A few high-end nvptx devices support the attribute
CU_DEVICE_ATTRIBUTE_PAGEABLE_MEMORY_ACCESS; for those, unified shared
memory is supported in hardware. This patch enables support for those -
if all installed nvptx devices have this feature (as the capabilities
are per device type).
This exposes a bug in gomp_copy_back_icvs as it did before use
omp_get_mapped_ptr to find mapped variables, but that returns
the unchanged pointer in cased of shared memory. But in this case,
we have a few actually mapped pointers - like the ICV variables.
Additionally, there was a mismatch with regards to '-1' for the
device number as gomp_copy_back_icvs and omp_get_mapped_ptr count
differently. Hence, do the lookup manually.
include/ChangeLog:
* cuda/cuda.h (CU_DEVICE_ATTRIBUTE_PAGEABLE_MEMORY_ACCESS): Add.
libgomp/ChangeLog:
* libgomp.texi (nvptx): Update USM description.
* plugin/plugin-nvptx.c (GOMP_OFFLOAD_get_num_devices):
Claim support when requesting USM and all devices support
CU_DEVICE_ATTRIBUTE_PAGEABLE_MEMORY_ACCESS.
* target.c (gomp_copy_back_icvs): Fix device ptr lookup.
(gomp_target_init): Set GOMP_OFFLOAD_CAP_SHARED_MEM is the
devices supports USM.
Recently -march=gfx{90c,1036,1103} support has been added, but corresponding
changes weren't done in the testsuite.
The following patch adds that.
Tested on x86_64-linux (with fiji and gfx1103 devices; had to use
OMP_DEFAULT_DEVICE=1 there, fiji doesn't really work due to LLVM dropping
support, but we still list those as offloading devices).
2024-05-02 Jakub Jelinek <jakub@redhat.com>
* testsuite/libgomp.c/declare-variant-4.h (gfx90c, gfx1036, gfx1103):
New functions.
(f): Add #pragma omp declare variant directives for those.
* testsuite/libgomp.c/declare-variant-4-gfx90c.c: New test.
* testsuite/libgomp.c/declare-variant-4-gfx1036.c: New test.
* testsuite/libgomp.c/declare-variant-4-gfx1103.c: New test.
Add support for gfx90c GCN5 APU integrated graphics devices.
The LLVM AMDGPU documentation does not list those devices as supported
by rocm-amdhsa, but it passes most libgomp offloading tests.
Although they are constrainted compared to dGPUs, they might be
interesting for learning, experimentation, and testing.
gcc/ChangeLog:
* config.gcc: Add gfx90c.
* config/gcn/gcn-hsa.h (NO_SRAM_ECC): Likewise.
* config/gcn/gcn-opts.h (enum processor_type): Likewise.
(TARGET_GFX90c): New macro.
* config/gcn/gcn.cc (gcn_option_override): Handle gfx90c.
(gcn_omp_device_kind_arch_isa): Likewise.
(output_file_start): Likewise.
* config/gcn/gcn.h: Add gfx90c.
* config/gcn/gcn.opt: Likewise.
* config/gcn/mkoffload.cc (EF_AMDGPU_MACH_AMDGCN_GFX90c): New macro.
(get_arch): Handle gfx90c.
(main): Handle EF_AMDGPU_MACH_AMDGCN_GFX90c
* config/gcn/t-omp-device: Add gfx90c.
* doc/install.texi: Likewise.
* doc/invoke.texi: Likewise.
libgomp/ChangeLog:
* plugin/plugin-gcn.c (isa_hsa_name): Handle EF_AMDGPU_MACH_AMDGCN_GFX90c.
(isa_code): Handle gfx90c.
(max_isa_vgprs): Handle EF_AMDGPU_MACH_AMDGCN_GFX90c.
Signed-off-by: Frederik Harwath <frederik@harwath.name>
This patch adjusts the implementation of acc_map_data/acc_unmap_data API library
routines to more fit the description in the OpenACC 2.7 specification.
Instead of using REFCOUNT_INFINITY, we now define a REFCOUNT_ACC_MAP_DATA
special value to mark acc_map_data-created mappings. Adjustment around
mapping related code to respect OpenACC semantics are also added.
libgomp/ChangeLog:
* libgomp.h (REFCOUNT_ACC_MAP_DATA): Define as (REFCOUNT_SPECIAL | 2).
* oacc-mem.c (acc_map_data): Adjust to use REFCOUNT_ACC_MAP_DATA,
initialize dynamic_refcount as 1.
