Partitioning

One can also handle partitioning by hand, by registering several views on the
same piece of data. The idea is then to manage the coherency of the various
views through the common buffer in the main memory.
<c>fmultiple_manual</c> is a complete example using this technique.

In short, we first register the same matrix several times:

\code{.c}
starpu_matrix_data_register(&handle, STARPU_MAIN_RAM, (uintptr_t)matrix, NX, NX, NY, sizeof(matrix[0]));

for (i = 0; i < PARTS; i++)
	starpu_matrix_data_register(&vert_handle[i], STARPU_MAIN_RAM, (uintptr_t)&matrix[0][i*(NX/PARTS)], NX, NX/PARTS, NY, sizeof(matrix[0][0]));
\endcode

Since StarPU is not aware that the two handles are actually pointing to the same
data, we have a danger of inadvertently submitting tasks to both views, which
will bring a mess since StarPU will not guarantee any coherency between the two
views.  To make sure we don't do this, we invalidate the view that we will not
use:

\code{.c}
for (i = 0; i < PARTS; i++)
	starpu_data_invalidate(vert_handle[i]);
\endcode

Then we can safely work on <c>handle</c>.

When we want to switch to the vertical slice view, all we need to do is bring
coherency between them by running an empty task on the home node of the data:

\code{.c}
struct starpu_codelet cl_switch =
{
	.where = STARPU_NOWHERE,
	.nbuffers = 3,
	.specific_nodes = 1,
	.nodes = { STARPU_MAIN_RAM, STARPU_MAIN_RAM, STARPU_MAIN_RAM },
};

ret = starpu_task_insert(&cl_switch, STARPU_RW, handle,
			STARPU_W, vert_handle[0],
			STARPU_W, vert_handle[1],
			0);
\endcode

The execution of the task <c>switch</c> will get back the matrix data into the
main memory, and thus the vertical slices will get the updated value there.

Again, we prefer to make sure that we don't accidentally access the matrix through the whole-matrix handle:

\code{.c}
starpu_data_invalidate_submit(handle);
\endcode

Note: when enabling a set of handles in this way, the set must not have any
overlapping, i.e. the handles of the set must not have any part of data in
common, otherwise StarPU will not properly handle concurrent accesses between
them.

And now we can start using vertical slices, etc.

\section DataPointers Handles data buffer pointers

A simple understanding of StarPU handles is that it's a collection of buffers on
each memory node of the machine, which contain the same data.  The picture is
however made more complex with the OpenCL support and with partitioning.

When partitioning a handle, the data buffers of the subhandles will indeed
be inside the data buffers of the main handle (to save transferring data
back and forth between the main handle and the subhandles). But in OpenCL,
a <c>cl_mem</c> is not a pointer, but an opaque value on which pointer
arithmetic can not be used. That is why data interfaces contain three fields:
<c>dev_handle</c>, <c>offset</c>, and <c>ptr</c>.
<ul>
<li> The field <c>dev_handle</c> is what the allocation function
returned, and one can not do arithmetic on it.
</li>
<li> The field <c>offset</c> is the offset inside the allocated area,
most often it will be 0 because data start at the beginning of the
allocated area, but when the handle is partitioned, the subhandles
will have varying <c>offset</c> values, for each subpiece.
</li>
<li> The field <c>ptr</c>, in the non-OpenCL case, i.e. when pointer
arithmetic can be used on <c>dev_handle</c>, is just the sum of
<c>dev_handle</c> and <c>offset</c>, provided for convenience.
</li>
</ul>

This means that:
<ul>
<li>computation kernels can use <c>ptr</c> in non-OpenCL implementations.</li>
<li>computation kernels have to use <c>dev_handle</c> and <c>offset</c> in the OpenCL implementation.</li>
<li>allocation methods of data interfaces have to store the value returned by starpu_malloc_on_node() in <c>dev_handle</c> and <c>ptr</c>, and set <c>offset</c> to 0.</li>
<li>partitioning filters have to copy over <c>dev_handle</c> without modifying it, set in the child different values of <c>offset</c>, and set <c>ptr</c> accordingly as the sum of <c>dev_handle</c> and <c>offset</c>.</li>
</ul>

