Migrate C2H library to top-level library (#2629)

Migrate CUB's c2h (catch2 helper) testing library to top-level.

c/parallel/test/test_for.cpp

@@ -12,7 +12,7 @@
 
 #include <algorithm>
 
-#include "c2h.h"
+#include "test_util.h"
 #include <cccl/c/for.h>
 
 void for_each(cccl_iterator_t input, unsigned long long num_items, cccl_op_t op)

c/parallel/test/test_reduce.cpp

@@ -10,7 +10,7 @@
 
 #include <cuda_runtime.h>
 
-#include "c2h.h"
+#include "test_util.h"
 
 void reduce(cccl_iterator_t input, cccl_iterator_t output, unsigned long long num_items, cccl_op_t op, cccl_value_t init)
 {

c2h/CMakeLists.txt

@@ -0,0 +1,41 @@
+cmake_minimum_required(VERSION 3.21)
+
+project(C2H LANGUAGES CXX CUDA)
+
+cccl_get_catch2()
+
+find_package(CCCL CONFIG REQUIRED
+  NO_DEFAULT_PATH # Only check the explicit HINTS below:
+  HINTS "${CCCL_SOURCE_DIR}/lib/cmake/cccl/"
+)
+
+find_package(CUDAToolkit)
+
+set(curand_default OFF)
+if (CUDA_curand_LIBRARY)
+  set(curand_default ON)
+endif()
+
+option(C2H_ENABLE_CURAND "Use CUDA CURAND library in c2h." ${curand_default})
+
+add_library(cccl.c2h STATIC
+  generators.cu
+)
+target_include_directories(cccl.c2h PUBLIC "${C2H_SOURCE_DIR}/include")
+target_link_libraries(cccl.c2h PUBLIC
+  CCCL::CCCL
+  Catch2::Catch2
+)
+
+if (C2H_ENABLE_CURAND)
+  target_link_libraries(cccl.c2h PRIVATE CUDA::curand)
+  target_compile_definitions(cccl.c2h PRIVATE C2H_HAS_CURAND=1)
+else()
+  target_compile_definitions(cccl.c2h PRIVATE C2H_HAS_CURAND=0)
+endif()
+
+add_library(cccl.c2h.main OBJECT
+  catch2_runner.cpp
+  catch2_runner_helper.cu
+)
+target_link_libraries(cccl.c2h.main PUBLIC cccl.c2h)

cub/test/catch2_runner.cpp → c2h/catch2_runner.cpp

@@ -31,4 +31,4 @@
 
 #define CUB_CONFIG_MAIN
 #define CUB_EXCLUDE_CATCH2_HELPER_IMPL
-#include "catch2_main.cuh"
+#include <c2h/catch2_main.cuh>

cub/test/c2h/generators.cu → c2h/generators.cu

@@ -47,9 +47,9 @@
 #include <c2h/custom_type.cuh>
 #include <c2h/device_policy.cuh>
 #include <c2h/extended_types.cuh>
+#include <c2h/fill_striped.cuh>
 #include <c2h/generators.cuh>
 #include <c2h/vector.cuh>
-#include <fill_striped.cuh>
 
 #if C2H_HAS_CURAND
 #  include <curand.h>

cub/test/catch2_main.cuh → c2h/include/c2h/catch2_main.cuh

@@ -41,7 +41,7 @@
 #include <catch2/catch.hpp>
 
 #if defined(CUB_CONFIG_MAIN)
-#  include "catch2_runner_helper.h"
+#  include <c2h/catch2_runner_helper.cuh>
 
 #  if !defined(CUB_EXCLUDE_CATCH2_HELPER_IMPL)
 #    include "catch2_runner_helper.inl"

cub/test/catch2_test_helper.h → c2h/include/c2h/catch2_test_helper.cuh

@@ -29,14 +29,13 @@
 
 #include <cub/util_compiler.cuh>
 
+#include <cuda/std/__cccl/diagnostic.h>
+
 #include <cstdint>
 #include <cstdlib>
 #include <tuple>
 #include <type_traits>
 
