Previously we added the `push_target_tripcount` function to send the
loop tripcount to the device runtime so we knew how to configure the
teams / threads for execute the loop for a teams distribute construct.
This was implemented as a separate function mostly to avoid changing the
interface for backwards compatbility. Now that we've changed it anyway
and the new interface can take an arbitrary number of arguments via the
struct without changing the ABI, we can move this to the new interface.
This will simplify the runtime by removing unnecessary state between
calls.
Depends on D128550
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D128816
This patch moves the logic for generating the offloading entries to the
OpenMPIRBuilder. This makes it easier to re-use in other places, such as
for OpenMP support in Flang or using the same method for generating
offloading entires for other languages like Cuda.
Reviewed By: tianshilei1992
Differential Revision: https://reviews.llvm.org/D123460
Rather than using a dummy void pointer type, we should specify the
correct private type and perform the bitcast beforehand rather than
afterwards. This way, the Address will have correct alignment
information.
The reduction initialization code creates a "naturally aligned null
pointer to void lvalue", which I found somewhat odd, even though it
works out in the end because it is not actually used. It doesn't
look like this code actually needs an LValue for anything though,
and we can use an invalid Address to represent this case instead.
Differential Revision: https://reviews.llvm.org/D116214
The problem with the old scheme is that we would need to keep track of
the "next region" and reset the num_threads value after it. The new RT
doesn't do it and an assertion is triggered. The old RT doesn't do it
either, I haven't tested it but I assume a num_threads clause might
impact multiple parallel regions "accidentally". Further, in SPMD mode
num_threads was simply ignored, for some reason beyond me.
In any case, parallel_51 is designed to take the clause value directly,
so let's do that instead.
Reviewed By: tianshilei1992
Differential Revision: https://reviews.llvm.org/D113623
Eachempati.
This patch adds clang (parsing, sema, serialization, codegen) support for the 'depend' clause on the 'taskwait' directive.
Reviewed By: ABataev
Differential Revision: https://reviews.llvm.org/D113540
This patch adds a new RTL function for worksharing. Currently we use
`__kmpc_for_static_init` for both the `distribute` and `parallel`
portion of the loop clause. This patch replaces the `distribute` portion
with a new runtime call `__kmpc_distribute_static_init`. Currently this
will be used exactly the same way, but will make it easier in the future
to fine-tune the distribute and parallel portion of the loop.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D110429
The device runtime contains several calls to __kmpc_get_hardware_num_threads_in_block
and __kmpc_get_hardware_num_blocks. If the thread_limit and the num_teams are constant,
these calls can be folded to the constant value.
In commit D106033 we have the optimization phase. This commit adds the attributes to
the outlined function for the grid size. the two attributes are `omp_target_num_teams` and
`omp_target_thread_limit`. These values are added as long as they are constant.
Two functions are created `getNumThreadsExprForTargetDirective` and
`getNumTeamsExprForTargetDirective`. The original functions `emitNumTeamsForTargetDirective`
and `emitNumThreadsForTargetDirective` identify the expresion and emit the code.
However, for the Device version of the outlined function, we cannot emit anything.
Therefore, this is a first attempt to separate emision of code from deduction of the
values.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D106298
Summary:
Currently the mapping names are not passed to the mapper components that set up
the array region. This means array mappings will not have their names availible
in the runtime. This patch fixes this by passing the argument name to the region
correctly. This means that the mapped variable's name will be the declared
mapper that placed it on the device.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D99681
Summary:
This patch adds support for passing in the original delcaration name in the source file to the libomptarget runtime. This will allow the runtime to provide more intelligent debugging messages. This patch takes the original expression parsed from the OpenMP map / update clause and provides a textual representation if it was explicitly mapped, otherwise it takes the name of the variable declaration as a fallback. The information in passed to the runtime in a global array of strings that matches the existing ident_t source location strings using ";name;filename;column;row;;"
Reviewers: jdoerfert
Differential Revision: https://reviews.llvm.org/D89802
Clang now asserts for the below case:
```
void clang::CodeGen::CGOpenMPRuntime::createOffloadEntriesAndInfoMetadata(): Assertion `std::get<0>(E) && "All ordered entries must exist!"' failed.