\section DefiningANewDataFilter Defining A New Data Filter

StarPU provides a series of predefined filters in \ref API_Data_Partition, but
additional filters can be defined by the application. The principle is that the
filter function just fills the memory location of the <c>i-th</c> subpart of a data.
Examples are provided in <c>src/datawizard/interfaces/*_filters.c</c>,
check \ref starpu_data_filter::filter_func for further details.
The helper function starpu_filter_nparts_compute_chunk_size_and_offset() can be used to
compute the division of pieces of data.

\section DataReduction Data Reduction

In various cases, some piece of data is used to accumulate intermediate
results. For instances, the dot product of a vector, maximum/minimum finding,
the histogram of a photograph, etc. When these results are produced along the
whole machine, it would not be efficient to accumulate them in only one place,
incurring data transmission each and access concurrency.

StarPU provides a mode ::STARPU_REDUX, which permits to optimize
this case: it will allocate a buffer on each worker (lazily), and accumulate
intermediate results there. When the data is eventually accessed in the normal
mode ::STARPU_R, StarPU will collect the intermediate results in just one
buffer.

For this to work, the user has to use the function
starpu_data_set_reduction_methods() to declare how to initialize these
buffers, and how to assemble partial results.

For instance, <c>cg</c> uses that to optimize its dot product: it first defines
the codelets for initialization and reduction:

\code{.c}
struct starpu_codelet bzero_variable_cl =
{
        .cpu_funcs = { bzero_variable_cpu },
        .cpu_funcs_name = { "bzero_variable_cpu" },
        .cuda_funcs = { bzero_variable_cuda },
        .nbuffers = 1,
}

static void accumulate_variable_cpu(void *descr[], void *cl_arg)
{
        double *v_dst = (double *)STARPU_VARIABLE_GET_PTR(descr[0]);
        double *v_src = (double *)STARPU_VARIABLE_GET_PTR(descr[1]);
        *v_dst = *v_dst + *v_src;
}

static void accumulate_variable_cuda(void *descr[], void *cl_arg)
{
        double *v_dst = (double *)STARPU_VARIABLE_GET_PTR(descr[0]);
        double *v_src = (double *)STARPU_VARIABLE_GET_PTR(descr[1]);
        cublasaxpy(1, (double)1.0, v_src, 1, v_dst, 1);
        cudaStreamSynchronize(starpu_cuda_get_local_stream());
}

struct starpu_codelet accumulate_variable_cl =
{
        .cpu_funcs = { accumulate_variable_cpu },
        .cpu_funcs_name = { "accumulate_variable_cpu" },
        .cuda_funcs = { accumulate_variable_cuda },
        .nbuffers = 2,
	.modes = {STARPU_RW|STARPU_COMMUTE, STARPU_R},
}
\endcode

and attaches them as reduction methods for its handle <c>dtq</c>:

\code{.c}
starpu_variable_data_register(&dtq_handle, -1, NULL, sizeof(type));
starpu_data_set_reduction_methods(dtq_handle, &accumulate_variable_cl, &bzero_variable_cl);
\endcode

and <c>dtq_handle</c> can now be used in mode ::STARPU_REDUX for the
dot products with partitioned vectors:

\code{.c}
for (b = 0; b < nblocks; b++)
    starpu_task_insert(&dot_kernel_cl,
        STARPU_REDUX, dtq_handle,
        STARPU_R, starpu_data_get_sub_data(v1, 1, b),
        STARPU_R, starpu_data_get_sub_data(v2, 1, b),
        0);
\endcode

During registration, we have here provided <c>NULL</c>, i.e. there is
no initial value to be taken into account during reduction. StarPU
will thus only take into account the contributions from the tasks
<c>dot_kernel_cl</c>. Also, it will not allocate any memory for
<c>dtq_handle</c> before tasks <c>dot_kernel_cl</c> are ready to run.

