* gfortran.dg/goacc/classify-kernels-unparallelized.f95: Add
one dg-message for additional -fopt-info-optimized-omp output.
* gfortran.dg/goacc/classify-kernels.f95: Likewise.
* gfortran.dg/goacc/kernels-decompose-1.f95: Change 'note' to
'optimized' in dg-message.
Julian Brown [Wed, 11 Sep 2019 15:31:38 +0000 (08:31 -0700)]
[og9] A couple of GCN-specific test fixes
libgomp/
* testsuite/libgomp.oacc-c-c++-common/async_queue-1.c: Only run
NVidia-specific test on NVidia hardware.
* testsuite/libgomp.oacc-c-c++-common/asyncwait-nop-1.c (main):
Initialise for acc_device_gcn if testing on AMD GCN.
* testsuite/libgomp.oacc-c-c++-common/function-not-offloaded.c: Support
AMD GCN.
* testsuite/libgomp.oacc-c-c++-common/loop-dim-default.c (check): Skip
vector dimension test for AMD GCN.
Julian Brown [Sun, 8 Sep 2019 23:04:54 +0000 (16:04 -0700)]
[og9] Improve async serialize implementation for AMD GCN libgomp plugin
libgomp/
* plugin/plugin-gcn.c (struct placeholder, struct asyncwait_info,
enum entry_type): New.
(queue_entry): Use entry_type enum for tag. Add asyncwait and
placeholder event type fields.
(wait_for_queue_nonfull): New function.
(queue_push_launch): Use above function instead of raising a fatal
error on queue-full condition. Use KERNEL_LAUNCH instead of hardwired
0.
(queue_push_callback): Use wait_for_queue_nonfull instead of open-coded
wait sequence. Use CALLBACK instead of hardwired 1.
(queue_push_asyncwait, queue_push_placeholder): New.
(execute_queue_entry): Implement ASYNC_WAIT and ASYNC_PLACEHOLDER event
types.
(GOMP_OFFLOAD_openacc_async_serialize): Use queue_push_placeholder and
queue_push_asyncwait instead of host-synchronized wait_queue calls.
* testsuite/libgomp.oacc-c-c++-common/data-2-lib.c (main): Add missing
asynchronous waits.
* testsuite/libgomp.oacc-c-c++-common/data-2.c (main): Likewise.
Andrew Stubbs [Fri, 19 Jul 2019 14:45:07 +0000 (15:45 +0100)]
Fix relocations with multiple devices.
2019-09-10 Andrew Stubbs <ams@codesourcery.com>
libgomp/
* plugin/plugin-gcn.c (obstack_chunk_alloc): Delete.
(obstack_chunk_free): Delete.
(obstack.h): Remove include.
(create_and_finalize_hsa_program): Remove all unmodified_sections_os
and use sections directly from the issue.
Use "or 0x80" instead of SHT_NOTE to hide relocations, and then
simply recognise that ourselves.
Andrew Stubbs [Fri, 19 Jul 2019 11:00:53 +0000 (12:00 +0100)]
Move offload data into GPU memory.
2019-09-09 Andrew Stubbs <ams@codesourcery.com>
gcc/
* config/gcn/gcn-run.c (heap_region): New global variable.
(struct hsa_runtime_fn_info): Add hsa_memory_assign_agent_fn.
(init_hsa_runtime_functions): Initialize hsa_memory_assign_agent.
(get_kernarg_region): Move contents to ....
(get_memory_region): .... here.
(get_heap_region): New function.
(init_device): Initialize the heap_region.
(device_malloc): Add region parameter.
(struct kernargs): Move heap ....
(heap): ... to global scope.
(main): Allocate heap separate to kernargs.
libgomp/
* plugin/plugin-gcn.c (struct hsa_runtime_fn_info): Add
hsa_memory_assign_agent_fn.
(struct agent_info): Add data_region.
(init_hsa_runtime_functions): Initialize hsa_memory_assign_agent.
(get_kernarg_memory_region): Move contents to new function ...
(get_memory_region): ... here.
(get_data_memory_region): New function.
(GOMP_OFFLOAD_get_property): Use data_region, not kernarg_region.
(GOMP_OFFLOAD_init_device): Initialize data_region.
(create_and_finalize_hsa_program): Use data_region, not
kernarg_region, and assign heap to device agent.
(GOMP_OFFLOAD_alloc_by_agent): Likewise.
(image_address_p): Delete function.
(struct copy_data): Remove use_hsa_memory_copy.
(copy_data): Always use hsa_memory_copy.
(queue_push_copy): Remove use_hsa_memory_copy.
