starpu-max/doc/starpu.texi

\input texinfo @c -*-texinfo-*-

@c %**start of header
@setfilename starpu.info
@settitle StarPU Handbook
@c %**end of header

@include version.texi

@setchapternewpage odd
@dircategory Development
@direntry
* StarPU: (starpu).             StarPU Handbook
@end direntry

@titlepage
@title StarPU Handbook
@subtitle for StarPU @value{VERSION}
@page
@vskip 0pt plus 1fill
@comment For the @value{version-GCC} Version*
@end titlepage

@c @summarycontents
@contents
@page

@node Top
@top Preface
@cindex Preface

This manual documents the usage of StarPU version @value{VERSION}.  It
was last updated on @value{UPDATED}.


@comment
@comment  When you add a new menu item, please keep the right hand
@comment  aligned to the same column.  Do not use tabs.  This provides
@comment  better formatting.
@comment
@menu
* Introduction::                A basic introduction to using StarPU
* Installing StarPU::           How to configure, build and install StarPU
* Using StarPU::                How to run StarPU application
* Basic Examples::              Basic examples of the use of StarPU
* Performance optimization::    How to optimize performance with StarPU
* Performance feedback::        Performance debugging tools
* StarPU MPI support::          How to combine StarPU with MPI
* Configuring StarPU::          How to configure StarPU
* StarPU API::                  The API to use StarPU
* Advanced Topics::             Advanced use of StarPU
* Full source code for the 'Scaling a Vector' example::  

* Function Index::              Index of C functions.
@end menu

@c ---------------------------------------------------------------------
@c Introduction to StarPU
@c ---------------------------------------------------------------------

@node Introduction
@chapter Introduction to StarPU

@menu
* Motivation::                  Why StarPU ?
* StarPU in a Nutshell::        The Fundamentals of StarPU
@end menu

@node Motivation
@section Motivation

@c complex machines with heterogeneous cores/devices
The use of specialized hardware such as accelerators or coprocessors offers an
interesting approach to overcome the physical limits encountered by processor
architects. As a result, many machines are now equipped with one or several
accelerators (e.g. a GPU), in addition to the usual processor(s). While a lot of
efforts have been devoted to offload computation onto such accelerators, very
little attention as been paid to portability concerns on the one hand, and to the
possibility of having heterogeneous accelerators and processors to interact on the other hand.

StarPU is a runtime system that offers support for heterogeneous multicore
architectures, it not only offers a unified view of the computational resources
(i.e. CPUs and accelerators at the same time), but it also takes care of
efficiently mapping and executing tasks onto an heterogeneous machine while
transparently handling low-level issues such as data transfers in a portable
fashion.

@c this leads to a complicated distributed memory design
@c which is not (easily) manageable by hand

@c added value/benefits of StarPU
@c   - portability
@c   - scheduling, perf. portability

@node StarPU in a Nutshell
@section StarPU in a Nutshell

@menu
* Codelet and Tasks::           
* StarPU Data Management Library::  
* Glossary::
* Research Papers::
@end menu

From a programming point of view, StarPU is not a new language but a library
that executes tasks explicitly submitted by the application.  The data that a
task manipulates are automatically transferred onto the accelerator so that the
programmer does not have to take care of complex data movements.  StarPU also
takes particular care of scheduling those tasks efficiently and allows
scheduling experts to implement custom scheduling policies in a portable
fashion.

@c explain the notion of codelet and task (i.e. g(A, B)
@node Codelet and Tasks
@subsection Codelet and Tasks

One of the StarPU primary data structures is the @b{codelet}. A codelet describes a
computational kernel that can possibly be implemented on multiple architectures
such as a CPU, a CUDA device or a Cell's SPU.

@c TODO insert illustration f : f_spu, f_cpu, ...

Another important data structure is the @b{task}. Executing a StarPU task
consists in applying a codelet on a data set, on one of the architectures on
which the codelet is implemented. A task thus describes the codelet that it
uses, but also which data are accessed, and how they are
accessed during the computation (read and/or write).
StarPU tasks are asynchronous: submitting a task to StarPU is a non-blocking
operation. The task structure can also specify a @b{callback} function that is
called once StarPU has properly executed the task. It also contains optional
fields that the application may use to give hints to the scheduler (such as
priority levels).

By default, task dependencies are inferred from data dependency (sequential
coherence) by StarPU. The application can however disable sequential coherency
for some data, and dependencies be expressed by hand.
A task may be identified by a unique 64-bit number chosen by the application
which we refer as a @b{tag}.
Task dependencies can be enforced by hand either by the means of callback functions, by
submitting other tasks, or by expressing dependencies
between tags (which can thus correspond to tasks that have not been submitted
yet).

@c TODO insert illustration f(Ar, Brw, Cr) + ..

@c DSM
@node StarPU Data Management Library
@subsection StarPU Data Management Library

Because StarPU schedules tasks at runtime, data transfers have to be
done automatically and ``just-in-time'' between processing units,
relieving the application programmer from explicit data transfers.
Moreover, to avoid unnecessary transfers, StarPU keeps data
where it was last needed, even if was modified there, and it
allows multiple copies of the same data to reside at the same time on
several processing units as long as it is not modified.

@node Glossary
@subsection Glossary

A @b{codelet} records pointers to various implementations of the same
theoretical function.

A @b{memory node} can be either the main RAM or GPU-embedded memory.

A @b{bus} is a link between memory nodes.

A @b{data handle} keeps track of replicates of the same data (@b{registered} by the
application) over various memory nodes. The data management library manages
keeping them coherent.

The @b{home} memory node of a data handle is the memory node from which the data
was registered (usually the main memory node).

A @b{task} represents a scheduled execution of a codelet on some data handles.

A @b{tag} is a rendez-vous point. Tasks typically have their own tag, and can
depend on other tags. The value is chosen by the application.

A @b{worker} execute tasks. There is typically one per CPU computation core and
one per accelerator (for which a whole CPU core is dedicated).

A @b{driver} drives a given kind of workers. There are currently CPU, CUDA,
OpenCL and Gordon drivers. They usually start several workers to actually drive
them.

A @b{performance model} is a (dynamic or static) model of the performance of a
given codelet. Codelets can have execution time performance model as well as
power consumption performance models.

A data @b{interface} describes the layout of the data: for a vector, a pointer
for the start, the number of elements and the size of elements ; for a matrix, a
pointer for the start, the number of elements per row, the offset between rows,
and the size of each element ; etc. To access their data, codelet functions are
given interfaces for the local memory node replicates of the data handles of the
scheduled task.

@b{Partitioning} data means dividing the data of a given data handle (called
@b{father}) into a series of @b{children} data handles which designate various
portions of the former.

A @b{filter} is the function which computes children data handles from a father
data handle, and thus describes how the partitioning should be done (horizontal,
vertical, etc.)

@b{Acquiring} a data handle can be done from the main application, to safely
access the data of a data handle from its home node, without having to
unregister it.


@node Research Papers
@subsection Research Papers

Research papers about StarPU can be found at

@indicateurl{http://runtime.bordeaux.inria.fr/Publis/Keyword/STARPU.html}

Notably a good overview in the research report

@indicateurl{http://hal.archives-ouvertes.fr/inria-00467677}

@c ---------------------------------------------------------------------
@c Installing StarPU
@c ---------------------------------------------------------------------

@node Installing StarPU
@chapter Installing StarPU

@menu
* Downloading StarPU::          
* Configuration of StarPU::     
* Building and Installing StarPU::  
@end menu

StarPU can be built and installed by the standard means of the GNU
autotools. The following chapter is intended to briefly remind how these tools
can be used to install StarPU.

@node Downloading StarPU
@section Downloading StarPU

@menu
* Getting Sources::             
* Optional dependencies::       
@end menu

@node Getting Sources
@subsection Getting Sources

The simplest way to get StarPU sources is to download the latest official
release tarball from @indicateurl{https://gforge.inria.fr/frs/?group_id=1570} ,
or the latest nightly snapshot from
@indicateurl{http://starpu.gforge.inria.fr/testing/} . The following documents
how to get the very latest version from the subversion repository itself, it
should be needed only if you need the very latest changes (i.e. less than a
day!)

The source code is managed by a Subversion server hosted by the
InriaGforge. To get the source code, you need:

@itemize
@item
To install the client side of the software Subversion if it is
not already available on your system. The software can be obtained from
@indicateurl{http://subversion.tigris.org} . If you are running
on Windows, you will probably prefer to use TortoiseSVN from
@indicateurl{http://tortoisesvn.tigris.org/} .

@item
You can check out the project's SVN repository through anonymous
access. This will provide you with a read access to the
repository.

If you need to have write access on the StarPU project, you can also choose to
become a member of the project @code{starpu}.  For this, you first need to get
an account to the gForge server. You can then send a request to join the project
(@indicateurl{https://gforge.inria.fr/project/request.php?group_id=1570}).

@item
More information on how to get a gForge account, to become a member of
a project, or on any other related task can be obtained from the
InriaGforge at @indicateurl{https://gforge.inria.fr/}. The most important
thing is to upload your public SSH key on the gForge server (see the
FAQ at @indicateurl{http://siteadmin.gforge.inria.fr/FAQ.html#Q6} for
instructions).
@end itemize

You can now check out the latest version from the Subversion server:
@itemize
@item
using the anonymous access via svn:
@example
% svn checkout svn://scm.gforge.inria.fr/svn/starpu/trunk
@end example
@item
using the anonymous access via https:
@example
% svn checkout --username anonsvn https://scm.gforge.inria.fr/svn/starpu/trunk
@end example
The password is @code{anonsvn}.
@item
using your gForge account
@example
% svn checkout svn+ssh://<login>@@scm.gforge.inria.fr/svn/starpu/trunk
@end example
@end itemize

The following step requires the availability of @code{autoconf} and
@code{automake} to generate the @code{./configure} script. This is
done by calling @code{./autogen.sh}. The required version for
@code{autoconf} is 2.60 or higher. You will also need @code{makeinfo}.

@example
% ./autogen.sh
@end example

If the autotools are not available on your machine or not recent
enough, you can choose to download the latest nightly tarball, which
is provided with a @code{configure} script.

@example
% wget http://starpu.gforge.inria.fr/testing/starpu-nightly-latest.tar.gz
@end example

@node Optional dependencies
@subsection Optional dependencies

The topology discovery library, @code{hwloc}, is not mandatory to use StarPU
but strongly recommended. It allows to increase performance, and to
perform some topology aware scheduling.

@code{hwloc} is available in major distributions and for most OSes and can be
downloaded from @indicateurl{http://www.open-mpi.org/software/hwloc}.

@node Configuration of StarPU
@section Configuration of StarPU

@menu
* Generating Makefiles and configuration scripts::  
* Running the configuration::   
@end menu

@node Generating Makefiles and configuration scripts
@subsection Generating Makefiles and configuration scripts

This step is not necessary when using the tarball releases of StarPU.  If you
are using the source code from the svn repository, you first need to generate
the configure scripts and the Makefiles.

@example
% ./autogen.sh
@end example

@node Running the configuration
@subsection Running the configuration

@example
% ./configure
@end example

Details about options that are useful to give to @code{./configure} are given in
@ref{Compilation configuration}.

@node Building and Installing StarPU
@section Building and Installing StarPU

@menu
* Building::                    
* Sanity Checks::               
* Installing::                  
@end menu

@node Building
@subsection Building

@example
% make
@end example

@node Sanity Checks
@subsection Sanity Checks

In order to make sure that StarPU is working properly on the system, it is also
possible to run a test suite.

@example
% make check
@end example

@node Installing
@subsection Installing

In order to install StarPU at the location that was specified during
configuration:

@example
% make install
@end example

@c ---------------------------------------------------------------------
@c Using StarPU
@c ---------------------------------------------------------------------

@node Using StarPU
@chapter Using StarPU

@menu
* Setting flags for compiling and linking applications::  
* Running a basic StarPU application::  
* Kernel threads started by StarPU::
* Using accelerators::          
@end menu

@node Setting flags for compiling and linking applications
@section Setting flags for compiling and linking applications

Compiling and linking an application against StarPU may require to use
specific flags or libraries (for instance @code{CUDA} or @code{libspe2}).
To this end, it is possible to use the @code{pkg-config} tool.

If StarPU was not installed at some standard location, the path of StarPU's
library must be specified in the @code{PKG_CONFIG_PATH} environment variable so
that @code{pkg-config} can find it. For example if StarPU was installed in
@code{$prefix_dir}:

@example
% PKG_CONFIG_PATH=$PKG_CONFIG_PATH:$prefix_dir/lib/pkgconfig
@end example

The flags required to compile or link against StarPU are then
accessible with the following commands:

@example
% pkg-config --cflags libstarpu  # options for the compiler
% pkg-config --libs libstarpu    # options for the linker
@end example

@node Running a basic StarPU application
@section Running a basic StarPU application

Basic examples using StarPU are built in the directory
@code{examples/basic_examples/} (and installed in
@code{$prefix_dir/lib/starpu/examples/}). You can for example run the example
@code{vector_scal}.

@example
% ./examples/basic_examples/vector_scal
BEFORE : First element was 1.000000
AFTER First element is 3.140000
%
@end example

When StarPU is used for the first time, the directory
@code{$HOME/.starpu/} is created, performance models will be stored in
that directory.

Please note that buses are benchmarked when StarPU is launched for the
first time. This may take a few minutes, or less if @code{hwloc} is
installed. This step is done only once per user and per machine.

@node Kernel threads started by StarPU
@section Kernel threads started by StarPU

TODO: StarPU starts one thread per CPU core and binds them there, uses one of
them per GPU. The application is not supposed to do computations in its own
threads. TODO: add a StarPU function to bind an application thread (e.g. the
main thread) to a dedicated core (and thus disable the corresponding StarPU CPU
worker).

@node Using accelerators
@section Using accelerators

When both CUDA and OpenCL drivers are enabled, StarPU will launch an
OpenCL worker for NVIDIA GPUs only if CUDA is not already running on them.
This design choice was necessary as OpenCL and CUDA can not run at the
same time on the same NVIDIA GPU, as there is currently no interoperability
between them.

Details on how to specify devices running OpenCL and the ones running
CUDA are given in @ref{Enabling OpenCL}.


@c ---------------------------------------------------------------------
@c Basic Examples
@c ---------------------------------------------------------------------

@node Basic Examples
@chapter Basic Examples

@menu
* Compiling and linking options::  
* Hello World::                 Submitting Tasks
* Scaling a Vector::            Manipulating Data
* Vector Scaling on an Hybrid CPU/GPU Machine::  Handling Heterogeneous Architectures
* Task and Worker Profiling::   
* Partitioning Data::           Partitioning Data
* Performance model example::   
* Theoretical lower bound on execution time::  
* Insert Task Utility::          
* More examples::               More examples shipped with StarPU
* Debugging::                   When things go wrong.
@end menu

@node Compiling and linking options
@section Compiling and linking options

Let's suppose StarPU has been installed in the directory
@code{$STARPU_DIR}. As explained in @ref{Setting flags for compiling and linking applications},
the variable @code{PKG_CONFIG_PATH} needs to be set. It is also
necessary to set the variable @code{LD_LIBRARY_PATH} to locate dynamic
libraries at runtime.

@example
% PKG_CONFIG_PATH=$STARPU_DIR/lib/pkgconfig:$PKG_CONFIG_PATH
% LD_LIBRARY_PATH=$STARPU_DIR/lib:$LD_LIBRARY_PATH
@end example

The Makefile could for instance contain the following lines to define which
options must be given to the compiler and to the linker:

@cartouche
@example
CFLAGS          +=      $$(pkg-config --cflags libstarpu)
LDFLAGS         +=      $$(pkg-config --libs libstarpu)
@end example
@end cartouche

@node Hello World
@section Hello World

@menu
* Required Headers::            
* Defining a Codelet::          
* Submitting a Task::           
* Execution of Hello World::    
@end menu

In this section, we show how to implement a simple program that submits a task to StarPU.

@node Required Headers
@subsection Required Headers

The @code{starpu.h} header should be included in any code using StarPU.

@cartouche
@smallexample
#include <starpu.h>
@end smallexample
@end cartouche


@node Defining a Codelet
@subsection Defining a Codelet

@cartouche
@smallexample
struct params @{
    int i;
    float f;
@};
void cpu_func(void *buffers[], void *cl_arg)
@{
    struct params *params = cl_arg;

    printf("Hello world (params = @{%i, %f@} )\n", params->i, params->f);
@}

starpu_codelet cl =
@{
    .where = STARPU_CPU,
    .cpu_func = cpu_func,
    .nbuffers = 0
@};
@end smallexample
@end cartouche

A codelet is a structure that represents a computational kernel. Such a codelet
may contain an implementation of the same kernel on different architectures
(e.g. CUDA, Cell's SPU, x86, ...).

The @code{nbuffers} field specifies the number of data buffers that are
manipulated by the codelet: here the codelet does not access or modify any data
that is controlled by our data management library. Note that the argument
passed to the codelet (the @code{cl_arg} field of the @code{starpu_task}
structure) does not count as a buffer since it is not managed by our data
management library, but just contain trivial parameters.

@c TODO need a crossref to the proper description of "where" see bla for more ...
We create a codelet which may only be executed on the CPUs. The @code{where}
field is a bitmask that defines where the codelet may be executed. Here, the
@code{STARPU_CPU} value means that only CPUs can execute this codelet
(@pxref{Codelets and Tasks} for more details on this field).
When a CPU core executes a codelet, it calls the @code{cpu_func} function,
which @emph{must} have the following prototype:

@code{void (*cpu_func)(void *buffers[], void *cl_arg);}

In this example, we can ignore the first argument of this function which gives a
description of the input and output buffers (e.g. the size and the location of
the matrices) since there is none.
The second argument is a pointer to a buffer passed as an
argument to the codelet by the means of the @code{cl_arg} field of the
@code{starpu_task} structure.

@c TODO rewrite so that it is a little clearer ?
Be aware that this may be a pointer to a
@emph{copy} of the actual buffer, and not the pointer given by the programmer:
if the codelet modifies this buffer, there is no guarantee that the initial
buffer will be modified as well: this for instance implies that the buffer
cannot be used as a synchronization medium. If synchronization is needed, data
has to be registered to StarPU, see @ref{Scaling a Vector}.

@node Submitting a Task
@subsection Submitting a Task

@cartouche
@smallexample
void callback_func(void *callback_arg)
@{
    printf("Callback function (arg %x)\n", callback_arg);
@}

int main(int argc, char **argv)
@{
    /* @b{initialize StarPU} */
    starpu_init(NULL);

    struct starpu_task *task = starpu_task_create();

    task->cl = &cl; /* @b{Pointer to the codelet defined above} */

    struct params params = @{ 1, 2.0f @};
    task->cl_arg = &params;
    task->cl_arg_size = sizeof(params);

    task->callback_func = callback_func;
    task->callback_arg = 0x42;

    /* @b{starpu_task_submit will be a blocking call} */
    task->synchronous = 1;

    /* @b{submit the task to StarPU} */
    starpu_task_submit(task);

    /* @b{terminate StarPU} */
    starpu_shutdown();

    return 0;
@}
@end smallexample
@end cartouche

Before submitting any tasks to StarPU, @code{starpu_init} must be called. The
@code{NULL} argument specifies that we use default configuration. Tasks cannot
be submitted after the termination of StarPU by a call to
@code{starpu_shutdown}.

In the example above, a task structure is allocated by a call to
@code{starpu_task_create}. This function only allocates and fills the
corresponding structure with the default settings (@pxref{Codelets and
Tasks, starpu_task_create}), but it does not submit the task to StarPU.

@c not really clear ;)
The @code{cl} field is a pointer to the codelet which the task will
execute: in other words, the codelet structure describes which computational
kernel should be offloaded on the different architectures, and the task
structure is a wrapper containing a codelet and the piece of data on which the
codelet should operate.

The optional @code{cl_arg} field is a pointer to a buffer (of size
@code{cl_arg_size}) with some parameters for the kernel
described by the codelet. For instance, if a codelet implements a computational
kernel that multiplies its input vector by a constant, the constant could be
specified by the means of this buffer, instead of registering it as a StarPU
data. It must however be noted that StarPU avoids making copy whenever possible
and rather passes the pointer as such, so the buffer which is pointed at must
kept allocated until the task terminates, and if several tasks are submitted
with various parameters, each of them must be given a pointer to their own
buffer.

Once a task has been executed, an optional callback function is be called.
While the computational kernel could be offloaded on various architectures, the
callback function is always executed on a CPU. The @code{callback_arg}
pointer is passed as an argument of the callback. The prototype of a callback
function must be:

@code{void (*callback_function)(void *);}

If the @code{synchronous} field is non-zero, task submission will be
synchronous: the @code{starpu_task_submit} function will not return until the
task was executed. Note that the @code{starpu_shutdown} method does not
guarantee that asynchronous tasks have been executed before it returns,
@code{starpu_task_wait_for_all} can be used to that effect, or data can be
unregistered (@code{starpu_data_unregister(vector_handle);}), which will
implicitly wait for all the tasks scheduled to work on it, unless explicitly
disabled thanks to @code{starpu_data_set_default_sequential_consistency_flag} or
@code{starpu_data_set_sequential_consistency_flag}.

@node Execution of Hello World
@subsection Execution of Hello World

@smallexample
% make hello_world
cc $(pkg-config --cflags libstarpu)  $(pkg-config --libs libstarpu) hello_world.c -o hello_world
% ./hello_world
Hello world (params = @{1, 2.000000@} )
Callback function (arg 42)
@end smallexample

@node Scaling a Vector
@section Manipulating Data: Scaling a Vector

The previous example has shown how to submit tasks. In this section,
we show how StarPU tasks can manipulate data. The full source code for
this example is given in @ref{Full source code for the 'Scaling a Vector' example}.