If another dot product has to be performed, one could unregister
<c>dtq_handle</c>, and re-register it. But one can also call
starpu_data_invalidate_submit() with the parameter <c>dtq_handle</c>,
which will clear all data from the handle, thus resetting it back to
the initial status <c>register(NULL)</c>.

The example <c>cg</c> also uses reduction for the blocked gemv kernel,
leading to yet more relaxed dependencies and more parallelism.

::STARPU_REDUX can also be passed to starpu_mpi_task_insert() in the MPI
case. This will however not produce any MPI communication, but just pass
::STARPU_REDUX to the underlying starpu_task_insert(). It is up to the
application to call starpu_mpi_redux_data(), which posts tasks which will
reduce the partial results among MPI nodes into the MPI node which owns the
data. For instance, some hypothetical application which collects partial results
into data <c>res</c>, then uses it for other computation, before looping again
with a new reduction:

\code{.c}
for (i = 0; i < 100; i++)
{
    starpu_mpi_task_insert(MPI_COMM_WORLD, &init_res, STARPU_W, res, 0);
    starpu_mpi_task_insert(MPI_COMM_WORLD, &work, STARPU_RW, A, STARPU_R, B, STARPU_REDUX, res, 0);
    starpu_mpi_redux_data(MPI_COMM_WORLD, res);
    starpu_mpi_task_insert(MPI_COMM_WORLD, &work2, STARPU_RW, B, STARPU_R, res, 0);
}
\endcode

\section DataCommute Commute Data Access

By default, the implicit dependencies computed from data access use the
sequential semantic. Notably, write accesses are always serialized in the order
of submission. In some applicative cases, the write contributions can actually
be performed in any order without affecting the eventual result. In this case
it is useful to drop the strictly sequential semantic, to improve parallelism
by allowing StarPU to reorder the write accesses. This can be done by using
the data access flag ::STARPU_COMMUTE. Accesses without this flag will however
properly be serialized against accesses with this flag. For instance:

\code{.c}
    starpu_task_insert(&cl1, STARPU_R, h, STARPU_RW, handle, 0);
    starpu_task_insert(&cl2, STARPU_R, handle1, STARPU_RW|STARPU_COMMUTE, handle, 0);
    starpu_task_insert(&cl2, STARPU_R, handle2, STARPU_RW|STARPU_COMMUTE, handle, 0);
    starpu_task_insert(&cl3, STARPU_R, g, STARPU_RW, handle, 0);
\endcode

The two tasks running <c>cl2</c> will be able to commute: depending on whether the
value of <c>handle1</c> or <c>handle2</c> becomes available first, the corresponding task
running <c>cl2</c> will start first. The task running <c>cl1</c> will however always be run
before them, and the task running <c>cl3</c> will always be run after them.

If a lot of tasks use the commute access on the same set of data and a lot of
them are ready at the same time, it may become interesting to use an arbiter,
see \ref ConcurrentDataAccess.

\section ConcurrentDataAccess Concurrent Data Accesses

When several tasks are ready and will work on several data, StarPU is faced with
the classical Dining Philosophers problem, and has to determine the order in
which it will run the tasks.

Data accesses usually use sequential ordering, so data accesses are usually
already serialized, and thus by default StarPU uses the Dijkstra solution which
scales very well in terms of overhead: tasks will just acquire data one by one
by data handle pointer value order.

When sequential ordering is disabled or the flag ::STARPU_COMMUTE is used, there
may be a lot of concurrent accesses to the same data, and the Dijkstra solution
gets only poor parallelism, typically in some pathological cases which do happen
in various applications. In this case, one can use a data access arbiter, which
implements the classical centralized solution for the Dining Philosophers
problem. This is more expensive in terms of overhead since it is centralized,
but it opportunistically gets a lot of parallelism. The centralization can also
be avoided by using several arbiters, thus separating sets of data for which
arbitration will be done.  If a task accesses data from different arbiters, it
will acquire them arbiter by arbiter, in arbiter pointer value order.

See the <c>tests/datawizard/test_arbiter.cpp</c> example.

