libgomp.texi: Document some of the device-memory routines

libgomp/ChangeLog:

	* libgomp.texi (Device Memory Routines): New.
This commit is contained in:
Tobias Burnus 2023-10-06 12:38:31 +02:00
parent e77428a9a3
commit e0786ba689

View file

@ -514,7 +514,7 @@ specification in version 5.2.
* Tasking Routines::
@c * Resource Relinquishing Routines::
* Device Information Routines::
@c * Device Memory Routines::
* Device Memory Routines::
* Lock Routines::
* Timing Routines::
* Event Routine::
@ -1658,25 +1658,298 @@ For OpenMP 5.1, this must be equal to the value returned by the
@c @node Device Memory Routines
@c @section Device Memory Routines
@c
@c Routines related to memory allocation and managing corresponding
@c pointers on devices. They have C linkage and do not throw exceptions.
@c
@c @menu
@c * omp_target_alloc:: <fixme>
@c * omp_target_free:: <fixme>
@c * omp_target_is_present:: <fixme>
@node Device Memory Routines
@section Device Memory Routines
Routines related to memory allocation and managing corresponding
pointers on devices. They have C linkage and do not throw exceptions.
@menu
* omp_target_alloc:: Allocate device memory
* omp_target_free:: Free device memory
* omp_target_is_present:: Check whether storage is mapped
@c * omp_target_is_accessible:: <fixme>
@c * omp_target_memcpy:: <fixme>
@c * omp_target_memcpy_rect:: <fixme>
@c * omp_target_memcpy_async:: <fixme>
@c * omp_target_memcpy_rect_async:: <fixme>
@c * omp_target_associate_ptr:: <fixme>
@c * omp_target_disassociate_ptr:: <fixme>
@c * omp_get_mapped_ptr:: <fixme>
@c @end menu
@c * omp_target_memset:: <fixme>/TR12
@c * omp_target_memset_async:: <fixme>/TR12
* omp_target_associate_ptr:: Associate a device pointer with a host pointer
* omp_target_disassociate_ptr:: Remove device--host pointer association
* omp_get_mapped_ptr:: Return device pointer to a host pointer
@end menu
@node omp_target_alloc
@subsection @code{omp_target_alloc} -- Allocate device memory
@table @asis
@item @emph{Description}:
This routine allocates @var{size} bytes of memory in the device environment
associated with the device number @var{device_num}. If successful, a device
pointer is returned, otherwise a null pointer.
In GCC, when the device is the host or the device shares memory with the host,
the memory is allocated on the host; in that case, when @var{size} is zero,
either NULL or a unique pointer value that can later be successfully passed to
@code{omp_target_free} is returned. When the allocation is not performed on
the host, a null pointer is returned when @var{size} is zero; in that case,
additionally a diagnostic might be printed to standard error (stderr).
Running this routine in a @code{target} region except on the initial device
is not supported.
@item @emph{C/C++}
@multitable @columnfractions .20 .80
@item @emph{Prototype}: @tab @code{void *omp_target_alloc(size_t size, int device_num)}
@end multitable
@item @emph{Fortran}:
@multitable @columnfractions .20 .80
@item @emph{Interface}: @tab @code{type(c_ptr) function omp_target_alloc(size, device_num) bind(C)}
@item @tab @code{use, intrinsic :: iso_c_binding, only: c_ptr, c_int, c_size_t}
@item @tab @code{integer(c_size_t), value :: size}
@item @tab @code{integer(c_int), value :: device_num}
@end multitable
@item @emph{See also}:
@ref{omp_target_free}, @ref{omp_target_associate_ptr}
@item @emph{Reference}:
@uref{https://www.openmp.org, OpenMP specification v5.1}, Section 18.8.1
@end table
@node omp_target_free
@subsection @code{omp_target_free} -- Free device memory
@table @asis
@item @emph{Description}:
This routine frees memory allocated by the @code{omp_target_alloc} routine.
The @var{device_ptr} argument must be either a null pointer or a device pointer
returned by @code{omp_target_alloc} for the specified @code{device_num}. The
device number @var{device_num} must be a conforming device number.
Running this routine in a @code{target} region except on the initial device
is not supported.
