Merge branch 'main' into main

Author: alliepiper

CMakeLists.txt

@@ -11,8 +11,13 @@ else()
   set(NVBench_TOPLEVEL_PROJECT OFF)
 endif()
 
+# Default to native if no cuda arches are specified:
+if (NOT (DEFINED CMAKE_CUDA_ARCHITECTURES OR DEFINED ENV{CUDAARCHS}))
+  set(CMAKE_CUDA_ARCHITECTURES "native")
+endif()
+
 include(cmake/NVBenchRapidsCMake.cmake)
-nvbench_load_rapids_cmake()
+nvbench_load_rapids_cmake(25.12)
 
 project(NVBench
   LANGUAGES CUDA CXX

cmake/NVBenchConfigTarget.cmake

@@ -33,13 +33,46 @@ function(nvbench_add_cxx_flag target_name type flag)
   endif()
 endfunction()
 
+# We test to see if C++ compiler options exist using try-compiles in the CXX lang, and then reuse those flags as
+# -Xcompiler flags for CUDA targets. This requires that the CXX compiler and CUDA_HOST compilers are the same when
+# using nvcc.
+if (NVBench_TOPLEVEL_PROJECT AND CMAKE_CUDA_COMPILER_ID STREQUAL "NVIDIA")
+  set(cuda_host_matches_cxx_compiler FALSE)
+  if (CMAKE_VERSION VERSION_GREATER_EQUAL 3.31)
+    set(host_info "${CMAKE_CUDA_HOST_COMPILER} (${CMAKE_CUDA_HOST_COMPILER_ID} ${CMAKE_CUDA_HOST_COMPILER_VERSION})")
+    set(cxx_info "${CMAKE_CXX_COMPILER} (${CMAKE_CXX_COMPILER_ID} ${CMAKE_CXX_COMPILER_VERSION})")
+    if (CMAKE_CUDA_HOST_COMPILER_ID STREQUAL CMAKE_CXX_COMPILER_ID AND
+        CMAKE_CUDA_HOST_COMPILER_VERSION VERSION_EQUAL CMAKE_CXX_COMPILER_VERSION)
+      set(cuda_host_matches_cxx_compiler TRUE)
+    endif()
+  else() # CMake < 3.31 doesn't have the CMAKE_CUDA_HOST_COMPILER_ID/VERSION variables
+    set(host_info "${CMAKE_CUDA_HOST_COMPILER}")
+    set(cxx_info "${CMAKE_CXX_COMPILER}")
+    if (CMAKE_CUDA_HOST_COMPILER STREQUAL CMAKE_CXX_COMPILER)
+      set(cuda_host_matches_cxx_compiler TRUE)
+    endif()
+  endif()
+
+  if (NOT cuda_host_matches_cxx_compiler)
+    message(FATAL_ERROR
+      "NVBench developer builds require that CMAKE_CUDA_HOST_COMPILER matches "
+      "CMAKE_CXX_COMPILER when using nvcc:\n"
+      "CMAKE_CUDA_COMPILER: ${CMAKE_CUDA_COMPILER}\n"
+      "CMAKE_CUDA_HOST_COMPILER: ${host_info}\n"
+      "CMAKE_CXX_COMPILER: ${cxx_info}\n"
+      "Rerun cmake with \"-DCMAKE_CUDA_HOST_COMPILER=${CMAKE_CXX_COMPILER}\".\n"
+      "Alternatively, configure the CUDAHOSTCXX and CXX environment variables to match.\n"
+    )
+  endif()
+endif()
+
 nvbench_add_cxx_flag(nvbench.build_interface INTERFACE "-Wall")
 nvbench_add_cxx_flag(nvbench.build_interface INTERFACE "-Wextra")
 nvbench_add_cxx_flag(nvbench.build_interface INTERFACE "-Wconversion")
 nvbench_add_cxx_flag(nvbench.build_interface INTERFACE "-Woverloaded-virtual")
 nvbench_add_cxx_flag(nvbench.build_interface INTERFACE "-Wcast-qual")
 nvbench_add_cxx_flag(nvbench.build_interface INTERFACE "-Wpointer-arith")
-nvbench_add_cxx_flag(nvbench.build_interface INTERFACE "-Wunused-local-typedef")
+nvbench_add_cxx_flag(nvbench.build_interface INTERFACE "-Wunused-local-typedefs")
 nvbench_add_cxx_flag(nvbench.build_interface INTERFACE "-Wunused-parameter")
 nvbench_add_cxx_flag(nvbench.build_interface INTERFACE "-Wvla")
 nvbench_add_cxx_flag(nvbench.build_interface INTERFACE "-Wgnu")