```
The reason why Clang hit the assert is because in
`emitTargetDataCalls`, both `BeginThenGen` and `BeginElseGen` call
`registerTargetRegionEntryInfo` and try to register the Entry in
OffloadEntriesTargetRegion with same key. If changing the expression in
if clause to any constant expression, then the assert disappear. (https://godbolt.org/z/TW7haj)
The assert itself is to avoid
user from accessing elements out of bound inside `OrderedEntries` in
`createOffloadEntriesAndInfoMetadata`.
In this patch, I add a check in `registerTargetRegionEntryInfo` to avoid
register the target region more than once.
A test case that triggers assert: https://godbolt.org/z/4cnGW8
Reviewed By: ABataev
Differential Revision: https://reviews.llvm.org/D90704
Summary:
This patch adds support for passing in the original delcaration name in the
source file to the libomptarget runtime. This will allow the runtime to provide
more intelligent debugging messages. This patch takes the original expression
parsed from the OpenMP map / update clause and provides a textual
representation if it was explicitly mapped, otherwise it takes the name of the
variable declaration as a fallback. The information in passed to the runtime in
a global array of strings that matches the existing ident_t source location
strings using ";name;filename;column;row;;". See
clang/test/OpenMP/target_map_names.cpp for an example of the generated output
for a given map clause.
Reviewers: jdoervert
Differential Revision: https://reviews.llvm.org/D89802
The patch adjusts the existing `llvm::DenseMap<unsigned, T>` and
`llvm::DenseSet<unsigned>` objects that store source locations, so
that they use `SourceLocation` directly instead of `unsigned`.
This patch relies on the `DenseMapInfo` trait added in D89719.
It also replaces the construction of `SourceLocation` objects from
the constants -1 and -2 with calls to the trait's methods `getEmptyKey`
and `getTombstoneKey` where appropriate.
Reviewed By: dexonsmith
Differential Revision: https://reviews.llvm.org/D69840
Summary:
Replace the OpenMP Runtime Library functions used in CGOpenMPRuntimeGPU
for OpenMP device code generation with ones in OMPKinds.def and use
OMPIRBuilder for generating runtime calls. This allows us to
consolidate more OpenMP code generation into the OMPIRBuilder. Future
additions to the GPU runtime functions should now go in OMPKinds.def
Reviewers: jdoerfert
Subscribers: aaron.ballman cfe-commits guansong llvm-commits sstefan1 yaxunl
Tags: #OpenMP #LLVM #clang
Differential Revision: https://reviews.llvm.org/D88430
Failing tests on Arm due to the tests automatically populating
incomatible pointer width architectures. Reverting until the tests are
updated. Failing tests:
OpenMP/distribute_parallel_for_num_threads_codegen.cpp
OpenMP/distribute_parallel_for_if_codegen.cpp
OpenMP/distribute_parallel_for_simd_if_codegen.cpp
OpenMP/distribute_parallel_for_simd_num_threads_codegen.cpp
OpenMP/target_teams_distribute_parallel_for_if_codegen.cpp
OpenMP/target_teams_distribute_parallel_for_simd_if_codegen.cpp
OpenMP/teams_distribute_parallel_for_if_codegen.cpp
OpenMP/teams_distribute_parallel_for_simd_if_codegen.cpp
This reverts commit 90eaedda9b.
Summary:
Replace the OpenMP Runtime Library functions used in CGOpenMPRuntimeGPU
for OpenMP device code generation with ones in OMPKinds.def and use
OMPIRBuilder for generating runtime calls. This allows us to consolidate
more OpenMP code generation into the OMPIRBuilder. This patch also
invalidates specifying target architectures with conflicting pointer
sizes.