@menu
* Source code of Vector Scaling::  
* Execution of Vector Scaling::  
@end menu

@node Source code of Vector Scaling
@subsection Source code of Vector Scaling

Programmers can describe the data layout of their application so that StarPU is
responsible for enforcing data coherency and availability across the machine.
Instead of handling complex (and non-portable) mechanisms to perform data
movements, programmers only declare which piece of data is accessed and/or
modified by a task, and StarPU makes sure that when a computational kernel
starts somewhere (e.g. on a GPU), its data are available locally.

Before submitting those tasks, the programmer first needs to declare the
different pieces of data to StarPU using the @code{starpu_*_data_register}
functions. To ease the development of applications for StarPU, it is possible
to describe multiple types of data layout. A type of data layout is called an
@b{interface}. There are different predefined interfaces available in StarPU:
here we will consider the @b{vector interface}.

The following lines show how to declare an array of @code{NX} elements of type
@code{float} using the vector interface:

@cartouche
@smallexample
float vector[NX];

starpu_data_handle vector_handle;
starpu_vector_data_register(&vector_handle, 0, (uintptr_t)vector, NX,
                            sizeof(vector[0]));
@end smallexample
@end cartouche

The first argument, called the @b{data handle}, is an opaque pointer which
designates the array in StarPU. This is also the structure which is used to
describe which data is used by a task. The second argument is the node number
where the data originally resides. Here it is 0 since the @code{vector} array is in
the main memory. Then comes the pointer @code{vector} where the data can be found in main memory,
the number of elements in the vector and the size of each element.
The following shows how to construct a StarPU task that will manipulate the
vector and a constant factor.

@cartouche
@smallexample
float factor = 3.14;
struct starpu_task *task = starpu_task_create();

task->cl = &cl;                          /* @b{Pointer to the codelet defined below} */
task->buffers[0].handle = vector_handle; /* @b{First parameter of the codelet} */
task->buffers[0].mode = STARPU_RW;
task->cl_arg = &factor;
task->cl_arg_size = sizeof(factor);
task->synchronous = 1;

starpu_task_submit(task);
@end smallexample
@end cartouche

Since the factor is a mere constant float value parameter,
it does not need a preliminary registration, and
can just be passed through the @code{cl_arg} pointer like in the previous
example.  The vector parameter is described by its handle.
There are two fields in each element of the @code{buffers} array.
@code{handle} is the handle of the data, and @code{mode} specifies how the
kernel will access the data (@code{STARPU_R} for read-only, @code{STARPU_W} for
write-only and @code{STARPU_RW} for read and write access).

The definition of the codelet can be written as follows:

@cartouche
@smallexample
void scal_cpu_func(void *buffers[], void *cl_arg)
@{
    unsigned i;
    float *factor = cl_arg;

    /* length of the vector */
    unsigned n = STARPU_VECTOR_GET_NX(buffers[0]);
    /* CPU copy of the vector pointer */
    float *val = (float *)STARPU_VECTOR_GET_PTR(buffers[0]);

    for (i = 0; i < n; i++)
        val[i] *= *factor;
@}

starpu_codelet cl = @{
    .where = STARPU_CPU,
    .cpu_func = scal_cpu_func,
    .nbuffers = 1
@};
@end smallexample
@end cartouche

The first argument is an array that gives
a description of all the buffers passed in the @code{task->buffers}@ array. The
size of this array is given by the @code{nbuffers} field of the codelet
structure. For the sake of genericity, this array contains pointers to the
different interfaces describing each buffer.  In the case of the @b{vector
interface}, the location of the vector (resp. its length) is accessible in the
@code{ptr} (resp. @code{nx}) of this array. Since the vector is accessed in a
read-write fashion, any modification will automatically affect future accesses
to this vector made by other tasks.

The second argument of the @code{scal_cpu_func} function contains a pointer to the
parameters of the codelet (given in @code{task->cl_arg}), so that we read the
constant factor from this pointer.

@node Execution of Vector Scaling
@subsection Execution of Vector Scaling

@smallexample
% make vector_scal
cc $(pkg-config --cflags libstarpu)  $(pkg-config --libs libstarpu)  vector_scal.c   -o vector_scal
% ./vector_scal
0.000000 3.000000 6.000000 9.000000 12.000000
@end smallexample

@node Vector Scaling on an Hybrid CPU/GPU Machine
@section Vector Scaling on an Hybrid CPU/GPU Machine

Contrary to the previous examples, the task submitted in this example may not
only be executed by the CPUs, but also by a CUDA device.

@menu
* Definition of the CUDA Kernel::  
* Definition of the OpenCL Kernel::  
* Definition of the Main Code::  
* Execution of Hybrid Vector Scaling::  
@end menu

@node Definition of the CUDA Kernel
@subsection Definition of the CUDA Kernel

The CUDA implementation can be written as follows. It needs to be compiled with
a CUDA compiler such as nvcc, the NVIDIA CUDA compiler driver. It must be noted
that the vector pointer returned by STARPU_VECTOR_GET_PTR is here a pointer in GPU
memory, so that it can be passed as such to the @code{vector_mult_cuda} kernel
call.

@cartouche
@smallexample
#include <starpu.h>

static __global__ void vector_mult_cuda(float *val, unsigned n,
                                        float factor)
@{
    unsigned i =  blockIdx.x*blockDim.x + threadIdx.x;
    if (i < n)
        val[i] *= factor;
@}

extern "C" void scal_cuda_func(void *buffers[], void *_args)
@{
    float *factor = (float *)_args;

    /* length of the vector */
    unsigned n = STARPU_VECTOR_GET_NX(buffers[0]);
    /* CUDA copy of the vector pointer */
    float *val = (float *)STARPU_VECTOR_GET_PTR(buffers[0]);
    unsigned threads_per_block = 64;
    unsigned nblocks = (n + threads_per_block-1) / threads_per_block;

@i{    vector_mult_cuda<<<nblocks,threads_per_block, 0, starpu_cuda_get_local_stream()>>>(val, n, *factor);}

@i{    cudaStreamSynchronize(starpu_cuda_get_local_stream());}
@}
@end smallexample
@end cartouche

@node Definition of the OpenCL Kernel
@subsection Definition of the OpenCL Kernel

The OpenCL implementation can be written as follows. StarPU provides
tools to compile a OpenCL kernel stored in a file.

@cartouche
@smallexample
__kernel void vector_mult_opencl(__global float* val, int nx, float factor)
@{
        const int i = get_global_id(0);
        if (i < nx) @{
                val[i] *= factor;
        @}
@}
@end smallexample
@end cartouche

Similarly to CUDA, the pointer returned by @code{STARPU_VECTOR_GET_PTR} is here
a device pointer, so that it is passed as such to the OpenCL kernel.

@cartouche
@smallexample
#include <starpu.h>
@i{#include <starpu_opencl.h>}

@i{extern struct starpu_opencl_program programs;}

void scal_opencl_func(void *buffers[], void *_args)
@{
    float *factor = _args;
@i{    int id, devid, err;}
@i{    cl_kernel kernel;}
@i{    cl_command_queue queue;}
@i{    cl_event event;}

    /* length of the vector */
    unsigned n = STARPU_VECTOR_GET_NX(buffers[0]);
    /* OpenCL copy of the vector pointer */
    cl_mem val = (cl_mem) STARPU_VECTOR_GET_PTR(buffers[0]);

@i{    id = starpu_worker_get_id();}
@i{    devid = starpu_worker_get_devid(id);}

@i{    err = starpu_opencl_load_kernel(&kernel, &queue, &programs,}
@i{                    "vector_mult_opencl", devid);   /* @b{Name of the codelet defined above} */}
@i{    if (err != CL_SUCCESS) STARPU_OPENCL_REPORT_ERROR(err);}

@i{    err = clSetKernelArg(kernel, 0, sizeof(val), &val);}
@i{    err |= clSetKernelArg(kernel, 1, sizeof(n), &n);}
@i{    err |= clSetKernelArg(kernel, 2, sizeof(*factor), factor);}
@i{    if (err) STARPU_OPENCL_REPORT_ERROR(err);}

@i{    @{}
@i{        size_t global=1;}
@i{        size_t local=1;}
@i{        err = clEnqueueNDRangeKernel(queue, kernel, 1, NULL, &global, &local, 0, NULL, &event);}
@i{        if (err != CL_SUCCESS) STARPU_OPENCL_REPORT_ERROR(err);}
@i{    @}}

@i{    clFinish(queue);}
@i{    starpu_opencl_collect_stats(event);}
@i{    clReleaseEvent(event);}

@i{    starpu_opencl_release_kernel(kernel);}
@}
@end smallexample
@end cartouche


@node Definition of the Main Code
@subsection Definition of the Main Code

The CPU implementation is the same as in the previous section.

Here is the source of the main application. You can notice the value of the
field @code{where} for the codelet. We specify
@code{STARPU_CPU|STARPU_CUDA|STARPU_OPENCL} to indicate to StarPU that the codelet
can be executed either on a CPU or on a CUDA or an OpenCL device.

@cartouche
@smallexample
#include <starpu.h>

#define NX 2048

extern void scal_cuda_func(void *buffers[], void *_args);
extern void scal_cpu_func(void *buffers[], void *_args);
extern void scal_opencl_func(void *buffers[], void *_args);

/* @b{Definition of the codelet} */
static starpu_codelet cl = @{
    .where = STARPU_CPU|STARPU_CUDA|STARPU_OPENCL; /* @b{It can be executed on a CPU,} */
                                     /* @b{on a CUDA device, or on an OpenCL device} */
    .cuda_func = scal_cuda_func;
    .cpu_func = scal_cpu_func;
    .opencl_func = scal_opencl_func;
    .nbuffers = 1;
@}

#ifdef STARPU_USE_OPENCL
/* @b{The compiled version of the OpenCL program} */
struct starpu_opencl_program programs;
#endif

int main(int argc, char **argv)
@{
    float *vector;
    int i, ret;
    float factor=3.0;
    struct starpu_task *task;
    starpu_data_handle vector_handle;

    starpu_init(NULL);                            /* @b{Initialising StarPU} */

#ifdef STARPU_USE_OPENCL
    starpu_opencl_load_opencl_from_file(
            "examples/basic_examples/vector_scal_opencl_codelet.cl",
            &programs, NULL);
#endif

    vector = malloc(NX*sizeof(vector[0]));
    assert(vector);
    for(i=0 ; i<NX ; i++) vector[i] = i;
@end smallexample
@end cartouche

@cartouche
@smallexample
    /* @b{Registering data within StarPU} */
    starpu_vector_data_register(&vector_handle, 0, (uintptr_t)vector,
                                NX, sizeof(vector[0]));

    /* @b{Definition of the task} */
    task = starpu_task_create();
    task->cl = &cl;
    task->buffers[0].handle = vector_handle;
    task->buffers[0].mode = STARPU_RW;
    task->cl_arg = &factor;
    task->cl_arg_size = sizeof(factor);
@end smallexample
@end cartouche

@cartouche
@smallexample
    /* @b{Submitting the task} */
    ret = starpu_task_submit(task);
    if (ret == -ENODEV) @{
            fprintf(stderr, "No worker may execute this task\n");
            return 1;
    @}

@c TODO: Mmm, should rather be an unregistration with an implicit dependency, no?
    /* @b{Waiting for its termination} */
    starpu_task_wait_for_all();

    /* @b{Update the vector in RAM} */
    starpu_data_acquire(vector_handle, STARPU_R);
@end smallexample
@end cartouche

@cartouche
@smallexample
    /* @b{Access the data} */
    for(i=0 ; i<NX; i++) @{
      fprintf(stderr, "%f ", vector[i]);
    @}
    fprintf(stderr, "\n");

    /* @b{Release the RAM view of the data before unregistering it and shutting down StarPU} */
    starpu_data_release(vector_handle);
    starpu_data_unregister(vector_handle);
    starpu_shutdown();

    return 0;
@}
@end smallexample
@end cartouche

@node Execution of Hybrid Vector Scaling
@subsection Execution of Hybrid Vector Scaling

The Makefile given at the beginning of the section must be extended to
give the rules to compile the CUDA source code. Note that the source
file of the OpenCL kernel does not need to be compiled now, it will
be compiled at run-time when calling the function
@code{starpu_opencl_load_opencl_from_file()} (@pxref{starpu_opencl_load_opencl_from_file}).

@cartouche
@smallexample
CFLAGS	+=	$(shell pkg-config --cflags libstarpu)
LDFLAGS	+=	$(shell pkg-config --libs libstarpu)
CC	=	gcc

vector_scal: vector_scal.o vector_scal_cpu.o vector_scal_cuda.o vector_scal_opencl.o

%.o: %.cu
       nvcc $(CFLAGS) $< -c $@

clean:
       rm -f vector_scal *.o
@end smallexample
@end cartouche

@smallexample
% make
@end smallexample

and to execute it, with the default configuration:

@smallexample
% ./vector_scal
0.000000 3.000000 6.000000 9.000000 12.000000
@end smallexample

or for example, by disabling CPU devices:

@smallexample
% STARPU_NCPUS=0 ./vector_scal
0.000000 3.000000 6.000000 9.000000 12.000000
@end smallexample

or by disabling CUDA devices (which may permit to enable the use of OpenCL,
see @ref{Using accelerators}):

@smallexample
% STARPU_NCUDA=0 ./vector_scal
0.000000 3.000000 6.000000 9.000000 12.000000
@end smallexample

@node Task and Worker Profiling
@section Task and Worker Profiling

A full example showing how to use the profiling API is available in
the StarPU sources in the directory @code{examples/profiling/}.

@cartouche
@smallexample
struct starpu_task *task = starpu_task_create();
task->cl = &cl;
task->synchronous = 1;
/* We will destroy the task structure by hand so that we can
 * query the profiling info before the task is destroyed. */
task->destroy = 0;

/* Submit and wait for completion (since synchronous was set to 1) */
starpu_task_submit(task);

/* The task is finished, get profiling information */
struct starpu_task_profiling_info *info = task->profiling_info;

/* How much time did it take before the task started ? */
double delay += starpu_timing_timespec_delay_us(&info->submit_time, &info->start_time);

/* How long was the task execution ? */
double length += starpu_timing_timespec_delay_us(&info->start_time, &info->end_time);

/* We don't need the task structure anymore */
starpu_task_destroy(task);
@end smallexample
@end cartouche

@cartouche
@smallexample
/* Display the occupancy of all workers during the test */
int worker;
for (worker = 0; worker < starpu_worker_get_count(); worker++)
@{
        struct starpu_worker_profiling_info worker_info;
        int ret = starpu_worker_get_profiling_info(worker, &worker_info);
        STARPU_ASSERT(!ret);

        double total_time = starpu_timing_timespec_to_us(&worker_info.total_time);
        double executing_time = starpu_timing_timespec_to_us(&worker_info.executing_time);
        double sleeping_time = starpu_timing_timespec_to_us(&worker_info.sleeping_time);

        float executing_ratio = 100.0*executing_time/total_time;
        float sleeping_ratio = 100.0*sleeping_time/total_time;

        char workername[128];
        starpu_worker_get_name(worker, workername, 128);
        fprintf(stderr, "Worker %s:\n", workername);
        fprintf(stderr, "\ttotal time : %.2lf ms\n", total_time*1e-3);
        fprintf(stderr, "\texec time  : %.2lf ms (%.2f %%)\n", executing_time*1e-3,
                executing_ratio);
        fprintf(stderr, "\tblocked time  : %.2lf ms (%.2f %%)\n", sleeping_time*1e-3,
                sleeping_ratio);
@}
@end smallexample
@end cartouche

@node Partitioning Data
@section Partitioning Data

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

@cartouche
@smallexample
int vector[NX];
starpu_data_handle handle;

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

/* Partition the vector in PARTS sub-vectors */
starpu_filter f =
@{
    .filter_func = starpu_block_filter_func_vector,
    .nchildren = PARTS,
    .get_nchildren = NULL,
    .get_child_ops = NULL
@};
starpu_data_partition(handle, &f);
@end smallexample
@end cartouche

@cartouche
@smallexample
/* 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 sub_handle = starpu_data_get_sub_data(handle, 1, i);
    struct starpu_task *task = starpu_task_create();

    task->buffers[0].handle = sub_handle;
    task->buffers[0].mode = STARPU_RW;
    task->cl = &cl;
    task->synchronous = 1;
    task->cl_arg = &factor;
    task->cl_arg_size = sizeof(factor);

    starpu_task_submit(task);
@}
@end smallexample
@end cartouche

Partitioning can be applied several times, see
@code{examples/basic_examples/mult.c} and @code{examples/filters/}.

@node Performance model example
@section Performance model example

To achieve good scheduling, StarPU scheduling policies need to be able to
estimate in advance the duration of a task. This is done by giving to codelets a
performance model. There are several kinds of performance models.

@itemize
@item
Providing an estimation from the application itself (@code{STARPU_COMMON} model type and @code{cost_model} field),
see for instance
@code{examples/common/blas_model.h} and @code{examples/common/blas_model.c}. It can also be provided for each architecture (@code{STARPU_PER_ARCH} model type and @code{per_arch} field)
@item
Measured at runtime (STARPU_HISTORY_BASED model type). This assumes that for a
given set of data input/output sizes, the performance will always be about the
same. This is very true for regular kernels on GPUs for instance (<0.1% error),
and just a bit less true on CPUs (~=1% error). This also assumes that there are
few different sets of data input/output sizes. StarPU will then keep record of
the average time of previous executions on the various processing units, and use
it as an estimation. History is done per task size, by using a hash of the input
and ouput sizes as an index.
It will also save it in @code{~/.starpu/sampling/codelets}
for further executions, and can be observed by using the
@code{starpu_perfmodel_display} command. The models are indexed by machine name. To share the models between machines (e.g. for a homogeneous cluster), use @code{export STARPU_HOSTNAME=some_global_name}.  The following is a small code example.

If e.g. the code is recompiled with other compilation options, or several
variants of the code are used, the symbol string should be changed to reflect
that, in order to recalibrate the model from zero.

@cartouche
@smallexample
static struct starpu_perfmodel_t mult_perf_model = @{
    .type = STARPU_HISTORY_BASED,
    .symbol = "mult_perf_model"
@};

starpu_codelet cl = @{
    .where = STARPU_CPU,
    .cpu_func = cpu_mult,
    .nbuffers = 3,
    /* for the scheduling policy to be able to use performance models */
    .model = &mult_perf_model
@};
@end smallexample
@end cartouche

@item
Measured at runtime and refined by regression (STARPU_REGRESSION_*_BASED
model type). This still assumes performance regularity, but can work
with various data input sizes, by applying regression over observed
execution times. STARPU_REGRESSION_BASED uses an a*n^b regression
form, STARPU_NL_REGRESSION_BASED uses an a*n^b+c (more precise than
STARPU_REGRESSION_BASED, but costs a lot more to compute)

@item
Provided explicitly by the application (STARPU_PER_ARCH model type): the
@code{.per_arch[i].cost_model} fields have to be filled with pointers to
functions which return the expected duration of the task in micro-seconds, one
per architecture.

@end itemize

How to use schedulers which can benefit from such performance model is explained
in @ref{Task scheduling policy}.

The same can be done for task power consumption estimation, by setting the
@code{power_model} field the same way as the @code{model} field. Note: for
now, the application has to give to the power consumption performance model
a name which is different from the execution time performance model.

@node Theoretical lower bound on execution time
@section Theoretical lower bound on execution time

For kernels with history-based performance models, StarPU can very easily provide a theoretical lower
bound for the execution time of a whole set of tasks. See for
instance @code{examples/lu/lu_example.c}: before submitting tasks,
call @code{starpu_bound_start}, and after complete execution, call
@code{starpu_bound_stop}. @code{starpu_bound_print_lp} or
@code{starpu_bound_print_mps} can then be used to output a Linear Programming
problem corresponding to the schedule of your tasks. Run it through
@code{lp_solve} or any other linear programming solver, and that will give you a
lower bound for the total execution time of your tasks. If StarPU was compiled
with the glpk library installed, @code{starpu_bound_compute} can be used to
solve it immediately and get the optimized minimum. Its @code{integer}
parameter allows to decide whether integer resolution should be computed
and returned.

The @code{deps} parameter tells StarPU whether to take tasks and implicit data
dependencies into account. It must be understood that the linear programming
problem size is quadratic with the number of tasks and thus the time to solve it
will be very long, it could be minutes for just a few dozen tasks. You should
probably use @code{lp_solve -timeout 1 test.pl -wmps test.mps} to convert the
problem to MPS format and then use a better solver, @code{glpsol} might be
better than @code{lp_solve} for instance (the @code{--pcost} option may be
useful), but sometimes doesn't manage to converge. @code{cbc} might look
slower, but it is parallel. Be sure to try at least all the @code{-B} options
of @code{lp_solve}. For instance, we often just use
@code{lp_solve -cc -B1 -Bb -Bg -Bp -Bf -Br -BG -Bd -Bs -BB -Bo -Bc -Bi} , and
the @code{-gr} option can also be quite useful.

Setting @code{deps} to 0 will only take into account the actual computations
on processing units. It however still properly takes into account the varying
performances of kernels and processing units, which is quite more accurate than
just comparing StarPU performances with the fastest of the kernels being used.

The @code{prio} parameter tells StarPU whether to simulate taking into account
the priorities as the StarPU scheduler would, i.e. schedule prioritized
tasks before less prioritized tasks, to check to which extend this results
to a less optimal solution. This increases even more computation time.

Note that for simplicity, all this however doesn't take into account data
transfers, which are assumed to be completely overlapped.

@node Insert Task Utility
@section Insert Task Utility

StarPU provides the wrapper function @code{starpu_insert_task} to ease
the creation and submission of tasks.

@deftypefun int starpu_insert_task (starpu_codelet *@var{cl}, ...)
Create and submit a task corresponding to @var{cl} with the following
arguments.  The argument list must be zero-terminated.