(acc_unmap_data): Adjust to use REFCOUNT_ACC_MAP_DATA,
(goacc_map_var_existing): Add REFCOUNT_ACC_MAP_DATA case.
(goacc_exit_datum_1): Add REFCOUNT_ACC_MAP_DATA case, respect
REFCOUNT_ACC_MAP_DATA when decrementing/finalizing. Force lowest
dynamic_refcount to be 1 for REFCOUNT_ACC_MAP_DATA.
(goacc_enter_data_internal): Add REFCOUNT_ACC_MAP_DATA case.
* target.c (gomp_increment_refcount): Return early for
REFCOUNT_ACC_MAP_DATA case.
(gomp_decrement_refcount): Likewise.
* testsuite/libgomp.oacc-c-c++-common/lib-96.c: New testcase.
* testsuite/libgomp.oacc-c-c++-common/unmap-infinity-1.c: Adjust
testcase error output scan test.
Currently, we silently disable libgomp GCN and nvptx plugins/devices in
presence of certain error conditions during device probing, thus typically
silently resorting to host-fallback execution. Make such errors fatal, similar
as for any other device access later on, so that we early and reliably notice
when things go wrong. (Keep just two cases non-fatal: (a) libgomp GCN or nvptx
plugins are available but 'libhsa-runtime64.so.1' or 'libcuda.so.1' are not,
and (b) those are available, but the corresponding devices are not.)
This resolves the issue that we've got execution test cases unexpectedly
PASSing, despite:
libgomp: GCN fatal error: Run-time could not be initialized
Runtime message: HSA_STATUS_ERROR_OUT_OF_RESOURCES: The runtime failed to allocate the necessary resources. This error may also occur when the core runtime library needs to spawn threads or create internal OS-specific events.
..., and therefore they were not offloaded to the GCN device, but ran in
host-fallback execution mode. What happend in that scenario is that in
'init_hsa_context' during the initial 'GOMP_OFFLOAD_get_num_devices' we ran
into 'HSA_STATUS_ERROR_OUT_OF_RESOURCES', but it wasn't fatal, but just
silently disabled the libgomp plugin/device.
Especially "entertaining" were cases where such unintended host-fallback
execution happened during effective-target checks like
'offload_device_available' (host-fallback execution there meaning: no offload
device available), but actual test cases then were running with an offload
device available, and therefore mis-configured.
include/
* cuda/cuda.h (CUresult): Add 'CUDA_ERROR_NO_DEVICE'.
libgomp/
* plugin/plugin-gcn.c (init_hsa_context): Add and handle
'bool probe' parameter. Adjust all users; errors during device
probing are fatal.
* plugin/plugin-nvptx.c (nvptx_get_num_devices): Aside from
'CUDA_ERROR_NO_DEVICE', errors during device probing are fatal.
With commit 7520a4992c
"nvptx: In mkoffload.cc, call diagnostic_color_init + gcc_init_libintl",
we regressed:
[-PASS:-]{+FAIL:+} libgomp.c/reverse-offload-sm30.c at line 15 (test for warnings, line )
[-PASS:-]{+FAIL:+} libgomp.c/reverse-offload-sm30.c (test for excess errors)
libgomp/
* testsuite/libgomp.c/reverse-offload-sm30.c: Set 'GCC_COLORS' to the empty string.
While ctors/dtors don't return anything (undeclared void or this pointer
on arm) and copy assignment operators normally return a reference to *this,
it isn't invalid to return uselessly some class object which might need
destructing, but the OpenMP clause handling code wasn't expecting that.
The following patch fixes that.
2024-04-05 Jakub Jelinek <jakub@redhat.com>
PR c++/114572
* cp-gimplify.cc (cxx_omp_clause_apply_fn): Call build_cplus_new
on build_call_a result if it has class type.
* testsuite/libgomp.c++/pr114572.C: New test.
Like in r12-7519-g027e30414492d50feb2854aff38227b14300dc4b, I've done
git grep -v 'long long\|optab optab\|template template\|double double' | grep ' \([a-zA-Z]\+\) \1 '
This is just part of the changes, mostly for non-gcc directories.
I'll try to get to the rest soon. Obviously, the above command also
finds cases which are correct as is and shouldn't be changed, so one
needs to manually inspect everything.
I'd hope most of it is pretty obvious, but the config/ and libstdc++-v3/
hunks include a tweak in a license wording, though other copies of the
similar license have the wording right.
