Introduce global setup/cleanup in APIs

examples/01_1DFFT/01_1DFFT.cpp

@@ -13,6 +13,7 @@ using View1D = Kokkos::View<T*, execution_space>;
 
 int main(int argc, char* argv[]) {
   Kokkos::initialize(argc, argv);
+  KokkosFFT::initialize();
   {
     const int n0 = 128;
     const Kokkos::complex<double> z(1.0, 1.0);
@@ -44,6 +45,7 @@ int main(int argc, char* argv[]) {
     KokkosFFT::irfft(exec, xc2r, xc2r_hat);
     exec.fence();
   }
+  KokkosFFT::finalize();
   Kokkos::finalize();
 
   return 0;

examples/02_2DFFT/02_2DFFT.cpp

@@ -13,6 +13,7 @@ using View2D = Kokkos::View<T**, execution_space>;
 
 int main(int argc, char* argv[]) {
   Kokkos::initialize(argc, argv);
+  KokkosFFT::initialize();
   {
     const int n0 = 128, n1 = 128;
     const Kokkos::complex<double> z(1.0, 1.0);
@@ -44,6 +45,7 @@ int main(int argc, char* argv[]) {
     KokkosFFT::irfft2(exec, xc2r, xc2r_hat);
     exec.fence();
   }
+  KokkosFFT::finalize();
   Kokkos::finalize();
 
   return 0;

examples/03_NDFFT/03_NDFFT.cpp

@@ -15,6 +15,7 @@ using axis_type = KokkosFFT::axis_type<DIM>;
 
 int main(int argc, char* argv[]) {
   Kokkos::initialize(argc, argv);
+  KokkosFFT::initialize();
   {
     const int n0 = 128, n1 = 128, n2 = 16;
     const Kokkos::complex<double> z(1.0, 1.0);
@@ -46,6 +47,7 @@ int main(int argc, char* argv[]) {
     KokkosFFT::irfftn(exec, xc2r, xc2r_hat, axis_type<3>{-3, -2, -1});
     exec.fence();
   }
+  KokkosFFT::finalize();
   Kokkos::finalize();
 
   return 0;

examples/04_batchedFFT/04_batchedFFT.cpp

@@ -13,6 +13,7 @@ using View3D = Kokkos::View<T***, execution_space>;
 
 int main(int argc, char* argv[]) {
   Kokkos::initialize(argc, argv);
+  KokkosFFT::initialize();
   {
     const int n0 = 128, n1 = 128, n2 = 16;
     const Kokkos::complex<double> z(1.0, 1.0);
@@ -48,6 +49,7 @@ int main(int argc, char* argv[]) {
                      /*axis=*/-1);
     exec.fence();
   }
+  KokkosFFT::finalize();
   Kokkos::finalize();
 
   return 0;

examples/05_1DFFT_HOST_DEVICE/05_1DFFT_HOST_DEVICE.cpp

@@ -16,6 +16,7 @@ using HostView1D = Kokkos::View<T*, host_execution_space>;
 
 int main(int argc, char* argv[]) {
   Kokkos::initialize(argc, argv);
+  KokkosFFT::initialize();
   {
     const int n0 = 128;
     const Kokkos::complex<double> z(1.0, 1.0);
@@ -79,6 +80,7 @@ int main(int argc, char* argv[]) {
     host_exec.fence();
 #endif
   }
+  KokkosFFT::finalize();
   Kokkos::finalize();
 
   return 0;

examples/06_1DFFT_reuse_plans/06_1DFFT_reuse_plans.cpp

@@ -13,6 +13,7 @@ using View1D = Kokkos::View<T*, execution_space>;
 
 int main(int argc, char* argv[]) {
   Kokkos::initialize(argc, argv);
+  KokkosFFT::initialize();
   {
     const int n0 = 128;
     const Kokkos::complex<double> z(1.0, 1.0);
@@ -54,6 +55,7 @@ int main(int argc, char* argv[]) {
     irfft_plan.execute(xc2r, xc2r_hat);
     exec.fence();
   }
+  KokkosFFT::finalize();
   Kokkos::finalize();
 
   return 0;

examples/07_unmanaged_views/07_unmanaged_views.cpp

@@ -25,6 +25,7 @@ using shape_type = KokkosFFT::shape_type<DIM>;
 
 int main(int argc, char* argv[]) {
   Kokkos::initialize(argc, argv);
+  KokkosFFT::initialize();
   {
     const int n0 = 128, n1 = 128, n2 = 16;
     const Kokkos::complex<double> z(1.0, 1.0);
@@ -63,6 +64,7 @@ int main(int argc, char* argv[]) {
                      KokkosFFT::Normalization::backward, shape);
     exec.fence();
   }
+  KokkosFFT::finalize();
   Kokkos::finalize();
 
   return 0;

examples/08_inplace_FFT/08_inplace_FFT.cpp

@@ -14,6 +14,7 @@ using RightView2D = Kokkos::View<T **, Kokkos::LayoutRight, execution_space>;
 
 int main(int argc, char *argv[]) {
   Kokkos::initialize(argc, argv);
+  KokkosFFT::initialize();
   {
     const int n0 = 128, n1 = 128;
     const Kokkos::complex<double> z(1.0, 1.0);
@@ -69,6 +70,7 @@ int main(int argc, char *argv[]) {
 
     exec.fence();
   }
+  KokkosFFT::finalize();
   Kokkos::finalize();
 
   return 0;

fft/src/CMakeLists.txt

@@ -2,15 +2,15 @@
 #
 # SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception
 
