omp_target_memcpy_rect uses Nvidia's (CUDA) and AMD's (ROCr/HSA)
features to transfer noncontiguous rectangular data efficient.
While for CUDA, the wording was already there, it was missing
for HSA. (Both landed in GCC 14, albeit HSA half a year later.)
This patch adds it also for AMD GPUs; for nvptx, I moved the
bullet point down; in the current version, the API call comes
between the stack memory and memory allocation bullets, which
seems to be misplaced.
Additionally, I added a crossref to the two API functions.
Tobias
commit 0c63c7524bd523ea82933e90689b63d80e16d67e
Author: Tobias Burnus <tbur...@baylibre.com>
Date: Mon Apr 7 09:04:53 2025 +0200
libgomp.texi: Add GCN doc for omp_target_memcpy_rect
libgomp/ChangeLog:
* libgomp.texi (omp_target_memcpy_rect_async,
omp_target_memcpy_rect): Add @ref to 'Offload-Target Specifics'.
(AMD Radeon (GCN)): Document how memcpy_rect is implemented.
(nvptx): Move item about memcpy_rect item down; use present tense.
---
libgomp/libgomp.texi | 19 ++++++++++++-------
1 file changed, 12 insertions(+), 7 deletions(-)
diff --git a/libgomp/libgomp.texi b/libgomp/libgomp.texi
index 4217c29dd37..fed9d5efb6a 100644
--- a/libgomp/libgomp.texi
+++ b/libgomp/libgomp.texi
@@ -2316,7 +2316,7 @@ the initial device.
@end multitable
@item @emph{See also}:
-@ref{omp_target_memcpy_rect_async}, @ref{omp_target_memcpy}
+@ref{omp_target_memcpy_rect_async}, @ref{omp_target_memcpy}, @ref{Offload-Target Specifics}
@item @emph{Reference}:
@uref{https://www.openmp.org, OpenMP specification v5.1}, Section 3.8.6
@@ -2391,7 +2391,7 @@ the initial device.
@end multitable
@item @emph{See also}:
-@ref{omp_target_memcpy_rect}, @ref{omp_target_memcpy_async}
+@ref{omp_target_memcpy_rect}, @ref{omp_target_memcpy_async}, @ref{Offload-Target Specifics}
@item @emph{Reference}:
@uref{https://www.openmp.org, OpenMP specification v5.1}, Section 3.8.8
@@ -6911,6 +6911,11 @@ The implementation remark:
@code{omp_thread_mem_alloc}, all use low-latency memory as first
preference, and fall back to main graphics memory when the low-latency
pool is exhausted.
+@item The OpenMP routines @code{omp_target_memcpy_rect} and
+ @code{omp_target_memcpy_rect_async} and the @code{target update}
+ directive for non-contiguous list items use the 3D memory-copy function
+ of the HSA library. Higher dimensions call this functions in a loop and
+ are therefore supported.
@item The unique identifier (UID), used with OpenMP's API UID routines, is the
value returned by the HSA runtime library for @code{HSA_AMD_AGENT_INFO_UUID}.
For GPUs, it is currently @samp{GPU-} followed by 16 lower-case hex digits,
@@ -7048,11 +7053,6 @@ The implementation remark:
devices (``host fallback'').
@item The default per-warp stack size is 128 kiB; see also @code{-msoft-stack}
in the GCC manual.
-@item The OpenMP routines @code{omp_target_memcpy_rect} and
- @code{omp_target_memcpy_rect_async} and the @code{target update}
- directive for non-contiguous list items will use the 2D and 3D
- memory-copy functions of the CUDA library. Higher dimensions will
- call those functions in a loop and are therefore supported.
@item Low-latency memory (@code{omp_low_lat_mem_space}) is supported when the
the @code{access} trait is set to @code{cgroup}, and libgomp has
been built for PTX ISA version 4.1 or higher (such as in GCC's
@@ -7070,6 +7070,11 @@ The implementation remark:
@code{omp_thread_mem_alloc}, all use low-latency memory as first
preference, and fall back to main graphics memory when the low-latency
pool is exhausted.
+@item The OpenMP routines @code{omp_target_memcpy_rect} and
+ @code{omp_target_memcpy_rect_async} and the @code{target update}
+ directive for non-contiguous list items use the 2D and 3D memory-copy
+ functions of the CUDA library. Higher dimensions call those functions
+ in a loop and are therefore supported.
@item The unique identifier (UID), used with OpenMP's API UID routines, consists
of the @samp{GPU-} prefix followed by the 16-bytes UUID as returned by
the CUDA runtime library. This UUID is output in grouped lower-case
