Adding Separate OpenMP Offloading Backend to libcxx/include/__algorithm/pstl_backends
#66968
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
llvmbot
added
the
libc++
|
@llvm/pr-subscribers-github-workflow @llvm/pr-subscribers-libcxx ChangesTo address some of the good feedback from the recent request for comments, I have added a separate OpenMP offloading backend in To begin with, I have only added I know there are still a lot of things missing, such as LIT tests, but I think we should try to agree on the project structure first. Full diff: https://github.com/llvm/llvm-project/pull/66968.diff 9 Files Affected:
diff --git a/libcxx/CMakeLists.txt b/libcxx/CMakeLists.txt
index bb2898b799bcef9..43d2a448de79584 100644
--- a/libcxx/CMakeLists.txt
+++ b/libcxx/CMakeLists.txt
@@ -290,6 +290,8 @@ option(LIBCXX_HAS_WIN32_THREAD_API "Ignore auto-detection and force use of win32
option(LIBCXX_HAS_EXTERNAL_THREAD_API
"Build libc++ with an externalized threading API.
This option may only be set to ON when LIBCXX_ENABLE_THREADS=ON." OFF)
+option(LIBCXX_ENABLE_GPU_OFFLOAD
+ "Build libc++ with support for GPU offload" OFF)
if (LIBCXX_ENABLE_THREADS)
set(LIBCXX_PSTL_CPU_BACKEND "std_thread" CACHE STRING "Which PSTL CPU backend to use")
@@ -297,6 +299,14 @@ else()
set(LIBCXX_PSTL_CPU_BACKEND "serial" CACHE STRING "Which PSTL CPU backend to use")
endif()
+if (NOT DEFINED LIBCXX_PSTL_GPU_BACKEND)
+ if (${LIBCXX_ENABLE_GPU_OFFLOAD})
+ set(LIBCXX_PSTL_GPU_BACKEND "omp_offload" CACHE STRING "Which PSTL GPU backend to use")
+ else()
+ set(LIBCXX_PSTL_GPU_BACKEND "none" CACHE STRING "Which PSTL GPU backend to use")
+ endif()
+endif()
+
# Misc options ----------------------------------------------------------------
# FIXME: Turn -pedantic back ON. It is currently off because it warns
# about #include_next which is used everywhere.
@@ -809,6 +819,10 @@ else()
Valid backends are: serial, std_thread and libdispatch")
endif()
+if (LIBCXX_PSTL_GPU_BACKEND STREQUAL "omp_offload")
+ config_define(1 _LIBCPP_PSTL_GPU_BACKEND_OMP_OFFLOAD)
+endif()
+
if (LIBCXX_ABI_DEFINES)
set(abi_defines)
foreach (abi_define ${LIBCXX_ABI_DEFINES})
diff --git a/libcxx/include/CMakeLists.txt b/libcxx/include/CMakeLists.txt
index 2ec755236dbaee2..a3d72df61a86dde 100644
--- a/libcxx/include/CMakeLists.txt
+++ b/libcxx/include/CMakeLists.txt
@@ -85,6 +85,11 @@ set(files
__algorithm/pstl_backends/cpu_backends/thread.h
__algorithm/pstl_backends/cpu_backends/transform.h
__algorithm/pstl_backends/cpu_backends/transform_reduce.h
+ __algorithm/pstl_backends/gpu_backend.h
+ __algorithm/pstl_backends/gpu_backends/backend.h
+ __algorithm/pstl_backends/gpu_backends/fill.h
+ __algorithm/pstl_backends/gpu_backends/for_each.h
+ __algorithm/pstl_backends/gpu_backends/omp_offload.h
__algorithm/pstl_copy.h
__algorithm/pstl_count.h
__algorithm/pstl_fill.h
diff --git a/libcxx/include/__algorithm/pstl_backend.h b/libcxx/include/__algorithm/pstl_backend.h
index 93372f019031b63..47f5191b48517ba 100644
--- a/libcxx/include/__algorithm/pstl_backend.h
+++ b/libcxx/include/__algorithm/pstl_backend.h
@@ -10,6 +10,7 @@
#define _LIBCPP___ALGORITHM_PSTL_BACKEND_H
#include <__algorithm/pstl_backends/cpu_backend.h>
+#include <__algorithm/pstl_backends/gpu_backend.h>
#include <__config>
#include <execution>
@@ -179,10 +180,17 @@ struct __select_backend<std::execution::parallel_policy> {
using type = __cpu_backend_tag;
};
+# if defined(_LIBCPP_PSTL_GPU_BACKEND_OMP_OFFLOAD)
