starpu-max/doc/chapters/advanced-api.texi

@c -*-texinfo-*-

@c This file is part of the StarPU Handbook.
@c Copyright (C) 2009--2011  Universit@'e de Bordeaux 1
@c Copyright (C) 2010, 2011, 2012  Centre National de la Recherche Scientifique
@c Copyright (C) 2011, 2012 Institut National de Recherche en Informatique et Automatique
@c See the file starpu.texi for copying conditions.

@menu
* Defining a new data interface::  
* Multiformat Data Interface::  
* Task Bundles::                
* Task Lists::                  
* Using Parallel Tasks::        
* Defining a new scheduling policy::  
* Expert mode::                 
@end menu

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

@menu
* Data Interface API::  Data Interface API
* An example of data interface::        An example of data interface
@end menu

@node Data Interface API
@subsection Data Interface API

@deftp {Data Type} {struct starpu_data_interface_ops}
@anchor{struct starpu_data_interface_ops}
Per-interface data transfer methods.

@table @asis
@item @code{void (*register_data_handle)(starpu_data_handle_t handle, uint32_t home_node, void *data_interface)}
Register an existing interface into a data handle.

@item @code{starpu_ssize_t (*allocate_data_on_node)(void *data_interface, uint32_t node)}
Allocate data for the interface on a given node.

@item @code{ void (*free_data_on_node)(void *data_interface, uint32_t node)}
Free data of the interface on a given node.

@item @code{ const struct starpu_data_copy_methods *copy_methods}
ram/cuda/spu/opencl synchronous and asynchronous transfer methods.

@item @code{ void * (*handle_to_pointer)(starpu_data_handle_t handle, uint32_t node)}
Return the current pointer (if any) for the handle on the given node.

@item @code{ size_t (*get_size)(starpu_data_handle_t handle)}
Return an estimation of the size of data, for performance models.

@item @code{ uint32_t (*footprint)(starpu_data_handle_t handle)}
Return a 32bit footprint which characterizes the data size.

@item @code{ int (*compare)(void *data_interface_a, void *data_interface_b)}
Compare the data size of two interfaces.

@item @code{ void (*display)(starpu_data_handle_t handle, FILE *f)}
Dump the sizes of a handle to a file.

@item @code{ int (*convert_to_gordon)(void *data_interface, uint64_t *ptr, gordon_strideSize_t *ss)}
Convert the data size to the spu size format. If no SPUs are used, this field can be seto NULL.

@item @code{enum starpu_data_interface_id interfaceid}
An identifier that is unique to each interface.

@item @code{size_t interface_size}
The size of the interface data descriptor.

@item @code{ void (*allocate_new_data)(starpu_data_handle_t handle, void **data_interface)}
Create a new data interface of the given type based on the handle @var{handle}.
@end table
@end deftp

@deftp {Data Type} {struct starpu_data_copy_methods}
Defines the per-interface methods.
@table @asis
@item @code{int @{ram,cuda,opencl,spu@}_to_@{ram,cuda,opencl,spu@}(void *src_interface, unsigned src_node, void *dst_interface, unsigned dst_node)}
These 16 functions define how to copy data from the @var{src_interface}
interface on the @var{src_node} node to the @var{dst_interface} interface
on the @var{dst_node} node. They return 0 on success.

@item @code{int (*ram_to_cuda_async)(void *src_interface, unsigned src_node, void *dst_interface, unsigned dst_node, cudaStream_t stream)}
Define how to copy data from the @var{src_interface} interface on the
@var{src_node} node (in RAM) to the @var{dst_interface} interface on the
@var{dst_node} node (on a CUDA device), using the given @var{stream}. Return 0
on success.

@item @code{int (*cuda_to_ram_async)(void *src_interface, unsigned src_node, void *dst_interface, unsigned dst_node, cudaStream_t stream)}
Define how to copy data from the @var{src_interface} interface on the
@var{src_node} node (on a CUDA device) to the @var{dst_interface} interface on the
@var{dst_node} node (in RAM), using the given @var{stream}. Return 0
on success.

@item @code{int (*cuda_to_cuda_async)(void *src_interface, unsigned src_node, void *dst_interface, unsigned dst_node, cudaStream_t stream)}
Define how to copy data from the @var{src_interface} interface on the
@var{src_node} node (on a CUDA device) to the @var{dst_interface} interface on
the @var{dst_node} node (on another CUDA device), using the given @var{stream}.
Return 0 on success.