2024-04-02 Jakub Jelinek <jakub@redhat.com>
* Makefile.tpl: Fix duplicated words; returns returns ->
returns.
config/
* lcmessage.m4: Fix duplicated words; can can -> can,
package package -> package.
libdecnumber/
* decCommon.c (decFinalize): Fix duplicated words in
comment; the the -> the.
libgcc/
* unwind-dw2-fde.c (struct fde_accumulator): Fix duplicated
words in comment; is is -> is.
libgfortran/
* configure.host: Fix duplicated words; the the -> the.
libgm2/
* configure.host: Fix duplicated words; the the -> the.
libgomp/
* libgomp.texi (OpenMP 5.2): Fix duplicated words; with with ->
with.
(omp_target_associate_ptr): Fix duplicated words; either either ->
either.
(omp_init_allocator): Fix duplicated words; be be -> be.
(omp_realloc): Fix duplicated words; is is -> is.
(OMP_ALLOCATOR): Fix duplicated words; other other -> other.
* priority_queue.h (priority_queue_multi_p): Fix duplicated words;
to to -> to.
libiberty/
* regex.c (byte_re_match_2_internal): Fix duplicated words in comment;
next next -> next.
* dyn-string.c (dyn_string_init): Fix duplicated words in comment;
of of -> of.
libitm/
* beginend.cc (GTM::gtm_thread::begin_transaction): Fix duplicated
words in comment; not not -> not to.
libobjc/
* init.c (duplicate_classes): Fix duplicated words in comment; in in
-> in.
* sendmsg.c (__objc_prepare_dtable_for_class): Fix duplicated words
in comment; the the -> the.
* encoding.c (objc_layout_structure): Likewise.
libstdc++-v3/
* acinclude.m4: Fix duplicated words; file file -> file can.
* configure.host: Fix duplicated words; the the -> the.
libvtv/
* vtv_rts.cc (vtv_fail): Fix duplicated words; to to -> to.
* vtv_fail.cc (vtv_fail): Likewise.
A splay-tree was previously used to lookup equivalent target addresses
for a given host address on offload targets. However, as splay-trees can
modify their structure on lookup, they are not suitable for concurrent
access from separate teams/threads without some form of locking. This
patch changes the lookup data structure to a hashtab instead, which does
not have these issues.
The call to build_indirect_map to initialize the data structure is now
called from just the first thread of the first team to avoid redundant
calls to this function.
2024-03-22 Kwok Cheung Yeung <kcyeung@baylibre.com>
libgomp/
* config/accel/target-indirect.c: Include string.h and hashtab.h.
Remove include of splay-tree.h. Update comments.
(splay_tree_prefix, splay_tree_c): Delete.
(struct indirect_map_t): New.
(hash_entry_type, htab_alloc, htab_free, htab_hash, htab_eq): New.
(GOMP_INDIRECT_ADD_MAP): Remove volatile qualifier.
(USE_SPLAY_TREE_LOOKUP): Rename to...
(USE_HASHTAB_LOOKUP): ..this.
(indirect_map, indirect_array): Delete.
(indirect_htab): New.
(build_indirect_map): Remove locking. Build indirect map using
hashtab.
(GOMP_target_map_indirect_ptr): Use indirect_htab to lookup target
address.
(GOMP_target_map_indirect_ptr): Remove volatile qualifier.
* config/gcn/team.c (gomp_gcn_enter_kernel): Call build_indirect_map
from first thread of first team only.
* config/nvptx/team.c (gomp_nvptx_main): Likewise.
* testsuite/libgomp.c-c++-common/declare-target-indirect-2.c (main):
Add missing break statements.
* testsuite/libgomp.fortran/declare-target-indirect-2.f90: Remove
xfail.
Our dear friend '-Wuninitialized' reported:
[...]/libgomp.oacc-fortran/acc-memcpy.f90:18:27:
18 | char(j) = int (j, int8)
| ^
Warning: ‘j’ may be used uninitialized [-Wmaybe-uninitialized]
[...]/libgomp.oacc-fortran/acc-memcpy.f90:14:20:
14 | integer(int8) :: j
| ^
note: ‘j’ was declared here
..., but actually there were other issues.
libgomp/
* testsuite/libgomp.oacc-fortran/acc-memcpy.f90: Fix 'char'
initialization, copy, check.
Dummy procedures look similar to variables but aren't - neither in Fortran
nor in OpenMP. As the middle end sees PARM_DECLs, mark them as predetermined
firstprivate for mapping (as already done in gfc_omp_predetermined_sharing).
This does not address the isses related to procedure pointers, which are
still discussed on spec level [see PR].