(GOMP_OFFLOAD_dev2host): Always use hsa_memory_copy.
(GOMP_OFFLOAD_host2dev): Likewise.
(GOMP_OFFLOAD_dev2dev): Likewise.
(gcn_exec): Use hsa_memory_copy.
(GOMP_OFFLOAD_openacc_async_host2dev): Always use hsa_memory_copy.
(GOMP_OFFLOAD_openacc_async_dev2host): Likewise.
Julian Brown [Fri, 9 Aug 2019 20:01:33 +0000 (13:01 -0700)]
[og9] Wait at end of OpenACC asynchronous kernels regions
gcc/
* omp-oacc-kernels.c (add_wait): New function, split out of...
(add_async_clauses_and_wait): ...here. Call new outlined function.
(decompose_kernels_region_body): Add wait at the end of
explicitly-asynchronous kernels regions.
Julian Brown [Wed, 31 Jul 2019 12:38:42 +0000 (05:38 -0700)]
[og9] Update parallel-dims.c and serial-dims.c warning line numbering.
* testsuite/libgomp.oacc-c-c++-common/parallel-dims.c: Use relative
line numbers for warning.
* testsuite/libgomp.oacc-c-c++-common/serial-dims.c: Likewise.
libatomic/
PR other/79543
* acinclude.m4 (LIBAT_CHECK_LINKER_FEATURES): Fix GNU ld --version
scanning to conform to the GNU Coding Standards.
* configure: Regenerate.
libffi/
PR other/79543
* acinclude.m4 (LIBAT_CHECK_LINKER_FEATURES): Fix GNU ld --version
scanning to conform to the GNU Coding Standards.
* configure: Regenerate.
libgomp/
PR other/79543
* acinclude.m4 (LIBGOMP_CHECK_LINKER_FEATURES): Fix GNU ld --version
scanning to conform to the GNU Coding Standards.
* configure: Regenerate.
libitm/
PR other/79543
* acinclude.m4 (LIBITM_CHECK_LINKER_FEATURES): Fix GNU ld --version
scanning to conform to the GNU Coding Standards.
* configure: Regenerate.
libstdc++-v3/
PR other/79543
* acinclude.m4 (GLIBCXX_CHECK_LINKER_FEATURES): Fix GNU ld --version
scanning to conform to the GNU Coding Standards.
* configure: Regenerate.
Julian Brown [Fri, 5 Jul 2019 01:14:41 +0000 (18:14 -0700)]
Assumed-size arrays with non-lexical data mappings
gcc/
* gimplify.c (gimplify_adjust_omp_clauses_1): Raise error for
assumed-size arrays in map clauses for Fortran/OpenMP.
* omp-low.c (lower_omp_target): Set the size of assumed-size Fortran
arrays to one to allow use of data already mapped on the offload device.
gcc/fortran/
* trans-openmp.c (gfc_omp_finish_clause): Change clauses mapping
assumed-size arrays to use the GOMP_MAP_FORCE_PRESENT map type.
Julian Brown [Wed, 20 Feb 2019 13:21:15 +0000 (05:21 -0800)]
Support Fortran 2003 class pointers in OpenACC
gcc/
* gimplify.c (insert_struct_comp_map): Handle GOMP_MAP_ATTACH_DETACH.
(gimplify_scan_omp_clauses): Separate out handling of OACC_ENTER_DATA
and OACC_EXIT_DATA. Remove GOMP_MAP_POINTER and GOMP_MAP_TO_PSET
mappings, apart from those following GOMP_MAP_DECLARE_{,DE}ALLOCATE.
Handle GOMP_MAP_ATTACH_DETACH.
* tree-pretty-print.c (dump_omp_clause): Support GOMP_MAP_ATTACH_DETACH.
Print "bias" not "len" for attach/detach clause types.
libgomp/
* oacc-parallel.c (GOACC_enter_exit_data): Fix optional arguments for
changes to clause stripping in enter data/exit data directives.
* testsuite/libgomp.oacc-fortran/class-ptr-param.f95: New test.
* testsuite/libgomp.oacc-fortran/classtypes-1.f95: New test.
* testsuite/libgomp.oacc-fortran/classtypes-2.f95: New test.
* testsuite/libgomp.oacc-fortran/derivedtype-1.f95: New test.
* testsuite/libgomp.oacc-fortran/derivedtype-2.f95: New test.
* testsuite/libgomp.oacc-fortran/multidim-slice.f95: New test.
Julian Brown [Tue, 28 May 2019 15:42:10 +0000 (08:42 -0700)]
Apply gangprivate attribute to innermost decl
...and fix parallelism-level calculation when applying the attribute.
gcc/
* omp-low.c (mark_oacc_gangprivate): Add CTX parameter. Use to look up
correct decl to add attribute to.
