Add support for Darwin 32-bit and PPC (#5916)

* Do not use 64-bit static_asserts on 32-bit build * TestScan: case for 32-bit * Disable team scratch failing test on 32-bit, for now * CMakeLists: status message instead of a fatal error for 32-bit build * Kokkos_ClockTic: add implementation for ppc * Use KOKKOS_IMPL_32BIT instead of KOKKOS_32_BIT * Add 32bit build to GitHub CI * Limit ScatterView test to 1GB * Fix signed compariosn in TestVector.hpp * Indentation * Signed compare * Disable std::complex<long double> test * Disable repeated_team_reduce and deep_copy_conversion * Update test restriction Co-authored-by: Damien L-G <[email protected]> * Update core/unit_test/TestTeamBasic.hpp * Add assert for sizeof(void*) * Try using integer comparison * Add reference for clock_tic_host implementation * Make FIXME_32BIT more uniform * Print platform information * Fix Kokkos_Core_fwd.hpp * Separate 32bit CI into its own workflow * Add new workflow file * Minimize workflow file * Try LIBDL=ON * Remove ccache commands and LIBDL --------- Co-authored-by: Daniel Arndt <[email protected]> Co-authored-by: Damien L-G <[email protected]>
sandialabs · May 2, 2023 · e5490e1 · e5490e1
1 parent 56ef02c
commit e5490e1
Show file tree

Hide file tree

Showing 14 changed files with 100 additions and 10 deletions.
diff --git a/.github/workflows/continuous-integration-workflow-32bit.yml b/.github/workflows/continuous-integration-workflow-32bit.yml
@@ -0,0 +1,37 @@
+name: github-Linux-32bit
+on: [push, pull_request]
+
+concurrency:
+  group: ${ {github.event_name }}-${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: ${{github.event_name == 'pull_request'}}
+
+jobs:
+  CI-32bit:
+    name: Linux-32bit
+    runs-on: ubuntu-latest
+    container:
+      image: ghcr.io/kokkos/ci-containers/ubuntu:latest
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v3
+      - name: install_multilib
+        run: sudo apt-get update && sudo apt-get install -y gcc-multilib g++-multilib gfortran-multilib
+      - name: Configure Kokkos
+        run: |
+          cmake -B builddir \
+            -DKokkos_ENABLE_OPENMP=ON \
+            -DKokkos_ENABLE_TESTS=ON \
+            -DKokkos_ENABLE_BENCHMARKS=ON \
+            -DKokkos_ENABLE_EXAMPLES=ON \
+            -DKokkos_ENABLE_DEPRECATED_CODE_4=ON \
+            -DKokkos_ENABLE_DEPRECATION_WARNINGS=OFF \
+            -DKokkos_ENABLE_COMPILER_WARNINGS=ON \
+            -DCMAKE_CXX_FLAGS="-Werror -m32 -DKOKKOS_IMPL_32BIT" \
+            -DCMAKE_CXX_COMPILER=g++ \
+            -DCMAKE_BUILD_TYPE=RelWithDebInfo
+      - name: Build
+        run: |
+          cmake --build builddir --parallel 2
+      - name: Tests
+        working-directory: builddir
+        run: ctest --output-on-failure
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -138,7 +138,12 @@ IF (NOT CMAKE_SIZEOF_VOID_P)
     MESSAGE(FATAL_ERROR "Kokkos did not configure correctly and failed to validate compiler. The most likely cause is linkage errors during CMake compiler validation. Please consult the CMake error log shown below for the exact error during compiler validation")
   ENDIF()
 ELSEIF (NOT CMAKE_SIZEOF_VOID_P EQUAL 8)
-  MESSAGE(FATAL_ERROR "Kokkos assumes a 64-bit build; i.e., 8-byte pointers, but found ${CMAKE_SIZEOF_VOID_P}-byte pointers instead")
+  IF(CMAKE_SIZEOF_VOID_P EQUAL 4)
+    MESSAGE(WARNING "32-bit builds are experimental and not officially supported.")
+    SET(KOKKOS_IMPL_32BIT ON)
+  ELSE()
+    MESSAGE(FATAL_ERROR "Kokkos assumes a 64-bit build, i.e., 8-byte pointers, but found ${CMAKE_SIZEOF_VOID_P}-byte pointers instead;")
+  ENDIF()
 ENDIF()
 