+template <>
+struct __select_backend<std::execution::parallel_unsequenced_policy> {
+ using type = __gpu_backend_tag;
+};
+# else
template <>
struct __select_backend<std::execution::parallel_unsequenced_policy> {
using type = __cpu_backend_tag;
};
+# endif
# else
diff --git a/libcxx/include/__algorithm/pstl_backends/gpu_backend.h b/libcxx/include/__algorithm/pstl_backends/gpu_backend.h
new file mode 100644
index 000000000000000..7237036156a1bf3
--- /dev/null
+++ b/libcxx/include/__algorithm/pstl_backends/gpu_backend.h
@@ -0,0 +1,21 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _LIBCPP___ALGORITHM_PSTL_BACKENDS_GPU_BACKEND_H
+#define _LIBCPP___ALGORITHM_PSTL_BACKENDS_GPU_BACKEND_H
+
+#include <__config>
+
+#include <__algorithm/pstl_backends/gpu_backends/backend.h>
+
+#if defined(_LIBCPP_PSTL_GPU_BACKEND_OMP_OFFLOAD)
+# include <__algorithm/pstl_backends/gpu_backends/fill.h>
+# include <__algorithm/pstl_backends/gpu_backends/for_each.h>
+#endif
+
+#endif // _LIBCPP___ALGORITHM_PSTL_BACKENDS_GPU_BACKEND_H
diff --git a/libcxx/include/__algorithm/pstl_backends/gpu_backends/backend.h b/libcxx/include/__algorithm/pstl_backends/gpu_backends/backend.h
new file mode 100644
index 000000000000000..a8b400afbb94d9d
--- /dev/null
+++ b/libcxx/include/__algorithm/pstl_backends/gpu_backends/backend.h
@@ -0,0 +1,33 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _LIBCPP___ALGORITHM_PSTL_BACKENDS_GPU_BACKEND_BACKEND_H
+#define _LIBCPP___ALGORITHM_PSTL_BACKENDS_GPU_BACKEND_BACKEND_H
+
+#include <__config>
+#include <cstddef>
+
+#if defined(_LIBCPP_PSTL_GPU_BACKEND_OMP_OFFLOAD)
+# include <__algorithm/pstl_backends/gpu_backends/omp_offload.h>
+#endif
+
+#if !defined(_LIBCPP_HAS_NO_PRAGMA_SYSTEM_HEADER)
+# pragma GCC system_header
+#endif
+
+#if _LIBCPP_STD_VER >= 17
+
+_LIBCPP_BEGIN_NAMESPACE_STD
+
+struct __gpu_backend_tag {};
+
+_LIBCPP_END_NAMESPACE_STD
+
+#endif // _LIBCPP_STD_VER >= 17
+
+#endif // _LIBCPP___ALGORITHM_PSTL_BACKENDS_GPU_BACKEND_BACKEND_H
diff --git a/libcxx/include/__algorithm/pstl_backends/gpu_backends/fill.h b/libcxx/include/__algorithm/pstl_backends/gpu_backends/fill.h
new file mode 100644
index 000000000000000..32926da87e2a083
--- /dev/null
+++ b/libcxx/include/__algorithm/pstl_backends/gpu_backends/fill.h
@@ -0,0 +1,59 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _LIBCPP___ALGORITHM_PSTL_BACKENDS_GPU_BACKNEDS_FILL_H
+#define _LIBCPP___ALGORITHM_PSTL_BACKENDS_GPU_BACKNEDS_FILL_H
+
+#include <__algorithm/fill.h>
+#include <__algorithm/pstl_backends/cpu_backends/backend.h>
+#include <__algorithm/pstl_backends/gpu_backends/backend.h>
+#include <__config>
+#include <__iterator/concepts.h>
+#include <__type_traits/is_execution_policy.h>
+#include <__utility/terminate_on_exception.h>
+#include <stdio.h>
+
+#if !defined(_LIBCPP_HAS_NO_PRAGMA_SYSTEM_HEADER)
+# pragma GCC system_header
+#endif
+
+#if !defined(_LIBCPP_HAS_NO_INCOMPLETE_PSTL) && _LIBCPP_STD_VER >= 17
+
+_LIBCPP_BEGIN_NAMESPACE_STD
+
+template <class _ExecutionPolicy, class _ForwardIterator, class _Tp>
+_LIBCPP_HIDE_FROM_ABI void
+__pstl_fill(__gpu_backend_tag, _ForwardIterator __first, _ForwardIterator __last, const _Tp& __value) {
+ // It is only safe to execute for_each on the GPU, it the execution policy is
+ // parallel unsequenced, as it is the only execution policy prohibiting throwing
+ // exceptions and allowing SIMD instructions
+ if constexpr (__is_unsequenced_execution_policy_v<_ExecutionPolicy> &&