(lower_omp_for): Move "oacc gangprivate" processing from here...
(process_oacc_gangprivate_1): ...to here. New function.
(lower_omp_target): Update call to mark_oacc_gangprivate.
(execute_lower_omp): Call process_oacc_gangprivate_1 for each OMP
context.
libgomp/
* testsuite/libgomp.oacc-fortran/gangprivate-attrib-2.f90: New test.
Andrew Jenner [Wed, 30 Jan 2019 17:38:46 +0000 (09:38 -0800)]
Link libquadmath in Fortran libgomp tests
When invoking gcc to compile fortran code, fortran.exp is currently adding the
options -lgfortran -foffload=-lgfortran to the gcc command line. libgfortran
statically links to libquadmath and the gfortran driver invokes the linker
with -lquadmath as well as -lgfortran so fortran.exp should do so too.
Julian Brown [Mon, 20 May 2019 23:31:41 +0000 (16:31 -0700)]
Fix warning syntax and typos in two libgomp tests
libgomp/
* testsuite/libgomp.oacc-c-c++-common/kernels-decompose-1.c: Expect
"optimized:" not "note:" in warnings.
* testsuite/libgomp.oacc-c-c++-common/serial-dims.c: Fix typos in
warnings.
Julian Brown [Sun, 19 May 2019 17:42:20 +0000 (10:42 -0700)]
Fix references declared in lexically-enclosing OpenACC data region
gcc/fortran/
* trans-openmp.c (gfc_omp_finish_clause): Guard addition of clauses for
pointers with DECL_P.
gcc/
* gimplify.c (oacc_array_mapping_info): Add REF field.
(gimplify_scan_omp_clauses): Initialise above field for data blocks
passed by reference.
(gomp_oacc_needs_data_present): Handle references.
(gimplify_adjust_omp_clauses_1): Handle references and optional
arguments for variables declared in lexically-enclosing OpenACC data
region.
Julian Brown [Thu, 16 May 2019 12:45:35 +0000 (05:45 -0700)]
Avoid introducing 'create' mapping clauses for loop index variables in kernels regions
gcc/
* omp-oacc-kernels.c (find_omp_for_index_vars_1,
find_omp_for_index_vars): New functions.
(maybe_build_inner_data_region): Add IDX_VARS argument. Don't add
CREATE mapping clauses for loop index variables. Set TREE_ADDRESSABLE
flag on newly-mapped declarations as a side effect.
(decompose_kernels_region_body): Call find_omp_for_index_vars. Don't
create PRESENT clause for loop index variables. Pass index variable
set to maybe_build_inner_data_region.
Julian Brown [Wed, 9 Jan 2019 11:41:04 +0000 (03:41 -0800)]
Update OpenACC version to 2.6
gcc/c-family/
* c-cppbuiltin.c (c_cpp_builtins): Update _OPENACC define to 201711.
gcc/doc/
* invoke.texi: Update mention of OpenACC version to 2.6.
gcc/fortran/
* cpp.c (cpp_define_builtins): Update _OPENACC define to 201711.
* gfortran.texi: Update mentions of OpenACC version to 2.6.
* intrinsic.texi: Likewise.
gcc/testsuite/
* c-c++-common/cpp/openacc-define-3.c: Update expected value for
_OPENACC define.
* gfortran.dg/openacc-define-3.f90: Likewise.
libgomp/
* libgomp.texi: Update mentions of OpenACC version to 2.6. Update
section numbers to match version 2.6 of the spec.
* openacc.f90 (openacc_version): Update to 201711.
* openacc_lib.h (openacc_version): Update to 201711.
* testsuite/libgomp.oacc-fortran/openacc_version-1.f: Update expected
openacc_version to 201711.
* testsuite/libgomp.oacc-fortran/openacc_version-2.f90: Likewise.
Thomas Schwinge [Tue, 8 Jan 2019 14:21:35 +0000 (15:21 +0100)]
Add OpenACC 2.6 `acc_get_property' support: restore Intel MIC offloading
The "OpenACC 2.6 `acc_get_property' support" changes regressed the relevant
libgomp OpenMP execution test cases to no longer consider Intel MIC offloading
because of:
libgomp: while loading libgomp-plugin-intelmic.so.1: [...]/libgomp-plugin-intelmic.so.1: undefined symbol: GOMP_OFFLOAD_get_property
liboffloadmic/
* plugin/libgomp-plugin-intelmic.cpp (GOMP_OFFLOAD_get_property):
New function.