@item @emph{C/C++}
@multitable @columnfractions .20 .80
@item @emph{Prototype}: @tab @code{void omp_target_free(void *device_ptr, int device_num)}
@end multitable
@item @emph{Fortran}:
@multitable @columnfractions .20 .80
@item @emph{Interface}: @tab @code{subroutine omp_target_free(device_ptr, device_num) bind(C)}
@item @tab @code{use, intrinsic :: iso_c_binding, only: c_ptr, c_int}
@item @tab @code{type(c_ptr), value :: device_ptr}
@item @tab @code{integer(c_int), value :: device_num}
@end multitable
@item @emph{See also}:
@ref{omp_target_alloc}, @ref{omp_target_disassociate_ptr}
@item @emph{Reference}:
@uref{https://www.openmp.org, OpenMP specification v5.1}, Section 18.8.2
@end table
@node omp_target_is_present
@subsection @code{omp_target_is_present} -- Check whether storage is mapped
@table @asis
@item @emph{Description}:
This routine tests whether storage, identified by the host pointer @var{ptr}
is mapped to the device specified by @var{device_num}. If so, it returns
@emph{true} and otherwise @emph{false}.
In GCC, this includes self mapping such that @code{omp_target_is_present}
returns @emph{true} when @var{device_num} specifies the host or when the host
and the device share memory. If @var{ptr} is a null pointer, @var{true} is
returned and if @var{device_num} is an invalid device number, @var{false} is
returned.
If those conditions do not apply, @emph{true} is returned if the association has
been established by an explicit or implicit @code{map} clause, the
@code{declare target} directive or a call to the @code{omp_target_associate_ptr}
routine.
Running this routine in a @code{target} region except on the initial device
is not supported.
@item @emph{C/C++}
@multitable @columnfractions .20 .80
@item @emph{Prototype}: @tab @code{int omp_target_is_present(const void *ptr,}
@item @tab @code{ int device_num)}
@end multitable
@item @emph{Fortran}:
@multitable @columnfractions .20 .80
@item @emph{Interface}: @tab @code{integer(c_int) function omp_target_is_present(ptr, &}
@item @tab @code{ device_num) bind(C)}
@item @tab @code{use, intrinsic :: iso_c_binding, only: c_ptr, c_int}
@item @tab @code{type(c_ptr), value :: ptr}
@item @tab @code{integer(c_int), value :: device_num}
@end multitable
@item @emph{See also}:
@ref{omp_target_associate_ptr}
@item @emph{Reference}:
@uref{https://www.openmp.org, OpenMP specification v5.1}, Section 18.8.3
@end table
@node omp_target_associate_ptr
@subsection @code{omp_target_associate_ptr} -- Associate a device pointer with a host pointer
@table @asis
@item @emph{Description}:
This routine associates storage on the host with storage on a device identified
by @var{device_num}. The device pointer is usually obtained by calling
@code{omp_target_alloc} or by other means (but not by using the @code{map}
clauses or the @code{declare target} directive). The host pointer should point
to memory that has a storage size of at least @var{size}.
The @var{device_offset} parameter specifies the offset into @var{device_ptr}
that is used as the base address for the device side of the mapping; the
storage size should be at least @var{device_offset} plus @var{size}.
After the association, the host pointer can be used in a @code{map} clause and
in the @code{to} and @code{from} clauses of the @code{target update} directive
to transfer data between the associated pointers. The reference count of such
associated storage is infinite. The association can be removed by calling
@code{omp_target_disassociate_ptr} which should be done before the lifetime
of either either storage ends.
The routine returns nonzero (@code{EINVAL}) when the @var{device_num} invalid,
for when the initial device or the associated device shares memory with the
host. @code{omp_target_associate_ptr} returns zero if @var{host_ptr} points
into already associated storage that is fully inside of a previously associated
memory. Otherwise, if the association was successful zero is returned; if none
of the cases above apply, nonzero (@code{EINVAL}) is returned.
The @code{omp_target_is_present} routine can be used to test whether
associated storage for a device pointer exists.
Running this routine in a @code{target} region except on the initial device
is not supported.
