Support NVidia HPC SDK (PGI)

.github/CONTRIBUTING.md

@@ -164,6 +164,29 @@ name, pre-commit):
 pre-commit install
 ```
 
+### Build recipes
+
+This builds with the Intel compiler (assuming it is in your path, along with a
+recent CMake and Python 3):
+
+```bash
+python3 -m venv venv
+. venv/bin/activate
+pip install pytest
+cmake -S . -B build-intel -DCMAKE_CXX_COMPILER=$(which icpc) -DDOWNLOAD_CATCH=ON -DDOWNLOAD_EIGEN=ON -DPYBIND11_WERROR=ON
+```
+
+This will test the PGI compilers:
+
+```bash
+docker run --rm -it -v $PWD:/pybind11 nvcr.io/hpc/pgi-compilers:ce
+apt-get update && apt-get install -y python3-dev python3-pip python3-pytest
+wget -qO- "https://cmake.org/files/v3.18/cmake-3.18.1-Linux-x86_64.tar.gz" | tar --strip-components=1 -xz -C /usr/local
+cmake -S pybind11/ -B build
+cmake --build build
+```
+
+
 [pre-commit]: https://pre-commit.com
 [pybind11.readthedocs.org]: http://pybind11.readthedocs.org/en/latest
 [issue tracker]: https://github.com/pybind/pybind11/issues

.github/workflows/ci.yml

@@ -69,7 +69,6 @@ jobs:
             python: 3.9-dev
             arch: x64
 
-
     name: "🐍 ${{ matrix.python }} • ${{ matrix.runs-on }} • ${{ matrix.arch }} ${{ matrix.args }}"
     runs-on: ${{ matrix.runs-on }}
     continue-on-error: ${{ endsWith(matrix.python, 'dev') }}
@@ -196,6 +195,109 @@ jobs:
     - name: Interface test
       run: cmake --build build --target test_cmake_build
 
+
+  cuda:
+    runs-on: ubuntu-latest
+    name: "🐍 3.8 • CUDA 11 • Ubuntu 20.04"
+    container: nvidia/cuda:11.0-devel-ubuntu20.04
+
+    steps:
+    - uses: actions/checkout@v2
+
+    # tzdata will try to ask for the timezone, so set the DEBIAN_FRONTEND
+    - name: Install 🐍 3
+      run: apt-get update && DEBIAN_FRONTEND="noninteractive" apt-get install -y cmake git python3-dev python3-pytest python3-numpy
+
+    - name: Configure
+      run: cmake -S . -B build -DPYBIND11_CUDA_TESTS=ON -DPYBIND11_WERROR=ON -DDOWNLOAD_CATCH=ON
+
+    - name: Build
+      run: cmake --build build -j2 -v
+
+    - name: Python tests
+      run: cmake --build build --target pytest
+
+
+  centos-nvhpc:
+    runs-on: ubuntu-latest
+    name: "🐍 3 • CentOS8/PGI • x64"
+    container: "centos:8"
+
+    steps:
+    - uses: actions/checkout@v2
+
+    - name: Add Python 3 and a few requirements
+      run: yum update -y && yum install -y git python3-devel python3-numpy python3-pytest make environment-modules
+
+    - name: Install CMake with pip
+      run: |
+        python3 -m pip install --upgrade pip
+        python3 -m pip install cmake --prefer-binary
+
+    - name: Install NVidia HPC SDK
+      run: yum -y install https://developer.download.nvidia.com/hpc-sdk/nvhpc-20-7-20.7-1.x86_64.rpm  https://developer.download.nvidia.com/hpc-sdk/nvhpc-2020-20.7-1.x86_64.rpm
+
+    - name: Configure
+      shell: bash
+      run: |
+        source /etc/profile.d/modules.sh
+        module load /opt/nvidia/hpc_sdk/modulefiles/nvhpc/20.7
+        cmake -S . -B build -DDOWNLOAD_CATCH=ON -DCMAKE_CXX_STANDARD=14 -DPYTHON_EXECUTABLE=$(python3 -c "import sys; print(sys.executable)")
+
+    - name: Build
+      run: cmake --build build -j 2
+
+    - name: Python tests
+      run: cmake --build build --target pytest
+
+    - name: C++ tests
+      run: cmake --build build --target cpptest
+
+    - name: Interface test
+      run: cmake --build build --target test_cmake_build
+
+  centos-nvhpc7:
+    runs-on: ubuntu-latest
+    name: "🐍 3 • CentOS7/PGI • x64"