Thomas Schwinge [Fri, 1 Feb 2019 17:12:05 +0000 (18:12 +0100)]
Adjust parallelism of loops in gang-single parts of OpenACC kernels regions: "struct adjust_nested_loop_clauses_wi_info"
The current code apparently is too freaky at least for for GCC 4.6:
[...]/gcc/omp-oacc-kernels.c: In function 'tree_node* transform_kernels_loop_clauses(gimple*, tree, tree, tree, tree)':
[...]/gcc/omp-oacc-kernels.c:584:10: error: expected identifier before numeric constant
[...]/gcc/omp-oacc-kernels.c: In lambda function:
[...]/gcc/omp-oacc-kernels.c:584:25: error: expected '{' before '=' token
[...]/gcc/omp-oacc-kernels.c: In function 'tree_node* transform_kernels_loop_clauses(gimple*, tree, tree, tree, tree)':
[...]/gcc/omp-oacc-kernels.c:584:25: warning: lambda expressions only available with -std=c++0x or -std=gnu++0x [enabled by default]
[...]/gcc/omp-oacc-kernels.c:584:28: error: no match for 'operator=' in '{} = & loop_gang_clause'
[...]
gcc/
* omp-oacc-kernels.c (struct adjust_nested_loop_clauses_wi_info): New.
(adjust_nested_loop_clauses, transform_kernels_loop_clauses): Use it.
Thomas Schwinge [Thu, 24 Jan 2019 16:40:03 +0000 (08:40 -0800)]
New OpenACC kernels region decompose algorithm
Previously, OpenACC kernels region bodies were decomposed into a sequence of
alternating gang-single and gang-parallel "parallel" regions. The new
algorithm in this patch introduces a third possibility: Loops that look like
they might benefit from the parloops pass are converted into old "kernels"
regions, exposing them to the parloops pass later on. This has the benefit
that loops that cannot be parallelized are not offloaded to the GPU.
gcc/
* omp-oacc-kernels.c (adjust_region_code_walk_stmt_fn)
(adjust_region_code): New functions.
(make_loops_gang_single): Update.
(make_gang_single_region): Rename to...
(make_region_seq): ... this, and update.
(make_gang_parallel_loop_region): Rename to...
(make_region_loop_nest): ... this, and update.
(is_unconditional_oacc_for_loop): Remove stmt parameter and check.
(decompose_kernels_region_body): Update.
gcc/testsuite/
* c-c++-common/goacc/kernels-conversion.c: Adjust test.
* gfortran.dg/goacc/kernels-conversion.f95: Likewise.
* c-c++-common/goacc/kernels-decompose-1.c: New file.
* gfortran.dg/goacc/kernels-decompose-1.f95: Likewise.
libgomp/
* testsuite/libgomp.oacc-c-c++-common/kernels-decompose-1.c: New
file.
Gergö Barany [Mon, 21 Jan 2019 20:50:14 +0000 (12:50 -0800)]
Launch kernels asynchronously in OpenACC kernels regions
Kernels regions are decomposed into one or more smaller regions that are to
be executed in sequence. With this patch, all of these regions are launched
asynchronously, and a wait directive is added after them. This means that
the host only waits once for the kernels to complete, not once per kernel.
If the original kernels region was marked async, that asynchronous behavior
is preserved, and no wait is added.
gcc/
* omp-oacc-kernels.c (add_async_clauses_and_wait): New function...
(decompose_kernels_region_body): ... called from here.
Gergö Barany [Thu, 24 Jan 2019 06:11:11 +0000 (22:11 -0800)]
Adjust parallelism of loops in gang-single parts of OpenACC kernels regions
Loops in gang-single parts of kernels regions cannot be executed in
gang-redundant mode. If the user specified gang clauses on such loops, emit
an error and remove these clauses. Adjust automatic partitioning to exclude
gang partitioning in gang-single regions.
gcc/
* omp-oacc-kernels.c (add_parent_or_loop_num_clause): New function.
(adjust_nested_loop_clauses): Likewise.
(transform_kernels_loop_clauses, make_gang_parallel_loop_region):
Add worker and vector clause parameters, emit error on illegal
nesting.
(visit_loops_in_gang_single_region): Emit warning on conditionally
executed code with a gang clause.
(make_loops_gang_single): New function.
(decompose_kernels_region_body): Separate out gang/worker/vector clauses
for separate handling; add call to make_loops_gang_single.
* omp-offload.c (oacc_loop_auto_partitions): Add and propagate
is_oacc_gang_single parameter.
(oacc_loop_partition): Likewise.
(execute_oacc_device_lower): Adjust call to oacc_loop_partition.