+ __has_random_access_iterator_category_or_concept<_ForwardIterator>::value) {
+ std::__par_backend::__parallel_for_simd_val_1(__first, __last - __first, __value);
+ }
+ // Else if the excution policy is parallel, we execute for_each on the CPU instead
+ else if constexpr (__is_parallel_execution_policy_v<_ExecutionPolicy> &&
+ __has_random_access_iterator_category_or_concept<_ForwardIterator>::value) {
+ std::__terminate_on_exception([&] {
+ __par_backend::__parallel_for(
+ __first, __last, [&__value](_ForwardIterator __brick_first, _ForwardIterator __brick_last) {
+ std::__pstl_fill<__remove_parallel_policy_t<_ExecutionPolicy>>(
+ __cpu_backend_tag{}, __brick_first, __brick_last, __value);
+ });
+ });
+ // Else we execute for_each in serial
+ } else {
+ std::fill(__first, __last, __value);
+ }
+}
+
+_LIBCPP_END_NAMESPACE_STD
+
+#endif // !defined(_LIBCPP_HAS_NO_INCOMPLETE_PSTL) && _LIBCPP_STD_VER >= 17
+
+#endif // _LIBCPP___ALGORITHM_PSTL_BACKENDS_GPU_BACKNEDS_FILL_H
diff --git a/libcxx/include/__algorithm/pstl_backends/gpu_backends/for_each.h b/libcxx/include/__algorithm/pstl_backends/gpu_backends/for_each.h
new file mode 100644
index 000000000000000..14de2af8e4a15c6
--- /dev/null
+++ b/libcxx/include/__algorithm/pstl_backends/gpu_backends/for_each.h
@@ -0,0 +1,59 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _LIBCPP___ALGORITHM_PSTL_BACKENDS_GPU_BACKNEDS_FOR_EACH_H
+#define _LIBCPP___ALGORITHM_PSTL_BACKENDS_GPU_BACKNEDS_FOR_EACH_H
+
+#include <__algorithm/for_each.h>
+#include <__algorithm/pstl_backends/cpu_backends/backend.h>
+#include <__algorithm/pstl_backends/gpu_backends/backend.h>
+#include <__config>
+#include <__iterator/concepts.h>
+#include <__type_traits/is_execution_policy.h>
+#include <__utility/terminate_on_exception.h>
+#include <stdio.h>
+
+#if !defined(_LIBCPP_HAS_NO_PRAGMA_SYSTEM_HEADER)
+# pragma GCC system_header
+#endif
+
+#if !defined(_LIBCPP_HAS_NO_INCOMPLETE_PSTL) && _LIBCPP_STD_VER >= 17
+
+_LIBCPP_BEGIN_NAMESPACE_STD
+
+template <class _ExecutionPolicy, class _ForwardIterator, class _Functor>
+_LIBCPP_HIDE_FROM_ABI void
+__pstl_for_each(__gpu_backend_tag, _ForwardIterator __first, _ForwardIterator __last, _Functor __func) {
+ // It is only safe to execute for_each on the GPU, it the execution policy is
+ // parallel unsequenced, as it is the only execution policy prohibiting throwing
+ // exceptions and allowing SIMD instructions
+ if constexpr (__is_unsequenced_execution_policy_v<_ExecutionPolicy> &&
+ __has_random_access_iterator_category_or_concept<_ForwardIterator>::value) {
+ std::__par_backend::__parallel_for_simd_1(__first, __last - __first, __func);
+ }
+ // Else if the excution policy is parallel, we execute for_each on the CPU instead
+ else if constexpr (__is_parallel_execution_policy_v<_ExecutionPolicy> &&
+ __has_random_access_iterator_category_or_concept<_ForwardIterator>::value) {
+ std::__terminate_on_exception([&] {
+ std::__par_backend::__parallel_for(
+ __first, __last, [__func](_ForwardIterator __brick_first, _ForwardIterator __brick_last) {
+ std::__pstl_for_each<__remove_parallel_policy_t<_ExecutionPolicy>>(
+ __cpu_backend_tag{}, __brick_first, __brick_last, __func);
+ });
+ });
+ // Else we execute for_each in serial
+ } else {
+ std::for_each(__first, __last, __func);
+ }
+}
+
+_LIBCPP_END_NAMESPACE_STD
+
+#endif // !defined(_LIBCPP_HAS_NO_INCOMPLETE_PSTL) && _LIBCPP_STD_VER >= 17
+