Reviewers: jdoerfert
Subscribers: aaron.ballman cfe-commits guansong llvm-commits sstefan1 yaxunl
Tags: #OpenMP #Clang #LLVM
Differential Revision: https://reviews.llvm.org/D88430
Local vars, marked with pragma allocate, mustbe allocate by the call of
the runtime function and cannot be allocated as other local variables.
Instead, we allocate a space for the pointer in private record and store
the address, returned by kmpc_alloc call in this pointer.
So, for untied tasks
```
#pragma omp task untied
{
S s;
#pragma omp allocate(s) allocator(allocator)
s = x;
}
```
compiler generates something like this:
```
struct task_with_privates {
S *ptr;
};
void entry(task_with_privates *p) {
S *s = p->s;
switch(partid) {
case 1:
p->s = (S*)kmpc_alloc();
kmpc_omp_task();
br exit;
case 2:
*s = x;
kmpc_omp_task();
br exit;
case 2:
~S(s);
kmpc_free((void*)s);
br exit;
}
exit:
}
```
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D86558
In untied tasks, need to allocate the space for local variales, declared
in task region, when the memory for task data is allocated. THe function
can be interrupted and we can exit from the function in untied task
switch. Need to keep the state of the local variables in this case.
Also, the compiler should not call cleanup when exiting in untied task
switch until the real exit out of the declaration scope is met during
execution.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D84457
Summary:
In untied tasks, need to allocate the space for local variales, declared
in task region, when the memory for task data is allocated. THe function
can be interrupted and we can exit from the function in untied task
switch. Need to keep the state of the local variables in this case.
Also, the compiler should not call cleanup when exiting in untied task
switch until the real exit out of the declaration scope is met during
execution.
Reviewers: jdoerfert
Subscribers: yaxunl, guansong, cfe-commits, sstefan1, caomhin
Tags: #clang
Differential Revision: https://reviews.llvm.org/D84457
Replace the `ident_t` handling in Clang with the methods offered by the
OMPIRBuilder. This cuts down on the clang code as well as the
differences between the two, making further transitions easier. Tests
have changed but there should not be a real functional change. The most
interesting difference is probably that we stop generating local ident_t
allocations for now and just use globals. Given that this happens only
with debug info, the location part of the `ident_t` is probably bigger
than the test anyway. As the location part is already a global, we can
avoid the allocation, memcpy, and store in favor of a constant global
that is slightly bigger. This can be revisited if there are
complications.
Reviewed By: ABataev
Differential Revision: https://reviews.llvm.org/D80735
Without this patch, the following example fails but shouldn't
according to OpenMP TR8:
```
#pragma omp target enter data map(alloc:i)
#pragma omp target data map(present, alloc: i)
{
#pragma omp target exit data map(delete:i)
} // fails presence check here
```
OpenMP TR8 sec. 2.22.7.1 "map Clause", p. 321, L23-26 states:
> If the map clause appears on a target, target data, target enter
> data or target exit data construct with a present map-type-modifier
> then on entry to the region if the corresponding list item does not
> appear in the device data environment an error occurs and the
> program terminates.
There is no corresponding statement about the exit from a region.
Thus, the `present` modifier should:
1. Check for presence upon entry into any region, including a `target
exit data` region. This behavior is already implemented correctly.
2. Should not check for presence upon exit from any region, including
a `target` or `target data` region. Without this patch, this
behavior is not implemented correctly, breaking the above example.
In the case of `target data`, this patch fixes the latter behavior by
removing the `present` modifier from the map types Clang generates for
the runtime call at the end of the region.
In the case of `target`, we have not found a valid OpenMP program for
which such a fix would matter. It appears that, if a program can
guarantee that data is present at the beginning of a `target` region
so that there's no error there, that data is also guaranteed to be
present at the end. This patch adds a comment to the runtime to
document this case.