 

diff --git a/cmake/KokkosCore_config.h.in b/cmake/KokkosCore_config.h.in
@@ -119,3 +119,5 @@
 #cmakedefine KOKKOS_ARCH_NAVI
 #cmakedefine KOKKOS_ARCH_NAVI1030
 #cmakedefine KOKKOS_ARCH_NAVI1100
+
+#cmakedefine KOKKOS_IMPL_32BIT
diff --git a/containers/unit_tests/TestScatterView.hpp b/containers/unit_tests/TestScatterView.hpp
@@ -726,9 +726,9 @@ void test_scatter_view(int64_t n) {
   }
 #endif
   // with hundreds of threads we were running out of memory.
-  // limit (n) so that duplication doesn't exceed 4GB
+  // limit (n) so that duplication doesn't exceed 1GB
   constexpr std::size_t maximum_allowed_total_bytes =
-      4ull * 1024ull * 1024ull * 1024ull;
+      1ull * 1024ull * 1024ull * 1024ull;
   std::size_t const maximum_allowed_copy_bytes =
       maximum_allowed_total_bytes /
       std::size_t(execution_space().concurrency());

diff --git a/containers/unit_tests/TestVector.hpp b/containers/unit_tests/TestVector.hpp
@@ -52,7 +52,7 @@ struct test_vector_insert {
     it_return = a.insert(it, n + 5, scalar_type(5));
 
     ASSERT_EQ(a.size(), n + 1 + n + 5);
-    ASSERT_EQ(std::distance(it_return, a.begin() + 17), 0u);
+    ASSERT_EQ(std::distance(it_return, a.begin() + 17), 0);
 
     Vector b;
 
@@ -65,7 +65,7 @@ struct test_vector_insert {
     it_return = a.insert(it, b.begin(), b.end());
 
     ASSERT_EQ(a.size(), n + 1 + n + 5 + 7);
-    ASSERT_EQ(std::distance(it_return, a.begin() + 27 + n), 0u);
+    ASSERT_EQ(std::distance(it_return, a.begin() + 27 + n), 0);
 
     // Testing insert at end via all three function interfaces
     a.insert(a.end(), 11);

diff --git a/core/src/Kokkos_Core_fwd.hpp b/core/src/Kokkos_Core_fwd.hpp
@@ -34,11 +34,15 @@
 #endif
 
 //----------------------------------------------------------------------------
-// Have assumed a 64bit build (8byte pointers) throughout the code base.
-
+// Have assumed a 64-bit build (8-byte pointers) throughout the code base.
+// 32-bit build allowed but unsupported.
+#ifdef KOKKOS_IMPL_32BIT
+static_assert(sizeof(void *) == 4,
+              "Kokkos assumes 64-bit build; i.e., 4-byte pointers");
+#else
 static_assert(sizeof(void *) == 8,
               "Kokkos assumes 64-bit build; i.e., 8-byte pointers");
-
+#endif
 //----------------------------------------------------------------------------
 
 namespace Kokkos {

diff --git a/core/src/impl/Kokkos_ClockTic.hpp b/core/src/impl/Kokkos_ClockTic.hpp
@@ -90,6 +90,27 @@ KOKKOS_IMPL_HOST_FUNCTION inline uint64_t clock_tic_host() noexcept {
 
   return (uint64_t)cycles;
 
+#elif defined(__ppc__)
+  // see : pages.cs.wisc.edu/~legault/miniproj-736.pdf or
+  // cmssdt.cern.ch/lxr/source/FWCore/Utilities/interface/HRRealTime.h
+
+  uint64_t result = 0;
+  uint32_t upper, lower, tmp;
+
+  __asm__ volatile(
+      "0: \n"
+      "\tmftbu %0     \n"
+      "\tmftb  %1     \n"
+      "\tmftbu %2     \n"
+      "\tcmpw  %2, %0 \n"
+      "\tbne   0b     \n"
+      : "=r"(upper), "=r"(lower), "=r"(tmp));
+  result = upper;
+  result = result << 32;
+  result = result | lower;
+
+  return (result);
+
 #else
 
   return std::chrono::high_resolution_clock::now().time_since_epoch().count();

diff --git a/core/src/impl/Kokkos_Core.cpp b/core/src/impl/Kokkos_Core.cpp
@@ -772,6 +772,12 @@ void pre_initialize_internal(const Kokkos::InitializationSettings& settings) {
 #else
   declare_configuration_metadata("architecture", "GPU architecture", "none");
 #endif
+
+#ifdef KOKKOS_IMPL_32BIT
+  declare_configuration_metadata("architecture", "platform", "32bit");
+#else
+  declare_configuration_metadata("architecture", "platform", "64bit");
+#endif
 }
 