+#endif // _LIBCPP___ALGORITHM_PSTL_BACKENDS_GPU_BACKNEDS_FOR_EACH_H
diff --git a/libcxx/include/__algorithm/pstl_backends/gpu_backends/omp_offload.h b/libcxx/include/__algorithm/pstl_backends/gpu_backends/omp_offload.h
new file mode 100644
index 000000000000000..4baa4e7f65859d1
--- /dev/null
+++ b/libcxx/include/__algorithm/pstl_backends/gpu_backends/omp_offload.h
@@ -0,0 +1,94 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _LIBCPP___ALGORITHM_PSTL_BACKENDS_CPU_BACKENDS_OMP_OFFLOAD_H
+#define _LIBCPP___ALGORITHM_PSTL_BACKENDS_CPU_BACKENDS_OMP_OFFLOAD_H
+
+#include <__assert>
+#include <__config>
+#include <__utility/move.h>
+#include <cstddef>
+
+#if !defined(_LIBCPP_HAS_NO_PRAGMA_SYSTEM_HEADER)
+# pragma GCC system_header
+#endif
+
+_LIBCPP_PUSH_MACROS
+#include <__undef_macros>
+
+#if !defined(_LIBCPP_HAS_NO_INCOMPLETE_PSTL) && _LIBCPP_STD_VER >= 17
+
+_LIBCPP_BEGIN_NAMESPACE_STD
+
+namespace __par_backend {
+inline namespace __omp_gpu_backend {
+
+// In OpenMP, we need to extract the pointer for the underlying data for data
+// structures like std::vector and std::array to be able to map the data to the
+// device.
+
+template <typename T>
+_LIBCPP_HIDE_FROM_ABI inline T __omp_extract_base_ptr(T p) {
+ return p;
+}
+
+template <typename T>
+_LIBCPP_HIDE_FROM_ABI inline T __omp_extract_base_ptr(std::__wrap_iter<T> w) {
+ std::pointer_traits<std::__wrap_iter<T>> PT;
+ return PT.to_address(w);
+}
+
+// Applying function or lambda in a loop
+
+template <class _Iterator, class _DifferenceType, class _Function>
+_LIBCPP_HIDE_FROM_ABI _Iterator
+__omp_parallel_for_simd_1(_Iterator __first, _DifferenceType __n, _Function __f) noexcept {
+# pragma omp target teams distribute parallel for simd map(tofrom : __first[0 : __n])
+ for (_DifferenceType __i = 0; __i < __n; ++__i)
+ __f(__first[__i]);
+
+ return __first + __n;
+}
+
+// Extracting the underlying pointer
+
+template <class _Iterator, class _DifferenceType, class _Function>
+_LIBCPP_HIDE_FROM_ABI _Iterator __parallel_for_simd_1(_Iterator __first, _DifferenceType __n, _Function __f) noexcept {
+ __omp_parallel_for_simd_1(__omp_gpu_backend::__omp_extract_base_ptr(__first), __n, __f);
+ return __first + __n;
+}
+
+// Assigning a value in a loop
+
+template <class _Index, class _DifferenceType, class _Tp>
+_LIBCPP_HIDE_FROM_ABI _Index
+__omp_parallel_for_simd_val_1(_Index __first, _DifferenceType __n, const _Tp& __value) noexcept {
+# pragma omp target teams distribute parallel for simd map(tofrom : __first[0 : __n]) map(to : __value)
+ for (_DifferenceType __i = 0; __i < __n; ++__i)
+ __first[__i] = __value;
+
+ return __first + __n;
+}
+
+template <class _Index, class _DifferenceType, class _Tp>
+_LIBCPP_HIDE_FROM_ABI _Index
+__parallel_for_simd_val_1(_Index __first, _DifferenceType __n, const _Tp& __value) noexcept {
+ __omp_parallel_for_simd_val_1(__omp_gpu_backend::__omp_extract_base_ptr(__first), __n, __value);
+ return __first + __n;
+}
+
+} // namespace __omp_gpu_backend
+} // namespace __par_backend
+
+_LIBCPP_END_NAMESPACE_STD
+
+#endif // !defined(_LIBCPP_HAS_NO_INCOMPLETE_PSTL) && && _LIBCPP_STD_VER >= 17
+
+_LIBCPP_POP_MACROS
+
+#endif // _LIBCPP___ALGORITHM_PSTL_BACKENDS_CPU_BACKENDS_OMP_OFFLOAD_H
diff --git a/libcxx/include/__config_site.in b/libcxx/include/__config_site.in
index c85cbcd02c441b9..e0edddce3afc3ff 100644
--- a/libcxx/include/__config_site.in
+++ b/libcxx/include/__config_site.in
@@ -34,6 +34,7 @@
#cmakedefine _LIBCPP_PSTL_CPU_BACKEND_SERIAL