PR fortran/114283
gcc/fortran/ChangeLog:
* trans-openmp.cc (gfc_omp_predetermined_mapping): Map dummy
procedures as firstprivate.
libgomp/ChangeLog:
* testsuite/libgomp.fortran/declare-target-indirect-4.f90: New test.
While texinfo 7.0.3 does not warn, an older texinfo did complain about:
libgomp.texi:1964: warning: node next `omp_target_memcpy' in menu
`omp_target_memcpy_rect' and in sectioning `omp_target_memcpy_async' differ
libgomp/
* libgomp.texi (Device Memory Routines): Swap item order to match
the order of the '@node's of the '@subsection's.
'GCN_SUPPRESS_HOST_FALLBACK' originated as 'HSA_SUPPRESS_HOST_FALLBACK' in the
libgomp HSA plugin, where the idea was -- in my understanding -- that you
wouldn't have device code available for all functions that may be called, and
in that case transparently (shared memory system!) do host-fallback execution.
Or, with 'HSA_SUPPRESS_HOST_FALLBACK' set, you'd get those diagnosed.
This has then been copied into the libgomp GCN plugin as
'GCN_SUPPRESS_HOST_FALLBACK'. However, the original meaning isn't applicable
for the libgomp GCN plugin anymore: we assume that we're generating device code
for all relevant functions, and we're implementing a non-shared memory system,
where we cannot transparently do host-fallback execution for individual
functions.
However, 'GCN_SUPPRESS_HOST_FALLBACK' has gained an additional meaning, to
enforce a fatal error in case that 'libhsa-runtime64.so.1' can't be dynamically
loaded; keep that meaning.
libgomp/
* plugin/plugin-gcn.c (GOMP_OFFLOAD_can_run): Don't consider
'GCN_SUPPRESS_HOST_FALLBACK' anymore (assume always-'true').
(init_hsa_context): Adjust 'GCN_SUPPRESS_HOST_FALLBACK' error
message.
Per commit 683f118439
"OpenMP: Move omp requires checks to libgomp", we're now using 'return -1'
from 'GOMP_OFFLOAD_get_num_devices' for 'omp_requires_mask' purposes. This
missed that via 'nvptx_get_num_devices', we could also 'return -1' for
'cuDeviceGetCount' failure. Before, this meant (in 'gomp_target_init') to
silently ignore the plugin/device -- which also has been doubtful behavior.
Let's instead turn 'cuDeviceGetCount' failure into a fatal error, similar to
other errors during device initialization.
libgomp/
* plugin/plugin-nvptx.c (nvptx_get_num_devices):
'cuDeviceGetCount' failure is fatal.
If 'libhsa-runtime64.so.1', 'libcuda.so.1' are not available, the corresponding
libgomp plugin/device gets disabled, as before. But if they are available,
report any inconsistencies such as missing symbols, similar to how we fail in
presence of other issues during device initialization.
libgomp/
* plugin/plugin-gcn.c (init_hsa_runtime_functions): Fatal error
for missing symbols.
* plugin/plugin-nvptx.c (init_cuda_lib): Likewise.
For the type of the target callbacks we use elsehwere void (*) (void *) and
IMHO should use that for the reverse offload fallback as well (where the actual
callback is emitted using the same code as for host fallback or device kernel
entry routines), even when it is also ok to use void (*) () before C23 and
we aren't building libgomp with C23 yet. On some arches perhaps void (*) ()
could result in worse code generation because calls in that case like casts
to unprototyped functions need to sometimes pass argument in two different spots
etc. so that it deals with both passing it through ... and as a named argument.
2024-03-04 Jakub Jelinek <jakub@redhat.com>
PR libgomp/114216
* target.c (gomp_target_rev): Change host_fn type and corresponding
cast from void (*)() to void (*) (void *).
OpenMP permits '(first)private' for C++ member variables, which GCC handles
by tagging those by DECL_OMP_PRIVATIZED_MEMBER, adding a temporary VAR_DECL
and DECL_VALUE_EXPR pointing to the 'this->member_var' in the C++ front end.
The idea is that in omp-low.cc, the DECL_VALUE_EXPR is used before the
region (for 'firstprivate'; ignored for 'private') while in the region,
the DECL itself is used.
In gimplify, the value expansion is suppressed and deferred if the
lang_hooks.decls.omp_disregard_value_expr (decl, shared)
returns true - which is never the case if 'shared' is true. In OpenMP 4.5,
only 'map' and 'use_device_ptr' was permitted for the 'target' directive.
And when OpenMP 5.0's 'private'/'firstprivate' clauses was added, the
the update that now 'shared' argument could be false was missed. The
respective check has now been added.