@item @code{int (*ram_to_opencl_async)(void *src_interface, unsigned src_node, void *dst_interface, unsigned dst_node, /* cl_event * */ void *event)}
Define how to copy data from the @var{src_interface} interface on the
@var{src_node} node (in RAM) to the @var{dst_interface} interface on the
@var{dst_node} node (on an OpenCL device), using @var{event}, a pointer to a
cl_event. Return 0 on success.

@item @code{int (*opencl_to_ram_async)(void *src_interface, unsigned src_node, void *dst_interface, unsigned dst_node, /* cl_event * */ void *event)}
Define how to copy data from the @var{src_interface} interface on the
@var{src_node} node (on an OpenCL device) to the @var{dst_interface} interface
on the @var{dst_node} node (in RAM), using the given @var{event}, a pointer to
a cl_event. Return 0 on success.

@item @code{int (*opencl_to_opencl_async)(void *src_interface, unsigned src_node, void *dst_interface, unsigned dst_node, /* cl_event * */ void *event)}
Define how to copy data from the @var{src_interface} interface on the
@var{src_node} node (on an OpenCL device) to the @var{dst_interface} interface
on the @var{dst_node} node (on another OpenCL device), using the given
@var{event}, a pointer to a cl_event. Return 0 on success.
@end table
@end deftp

@deftypefun uint32_t starpu_crc32_be_n ({void *}@var{input}, size_t @var{n}, uint32_t @var{inputcrc})
Compute the CRC of a byte buffer seeded by the inputcrc "current
state". The return value should be considered as the new "current
state" for future CRC computation. This is used for computing data size
footprint.
@end deftypefun

@deftypefun uint32_t starpu_crc32_be (uint32_t @var{input}, uint32_t @var{inputcrc})
Compute the CRC of a 32bit number seeded by the inputcrc "current
state". The return value should be considered as the new "current
state" for future CRC computation. This is used for computing data size
footprint.
@end deftypefun

@deftypefun uint32_t starpu_crc32_string ({char *}@var{str}, uint32_t @var{inputcrc})
Compute the CRC of a string seeded by the inputcrc "current state".
The return value should be considered as the new "current state" for
future CRC computation. This is used for computing data size footprint.
@end deftypefun

@node An example of data interface
@subsection An example of data interface

@deftypefun int starpu_data_interface_get_next_id ()
Returns the next available id for a newly created data interface.
@end deftypefun

TODO
See @code{src/datawizard/interfaces/vector_interface.c} for now.

@node Multiformat Data Interface
@section Multiformat Data Interface

@deftp {Data Type} {struct starpu_multiformat_data_interface_ops}
todo. The different fields are:
@table @asis
@item @code{size_t cpu_elemsize}
the size of each element on CPUs,

@item @code{size_t opencl_elemsize}
the size of each element on OpenCL devices,

@item @code{struct starpu_codelet *cpu_to_opencl_cl}
pointer to a codelet which converts from CPU to OpenCL

@item @code{struct starpu_codelet *opencl_to_cpu_cl}
pointer to a codelet which converts from OpenCL to CPU

@item @code{size_t cuda_elemsize}
the size of each element on CUDA devices,

@item @code{struct starpu_codelet *cpu_to_cuda_cl}
pointer to a codelet which converts from CPU to CUDA

@item @code{struct starpu_codelet *cuda_to_cpu_cl}
pointer to a codelet which converts from CUDA to CPU
@end table
@end deftp

@deftypefun void starpu_multiformat_data_register (starpu_data_handle_t *@var{handle}, uint32_t @var{home_node}, void *@var{ptr}, uint32_t @var{nobjects}, struct starpu_multiformat_data_interface_ops *@var{format_ops})
Register a piece of data that can be represented in different ways, depending upon
the processing unit that manipulates it. It allows the programmer, for instance, to
use an array of structures when working on a CPU, and a structure of arrays when
working on a GPU.