Arbiters however do not support the flag ::STARPU_REDUX yet.

\section TemporaryBuffers Temporary Buffers

There are two kinds of temporary buffers: temporary data which just pass results
from a task to another, and scratch data which are needed only internally by
tasks.

\subsection TemporaryData Temporary Data

Data can sometimes be entirely produced by a task, and entirely consumed by
another task, without the need for other parts of the application to access
it. In such case, registration can be done without prior allocation, by using
the special memory node number <c>-1</c>, and passing a zero pointer. StarPU will
actually allocate memory only when the task creating the content gets scheduled,
and destroy it on unregistration.

In addition to this, it can be tedious for the application to have to unregister
the data, since it will not use its content anyway. The unregistration can be
done lazily by using the function starpu_data_unregister_submit(),
which will record that no more tasks accessing the handle will be submitted, so
that it can be freed as soon as the last task accessing it is over.

The following code examplifies both points: it registers the temporary
data, submits three tasks accessing it, and records the data for automatic
unregistration.

\code{.c}
starpu_vector_data_register(&handle, -1, 0, n, sizeof(float));
starpu_task_insert(&produce_data, STARPU_W, handle, 0);
starpu_task_insert(&compute_data, STARPU_RW, handle, 0);
starpu_task_insert(&summarize_data, STARPU_R, handle, STARPU_W, result_handle, 0);
starpu_data_unregister_submit(handle);
\endcode

The application may also want to see the temporary data initialized
on the fly before being used by the task. This can be done by using
starpu_data_set_reduction_methods() to set an initialization codelet (no redux
codelet is needed).

\subsection ScratchData Scratch Data

Some kernels sometimes need temporary data to achieve the computations, i.e. a
workspace. The application could allocate it at the start of the codelet
function, and free it at the end, but this would be costly. It could also
allocate one buffer per worker (similarly to \ref HowToInitializeAComputationLibraryOnceForEachWorker),
but this would
make them systematic and permanent. A more  optimized way is to use
the data access mode ::STARPU_SCRATCH, as examplified below, which
provides per-worker buffers without content consistency. The buffer is
registered only once, using memory node <c>-1</c>, i.e. the application didn't allocate
memory for it, and StarPU will allocate it on demand at task execution.

\code{.c}
starpu_vector_data_register(&workspace, -1, 0, sizeof(float));
for (i = 0; i < N; i++)
    starpu_task_insert(&compute, STARPU_R, input[i], STARPU_SCRATCH, workspace, STARPU_W, output[i], 0);
\endcode

StarPU will make sure that the buffer is allocated before executing the task,
and make this allocation per-worker: for CPU workers, notably, each worker has
its own buffer. This means that each task submitted above will actually have its
own workspace, which will actually be the same for all tasks running one after
the other on the same worker. Also, if for instance memory becomes scarce,
StarPU will notice that it can free such buffers easily, since the content does
not matter.

The example <c>examples/pi</c> uses scratches for some temporary buffer.

\section TheMultiformatInterface The Multiformat Interface

It may be interesting to represent the same piece of data using two different
data structures: one only used on CPUs, and one only used on GPUs.
This can be done by using the multiformat interface. StarPU
will be able to convert data from one data structure to the other when needed.
Note that the scheduler <c>dmda</c> is the only one optimized for this
interface. The user must provide StarPU with conversion codelets:

\snippet multiformat.c To be included. You should update doxygen if you see this text.

Kernels can be written almost as for any other interface. Note that
::STARPU_MULTIFORMAT_GET_CPU_PTR shall only be used for CPU kernels. CUDA kernels
must use ::STARPU_MULTIFORMAT_GET_CUDA_PTR, and OpenCL kernels must use
::STARPU_MULTIFORMAT_GET_OPENCL_PTR. ::STARPU_MULTIFORMAT_GET_NX may
be used in any kind of kernel.