+    container: "centos:7"
+
+    steps:
+    - uses: actions/checkout@v2
+
+    - name: Add Python 3 and a few requirements
+      run: yum update -y && yum install -y git python3-devel  make environment-modules
+
+    - name: Install CMake with pip
+      run: |
+        python3 -m pip install --upgrade pip
+        python3 -m pip install pytest cmake --prefer-binary
+
+    - name: Install NVidia HPC SDK
+      run: yum -y install https://developer.download.nvidia.com/hpc-sdk/nvhpc-20-7-20.7-1.x86_64.rpm  https://developer.download.nvidia.com/hpc-sdk/nvhpc-2020-20.7-1.x86_64.rpm
+
+    - name: Configure
+      shell: bash
+      run: |
+        source /etc/profile.d/modules.sh
+        module load /opt/nvidia/hpc_sdk/modulefiles/nvhpc/20.7
+        cmake -S . -B build -DDOWNLOAD_CATCH=ON \
+                            -DCMAKE_CXX_STANDARD=11 \
+                            -DPYTHON_EXECUTABLE=$(python3 -c "import sys; print(sys.executable)") \
+                            -DCMAKE_CXX_FLAGS="-Wc,--pending_instantiations=0" \
+                            -DPYBIND11_TEST_FILTER="test_smart_ptr.cpp;test_virtual_functions.cpp"
+
+    - name: Build
+      run: cmake --build build -j 2
+
+    - name: Python tests
+      run: cmake --build build --target pytest
+
+    - name: C++ tests
+      run: cmake --build build --target cpptest
+
+    - name: Interface test
+      run: cmake --build build --target test_cmake_build
+
   gcc:
     runs-on: ubuntu-latest
     strategy:
@@ -243,6 +345,7 @@ jobs:
     - name: Interface test
       run: cmake --build build --target test_cmake_build
 
+
   centos:
     runs-on: ubuntu-latest
     strategy:
@@ -289,27 +392,6 @@ jobs:
     - name: Interface test
       run: cmake --build build --target test_cmake_build
 
-  cuda:
-    runs-on: ubuntu-latest
-    name: "🐍 3.8 • CUDA 11 • Ubuntu 20.04"
-    container: nvidia/cuda:11.0-devel-ubuntu20.04
-
-    steps:
-    - uses: actions/checkout@v2
-
-    # tzdata will try to ask for the timezone, so set the DEBIAN_FRONTEND
-    - name: Install 🐍 3
-      run: apt-get update && DEBIAN_FRONTEND="noninteractive" apt-get install -y cmake python3-dev python3-pytest
-
-    - name: Configure
-      run: cmake -S . -B build -DPYBIND11_CUDA_TESTS=ON -DPYBIND11_WERROR=ON -DDOWNLOAD_CATCH=ON
-
-    - name: Build
-      run: cmake --build build -j2 -v
-
-    - name: Python tests
-      run: cmake --build build --target pytest
-
 
   install-classic:
     name: "🐍 3.5 • Debian • x86 •  Install"

README.md

@@ -99,6 +99,7 @@ In addition to the core functionality, pybind11 provides some extra goodies:
    v2.0 and a [workaround][intel-15-workaround])
 5. Cygwin/GCC (tested on 2.5.1)
 6. NVCC (CUDA 11 tested)
+7. NVIDIA PGI (20.7 tested)
 
 ## About
 

include/pybind11/detail/common.h

@@ -154,6 +154,7 @@
 #include <vector>
 #include <string>
 #include <stdexcept>
+#include <exception>
 #include <unordered_set>
 #include <unordered_map>
 #include <memory>
@@ -501,8 +502,16 @@ template <bool... Bs> using select_indices = typename select_indices_impl<index_
 template <bool B> using bool_constant = std::integral_constant<bool, B>;
 template <typename T> struct negation : bool_constant<!T::value> { };
 
+// PGI cannot detect operator delete with the "compatible" void_t impl, so
+// using the new one (C++14 defect, so generally works on newer compilers, even
+// if not in C++17 mode)
+#if defined(__PGIC__)
+template<typename... > using void_t = void;
+#else