Reviewed By: grokos, RaviNarayanaswamy, ABataev
Differential Revision: https://reviews.llvm.org/D84422
This patch implements the code generation to use OpenMP 5.0 declare mapper (a.k.a. user-defined mapper) constructs.
Patch written by Lingda Li.
Differential Revision: https://reviews.llvm.org/D67833
Summary:
D82193 exposed a problem with global type definitions in
`OMPConstants.h`. This causes a race when running in thinLTO mode.
Types now live inside of OpenMPIRBuilder to prevent this from happening.
Reviewers: jdoerfert
Subscribers: yaxunl, hiraditya, guansong, dexonsmith, aaron.ballman, cfe-commits, llvm-commits
Tags: #clang, #llvm
Differential Revision: https://reviews.llvm.org/D83176
Summary:
Added initial codegen for 'affinity' clauses on task directives.
Emits next code:
```
kmp_task_affinity_info_t affs[<num_elems>];
void *td = __kmpc_task_alloc(..);
affs[<i>].base = &data_i;
affs[<i>].size = sizeof(data_i);
__kmpc_omp_reg_task_with_affinity(&loc, <gtid>, td, <num_elems>, affs);
```
The result returned by the call of `__kmpc_omp_reg_task_with_affinity`
function is ignored currently sincethe runtime currently ignores args
and returns 0 uncoditionally.
Reviewers: jdoerfert
Subscribers: yaxunl, guansong, sstefan1, llvm-commits, cfe-commits, caomhin
Tags: #clang, #llvm
Differential Revision: https://reviews.llvm.org/D80240
Summary: This changes Clang's generation of OpenMP runtime functions to use the types and functions defined in OpenMPKinds and OpenMPConstants. New OpenMP runtime function information should now be added to OMPKinds.def. This patch also changed the definitions of __kmpc_push_num_teams and __kmpc_copyprivate to match those found in the runtime.
Reviewers: jdoerfert
Reviewed By: jdoerfert
Subscribers: jfb, AndreyChurbanov, openmp-commits, fghanim, hiraditya, sstefan1, cfe-commits, llvm-commits
Tags: #openmp, #clang, #llvm
Differential Revision: https://reviews.llvm.org/D80222
Summary:
Predefined allocators should not be mapped at all (they are just enumeric
constants). FOr user-defined allocators need to map the traits only as
firstprivates, the allocator itself is private.
At the beginning of the target region the user-defined allocatores must
be created and then destroyed at the end of the target region:
```
omp_allocator_handle_t my_allocator = __kmpc_init_allocator(<gtid>,
/*default memhandle*/ 0, <number_of_traits>, &<traits>);
...
call void @__kmpc_destroy_allocator(<gtid>, my_allocator);
```
Reviewers: jdoerfert, aaron.ballman
Subscribers: jholewinski, yaxunl, guansong, cfe-commits, caomhin
Tags: #clang
Differential Revision: https://reviews.llvm.org/D79257
Summary:
Patch forces codegen to use the new runtime functions for task reductions where
the issue with passing the address of the original variables to the UDR
initializers is fixed. Also, this patch is required for upcoming
support of task modifier inreduction clause.
Reviewers: jdoerfert
Subscribers: yaxunl, guansong, cfe-commits, caomhin
Tags: #clang
Differential Revision: https://reviews.llvm.org/D78733
Implemented codegen for the iterator expression in the depend clauses.
Iterator construct is emitted the following way:
iterator(cnt1, cnt2, ...), in : <dep>
<TotalNumDeps> = <cnt1_size> * <cnt2_size> * ...;
kmp_depend_t deps[<TotalNumDeps>];
deps_counter = 0;
for (cnt1) {
for (cnt2) {
...
deps[deps_counter].base_addr = &<dep>;
deps[deps_counter].size = sizeof(<dep>);
deps[deps_counter].flags = in;
deps_counter += 1;
...