The arguments following the codelets can be of the following types:

@itemize
@item
@code{STARPU_R}, @code{STARPU_W}, @code{STARPU_RW}, @code{STARPU_SCRATCH}, @code{STARPU_REDUX} an access mode followed by a data handle;
@item
@code{STARPU_VALUE} followed  by a pointer to a constant value and
the size of the constant;
@item
@code{STARPU_CALLBACK} followed by a pointer to a callback function;
@item
@code{STARPU_CALLBACK_ARG} followed by a pointer to be given as an
argument to the callback function;
@item
@code{STARPU_PRIORITY} followed by a integer defining a priority level.
@end itemize

Parameters to be passed to the codelet implementation are defined
through the type @code{STARPU_VALUE}. The function
@code{starpu_unpack_cl_args} must be called within the codelet
implementation to retrieve them.
@end deftypefun

Here the implementation of the codelet:

@smallexample
void func_cpu(void *descr[], void *_args)
@{
        int *x0 = (int *)STARPU_VARIABLE_GET_PTR(descr[0]);
        float *x1 = (float *)STARPU_VARIABLE_GET_PTR(descr[1]);
        int ifactor;
        float ffactor;

        starpu_unpack_cl_args(_args, &ifactor, &ffactor);
        *x0 = *x0 * ifactor;
        *x1 = *x1 * ffactor;
@}

starpu_codelet mycodelet = @{
        .where = STARPU_CPU,
        .cpu_func = func_cpu,
        .nbuffers = 2
@};
@end smallexample

And the call to the @code{starpu_insert_task} wrapper:

@smallexample
starpu_insert_task(&mycodelet,
                   STARPU_VALUE, &ifactor, sizeof(ifactor),
                   STARPU_VALUE, &ffactor, sizeof(ffactor),
                   STARPU_RW, data_handles[0], STARPU_RW, data_handles[1],
                   0);
@end smallexample

The call to @code{starpu_insert_task} is equivalent to the following
code:

@smallexample
struct starpu_task *task = starpu_task_create();
task->cl = &mycodelet;
task->buffers[0].handle = data_handles[0];
task->buffers[0].mode = STARPU_RW;
task->buffers[1].handle = data_handles[1];
task->buffers[1].mode = STARPU_RW;
char *arg_buffer;
size_t arg_buffer_size;
starpu_pack_cl_args(&arg_buffer, &arg_buffer_size,
		    STARPU_VALUE, &ifactor, sizeof(ifactor),
		    STARPU_VALUE, &ffactor, sizeof(ffactor),
		    0);
task->cl_arg = arg_buffer;
task->cl_arg_size = arg_buffer_size;
int ret = starpu_task_submit(task);
@end smallexample

@node Debugging
@section Debugging

StarPU provides several tools to help debugging aplications. Execution traces
can be generated and displayed graphically, see @ref{Generating traces}. Some
gdb helpers are also provided to show the whole StarPU state:

@smallexample
(gdb) source tools/gdbinit
(gdb) help starpu
@end smallexample

@node More examples
@section More examples

More examples are available in the StarPU sources in the @code{examples/}
directory. Simple examples include:

@table @asis
@item @code{incrementer/}:
	Trivial incrementation test.
@item @code{basic_examples/}:
        Simple documented Hello world (as shown in @ref{Hello World}), vector/scalar product (as shown
        in @ref{Vector Scaling on an Hybrid CPU/GPU Machine}), matrix
        product examples (as shown in @ref{Performance model example}), an example using the blocked matrix data
        interface, and an example using the variable data interface.
@item @code{matvecmult/}:
	OpenCL example from NVidia, adapted to StarPU.
@item @code{axpy/}:
	AXPY CUBLAS operation adapted to StarPU.
@item @code{fortran/}:
	Example of Fortran bindings.
@end table

More advanced examples include:

@table @asis
@item @code{filters/}:
	Examples using filters, as shown in @ref{Partitioning Data}.
@item @code{lu/}:
	LU matrix factorization, see for instance @code{xlu_implicit.c}
@item @code{cholesky/}:
	Cholesky matrix factorization, see for instance @code{cholesky_implicit.c}.
@end table

@c ---------------------------------------------------------------------
@c Performance options
@c ---------------------------------------------------------------------

@node Performance optimization
@chapter How to optimize performance with StarPU

TODO: improve!

@menu
* Data management::
* Task submission::
* Task priorities::
* Task scheduling policy::
* Performance model calibration::
* Task distribution vs Data transfer::
* Data prefetch::
* Power-based scheduling::
* Profiling::
* CUDA-specific optimizations::
@end menu

Simply encapsulating application kernels into tasks already permits to
seamlessly support CPU and GPUs at the same time. To achieve good performance, a
few additional changes are needed.

@node Data management
@section Data management

When the application allocates data, whenever possible it should use the
@code{starpu_malloc} function, which will ask CUDA or
OpenCL to make the allocation itself and pin the corresponding allocated
memory. This is needed to permit asynchronous data transfer, i.e. permit data
transfer to overlap with computations.

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 that 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.

@example
starpu_data_set_wt_mask(img_handle, 1<<0);
@end example

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

@node Task submission
@section Task submission

To let StarPU make online optimizations, tasks should be submitted
asynchronously as much as possible. Ideally, all the tasks should be
submitted, and mere calls to @code{starpu_task_wait_for_all} or
@code{starpu_data_unregister} be done to wait for
termination. StarPU will then be able to rework the whole schedule, overlap
computation with communication, manage accelerator local memory usage, etc.

@node Task priorities
@section Task priorities

By default, StarPU will consider the tasks in the order they are submitted by
the application. If the application programmer knows that some tasks should
be performed in priority (for instance because their output is needed by many
other tasks and may thus be a bottleneck if not executed early enough), the
@code{priority} field of the task structure should be set to transmit the
priority information to StarPU.

@node Task scheduling policy
@section Task scheduling policy

By default, StarPU uses the @code{eager} simple greedy scheduler. This is
because it provides correct load balance even if the application codelets do not
have performance models. If your application codelets have performance models
(@pxref{Performance model example} for examples showing how to do it),
you should change the scheduler thanks to the @code{STARPU_SCHED} environment
variable. For instance @code{export STARPU_SCHED=dmda} . Use @code{help} to get
the list of available schedulers.

The @b{eager} scheduler uses a central task queue, from which workers draw tasks
to work on. This however does not permit to prefetch data since the scheduling
decision is taken late. If a task has a non-0 priority, it is put at the front of the queue.

The @b{prio} scheduler also uses a central task queue, but sorts tasks by
priority (between -5 and 5).

The @b{random} scheduler distributes tasks randomly according to assumed worker
overall performance.

The @b{ws} (work stealing) scheduler schedules tasks on the local worker by
default. When a worker becomes idle, it steals a task from the most loaded
worker.

The @b{dm} (deque model) scheduler uses task execution performance models into account to
perform an HEFT-similar scheduling strategy: it schedules tasks where their
termination time will be minimal.

The @b{dmda} (deque model data aware) scheduler is similar to dm, it also takes
into account data transfer time.

The @b{dmdar} (deque model data aware ready) scheduler is similar to dmda,
it also sorts tasks on per-worker queues by number of already-available data
buffers.

The @b{dmdas} (deque model data aware sorted) scheduler is similar to dmda, it
also supports arbitrary priority values.

The @b{heft} (HEFT) scheduler is similar to dmda, it also supports task bundles.

The @b{pheft} (parallel HEFT) scheduler is similar to heft, it also supports
parallel tasks (still experimental).

The @b{pgreedy} (parallel greedy) scheduler is similar to greedy, it also
supports parallel tasks (still experimental).

@node Performance model calibration
@section Performance model calibration

Most schedulers are based on an estimation of codelet duration on each kind
of processing unit. For this to be possible, the application programmer needs
to configure a performance model for the codelets of the application (see
@ref{Performance model example} for instance). History-based performance models
use on-line calibration.  StarPU will automatically calibrate codelets
which have never been calibrated yet, and save the result in
@code{~/.starpu/sampling/codelets}.
The models are indexed by machine name. To share the models between machines (e.g. for a homogeneous cluster), use @code{export STARPU_HOSTNAME=some_global_name}. To force continuing calibration, use
@code{export STARPU_CALIBRATE=1} . This may be necessary if your application
has not-so-stable performance. Details on the current performance model status
can be obtained from the @code{starpu_perfmodel_display} command: the @code{-l}
option lists the available performance models, and the @code{-s} option permits
to choose the performance model to be displayed. The result looks like:

@example
$ starpu_perfmodel_display -s starpu_dlu_lu_model_22
performance model for cpu
# hash		size		mean		dev		n
5c6c3401	1572864        	1.216300e+04   	2.277778e+03   	1240
@end example

Which shows that for the LU 22 kernel with a 1.5MiB matrix, the average
execution time on CPUs was about 12ms, with a 2ms standard deviation, over
1240 samples. It is a good idea to check this before doing actual performance
measurements.

If a kernel source code was modified (e.g. performance improvement), the
calibration information is stale and should be dropped, to re-calibrate from
start. This can be done by using @code{export STARPU_CALIBRATE=2}.

Note: due to CUDA limitations, to be able to measure kernel duration,
calibration mode needs to disable asynchronous data transfers. Calibration thus
disables data transfer / computation overlapping, and should thus not be used
for eventual benchmarks. Note 2: history-based performance models get calibrated
only if a performance-model-based scheduler is chosen.

@node Task distribution vs Data transfer
@section Task distribution vs Data transfer

Distributing tasks to balance the load induces data transfer penalty. StarPU
thus needs to find a balance between both. The target function that the
@code{dmda} scheduler of StarPU
tries to minimize is @code{alpha * T_execution + beta * T_data_transfer}, where
@code{T_execution} is the estimated execution time of the codelet (usually
accurate), and @code{T_data_transfer} is the estimated data transfer time. The
latter is estimated based on bus calibration before execution start,
i.e. with an idle machine, thus without contention. You can force bus re-calibration by running
@code{starpu_calibrate_bus}. The beta parameter defaults to 1, but it can be
worth trying to tweak it by using @code{export STARPU_BETA=2} for instance,
since during real application execution, contention makes transfer times bigger.
This is of course imprecise, but in practice, a rough estimation already gives
the good results that a precise estimation would give.

@node Data prefetch
@section Data prefetch

The @code{heft}, @code{dmda} and @code{pheft} scheduling policies 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 needeed, 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 @code{starpu_data_prefetch_on_node} function
the handle and the desired target memory node.

@node Power-based scheduling
@section Power-based scheduling

If the application can provide some power performance model (through
the @code{power_model} field of the codelet structure), StarPU will
take it into account when distributing tasks. The target function that
the @code{dmda} scheduler minimizes becomes @code{alpha * T_execution +
beta * T_data_transfer + gamma * Consumption} , where @code{Consumption}
is the estimated task consumption in Joules. To tune this parameter, use
@code{export STARPU_GAMMA=3000} for instance, to express that each Joule
(i.e kW during 1000us) is worth 3000us execution time penalty. Setting
@code{alpha} and @code{beta} to zero permits to only take into account power consumption.

This is however not sufficient to correctly optimize power: the scheduler would
simply tend to run all computations on the most energy-conservative processing
unit. To account for the consumption of the whole machine (including idle
processing units), the idle power of the machine should be given by setting
@code{export STARPU_IDLE_POWER=200} for 200W, for instance. This value can often
be obtained from the machine power supplier.

The power actually consumed by the total execution can be displayed by setting
@code{export STARPU_PROFILING=1 STARPU_WORKER_STATS=1} .

@node Profiling
@section Profiling

A quick view of how many tasks each worker has executed can be obtained by setting 
@code{export STARPU_WORKER_STATS=1} This is a convenient way to check that
execution did happen on accelerators without penalizing performance with
the profiling overhead.

A quick view of how much data transfers have been issued can be obtained by setting 
@code{export STARPU_BUS_STATS=1} .

More detailed profiling information can be enabled by using @code{export STARPU_PROFILING=1} or by
calling @code{starpu_profiling_status_set} from the source code.
Statistics on the execution can then be obtained by using @code{export
STARPU_BUS_STATS=1} and @code{export STARPU_WORKER_STATS=1} .
 More details on performance feedback are provided by the next chapter.

@node CUDA-specific optimizations
@section CUDA-specific optimizations

Due to CUDA limitations, StarPU will have a hard time overlapping its own
communications and the codelet computations if the application does not use a
dedicated CUDA stream for its computations. StarPU provides one by the use of
@code{starpu_cuda_get_local_stream()} which should be used by all CUDA codelet
operations. For instance:

@example
func <<<grid,block,0,starpu_cuda_get_local_stream()>>> (foo, bar);
cudaStreamSynchronize(starpu_cuda_get_local_stream());
@end example

StarPU already does appropriate calls for the CUBLAS library.

Unfortunately, some CUDA libraries do not have stream variants of
kernels. That will lower the potential for overlapping.

@c ---------------------------------------------------------------------
@c Performance feedback
@c ---------------------------------------------------------------------

@node Performance feedback
@chapter Performance feedback

@menu
* On-line::       On-line performance feedback
* Off-line::      Off-line performance feedback
* Codelet performance::      Performance of codelets
@end menu

@node On-line
@section On-line performance feedback

@menu
* Enabling monitoring::     Enabling on-line performance monitoring
* Task feedback::           Per-task feedback
* Codelet feedback::        Per-codelet feedback
* Worker feedback::         Per-worker feedback
* Bus feedback::            Bus-related feedback
* StarPU-Top::              StarPU-Top interface
@end menu

@node Enabling monitoring
@subsection Enabling on-line performance monitoring

In order to enable online performance monitoring, the application can call
@code{starpu_profiling_status_set(STARPU_PROFILING_ENABLE)}. It is possible to
detect whether monitoring is already enabled or not by calling
@code{starpu_profiling_status_get()}. Enabling monitoring also reinitialize all
previously collected feedback. The @code{STARPU_PROFILING} environment variable
can also be set to 1 to achieve the same effect.

Likewise, performance monitoring is stopped by calling
@code{starpu_profiling_status_set(STARPU_PROFILING_DISABLE)}. Note that this
does not reset the performance counters so that the application may consult
them later on.

More details about the performance monitoring API are available in section
@ref{Profiling API}.

@node Task feedback
@subsection Per-task feedback

If profiling is enabled, a pointer to a @code{starpu_task_profiling_info}
structure is put in the @code{.profiling_info} field of the @code{starpu_task}
structure when a task terminates.
This structure is automatically destroyed when the task structure is destroyed,
either automatically or by calling @code{starpu_task_destroy}.

The @code{starpu_task_profiling_info} structure indicates the date when the
task was submitted (@code{submit_time}), started (@code{start_time}), and
terminated (@code{end_time}), relative to the initialization of
StarPU with @code{starpu_init}. It also specifies the identifier of the worker
that has executed the task (@code{workerid}).
These date are stored as @code{timespec} structures which the user may convert
into micro-seconds using the @code{starpu_timing_timespec_to_us} helper
function.

It it worth noting that the application may directly access this structure from
the callback executed at the end of the task. The @code{starpu_task} structure
associated to the callback currently being executed is indeed accessible with
the @code{starpu_get_current_task()} function.

@node Codelet feedback
@subsection Per-codelet feedback

The @code{per_worker_stats} field of the @code{starpu_codelet_t} structure is
an array of counters. The i-th entry of the array is incremented every time a
task implementing the codelet is executed on the i-th worker.
This array is not reinitialized when profiling is enabled or disabled.

@node Worker feedback
@subsection Per-worker feedback

The second argument returned by the @code{starpu_worker_get_profiling_info}
function is a @code{starpu_worker_profiling_info} structure that gives
statistics about the specified worker. This structure specifies when StarPU
started collecting profiling information for that worker (@code{start_time}),
the duration of the profiling measurement interval (@code{total_time}), the
time spent executing kernels (@code{executing_time}), the time spent sleeping
because there is no task to execute at all (@code{sleeping_time}), and the
number of tasks that were executed while profiling was enabled.
These values give an estimation of the proportion of time spent do real work,
and the time spent either sleeping because there are not enough executable
tasks or simply wasted in pure StarPU overhead. 

Calling @code{starpu_worker_get_profiling_info} resets the profiling
information associated to a worker.

When an FxT trace is generated (see @ref{Generating traces}), it is also
possible to use the @code{starpu_top} script (described in @ref{starpu-top}) to
generate a graphic showing the evolution of these values during the time, for
the different workers.

@node Bus feedback
@subsection Bus-related feedback 

TODO

@c how to enable/disable performance monitoring

@c what kind of information do we get ?

@node StarPU-Top
@subsection StarPU-Top interface

StarPU-Top is an interface which remotely displays the on-line state of a StarPU
application and permits the user to change parameters on the fly.

Variables to be monitored can be registered by calling the
@code{starputop_add_data_boolean}, @code{starputop_add_data_integer},
@code{starputop_add_data_float} functions, e.g.:

@example
starputop_data *data = starputop_add_data_integer("mynum", 0, 100, 1);
@end example

The application should then call @code{starputop_init_and_wait} to give its name
and wait for StarPU-Top to get a start request from the user. The name is used
by StarPU-Top to quickly reload a previously-saved layout of parameter display.

@example
starputop_init_and_wait("the application");
@end example

The new values can then be provided thanks to
@code{starputop_update_data_boolean}, @code{starputop_update_data_integer},
@code{starputop_update_data_float}, e.g.:

@example
starputop_update_data_integer(data, mynum);
@end example

Updateable parameters can be registered thanks to @code{starputop_register_parameter_boolean}, @code{starputop_register_parameter_integer}, @code{starputop_register_parameter_float}, e.g.:

@example
float apha;
starputop_register_parameter_float("alpha", &alpha, 0, 10, modif_hook);
@end example

@code{modif_hook} is a function which will be called when the parameter is being modified, it can for instance print the new value:

@example
void modif_hook(struct starputop_param_t *d) @{
    fprintf(stderr,"%s has been modified: %f\n", d->name, alpha);
@}
@end example

Task schedulers should notify StarPU-Top when it has decided when a task will be
scheduled, so that it can show it in its Gantt chart, for instance:

@example
starputop_task_prevision(task, workerid, begin, end);
@end example

Starting StarPU-Top and the application can be done two ways:

@itemize
@item The application is started by hand on some machine (and thus already
waiting for the start event). In the Preference dialog of StarPU-Top, the SSH
checkbox should be unchecked, and the hostname and port (default is 2011) on
which the application is already running should be specified. Clicking on the
connection button will thus connect to the already-running application.
@item StarPU-Top is started first, and clicking on the connection button will
start the application itself (possibly on a remote machine). The SSH checkbox
should be checked, and a command line provided, e.g.:

@example
ssh myserver STARPU_SCHED=heft ./application
@end example

If port 2011 of the remote machine can not be accessed directly, an ssh port bridge should be added:

@example
ssh -L 2011:localhost:2011 myserver STARPU_SCHED=heft ./application
@end example

and "localhost" should be used as IP Address to connect to.
@end itemize

@node Off-line
@section Off-line performance feedback

@menu
* Generating traces::       Generating traces with FxT
* Gantt diagram::           Creating a Gantt Diagram
* DAG::                     Creating a DAG with graphviz
* starpu-top::              Monitoring activity
@end menu

@node Generating traces
@subsection Generating traces with FxT

StarPU can use the FxT library (see
@indicateurl{https://savannah.nongnu.org/projects/fkt/}) to generate traces
with a limited runtime overhead.

You can either get a tarball:
@example
% wget http://download.savannah.gnu.org/releases/fkt/fxt-0.2.2.tar.gz
@end example

or use the FxT library from CVS (autotools are required):
@example
% cvs -d :pserver:anonymous@@cvs.sv.gnu.org:/sources/fkt co FxT
% ./bootstrap
@end example

Compiling and installing the FxT library in the @code{$FXTDIR} path is
done following the standard procedure:
@example
% ./configure --prefix=$FXTDIR
% make
% make install
@end example

In order to have StarPU to generate traces, StarPU should be configured with
the @code{--with-fxt} option:
@example
$ ./configure --with-fxt=$FXTDIR
@end example

Or you can simply point the @code{PKG_CONFIG_PATH} to
@code{$FXTDIR/lib/pkgconfig} and pass @code{--with-fxt} to @code{./configure}

When FxT is enabled, a trace is generated when StarPU is terminated by calling
@code{starpu_shutdown()}). The trace is a binary file whose name has the form
@code{prof_file_XXX_YYY} where @code{XXX} is the user name, and
@code{YYY} is the pid of the process that used StarPU. This file is saved in the
@code{/tmp/} directory by default, or by the directory specified by
the @code{STARPU_FXT_PREFIX} environment variable.

@node Gantt diagram
@subsection Creating a Gantt Diagram

When the FxT trace file @code{filename} has been generated, it is possible to
generate a trace in the Paje format by calling:
@example
% starpu_fxt_tool -i filename
@end example

Or alternatively, setting the @code{STARPU_GENERATE_TRACE} environment variable
to 1 before application execution will make StarPU do it automatically at
application shutdown.

This will create a @code{paje.trace} file in the current directory that can be
inspected with the ViTE trace visualizing open-source tool. More information
about ViTE is available at @indicateurl{http://vite.gforge.inria.fr/}. It is
possible to open the @code{paje.trace} file with ViTE by using the following
command:
@example
% vite paje.trace
@end example

@node DAG
@subsection Creating a DAG with graphviz

When the FxT trace file @code{filename} has been generated, it is possible to
generate a task graph in the DOT format by calling:
@example
$ starpu_fxt_tool -i filename
@end example

This will create a @code{dag.dot} file in the current directory. This file is a
task graph described using the DOT language. It is possible to get a
graphical output of the graph by using the graphviz library:
@example
$ dot -Tpdf dag.dot -o output.pdf
@end example

@node starpu-top
@subsection Monitoring activity

When the FxT trace file @code{filename} has been generated, it is possible to
generate a activity trace by calling:
@example
$ starpu_fxt_tool -i filename
@end example

This will create an @code{activity.data} file in the current
directory. A profile of the application showing the activity of StarPU
during the execution of the program can be generated:
@example
$ starpu_top.sh activity.data
@end example

This will create a file named @code{activity.eps} in the current directory.
This picture is composed of two parts.
The first part shows the activity of the different workers. The green sections
indicate which proportion of the time was spent executed kernels on the
processing unit. The red sections indicate the proportion of time spent in
StartPU: an important overhead may indicate that the granularity may be too
low, and that bigger tasks may be appropriate to use the processing unit more
efficiently. The black sections indicate that the processing unit was blocked
because there was no task to process: this may indicate a lack of parallelism
which may be alleviated by creating more tasks when it is possible.