Gergö Barany [Wed, 23 Jan 2019 22:32:57 +0000 (14:32 -0800)]
Handle conditional execution of loops in OpenACC kernels regions
Any OpenACC loop controlled by an if statement or a non-OpenACC loop must be
executed in a gang-single region. Detecting such loops is not trivial as
OpenACC kernels expansion is done on GIMPLE but before computation of the
control flow graph. This patch adds an auxiliary analysis for determining
whether a statement is inside a conditionally executed region (relative to
the kernels region's entry).
gcc/
* omp-oacc-kernels.c (control_flow_regions): New class.
(control_flow_regions::control_flow_regions): New constructor.
(control_flow_regions::is_unconditional_oacc_for_loop): New method.
(control_flow_regions::find_rep): Likewise.
(control_flow_regions::union_reps): Likewise.
(control_flow_regions::compute_regions): Likewise.
(decompose_kernels_region_body): Use test for conditional execution.
gcc/testsuite/
* c-c++-common/goacc/kernels-conversion.c: Add test for conditionally
executed code.
* gfortran.dg/goacc/kernels-conversion.f95: Likewise.
Gergö Barany [Mon, 21 Jan 2019 15:16:06 +0000 (07:16 -0800)]
Turn OpenACC kernels regions into a sequence of parallel regions
This patch decomposes each OpenACC kernels region into a sequence of
parallel regions. Each OpenACC loop nest turns into its own region; any code
between such loop nests is gathered up into a region as well. The loop
regions can be distributed across gangs if the original kernels region had a
num_gangs clause, while the other regions are executed in "gang-single"
mode. The implied default "auto" clause on kernels loops is made explicit
unless there is a conflicting clause.
gcc/
* omp-oacc-kernels.c (top_level_omp_for_in_stmt): New function.
(make_gang_single_region): Likewise.
(transform_kernels_loop_clauses, make_gang_parallel_loop_region):
Likewise.
(flatten_binds): Likewise.
(make_data_region_try_statement): Likewise.
(maybe_build_inner_data_region): Likewise.
(decompose_kernels_region_body): Likewise.
(transform_kernels_region): Delegate to decompose_kernels_region_body
and make_data_region_try_statement.
gcc/testsuite/
* c-c++-common/goacc/kernels-conversion.c: Test for a gang-single
region.
* gfortran.dg/goacc/kernels-conversion.f95: Likewise.
Gergö Barany [Mon, 21 Jan 2019 13:28:20 +0000 (05:28 -0800)]
Separate OpenACC kernels regions in data and parallel parts
This is the first in a series of patches that completely rework the handling
of the OpenACC "kernels" directive. In the future, kernels regions will be
transformed into data regions containing a sequence of serial and parallel
offloaded regions. This first patch sets up a new pass that is responsible
for this transformation, and in a first step constructs the new data region
containing a parallel region with the original kernels region's body.
gcc/
* Makefile.in: Add...
* omp-oacc-kernels.c: ... this new file for the kernels conversion
pass.
* flag-types.h (enum openacc_kernels): Add "split" style. Adjust
all users.
* doc/invoke.texi (-fopenacc-kernels): Update.
* passes.def: Add pass_convert_oacc_kernels to pipeline.
* tree-pass.h (make_pass_convert_oacc_kernels): Add declaration.
gcc/testsuite/
* c-c++-common/goacc/kernels-conversion.c: New test.
* gfortran.dg/goacc/kernels-conversion.f95: Likewise.
* c-c++-common/goacc/if-clause-2.c: Update.
* gfortran.dg/goacc/kernels-tree.f95: Likewise.
Thomas Schwinge [Wed, 23 Jan 2019 14:56:52 +0000 (06:56 -0800)]
Add OpenACC target kinds for decomposed kernels regions
This patch is in preparation for changes that will cut up OpenACC kernels
regions into individual parts. For the new sub-regions that will be
generated, this adds the following new kinds of OpenACC regions for internal
use:
- GF_OMP_TARGET_KIND_OACC_PARALLEL_KERNELS_PARALLELIZED for parts of kernels
regions to be executed in gang-redundant mode
- GF_OMP_TARGET_KIND_OACC_PARALLEL_KERNELS_GANG_SINGLE for parts of kernels
regions to be executed in gang-single mode
- GF_OMP_TARGET_KIND_OACC_DATA_KERNELS for data regions generated around the
body of a kernels region
gcc/
* gimple.h (enum gf_mask): Add new target kinds
GF_OMP_TARGET_KIND_OACC_PARALLEL_KERNELS_PARALLELIZED,
GF_OMP_TARGET_KIND_OACC_PARALLEL_KERNELS_GANG_SINGLE, and
GF_OMP_TARGET_KIND_OACC_DATA_KERNELS.