-#include "cuda/std/__cccl/diagnostic.h"
-#include "test_util_vec.h"
-
 #if __CUDACC_VER_MAJOR__ == 11
 _CCCL_NV_DIAG_SUPPRESS(177) // catch2 may contain unused variableds
 #endif // nvcc-11
@@ -46,9 +45,9 @@ _CCCL_NV_DIAG_SUPPRESS(177) // catch2 may contain unused variableds
 #include <cuda/std/type_traits>
 #include <cuda/std/utility>
 
-#include "catch2_main.cuh"
-#include "test_warning_suppression.cuh"
+#include <c2h/catch2_main.cuh>
 #include <c2h/device_policy.cuh>
+#include <c2h/test_util_vec.cuh>
 #include <c2h/utility.cuh>
 #include <c2h/vector.cuh>
 
@@ -234,26 +233,26 @@ struct Catch::StringMaker<cudaError>
 #include <c2h/custom_type.cuh>
 #include <c2h/generators.cuh>
 
-#define CUB_TEST_NAME_IMPL(NAME, PARAM) CUB_TEST_STR(NAME) "(" CUB_TEST_STR(PARAM) ")"
+#define C2H_TEST_NAME_IMPL(NAME, PARAM) C2H_TEST_STR(NAME) "(" C2H_TEST_STR(PARAM) ")"
 
-#define CUB_TEST_NAME(NAME) CUB_TEST_NAME_IMPL(NAME, VAR_IDX)
+#define C2H_TEST_NAME(NAME) C2H_TEST_NAME_IMPL(NAME, VAR_IDX)
 
-#define CUB_TEST_CONCAT(A, B)       CUB_TEST_CONCAT_INNER(A, B)
-#define CUB_TEST_CONCAT_INNER(A, B) A##B
+#define C2H_TEST_CONCAT(A, B)       C2H_TEST_CONCAT_INNER(A, B)
+#define C2H_TEST_CONCAT_INNER(A, B) A##B
 
-#define CUB_TEST_IMPL(ID, NAME, TAG, ...)                                  \
-  using CUB_TEST_CONCAT(types_, ID) = c2h::cartesian_product<__VA_ARGS__>; \
-  TEMPLATE_LIST_TEST_CASE(CUB_TEST_NAME(NAME), TAG, CUB_TEST_CONCAT(types_, ID))
+#define C2H_TEST_IMPL(ID, NAME, TAG, ...)                                  \
+  using C2H_TEST_CONCAT(types_, ID) = c2h::cartesian_product<__VA_ARGS__>; \
+  TEMPLATE_LIST_TEST_CASE(C2H_TEST_NAME(NAME), TAG, C2H_TEST_CONCAT(types_, ID))
 
-#define CUB_TEST(NAME, TAG, ...) CUB_TEST_IMPL(__LINE__, NAME, TAG, __VA_ARGS__)
+#define C2H_TEST(NAME, TAG, ...) C2H_TEST_IMPL(__LINE__, NAME, TAG, __VA_ARGS__)
 
-#define CUB_TEST_LIST_IMPL(ID, NAME, TAG, ...)                     \
-  using CUB_TEST_CONCAT(types_, ID) = c2h::type_list<__VA_ARGS__>; \
-  TEMPLATE_LIST_TEST_CASE(CUB_TEST_NAME(NAME), TAG, CUB_TEST_CONCAT(types_, ID))
+#define C2H_TEST_LIST_IMPL(ID, NAME, TAG, ...)                     \
+  using C2H_TEST_CONCAT(types_, ID) = c2h::type_list<__VA_ARGS__>; \
+  TEMPLATE_LIST_TEST_CASE(C2H_TEST_NAME(NAME), TAG, C2H_TEST_CONCAT(types_, ID))
 
-#define CUB_TEST_LIST(NAME, TAG, ...) CUB_TEST_LIST_IMPL(__LINE__, NAME, TAG, __VA_ARGS__)
+#define C2H_TEST_LIST(NAME, TAG, ...) C2H_TEST_LIST_IMPL(__LINE__, NAME, TAG, __VA_ARGS__)
 
-#define CUB_TEST_STR(a) #a
+#define C2H_TEST_STR(a) #a
 
 namespace detail
 {
@@ -267,7 +266,7 @@ inline std::size_t adjust_seed_count(std::size_t requested)
 }
 } // namespace detail
 