 void post_initialize_internal(const Kokkos::InitializationSettings& settings) {

diff --git a/core/src/impl/Kokkos_StringManipulation.hpp b/core/src/impl/Kokkos_StringManipulation.hpp
@@ -173,7 +173,7 @@ KOKKOS_FUNCTION constexpr to_chars_result to_chars_i(char *first, char *last,
       unsigned_val = Unsigned(~value) + Unsigned(1);
     }
   }
-  unsigned int const len = to_chars_len(unsigned_val);
+  std::ptrdiff_t const len = to_chars_len(unsigned_val);
   if (last - first < len) {
     return {last, errc::value_too_large};
   }

diff --git a/core/src/impl/Kokkos_TaskBase.hpp b/core/src/impl/Kokkos_TaskBase.hpp
@@ -206,6 +206,7 @@ class TaskBase {
 // the number of full task types that fit into a cache line.  We'll leave it
 // here for now, though, since we're probably going to be ripping all of the
 // old TaskBase stuff out eventually anyway.
+#ifndef KOKKOS_IMPL_32BIT
 constexpr size_t unpadded_task_base_size = 44 + 2 * sizeof(int16_t);
 // don't forget padding:
 constexpr size_t task_base_misalignment =
@@ -229,7 +230,7 @@ static constexpr
 
 static_assert(sizeof(TaskBase) == expected_task_base_size,
               "Verifying expected sizeof(TaskBase)");
-
+#endif
 // </editor-fold> end Verify the size of TaskBase is as expected }}}2
 //------------------------------------------------------------------------------
 

diff --git a/core/unit_test/TestComplex.hpp b/core/unit_test/TestComplex.hpp
@@ -519,9 +519,13 @@ TEST(TEST_CATEGORY, complex_operations_arithmetic_types_overloads) {
   ASSERT_EQ(Kokkos::conj(1), Kokkos::complex<double>(1));
   ASSERT_EQ(Kokkos::conj(2.f), Kokkos::complex<float>(2.f));
   ASSERT_EQ(Kokkos::conj(3.), Kokkos::complex<double>(3.));
+// long double has size 12 but Kokkos::complex requires 2*sizeof(T) to be a
+// power of two.
+#ifndef KOKKOS_IMPL_32BIT
   ASSERT_EQ(Kokkos::conj(4.l), Kokkos::complex<long double>(4.l));
   static_assert((
       std::is_same<decltype(Kokkos::conj(1)), Kokkos::complex<double>>::value));
+#endif
   static_assert((std::is_same<decltype(Kokkos::conj(2.f)),
                               Kokkos::complex<float>>::value));
   static_assert((std::is_same<decltype(Kokkos::conj(3.)),

diff --git a/core/unit_test/TestDeepCopyAlignment.hpp b/core/unit_test/TestDeepCopyAlignment.hpp
@@ -341,6 +341,9 @@ struct TestDeepCopyScalarConversion {
 }  // namespace Impl
 
 TEST(TEST_CATEGORY, deep_copy_conversion) {
+#ifdef KOKKOS_IMPL_32BIT
+  GTEST_SKIP() << "Failing KOKKOS_IMPL_32BIT";  // FIXME_32BIT
+#endif
   int64_t N0 = 19381;
   int64_t N1 = 17;
 

diff --git a/core/unit_test/TestTeamBasic.hpp b/core/unit_test/TestTeamBasic.hpp
@@ -182,6 +182,9 @@ struct LargeTeamScratchFunctor {
 };
 
 TEST(TEST_CATEGORY, large_team_scratch_size) {
+#ifdef KOKKOS_IMPL_32BIT
+  GTEST_SKIP() << "Fails on 32-bit";  // FIXME_32BIT
+#endif
   const int level   = 1;
   const int n_teams = 1;
 

diff --git a/core/unit_test/TestTeamReductionScan.hpp b/core/unit_test/TestTeamReductionScan.hpp
@@ -113,6 +113,10 @@ TEST(TEST_CATEGORY, repeated_team_reduce) {
                     "properly implemented";
 #endif
 
+#ifdef KOKKOS_IMPL_32BIT
+  GTEST_SKIP() << "Failing KOKKOS_IMPL_32BIT";  // FIXME_32BIT
+#endif
+
   TestRepeatedTeamReduce<TEST_EXECSPACE>();
 }