-add_library(fft INTERFACE)
+add_library(fft STATIC KokkosFFT_Core.cpp)
 
 target_link_libraries(fft
-    INTERFACE
+    PUBLIC
     common
     Kokkos::kokkos
 )
 
-target_include_directories(fft INTERFACE
+target_include_directories(fft PUBLIC
                            $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}>
                            $<INSTALL_INTERFACE:${INSTALL_INCLUDEDIR}>
 )

fft/src/KokkosFFT.hpp

@@ -10,5 +10,6 @@
 #include "KokkosFFT_Helpers.hpp"
 #include "KokkosFFT_Plans.hpp"
 #include "KokkosFFT_Transform.hpp"
+#include "KokkosFFT_Core.hpp"
 
 #endif

fft/src/KokkosFFT_Core.cpp

@@ -0,0 +1,79 @@
+// SPDX-FileCopyrightText: (C) The kokkos-fft development team, see COPYRIGHT.md file
+//
+// SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception
+
+#include <Kokkos_Core.hpp>
+#include "KokkosFFT_Core.hpp"
+#include "KokkosFFT_default_types.hpp"
+
+namespace {
+bool g_is_initialized = false;
+bool g_is_finalized   = false;
+
+bool kokkosfft_initialize_was_called() {
+  return KokkosFFT::is_initialized() || KokkosFFT::is_finalized();
+}
+bool kokkosfft_finalize_was_called() { return KokkosFFT::is_finalized(); }
+
+void initialize_internal() {
+  KokkosFFT::Impl::initialize_host();
+  KokkosFFT::Impl::initialize_device();
+}
+
+void finalize_internal() {
+  KokkosFFT::Impl::finalize_host();
+  KokkosFFT::Impl::finalize_device();
+}
+}  // namespace
+
+[[nodiscard]] bool KokkosFFT::is_initialized() noexcept {
+  return g_is_initialized;
+}
+
+[[nodiscard]] bool KokkosFFT::is_finalized() noexcept { return g_is_finalized; }
+
+void KokkosFFT::initialize() {
+  if (!(Kokkos::is_initialized() || Kokkos::is_finalized())) {
+    Kokkos::abort(
+        "Error: KokkosFFT::initialize() must not be called before initializing "
+        "Kokkos.\n");
+  }
+  if (Kokkos::is_finalized()) {
+    Kokkos::abort(
+        "Error: KokkosFFT::initialize() must not be called after finalizing "
+        "Kokkos.\n");
+  }
+  if (kokkosfft_initialize_was_called()) {
+    Kokkos::abort(
+        "Error: KokkosFFT::initialize() has already been called."
+        " KokkosFFT can be initialized at most once.\n");
+  }
+  initialize_internal();
+  g_is_initialized = true;
+}
+
+void KokkosFFT::finalize() {
+  if (!(Kokkos::is_initialized() || Kokkos::is_finalized())) {
+    Kokkos::abort(
+        "Error: KokkosFFT::finalize() may only be called after Kokkos has been "
+        "initialized.\n");
+  }
+  if (Kokkos::is_finalized()) {
+    Kokkos::abort(
+        "Error: KokkosFFT::finalize() must be called before finalizing "
+        "Kokkos.\n");
+  }
+
+  if (!kokkosfft_initialize_was_called()) {
+    Kokkos::abort(
+        "Error: KokkosFFT::finalize() may only be called after KokkosFFT has "
+        "been "
+        "initialized.\n");
+  }
+  if (kokkosfft_finalize_was_called()) {
+    Kokkos::abort("Error: KokkosFFT::finalize() has already been called.\n");
+  }
+  finalize_internal();
+  g_is_initialized = false;
+  g_is_finalized   = true;
+}

fft/src/KokkosFFT_Core.hpp

@@ -0,0 +1,17 @@
+// SPDX-FileCopyrightText: (C) The kokkos-fft development team, see COPYRIGHT.md file
+//
+// SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception
+
+#ifndef KOKKOSFFT_CORE_HPP
+#define KOKKOSFFT_CORE_HPP
+
+namespace KokkosFFT {
+void initialize();
+
+[[nodiscard]] bool is_initialized() noexcept;
+[[nodiscard]] bool is_finalized() noexcept;
+
+void finalize();
+}  // namespace KokkosFFT
+
+#endif