 template <typename...> struct void_t_impl { using type = void; };
 template <typename... Ts> using void_t = typename void_t_impl<Ts...>::type;
+#endif
+
 
 /// Compile-time all/any/none of that check the boolean value of all template types
 #if defined(__cpp_fold_expressions) && !(defined(_MSC_VER) && (_MSC_VER < 1916))

tests/CMakeLists.txt

@@ -16,6 +16,29 @@ else()
   cmake_policy(VERSION 3.18)
 endif()
 
+# Only needed for CMake < 3.5 support
+include(CMakeParseArguments)
+
+# Filter out items; print an optional message if any items filtered
+#
+# Usage:
+#   pybind11_filter_tests(LISTNAME file1.cpp file2.cpp ... MESSAGE "")
+#
+macro(PYBIND11_FILTER_TESTS LISTNAME)
+  cmake_parse_arguments(ARG "" "MESSAGE" "" ${ARGN})
+  set(PYBIND11_FILTER_TESTS_FOUND OFF)
+  foreach(filename IN LISTS ARG_UNPARSED_ARGUMENTS)
+    list(FIND ${LISTNAME} ${filename} _FILE_FOUND)
+    if(_FILE_FOUND GREATER -1)
+      list(REMOVE_AT ${LISTNAME} ${_FILE_FOUND})
+      set(PYBIND11_FILTER_TESTS_FOUND ON)
+    endif()
+  endforeach()
+  if(PYBIND11_FILTER_TESTS_FOUND AND ARG_MESSAGE)
+    message(STATUS "${ARG_MESSAGE}")
+  endif()
+endmacro()
+
 # New Python support
 if(DEFINED Python_EXECUTABLE)
   set(PYTHON_EXECUTABLE "${Python_EXECUTABLE}")
@@ -34,6 +57,9 @@ option(PYBIND11_CUDA_TESTS "Enable building CUDA tests (requires CMake 3.12+)" O
 set(PYBIND11_TEST_OVERRIDE
     ""
     CACHE STRING "Tests from ;-separated list of *.cpp files will be built instead of all tests")
+set(PYBIND11_TEST_FILTER
+    ""
+    CACHE STRING "Tests from ;-separated list of *.cpp files will be removed from all tests")
 
 if(CMAKE_CURRENT_SOURCE_DIR STREQUAL CMAKE_SOURCE_DIR)
   # We're being loaded directly, i.e. not via add_subdirectory, so make this
@@ -106,21 +132,23 @@ if(PYBIND11_TEST_OVERRIDE)
   set(PYBIND11_TEST_FILES ${PYBIND11_TEST_OVERRIDE})
 endif()
 
-# Skip test_async for Python < 3.5
-list(FIND PYBIND11_TEST_FILES test_async.cpp PYBIND11_TEST_FILES_ASYNC_I)
-if((PYBIND11_TEST_FILES_ASYNC_I GREATER -1) AND (PYTHON_VERSION VERSION_LESS 3.5))
-  message(STATUS "Skipping test_async because Python version ${PYTHON_VERSION} < 3.5")
-  list(REMOVE_AT PYBIND11_TEST_FILES ${PYBIND11_TEST_FILES_ASYNC_I})
+# You can also filter tests:
+if(PYBIND11_TEST_FILTER)
+  pybind11_filter_tests(PYBIND11_TEST_FILES ${PYBIND11_TEST_FILTER})
+endif()
+
+if(PYTHON_VERSION VERSION_LESS 3.5)
+  pybind11_filter_tests(PYBIND11_TEST_FILES test_async.cpp MESSAGE
+                        "Skipping test_async on Python 2")
 endif()
 
 # Skip tests for CUDA check:
 # /pybind11/tests/test_constants_and_functions.cpp(125):
 #   error: incompatible exception specifications
-list(FIND PYBIND11_TEST_FILES test_constants_and_functions.cpp PYBIND11_TEST_FILES_CAF_I)
-if((PYBIND11_TEST_FILES_CAF_I GREATER -1) AND PYBIND11_CUDA_TESTS)
-  message(
-    STATUS "Skipping test_constants_and_functions due to incompatible exception specifications")
-  list(REMOVE_AT PYBIND11_TEST_FILES ${PYBIND11_TEST_FILES_CAF_I})
+if(PYBIND11_CUDA_TESTS)
+  pybind11_filter_tests(
+    PYBIND11_TEST_FILES test_constants_and_functions.cpp MESSAGE
+    "Skipping test_constants_and_functions due to incompatible exception specifications")
 endif()
 
 string(REPLACE ".cpp" ".py" PYBIND11_PYTEST_FILES "${PYBIND11_TEST_FILES}")

tests/test_factory_constructors.cpp

@@ -11,6 +11,7 @@
 #include "pybind11_tests.h"
 #include "constructor_stats.h"
 #include <cmath>
+#include <new>
 
 // Classes for testing python construction via C++ factory function:
 // Not publicly constructible, copyable, or movable:
@@ -298,11 +299,12 @@ TEST_SUBMODULE(factory_constructors, m) {
         static void *operator new(size_t, void *p) { py::print("noisy placement new"); return p; }
         static void operator delete(void *p, size_t) { py::print("noisy delete"); ::operator delete(p); }
         static void operator delete(void *, void *) { py::print("noisy placement delete"); }
-#if defined(_MSC_VER) && _MSC_VER < 1910
+#       if defined(_MSC_VER) && _MSC_VER < 1910
         // MSVC 2015 bug: the above "noisy delete" isn't invoked (fixed in MSVC 2017)
-        static void operator delete(void *p) { py::print("noisy delete"); ::operator delete(p); }
-#endif
+            static void operator delete(void *p) { py::print("noisy delete"); ::operator delete(p); }
+#       endif
     };
+
     py::class_<NoisyAlloc>(m, "NoisyAlloc")
         // Since these overloads have the same number of arguments, the dispatcher will try each of
         // them until the arguments convert.  Thus we can get a pre-allocation here when passing a

tests/test_factory_constructors.py

@@ -336,10 +336,10 @@ def strip_comments(s):
     return re.sub(r'\s+#.*', '', s)
 
 
-def test_reallocations(capture, msg):
+def test_reallocation_a(capture, msg):
     """When the constructor is overloaded, previous overloads can require a preallocated value.
     This test makes sure that such preallocated values only happen when they might be necessary,
-    and that they are deallocated properly"""
+    and that they are deallocated properly."""
 
     pytest.gc_collect()
 
@@ -353,6 +353,9 @@ def test_reallocations(capture, msg):
         ~NoisyAlloc()
         noisy delete
     """
+
+
+def test_reallocation_b(capture, msg):
     with capture:
         create_and_destroy(1.5)
     assert msg(capture) == strip_comments("""
@@ -365,6 +368,8 @@ def test_reallocations(capture, msg):
         noisy delete   # operator delete
     """)
 
+
+def test_reallocation_c(capture, msg):
     with capture:
         create_and_destroy(2, 3)
     assert msg(capture) == strip_comments("""
@@ -375,6 +380,8 @@ def test_reallocations(capture, msg):
         noisy delete   # operator delete
     """)
 
+
+def test_reallocation_d(capture, msg):
     with capture:
         create_and_destroy(2.5, 3)
     assert msg(capture) == strip_comments("""
@@ -386,6 +393,8 @@ def test_reallocations(capture, msg):
         noisy delete   # operator delete
     """)
 
+
+def test_reallocation_e(capture, msg):
     with capture:
         create_and_destroy(3.5, 4.5)
     assert msg(capture) == strip_comments("""
@@ -397,6 +406,8 @@ def test_reallocations(capture, msg):
         noisy delete   # operator delete
     """)
 
+
+def test_reallocation_f(capture, msg):
     with capture:
         create_and_destroy(4, 0.5)
     assert msg(capture) == strip_comments("""
@@ -409,6 +420,8 @@ def test_reallocations(capture, msg):
         noisy delete   # operator delete
     """)
 
+
+def test_reallocation_g(capture, msg):
     with capture:
         create_and_destroy(5, "hi")
     assert msg(capture) == strip_comments("""

tests/test_smart_ptr.py

@@ -1,7 +1,8 @@
 # -*- coding: utf-8 -*-
 import pytest
-from pybind11_tests import smart_ptr as m
-from pybind11_tests import ConstructorStats
+
+m = pytest.importorskip("pybind11_tests.smart_ptr")
+from pybind11_tests import ConstructorStats  # noqa: E402
 
 
 def test_smart_ptr(capture):