cmake/NVBenchDependencies.cmake

@@ -9,7 +9,6 @@ if(NOT BUILD_SHARED_LIBS AND NVBench_ENABLE_INSTALL_RULES)
 endif()
 
 rapids_cpm_find(fmt 11.2.0 ${export_set_details}
-  GLOBAL_TARGETS fmt::fmt fmt::fmt-header-only
   CPM_ARGS
     GIT_REPOSITORY "https://github.com/fmtlib/fmt.git"
     GIT_TAG "11.2.0"

cmake/NVBenchRapidsCMake.cmake

@@ -1,12 +1,13 @@
 # Called before project(...)
-macro(nvbench_load_rapids_cmake)
+macro(nvbench_load_rapids_cmake version)
   # - Including directly, see https://github.com/rapidsai/rmm/pull/1886
   # - Versioned download URL:
   #   https://raw.githubusercontent.com/rapidsai/rapids-cmake/branch-XX.YY/RAPIDS.cmake
   # - This macro is always called before project() in the root CMakeLists.txt, so:
   #   - we can't just use NVBench_SOURCE_DIR, it's not defined yet.
   #   - We can't rely on CMAKE_CURRENT_LIST_DIR because of macro expansion.
   #   - We can fallback to CURRENT_SOURCE_DIR because we know this will be expanded in the root:
+  set(rapids-cmake-version ${version})
   include("${CMAKE_CURRENT_SOURCE_DIR}/cmake/RAPIDS.cmake")
 
   include(rapids-cmake)

cmake/RAPIDS.cmake

@@ -16,11 +16,13 @@
 #
 # This is the preferred entry point for projects using rapids-cmake
 #
+# Enforce the minimum required CMake version for all users
+cmake_minimum_required(VERSION 3.30.4 FATAL_ERROR)
 
 # Allow users to control which version is used
-if(NOT rapids-cmake-version)
-  # Define a default version if the user doesn't set one
-  set(rapids-cmake-version 25.04)
+if(NOT (rapids-cmake-branch OR rapids-cmake-version))
+  message(FATAL_ERROR "The CMake variable `rapids-cmake-branch` or `rapids-cmake-version` must be defined"
+  )
 endif()
 
 # Allow users to control which GitHub repo is fetched
@@ -66,26 +68,18 @@ if(NOT rapids-cmake-url)
   endif()
 endif()
 
-if(POLICY CMP0135)
-  cmake_policy(PUSH)
-  cmake_policy(SET CMP0135 NEW)
-endif()
 include(FetchContent)
 if(rapids-cmake-fetch-via-git)
-  FetchContent_Declare(rapids-cmake
-    GIT_REPOSITORY "${rapids-cmake-url}"
-    GIT_TAG "${rapids-cmake-value-to-clone}")
+  FetchContent_Declare(rapids-cmake GIT_REPOSITORY "${rapids-cmake-url}"
+                       GIT_TAG "${rapids-cmake-value-to-clone}")
 else()
   string(APPEND rapids-cmake-url "${rapids-cmake-value-to-clone}")
   FetchContent_Declare(rapids-cmake URL "${rapids-cmake-url}")
 endif()
-if(POLICY CMP0135)
-  cmake_policy(POP)
-endif()
 FetchContent_GetProperties(rapids-cmake)
 if(rapids-cmake_POPULATED)
-  # Something else has already populated rapids-cmake, only thing
-  # we need to do is setup the CMAKE_MODULE_PATH
+  # Something else has already populated rapids-cmake, only thing we need to do is setup the
+  # CMAKE_MODULE_PATH
   if(NOT "${rapids-cmake-dir}" IN_LIST CMAKE_MODULE_PATH)
     list(APPEND CMAKE_MODULE_PATH "${rapids-cmake-dir}")
   endif()

scripts/nvbench_compare.py

@@ -5,6 +5,7 @@
 import os
 import sys
 
+import jsondiff
 import tabulate
 from colorama import Fore
 from nvbench_json import reader
@@ -371,9 +372,13 @@ def main():
         global all_devices
         all_devices = cmp_root["devices"]
 
-        # This is blunt but works for now:
         if ref_root["devices"] != cmp_root["devices"]:
             print("Device sections do not match.")
+            print(
+                jsondiff.diff(
+                    ref_root["devices"], cmp_root["devices"], syntax="symmetric"
+                )
+            )
             sys.exit(1)
 
         compare_benches(