fft/src/KokkosFFT_Cuda_types.hpp

@@ -238,7 +238,13 @@ template <typename ExecutionSpace>
 auto direction_type(Direction direction) {
   return direction == Direction::forward ? CUFFT_FORWARD : CUFFT_INVERSE;
 }
+
+inline void initialize_host() {}
+inline void finalize_host() {}
 #endif
+
+inline void initialize_device() {}
+inline void finalize_device() {}
 }  // namespace Impl
 }  // namespace KokkosFFT
 

fft/src/KokkosFFT_FFTW_Types.hpp

@@ -138,6 +138,28 @@ struct ScopedFFTWPlan {
   }
 };
 
+#if defined(KOKKOS_ENABLE_OPENMP) || defined(KOKKOS_ENABLE_THREADS)
+inline void initialize_host() {
+  fftwf_init_threads();
+  fftw_init_threads();
+}
+inline void finalize_host() {
+  fftwf_cleanup_threads();
+  fftw_cleanup_threads();
+}
+#else
+inline void initialize_host() {}
+inline void finalize_host() {}
+#endif
+
+// If non of device backend is enabled, then FFTW is responsible
+// for the device cleanup. Otherwise, device backend will cleanup
+#if !(defined(KOKKOS_ENABLE_CUDA) || defined(KOKKOS_ENABLE_HIP) || \
+      defined(KOKKOS_ENABLE_SYCL))
+inline void initialize_device() {}
+inline void finalize_device() {}
+#endif
+
 }  // namespace Impl
 }  // namespace KokkosFFT
 

fft/src/KokkosFFT_HIP_types.hpp

@@ -238,7 +238,13 @@ template <typename ExecutionSpace>
 auto direction_type(Direction direction) {
   return direction == Direction::forward ? HIPFFT_FORWARD : HIPFFT_BACKWARD;
 }
+
+inline void initialize_host() {}
+inline void finalize_host() {}
 #endif
+
+inline void initialize_device() {}
+inline void finalize_device() {}
 }  // namespace Impl
 }  // namespace KokkosFFT
 

fft/src/KokkosFFT_ROCM_types.hpp

@@ -363,7 +363,19 @@ template <typename ExecutionSpace>
 auto direction_type(Direction direction) {
   return direction == Direction::forward ? ROCFFT_FORWARD : ROCFFT_BACKWARD;
 }
+
+inline void initialize_host() {}
+inline void finalize_host() {}
 #endif
+
+inline void initialize_device() {
+  rocfft_status status = rocfft_setup();
+  if (status != rocfft_status_success) Kokkos::abort("rocfft_setup failed");
+}
+inline void finalize_device() {
+  rocfft_status status = rocfft_cleanup();
+  if (status != rocfft_status_success) Kokkos::abort("rocfft_cleanup failed");
+}
 }  // namespace Impl
 }  // namespace KokkosFFT
 

fft/src/KokkosFFT_SYCL_types.hpp

@@ -225,7 +225,13 @@ template <typename ExecutionSpace>
 auto direction_type(Direction direction) {
   return direction == Direction::forward ? MKL_FFT_FORWARD : MKL_FFT_BACKWARD;
 }
+
+inline void initialize_host() {}
+inline void finalize_host() {}
 #endif
+
+inline void initialize_device() {}
+inline void finalize_device() {}
 }  // namespace Impl
 }  // namespace KokkosFFT
 

fft/unit_test/CMakeLists.txt

@@ -8,10 +8,18 @@ add_executable(unit-tests-kokkos-fft-core
     Test_Transform.cpp
 )
 
-target_compile_features(unit-tests-kokkos-fft-core PUBLIC cxx_std_17)
+add_executable(unit-tests-kokkos-fft-core-no-init
+    Test_MainNoInit.cpp
+    Test_InitializeFinalize.cpp
+)
 
+target_compile_features(unit-tests-kokkos-fft-core PUBLIC cxx_std_17)
 target_link_libraries(unit-tests-kokkos-fft-core PUBLIC KokkosFFT::fft GTest::gtest)
 
+target_compile_features(unit-tests-kokkos-fft-core-no-init PUBLIC cxx_std_17)
+target_link_libraries(unit-tests-kokkos-fft-core-no-init PUBLIC KokkosFFT::fft GTest::gtest)
+
 # Enable GoogleTest
 include(GoogleTest)
 gtest_discover_tests(unit-tests-kokkos-fft-core PROPERTIES DISCOVERY_TIMEOUT 600 DISCOVERY_MODE PRE_TEST)
+gtest_discover_tests(unit-tests-kokkos-fft-core-no-init PROPERTIES DISCOVERY_TIMEOUT 600 DISCOVERY_MODE PRE_TEST)