The second part of the @code{activity.eps} picture is a graph showing the
evolution of the number of tasks available in the system during the execution.
Ready tasks are shown in black, and tasks that are submitted but not
schedulable yet are shown in grey.

@node Codelet performance
@section Performance of codelets

The performance model of codelets can be examined by using the
@code{starpu_perfmodel_display} tool:

@example
$ starpu_perfmodel_display -l
file: <malloc_pinned.hannibal>
file: <starpu_slu_lu_model_21.hannibal>
file: <starpu_slu_lu_model_11.hannibal>
file: <starpu_slu_lu_model_22.hannibal>
file: <starpu_slu_lu_model_12.hannibal>
@end example

Here, the codelets of the lu example are available. We can examine the
performance of the 22 kernel:

@example
$ starpu_perfmodel_display -s starpu_slu_lu_model_22
performance model for cpu
# hash		size		mean		dev		n
57618ab0	19660800       	2.851069e+05   	1.829369e+04   	109
performance model for cuda_0
# hash		size		mean		dev		n
57618ab0	19660800       	1.164144e+04   	1.556094e+01   	315
performance model for cuda_1
# hash		size		mean		dev		n
57618ab0	19660800       	1.164271e+04   	1.330628e+01   	360
performance model for cuda_2
# hash		size		mean		dev		n
57618ab0	19660800       	1.166730e+04   	3.390395e+02   	456
@end example

We can see that for the given size, over a sample of a few hundreds of
execution, the GPUs are about 20 times faster than the CPUs (numbers are in
us). The standard deviation is extremely low for the GPUs, and less than 10% for
CPUs.

@c ---------------------------------------------------------------------
@c MPI support
@c ---------------------------------------------------------------------

@node StarPU MPI support
@chapter StarPU MPI support

The integration of MPI transfers within task parallelism is done in a
very natural way by the means of asynchronous interactions between the
application and StarPU.  This is implemented in a separate libstarpumpi library
which basically provides "StarPU" equivalents of @code{MPI_*} functions, where
@code{void *} buffers are replaced with @code{starpu_data_handle}s, and all
GPU-RAM-NIC transfers are handled efficiently by StarPU-MPI.

@menu
* The API::                     
* Simple Example::              
* MPI Insert Task Utility::         
* MPI Collective Operations::         
@end menu

@node The API
@section The API

@subsection Initialisation

@deftypefun int starpu_mpi_initialize (void)
Initializes the starpumpi library. This must be called between calling
@code{starpu_init} and other @code{starpu_mpi} functions. This
function does not call @code{MPI_Init}, it should be called beforehand.
@end deftypefun

@deftypefun int starpu_mpi_initialize_extended (int *@var{rank}, int *@var{world_size})
Initializes the starpumpi library. This must be called between calling
@code{starpu_init} and other @code{starpu_mpi} functions.
This function calls @code{MPI_Init}, and therefore should be prefered
to the previous one for MPI implementations which are not thread-safe.
Returns the current MPI node rank and world size.
@end deftypefun

@deftypefun int starpu_mpi_shutdown (void)
Cleans the starpumpi library. This must be called between calling
@code{starpu_mpi} functions and @code{starpu_shutdown}.
@code{MPI_Finalize} will be called if StarPU-MPI has been initialized
by calling @code{starpu_mpi_initialize_extended}.
@end deftypefun

@subsection Communication

@deftypefun int starpu_mpi_send (starpu_data_handle @var{data_handle}, int @var{dest}, int @var{mpi_tag}, MPI_Comm @var{comm})
@end deftypefun

@deftypefun int starpu_mpi_recv (starpu_data_handle @var{data_handle}, int @var{source}, int @var{mpi_tag}, MPI_Comm @var{comm}, MPI_Status *@var{status})
@end deftypefun

@deftypefun int starpu_mpi_isend (starpu_data_handle @var{data_handle}, starpu_mpi_req *@var{req}, int @var{dest}, int @var{mpi_tag}, MPI_Comm @var{comm})

@end deftypefun

@deftypefun int starpu_mpi_irecv (starpu_data_handle @var{data_handle}, starpu_mpi_req *@var{req}, int @var{source}, int @var{mpi_tag}, MPI_Comm @var{comm})
@end deftypefun

@deftypefun int starpu_mpi_isend_detached (starpu_data_handle @var{data_handle}, int @var{dest}, int @var{mpi_tag}, MPI_Comm @var{comm}, void (*@var{callback})(void *), void *@var{arg})
@end deftypefun

@deftypefun int starpu_mpi_irecv_detached (starpu_data_handle @var{data_handle}, int @var{source}, int @var{mpi_tag}, MPI_Comm @var{comm}, void (*@var{callback})(void *), void *@var{arg})
@end deftypefun

@deftypefun int starpu_mpi_wait (starpu_mpi_req *@var{req}, MPI_Status *@var{status})
@end deftypefun

@deftypefun int starpu_mpi_test (starpu_mpi_req *@var{req}, int *@var{flag}, MPI_Status *@var{status})
@end deftypefun

@deftypefun int starpu_mpi_barrier (MPI_Comm @var{comm})
@end deftypefun

@deftypefun int starpu_mpi_isend_detached_unlock_tag (starpu_data_handle @var{data_handle}, int @var{dest}, int @var{mpi_tag}, MPI_Comm @var{comm}, starpu_tag_t @var{tag})
When the transfer is completed, the tag is unlocked
@end deftypefun

@deftypefun int starpu_mpi_irecv_detached_unlock_tag (starpu_data_handle @var{data_handle}, int @var{source}, int @var{mpi_tag}, MPI_Comm @var{comm}, starpu_tag_t @var{tag})
@end deftypefun

@deftypefun int starpu_mpi_isend_array_detached_unlock_tag (unsigned @var{array_size}, starpu_data_handle *@var{data_handle}, int *@var{dest}, int *@var{mpi_tag}, MPI_Comm *@var{comm}, starpu_tag_t @var{tag})
Asynchronously send an array of buffers, and unlocks the tag once all
of them are transmitted.
@end deftypefun

@deftypefun int starpu_mpi_irecv_array_detached_unlock_tag (unsigned @var{array_size}, starpu_data_handle *@var{data_handle}, int *@var{source}, int *@var{mpi_tag}, MPI_Comm *@var{comm}, starpu_tag_t @var{tag})
@end deftypefun

@page
@node Simple Example
@section Simple Example

@cartouche
@smallexample
void increment_token(void)
@{
    struct starpu_task *task = starpu_task_create();

    task->cl = &increment_cl;
    task->buffers[0].handle = token_handle;
    task->buffers[0].mode = STARPU_RW;

    starpu_task_submit(task);
@}
@end smallexample
@end cartouche

@cartouche
@smallexample
int main(int argc, char **argv)
@{
    int rank, size;

    starpu_init(NULL);
    starpu_mpi_initialize_extended(&rank, &size);

    starpu_vector_data_register(&token_handle, 0, (uintptr_t)&token, 1, sizeof(unsigned));

    unsigned nloops = NITER;
    unsigned loop;

    unsigned last_loop = nloops - 1;
    unsigned last_rank = size - 1;
@end smallexample
@end cartouche

@cartouche
@smallexample
    for (loop = 0; loop < nloops; loop++) @{
        int tag = loop*size + rank;

        if (loop == 0 && rank == 0)
        @{
            token = 0;
            fprintf(stdout, "Start with token value %d\n", token);
        @}
        else
        @{
            starpu_mpi_irecv_detached(token_handle, (rank+size-1)%size, tag,
                    MPI_COMM_WORLD, NULL, NULL);
        @}

        increment_token();

        if (loop == last_loop && rank == last_rank)
        @{
            starpu_data_acquire(token_handle, STARPU_R);
            fprintf(stdout, "Finished : token value %d\n", token);
            starpu_data_release(token_handle);
        @}
        else
        @{
            starpu_mpi_isend_detached(token_handle, (rank+1)%size, tag+1,
                    MPI_COMM_WORLD, NULL, NULL);
        @}
    @}

    starpu_task_wait_for_all();
@end smallexample
@end cartouche

@cartouche
@smallexample
    starpu_mpi_shutdown();
    starpu_shutdown();

    if (rank == last_rank)
    @{
        fprintf(stderr, "[%d] token = %d == %d * %d ?\n", rank, token, nloops, size);
        STARPU_ASSERT(token == nloops*size);
    @}
@end smallexample
@end cartouche

@page
@node MPI Insert Task Utility
@section MPI Insert Task Utility

@deftypefun void starpu_mpi_insert_task (MPI_Comm @var{comm}, starpu_codelet *@var{cl}, ...)
Create and submit a task corresponding to @var{cl} with the following
arguments.  The argument list must be zero-terminated.

The arguments following the codelets are the same types as for the
function @code{starpu_insert_task} defined in @ref{Insert Task
Utility}. The extra argument @code{STARPU_EXECUTE_ON_NODE} followed by an
integer allows to specify the node to execute the codelet. It is also
possible to specify that the node owning a specific data will execute
the codelet, by using @code{STARPU_EXECUTE_ON_DATA} followed by a data
handle.

The algorithm is as follows:
@enumerate
@item Find out whether we are to execute the codelet because we own the
data to be written to. If different tasks own data to be written to,
the argument @code{STARPU_EXECUTE_ON_NODE} or
@code{STARPU_EXECUTE_ON_DATA} should be used to specify the executing
task @code{ET}.
@item Send and receive data as requested. Tasks owning data which need
to be read by the executing task @code{ET} are sending them to @code{ET}.
@item Execute the codelet. This is done by the task selected in the
1st step of the algorithm.
@item In the case when different tasks own data to be written to, send
W data back to their owners.
@end enumerate

The algorithm also includes a cache mechanism that allows not to send
data twice to the same task, unless the data has been modified.

@end deftypefun

@deftypefun void starpu_mpi_get_data_on_node (MPI_Comm @var{comm}, starpu_data_handle @var{data_handle}, int @var{node})
@end deftypefun

@page

Here an example showing how to use @code{starpu_mpi_insert_task}. One
first needs to define a distribution function which specifies the
locality of the data. Note that that distribution information needs to
be given to StarPU by calling @code{starpu_data_set_rank}.

@cartouche
@smallexample
/* Returns the MPI node number where data is */
int my_distrib(int x, int y, int nb_nodes) @{
	/* Cyclic distrib */
	return ((int)(x / sqrt(nb_nodes) + (y / sqrt(nb_nodes)) * sqrt(nb_nodes))) % nb_nodes;
        //	/* Linear distrib */
        //	return x / sqrt(nb_nodes) + (y / sqrt(nb_nodes)) * X;
@}
@end smallexample
@end cartouche

Now the data can be registered within StarPU. Data which are not
owned but will be needed for computations can be registered through
the lazy allocation mechanism, i.e. with a @code{home_node} set to -1.
StarPU will automatically allocate the memory when it is used for the
first time.

@cartouche
@smallexample
    unsigned matrix[X][Y];
    starpu_data_handle data_handles[X][Y];

    for(x = 0; x < X; x++) @{
        for (y = 0; y < Y; y++) @{
            int mpi_rank = my_distrib(x, y, size);
             if (mpi_rank == rank)
                /* Owning data */
                starpu_variable_data_register(&data_handles[x][y], 0,
                                              (uintptr_t)&(matrix[x][y]), sizeof(unsigned));
            else if (rank == mpi_rank+1 || rank == mpi_rank-1)
                /* I don't own that index, but will need it for my computations */
                starpu_variable_data_register(&data_handles[x][y], -1,
                                              (uintptr_t)NULL, sizeof(unsigned));
            else
                /* I know it's useless to allocate anything for this */
                data_handles[x][y] = NULL;
            if (data_handles[x][y])
                starpu_data_set_rank(data_handles[x][y], mpi_rank);
        @}
    @}
@end smallexample
@end cartouche

Now @code{starpu_mpi_insert_task()} can be called for the different
steps of the application.

@cartouche
@smallexample
    for(loop=0 ; loop<niter; loop++)
        for (x = 1; x < X-1; x++)
            for (y = 1; y < Y-1; y++)
                starpu_mpi_insert_task(MPI_COMM_WORLD, &stencil5_cl,
                                       STARPU_RW, data_handles[x][y],
                                       STARPU_R, data_handles[x-1][y],
                                       STARPU_R, data_handles[x+1][y],
                                       STARPU_R, data_handles[x][y-1],
                                       STARPU_R, data_handles[x][y+1],
                                       0);
    starpu_task_wait_for_all();
@end smallexample
@end cartouche

@node MPI Collective Operations
@section MPI Collective Operations

@deftypefun int starpu_mpi_scatter_detached (starpu_data_handle *@var{data_handles}, int @var{count}, int @var{root}, MPI_Comm @var{comm})
Scatter data among processes of the communicator based on the ownership of
the data. For each data of the array @var{data_handles}, the
process @var{root} sends the data to the process owning this data.
Processes receiving data must have valid data handles to receive them.
@end deftypefun

@deftypefun int starpu_mpi_gather_detached (starpu_data_handle *@var{data_handles}, int @var{count}, int @var{root}, MPI_Comm @var{comm})
Gather data from the different processes of the communicator onto the
process @var{root}. Each process owning data handle in the array
@var{data_handles} will send them to the process @var{root}. The
process @var{root} must have valid data handles to receive the data.
@end deftypefun

@page
@cartouche
@smallexample
if (rank == root)
@{
    /* Allocate the vector */
    vector = malloc(nblocks * sizeof(float *));
    for(x=0 ; x<nblocks ; x++)
    @{
        starpu_malloc((void **)&vector[x], block_size*sizeof(float));
    @}
@}

/* Allocate data handles and register data to StarPU */
data_handles = malloc(nblocks*sizeof(starpu_data_handle *));
for(x = 0; x < nblocks ;  x++)
@{
    int mpi_rank = my_distrib(x, nodes);
    if (rank == root) @{
        starpu_vector_data_register(&data_handles[x], 0, (uintptr_t)vector[x],
                                    blocks_size, sizeof(float));
    @}
    else if ((mpi_rank == rank) || ((rank == mpi_rank+1 || rank == mpi_rank-1))) @{
        /* I own that index, or i will need it for my computations */
        starpu_vector_data_register(&data_handles[x], -1, (uintptr_t)NULL,
                                   block_size, sizeof(float));
    @}
    else @{
        /* I know it's useless to allocate anything for this */
        data_handles[x] = NULL;
    @}
    if (data_handles[x]) @{
        starpu_data_set_rank(data_handles[x], mpi_rank);
    @}
@}

/* Scatter the matrix among the nodes */
starpu_mpi_scatter_detached(data_handles, nblocks, root, MPI_COMM_WORLD);

/* Calculation */
for(x = 0; x < nblocks ;  x++) @{
    if (data_handles[x]) @{
        int owner = starpu_data_get_rank(data_handles[x]);
        if (owner == rank) @{
            starpu_insert_task(&cl, STARPU_RW, data_handles[x], 0);
        @}
    @}
@}

/* Gather the matrix on main node */
starpu_mpi_gather_detached(data_handles, nblocks, 0, MPI_COMM_WORLD);
@end smallexample
@end cartouche


@c ---------------------------------------------------------------------
@c Configuration options
@c ---------------------------------------------------------------------

@node Configuring StarPU
@chapter Configuring StarPU


@menu
* Compilation configuration::   
* Execution configuration through environment variables::  
@end menu

@node Compilation configuration
@section Compilation configuration

The following arguments can be given to the @code{configure} script.

@menu
* Common configuration::        
* Configuring workers::         
* Advanced configuration::      
@end menu

@node Common configuration
@subsection Common configuration


@menu
* --enable-debug::              
* --enable-fast::               
* --enable-verbose::            
* --enable-coverage::           
@end menu

@node --enable-debug
@subsubsection @code{--enable-debug}
@table @asis
@item @emph{Description}:
Enable debugging messages.
@end table

@node --enable-fast
@subsubsection @code{--enable-fast}
@table @asis
@item @emph{Description}:
Do not enforce assertions, saves a lot of time spent to compute them otherwise.
@end table

@node --enable-verbose
@subsubsection @code{--enable-verbose}
@table @asis
@item @emph{Description}:
Augment the verbosity of the debugging messages. This can be disabled
at runtime by setting the environment variable @code{STARPU_SILENT} to
any value.

@smallexample
% STARPU_SILENT=1 ./vector_scal
@end smallexample
@end table

@node --enable-coverage
@subsubsection @code{--enable-coverage}
@table @asis
@item @emph{Description}:
Enable flags for the @code{gcov} coverage tool.
@end table

@node Configuring workers
@subsection Configuring workers

@menu
* --enable-nmaxcpus::         
* --disable-cpu::               
* --enable-maxcudadev::         
* --disable-cuda::              
* --with-cuda-dir::             
* --with-cuda-include-dir::             
* --with-cuda-lib-dir::             
* --enable-maxopencldev::       
* --disable-opencl::            
* --with-opencl-dir::           
* --with-opencl-include-dir::           
* --with-opencl-lib-dir::           
* --enable-gordon::             
* --with-gordon-dir::           
@end menu

@node --enable-nmaxcpus
@subsubsection @code{--enable-nmaxcpus=<number>}
@table @asis
@item @emph{Description}:
Defines the maximum number of CPU cores that StarPU will support, then
available as the @code{STARPU_NMAXCPUS} macro.
@end table

@node --disable-cpu
@subsubsection @code{--disable-cpu}
@table @asis
@item @emph{Description}:
Disable the use of CPUs of the machine. Only GPUs etc. will be used.
@end table

@node --enable-maxcudadev
@subsubsection @code{--enable-maxcudadev=<number>}
@table @asis
@item @emph{Description}:
Defines the maximum number of CUDA devices that StarPU will support, then
available as the @code{STARPU_MAXCUDADEVS} macro.
@end table

@node --disable-cuda
@subsubsection @code{--disable-cuda}
@table @asis
@item @emph{Description}:
Disable the use of CUDA, even if a valid CUDA installation was detected.
@end table

@node --with-cuda-dir
@subsubsection @code{--with-cuda-dir=<path>}
@table @asis
@item @emph{Description}:
Specify the directory where CUDA is installed. This directory should notably contain
@code{include/cuda.h}.
@end table

@node --with-cuda-include-dir
@subsubsection @code{--with-cuda-include-dir=<path>}
@table @asis
@item @emph{Description}:
Specify the directory where CUDA headers are installed. This directory should
notably contain @code{cuda.h}. This defaults to @code{/include} appended to the
value given to @code{--with-cuda-dir}.
@end table

@node --with-cuda-lib-dir
@subsubsection @code{--with-cuda-lib-dir=<path>}
@table @asis
@item @emph{Description}:
Specify the directory where the CUDA library is installed. This directory should
notably contain the CUDA shared libraries (e.g. libcuda.so). This defaults to
@code{/lib} appended to the value given to @code{--with-cuda-dir}.

@end table

@node --enable-maxopencldev
@subsubsection @code{--enable-maxopencldev=<number>}
@table @asis
@item @emph{Description}:
Defines the maximum number of OpenCL devices that StarPU will support, then
available as the @code{STARPU_MAXOPENCLDEVS} macro.
@end table

@node --disable-opencl
@subsubsection @code{--disable-opencl}
@table @asis
@item @emph{Description}:
Disable the use of OpenCL, even if the SDK is detected.
@end table

@node --with-opencl-dir
@subsubsection @code{--with-opencl-dir=<path>}
@table @asis
@item @emph{Description}:
Specify the location of the OpenCL SDK. This directory should notably contain
@code{include/CL/cl.h} (or @code{include/OpenCL/cl.h} on Mac OS).
@end table

@node --with-opencl-include-dir
@subsubsection @code{--with-opencl-include-dir=<path>}
@table @asis
@item @emph{Description}:
Specify the location of OpenCL headers. This directory should notably contain
@code{CL/cl.h} (or @code{OpenCL/cl.h} on Mac OS). This defaults to
@code{/include} appended to the value given to @code{--with-opencl-dir}.