\code{.c}
static void
multiformat_scal_cpu_func(void *buffers[], void *args)
{
    struct point *aos;
    unsigned int n;

    aos = STARPU_MULTIFORMAT_GET_CPU_PTR(buffers[0]);
    n = STARPU_MULTIFORMAT_GET_NX(buffers[0]);
    ...
}

extern "C" void multiformat_scal_cuda_func(void *buffers[], void *_args)
{
    unsigned int n;
    struct struct_of_arrays *soa;

    soa = (struct struct_of_arrays *) STARPU_MULTIFORMAT_GET_CUDA_PTR(buffers[0]);
    n = STARPU_MULTIFORMAT_GET_NX(buffers[0]);

    ...
}
\endcode

A full example may be found in <c>examples/basic_examples/multiformat.c</c>.

\section DefiningANewDataInterface Defining A New Data Interface

This section proposes an example how to define your own interface, when the
StarPU-provided interface do not fit your needs. Here we take a dumb example of
an array of complex numbers represented by two arrays of double values.

Let's thus define a new data interface to manage arrays of complex numbers:

\code{.c}
/* interface for complex numbers */
struct starpu_complex_interface
{
        double *real;
        double *imaginary;
        int nx;
};
\endcode

That structure stores enough to describe <b>one</b> buffer of such kind of
data. It is used for the buffer stored in the main memory, another instance
is used for the buffer stored in a GPU, etc. A <i>data handle</i> is thus a
collection of such structures, to describe each buffer on each memory node.

Note: one should not take pointers into such structures, because StarPU needs
to be able to copy over the content of it to various places, for instance to
efficiently migrate a data buffer from one data handle to another data handle.

\subsection DefiningANewDataInterface_registration Data registration

Registering such a data to StarPU is easily done using the function
starpu_data_register(). The last
parameter of the function, <c>interface_complex_ops</c>, will be
described below.

\code{.c}
void starpu_complex_data_register(starpu_data_handle_t *handleptr,
     unsigned home_node, double *real, double *imaginary, int nx)
{
        struct starpu_complex_interface complex =
        {
                .real = real,
                .imaginary = imaginary,
                .nx = nx
        };

        if (interface_complex_ops.interfaceid == STARPU_UNKNOWN_INTERFACE_ID)
        {
                interface_complex_ops.interfaceid = starpu_data_interface_get_next_id();
        }

        starpu_data_register(handleptr, home_node, &complex, &interface_complex_ops);
}
\endcode

The <c>struct starpu_complex_interface complex</c> is here used just to store the
parameters that the user provided to <c>starpu_complex_data_register</c>.
starpu_data_register() will first allocate the handle, and
then pass the structure <c>starpu_complex_interface</c> to the method
starpu_data_interface_ops::register_data_handle, which records them
within the data handle (it is called once per node by starpu_data_register()):

\code{.c}
static void complex_register_data_handle(starpu_data_handle_t handle, unsigned home_node, void *data_interface)
{
	struct starpu_complex_interface *complex_interface = (struct starpu_complex_interface *) data_interface;

	unsigned node;
	for (node = 0; node < STARPU_MAXNODES; node++)
	{
		struct starpu_complex_interface *local_interface = (struct starpu_complex_interface *)
			starpu_data_get_interface_on_node(handle, node);

		local_interface->nx = complex_interface->nx;
		if (node == home_node)
		{
			local_interface->real = complex_interface->real;
			local_interface->imaginary = complex_interface->imaginary;
		}
		else
		{
			local_interface->real = NULL;
			local_interface->imaginary = NULL;
		}
	}
}
\endcode

If the application provided a home node, the corresponding pointers will be
recorded for that node. Others have no buffer allocated yet.
Possibly the interface needs some dynamic allocation (e.g. to store an array of
dimensions that can have variable size). The corresponding deallocation will then be
done in starpu_data_interface_ops::unregister_data_handle.

Different operations need to be defined for a data interface through
the type starpu_data_interface_ops. We only define here the basic
operations needed to run simple applications. The source code for the
different functions can be found in the file
<c>examples/interface/complex_interface.c</c>, the details of the hooks to be
provided are documented in \ref starpu_data_interface_ops .