-#define CUB_SEED(N)                                                                                                    \
+#define C2H_SEED(N)                                                                                                    \
   c2h::seed_t                                                                                                          \
   {                                                                                                                    \
     GENERATE_COPY(take(                                                                                                \

cub/test/c2h/extended_types.cuh → c2h/include/c2h/extended_types.cuh

@@ -40,11 +40,11 @@
 #ifdef TEST_HALF_T
 #  include <cuda_fp16.h>
 
-#  include "../half.h"
+#  include <c2h/half.cuh>
 #endif
 
 #ifdef TEST_BF_T
 #  include <cuda_bf16.h>
 
-#  include "../bfloat16.h"
+#  include <c2h/bfloat16.cuh>
 #endif

cub/test/test_util_vec.h → c2h/include/c2h/test_util_vec.cuh

@@ -68,7 +68,7 @@ inline int CoutCast(signed char val)
 /**
  * Vector1 overloads
  */
-#define CUB_VEC_OVERLOAD_1(T, BaseT)                                        \
+#define C2H_VEC_OVERLOAD_1(T, BaseT)                                        \
   /* Ostream output */                                                      \
   inline std::ostream& operator<<(std::ostream& os, const T& val)           \
   {                                                                         \
@@ -105,7 +105,7 @@ inline int CoutCast(signed char val)
 /**
  * Vector2 overloads
  */
-#define CUB_VEC_OVERLOAD_2(T, BaseT)                                        \
+#define C2H_VEC_OVERLOAD_2(T, BaseT)                                        \
   /* Ostream output */                                                      \
   inline std::ostream& operator<<(std::ostream& os, const T& val)           \
   {                                                                         \
@@ -150,7 +150,7 @@ inline int CoutCast(signed char val)
 /**
  * Vector3 overloads
  */
-#define CUB_VEC_OVERLOAD_3(T, BaseT)                                                         \
+#define C2H_VEC_OVERLOAD_3(T, BaseT)                                                         \
   /* Ostream output */                                                                       \
   inline std::ostream& operator<<(std::ostream& os, const T& val)                            \
   {                                                                                          \
@@ -203,7 +203,7 @@ inline int CoutCast(signed char val)
 /**
  * Vector4 overloads
  */
-#define CUB_VEC_OVERLOAD_4(T, BaseT)                                                                                   \
+#define C2H_VEC_OVERLOAD_4(T, BaseT)                                                                                   \
   /* Ostream output */                                                                                                 \
   inline std::ostream& operator<<(std::ostream& os, const T& val)                                                      \
   {                                                                                                                    \
@@ -264,27 +264,27 @@ inline int CoutCast(signed char val)
 /**
  * All vector overloads
  */
-#define CUB_VEC_OVERLOAD(COMPONENT_T, BaseT) \
-  CUB_VEC_OVERLOAD_1(COMPONENT_T##1, BaseT)  \
-  CUB_VEC_OVERLOAD_2(COMPONENT_T##2, BaseT)  \
-  CUB_VEC_OVERLOAD_3(COMPONENT_T##3, BaseT)  \
-  CUB_VEC_OVERLOAD_4(COMPONENT_T##4, BaseT)
+#define C2H_VEC_OVERLOAD(COMPONENT_T, BaseT) \
+  C2H_VEC_OVERLOAD_1(COMPONENT_T##1, BaseT)  \
+  C2H_VEC_OVERLOAD_2(COMPONENT_T##2, BaseT)  \
+  C2H_VEC_OVERLOAD_3(COMPONENT_T##3, BaseT)  \
+  C2H_VEC_OVERLOAD_4(COMPONENT_T##4, BaseT)
 
 /**
  * Define for types
  */
-CUB_VEC_OVERLOAD(char, char)
-CUB_VEC_OVERLOAD(short, short)
-CUB_VEC_OVERLOAD(int, int)
-CUB_VEC_OVERLOAD(long, long)
-CUB_VEC_OVERLOAD(longlong, long long)
-CUB_VEC_OVERLOAD(uchar, unsigned char)
-CUB_VEC_OVERLOAD(ushort, unsigned short)
-CUB_VEC_OVERLOAD(uint, unsigned int)
-CUB_VEC_OVERLOAD(ulong, unsigned long)
-CUB_VEC_OVERLOAD(ulonglong, unsigned long long)
-CUB_VEC_OVERLOAD(float, float)
-CUB_VEC_OVERLOAD(double, double)
+C2H_VEC_OVERLOAD(char, char)
+C2H_VEC_OVERLOAD(short, short)
+C2H_VEC_OVERLOAD(int, int)
+C2H_VEC_OVERLOAD(long, long)
+C2H_VEC_OVERLOAD(longlong, long long)
+C2H_VEC_OVERLOAD(uchar, unsigned char)
+C2H_VEC_OVERLOAD(ushort, unsigned short)
+C2H_VEC_OVERLOAD(uint, unsigned int)
+C2H_VEC_OVERLOAD(ulong, unsigned long)
+C2H_VEC_OVERLOAD(ulonglong, unsigned long long)
+C2H_VEC_OVERLOAD(float, float)
+C2H_VEC_OVERLOAD(double, double)
 