@end table

@node --with-opencl-lib-dir
@subsubsection @code{--with-opencl-lib-dir=<path>}
@table @asis
@item @emph{Description}:
Specify the location of the OpenCL library. This directory should notably
contain the OpenCL shared libraries (e.g. libOpenCL.so). This defaults to
@code{/lib} appended to the value given to @code{--with-opencl-dir}.
@end table

@node --enable-gordon
@subsubsection @code{--enable-gordon}
@table @asis
@item @emph{Description}:
Enable the use of the Gordon runtime for Cell SPUs.
@c TODO: rather default to enabled when detected
@end table

@node --with-gordon-dir
@subsubsection @code{--with-gordon-dir=<path>}
@table @asis
@item @emph{Description}:
Specify the location of the Gordon SDK.
@end table

@node Advanced configuration
@subsection Advanced configuration

@menu
* --enable-perf-debug::         
* --enable-model-debug::        
* --enable-stats::              
* --enable-maxbuffers::         
* --enable-allocation-cache::   
* --enable-opengl-render::      
* --enable-blas-lib::           
* --with-magma::                
* --with-fxt::                  
* --with-perf-model-dir::       
* --with-mpicc::                
* --with-goto-dir::             
* --with-atlas-dir::            
* --with-mkl-cflags::
* --with-mkl-ldflags::
@end menu

@node --enable-perf-debug
@subsubsection @code{--enable-perf-debug}
@table @asis
@item @emph{Description}:
Enable performance debugging through gprof.
@end table

@node --enable-model-debug
@subsubsection @code{--enable-model-debug}
@table @asis
@item @emph{Description}:
Enable performance model debugging.
@end table

@node --enable-stats
@subsubsection @code{--enable-stats}
@table @asis
@item @emph{Description}:
Enable statistics.
@end table

@node --enable-maxbuffers
@subsubsection @code{--enable-maxbuffers=<nbuffers>}
@table @asis
@item @emph{Description}:
Define the maximum number of buffers that tasks will be able to take
as parameters, then available as the @code{STARPU_NMAXBUFS} macro.
@end table

@node --enable-allocation-cache
@subsubsection @code{--enable-allocation-cache}
@table @asis
@item @emph{Description}:
Enable the use of a data allocation cache to avoid the cost of it with
CUDA. Still experimental.
@end table

@node --enable-opengl-render
@subsubsection @code{--enable-opengl-render}
@table @asis
@item @emph{Description}:
Enable the use of OpenGL for the rendering of some examples.
@c TODO: rather default to enabled when detected
@end table

@node --enable-blas-lib
@subsubsection @code{--enable-blas-lib=<name>}
@table @asis
@item @emph{Description}:
Specify the blas library to be used by some of the examples. The
library has to be 'atlas' or 'goto'.
@end table

@node --with-magma
@subsubsection @code{--with-magma=<path>}
@table @asis
@item @emph{Description}:
Specify where magma is installed. This directory should notably contain
@code{include/magmablas.h}.
@end table

@node --with-fxt
@subsubsection @code{--with-fxt=<path>}
@table @asis
@item @emph{Description}:
Specify the location of FxT (for generating traces and rendering them
using ViTE). This directory should notably contain
@code{include/fxt/fxt.h}.
@c TODO add ref to other section
@end table

@node --with-perf-model-dir
@subsubsection @code{--with-perf-model-dir=<dir>}
@table @asis
@item @emph{Description}:
Specify where performance models should be stored (instead of defaulting to the
current user's home).
@end table

@node --with-mpicc
@subsubsection @code{--with-mpicc=<path to mpicc>}
@table @asis
@item @emph{Description}:
Specify the location of the @code{mpicc} compiler to be used for starpumpi.
@end table

@node --with-goto-dir
@subsubsection @code{--with-goto-dir=<dir>}
@table @asis
@item @emph{Description}:
Specify the location of GotoBLAS.
@end table

@node --with-atlas-dir
@subsubsection @code{--with-atlas-dir=<dir>}
@table @asis
@item @emph{Description}:
Specify the location of ATLAS. This directory should notably contain
@code{include/cblas.h}.
@end table

@node --with-mkl-cflags
@subsubsection @code{--with-mkl-cflags=<cflags>}
@table @asis
@item @emph{Description}:
Specify the compilation flags for the MKL Library.
@end table

@node --with-mkl-ldflags
@subsubsection @code{--with-mkl-ldflags=<ldflags>}
@table @asis
@item @emph{Description}:
Specify the linking flags for the MKL Library. Note that the
@url{http://software.intel.com/en-us/articles/intel-mkl-link-line-advisor/}
website provides a script to determine the linking flags.
@end table


@c ---------------------------------------------------------------------
@c Environment variables
@c ---------------------------------------------------------------------

@node Execution configuration through environment variables
@section Execution configuration through environment variables

@menu
* Workers::                     Configuring workers
* Scheduling::                  Configuring the Scheduling engine
* Misc::                        Miscellaneous and debug
@end menu

Note: the values given in @code{starpu_conf} structure passed when
calling @code{starpu_init} will override the values of the environment
variables.

@node Workers
@subsection Configuring workers

@menu
* STARPU_NCPUS::                Number of CPU workers
* STARPU_NCUDA::                Number of CUDA workers
* STARPU_NOPENCL::              Number of OpenCL workers
* STARPU_NGORDON::              Number of SPU workers (Cell)
* STARPU_WORKERS_CPUID::        Bind workers to specific CPUs
* STARPU_WORKERS_CUDAID::       Select specific CUDA devices
* STARPU_WORKERS_OPENCLID::     Select specific OpenCL devices
@end menu

@node STARPU_NCPUS
@subsubsection @code{STARPU_NCPUS} -- Number of CPU workers
@table @asis

@item @emph{Description}:
Specify the number of CPU workers (thus not including workers dedicated to control acceleratores). Note that by default, StarPU will not allocate
more CPU workers than there are physical CPUs, and that some CPUs are used to control
the accelerators.

@end table

@node STARPU_NCUDA
@subsubsection @code{STARPU_NCUDA} -- Number of CUDA workers
@table @asis

@item @emph{Description}:
Specify the number of CUDA devices that StarPU can use. If
@code{STARPU_NCUDA} is lower than the number of physical devices, it is
possible to select which CUDA devices should be used by the means of the
@code{STARPU_WORKERS_CUDAID} environment variable. By default, StarPU will
create as many CUDA workers as there are CUDA devices.

@end table

@node STARPU_NOPENCL
@subsubsection @code{STARPU_NOPENCL} -- Number of OpenCL workers
@table @asis

@item @emph{Description}:
OpenCL equivalent of the @code{STARPU_NCUDA} environment variable.
@end table

@node STARPU_NGORDON
@subsubsection @code{STARPU_NGORDON} -- Number of SPU workers (Cell)
@table @asis

@item @emph{Description}:
Specify the number of SPUs that StarPU can use.
@end table


@node STARPU_WORKERS_CPUID
@subsubsection @code{STARPU_WORKERS_CPUID} -- Bind workers to specific CPUs
@table @asis

@item @emph{Description}:
Passing an array of integers (starting from 0) in @code{STARPU_WORKERS_CPUID}
specifies on which logical CPU the different workers should be
bound. For instance, if @code{STARPU_WORKERS_CPUID = "0 1 4 5"}, the first
worker will be bound to logical CPU #0, the second CPU worker will be bound to
logical CPU #1 and so on.  Note that the logical ordering of the CPUs is either
determined by the OS, or provided by the @code{hwloc} library in case it is
available.

Note that the first workers correspond to the CUDA workers, then come the
OpenCL and the SPU, and finally the CPU workers. For example if
we have @code{STARPU_NCUDA=1}, @code{STARPU_NOPENCL=1}, @code{STARPU_NCPUS=2}
and @code{STARPU_WORKERS_CPUID = "0 2 1 3"}, the CUDA device will be controlled
by logical CPU #0, the OpenCL device will be controlled by logical CPU #2, and
the logical CPUs #1 and #3 will be used by the CPU workers.

If the number of workers is larger than the array given in
@code{STARPU_WORKERS_CPUID}, the workers are bound to the logical CPUs in a
round-robin fashion: if @code{STARPU_WORKERS_CPUID = "0 1"}, the first and the
third (resp. second and fourth) workers will be put on CPU #0 (resp. CPU #1).

This variable is ignored if the @code{use_explicit_workers_bindid} flag of the
@code{starpu_conf} structure passed to @code{starpu_init} is set.

@end table

@node STARPU_WORKERS_CUDAID
@subsubsection @code{STARPU_WORKERS_CUDAID} -- Select specific CUDA devices
@table @asis

@item @emph{Description}:
Similarly to the @code{STARPU_WORKERS_CPUID} environment variable, it is
possible to select which CUDA devices should be used by StarPU. On a machine
equipped with 4 GPUs, setting @code{STARPU_WORKERS_CUDAID = "1 3"} and
@code{STARPU_NCUDA=2} specifies that 2 CUDA workers should be created, and that
they should use CUDA devices #1 and #3 (the logical ordering of the devices is
the one reported by CUDA).

This variable is ignored if the @code{use_explicit_workers_cuda_gpuid} flag of
the @code{starpu_conf} structure passed to @code{starpu_init} is set.
@end table

@node STARPU_WORKERS_OPENCLID
@subsubsection @code{STARPU_WORKERS_OPENCLID} -- Select specific OpenCL devices
@table @asis

@item @emph{Description}:
OpenCL equivalent of the @code{STARPU_WORKERS_CUDAID} environment variable.

This variable is ignored if the @code{use_explicit_workers_opencl_gpuid} flag of
the @code{starpu_conf} structure passed to @code{starpu_init} is set.
@end table

@node Scheduling
@subsection Configuring the Scheduling engine

@menu
* STARPU_SCHED::                Scheduling policy
* STARPU_CALIBRATE::            Calibrate performance models
* STARPU_PREFETCH::             Use data prefetch
* STARPU_SCHED_ALPHA::          Computation factor
* STARPU_SCHED_BETA::           Communication factor
@end menu

@node STARPU_SCHED
@subsubsection @code{STARPU_SCHED} -- Scheduling policy
@table @asis

@item @emph{Description}:

This chooses between the different scheduling policies proposed by StarPU: work
random, stealing, greedy, with performance models, etc.

Use @code{STARPU_SCHED=help} to get the list of available schedulers.

@end table

@node STARPU_CALIBRATE
@subsubsection @code{STARPU_CALIBRATE} -- Calibrate performance models
@table @asis

@item @emph{Description}:
If this variable is set to 1, the performance models are calibrated during
the execution. If it is set to 2, the previous values are dropped to restart
calibration from scratch. Setting this variable to 0 disable calibration, this
is the default behaviour.

Note: this currently only applies to @code{dm}, @code{dmda} and @code{heft} scheduling policies.

@end table

@node STARPU_PREFETCH
@subsubsection @code{STARPU_PREFETCH} -- Use data prefetch
@table @asis

@item @emph{Description}:
This variable indicates whether data prefetching should be enabled (0 means
that it is disabled). If prefetching is enabled, when a task is scheduled to be
executed e.g. on a GPU, StarPU will request an asynchronous transfer in
advance, so that data is already present on the GPU when the task starts. As a
result, computation and data transfers are overlapped.
Note that prefetching is enabled by default in StarPU.

@end table

@node STARPU_SCHED_ALPHA
@subsubsection @code{STARPU_SCHED_ALPHA} -- Computation factor
@table @asis

@item @emph{Description}:
To estimate the cost of a task StarPU takes into account the estimated
computation time (obtained thanks to performance models). The alpha factor is
the coefficient to be applied to it before adding it to the communication part.

@end table

@node STARPU_SCHED_BETA
@subsubsection @code{STARPU_SCHED_BETA} -- Communication factor
@table @asis

@item @emph{Description}:
To estimate the cost of a task StarPU takes into account the estimated
data transfer time (obtained thanks to performance models). The beta factor is
the coefficient to be applied to it before adding it to the computation part.

@end table

@node Misc
@subsection Miscellaneous and debug

@menu
* STARPU_SILENT::               Disable verbose mode
* STARPU_LOGFILENAME::          Select debug file name
* STARPU_FXT_PREFIX::           FxT trace location
* STARPU_LIMIT_GPU_MEM::        Restrict memory size on the GPUs
* STARPU_GENERATE_TRACE::       Generate a Paje trace when StarPU is shut down
@end menu

@node STARPU_SILENT
@subsubsection @code{STARPU_SILENT} -- Disable verbose mode
@table @asis

@item @emph{Description}:
This variable allows to disable verbose mode at runtime when StarPU
has been configured with the option @code{--enable-verbose}.
@end table

@node STARPU_LOGFILENAME
@subsubsection @code{STARPU_LOGFILENAME} -- Select debug file name
@table @asis

@item @emph{Description}:
This variable specifies in which file the debugging output should be saved to.
@end table

@node STARPU_FXT_PREFIX
@subsubsection @code{STARPU_FXT_PREFIX} -- FxT trace location
@table @asis

@item @emph{Description}
This variable specifies in which directory to save the trace generated if FxT is enabled. It needs to have a trailing '/' character.
@end table

@node STARPU_LIMIT_GPU_MEM
@subsubsection @code{STARPU_LIMIT_GPU_MEM} -- Restrict memory size on the GPUs
@table @asis

@item @emph{Description}
This variable specifies the maximum number of megabytes that should be
available to the application on each GPUs. In case this value is smaller than
the size of the memory of a GPU, StarPU pre-allocates a buffer to waste memory
on the device. This variable is intended to be used for experimental purposes
as it emulates devices that have a limited amount of memory.
@end table

@node STARPU_GENERATE_TRACE
@subsubsection @code{STARPU_GENERATE_TRACE} -- Generate a Paje trace when StarPU is shut down
@table @asis

@item @emph{Description}
When set to 1, this variable indicates that StarPU should automatically
generate a Paje trace when starpu_shutdown is called.
@end table


@c ---------------------------------------------------------------------
@c StarPU API
@c ---------------------------------------------------------------------

@node StarPU API
@chapter StarPU API

@menu
* Initialization and Termination::  Initialization and Termination methods
* Workers' Properties::         Methods to enumerate workers' properties
* Data Library::                Methods to manipulate data
* Data Interfaces::             
* Data Partition::              
* Codelets and Tasks::          Methods to construct tasks
* Explicit Dependencies::       Explicit Dependencies
* Implicit Data Dependencies::  Implicit Data Dependencies
* Performance Model API::       
* Profiling API::               Profiling API
* CUDA extensions::             CUDA extensions
* OpenCL extensions::           OpenCL extensions
* Cell extensions::             Cell extensions
* Miscellaneous helpers::       
@end menu

@node Initialization and Termination
@section Initialization and Termination

@menu
* starpu_init::                 Initialize StarPU
* struct starpu_conf::          StarPU runtime configuration
* starpu_conf_init::     Initialize starpu_conf structure
* starpu_shutdown::             Terminate StarPU
@end menu

@node starpu_init
@subsection @code{starpu_init} -- Initialize StarPU
@table @asis

@item @emph{Description}:
This is StarPU initialization method, which must be called prior to any other
StarPU call.  It is possible to specify StarPU's configuration (e.g. scheduling
policy, number of cores, ...) by passing a non-null argument. Default
configuration is used if the passed argument is @code{NULL}.
@item @emph{Return value}:
Upon successful completion, this function returns 0. Otherwise, @code{-ENODEV}
indicates that no worker was available (so that StarPU was not initialized).

@item @emph{Prototype}:
@code{int starpu_init(struct starpu_conf *conf);}

@end table

@node struct starpu_conf
@subsection @code{struct starpu_conf} -- StarPU runtime configuration

@table @asis
@item @emph{Description}:
This structure is passed to the @code{starpu_init} function in order
to configure StarPU.
When the default value is used, StarPU automatically selects the number
of processing units and takes the default scheduling policy. This parameter
overwrites the equivalent environment variables.

@item @emph{Fields}:
@table @asis
@item @code{sched_policy_name} (default = NULL):
This is the name of the scheduling policy. This can also be specified with the
@code{STARPU_SCHED} environment variable.
@item @code{sched_policy} (default = NULL):
This is the definition of the scheduling policy. This field is ignored
if @code{sched_policy_name} is set.

@item @code{ncpus} (default = -1):
This is the number of CPU cores that StarPU can use. This can also be
specified with the @code{STARPU_NCPUS} environment variable.
@item @code{ncuda} (default = -1):
This is the number of CUDA devices that StarPU can use. This can also be
specified with the @code{STARPU_NCUDA} environment variable.
@item @code{nopencl} (default = -1):
This is the number of OpenCL devices that StarPU can use. This can also be
specified with the @code{STARPU_NOPENCL} environment variable.
@item @code{nspus} (default = -1):
This is the number of Cell SPUs that StarPU can use. This can also be
specified with the @code{STARPU_NGORDON} environment variable.

@item @code{use_explicit_workers_bindid} (default = 0)
If this flag is set, the @code{workers_bindid} array indicates where the
different workers are bound, otherwise StarPU automatically selects where to
bind the different workers unless the @code{STARPU_WORKERS_CPUID} environment
variable is set. The @code{STARPU_WORKERS_CPUID} environment variable is
ignored if the @code{use_explicit_workers_bindid} flag is set.
@item @code{workers_bindid[STARPU_NMAXWORKERS]}
If the @code{use_explicit_workers_bindid} flag is set, this array indicates
where to bind the different workers. The i-th entry of the
@code{workers_bindid} indicates the logical identifier of the processor which
should execute the i-th worker. Note that the logical ordering of the CPUs is
either determined by the OS, or provided by the @code{hwloc} library in case it
is available.
When this flag is set, the @ref{STARPU_WORKERS_CPUID} environment variable is
ignored.
 
@item @code{use_explicit_workers_cuda_gpuid} (default = 0)
If this flag is set, the CUDA workers will be attached to the CUDA devices
specified in the @code{workers_cuda_gpuid} array. Otherwise, StarPU affects the
CUDA devices in a round-robin fashion.
When this flag is set, the @ref{STARPU_WORKERS_CUDAID} environment variable is
ignored.
@item @code{workers_cuda_gpuid[STARPU_NMAXWORKERS]}
If the @code{use_explicit_workers_cuda_gpuid} flag is set, this array contains
the logical identifiers of the CUDA devices (as used by  @code{cudaGetDevice}).
@item @code{use_explicit_workers_opencl_gpuid} (default = 0)
If this flag is set, the OpenCL workers will be attached to the OpenCL devices
specified in the @code{workers_opencl_gpuid} array. Otherwise, StarPU affects the
OpenCL devices in a round-robin fashion.
@item @code{workers_opencl_gpuid[STARPU_NMAXWORKERS]}:

@item @code{calibrate} (default = 0):
If this flag is set, StarPU will calibrate the performance models when
executing tasks. If this value is equal to -1, the default value is used. The
default value is overwritten by the @code{STARPU_CALIBRATE} environment
variable when it is set.
@end table

@end table


@node starpu_conf_init
@subsection @code{starpu_conf_init} -- Initialize starpu_conf structure
@table @asis

This function initializes the @code{starpu_conf} structure passed as argument
with the default values. In case some configuration parameters are already
specified through environment variables, @code{starpu_conf_init} initializes
the fields of the structure according to the environment variables. For
instance if @code{STARPU_CALIBRATE} is set, its value is put in the
@code{.ncuda} field of the structure passed as argument.

@item @emph{Return value}:
Upon successful completion, this function returns 0. Otherwise, @code{-EINVAL}
indicates that the argument was NULL.

@item @emph{Prototype}:
@code{int starpu_conf_init(struct starpu_conf *conf);}

@end table



@node starpu_shutdown
@subsection @code{starpu_shutdown} -- Terminate StarPU
@deftypefun void starpu_shutdown (void)
This is StarPU termination method. It must be called at the end of the
application: statistics and other post-mortem debugging information are not
guaranteed to be available until this method has been called.
@end deftypefun

@node Workers' Properties
@section Workers' Properties

@menu
* starpu_worker_get_count::     Get the number of processing units
* starpu_worker_get_count_by_type:: Get the number of processing units of a given type
* starpu_cpu_worker_get_count::  Get the number of CPU controlled by StarPU
* starpu_cuda_worker_get_count::  Get the number of CUDA devices controlled by StarPU
* starpu_opencl_worker_get_count::  Get the number of OpenCL devices controlled by StarPU
* starpu_spu_worker_get_count::  Get the number of Cell SPUs controlled by StarPU
* starpu_worker_get_id::        Get the identifier of the current worker
* starpu_worker_get_ids_by_type:: Get the list of identifiers of workers with a given type
* starpu_worker_get_devid::        Get the device identifier of a worker
* starpu_worker_get_type::      Get the type of processing unit associated to a worker
* starpu_worker_get_name::      Get the name of a worker
* starpu_worker_get_memory_node:: Get the memory node of a worker 
@end menu

@node starpu_worker_get_count
@subsection @code{starpu_worker_get_count} -- Get the number of processing units
@deftypefun unsigned starpu_worker_get_count (void)
This function returns the number of workers (i.e. processing units executing
StarPU tasks). The returned value should be at most @code{STARPU_NMAXWORKERS}.
@end deftypefun

@node starpu_worker_get_count_by_type
@subsection @code{starpu_worker_get_count_by_type} -- Get the number of processing units of a given type
@deftypefun int starpu_worker_get_count_by_type ({enum starpu_archtype} @var{type})
Returns the number of workers of the type indicated by the argument. A positive
(or null) value is returned in case of success, @code{-EINVAL} indicates that
the type is not valid otherwise.
@end deftypefun

@node starpu_cpu_worker_get_count
@subsection @code{starpu_cpu_worker_get_count} -- Get the number of CPU controlled by StarPU
@deftypefun unsigned starpu_cpu_worker_get_count (void)
This function returns the number of CPUs controlled by StarPU. The returned
value should be at most @code{STARPU_NMAXCPUS}.
@end deftypefun

@node starpu_cuda_worker_get_count
@subsection @code{starpu_cuda_worker_get_count} -- Get the number of CUDA devices controlled by StarPU
@deftypefun unsigned starpu_cuda_worker_get_count (void)
This function returns the number of CUDA devices controlled by StarPU. The returned
value should be at most @code{STARPU_MAXCUDADEVS}.
@end deftypefun

@node starpu_opencl_worker_get_count
@subsection @code{starpu_opencl_worker_get_count} -- Get the number of OpenCL devices controlled by StarPU
@deftypefun unsigned starpu_opencl_worker_get_count (void)
This function returns the number of OpenCL devices controlled by StarPU. The returned
value should be at most @code{STARPU_MAXOPENCLDEVS}.
@end deftypefun

@node starpu_spu_worker_get_count
@subsection @code{starpu_spu_worker_get_count} -- Get the number of Cell SPUs controlled by StarPU
@deftypefun unsigned starpu_opencl_worker_get_count (void)
This function returns the number of Cell SPUs controlled by StarPU.
@end deftypefun