#cmakedefine _LIBCPP_PSTL_CPU_BACKEND_THREAD
#cmakedefine _LIBCPP_PSTL_CPU_BACKEND_LIBDISPATCH
+#cmakedefine _LIBCPP_PSTL_GPU_BACKEND_OMP_OFFLOAD
// Hardening.
#cmakedefine01 _LIBCPP_ENABLE_HARDENED_MODE_DEFAULT
|
|
Do we want to enable GPU offloading when compiling |
|
I made it depend on CMake options only. |
|
✅ With the latest revision this PR passed the C/C++ code formatter. |
AntonRydahl
force-pushed
the
libcxx_pstl_omp_offload_backend
branch
from
a5f7752 to
96adadf
Compare
ldionne
requested changes
Oct 4, 2023
There was a problem hiding this comment.
Thanks a lot for the patch! I think this is really great and I'm glad we're paving the way for new backends being added in the library. Since this is the first such backend, there are some thing to change as we discussed during our live review, but I definitely see a path forward.
Note to self: did not have time to review omp_offload.h yet, will do next time.
| @@ -0,0 +1,349 @@ | |||
| //===----------------------------------------------------------------------===// | |||
There was a problem hiding this comment.
Not attached to this line:
We need to discuss how exceptions are handled when we offload to the GPU. @AntonRydahl mentioned that throw was compiled to __builtin_trap(). This is going to both fail our tests and make us non-conforming. Our preference would be that the OpenMP backend doesn't try to offload if there are exceptions in the code path we're considering offloading.
Needs more discussion.
There was a problem hiding this comment.
I have written a section about this in the libc++ documentation. Do you think the description is sufficiently detailed?
| // See https://llvm.org/LICENSE.txt for license information. | ||
| // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception | ||
| // | ||
| //===----------------------------------------------------------------------===// |
There was a problem hiding this comment.
You can actually remove this header entirely until you have an implementation for it, since it'll be implemented using std::find_if. Note that if you run into issues while doing that, it's a pre-existing PSTL bug.
There was a problem hiding this comment.
I think there is unfortunately a bug in the PSTL logic. I cannot compile libcxx if I remove the overlay header files for any_of, find_if, or stable_sort.
libcxx/include/__algorithm/pstl_backends/gpu_backends/for_each.h
Outdated
Show resolved
Hide resolved
libcxx/include/__algorithm/pstl_backends/gpu_backends/transform.h
Outdated
Show resolved
Hide resolved
libcxx/include/__algorithm/pstl_backends/gpu_backends/transform.h
Outdated
Show resolved
Hide resolved
libcxx/include/__algorithm/pstl_backends/gpu_backends/transform_reduce.h
Outdated
Show resolved
Hide resolved
| template <class _T1, class _T2, class _T3> | ||
| struct _LIBCPP_HIDE_FROM_ABI __is_supported_reduction : std::false_type {}; | ||
|
|
||
| # define __PSTL_IS_SUPPORTED_REDUCTION(funname) \ |
There was a problem hiding this comment.
There's something trying to emerge here. We already have libcxx/include/__type_traits/predicate_traits.h and libcxx/include/__type_traits/operation_traits.h. IMO we could unify those into something like:
template <>
struct __desugars_to<__equal_to, std::equal_to<>> : true_type {};
template <>
struct __desugars_to<ranges::equal_to, std::equal_to<>> : true_type {};
template <>
struct __desugars_to<ranges::plus, std::plus<>> : std::true_type {};
// etc...