@var{nobjects} is the number of elements in the data. @var{format_ops} describes
the format.
@end deftypefun

@defmac STARPU_MULTIFORMAT_GET_CPU_PTR ({void *}@var{interface})
returns the local pointer to the data with CPU format.
@end defmac

@defmac STARPU_MULTIFORMAT_GET_CUDA_PTR ({void *}@var{interface})
returns the local pointer to the data with CUDA format.
@end defmac

@defmac STARPU_MULTIFORMAT_GET_OPENCL_PTR ({void *}@var{interface})
returns the local pointer to the data with OpenCL format.
@end defmac

@defmac STARPU_MULTIFORMAT_GET_NX  ({void *}@var{interface})
returns the number of elements in the data.
@end defmac


@node Task Bundles
@section Task Bundles

@deftp {Data Type} {starpu_task_bundle_t}
Opaque structure describing a list of tasks that should be scheduled
on the same worker whenever it's possible. It must be considered as a
hint given to the scheduler as there is no guarantee that they will be
executed on the same worker.
@end deftp

@deftypefun void starpu_task_bundle_create ({starpu_task_bundle_t *}@var{bundle})
Factory function creating and initializing @var{bundle}, when the call returns, memory needed is allocated and @var{bundle} is ready to use.
@end deftypefun

@deftypefun int starpu_task_bundle_insert (starpu_task_bundle_t @var{bundle}, {struct starpu_task *}@var{task})
Insert @var{task} in @var{bundle}. Until @var{task} is removed from @var{bundle} its expected length and data transfer time will be considered along those of the other tasks of @var{bundle}.
This function mustn't be called if @var{bundle} is already closed and/or @var{task} is already submitted.
@end deftypefun

@deftypefun int starpu_task_bundle_remove (starpu_task_bundle_t @var{bundle}, {struct starpu_task *}@var{task})
Remove @var{task} from @var{bundle}.
Of course @var{task} must have been previously inserted @var{bundle}.
This function mustn't be called if @var{bundle} is already closed and/or @var{task} is already submitted. Doing so would result in undefined behaviour.
@end deftypefun

@deftypefun void starpu_task_bundle_close (starpu_task_bundle_t @var{bundle})
Inform the runtime that the user won't modify @var{bundle} anymore, it means no more inserting or removing task. Thus the runtime can destroy it when possible.
@end deftypefun


@node Task Lists
@section Task Lists

@deftp {Data Type} {struct starpu_task_list}
Stores a double-chained list of tasks
@end deftp

@deftypefun void starpu_task_list_init ({struct starpu_task_list *}@var{list})
Initialize a list structure
@end deftypefun

@deftypefun void starpu_task_list_push_front ({struct starpu_task_list *}@var{list}, {struct starpu_task *}@var{task})
Push a task at the front of a list
@end deftypefun

@deftypefun void starpu_task_list_push_back ({struct starpu_task_list *}@var{list}, {struct starpu_task *}@var{task})
Push a task at the back of a list
@end deftypefun

@deftypefun {struct starpu_task *} starpu_task_list_front ({struct starpu_task_list *}@var{list})
Get the front of the list (without removing it)
@end deftypefun

@deftypefun {struct starpu_task *} starpu_task_list_back ({struct starpu_task_list *}@var{list})
Get the back of the list (without removing it)
@end deftypefun

@deftypefun int starpu_task_list_empty ({struct starpu_task_list *}@var{list})
Test if a list is empty
@end deftypefun

@deftypefun void starpu_task_list_erase ({struct starpu_task_list *}@var{list}, {struct starpu_task *}@var{task})
Remove an element from the list
@end deftypefun

@deftypefun {struct starpu_task *} starpu_task_list_pop_front ({struct starpu_task_list *}@var{list})
Remove the element at the front of the list
@end deftypefun

@deftypefun {struct starpu_task *} starpu_task_list_pop_back ({struct starpu_task_list *}@var{list})
Remove the element at the back of the list
@end deftypefun

@deftypefun {struct starpu_task *} starpu_task_list_begin ({struct starpu_task_list *}@var{list})
Get the first task of the list.
@end deftypefun

@deftypefun {struct starpu_task *} starpu_task_list_end ({struct starpu_task_list *}@var{list})
Get the end of the list.
@end deftypefun

@deftypefun {struct starpu_task *} starpu_task_list_next ({struct starpu_task *}@var{task})
Get the next task of the list. This is not erase-safe.
@end deftypefun

@node Using Parallel Tasks
@section Using Parallel Tasks

These are used by parallel tasks:

@deftypefun int starpu_combined_worker_get_size (void)
Return the size of the current combined worker, i.e. the total number of cpus
running the same task in the case of SPMD parallel tasks, or the total number
of threads that the task is allowed to start in the case of FORKJOIN parallel
tasks.
@end deftypefun

@deftypefun int starpu_combined_worker_get_rank (void)
Return the rank of the current thread within the combined worker. Can only be
used in FORKJOIN parallel tasks, to know which part of the task to work on.
@end deftypefun

Most of these are used for schedulers which support parallel tasks.