@item @emph{C/C++}
@multitable @columnfractions .20 .80
@item @emph{Prototype}: @tab @code{int omp_target_associate_ptr(const void *host_ptr,}
@item @tab @code{ const void *device_ptr,}
@item @tab @code{ size_t size,}
@item @tab @code{ size_t device_offset,}
@item @tab @code{ int device_num)}
@end multitable
@item @emph{Fortran}:
@multitable @columnfractions .20 .80
@item @emph{Interface}: @tab @code{integer(c_int) function omp_target_associate_ptr(host_ptr, &}
@item @tab @code{ device_ptr, size, device_offset, device_num) bind(C)}
@item @tab @code{use, intrinsic :: iso_c_binding, only: c_ptr, c_int, c_size_t}
@item @tab @code{type(c_ptr), value :: host_ptr, device_ptr}
@item @tab @code{integer(c_size_t), value :: size, device_offset}
@item @tab @code{integer(c_int), value :: device_num}
@end multitable
@item @emph{See also}:
@ref{omp_target_disassociate_ptr}, @ref{omp_target_is_present},
@ref{omp_target_alloc}
@item @emph{Reference}:
@uref{https://www.openmp.org, OpenMP specification v5.1}, Section 18.8.9
@end table
@node omp_target_disassociate_ptr
@subsection @code{omp_target_disassociate_ptr} -- Remove device--host pointer association
@table @asis
@item @emph{Description}:
This routine removes the storage association established by calling
@code{omp_target_associate_ptr} and sets the reference count to zero,
even if @code{omp_target_associate_ptr} was invoked multiple times for
for host pointer @code{ptr}. If applicable, the device memory needs
to be freed by the user.
If an associated device storage location for the @var{device_num} was
found and has infinite reference count, the association is removed and
zero is returned. In all other cases, nonzero (@code{EINVAL}) is returned
and no other action is taken.
Note that passing a host pointer where the association to the device pointer
was established with the @code{declare target} directive yields undefined
behavior.
Running this routine in a @code{target} region except on the initial device
is not supported.
@item @emph{C/C++}
@multitable @columnfractions .20 .80
@item @emph{Prototype}: @tab @code{int omp_target_disassociate_ptr(const void *ptr,}
@item @tab @code{ int device_num)}
@end multitable
@item @emph{Fortran}:
@multitable @columnfractions .20 .80
@item @emph{Interface}: @tab @code{integer(c_int) function omp_target_disassociate_ptr(ptr, &}
@item @tab @code{ device_num) bind(C)}
@item @tab @code{use, intrinsic :: iso_c_binding, only: c_ptr, c_int}
@item @tab @code{type(c_ptr), value :: ptr}
@item @tab @code{integer(c_int), value :: device_num}
@end multitable
@item @emph{See also}:
@ref{omp_target_associate_ptr}
@item @emph{Reference}:
@uref{https://www.openmp.org, OpenMP specification v5.1}, Section 18.8.10
@end table
@node omp_get_mapped_ptr
@subsection @code{omp_get_mapped_ptr} -- Return device pointer to a host pointer
@table @asis
@item @emph{Description}:
If the device number is refers to the initial device or to a device with
memory accessible from the host (shared memory), the @code{omp_get_mapped_ptr}
routines returnes the value of the passed @var{ptr}. Otherwise, if associated
storage to the passed host pointer @var{ptr} exists on device associated with
@var{device_num}, it returns that pointer. In all other cases and in cases of
an error, a null pointer is returned.
The association of storage location is established either via an explicit or
implicit @code{map} clause, the @code{declare target} directive or the
@code{omp_target_associate_ptr} routine.
Running this routine in a @code{target} region except on the initial device
is not supported.
@item @emph{C/C++}
@multitable @columnfractions .20 .80
@item @emph{Prototype}: @tab @code{void *omp_get_mapped_ptr(const void *ptr, int device_num);}
@end multitable
@item @emph{Fortran}:
@multitable @columnfractions .20 .80
@item @emph{Interface}: @tab @code{type(c_ptr) function omp_get_mapped_ptr(ptr, device_num) bind(C)}
@item @tab @code{use, intrinsic :: iso_c_binding, only: c_ptr, c_int}
@item @tab @code{type(c_ptr), value :: ptr}
@item @tab @code{integer(c_int), value :: device_num}
@end multitable
@item @emph{See also}:
@ref{omp_target_associate_ptr}
@item @emph{Reference}:
@uref{https://www.openmp.org, OpenMP specification v5.1}, Section 18.8.11
@end table
@node Lock Routines
@section Lock Routines