This series consists of support for accelerated 2D/3D memory copies for
omp_target_memcpy_rect and "target update" directives, using underlying
API-provided routines (CUDA for nVidia, or via an AMD-specific extension
for HSA).
One of the patches (by Tobias) is already on mainline, so this is a
backport, and another is a bug fix that has been submitted for mainline
but not yet reviewed. The final patch is new (the AMD support for
2D/3D copies).
Tested with offloading to both NVPTX and AMD GCN. I will push (to the
og13 branch) shortly.
Julian Brown (2):
[og13] OpenMP, NVPTX: memcpy[23]D bias correction
[og13] OpenMP: Support accelerated 2D/3D memory copies for AMD GCN
Tobias Burnus (1):
[og13] OpenMP: Call cuMemcpy2D/cuMemcpy3D for nvptx for
omp_target_memcpy_rect
include/cuda/cuda.h | 87 +++
libgomp/libgomp-plugin.h | 7 +
libgomp/libgomp.h | 2 +
libgomp/libgomp.texi | 5 +
libgomp/oacc-host.c | 2 +
libgomp/plugin/cuda-lib.def | 3 +
libgomp/plugin/plugin-gcn.c | 359 ++++++++++++
libgomp/plugin/plugin-nvptx.c | 185 ++++++
libgomp/target.c | 164 +++++-
libgomp/testsuite/libgomp.c/target-12.c | 6 +-
.../testsuite/libgomp.fortran/target-12.f90 | 6 +-
.../libgomp.fortran/target-memcpy-rect-1.f90 | 531 ++++++++++++++++++
12 files changed, 1332 insertions(+), 25 deletions(-)
create mode 100644 libgomp/testsuite/libgomp.fortran/target-memcpy-rect-1.f90
--
2.25.1
