Improve: Scaling CUDA kernels

Add support for larger input arrays except
for CUB. Adds a new interleaving Thrust variant,
combining additions with FMA, that doesn't
yield any performance improvements on 4060.

CMakeLists.txt

@@ -1,24 +1,78 @@
-cmake_minimum_required(VERSION 3.14 FATAL_ERROR)
-
-# Define the CUDA architectures.
-# Empty value is forbidden.
+# Let's use a recent CMake version:
+# 3.16+ for native sanitizers support
+# 3.17+ for `FindCUDAToolkit`
+# 3.25.2 for CUDA20 support
+# The good news is that Ubuntu 24.04 comes with 3.28!
+cmake_minimum_required(VERSION 3.25.2 FATAL_ERROR)
+
+# ------------------------------------------------------------------------------
+# Project Setup
+# ------------------------------------------------------------------------------
 project(
     ParallelReductionsBenchmark
-    LANGUAGES CXX
     VERSION 0.1.0
-)
+    LANGUAGES CXX
+    DESCRIPTION "Parallel Reductions Benchmark for CPUs & GPUs"
+    HOMEPAGE_URL "https://github.com/ashvardanian/ParallelReductionsBenchmark")
 
-set(CMAKE_CUDA_STANDARD 17)
 set(CMAKE_CXX_STANDARD 17)
 set(CMAKE_CXX_STANDARD_REQUIRED YES)
 set(CMAKE_CXX_EXTENSIONS NO)
+set(CMAKE_CUDA_STANDARD 17)
+set(CMAKE_CUDA_STANDARD_REQUIRED YES)
 
-# Make Release by default
+# Some extra logging for the user:
+message(STATUS "----------------------------------------")
+message(STATUS "CMAKE_SYSTEM_NAME: ${CMAKE_SYSTEM_NAME}")
+message(STATUS "CMAKE_CXX_COMPILER_ID: ${CMAKE_CXX_COMPILER_ID}")
+message(STATUS "CMAKE_BUILD_TYPE: ${CMAKE_BUILD_TYPE}")
+message(STATUS "----------------------------------------")
+
+# Default to Release if no build type is set:
 if(NOT CMAKE_BUILD_TYPE)
     set(CMAKE_BUILD_TYPE Release)
 endif()
 
-# Get external content: Google Benchmark, cccl
+# ------------------------------------------------------------------------------
+# Detect CUDA Support
+# ------------------------------------------------------------------------------
+set(ENABLE_CUDA OFF)
+include(CheckLanguage)
+check_language(CUDA)
+
+if(CMAKE_CUDA_COMPILER)
+    enable_language(CUDA)
+    set(ENABLE_CUDA ON)
+    message(STATUS "CUDA detected! Using compiler: ${CMAKE_CUDA_COMPILER}")
+else()
+    message(STATUS "CUDA not detected. Skipping CUDA-specific builds.")
+endif()
+
+# ------------------------------------------------------------------------------
+# Options
+# ------------------------------------------------------------------------------
+option(USE_INTEL_TBB "Use Intel TBB for parallel STL algorithms" ON)
+option(USE_NVIDIA_CCCL "Use Nvidia CCCL for CUDA acceleration" ON)
+
+# Enable or disable options based on system and CUDA support
+if(ENABLE_CUDA)
+    set(USE_NVIDIA_CCCL ON)
+    set(USE_INTEL_TBB OFF) # Prioritize CUDA acceleration
+elseif(CMAKE_SYSTEM_NAME STREQUAL "Linux")
+    set(USE_INTEL_TBB ON) # Default to TBB on Linux without CUDA
+endif()
+
+message(STATUS "USE_INTEL_TBB: ${USE_INTEL_TBB}")
+message(STATUS "USE_NVIDIA_CCCL: ${USE_NVIDIA_CCCL}")
+
+# ------------------------------------------------------------------------------
+# Dependencies
+# ------------------------------------------------------------------------------
+find_package(Threads REQUIRED)
+find_package(OpenMP QUIET)
+find_package(OpenCL QUIET)
+
+set(FETCHCONTENT_QUIET OFF)
 include(FetchContent)
 
 FetchContent_Declare(
@@ -44,26 +98,43 @@ set(BENCHMARK_ENABLE_GTEST_TESTS OFF)
 set(BENCHMARK_USE_BUNDLED_GTEST ON)
 FetchContent_MakeAvailable(benchmark)
 