@deftypefun unsigned starpu_combined_worker_get_count (void)
Return the number of different combined workers.
@end deftypefun

@deftypefun int starpu_combined_worker_get_id (void)
Return the identifier of the current combined worker.
@end deftypefun

@deftypefun int starpu_combined_worker_assign_workerid (int @var{nworkers}, int @var{workerid_array}[])
Register a new combined worker and get its identifier
@end deftypefun

@deftypefun int starpu_combined_worker_get_description (int @var{workerid}, {int *}@var{worker_size}, {int **}@var{combined_workerid})
Get the description of a combined worker
@end deftypefun

@deftypefun int starpu_combined_worker_can_execute_task (unsigned @var{workerid}, {struct starpu_task *}@var{task}, unsigned @var{nimpl})
Variant of starpu_worker_can_execute_task compatible with combined workers
@end deftypefun


@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
* Scheduling Policy API:: Scheduling Policy API
* Source code::
@end menu

@node Scheduling Policy API
@subsection Scheduling Policy API

While StarPU comes with a variety of scheduling policies (@pxref{Task
scheduling policy}), it may sometimes be desirable to implement custom
policies to address specific problems.  The API described below allows
users to write their own scheduling policy.

@deftp {Data Type} {struct starpu_machine_topology}
@table @asis
@item @code{unsigned nworkers}
Total number of workers.

@item @code{unsigned ncombinedworkers}
Total number of combined workers.

@item @code{hwloc_topology_t hwtopology}
Topology as detected by hwloc.

To maintain ABI compatibility when hwloc is not available, the field
is replaced with @code{void *dummy}

@item @code{unsigned nhwcpus}
Total number of CPUs, as detected by the topology code. May be different from
the actual number of CPU workers.

@item @code{unsigned nhwcudagpus}
Total number of CUDA devices, as detected. May be different from the actual
number of CUDA workers.

@item @code{unsigned nhwopenclgpus}
Total number of OpenCL devices, as detected. May be different from the actual
number of CUDA workers.

@item @code{unsigned ncpus}
Actual number of CPU workers used by StarPU.

@item @code{unsigned ncudagpus}
Actual number of CUDA workers used by StarPU.

@item @code{unsigned nopenclgpus}
Actual number of OpenCL workers used by StarPU.

@item @code{unsigned ngordon_spus}
Actual number of Gordon workers used by StarPU.

@item @code{unsigned workers_bindid[STARPU_NMAXWORKERS]}
Indicates the successive cpu identifier that should be used to bind the
workers. It is either filled according to the user's explicit
parameters (from starpu_conf) or according to the STARPU_WORKERS_CPUID env.
variable. Otherwise, a round-robin policy is used to distributed the workers
over the cpus.

@item @code{unsigned workers_cuda_gpuid[STARPU_NMAXWORKERS]}
Indicates the successive cpu identifier that should be used by the CUDA
driver.  It is either filled according to the user's explicit parameters (from
starpu_conf) or according to the STARPU_WORKERS_CUDAID env. variable. Otherwise,
they are taken in ID order.

@item @code{unsigned workers_opencl_gpuid[STARPU_NMAXWORKERS]}
Indicates the successive cpu identifier that should be used by the OpenCL
driver.  It is either filled according to the user's explicit parameters (from
starpu_conf) or according to the STARPU_WORKERS_OPENCLID env. variable. Otherwise,
they are taken in ID order.


@end table
@end deftp

@deftp {Data Type} {struct starpu_sched_policy}
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. The different fields are:

@table @asis
@item @code{void (*init_sched)(struct starpu_machine_topology *, struct starpu_sched_policy *)}
Initialize the scheduling policy.