 /*
  * The following section defines macros to overload cub::NumericTraits<T>::{Max,Lowest}() for vector
@@ -294,7 +294,7 @@ CUB_VEC_OVERLOAD(double, double)
 /**
  * Vector1 overloads
  */
-#define CUB_VEC_1_TRAITS_OVERLOAD(T, BaseT)            \
+#define C2H_VEC_1_TRAITS_OVERLOAD(T, BaseT)            \
   CUB_NAMESPACE_BEGIN                                  \
   template <>                                          \
   struct NumericTraits<T>                              \
@@ -321,7 +321,7 @@ CUB_VEC_OVERLOAD(double, double)
 /**
  * Vector2 overloads
  */
-#define CUB_VEC_2_TRAITS_OVERLOAD(T, BaseT)                                        \
+#define C2H_VEC_2_TRAITS_OVERLOAD(T, BaseT)                                        \
   CUB_NAMESPACE_BEGIN                                                              \
   template <>                                                                      \
   struct NumericTraits<T>                                                          \
@@ -348,7 +348,7 @@ CUB_VEC_OVERLOAD(double, double)
 /**
  * Vector3 overloads
  */
-#define CUB_VEC_3_TRAITS_OVERLOAD(T, BaseT)                                                                        \
+#define C2H_VEC_3_TRAITS_OVERLOAD(T, BaseT)                                                                        \
   CUB_NAMESPACE_BEGIN                                                                                              \
   template <>                                                                                                      \
   struct NumericTraits<T>                                                                                          \
@@ -375,7 +375,7 @@ CUB_VEC_OVERLOAD(double, double)
 /**
  * Vector4 overloads
  */
-#define CUB_VEC_4_TRAITS_OVERLOAD(T, BaseT)            \
+#define C2H_VEC_4_TRAITS_OVERLOAD(T, BaseT)            \
   CUB_NAMESPACE_BEGIN                                  \
   template <>                                          \
   struct NumericTraits<T>                              \
@@ -408,24 +408,24 @@ CUB_VEC_OVERLOAD(double, double)
 /**
  * All vector overloads
  */
-#define CUB_VEC_TRAITS_OVERLOAD(COMPONENT_T, BaseT) \
-  CUB_VEC_1_TRAITS_OVERLOAD(COMPONENT_T##1, BaseT)  \
-  CUB_VEC_2_TRAITS_OVERLOAD(COMPONENT_T##2, BaseT)  \
-  CUB_VEC_3_TRAITS_OVERLOAD(COMPONENT_T##3, BaseT)  \
-  CUB_VEC_4_TRAITS_OVERLOAD(COMPONENT_T##4, BaseT)
+#define C2H_VEC_TRAITS_OVERLOAD(COMPONENT_T, BaseT) \
+  C2H_VEC_1_TRAITS_OVERLOAD(COMPONENT_T##1, BaseT)  \
+  C2H_VEC_2_TRAITS_OVERLOAD(COMPONENT_T##2, BaseT)  \
+  C2H_VEC_3_TRAITS_OVERLOAD(COMPONENT_T##3, BaseT)  \
+  C2H_VEC_4_TRAITS_OVERLOAD(COMPONENT_T##4, BaseT)
 