@node starpu_worker_get_id
@subsection @code{starpu_worker_get_id} -- Get the identifier of the current worker
@deftypefun int starpu_worker_get_id (void)
This function returns the identifier of the worker associated to the calling
thread. The returned value is either -1 if the current context is not a StarPU
worker (i.e. when called from the application outside a task or a callback), or
an integer between 0 and @code{starpu_worker_get_count() - 1}.
@end deftypefun

@node starpu_worker_get_ids_by_type
@subsection @code{starpu_worker_get_ids_by_type} -- Get the list of identifiers of workers with a given type
@deftypefun int starpu_worker_get_ids_by_type ({enum starpu_archtype} @var{type}, int *@var{workerids}, int @var{maxsize})
Fill the workerids array with the identifiers of the workers that have the type
indicated in the first argument. The maxsize argument indicates the size of the
workids array. The returned value gives the number of identifiers that were put
in the array. @code{-ERANGE} is returned is maxsize is lower than the number of
workers with the appropriate type: in that case, the array is filled with the
maxsize first elements. To avoid such overflows, the value of maxsize can be
chosen by the means of the @code{starpu_worker_get_count_by_type} function, or
by passing a value greater or equal to @code{STARPU_NMAXWORKERS}.
@end deftypefun

@node starpu_worker_get_devid
@subsection @code{starpu_worker_get_devid} -- Get the device identifier of a worker
@deftypefun int starpu_worker_get_devid (int @var{id})
This functions returns the device id of the worker associated to an identifier
(as returned by the @code{starpu_worker_get_id} function). In the case of a
CUDA worker, this device identifier is the logical device identifier exposed by
CUDA (used by the @code{cudaGetDevice} function for instance). The device
identifier of a CPU worker is the logical identifier of the core on which the
worker was bound; this identifier is either provided by the OS or by the
@code{hwloc} library in case it is available.
@end deftypefun

@node starpu_worker_get_type
@subsection @code{starpu_worker_get_type} -- Get the type of processing unit associated to a worker
@deftypefun {enum starpu_archtype} starpu_worker_get_type (int @var{id})
This function returns the type of worker associated to an identifier (as
returned by the @code{starpu_worker_get_id} function). The returned value
indicates the architecture of the worker: @code{STARPU_CPU_WORKER} for a CPU
core, @code{STARPU_CUDA_WORKER} for a CUDA device,
@code{STARPU_OPENCL_WORKER} for a OpenCL device, and
@code{STARPU_GORDON_WORKER} for a Cell SPU. The value returned for an invalid
identifier is unspecified.
@end deftypefun

@node starpu_worker_get_name
@subsection @code{starpu_worker_get_name} -- Get the name of a worker

@deftypefun void starpu_worker_get_name (int @var{id}, char *@var{dst}, size_t @var{maxlen})
StarPU associates a unique human readable string to each processing unit. This
function copies at most the @var{maxlen} first bytes of the unique string
associated to a worker identified by its identifier @var{id} into the
@var{dst} buffer. The caller is responsible for ensuring that the @var{dst}
is a valid pointer to a buffer of @var{maxlen} bytes at least. Calling this
function on an invalid identifier results in an unspecified behaviour.
@end deftypefun

@node starpu_worker_get_memory_node
@subsection @code{starpu_worker_get_memory_node} -- Get the memory node of a worker
@deftypefun unsigned starpu_worker_get_memory_node (unsigned @var{workerid})
This function returns the identifier of the memory node associated to the
worker identified by @var{workerid}.
@end deftypefun


@node Data Library
@section Data Library

This section describes the data management facilities provided by StarPU.

We show how to use existing data interfaces in @ref{Data Interfaces}, but developers can
design their own data interfaces if required.

@menu
* starpu_malloc::          Allocate data and pin it
* starpu_access_mode::          Data access mode
* unsigned memory_node::        Memory node
* starpu_data_handle::          StarPU opaque data handle
* void *interface::             StarPU data interface
* starpu_data_register::        Register a piece of data to StarPU
* starpu_data_unregister::      Unregister a piece of data from StarPU
* starpu_data_invalidate::      Invalidate all data replicates
* starpu_data_acquire::         Access registered data from the application
* starpu_data_acquire_cb::      Access registered data from the application asynchronously
* starpu_data_release::         Release registered data from the application
* starpu_data_set_wt_mask::     Set the Write-Through mask
* starpu_data_prefetch_on_node:: Prefetch data to a given node
@end menu

@node starpu_malloc
@subsection @code{starpu_malloc} -- Allocate data and pin it
@deftypefun int starpu_malloc (void **@var{A}, size_t @var{dim})
This function allocates data of the given size in main memory. It will also try to pin it in
CUDA or OpenCL, so that data transfers from this buffer can be asynchronous, and
thus permit data transfer and computation overlapping. The allocated buffer must
be freed thanks to the @code{starpu_free} function.
@end deftypefun

@node starpu_access_mode
@subsection @code{starpu_access_mode} -- Data access mode
This datatype describes a data access mode. The different available modes are:
@table @asis
@table @asis 
@item @code{STARPU_R} read-only mode.
@item @code{STARPU_W} write-only mode.
@item @code{STARPU_RW} read-write mode. This is equivalent to @code{STARPU_R|STARPU_W}.
@item @code{STARPU_SCRATCH} scratch memory. A temporary buffer is allocated for the task, but StarPU does not enforce data consistency, i.e. each device has its own buffer, independently from each other (even for CPUs). This is useful for temporary variables. For now, no behaviour is defined concerning the relation with STARPU_R/W modes and the value provided at registration, i.e. the value of the scratch buffer is undefined at entry of the codelet function, but this is being considered for future extensions.
@item @code{STARPU_REDUX} reduction mode. TODO: document, as well as @code{starpu_data_set_reduction_methods}
@end table
@end table

@node unsigned memory_node
@subsection @code{unsigned memory_node} -- Memory node
@table @asis
@item @emph{Description}:
Every worker is associated to a memory node which is a logical abstraction of
the address space from which the processing unit gets its data. For instance,
the memory node associated to the different CPU workers represents main memory
(RAM), the memory node associated to a GPU is DRAM embedded on the device.
Every memory node is identified by a logical index which is accessible from the
@code{starpu_worker_get_memory_node} function. When registering a piece of data
to StarPU, the specified memory node indicates where the piece of data
initially resides (we also call this memory node the home node of a piece of
data).
@end table


@node starpu_data_handle
@subsection @code{starpu_data_handle} -- StarPU opaque data handle
@table @asis
@item @emph{Description}:
StarPU uses @code{starpu_data_handle} as an opaque handle to manage a piece of
data. Once a piece of data has been registered to StarPU, it is associated to a
@code{starpu_data_handle} which keeps track of the state of the piece of data
over the entire machine, so that we can maintain data consistency and locate
data replicates for instance.
@end table

@node void *interface
@subsection @code{void *interface} -- StarPU data interface
@table @asis
@item @emph{Description}:
Data management is done at a high-level in StarPU: rather than accessing a mere
list of contiguous buffers, the tasks may manipulate data that are described by
a high-level construct which we call data interface.

An example of data interface is the "vector" interface which describes a
contiguous data array on a spefic memory node. This interface is a simple
structure containing the number of elements in the array, the size of the
elements, and the address of the array in the appropriate address space (this
address may be invalid if there is no valid copy of the array in the memory
node). More informations on the data interfaces provided by StarPU are
given in @ref{Data Interfaces}.

When a piece of data managed by StarPU is used by a task, the task
implementation is given a pointer to an interface describing a valid copy of
the data that is accessible from the current processing unit.
@end table

@node starpu_data_register
@subsection @code{starpu_data_register} -- Register a piece of data to StarPU
@deftypefun void starpu_data_register (starpu_data_handle *@var{handleptr}, uint32_t @var{home_node}, void *@var{interface}, {struct starpu_data_interface_ops_t} *@var{ops})
Register a piece of data into the handle located at the @var{handleptr}
address. The @var{interface} buffer contains the initial description of the
data in the home node. The @var{ops} argument is a pointer to a structure
describing the different methods used to manipulate this type of interface. See
@ref{struct starpu_data_interface_ops_t} for more details on this structure.

If @code{home_node} is -1, StarPU will automatically
allocate the memory when it is used for the
first time in write-only mode. Once such data handle has been automatically
allocated, it is possible to access it using any access mode.

Note that StarPU supplies a set of predefined types of interface (e.g. vector or
matrix) which can be registered by the means of helper functions (e.g.
@code{starpu_vector_data_register} or @code{starpu_matrix_data_register}).
@end deftypefun

@node starpu_data_unregister
@subsection @code{starpu_data_unregister} -- Unregister a piece of data from StarPU
@deftypefun void starpu_data_unregister (starpu_data_handle @var{handle})
This function unregisters a data handle from StarPU. If the data was
automatically allocated by StarPU because the home node was -1, all
automatically allocated buffers are freed. Otherwise, a valid copy of the data
is put back into the home node in the buffer that was initially registered.
Using a data handle that has been unregistered from StarPU results in an
undefined behaviour.
@end deftypefun

@node starpu_data_invalidate
@subsection @code{starpu_data_invalidate} -- Invalidate all data replicates
@deftypefun void starpu_data_invalidate (starpu_data_handle @var{handle})
Destroy all replicates of the data handle. After data invalidation, the first
access to the handle must be performed in write-only mode. Accessing an
invalidated data in read-mode results in undefined behaviour.
@end deftypefun

@c TODO create a specific sections about user interaction with the DSM ?

@node starpu_data_acquire
@subsection @code{starpu_data_acquire} -- Access registered data from the application
@deftypefun int starpu_data_acquire (starpu_data_handle @var{handle}, starpu_access_mode @var{mode})
The application must call this function prior to accessing registered data from
main memory outside tasks. StarPU ensures that the application will get an
up-to-date copy of the data in main memory located where the data was
originally registered, and that all concurrent accesses (e.g. from tasks) will
be consistent with the access mode specified in the @var{mode} argument.
@code{starpu_data_release} must be called once the application does not need to
access the piece of data anymore.  Note that implicit data
dependencies are also enforced by @code{starpu_data_acquire}, i.e.
@code{starpu_data_acquire} will wait for all tasks scheduled to work on
the data, unless that they have not been disabled explictly by calling
@code{starpu_data_set_default_sequential_consistency_flag} or
@code{starpu_data_set_sequential_consistency_flag}.
@code{starpu_data_acquire} is a blocking call, so that it cannot be called from
tasks or from their callbacks (in that case, @code{starpu_data_acquire} returns
@code{-EDEADLK}). Upon successful completion, this function returns 0. 
@end deftypefun

@node starpu_data_acquire_cb
@subsection @code{starpu_data_acquire_cb} -- Access registered data from the application asynchronously
@deftypefun int starpu_data_acquire_cb (starpu_data_handle @var{handle}, starpu_access_mode @var{mode}, void (*@var{callback})(void *), void *@var{arg})
@code{starpu_data_acquire_cb} is the asynchronous equivalent of
@code{starpu_data_release}. When the data specified in the first argument is
available in the appropriate access mode, the callback function is executed.
The application may access the requested data during the execution of this
callback. The callback function must call @code{starpu_data_release} once the
application does not need to access the piece of data anymore. 
Note that implicit data dependencies are also enforced by
@code{starpu_data_acquire_cb} in case they are enabled.
 Contrary to @code{starpu_data_acquire}, this function is non-blocking and may
be called from task callbacks. Upon successful completion, this function
returns 0.
@end deftypefun

@node starpu_data_release
@subsection @code{starpu_data_release} -- Release registered data from the application
@deftypefun void starpu_data_release (starpu_data_handle @var{handle})
This function releases the piece of data acquired by the application either by
@code{starpu_data_acquire} or by @code{starpu_data_acquire_cb}.
@end deftypefun

@node starpu_data_set_wt_mask
@subsection @code{starpu_data_set_wt_mask} -- Set the Write-Through mask
@deftypefun void starpu_data_set_wt_mask (starpu_data_handle @var{handle}, uint32_t @var{wt_mask})
This function sets the write-through mask of a given data, i.e. a bitmask of
nodes where the data should be always replicated after modification.
@end deftypefun

@node starpu_data_prefetch_on_node
@subsection @code{starpu_data_prefetch_on_node} -- Prefetch data to a given node

@deftypefun int starpu_data_prefetch_on_node (starpu_data_handle @var{handle}, unsigned @var{node}, unsigned @var{async})
Issue a prefetch request for a given data to a given node, i.e.
requests that the data be replicated to the given node, so that it is available
there for tasks. If the @var{async} parameter is 0, the call will block until
the transfer is achieved, else the call will return as soon as the request is
scheduled (which may however have to wait for a task completion).
@end deftypefun

@node Data Interfaces
@section Data Interfaces

@menu
* Variable Interface::          
* Vector Interface::            
* Matrix Interface::            
* 3D Matrix Interface::             
* BCSR Interface for Sparse Matrices (Blocked Compressed Sparse Row Representation)::  
* CSR Interface for Sparse Matrices (Compressed Sparse Row Representation)::  
@end menu

@node Variable Interface
@subsection Variable Interface

@table @asis
@item @emph{Description}:
This variant of @code{starpu_data_register} uses the variable interface,
i.e. for a mere single variable. @code{ptr} is the address of the variable,
and @code{elemsize} is the size of the variable.
@item @emph{Prototype}:
@code{void starpu_variable_data_register(starpu_data_handle *handle,
                                   uint32_t home_node,
                                   uintptr_t ptr, size_t elemsize);}
@item @emph{Example}:
@cartouche
@smallexample
float var;
starpu_data_handle var_handle;
starpu_variable_data_register(&var_handle, 0, (uintptr_t)&var, sizeof(var));
@end smallexample
@end cartouche
@end table

@node Vector Interface
@subsection Vector Interface

@table @asis
@item @emph{Description}:
This variant of @code{starpu_data_register} uses the vector interface,
i.e. for mere arrays of elements. @code{ptr} is the address of the first
element in the home node. @code{nx} is the number of elements in the vector.
@code{elemsize} is the size of each element.
@item @emph{Prototype}:
@code{void starpu_vector_data_register(starpu_data_handle *handle, uint32_t home_node,
                        uintptr_t ptr, uint32_t nx, size_t elemsize);}
@item @emph{Example}:
@cartouche
@smallexample
float vector[NX];
starpu_data_handle vector_handle;
starpu_vector_data_register(&vector_handle, 0, (uintptr_t)vector, NX,
                            sizeof(vector[0]));
@end smallexample
@end cartouche
@end table

@node Matrix Interface
@subsection Matrix Interface

@table @asis
@item @emph{Description}:
This variant of @code{starpu_data_register} uses the matrix interface, i.e. for
matrices of elements. @code{ptr} is the address of the first element in the home
node. @code{ld} is the number of elements between rows. @code{nx} is the number
of elements in a row (this can be different from @code{ld} if there are extra
elements for alignment for instance). @code{ny} is the number of rows.
@code{elemsize} is the size of each element.
@item @emph{Prototype}:
@code{void starpu_matrix_data_register(starpu_data_handle *handle, uint32_t home_node,
                                       uintptr_t ptr, uint32_t ld, uint32_t nx,
                                       uint32_t ny, size_t elemsize);}
@item @emph{Example}:
@cartouche
@smallexample
float *matrix;
starpu_data_handle matrix_handle;
matrix = (float*)malloc(width * height * sizeof(float));
starpu_matrix_data_register(&matrix_handle, 0, (uintptr_t)matrix,
                            width, width, height, sizeof(float));
@end smallexample
@end cartouche
@end table

@node 3D Matrix Interface
@subsection 3D Matrix Interface

@table @asis
@item @emph{Description}:
This variant of @code{starpu_data_register} uses the 3D matrix interface.
@code{ptr} is the address of the array of first element in the home node.
@code{ldy} is the number of elements between rows. @code{ldz} is the number
of rows between z planes. @code{nx} is the number of elements in a row (this
can be different from @code{ldy} if there are extra elements for alignment
for instance). @code{ny} is the number of rows in a z plane (likewise with
@code{ldz}). @code{nz} is the number of z planes. @code{elemsize} is the size of
each element.
@item @emph{Prototype}:
@code{void starpu_block_data_register(starpu_data_handle *handle, uint32_t home_node,
                        uintptr_t ptr, uint32_t ldy, uint32_t ldz, uint32_t nx,
                        uint32_t ny, uint32_t nz, size_t elemsize);}
@item @emph{Example}:
@cartouche
@smallexample
float *block;
starpu_data_handle block_handle;
block = (float*)malloc(nx*ny*nz*sizeof(float));
starpu_block_data_register(&block_handle, 0, (uintptr_t)block,
                           nx, nx*ny, nx, ny, nz, sizeof(float));
@end smallexample
@end cartouche
@end table

@node BCSR Interface for Sparse Matrices (Blocked Compressed Sparse Row Representation)
@subsection BCSR Interface for Sparse Matrices (Blocked Compressed Sparse Row Representation)

@deftypefun void starpu_bcsr_data_register (starpu_data_handle *@var{handle}, uint32_t @var{home_node}, uint32_t @var{nnz}, uint32_t @var{nrow}, uintptr_t @var{nzval}, uint32_t *@var{colind}, uint32_t *@var{rowptr}, uint32_t @var{firstentry}, uint32_t @var{r}, uint32_t @var{c}, size_t @var{elemsize})
This variant of @code{starpu_data_register} uses the BCSR sparse matrix interface.
TODO
@end deftypefun

@node CSR Interface for Sparse Matrices (Compressed Sparse Row Representation)
@subsection CSR Interface for Sparse Matrices (Compressed Sparse Row Representation)

@deftypefun void starpu_csr_data_register (starpu_data_handle *@var{handle}, uint32_t @var{home_node}, uint32_t @var{nnz}, uint32_t @var{nrow}, uintptr_t @var{nzval}, uint32_t *@var{colind}, uint32_t *@var{rowptr}, uint32_t @var{firstentry}, size_t @var{elemsize})
This variant of @code{starpu_data_register} uses the CSR sparse matrix interface.
TODO
@end deftypefun

@node Data Partition
@section Data Partition

@menu
* struct starpu_data_filter::   StarPU filter structure
* starpu_data_partition::       Partition Data
* starpu_data_unpartition::     Unpartition Data
* starpu_data_get_nb_children::  
* starpu_data_get_sub_data::    
* Predefined filter functions::  
@end menu

@node struct starpu_data_filter
@subsection @code{struct starpu_data_filter} -- StarPU filter structure
@table @asis
@item @emph{Description}:
The filter structure describes a data partitioning operation, to be given to the
@code{starpu_data_partition} function, see @ref{starpu_data_partition} for an example.
@item @emph{Fields}:
@table @asis
@item @code{filter_func}:
This function fills the @code{child_interface} structure with interface
information for the @code{id}-th child of the parent @code{father_interface} (among @code{nparts}).
@code{void (*filter_func)(void *father_interface, void* child_interface, struct starpu_data_filter *, unsigned id, unsigned nparts);}
@item @code{nchildren}:
This is the number of parts to partition the data into.
@item @code{get_nchildren}:
This returns the number of children. This can be used instead of @code{nchildren} when the number of
children depends on the actual data (e.g. the number of blocks in a sparse
matrix).
@code{unsigned (*get_nchildren)(struct starpu_data_filter *, starpu_data_handle initial_handle);}
@item @code{get_child_ops}:
In case the resulting children use a different data interface, this function
returns which interface is used by child number @code{id}.
@code{struct starpu_data_interface_ops_t *(*get_child_ops)(struct starpu_data_filter *, unsigned id);}
@item @code{filter_arg}:
Some filters take an addition parameter, but this is usually unused.
@item @code{filter_arg_ptr}:
Some filters take an additional array parameter like the sizes of the parts, but
this is usually unused.
@end table
@end table

@node starpu_data_partition
@subsection starpu_data_partition -- Partition Data

@table @asis
@item @emph{Description}:
This requests partitioning one StarPU data @code{initial_handle} into several
subdata according to the filter @code{f}
@item @emph{Prototype}:
@code{void starpu_data_partition(starpu_data_handle initial_handle, struct starpu_data_filter *f);}
@item @emph{Example}:
@cartouche
@smallexample
struct starpu_data_filter f = @{
    .filter_func = starpu_vertical_block_filter_func,
    .nchildren = nslicesx,
    .get_nchildren = NULL,
    .get_child_ops = NULL
@};
starpu_data_partition(A_handle, &f);
@end smallexample
@end cartouche
@end table

@node starpu_data_unpartition
@subsection starpu_data_unpartition -- Unpartition data

@table @asis
@item @emph{Description}:
This unapplies one filter, thus unpartitioning the data. The pieces of data are
collected back into one big piece in the @code{gathering_node} (usually 0).
@item @emph{Prototype}:
@code{void starpu_data_unpartition(starpu_data_handle root_data, uint32_t gathering_node);}
@item @emph{Example}:
@cartouche
@smallexample
starpu_data_unpartition(A_handle, 0);
@end smallexample
@end cartouche
@end table

@node starpu_data_get_nb_children
@subsection starpu_data_get_nb_children

@table @asis
@item @emph{Description}:
This function returns the number of children.
@item @emph{Return value}:
The number of children.
@item @emph{Prototype}:
@code{int starpu_data_get_nb_children(starpu_data_handle handle);}
@end table

@c starpu_data_handle starpu_data_get_child(starpu_data_handle handle, unsigned i);

@node starpu_data_get_sub_data
@subsection starpu_data_get_sub_data

@table @asis
@item @emph{Description}:
After partitioning a StarPU data by applying a filter,
@code{starpu_data_get_sub_data} can be used to get handles for each of the data
portions. @code{root_data} is the parent data that was partitioned. @code{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.
@item @emph{Return value}:
A handle to the subdata.
@item @emph{Prototype}:
@code{starpu_data_handle starpu_data_get_sub_data(starpu_data_handle root_data, unsigned depth, ... );}
@item @emph{Example}:
@cartouche
@smallexample
h = starpu_data_get_sub_data(A_handle, 1, taskx);
@end smallexample
@end cartouche
@end table

@node Predefined filter functions
@subsection Predefined filter functions

@menu
* Partitioning BCSR Data::      
* Partitioning BLAS interface::  
* Partitioning Vector Data::    
* Partitioning Block Data::     
@end menu

This section gives a partial list of the predefined partitioning functions.
Examples on how to use them are shown in @ref{Partitioning Data}. The complete
list can be found in @code{starpu_data_filters.h} .