-# Nvidia CCCL can be configured with Intel TBB.
-# https://github.com/rapidsai/cuml/issues/3540
-FetchContent_Declare(
-    CCCL
-    GIT_REPOSITORY https://github.com/nvidia/cccl.git
-    GIT_TAG v2.7.0
-    GIT_SHALLOW TRUE
-)
-FetchContent_MakeAvailable(CCCL)
+# Intel TBB for "Parallel STL" algorithms
+# https://github.com/oneapi-src/oneTBB/tree/onetbb_2021
+if(USE_INTEL_TBB)
+    FetchContent_Declare(
+        IntelTBB
+        GIT_REPOSITORY https://github.com/uxlfoundation/oneTBB.git
+        GIT_TAG master
+    )
 
-# We need TBB for Parallel CPU Algorithms in GCC.
-# https://github.com/oneapi-src/oneTBB/blob/onetbb_2021/cmake/README.md
-FetchContent_Declare(
-    TBB
-    GIT_REPOSITORY https://github.com/oneapi-src/oneTBB.git
-    GIT_TAG v2022.0.0
-    GIT_SHALLOW TRUE
-)
-set(TBB_TEST OFF CACHE BOOL "Do not build TBB tests" FORCE)
-FetchContent_MakeAvailable(TBB)
+    # Suppress TBB's own tests:
+    set(TBB_TEST OFF CACHE BOOL "Do not build TBB tests" FORCE)
+    FetchContent_MakeAvailable(IntelTBB)
+
+    # ------------------------------------------------------------------------------
+    # TBB fix for -Wstringop-overflow warnings treated as errors
+    # ------------------------------------------------------------------------------
+    # The TBB library target is typically called "tbb". We can explicitly disable
+    # the `stringop-overflow` warning for TBB only:
+    if(TARGET tbb)
+        if(CMAKE_CXX_COMPILER_ID STREQUAL "GNU")
+            target_compile_options(tbb PRIVATE -Wno-stringop-overflow)
+        endif()
+    endif()
+endif()
+
+# Nvidia's CUDA Core Compute Libraries for GPU acceleration
+if(USE_NVIDIA_CCCL)
+    # CUB, Thrust, and other libraries of interest are now included into the
+    # CUDA Toolkit, so we don't need this anymore:
+    #
+    # FetchContent_Declare(NvidiaCCCL GIT_REPOSITORY https://github.com/nvidia/cccl.git)
+    # FetchContent_MakeAvailable(NvidiaCCCL)
+    find_package(CUDAToolkit REQUIRED)
+    message(STATUS "CUDA Toolkit Version: ${CUDAToolkit_VERSION}")
+    message(STATUS "CUDA Toolkit Include Path: ${CUDAToolkit_INCLUDE_DIRS}")
+    message(STATUS "CUDA Toolkit Libraries Path: ${CUDAToolkit_LIBRARY_DIR}")
+endif()
 
 set(CMAKE_CXX_FLAGS_DEBUG "${CMAKE_CXX_FLAGS_DEBUG} -g")
 set(CMAKE_CUDA_FLAGS_DEBUG "${CMAKE_CXX_FLAGS_DEBUG} -g")
@@ -72,20 +143,21 @@ set(CMAKE_CUDA_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} -O2")
 
 set(CMAKE_GCC_FLAGS "${CMAKE_GCC_FLAGS} -march=native -fopenmp")
 
-# Add CUDA and CCCL dependencies
-find_package(Threads REQUIRED)
-
 add_executable(reduce_bench reduce_bench.cpp)
-target_link_libraries(reduce_bench benchmark::benchmark fmt::fmt Threads::Threads TBB::tbb)
+target_link_libraries(reduce_bench benchmark::benchmark fmt::fmt Threads::Threads)
 
-find_package(OpenMP)
+if(USE_INTEL_TBB)
+    target_link_libraries(reduce_bench tbb)
+endif()
+
+if(USE_NVIDIA_CCCL)
+    target_link_libraries(reduce_bench CUDA::cudart CUDA::cublas)
+endif()
 
 if(OpenMP_FOUND)
     target_link_libraries(reduce_bench OpenMP::OpenMP_CXX)
 endif()
 
-find_package(OpenCL)
-
 if(OpenCL_FOUND)
     target_link_libraries(reduce_bench OpenCL::OpenCL)
 endif()