Implement C API for device reduction (#2256)

* Implement C device reduce

* Format

* Fix device-specific module loading

* Don't need a context at build step now

* Address review feedback

CMakeLists.txt

@@ -38,6 +38,7 @@ option(CCCL_ENABLE_THRUST "Enable the Thrust developer build." ${CCCL_TOPLEVEL_P
 option(CCCL_ENABLE_TESTING "Enable CUDA C++ Core Library tests." ${CCCL_TOPLEVEL_PROJECT})
 option(CCCL_ENABLE_EXAMPLES "Enable CUDA C++ Core Library examples." ${CCCL_TOPLEVEL_PROJECT})
 option(CCCL_ENABLE_BENCHMARKS "Enable CUDA C++ Core Library benchmarks." OFF)
+option(CCCL_ENABLE_C "Enable CUDA C Core Library." OFF)
 
 option(CCCL_ENABLE_UNSTABLE "Enable targets and developer build options for unstable projects." OFF)
 
@@ -77,6 +78,10 @@ if (CCCL_ENABLE_UNSTABLE)
   add_subdirectory(cudax)
 endif()
 
+if (CCCL_ENABLE_C)
+  add_subdirectory(c)
+endif()
+
 if (CCCL_ENABLE_TESTING)
   add_subdirectory(test)
 endif()

CMakePresets.json

@@ -38,6 +38,7 @@
         "CCCL_ENABLE_TESTING": true,
         "CCCL_ENABLE_EXAMPLES": true,
         "CCCL_ENABLE_BENCHMARKS": true,
+        "CCCL_ENABLE_C": true,
         "LIBCUDACXX_ENABLE_LIBCUDACXX_TESTS": true,
         "CUB_ENABLE_TESTING": true,
         "CUB_ENABLE_EXAMPLES": true,

c/CMakeLists.txt

@@ -0,0 +1,20 @@
+cmake_minimum_required(VERSION 3.30)
+
+project(cccl.c LANGUAGES CUDA CXX)
+
+add_library(cccl.c SHARED src/reduce.cu)
+set_property(TARGET cccl.c PROPERTY POSITION_INDEPENDENT_CODE ON)
+set_property(TARGET cccl.c PROPERTY CXX_STANDARD 20)
+set_property(TARGET cccl.c PROPERTY CUDA_STANDARD 20)
+
+find_package(CUDAToolkit REQUIRED)
+
+# TODO Use static versions of cudart, nvrtc, and nvJitLink
+target_link_libraries(cccl.c PRIVATE CUDA::cudart
+                                     CUDA::nvrtc
+                                     CUDA::nvJitLink
+                                     CUDA::cuda_driver)
+target_compile_definitions(cccl.c PRIVATE NVRTC_GET_TYPE_NAME=1 CCCL_C_EXPERIMENTAL=1)
+target_include_directories(cccl.c PUBLIC "include")
+
+add_subdirectory(test)

c/include/cccl/reduce.h

@@ -0,0 +1,59 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of CUDA Experimental in CUDA C++ Core Libraries,
+// 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
+// SPDX-FileCopyrightText: Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#pragma once
+
+#ifndef CCCL_C_EXPERIMENTAL
+#  warning "C exposure is experimental and subject to change. Define CCCL_C_EXPERIMENTAL to acknowledge this warning."
+#else // ^^^ !CCCL_C_EXPERIMENTAL ^^^ / vvv CCCL_C_EXPERIMENTAL vvv
+
+#  include <cuda.h>
+
+#  include <cccl/types.h>
+
+struct cccl_device_reduce_build_result_t
+{
+  int cc;
+  void* cubin;
+  size_t cubin_size;
+  CUlibrary library;
+  CUkernel single_tile_kernel;
+  CUkernel single_tile_second_kernel;
+  CUkernel reduction_kernel;
+};
+
+// TODO return a union of nvtx/cuda/nvrtc errors or a string?
+extern "C" CCCL_C_API CUresult cccl_device_reduce_build(
+  cccl_device_reduce_build_result_t* build,
+  cccl_iterator_t d_in,
+  cccl_iterator_t d_out,
+  cccl_op_t op,
+  cccl_value_t init,
+  int cc_major,
+  int cc_minor,
+  const char* cub_path,
+  const char* thrust_path,
+  const char* libcudacxx_path,
+  const char* ctk_path) noexcept;
+
+extern "C" CCCL_C_API CUresult cccl_device_reduce(
+  cccl_device_reduce_build_result_t build,
+  void* d_temp_storage,
+  size_t* temp_storage_bytes,
+  cccl_iterator_t d_in,
+  cccl_iterator_t d_out,
+  unsigned long long num_items,
+  cccl_op_t op,
+  cccl_value_t init,
+  CUstream stream) noexcept;
+
+extern "C" CCCL_C_API CUresult cccl_device_reduce_cleanup(cccl_device_reduce_build_result_t* bld_ptr);
+
+#endif // CCCL_C_EXPERIMENTAL

c/include/cccl/types.h

@@ -0,0 +1,85 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of CUDA Experimental in CUDA C++ Core Libraries,
+// 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
+// SPDX-FileCopyrightText: Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#pragma once
+
+#ifndef CCCL_C_EXPERIMENTAL
+#  warning "C exposure is experimental and subject to change. Define CCCL_C_EXPERIMENTAL to acknowledge this warning."
+#else // ^^^ !CCCL_C_EXPERIMENTAL ^^^ / vvv CCCL_C_EXPERIMENTAL vvv
+
+#  if defined(_WIN32)
+#    define CCCL_C_API __declspec(dllexport)
+#  else
+#    define CCCL_C_API __attribute__((visibility("default")))
+#  endif
+
+enum class cccl_type_enum
+{
+  INT8    = 0,
+  INT16   = 1,
+  INT32   = 2,
+  INT64   = 3,
+  UINT8   = 4,
+  UINT16  = 5,
+  UINT32  = 6,
+  UINT64  = 7,
+  FLOAT32 = 8,
+  FLOAT64 = 9,
+  STORAGE = 10
+};
+
+struct cccl_type_info
+{
+  int size;
+  int alignment;
+  cccl_type_enum type;
+};
+
+enum class cccl_op_kind_t
+{
+  stateless,
+  stateful
+};
+
+struct cccl_op_t
+{
+  cccl_op_kind_t type;
+  const char* name;
+  const char* ltoir;
+  int ltoir_size;
+  int size;
+  int alignment;
+  void* state;
+};
+
+enum class cccl_iterator_kind_t
+{
+  pointer,
+  iterator
+};
+
+struct cccl_value_t
+{
+  cccl_type_info type;
+  void* state;
+};
+
+struct cccl_iterator_t
+{
+  int size;
+  int alignment;
+  cccl_iterator_kind_t type;
+  cccl_op_t advance;
+  cccl_op_t dereference;
+  cccl_type_info value_type;
+  void* state;
+};
+
+#endif // CCCL_C_EXPERIMENTAL