@node Partitioning BCSR Data
@subsubsection Partitioning BCSR Data

@deftypefun void starpu_canonical_block_filter_bcsr (void *@var{father_interface}, void *@var{child_interface}, {struct starpu_data_filter} *@var{f}, unsigned @var{id}, unsigned @var{nparts})
TODO
@end deftypefun

@deftypefun void starpu_vertical_block_filter_func_csr (void *@var{father_interface}, void *@var{child_interface}, {struct starpu_data_filter} *@var{f}, unsigned @var{id}, unsigned @var{nparts})
TODO
@end deftypefun

@node Partitioning BLAS interface
@subsubsection Partitioning BLAS interface

@deftypefun void starpu_block_filter_func (void *@var{father_interface}, void *@var{child_interface}, {struct starpu_data_filter} *@var{f}, unsigned @var{id}, unsigned @var{nparts})
This partitions a dense Matrix into horizontal blocks.
@end deftypefun

@deftypefun void starpu_vertical_block_filter_func (void *@var{father_interface}, void *@var{child_interface}, {struct starpu_data_filter} *@var{f}, unsigned @var{id}, unsigned @var{nparts})
This partitions a dense Matrix into vertical blocks.
@end deftypefun

@node Partitioning Vector Data
@subsubsection Partitioning Vector Data

@deftypefun void starpu_block_filter_func_vector (void *@var{father_interface}, void *@var{child_interface}, {struct starpu_data_filter} *@var{f}, unsigned @var{id}, unsigned @var{nparts})
This partitions a vector into blocks of the same size.
@end deftypefun


@deftypefun void starpu_vector_list_filter_func (void *@var{father_interface}, void *@var{child_interface}, {struct starpu_data_filter} *@var{f}, unsigned @var{id}, unsigned @var{nparts})
This partitions a vector into blocks of sizes given in @var{filter_arg_ptr}.
@end deftypefun

@deftypefun void starpu_vector_divide_in_2_filter_func (void *@var{father_interface}, void *@var{child_interface}, {struct starpu_data_filter} *@var{f}, unsigned @var{id}, unsigned @var{nparts})
This partitions a vector into two blocks, the first block size being given in @var{filter_arg}.
@end deftypefun


@node Partitioning Block Data
@subsubsection Partitioning Block Data

@deftypefun void starpu_block_filter_func_block (void *@var{father_interface}, void *@var{child_interface}, {struct starpu_data_filter} *@var{f}, unsigned @var{id}, unsigned @var{nparts})
This partitions a 3D matrix along the X axis.
@end deftypefun

@node Codelets and Tasks
@section Codelets and Tasks

This section describes the interface to manipulate codelets and tasks.

@deftp {Data Type} {struct starpu_codelet}
The codelet structure describes a kernel that is possibly implemented on various
targets. For compatibility, make sure to initialize the whole structure to zero.

@table @asis
@item @code{where}
Indicates which types of processing units are able to execute the codelet.
@code{STARPU_CPU|STARPU_CUDA} for instance indicates that the codelet is
implemented for both CPU cores and CUDA devices while @code{STARPU_GORDON}
indicates that it is only available on Cell SPUs.

@item @code{cpu_func} (optional)
Is a function pointer to the CPU implementation of the codelet. Its prototype
must be: @code{void cpu_func(void *buffers[], void *cl_arg)}. The first
argument being the array of data managed by the data management library, and
the second argument is a pointer to the argument passed from the @code{cl_arg}
field of the @code{starpu_task} structure.
The @code{cpu_func} field is ignored if @code{STARPU_CPU} does not appear in
the @code{where} field, it must be non-null otherwise.

@item @code{cuda_func} (optional)
Is a function pointer to the CUDA implementation of the codelet. @emph{This
must be a host-function written in the CUDA runtime API}. Its prototype must
be: @code{void cuda_func(void *buffers[], void *cl_arg);}. The @code{cuda_func}
field is ignored if @code{STARPU_CUDA} does not appear in the @code{where}
field, it must be non-null otherwise.

@item @code{opencl_func} (optional)
Is a function pointer to the OpenCL implementation of the codelet. Its
prototype must be:
@code{void opencl_func(starpu_data_interface_t *descr, void *arg);}.
This pointer is ignored if @code{STARPU_OPENCL} does not appear in the
@code{where} field, it must be non-null otherwise.

@item @code{gordon_func} (optional)
This is the index of the Cell SPU implementation within the Gordon library.
See Gordon documentation for more details on how to register a kernel and
retrieve its index.

@item @code{nbuffers}
Specifies the number of arguments taken by the codelet. These arguments are
managed by the DSM and are accessed from the @code{void *buffers[]}
array. The constant argument passed with the @code{cl_arg} field of the
@code{starpu_task} structure is not counted in this number.  This value should
not be above @code{STARPU_NMAXBUFS}.

@item @code{model} (optional)
This is a pointer to the task duration performance model associated to this
codelet. This optional field is ignored when set to @code{NULL}.

TODO

@item @code{power_model} (optional)
This is a pointer to the task power consumption performance model associated
to this codelet. This optional field is ignored when set to @code{NULL}.
In the case of parallel codelets, this has to account for all processing units
involved in the parallel execution.

TODO

@end table
@end deftp

@deftp {Data Type} {struct starpu_task}
The @code{starpu_task} structure describes a task that can be offloaded on the various
processing units managed by StarPU. It instantiates a codelet. It can either be
allocated dynamically with the @code{starpu_task_create} method, or declared
statically. In the latter case, the programmer has to zero the
@code{starpu_task} structure and to fill the different fields properly. The
indicated default values correspond to the configuration of a task allocated
with @code{starpu_task_create}.

@table @asis
@item @code{cl}
Is a pointer to the corresponding @code{starpu_codelet} data structure. This
describes where the kernel should be executed, and supplies the appropriate
implementations. When set to @code{NULL}, no code is executed during the tasks,
such empty tasks can be useful for synchronization purposes.

@item @code{buffers}
Is an array of @code{starpu_buffer_descr_t} structures. It describes the
different pieces of data accessed by the task, and how they should be accessed.
The @code{starpu_buffer_descr_t} structure is composed of two fields, the
@code{handle} field specifies the handle of the piece of data, and the
@code{mode} field is the required access mode (eg @code{STARPU_RW}). The number
of entries in this array must be specified in the @code{nbuffers} field of the
@code{starpu_codelet} structure, and should not excede @code{STARPU_NMAXBUFS}.
If unsufficient, this value can be set with the @code{--enable-maxbuffers}
option when configuring StarPU.

@item @code{cl_arg} (optional; default: @code{NULL})
This pointer is passed to the codelet through the second argument
of the codelet implementation (e.g. @code{cpu_func} or @code{cuda_func}).
In the specific case of the Cell processor, see the @code{cl_arg_size}
argument.

@item @code{cl_arg_size} (optional, Cell-specific)
In the case of the Cell processor, the @code{cl_arg} pointer is not directly
given to the SPU function. A buffer of size @code{cl_arg_size} is allocated on
the SPU. This buffer is then filled with the @code{cl_arg_size} bytes starting
at address @code{cl_arg}. In this case, the argument given to the SPU codelet
is therefore not the @code{cl_arg} pointer, but the address of the buffer in
local store (LS) instead. This field is ignored for CPU, CUDA and OpenCL
codelets, where the @code{cl_arg} pointer is given as such.

@item @code{callback_func} (optional) (default: @code{NULL})
This is a function pointer of prototype @code{void (*f)(void *)} which
specifies a possible callback. If this pointer is non-null, the callback
function is executed @emph{on the host} after the execution of the task. The
callback is passed the value contained in the @code{callback_arg} field. No
callback is executed if the field is set to @code{NULL}.

@item @code{callback_arg} (optional) (default: @code{NULL})
This is the pointer passed to the callback function. This field is ignored if
the @code{callback_func} is set to @code{NULL}.

@item @code{use_tag} (optional) (default: @code{0})
If set, this flag indicates that the task should be associated with the tag
contained in the @code{tag_id} field. Tag allow the application to synchronize
with the task and to express task dependencies easily.

@item @code{tag_id}
This fields contains the tag associated to the task if the @code{use_tag} field
was set, it is ignored otherwise.

@item @code{synchronous}
If this flag is set, the @code{starpu_task_submit} function is blocking and
returns only when the task has been executed (or if no worker is able to
process the task). Otherwise, @code{starpu_task_submit} returns immediately.

@item @code{priority} (optional) (default: @code{STARPU_DEFAULT_PRIO})
This field indicates a level of priority for the task. This is an integer value
that must be set between the return values of the
@code{starpu_sched_get_min_priority} function for the least important tasks,
and that of the @code{starpu_sched_get_max_priority} for the most important
tasks (included). The @code{STARPU_MIN_PRIO} and @code{STARPU_MAX_PRIO} macros
are provided for convenience and respectively returns value of
@code{starpu_sched_get_min_priority} and @code{starpu_sched_get_max_priority}.
Default priority is @code{STARPU_DEFAULT_PRIO}, which is always defined as 0 in
order to allow static task initialization.  Scheduling strategies that take
priorities into account can use this parameter to take better scheduling
decisions, but the scheduling policy may also ignore it.

@item @code{execute_on_a_specific_worker} (default: @code{0})
If this flag is set, StarPU will bypass the scheduler and directly affect this
task to the worker specified by the @code{workerid} field.

@item @code{workerid} (optional)
If the @code{execute_on_a_specific_worker} field is set, this field indicates
which is the identifier of the worker that should process this task (as
returned by @code{starpu_worker_get_id}). This field is ignored if
@code{execute_on_a_specific_worker} field is set to 0.

@item @code{detach} (optional) (default: @code{1})
If this flag is set, it is not possible to synchronize with the task
by the means of @code{starpu_task_wait} later on. Internal data structures
are only guaranteed to be freed once @code{starpu_task_wait} is called if the
flag is not set.

@item @code{destroy} (optional) (default: @code{1})
If this flag is set, the task structure will automatically be freed, either
after the execution of the callback if the task is detached, or during
@code{starpu_task_wait} otherwise. If this flag is not set, dynamically
allocated data structures will not be freed until @code{starpu_task_destroy} is
called explicitly. Setting this flag for a statically allocated task structure
will result in undefined behaviour.

@item @code{predicted} (output field)
Predicted duration of the task. This field is only set if the scheduling
strategy used performance models.

@end table
@end deftp

@deftypefun void starpu_task_init ({struct starpu_task} *@var{task})
Initialize @var{task} with default values. This function is implicitly
called by @code{starpu_task_create}. By default, tasks initialized with
@code{starpu_task_init} must be deinitialized explicitly with
@code{starpu_task_deinit}. Tasks can also be initialized statically, using the
constant @code{STARPU_TASK_INITIALIZER}.
@end deftypefun

@deftypefun {struct starpu_task *} starpu_task_create (void)
Allocate a task structure and initialize it with default values. Tasks
allocated dynamically with @code{starpu_task_create} are automatically freed when the
task is terminated. If the destroy flag is explicitly unset, the resources used
by the task are freed by calling
@code{starpu_task_destroy}.
@end deftypefun

@deftypefun void starpu_task_deinit ({struct starpu_task} *@var{task})
Release all the structures automatically allocated to execute @var{task}. This is
called automatically by @code{starpu_task_destroy}, but the task structure itself is not
freed. This should be used for statically allocated tasks for instance.
@end deftypefun

@deftypefun void starpu_task_destroy ({struct starpu_task} *@var{task})
Free the resource allocated during @code{starpu_task_create} and
associated with @var{task}. This function can be called automatically
after the execution of a task by setting the @code{destroy} flag of the
@code{starpu_task} structure (default behaviour).  Calling this function
on a statically allocated task results in an undefined behaviour.
@end deftypefun

@deftypefun int starpu_task_wait ({struct starpu_task} *@var{task})
This function blocks until @var{task} has been executed. It is not possible to
synchronize with a task more than once. It is not possible to wait for
synchronous or detached tasks.

Upon successful completion, this function returns 0. Otherwise, @code{-EINVAL}
indicates that the specified task was either synchronous or detached.
@end deftypefun

@deftypefun int starpu_task_submit ({struct starpu_task} *@var{task})
This function submits @var{task} to StarPU. Calling this function does
not mean that the task will be executed immediately as there can be data or task
(tag) dependencies that are not fulfilled yet: StarPU will take care of
scheduling this task with respect to such dependencies.
This function returns immediately if the @code{synchronous} field of the
@code{starpu_task} structure was set to 0, and block until the termination of
the task otherwise. It is also possible to synchronize the application with
asynchronous tasks by the means of tags, using the @code{starpu_tag_wait}
function for instance.

In case of success, this function returns 0, a return value of @code{-ENODEV}
means that there is no worker able to process this task (e.g. there is no GPU
available and this task is only implemented for CUDA devices).
@end deftypefun

@deftypefun int starpu_task_wait_for_all (void)
This function blocks until all the tasks that were submitted are terminated.
@end deftypefun

@deftypefun {struct starpu_task *} starpu_get_current_task (void)
This function returns the task currently executed by the worker, or
NULL if it is called either from a thread that is not a task or simply
because there is no task being executed at the moment.
@end deftypefun

@deftypefun void starpu_display_codelet_stats ({struct starpu_codelet_t} *@var{cl})
Output on @code{stderr} some statistics on the codelet @var{cl}.
@end deftypefun


@c Callbacks : what can we put in callbacks ?

@node Explicit Dependencies
@section Explicit Dependencies

@menu
* starpu_task_declare_deps_array::        starpu_task_declare_deps_array
* starpu_tag_t::                Task logical identifier
* starpu_tag_declare_deps::     Declare the Dependencies of a Tag
* starpu_tag_declare_deps_array::  Declare the Dependencies of a Tag
* starpu_tag_wait::             Block until a Tag is terminated
* starpu_tag_wait_array::       Block until a set of Tags is terminated
* starpu_tag_remove::           Destroy a Tag
* starpu_tag_notify_from_apps::  Feed a tag explicitly
@end menu

@node starpu_task_declare_deps_array
@subsection @code{starpu_task_declare_deps_array} -- Declare task dependencies
@deftypefun void starpu_task_declare_deps_array ({struct starpu_task} *@var{task}, unsigned @var{ndeps}, {struct starpu_task} *@var{task_array}[])
Declare task dependencies between a @var{task} and an array of tasks of length
@var{ndeps}. This function must be called prior to the submission of the task,
but it may called after the submission or the execution of the tasks in the
array provided the tasks are still valid (ie. they were not automatically
destroyed). Calling this function on a task that was already submitted or with
an entry of @var{task_array} that is not a valid task anymore results in an
undefined behaviour. If @var{ndeps} is null, no dependency is added. It is
possible to call @code{starpu_task_declare_deps_array} multiple times on the
same task, in this case, the dependencies are added. It is possible to have
redundancy in the task dependencies.
@end deftypefun



@node starpu_tag_t
@subsection @code{starpu_tag_t} -- Task logical identifier
@table @asis
@item @emph{Description}:
It is possible to associate a task with a unique ``tag'' chosen by the application, and to express
dependencies between tasks by the means of those tags. To do so, fill the
@code{tag_id} field of the @code{starpu_task} structure with a tag number (can
be arbitrary) and set the @code{use_tag} field to 1.

If @code{starpu_tag_declare_deps} is called with this tag number, the task will
not be started until the tasks which holds the declared dependency tags are
completed.
@end table

@node starpu_tag_declare_deps
@subsection @code{starpu_tag_declare_deps} -- Declare the Dependencies of a Tag
@table @asis
@item @emph{Description}:
Specify the dependencies of the task identified by tag @code{id}. The first
argument specifies the tag which is configured, the second argument gives the
number of tag(s) on which @code{id} depends. The following arguments are the
tags which have to be terminated to unlock the task.

This function must be called before the associated task is submitted to StarPU
with @code{starpu_task_submit}.

@item @emph{Remark}
Because of the variable arity of @code{starpu_tag_declare_deps}, note that the
last arguments @emph{must} be of type @code{starpu_tag_t}: constant values
typically need to be explicitly casted. Using the
@code{starpu_tag_declare_deps_array} function avoids this hazard.

@item @emph{Prototype}:
@code{void starpu_tag_declare_deps(starpu_tag_t id, unsigned ndeps, ...);}

@item @emph{Example}:
@cartouche
@example
/*  Tag 0x1 depends on tags 0x32 and 0x52 */
starpu_tag_declare_deps((starpu_tag_t)0x1,
        2, (starpu_tag_t)0x32, (starpu_tag_t)0x52);
@end example
@end cartouche

@end table

@node starpu_tag_declare_deps_array
@subsection @code{starpu_tag_declare_deps_array} -- Declare the Dependencies of a Tag
@table @asis
@item @emph{Description}:
This function is similar to @code{starpu_tag_declare_deps}, except that its
does not take a variable number of arguments but an array of tags of size
@code{ndeps}.
@item @emph{Prototype}:
@code{void starpu_tag_declare_deps_array(starpu_tag_t id, unsigned ndeps, starpu_tag_t *array);}
@item @emph{Example}:
@cartouche
@example
/*  Tag 0x1 depends on tags 0x32 and 0x52 */
starpu_tag_t tag_array[2] = @{0x32, 0x52@};
starpu_tag_declare_deps_array((starpu_tag_t)0x1, 2, tag_array);
@end example
@end cartouche


@end table


@node starpu_tag_wait
@subsection @code{starpu_tag_wait} -- Block until a Tag is terminated
@deftypefun void starpu_tag_wait (starpu_tag_t @var{id})
This function blocks until the task associated to tag @var{id} has been
executed. This is a blocking call which must therefore not be called within
tasks or callbacks, but only from the application directly.  It is possible to
synchronize with the same tag multiple times, as long as the
@code{starpu_tag_remove} function is not called.  Note that it is still
possible to synchronize with a tag associated to a task which @code{starpu_task}
data structure was freed (e.g. if the @code{destroy} flag of the
@code{starpu_task} was enabled).
@end deftypefun

@node starpu_tag_wait_array
@subsection @code{starpu_tag_wait_array} -- Block until a set of Tags is terminated
@deftypefun void starpu_tag_wait_array (unsigned @var{ntags}, starpu_tag_t *@var{id})
This function is similar to @code{starpu_tag_wait} except that it blocks until
@emph{all} the @var{ntags} tags contained in the @var{id} array are
terminated.
@end deftypefun

@node starpu_tag_remove
@subsection @code{starpu_tag_remove} -- Destroy a Tag
@deftypefun void starpu_tag_remove (starpu_tag_t @var{id})
This function releases the resources associated to tag @var{id}. It can be
called once the corresponding task has been executed and when there is
no other tag that depend on this tag anymore.
@end deftypefun

@node starpu_tag_notify_from_apps
@subsection @code{starpu_tag_notify_from_apps} -- Feed a Tag explicitly
@deftypefun void starpu_tag_notify_from_apps (starpu_tag_t @var{id})
This function explicitly unlocks tag @var{id}. It may be useful in the
case of applications which execute part of their computation outside StarPU
tasks (e.g. third-party libraries).  It is also provided as a
convenient tool for the programmer, for instance to entirely construct the task
DAG before actually giving StarPU the opportunity to execute the tasks.
@end deftypefun

@node Implicit Data Dependencies
@section Implicit Data Dependencies

@menu
* starpu_data_set_default_sequential_consistency_flag::        starpu_data_set_default_sequential_consistency_flag
* starpu_data_get_default_sequential_consistency_flag::        starpu_data_get_default_sequential_consistency_flag
* starpu_data_set_sequential_consistency_flag::                starpu_data_set_sequential_consistency_flag
@end menu

In this section, we describe how StarPU makes it possible to insert implicit
task dependencies in order to enforce sequential data consistency. When this
data consistency is enabled on a specific data handle, any data access will
appear as sequentially consistent from the application. For instance, if the
application submits two tasks that access the same piece of data in read-only
mode, and then a third task that access it in write mode, dependencies will be
added between the two first tasks and the third one. Implicit data dependencies
are also inserted in the case of data accesses from the application.