\code{.c}
static struct starpu_data_interface_ops interface_complex_ops =
{
        .register_data_handle = complex_register_data_handle,
        .allocate_data_on_node = complex_allocate_data_on_node,
        .copy_methods = &complex_copy_methods,
        .get_size = complex_get_size,
        .footprint = complex_footprint,
        .interfaceid = STARPU_UNKNOWN_INTERFACE_ID,
        .interface_size = sizeof(struct starpu_complex_interface),
};
\endcode

Convenience functions can defined to access the different fields of the
complex interface from a StarPU data handle after a call to starpu_data_acquire():

\code{.c}
double *starpu_complex_get_real(starpu_data_handle_t handle)
{
        struct starpu_complex_interface *complex_interface =
          (struct starpu_complex_interface *) starpu_data_get_interface_on_node(handle, STARPU_MAIN_RAM);
        return complex_interface->real;
}

double *starpu_complex_get_imaginary(starpu_data_handle_t handle);
int starpu_complex_get_nx(starpu_data_handle_t handle);
\endcode

Similar functions need to be defined to access the different fields of the
complex interface from a <c>void *</c> pointer to be used within codelet
implemetations.

\snippet complex.c To be included. You should update doxygen if you see this text.

Complex data interfaces can then be registered to StarPU.

\code{.c}
double real = 45.0;
double imaginary = 12.0;
starpu_complex_data_register(&handle1, STARPU_MAIN_RAM, &real, &imaginary, 1);
starpu_task_insert(&cl_display, STARPU_R, handle1, 0);
\endcode

and used by codelets.

\code{.c}
void display_complex_codelet(void *descr[], void *_args)
{
        int nx = STARPU_COMPLEX_GET_NX(descr[0]);
        double *real = STARPU_COMPLEX_GET_REAL(descr[0]);
        double *imaginary = STARPU_COMPLEX_GET_IMAGINARY(descr[0]);
        int i;

        for(i=0 ; i<nx ; i++)
        {
                fprintf(stderr, "Complex[%d] = %3.2f + %3.2f i\n", i, real[i], imaginary[i]);
        }
}
\endcode

The whole code for this complex data interface is available in the
directory <c>examples/interface/</c>.


\subsection DefiningANewDataInterface_allocation Data allocation

To be able to run tasks on GPUs etc. StarPU needs to know how to allocate a
buffer for the interface. In our example, two allocations are needed in the
allocation method \c complex_allocate_data_on_node(): one for the real part and one
for the imaginary part.

\code{.c}
static starpu_ssize_t complex_allocate_data_on_node(void *data_interface, unsigned node)
{
	struct starpu_complex_interface *complex_interface = (struct starpu_complex_interface *) data_interface;

	double *addr_real = NULL;
	double *addr_imaginary = NULL;
	starpu_ssize_t requested_memory = complex_interface->nx * sizeof(complex_interface->real[0]);

	addr_real = (double*) starpu_malloc_on_node(node, requested_memory);
	if (!addr_real)
		goto fail_real;
	addr_imaginary = (double*) starpu_malloc_on_node(node, requested_memory);
	if (!addr_imaginary)
		goto fail_imaginary;

	/* update the data properly in consequence */
	complex_interface->real = addr_real;
	complex_interface->imaginary = addr_imaginary;

	return 2*requested_memory;

fail_imaginary:
	starpu_free_on_node(node, (uintptr_t) addr_real, requested_memory);
fail_real:
	return -ENOMEM;
}
\endcode

Here we try to allocate the two parts. If either of them fails, we return
\c -ENOMEM. If they succeed, we can record the obtained pointers and returned the
amount of allocated memory (for memory usage accounting).