}
}
For depobj construct the codegen is very similar, but the memory is
allocated dynamically and added extra first item reserved for internal use.
This is the second part loosely extracted from D71179 and cleaned up.
This patch provides semantic analysis support for `omp begin/end declare
variant`, mostly as defined in OpenMP technical report 8 (TR8) [0].
The sema handling makes code generation obsolete as we generate "the
right" calls that can just be handled as usual. This handling also
applies to the existing, albeit problematic, `omp declare variant
support`. As a consequence a lot of unneeded code generation and
complexity is removed.
A major purpose of this patch is to provide proper `math.h`/`cmath`
support for OpenMP target offloading. See PR42061, PR42798, PR42799. The
current code was developed with this feature in mind, see [1].
The logic is as follows:
If we have seen a `#pragma omp begin declare variant match(<SELECTOR>)`
but not the corresponding `end declare variant`, and we find a function
definition we will:
1) Create a function declaration for the definition we were about to generate.
2) Create a function definition but with a mangled name (according to
`<SELECTOR>`).
3) Annotate the declaration with the `OMPDeclareVariantAttr`, the same
one used already for `omp declare variant`, using and the mangled
function definition as specialization for the context defined by
`<SELECTOR>`.
When a call is created we inspect it. If the target has an
`OMPDeclareVariantAttr` attribute we try to specialize the call. To this
end, all variants are checked, the best applicable one is picked and a
new call to the specialization is created. The new call is used instead
of the original one to the base function. To keep the AST printing and
tooling possible we utilize the PseudoObjectExpr. The original call is
the syntactic expression, the specialized call is the semantic
expression.
[0] https://www.openmp.org/wp-content/uploads/openmp-TR8.pdf
[1] https://reviews.llvm.org/D61399#change-496lQkg0mhRN
Reviewers: kiranchandramohan, ABataev, RaviNarayanaswamy, gtbercea, grokos, sdmitriev, JonChesterfield, hfinkel, fghanim, aaron.ballman
Subscribers: bollu, guansong, openmp-commits, cfe-commits
Tags: #clang
Differential Revision: https://reviews.llvm.org/D75779
If the ancestor device modifier is used and the value of the device
clause is evaluated to 1, the ancestor device shall be used for the
execution.
Since the reverse offloading is not supported yet, the target construct
execution is always initiated from the host, not from the device. So, if
the ancestor modifier is specified, just execute target region on the
host.
Added codegen for update clause in depobj. Reads the number of the
elements from the first element and updates flags for each element in
the loop.
```
omp_depend_t x;
kmp_depend_info *base = (kmp_depend_info *)x;
intptr_t num = x[-1].base_addr;
kmp_depend_info *end = x + num;
kmp_depend_info *el = base;
do {
el.flags = new_flag;
el = &el[1];
} while (el != end);
```
Added codegen for 'depend' clause in depobj directive. The depend clause
is emitted as kmp_depend_info <deps>[<number_of_items_in_clause> + 1]. The
first element in this array is reserved for storing the number of
elements in this array: <deps>[0].base_addr =
<number_of_items_in_clause>;
This extra element is required to implement 'update' and 'destroy'
clauses. It is required to know the size of array to destroy it
correctly and to update depency kind.
directive.
According to OpenMP 5.0, The atomic_default_mem_order clause specifies the default memory ordering behavior for atomic constructs that must be provided by an implementation. If the default memory ordering is specified as seq_cst, all atomic constructs on which memory-order-clause is not specified behave as if the seq_cst clause appears. If the default memory ordering is specified as relaxed, all atomic constructs on which memory-order-clause is not specified behave as if the relaxed clause appears.
If the default memory ordering is specified as acq_rel, atomic constructs on which memory-order-clause is not specified behave as if the release clause appears if the atomic write or atomic update operation is specified, as if the acquire clause appears if the atomic read operation is specified, and as if the acq_rel clause appears if the atomic captured update operation is specified.
Kazu Hirata