 /**
  * Define for types
  */
-CUB_VEC_TRAITS_OVERLOAD(char, signed char)
-CUB_VEC_TRAITS_OVERLOAD(short, short)
-CUB_VEC_TRAITS_OVERLOAD(int, int)
-CUB_VEC_TRAITS_OVERLOAD(long, long)
-CUB_VEC_TRAITS_OVERLOAD(longlong, long long)
-CUB_VEC_TRAITS_OVERLOAD(uchar, unsigned char)
-CUB_VEC_TRAITS_OVERLOAD(ushort, unsigned short)
-CUB_VEC_TRAITS_OVERLOAD(uint, unsigned int)
-CUB_VEC_TRAITS_OVERLOAD(ulong, unsigned long)
-CUB_VEC_TRAITS_OVERLOAD(ulonglong, unsigned long long)
-CUB_VEC_TRAITS_OVERLOAD(float, float)
-CUB_VEC_TRAITS_OVERLOAD(double, double)
+C2H_VEC_TRAITS_OVERLOAD(char, signed char)
+C2H_VEC_TRAITS_OVERLOAD(short, short)
+C2H_VEC_TRAITS_OVERLOAD(int, int)
+C2H_VEC_TRAITS_OVERLOAD(long, long)
+C2H_VEC_TRAITS_OVERLOAD(longlong, long long)
+C2H_VEC_TRAITS_OVERLOAD(uchar, unsigned char)
+C2H_VEC_TRAITS_OVERLOAD(ushort, unsigned short)
+C2H_VEC_TRAITS_OVERLOAD(uint, unsigned int)
+C2H_VEC_TRAITS_OVERLOAD(ulong, unsigned long)
+C2H_VEC_TRAITS_OVERLOAD(ulonglong, unsigned long long)
+C2H_VEC_TRAITS_OVERLOAD(float, float)
+C2H_VEC_TRAITS_OVERLOAD(double, double)

ci/inspect_changes.sh

@@ -28,17 +28,19 @@ subprojects=(
   cudax
   python
   cccl_c_parallel
+  c2h
 )
 
 # ...and their dependencies:
 declare -A dependencies=(
   [cccl]=""
   [libcudacxx]="cccl"
-  [cub]="cccl libcudacxx thrust"
+  [cub]="cccl libcudacxx thrust c2h"
   [thrust]="cccl libcudacxx cub"
   [cudax]="cccl libcudacxx"
   [python]="cccl libcudacxx cub thrust cccl_c_parallel"
   [cccl_c_parallel]="cccl libcudacxx cub thrust"
+  [c2h]="cccl libcudacxx cub thrust"
 )
 
 declare -A project_names=(
@@ -49,6 +51,7 @@ declare -A project_names=(
   [cudax]="CUDA Experimental"
   [python]="python"
   [cccl_c_parallel]="CCCL C Parallel Library"
+  [c2h]="Catch2Helper"
 )
 
 # By default, the project directory is assumed to be the same as the subproject name,

cmake/CCCLGetDependencies.cmake

@@ -5,6 +5,13 @@ macro(cccl_get_boost)
   CPMAddPackage("gh:boostorg/boost#boost-1.83.0")
 endmacro()
 
+# The CCCL Catch2Helper library:
+macro(cccl_get_c2h)
+  if (NOT TARGET cccl.c2h)
+    add_subdirectory("${CCCL_SOURCE_DIR}/c2h" "${CCCL_BINARY_DIR}/c2h")
+  endif()
+endmacro()
+
 macro(cccl_get_catch2)
   include("${_cccl_cpm_file}")
   CPMAddPackage("gh:catchorg/[email protected]")

cub/examples/CMakeLists.txt

@@ -1,3 +1,7 @@
+# Some of the examples include the `cub/test/test_util.h` header, which
+# depends on c2h:
+cccl_get_c2h()
+
 # Create meta targets that build all examples for a single configuration:
 foreach(cub_target IN LISTS CUB_TARGETS)
   cub_get_target_property(config_prefix ${cub_target} PREFIX)
@@ -30,7 +34,10 @@ function(cub_add_example target_name_var example_name example_src cub_target)
   set(example_meta_target cub.all.example.${example_name})
 
   add_executable(${example_target} "${example_src}")
-  target_link_libraries(${example_target} ${cub_target})
+  target_link_libraries(${example_target} PRIVATE
+    ${cub_target}
+    cccl.c2h
+  )
   cub_clone_target_properties(${example_target} ${cub_target})
   cub_configure_cuda_target(${example_target} RDC ${CUB_FORCE_RDC})
   target_include_directories(${example_target} PRIVATE "${CUB_SOURCE_DIR}/examples")