(is_gimple_omp_oacc): Handle new target kinds.
(is_gimple_omp_offloaded): Likewise.
* gimple-pretty-print.c (dump_gimple_omp_target): Likewise.
* omp-expand.c (expand_omp_target): Likewise.
(build_omp_regions_1): Likewise.
(omp_make_gimple_edges): Likewise.
* omp-low.c (is_oacc_parallel_or_serial): Likewise.
(was_originally_oacc_kernels): New function.
(scan_omp_for): Update check for illegal nesting.
(check_omp_nesting_restrictions): Handle new target kinds.
(lower_oacc_reductions): Likewise.
(lower_omp_target): Likewise.
* omp-offload.c (execute_oacc_device_lower): Likewise.
Add generic support for the OpenACC 2.6 `acc_get_property' and
`acc_get_property_string' routines, as well as full handlers for the
host and the NVPTX offload targets and a minimal handler for the HSA
offload target.
Include test cases for both C/C++ and Fortran support, both producing:
OpenACC vendor: GNU
OpenACC name: GOMP
OpenACC driver: 1.0
with the host driver and output like:
OpenACC vendor: Nvidia
OpenACC total memory: 12651462656
OpenACC free memory: 12202737664
OpenACC name: TITAN V
OpenACC driver: 9.1
The clause makes any device code use the local memory address for each
of the variables specified unless the given variable is already present
on the current device.
2018-12-19 Julian Brown <julian@codesourcery.com>
Maciej W. Rozycki <macro@codesourcery.com>
gcc/
* omp-low.c (lower_omp_target): Support GOMP_MAP_NO_ALLOC.
* tree-pretty-print.c (dump_omp_clause): Likewise.
include/
* gomp-constants.h (gomp_map_kind): Support GOMP_MAP_NO_ALLOC.
libgomp/
* target.c (gomp_map_vars_async): Support GOMP_MAP_NO_ALLOC.
* testsuite/libgomp.oacc-c-c++-common/nocreate-1.c: New test.
* testsuite/libgomp.oacc-c-c++-common/nocreate-2.c: New test.
* testsuite/libgomp.oacc-c-c++-common/nocreate-3.c: New test.
* testsuite/libgomp.oacc-c-c++-common/nocreate-4.c: New test.
* testsuite/libgomp.oacc-fortran/nocreate-1.f90: New test.
* testsuite/libgomp.oacc-fortran/nocreate-2.f90: New test.
The `serial' construct is equivalent to a `parallel' construct with
clauses `num_gangs(1) num_workers(1) vector_length(1)' implied.
Naturally these clauses are therefore not supported with the `serial'
construct. All the remaining clauses accepted with `parallel' are also
accepted with `serial'.
Consequently implementation is straightforward, by handling `serial'
exactly like `parallel', except for hardcoding dimensions rather than
taking them from the relevant clauses, in `expand_omp_target'.
Separate codes are used to denote the `serial' construct throughout the
middle end, even though the mapping of `serial' to an equivalent
`parallel' construct could have been done in the individual language
frontends, saving a lot of mechanical changes and avoiding middle-end
code expansion. This is so that any reporting such as with warning or
error messages and in diagnostic dumps use `serial' rather than
`parallel', therefore avoiding user confusion.
gcc/fortran/
* openmp.c (OACC_HOST_DATA_CLAUSES): Add OMP_CLAUSE_IF and
OMP_CLAUSE_IF_PRESENT.
gcc/
* omp-expand.c (expand_omp_target): Handle if_present flag on
OpenACC host_data construct.
gcc/testsuite/
* c-c++-common/goacc/host_data-1.c: Add tests of if and if_present
clauses on host_data.
* gfortran.dg/goacc/host_data-tree.f95: Likewise.
include/
* gomp-constants.h (GOACC_FLAG_HOST_DATA_IF_PRESENT): New constant.
Gergö Barany [Thu, 20 Dec 2018 14:07:34 +0000 (15:07 +0100)]
Report errors on missing OpenACC reduction clauses in nested reductions
..., as suggested by OpenACC 2.6, 2.9.11. "reduction clause".
In gcc/testsuite/c-c++-common/goacc/reduction-6.c, we remove the erroneous
reductions on variable b; adding a reduction clause to make it compile cleanly
would make it a duplicate of the test for variable c.
gcc/
* omp-low.c (struct omp_context): New fields
local_reduction_clauses, outer_reduction_clauses.
(new_omp_context): Initialize these.