Then, this becomes simply:
template <class _Func>
struct __is_supported_reduction : bool_constant<
__desugars_to<_Func, std::minus<>>::value ||
__desugars_to<_Func, std::multiplies<>>::value ||
__desugars_to<_Func, std::logical_and<>>::value ||
__desugars_to<_Func, std::logical_or<>>::value ||
__desugars_to<_Func, std::bit_and<>>::value ||
__desugars_to<_Func, std::bit_or<>>::value ||
__desugars_to<_Func, std::bit_xor<>>::value
> {};
This would mandate some prior refactoring, but it would be quite nice.
There was a problem hiding this comment.
To not make this PR even bigger, I have added a refactor to use __desugars_to here: #68642
I hope that is what you meant. 😄
ldionne
added a commit
to ldionne/llvm-project
that referenced
this pull request
Oct 4, 2023
…TION_POINT The _LIBCPP_PSTL_CUSTOMIZATION_POINT macro was assuming that the policy was called _RawPolicy and referencing it by name. It happened to always work but this was definitely accidental and an oversight in the original implementation. This patch fixes that by passing the policy to the macro explicitly. Noticed while reviewing llvm#66968.
ldionne
added a commit
that referenced
this pull request
Oct 4, 2023
…TION_POINT (#68238) The _LIBCPP_PSTL_CUSTOMIZATION_POINT macro was assuming that the policy was called _RawPolicy and referencing it by name. It happened to always work but this was definitely accidental and an oversight in the original implementation. This patch fixes that by passing the policy to the macro explicitly. Noticed while reviewing #66968.
|
@philnik777 Last we spoke, the design work around the customization points was still ongoing. Have we locked it in now? |
Kind-of, but not really. We'll probably get more insights while working on this. We're working closely with @AntonRydahl and others to get this backend into libc++. |
ldionne
requested changes
Oct 13, 2023
ldionne
pushed a commit
that referenced
this pull request
Oct 14, 2023
This PR addresses a smaller detail discussed in the code review for #66968. Currently, some functions in the `libc++` PSTL CPU backend have been appended with a digit to indicate the number of input iterator arguments. However, there is no need to change the name for each version as overloading can be used instead. This PR will make the naming more consistent in the the CPU and the proposed OpenMP backend.
ldionne
requested changes
Oct 16, 2023
Comment on lines
57
to
58
| // parallel unsequenced, as it is the only execution policy prohibiting throwing | ||
| // exceptions and allowing SIMD instructions |
There was a problem hiding this comment.
This comment is not technically correct, since exceptions can be thrown from user code even when par_unseq is used.
And unseq also allows for SIMD. So I think we need to explain that we don't have sequencing between the teams (thanks @jdoerfert !).
Also, I think it might be worth promoting this comment to openmp.h since it is not specific for for_each.
blueboxd
pushed a commit
to blueboxd/libcxx
that referenced
this pull request
Oct 17, 2023
This PR addresses a smaller detail discussed in the code review for llvm/llvm-project#66968. Currently, some functions in the `libc++` PSTL CPU backend have been appended with a digit to indicate the number of input iterator arguments. However, there is no need to change the name for each version as overloading can be used instead. This PR will make the naming more consistent in the the CPU and the proposed OpenMP backend. NOKEYCHECK=True GitOrigin-RevId: f2b79ed9c6c858426b15a0374103ab901b5b2ef3
|
When adding the logic to error out when OpenMP is not enabled, I had to add a bit of CMake logic to allow installing the library without getting that same error. On the good side, this also made it possible to enable the OpenMP toolchain when running the tests, if the OpenMP backend was selected. I now found that a few of the tests failed, because I had not thought of handling cases where the base type of the iterator is not trivially copyable. For instance, some of the tests may throw exceptions in the copy constructor. The commit I pushed today makes those tests pass while still allowing my local tests to be executed on the GPU. What do you think of this change? |
|
I have also added a number of tests that verify that vectors of primitive data types are offloaded to the GPU. They use If there are no available devices, the new tests simply don't run: if (omp_get_num_devices() < 1)
return 0;Also, I have introduced a new feature such that they only run when the OpenMP PSTL backend is enabled: // REQUIRES: openmp_pstl_backend |
|