2024-03-01 Jakub Jelinek <jakub@redhat.com>
Tobias Burnus <tburnus@baylibre.com>
PR c++/110347
gcc/ChangeLog:
* gimplify.cc (omp_notice_variable): Fix 'shared' arg to
lang_hooks.decls.omp_disregard_value_expr for
(first)private in target regions.
libgomp/ChangeLog:
* testsuite/libgomp.c++/target-lambda-3.C: Moved from
gcc/testsuite/g++.dg/gomp/ and fixed is-mapped handling.
* testsuite/libgomp.c++/target-lambda-1.C: Modify to also
also work without offloading.
* testsuite/libgomp.c++/firstprivate-1.C: New test.
* testsuite/libgomp.c++/firstprivate-2.C: New test.
* testsuite/libgomp.c++/private-1.C: New test.
* testsuite/libgomp.c++/private-2.C: New test.
* testsuite/libgomp.c++/target-lambda-4.C: New test.
* testsuite/libgomp.c++/use_device_ptr-1.C: New test.
gcc/testsuite/ChangeLog:
* g++.dg/gomp/target-lambda-1.C: Moved to become a
run-time test under testsuite/libgomp.c++.
Co-authored-by: Tobias Burnus <tburnus@baylibre.com>
These routines map simply to the C counterpart and are meanwhile
defined in OpenACC 3.3. (There are additional routine changes,
including the Fortran addition of acc_attach/acc_detach, that
require more work than a simple addition of an interface and
are therefore excluded.)
libgomp/ChangeLog:
* libgomp.texi (OpenACC Runtime Library Routines): Document new 3.3
routines that simply map to their C counterpart.
* openacc.f90 (openacc): Add them.
* openacc_lib.h: Likewise.
* testsuite/libgomp.oacc-fortran/acc_host_device_ptr.f90: New test.
* testsuite/libgomp.oacc-fortran/acc-memcpy.f90: New test.
* testsuite/libgomp.oacc-fortran/acc-memcpy-2.f90: New test.
* testsuite/libgomp.oacc-c-c++-common/lib-59.c: Crossref to f90 test.
* testsuite/libgomp.oacc-c-c++-common/lib-60.c: Likewise.
* testsuite/libgomp.oacc-c-c++-common/lib-95.c: Likewise.
The main 'arch' context selector for nvptx is, well, 'nvptx';
however, as 'nvptx64' is used as by LLVM, it makes sense
to support it as well.
Note that LLVM has: "The triple architecture can be one of
``nvptx`` (32-bit PTX) or ``nvptx64`` (64-bit PTX)."
GCC effectively only supports the 64bit variant (at least for
offloading). Thus, GCC's 'nvptx' is not quite the same as LLVM's.
The device-compiler part (nvptx_omp_device_kind_arch_isa) uses
TARGET_ABI64 such that nvptx64 is only defined with -m64.
gcc/ChangeLog:
* config/nvptx/gen-omp-device-properties.sh: Add 'nvptx64' to arch.
* config/nvptx/nvptx.cc (nvptx_omp_device_kind_arch_isa): Likewise.
libgomp/ChangeLog:
* libgomp.texi (OpenMP Context Selectors): Add 'nvptx64' as additional
'arch' value for nvptx.
Support for indirect calls to procedures/functions in offloaded target
regions is now available for C, C++ and Fortran.
2024-02-15 Kwok Cheung Yeung <kcyeung@baylibre.com>
libgomp/
* libgomp.texi (OpenMP 5.1): Mark indirect call support as fully
implemented.
Two libgomp tests XPASS on Solaris (any non-Linux target actually) since
their introduction:
XPASS: libgomp.c/alloc-pinned-1.c execution test
XPASS: libgomp.c/alloc-pinned-2.c execution test
The problem is that the test just prints
OS unsupported
and exits successfully, while the test is XFAILed:
/* { dg-xfail-run-if "Pinning not implemented on this host" { ! *-*-linux-gnu } } */
Fixed by aborting immediately after the message above in the non-Linux
case.
Tested on i386-pc-solaris2.11 and i686-pc-linux-gnu.
2024-02-02 Rainer Orth <ro@CeBiTec.Uni-Bielefeld.DE>
libgomp:
PR testsuite/113448
* testsuite/libgomp.c/alloc-pinned-1.c [!__linux__] (CHECK_SIZE):
Call abort.
* testsuite/libgomp.c/alloc-pinned-2.c [!__linux__] (CHECK_SIZE):
Likewise.