(scan_sharing_clauses): Record reduction clauses on OpenACC
constructs.
(scan_omp_for): Check reduction clauses for incorrect nesting.
gcc/testsuite/
* c-c++-common/goacc/nested-reductions-fail.c: New test.
* c-c++-common/goacc/nested-reductions.c: New test.
* c-c++-common/goacc/reduction-6.c: Adjust.
libgomp/
* testsuite/libgomp.oacc-c-c++-common/par-loop-comb-reduction-1.c:
Add missing reduction clauses.
* testsuite/libgomp.oacc-c-c++-common/par-loop-comb-reduction-2.c:
Likewise.
* testsuite/libgomp.oacc-c-c++-common/par-loop-comb-reduction-3.c:
Likewise.
* testsuite/libgomp.oacc-c-c++-common/par-loop-comb-reduction-4.c:
Likewise.
Disable AC_PROG_CXX and consequently a libstdc++ dependency for libffi,
introduced with upstream libffi commit 7d698125b1f0 ("Use the proper C++
compiler to run C++ tests"). This is only needed for the libffi test
suite, which we don't have to support in the GCC tree, as libffi is
maintained as a separate project. The dependency causes a build failure
with the `powerpc64le-linux-gnu' target due to a circular dependency:
due to a libgomp dependency for libstdc++ and then a libffi dependency
for libgomp, introduced with commit 998eb38b265d ("Use functional
parameters for data mappings in OpenACC child functions").
Allow optional arguments to be used in the use_device OpenACC clause
Optional arguments should be treated as references rather than pointers
in the lowering. However, for non-present arguments, this would result
in a null dereference, so conditionals need to be added to detect and
handle this.
gcc/
* omp-low.c (lower_omp_target): For use_device clauses, generate
conditional statements to treat Fortran optional arguments like
references if non-null, or propogate null arguments into offloaded
code otherwise.
Reviewed-by: Julian Brown <julian@codesourcery.com> Reviewed-by: Thomas Schwinge <thomas@codesourcery.com>
(cherry picked from openacc-gcc-9-branch commit b23eb4c2405aeaf877b441d76e998034752a9c7f)
Add support for allocatable arrays as optional arguments
This patch allows allocatable arrays to be used as Fortran optional
arguments. When an optional argument is detected, the Fortran
front-end now generates extra code to test if the argument is null.
If so, it sets the size of the array contents to zero, and the
pointers to data to null. This prevents libgomp from trying to
copy non-existant data, and preserves the null pointer used by
PRESENT to detect non-present arguments.
gcc/fortran/
* trans-openmp.c (gfc_build_conditional_assign): New.
(gfc_build_conditional_assign_expr): New.
(gfc_omp_finish_clause): Add conditionals to set the clause
declaration to null and size to zero if the declaration is a
non-present optional argument.
(gfc_trans_omp_clauses_1): Likewise.
Calculate correct size for optional arguments used in the firstprivate clause
The lowering for firstprivate uses the pointer size rather than the size
of the referenced object when passed an optional argument. This patch
detects optional arguments as a special case and treats them as reference
types.
gcc/
* omp-general.c (omp_is_optional_argument): New.
* omp-general.h (omp_is_optional_argument): New.
* omp-low.c (lower_omp_target): Use size of referenced object when
optional argument used as argument to firstprivate.
A non-present passed-by-reference Fortran optional argument is represented
by a null pointer. When passed to an update directive, it should be ignored
as variable mappings are not created for null pointers. This should be
safe as it is not possible to change a non-present argument into a present
one (or vice-versa) in Fortran.
libgomp/
* oacc-mem.c (update_dev_host): Return early if the host address
is NULL.
* testsuite/libgomp.oacc-c-c++-common/lib-43.c: Remove.
* testsuite/libgomp.oacc-c-c++-common/lib-47.c: Likewise.
Reviewed-by: Julian Brown <julian@codesourcery.com> Reviewed-by: Thomas Schwinge <thomas@codesourcery.com>
(cherry picked from openacc-gcc-9-branch commit b930a8a1ee826ca52081c7c0e88554b3091afb62)
Julian Brown [Thu, 21 Mar 2019 22:09:24 +0000 (15:09 -0700)]
Add support for gang local storage allocation in shared memory
2018-12-11 Julian Brown <julian@codesourcery.com>
Chung-Lin Tang <cltang@codesourcery.com>
gcc/
* config/nvptx/nvptx.c (tree-hash-traits.h): Include.
(gangprivate_shared_size): New global variable.
(gangprivate_shared_align): Likewise.
(gangprivate_shared_sym): Likewise.
(gangprivate_shared_hmap): Likewise.