@AntonRydahl I tried to rebase this but there are conflicts that I cannot easily resolve. |
ldionne
force-pushed
the
libcxx_pstl_omp_offload_backend
branch
from
2ecd90a to
f87dfc5
Compare
llvmbot
added
libc++abi
|
@jdoerfert Rebase done. There will probably be many small issues because I couldn't build this locally, but I basically resolved the rebasing issues and reformulated the backend on top of the new dispatching mechanism. |
ldionne
force-pushed
the
libcxx_pstl_omp_offload_backend
branch
from
f87dfc5 to
9594623
Compare
|
Thank you so, so much for spending time fixing this, @ldionne! I will re-run the tests and see if I can replicate the errors from the build bot. |
ldionne
force-pushed
the
libcxx_pstl_omp_offload_backend
branch
from
9594623 to
736262f
Compare
|
Thanks a lot! Bots seem clean, right? FWIW, I can (re-)run stuff on linux machines, at least try. We also have people, incl. a GSOC student who will pick this up now. |
|
I think we need to wait and see if stage3 (generic-pstl-openmp, libcxx-runners-8-set) passes. |
|
@AntonRydahl I don't think we ever had a clean run of the OpenMP bot, did we? IMO the testing + CI story for the OpenMP backend is the only thing blocking this PR from being merged, everything else (like improving the backend) can and should be done in separate PRs so we can capture progress here. |
AntonRydahl
force-pushed
the
libcxx_pstl_omp_offload_backend
branch
from
736262f to
10d408f
Compare
|
The tests should not be able to locate libomptarget.so. Let's see what other failures we get from the build bot. |
ldionne
reviewed
Jun 18, 2024
| __libcpp_is_contiguous_iterator<_ForwardOutIterator>::value && is_trivially_copyable_v<_ValueType>) { | ||
| return std::__rewrap_iter( | ||
| __outit, | ||
| __omp_transform( |
There was a problem hiding this comment.
Suggested change
| __omp_transform( | |
| __pstl::__omp_transform( |
|
Just pushed header files for ranges shift left and shift right for review, I was wondering if the directives and the argument types are correct, a specific format to follow? To test the files in libcxx we would have to run the executable llvm-lit in the openmp test directory? |
|
Thanks for the patch! I wanted to contribute to dev of openMP offload of std algorithms. Is this patch LGTM? and is it safe to base my work on top of this? |
|
Sure will do!
Best,
Vedant
…On Thu, Nov 14, 2024, 11:31 AM Johannes Doerfert ***@***.***> wrote:
You should rebase on top of main, that will solve the issues you're
seeing on the macOS builders.
@xevor11 <https://github.com/xevor11> can you rebase this on top of main,
as @ldionne <https://github.com/ldionne> mentioned?
Also, let's split your additions off into a new PR so we can merge this
first.
—
Reply to this email directly, view it on GitHub
<#66968 (comment)>,
or unsubscribe
<https://github.com/notifications/unsubscribe-auth/ARZYSZ4JDTTLW5ZXQFJR25D2ATM6JAVCNFSM6AAAAAA5AXJ4TCVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDINZXGAZDANZQGU>
.
You are receiving this because you were mentioned.Message ID:
***@***.***>
|
|
For all the conflicting files I just went with the changes made in upstream, since there were no modifications made on my end within these directories or files:
I have rebased on top of main will continue the rest of the work in a different PR as requested |
Did you update the PR after the rebase, or created a new one? |
The PR was not updated I'll create a new one |
|
Hello @AntonRydahl , I am having a difficult time to compile the tests in the PR. I build this PR code with the following command I am trying to compile of the tests with the following command: however, it throws an error about tbb, which is interesting, since i specifically compiled for the Openmp backend. I can fix this error if i add Do you have any clue what is happening here? |
Ping! could someone help with this! |
|
I get this error because the host pstl is included. Adding |
For now, use -fno-exceptions for the compilation, otherwise the host and device symbols are out of whack. That will be resolved in the future but is required now and needs to be documented. |
|
This PR was rebased in #122180, for anyone curious. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
To address some of the good feedback from the recent request for comments, I have added a separate OpenMP offloading backend in
libcxx/include/__algorithm/pstl_backends.To begin with, I have added
std::for_each,std::transform,std::transform_reduce, andstd::fill.I know there are still a lot of things missing, such as LIT tests, but I think we should try to agree on the project structure first.