Conversely, \c complex_free_data_on_node() frees the two parts:

\code{.c}
static void complex_free_data_on_node(void *data_interface, unsigned node)
{
	struct starpu_complex_interface *complex_interface = (struct starpu_complex_interface *) data_interface;
	starpu_ssize_t requested_memory = complex_interface->nx * sizeof(complex_interface->real[0]);

	starpu_free_on_node(node, (uintptr_t) complex_interface->real, requested_memory);
	starpu_free_on_node(node, (uintptr_t) complex_interface->imaginary, requested_memory);
}
\endcode

We we have not made anything particular for GPUs or whatsoever: it is
starpu_free_on_node() which knows how to actually make the allocation, and
returns the resulting pointer, be it in main memory, in GPU memory, etc.

\subsection DefiningANewDataInterface_copy Data copy

Now that StarPU knows how to allocate/free a buffer, it needs to be able to
copy over data into/from it. Defining a method \c copy_any_to_any() allows StarPU to
perform direct transfers between main memory and GPU memory.

\code{.c}
static int copy_any_to_any(void *src_interface, unsigned src_node,
			   void *dst_interface, unsigned dst_node,
			   void *async_data)
{
	struct starpu_complex_interface *src_complex = src_interface;
	struct starpu_complex_interface *dst_complex = dst_interface;
	int ret = 0;


	if (starpu_interface_copy((uintptr_t) src_complex->real, 0, src_node,
				    (uintptr_t) dst_complex->real, 0, dst_node,
				     src_complex->nx*sizeof(src_complex->real[0]),
				     async_data))
		ret = -EAGAIN;
	if (starpu_interface_copy((uintptr_t) src_complex->imaginary, 0, src_node,
				    (uintptr_t) dst_complex->imaginary, 0, dst_node,
				     src_complex->nx*sizeof(src_complex->imaginary[0]),
				     async_data))
		ret = -EAGAIN;
	return ret;
}
\endcode

We here again have no idea what is main memory or GPU memory, or even if the
copy is synchronous or asynchronous: we just call starpu_interface_copy()
according to the interface, passing it the pointers, and checking whether it
returned \c -EAGAIN, which means the copy is asynchronous, and StarPU will
appropriately wait for it thanks to the pointer \c async_data.

This copy method is referenced in a structure \ref starpu_data_copy_methods

\code{.c}
static const struct starpu_data_copy_methods complex_copy_methods =
{
	.any_to_any = copy_any_to_any
};
\endcode

which was referenced in the structure \ref starpu_data_interface_ops above.

Other fields of \ref starpu_data_copy_methods allow to provide optimized
variants, notably for the case of 2D or 3D matrix tiles with non-trivial ld.

\subsection DefiningANewDataInterface_pack Data pack/peek/unpack

The copy methods allow for RAM/GPU transfers, but is not enough for e.g.
transferring over MPI. That requires defining the pack/peek/unpack methods. The
principle is that the method starpu_data_interface_ops::pack_data concatenates
the buffer data into a newly-allocated contiguous bytes array, conversely
starpu_data_interface_ops::peek_data extracts from a bytes array into the
buffer data, and starpu_data_interface_ops::unpack_data does the same as
starpu_data_interface_ops::peek_data but also frees the bytes array.

\code{.c}
static int complex_pack_data(starpu_data_handle_t handle, unsigned node, void **ptr, starpu_ssize_t *count)
{
	STARPU_ASSERT(starpu_data_test_if_allocated_on_node(handle, node));

	struct starpu_complex_interface *complex_interface = (struct starpu_complex_interface *)
		starpu_data_get_interface_on_node(handle, node);

	*count = complex_get_size(handle);
	if (ptr != NULL)
	{
		char *data;
		data = (void*) starpu_malloc_on_node_flags(node, *count, 0);
		*ptr = data;
		memcpy(data, complex_interface->real, complex_interface->nx*sizeof(double));
		memcpy(data+complex_interface->nx*sizeof(double), complex_interface->imaginary, complex_interface->nx*sizeof(double));
	}

	return 0;
}
\endcode

\c complex_pack_data() first computes the size to be allocated, then allocates it,
and copies over into it the content of the two real and imaginary arrays.