(nvptx_option_override): Initialize gangprivate_shared_sym,
gangprivate_shared_align.
(nvptx_file_end): Output gangprivate_shared_sym.
(nvptx_goacc_expand_accel_var): New function.
(nvptx_set_current_function): New function.
(TARGET_SET_CURRENT_FUNCTION): Define hook.
(TARGET_GOACC_EXPAND_ACCEL): Likewise.
* doc/tm.texi (TARGET_GOACC_EXPAND_ACCEL_VAR): Document new hook.
* doc/tm.texi.in (TARGET_GOACC_EXPAND_ACCEL_VAR): Likewise.
* expr.c (expand_expr_real_1): Remap decls marked with the
"oacc gangprivate" attribute.
* omp-low.c (omp_context): Add oacc_partitioning_level and
oacc_addressable_var_decls fields.
(new_omp_context): Initialize oacc_addressable_var_decls in new
omp_context.
(delete_omp_context): Delete oacc_addressable_var_decls in old
omp_context.
(lower_oacc_head_tail): Record partitioning-level count in omp context.
(oacc_record_private_var_clauses, oacc_record_vars_in_bind)
(mark_oacc_gangprivate): New functions.
(lower_omp_for): Call oacc_record_private_var_clauses with "for"
clauses. Call mark_oacc_gangprivate for gang-partitioned loops.
(lower_omp_target): Call oacc_record_private_var_clauses with "target"
clauses.
Call mark_oacc_gangprivate for offloaded target regions.
(lower_omp_1): Call vars_in_bind for GIMPLE_BIND within OMP regions.
* target.def (expand_accel_var): New hook.
libgomp/
* testsuite/libgomp.oacc-c-c++-common/gang-private-1.c: New test.
* testsuite/libgomp.oacc-c-c++-common/loop-gwv-2.c: New test.
* testsuite/libgomp.oacc-c/pr85465.c: New test.
* testsuite/libgomp.oacc-fortran/gangprivate-attrib-1.f90: New test.kk
Julian Brown [Mon, 18 Mar 2019 22:32:57 +0000 (15:32 -0700)]
New tests from pr72741 only
Diagnostics updated relative to last posted version.
2018-10-02 Cesar Philippidis <cesar@codesourcery.com>
gcc/testsuite/
* gfortran.dg/goacc/pr72741.f90: Update test for current diagnostics.
* gfortran.dg/goacc/pr72741-2.f: New test.
* gfortran.dg/goacc/pr72741-intrinsic-1.f: New test.
* gfortran.dg/goacc/pr72741-intrinsic-2.f: New test.
Julian Brown [Wed, 27 Feb 2019 00:00:54 +0000 (16:00 -0800)]
Fix hang when running oacc exec with CUDA 9.0 nvprof
2018-09-20 Tom de Vries <tdevries@suse.de>
Cesar Philippidis <cesar@codesourcery.com>
libgomp/
* oacc-init.c (acc_init_state_lock, acc_init_state, acc_init_thread):
New variable.
(acc_init_1): Set acc_init_thread to pthread_self (). Set
acc_init_state to initializing at the start, and to initialized at the
end.
(self_initializing_p): New function.
(acc_get_device_type): Return acc_device_none if called by thread that
is currently executing acc_init_1.
gcc/testsuite/
* c-c++-common/goacc/reduction-8.c: New test.
libgomp/
* testsuite/libgomp.oacc-c-c++-common/privatize-reduction-1.c: New
test.
* testsuite/libgomp.oacc-c-c++-common/privatize-reduction-2.c: New
test.
Julian Brown [Tue, 26 Feb 2019 23:55:23 +0000 (15:55 -0800)]
Don't mark OpenACC auto loops as independent inside acc parallel regions
2018-09-20 Cesar Philippidis <cesar@codesourcery.com>
gcc/
* omp-low.c (lower_oacc_head_mark): Don't mark OpenACC auto
loops as independent inside acc parallel regions.
gcc/testsuite/
* c-c++-common/goacc/loop-auto-1.c: Adjust test case to conform to
the new behavior of the auto clause in OpenACC 2.5.
* c-c++-common/goacc/loop-auto-2.c: Likewise.
* gcc.dg/goacc/loop-processing-1.c: Likewise.
* c-c++-common/goacc/loop-auto-3.c: New test.
* gfortran.dg/goacc/loop-auto-1.f90: New test.
libgomp/
* testsuite/libgomp.oacc-c-c++-common/loop-auto-1.c: Adjust test case
to conform to the new behavior of the auto clause in OpenACC 2.5.
projects / thirdparty / gcc.git / log