@node starpu_data_set_default_sequential_consistency_flag
@subsection @code{starpu_data_set_default_sequential_consistency_flag} -- Set default sequential consistency flag
@deftypefun void starpu_data_set_default_sequential_consistency_flag (unsigned @var{flag})
Set the default sequential consistency flag. If a non-zero value is passed, a
sequential data consistency will be enforced for all handles registered after
this function call, otherwise it is disabled. By default, StarPU enables
sequential data consistency. It is also possible to select the data consistency
mode of a specific data handle with the
@code{starpu_data_set_sequential_consistency_flag} function.
@end deftypefun

@node starpu_data_get_default_sequential_consistency_flag
@subsection @code{starpu_data_get_default_sequential_consistency_flag} -- Get current default sequential consistency flag
@deftypefun unsigned starpu_data_set_default_sequential_consistency_flag (void)
This function returns the current default sequential consistency flag.
@end deftypefun

@node starpu_data_set_sequential_consistency_flag
@subsection @code{starpu_data_set_sequential_consistency_flag} -- Set data sequential consistency mode
@deftypefun void starpu_data_set_sequential_consistency_flag (starpu_data_handle @var{handle}, unsigned @var{flag})
Select the data consistency mode associated to a data handle. The consistency
mode set using this function has the priority over the default mode which can
be set with @code{starpu_data_set_sequential_consistency_flag}.
@end deftypefun

@node Performance Model API
@section Performance Model API

@menu
* starpu_load_history_debug::   
* starpu_perfmodel_debugfilepath::  
* starpu_perfmodel_get_arch_name::  
* starpu_force_bus_sampling::   
@end menu

@node starpu_load_history_debug
@subsection @code{starpu_load_history_debug}
@deftypefun int starpu_load_history_debug ({const char} *@var{symbol}, {struct starpu_perfmodel_t} *@var{model})
TODO
@end deftypefun

@node starpu_perfmodel_debugfilepath
@subsection @code{starpu_perfmodel_debugfilepath}
@deftypefun void starpu_perfmodel_debugfilepath ({struct starpu_perfmodel_t} *@var{model}, {enum starpu_perf_archtype} @var{arch}, char *@var{path}, size_t @var{maxlen})
TODO
@end deftypefun

@node starpu_perfmodel_get_arch_name
@subsection @code{starpu_perfmodel_get_arch_name}
@deftypefun void starpu_perfmodel_get_arch_name ({enum starpu_perf_archtype} @var{arch}, char *@var{archname}, size_t @var{maxlen})
TODO
@end deftypefun

@node starpu_force_bus_sampling
@subsection @code{starpu_force_bus_sampling}
@deftypefun void starpu_force_bus_sampling (void)
This forces sampling the bus performance model again.
@end deftypefun


@node Profiling API
@section Profiling API

@menu
* starpu_profiling_status_set::  starpu_profiling_status_set
* starpu_profiling_status_get::  starpu_profiling_status_get
* struct starpu_task_profiling_info::  task profiling information
* struct starpu_worker_profiling_info::  worker profiling information
* starpu_worker_get_profiling_info::  starpu_worker_get_profiling_info
* struct starpu_bus_profiling_info::  bus profiling information
* starpu_bus_get_count::        
* starpu_bus_get_id::           
* starpu_bus_get_src::          
* starpu_bus_get_dst::          
* starpu_timing_timespec_delay_us::  
* starpu_timing_timespec_to_us::  
* starpu_bus_profiling_helper_display_summary::  
* starpu_worker_profiling_helper_display_summary::  
@end menu

@node starpu_profiling_status_set
@subsection @code{starpu_profiling_status_set} -- Set current profiling status
@table @asis
@item @emph{Description}:
Thie function sets the profiling status. Profiling is activated by passing
@code{STARPU_PROFILING_ENABLE} in @code{status}. Passing
@code{STARPU_PROFILING_DISABLE} disables profiling. Calling this function
resets all profiling measurements. When profiling is enabled, the
@code{profiling_info} field of the @code{struct starpu_task} structure points
to a valid @code{struct starpu_task_profiling_info} structure containing
information about the execution of the task.
@item @emph{Return value}:
Negative return values indicate an error, otherwise the previous status is
returned.
@item @emph{Prototype}:
@code{int starpu_profiling_status_set(int status);}
@end table

@node starpu_profiling_status_get
@subsection @code{starpu_profiling_status_get} -- Get current profiling status
@deftypefun int starpu_profiling_status_get (void)
Return the current profiling status or a negative value in case there was an error.
@end deftypefun

@node struct starpu_task_profiling_info
@subsection @code{struct starpu_task_profiling_info} -- Task profiling information
@table @asis
@item @emph{Description}:
This structure contains information about the execution of a task. It is
accessible from the @code{.profiling_info} field of the @code{starpu_task}
structure if profiling was enabled.
@item @emph{Fields}:
@table @asis
@item @code{submit_time}:
Date of task submission (relative to the initialization of StarPU).
@item @code{start_time}:
Date of task execution beginning (relative to the initialization of StarPU).
@item @code{end_time}:
Date of task execution termination (relative to the initialization of StarPU).
@item @code{workerid}:
Identifier of the worker which has executed the task.
@end table
@end table

@node struct starpu_worker_profiling_info
@subsection @code{struct starpu_worker_profiling_info} -- Worker profiling information
@table @asis
@item @emph{Description}:
This structure contains the profiling information associated to a worker.
@item @emph{Fields}:
@table @asis
@item @code{start_time}:
Starting date for the reported profiling measurements.
@item @code{total_time}:
Duration of the profiling measurement interval.
@item @code{executing_time}:
Time spent by the worker to execute tasks during the profiling measurement interval.
@item @code{sleeping_time}:
Time spent idling by the worker during the profiling measurement interval.
@item @code{executed_tasks}:
Number of tasks executed by the worker during the profiling measurement interval.
@end table
@end table

@node starpu_worker_get_profiling_info
@subsection @code{starpu_worker_get_profiling_info} -- Get worker profiling info
@table @asis

@item @emph{Description}:
Get the profiling info associated to the worker identified by @code{workerid},
and reset the profiling measurements. If the @code{worker_info} argument is
NULL, only reset the counters associated to worker @code{workerid}.
@item @emph{Return value}:
Upon successful completion, this function returns 0. Otherwise, a negative
value is returned.

@item @emph{Prototype}:
@code{int starpu_worker_get_profiling_info(int workerid, struct starpu_worker_profiling_info *worker_info);}
@end table

@node struct starpu_bus_profiling_info
@subsection @code{struct starpu_bus_profiling_info} -- Bus profiling information
@table @asis
@item @emph{Description}:
TODO
@item @emph{Fields}:
@table @asis
@item @code{start_time}:
TODO
@item @code{total_time}:
TODO
@item @code{transferred_bytes}:
TODO
@item @code{transfer_count}:
TODO
@end table
@end table

@node starpu_bus_get_count
@subsection @code{starpu_bus_get_count}
@deftypefun int starpu_bus_get_count (void)
TODO
@end deftypefun

@node starpu_bus_get_id
@subsection @code{starpu_bus_get_id}
@deftypefun int starpu_bus_get_id (int @var{src}, int @var{dst})
TODO
@end deftypefun

@node starpu_bus_get_src
@subsection @code{starpu_bus_get_src}
@deftypefun int starpu_bus_get_src (int @var{busid})
TODO
@end deftypefun

@node starpu_bus_get_dst
@subsection @code{starpu_bus_get_dst}
@deftypefun int starpu_bus_get_dst (int @var{busid})
TODO
@end deftypefun

@node starpu_timing_timespec_delay_us
@subsection @code{starpu_timing_timespec_delay_us}
@deftypefun double starpu_timing_timespec_delay_us ({struct timespec} *@var{start}, {struct timespec} *@var{end})
TODO
@end deftypefun

@node starpu_timing_timespec_to_us
@subsection @code{starpu_timing_timespec_to_us}
@deftypefun double starpu_timing_timespec_to_us ({struct timespec} *@var{ts})
TODO
@end deftypefun

@node starpu_bus_profiling_helper_display_summary
@subsection @code{starpu_bus_profiling_helper_display_summary}
@deftypefun void starpu_bus_profiling_helper_display_summary (void)
TODO
@end deftypefun

@node starpu_worker_profiling_helper_display_summary
@subsection @code{starpu_worker_profiling_helper_display_summary}
@deftypefun void starpu_worker_profiling_helper_display_summary (void)
TODO
@end deftypefun



@node CUDA extensions
@section CUDA extensions

@c void starpu_malloc(float **A, size_t dim);

@menu
* starpu_cuda_get_local_stream::  Get current worker's CUDA stream
* starpu_helper_cublas_init::   Initialize CUBLAS on every CUDA device
* starpu_helper_cublas_shutdown::  Deinitialize CUBLAS on every CUDA device
@end menu

@node starpu_cuda_get_local_stream
@subsection @code{starpu_cuda_get_local_stream} -- Get current worker's CUDA stream
@deftypefun {cudaStream_t *} starpu_cuda_get_local_stream (void)
StarPU provides a stream for every CUDA device controlled by StarPU. This
function is only provided for convenience so that programmers can easily use
asynchronous operations within codelets without having to create a stream by
hand. Note that the application is not forced to use the stream provided by
@code{starpu_cuda_get_local_stream} and may also create its own streams.
Synchronizing with @code{cudaThreadSynchronize()} is allowed, but will reduce
the likelihood of having all transfers overlapped.
@end deftypefun

@node starpu_helper_cublas_init
@subsection @code{starpu_helper_cublas_init} -- Initialize CUBLAS on every CUDA device
@deftypefun void starpu_helper_cublas_init (void)
The CUBLAS library must be initialized prior to any CUBLAS call. Calling
@code{starpu_helper_cublas_init} will initialize CUBLAS on every CUDA device
controlled by StarPU. This call blocks until CUBLAS has been properly
initialized on every device.
@end deftypefun

@node starpu_helper_cublas_shutdown
@subsection @code{starpu_helper_cublas_shutdown} -- Deinitialize CUBLAS on every CUDA device
@deftypefun void starpu_helper_cublas_shutdown (void)
This function synchronously deinitializes the CUBLAS library on every CUDA device.
@end deftypefun

@node OpenCL extensions
@section OpenCL extensions

@menu
* Enabling OpenCL::            Enabling OpenCL
* Compiling OpenCL kernels::   Compiling OpenCL kernels
* Loading OpenCL kernels::     Loading OpenCL kernels
* OpenCL statistics::          Collecting statistics from OpenCL
@end menu

@node Enabling OpenCL
@subsection Enabling OpenCL

On GPU devices which can run both CUDA and OpenCL, CUDA will be
enabled by default. To enable OpenCL, you need either to disable CUDA
when configuring StarPU:

@example
% ./configure --disable-cuda
@end example

or when running applications:

@example
% STARPU_NCUDA=0 ./application
@end example

OpenCL will automatically be started on any device not yet used by
CUDA. So on a machine running 4 GPUS, it is therefore possible to
enable CUDA on 2 devices, and OpenCL on the 2 other devices by doing
so:

@example
% STARPU_NCUDA=2 ./application
@end example

@node Compiling OpenCL kernels
@subsection Compiling OpenCL kernels

Source codes for OpenCL kernels can be stored in a file or in a
string. StarPU provides functions to build the program executable for
each available OpenCL device as a @code{cl_program} object. This
program executable can then be loaded within a specific queue as
explained in the next section. These are only helpers, Applications
can also fill a @code{starpu_opencl_program} array by hand for more advanced
use (e.g. different programs on the different OpenCL devices, for
relocation purpose for instance).

@menu
* starpu_opencl_load_opencl_from_file::  Compiling OpenCL source code
* starpu_opencl_load_opencl_from_string::  Compiling OpenCL source code
* starpu_opencl_unload_opencl::  Releasing OpenCL code
@end menu

@node starpu_opencl_load_opencl_from_file
@subsubsection @code{starpu_opencl_load_opencl_from_file} -- Compiling OpenCL source code
@deftypefun int starpu_opencl_load_opencl_from_file (char *@var{source_file_name}, {struct starpu_opencl_program} *@var{opencl_programs}, {const char}* @var{build_options})
TODO
@end deftypefun

@node starpu_opencl_load_opencl_from_string
@subsubsection @code{starpu_opencl_load_opencl_from_string} -- Compiling OpenCL source code
@deftypefun int starpu_opencl_load_opencl_from_string (char *@var{opencl_program_source}, {struct starpu_opencl_program} *@var{opencl_programs}, {const char}* @var{build_options})
TODO
@end deftypefun

@node starpu_opencl_unload_opencl
@subsubsection @code{starpu_opencl_unload_opencl} -- Releasing OpenCL code
@deftypefun int starpu_opencl_unload_opencl ({struct starpu_opencl_program} *@var{opencl_programs})
TODO
@end deftypefun

@node Loading OpenCL kernels
@subsection Loading OpenCL kernels

@menu
* starpu_opencl_load_kernel::   Loading a kernel
* starpu_opencl_relase_kernel::  Releasing a kernel
@end menu

@node starpu_opencl_load_kernel
@subsubsection @code{starpu_opencl_load_kernel} -- Loading a kernel
@deftypefun int starpu_opencl_load_kernel (cl_kernel *@var{kernel}, cl_command_queue *@var{queue}, {struct starpu_opencl_program} *@var{opencl_programs}, char *@var{kernel_name}, int @var{devid})
TODO
@end deftypefun

@node starpu_opencl_relase_kernel
@subsubsection @code{starpu_opencl_release_kernel} -- Releasing a kernel
@deftypefun int starpu_opencl_release_kernel (cl_kernel @var{kernel})
TODO
@end deftypefun

@node OpenCL statistics
@subsection OpenCL statistics

@menu
* starpu_opencl_collect_stats::   Collect statistics on a kernel execution
@end menu

@node starpu_opencl_collect_stats
@subsubsection @code{starpu_opencl_collect_stats} -- Collect statistics on a kernel execution
@deftypefun int starpu_opencl_collect_stats (cl_event @var{event})
After termination of the kernels, the OpenCL codelet should call this function
to pass it the even returned by @code{clEnqueueNDRangeKernel}, to let StarPU
collect statistics about the kernel execution (used cycles, consumed power).
@end deftypefun


@node Cell extensions
@section Cell extensions

nothing yet.

@node Miscellaneous helpers
@section Miscellaneous helpers

@menu
* starpu_data_cpy::                Copy a data handle into another data handle
* starpu_execute_on_each_worker::  Execute a function on a subset of workers
@end menu

@node starpu_data_cpy
@subsection @code{starpu_data_cpy} -- Copy a data handle into another data handle
@deftypefun int starpu_data_cpy (starpu_data_handle @var{dst_handle}, starpu_data_handle @var{src_handle}, int @var{asynchronous}, void (*@var{callback_func})(void*), void *@var{callback_arg})
Copy the content of the @var{src_handle} into the @var{dst_handle} handle.
The @var{asynchronous} parameter indicates whether the function should 
block or not. In the case of an asynchronous call, it is possible to
synchronize with the termination of this operation either by the means of
implicit dependencies (if enabled) or by calling
@code{starpu_task_wait_for_all()}. If @var{callback_func} is not @code{NULL},
this callback function is executed after the handle has been copied, and it is
given the @var{callback_arg} pointer as argument.
@end deftypefun



@node starpu_execute_on_each_worker
@subsection @code{starpu_execute_on_each_worker} -- Execute a function on a subset of workers
@deftypefun void starpu_execute_on_each_worker (void (*@var{func})(void *), void *@var{arg}, uint32_t @var{where})
When calling this method, the offloaded function specified by the first argument is
executed by every StarPU worker that may execute the function.
The second argument is passed to the offloaded function.
The last argument specifies on which types of processing units the function
should be executed. Similarly to the @var{where} field of the
@code{starpu_codelet} structure, it is possible to specify that the function
should be executed on every CUDA device and every CPU by passing
@code{STARPU_CPU|STARPU_CUDA}.
This function blocks until the function has been executed on every appropriate
processing units, so that it may not be called from a callback function for
instance.
@end deftypefun


@c ---------------------------------------------------------------------
@c Advanced Topics
@c ---------------------------------------------------------------------

@node Advanced Topics
@chapter Advanced Topics

@menu
* Defining a new data interface::
* Defining a new scheduling policy::
@end menu

@node Defining a new data interface
@section Defining a new data interface

@menu
* struct starpu_data_interface_ops_t::  Per-interface methods
* struct starpu_data_copy_methods::	Per-interface data transfer methods
* An example of data interface::        An example of data interface
@end menu

@c void *starpu_data_get_interface_on_node(starpu_data_handle handle, unsigned memory_node); TODO

@node struct starpu_data_interface_ops_t
@subsection @code{struct starpu_data_interface_ops_t} -- Per-interface methods
@table @asis
@item @emph{Description}:
TODO describe all the different fields
@end table

@node struct starpu_data_copy_methods
@subsection @code{struct starpu_data_copy_methods} -- Per-interface data transfer methods
@table @asis
@item @emph{Description}:
TODO describe all the different fields
@end table

@node An example of data interface
@subsection An example of data interface
@table @asis
TODO
See @code{src/datawizard/interfaces/vector_interface.c} for now.
@end table

@node Defining a new scheduling policy
@section Defining a new scheduling policy

TODO

A full example showing how to define a new scheduling policy is available in
the StarPU sources in the directory @code{examples/scheduler/}.

@menu
* struct starpu_sched_policy_s::  
* starpu_worker_set_sched_condition::
* starpu_sched_set_min_priority::       Set the minimum priority level
* starpu_sched_set_max_priority::       Set the maximum priority level
* starpu_push_local_task::		Assign a task to a worker
* Source code::                 
@end menu

@node struct starpu_sched_policy_s
@subsection @code{struct starpu_sched_policy_s} -- Scheduler methods
@table @asis
@item @emph{Description}:
This structure contains all the methods that implement a scheduling policy.  An
application may specify which scheduling strategy in the @code{sched_policy}
field of the @code{starpu_conf} structure passed to the @code{starpu_init}
function.

@item @emph{Fields}:
@table @asis
@item @code{init_sched}:
Initialize the scheduling policy.
@item @code{deinit_sched}:
Cleanup the scheduling policy.
@item @code{push_task}:
Insert a task into the scheduler.
@item @code{push_prio_task}:
Insert a priority task into the scheduler.
@item @code{push_prio_notify}:
Notify the scheduler that a task was pushed on the worker. This method is
called when a task that was explicitely assigned to a worker is scheduled. This
method therefore permits to keep the state of of the scheduler coherent even
when StarPU bypasses the scheduling strategy.
@item @code{pop_task}:
Get a task from the scheduler. The mutex associated to the worker is already
taken when this method is called. If this method is defined as @code{NULL}, the
worker will only execute tasks from its local queue. In this case, the
@code{push_task} method should use the @code{starpu_push_local_task} method to
assign tasks to the different workers.
@item @code{pop_every_task}:
Remove all available tasks from the scheduler (tasks are chained by the means
of the prev and next fields of the starpu_task structure). The mutex associated
to the worker is already taken when this method is called. 
@item @code{post_exec_hook} (optionnal):
This method is called every time a task has been executed.
@item @code{policy_name}:
Name of the policy (optionnal).
@item @code{policy_description}:
Description of the policy (optionnal).
@end table
@end table


@node starpu_worker_set_sched_condition
@subsection @code{starpu_worker_set_sched_condition} -- Specify the condition variable associated to a worker
@deftypefun void starpu_worker_set_sched_condition (int @var{workerid}, pthread_cond_t *@var{sched_cond}, pthread_mutex_t *@var{sched_mutex})
When there is no available task for a worker, StarPU blocks this worker on a
condition variable. This function specifies which condition variable (and the
associated mutex) should be used to block (and to wake up) a worker. Note that
multiple workers may use the same condition variable. For instance, in the case
of a scheduling strategy with a single task queue, the same condition variable
would be used to block and wake up all workers.
The initialization method of a scheduling strategy (@code{init_sched}) must
call this function once per worker.
@end deftypefun

@node starpu_sched_set_min_priority
@subsection @code{starpu_sched_set_min_priority}
@deftypefun void starpu_sched_set_min_priority (int @var{min_prio})
Defines the minimum priority level supported by the scheduling policy. The
default minimum priority level is the same as the default priority level which
is 0 by convention.  The application may access that value by calling the
@code{starpu_sched_get_min_priority} function. This function should only be
called from the initialization method of the scheduling policy, and should not
be used directly from the application.
@end deftypefun

@node starpu_sched_set_max_priority
@subsection @code{starpu_sched_set_max_priority}
@deftypefun void starpu_sched_set_min_priority (int @var{max_prio})
Defines the maximum priority level supported by the scheduling policy. The
default maximum priority level is 1.  The application may access that value by
calling the @code{starpu_sched_get_max_priority} function. This function should
only be called from the initialization method of the scheduling policy, and
should not be used directly from the application.
@end deftypefun

@node starpu_push_local_task
@subsection @code{starpu_push_local_task}
@deftypefun int starpu_push_local_task (int @var{workerid}, {struct starpu_task} *@var{task}, int @var{back})
The scheduling policy may put tasks directly into a worker's local queue so
that it is not always necessary to create its own queue when the local queue
is sufficient. If "back" not null, the task is put at the back of the queue
where the worker will pop tasks first. Setting "back" to 0 therefore ensures
a FIFO ordering. 
@end deftypefun




@node Source code
@subsection Source code

@cartouche
@smallexample
static struct starpu_sched_policy_s dummy_sched_policy = @{
    .init_sched = init_dummy_sched,
    .deinit_sched = deinit_dummy_sched,
    .push_task = push_task_dummy,
    .push_prio_task = NULL,
    .pop_task = pop_task_dummy,
    .post_exec_hook = NULL,
    .pop_every_task = NULL,
    .policy_name = "dummy",
    .policy_description = "dummy scheduling strategy"
@};
@end smallexample
@end cartouche


@c ---------------------------------------------------------------------
@c Appendices
@c ---------------------------------------------------------------------

@c ---------------------------------------------------------------------
@c Full source code for the 'Scaling a Vector' example
@c ---------------------------------------------------------------------

@node Full source code for the 'Scaling a Vector' example
@appendix Full source code for the 'Scaling a Vector' example

@menu
* Main application::            
* CPU Kernel::                 
* CUDA Kernel::                
* OpenCL Kernel::              
@end menu

@node Main application
@section Main application

@smallexample
@include vector_scal_c.texi
@end smallexample

@node CPU Kernel
@section CPU Kernel

@smallexample
@include vector_scal_cpu.texi
@end smallexample

@node CUDA Kernel
@section CUDA Kernel

@smallexample
@include vector_scal_cuda.texi
@end smallexample

@node OpenCL Kernel
@section OpenCL Kernel

@menu
* Invoking the kernel::         
* Source of the kernel::        
@end menu

@node Invoking the kernel
@subsection Invoking the kernel

@smallexample
@include vector_scal_opencl.texi
@end smallexample

@node Source of the kernel
@subsection Source of the kernel

@smallexample
@include vector_scal_opencl_codelet.texi
@end smallexample

@c
@c Indices.
@c

@node Function Index
@unnumbered Function Index
@printindex fn

@bye