\code{.c}
static int complex_peek_data(starpu_data_handle_t handle, unsigned node, void *ptr, size_t count)
{
	char *data = ptr;
	STARPU_ASSERT(starpu_data_test_if_allocated_on_node(handle, node));

	struct starpu_complex_interface *complex_interface = (struct starpu_complex_interface *)
		starpu_data_get_interface_on_node(handle, node);

	STARPU_ASSERT(count == 2 * complex_interface->nx * sizeof(double));
	memcpy(complex_interface->real, data, complex_interface->nx*sizeof(double));
	memcpy(complex_interface->imaginary, data+complex_interface->nx*sizeof(double), complex_interface->nx*sizeof(double));

	return 0;
}
\endcode

\c complex_peek_data() simply uses \c memcpy() to copy over from the bytes array into the data buffer.

\code{.c}
static int complex_unpack_data(starpu_data_handle_t handle, unsigned node, void *ptr, size_t count)
{
	complex_peek_data(handle, node, ptr, count);

	starpu_free_on_node_flags(node, (uintptr_t) ptr, count, 0);

	return 0;
}
\endcode

And \c complex_unpack_data() just calls \c complex_peek_data() and releases the bytes array.


\section SpecifyingATargetNode Specifying A Target Node For Task Data

When executing a task on a GPU for instance, StarPU would normally copy all the
needed data for the tasks on the embedded memory of the GPU.  It may however
happen that the task kernel would rather have some of the datas kept in the
main memory instead of copied in the GPU, a pivoting vector for instance.
This can be achieved by setting the flag starpu_codelet::specific_nodes to
<c>1</c>, and then fill the array starpu_codelet::nodes (or starpu_codelet::dyn_nodes when
starpu_codelet::nbuffers is greater than \ref STARPU_NMAXBUFS) with the node numbers
where data should be copied to, or ::STARPU_SPECIFIC_NODE_LOCAL to let
StarPU copy it to the memory node where the task will be executed.

::STARPU_SPECIFIC_NODE_CPU can also be used to request data to be
put in CPU-accessible memory (and let StarPU choose the NUMA node).
::STARPU_SPECIFIC_NODE_FAST and ::STARPU_SPECIFIC_NODE_SLOW can also be
used

For instance,
with the following codelet:

\code{.c}
struct starpu_codelet cl =
{
	.cuda_funcs = { kernel },
	.nbuffers = 2,
	.modes = {STARPU_RW, STARPU_RW},
	.specific_nodes = 1,
	.nodes = {STARPU_SPECIFIC_NODE_CPU, STARPU_SPECIFIC_NODE_LOCAL},
};
\endcode

the first data of the task will be kept in the CPU memory, while the second
data will be copied to the CUDA GPU as usual. A working example is available in
<c>tests/datawizard/specific_node.c</c>

With the following codelet:

\code{.c}
struct starpu_codelet cl =
{
	.cuda_funcs = { kernel },
	.nbuffers = 2,
	.modes = {STARPU_RW, STARPU_RW},
	.specific_nodes = 1,
	.nodes = {STARPU_SPECIFIC_NODE_LOCAL, STARPU_SPECIFIC_NODE_SLOW},
};
\endcode

The first data will be copied into fast (but probably size-limited) local memory
while the second data will be left in slow (but large) memory. This makes sense
when the kernel does not make so many accesses to the second data, and thus data
being remote e.g. over a PCI bus is not a performance problem, and avoids
filling the fast local memory with data which does not need the performance.

In cases where the kernel is fine with some data being either local or in the
main memory, ::STARPU_SPECIFIC_NODE_LOCAL_OR_CPU can be used. StarPU will then
be free to leave the data in the main memory and let the kernel access it from
accelerators, or to move it to the accelerator before starting the kernel, for
instance:

\code{.c}
struct starpu_codelet cl =
{
	.cuda_funcs = { kernel },
	.nbuffers = 2,
	.modes = {STARPU_RW, STARPU_R},
	.specific_nodes = 1,
	.nodes = {STARPU_SPECIFIC_NODE_LOCAL, STARPU_SPECIFIC_NODE_LOCAL_OR_CPU},
};
\endcode

*/