@item @code{void (*deinit_sched)(struct starpu_machine_topology *, struct starpu_sched_policy *)}
Cleanup the scheduling policy.

@item @code{int (*push_task)(struct starpu_task *)}
Insert a task into the scheduler.

@item @code{void (*push_task_notify)(struct starpu_task *, int workerid)}
Notify the scheduler that a task was pushed on a given worker. This method is
called when a task that was explicitely assigned to a worker becomes ready and
is about to be executed by the worker. This method therefore permits to keep
the state of of the scheduler coherent even when StarPU bypasses the scheduling
strategy.

@item @code{struct starpu_task *(*pop_task)(void)} (optional)
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{struct starpu_task *(*pop_every_task)(void)}
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. This is currently
only used by the Gordon driver.

@item @code{void (*pre_exec_hook)(struct starpu_task *)} (optional)
This method is called every time a task is starting.

@item @code{void (*post_exec_hook)(struct starpu_task *)} (optional)
This method is called every time a task has been executed.

@item @code{const char *policy_name} (optional)
Name of the policy.

@item @code{const char *policy_description} (optional)
Description of the policy.
@end table
@end deftp

@deftypefun void starpu_worker_set_sched_condition (int @var{workerid}, pthread_cond_t *@var{sched_cond}, pthread_mutex_t *@var{sched_mutex})
This function specifies the condition variable associated to a worker
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

@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

@deftypefun void starpu_sched_set_max_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

@deftypefun int starpu_sched_get_min_priority (void)
Returns the current minimum priority level supported by the
scheduling policy
@end deftypefun

@deftypefun int starpu_sched_get_max_priority (void)
Returns the current maximum priority level supported by the
scheduling policy
@end deftypefun

@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 @var{back} not null, @var{task} is put at the back of the queue
where the worker will pop tasks first. Setting @var{back} to 0 therefore ensures
a FIFO ordering.
@end deftypefun

@deftypefun int starpu_worker_can_execute_task (unsigned @var{workerid}, {struct starpu_task *}@var{task}, unsigned {nimpl})
Check if the worker specified by workerid can execute the codelet. Schedulers need to call it before assigning a task to a worker, otherwise the task may fail to execute.
@end deftypefun

@deftypefun double starpu_timing_now (void)
Return the current date in µs
@end deftypefun

@deftypefun double starpu_task_expected_length ({struct starpu_task *}@var{task}, {enum starpu_perf_archtype} @var{arch}, unsigned @var{nimpl})
Returns expected task duration in µs
@end deftypefun

@deftypefun double starpu_worker_get_relative_speedup ({enum starpu_perf_archtype} @var{perf_archtype})
Returns an estimated speedup factor relative to CPU speed
@end deftypefun

@deftypefun double starpu_task_expected_data_transfer_time (uint32_t @var{memory_node}, {struct starpu_task *}@var{task})
Returns expected data transfer time in µs
@end deftypefun

@deftypefun double starpu_data_expected_transfer_time (starpu_data_handle_t @var{handle}, unsigned @var{memory_node}, {enum starpu_access_mode} @var{mode})
Predict the transfer time (in µs) to move a handle to a memory node
@end deftypefun

@deftypefun double starpu_task_expected_power ({struct starpu_task *}@var{task}, {enum starpu_perf_archtype} @var{arch}, unsigned @var{nimpl})
Returns expected power consumption in J
@end deftypefun

@deftypefun double starpu_task_expected_conversion_time ({struct starpu_task *}@var{task}, {enum starpu_perf_archtype} @var{arch}, unsigned {nimpl})
Returns expected conversion time in ms (multiformat interface only)
@end deftypefun

@node Source code
@subsection Source code

@cartouche
@smallexample
static struct starpu_sched_policy 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

@node Expert mode
@section Expert mode

@deftypefun void starpu_wake_all_blocked_workers (void)
Wake all the workers, so they can inspect data requests and task submissions
again.
@end deftypefun

@deftypefun int starpu_progression_hook_register (unsigned (*@var{func})(void *arg), void *@var{arg})
Register a progression hook, to be called when workers are idle.
@end deftypefun

@deftypefun void starpu_progression_hook_deregister (int @var{hook_id})
Unregister a given progression hook.
@end deftypefun