Build failure when using conda

[oleksandr-pavlyk]

I use conda environment to provide cmake and ninja needed to build nvbench.

I created the environment using conda create -n dev --yes python ipython numpy scipy pandas cmake ninja.

I attempted to build using

git clone https://github.com/NVIDIA/nvbench
cd nvbench
conda activate dev
cmake -B build --preset=nvbench-dev -DCMAKE_VERBOSE_MAKEFILE=ON
cmake --build build

Configuration log

(dev) local-opavlyk@2u1g-b650-1311:~/repos/nvbench$ cmake -B build --preset nvbench-dev -DCMAKE_VERBOSE_MAKEFILE=ON
-- [download 0% complete]
-- [download 1% complete]
-- [download 2% complete]
-- [download 6% complete]
-- [download 7% complete]
-- [download 9% complete]
-- [download 12% complete]
-- [download 14% complete]
-- [download 17% complete]
-- [download 18% complete]
-- [download 20% complete]
-- [download 23% complete]
-- [download 26% complete]
-- [download 29% complete]
-- [download 32% complete]
-- [download 35% complete]
-- [download 36% complete]
-- [download 39% complete]
-- [download 41% complete]
-- [download 43% complete]
-- [download 45% complete]
-- [download 48% complete]
-- [download 49% complete]
-- [download 50% complete]
-- [download 53% complete]
-- [download 54% complete]
-- [download 57% complete]
-- [download 60% complete]
-- [download 63% complete]
-- [download 65% complete]
-- [download 68% complete]
-- [download 71% complete]
-- [download 74% complete]
-- [download 77% complete]
-- [download 78% complete]
-- [download 81% complete]
-- [download 84% complete]
-- [download 87% complete]
-- [download 90% complete]
-- [download 91% complete]
-- [download 92% complete]
-- [download 93% complete]
-- [download 94% complete]
-- [download 95% complete]
-- [download 97% complete]
-- [download 100% complete]
-- The CUDA compiler identification is NVIDIA 13.1.115 with host compiler GNU 13.3.0
-- The CXX compiler identification is GNU 13.3.0
-- Detecting CUDA compiler ABI info
-- Detecting CUDA compiler ABI info - done
-- Check for working CUDA compiler: /usr/local/cuda/bin/nvcc - skipped
-- Detecting CUDA compile features
-- Detecting CUDA compile features - done
-- Detecting CXX compiler ABI info
-- Detecting CXX compiler ABI info - done
-- Check for working CXX compiler: /usr/bin/c++ - skipped
-- Detecting CXX compile features
-- Detecting CXX compile features - done
-- Testing CXX17 Support: TRUE
-- Testing CXX20 Support: TRUE
-- Testing CUDA17 Support: TRUE
-- Testing CUDA20 Support: TRUE
-- Found CUDAToolkit: /usr/local/cuda/targets/x86_64-linux/include;/usr/local/cuda/targets/x86_64-linux/include/cccl (found version "13.1.115")
-- Performing Test CMAKE_HAVE_LIBC_PTHREAD
-- Performing Test CMAKE_HAVE_LIBC_PTHREAD - Success
-- Found Threads: TRUE
-- Creating symlink from /localhome/local-opavlyk/repos/nvbench/compile_commands.json to /localhome/local-opavlyk/repos/nvbench/build/compile_commands.json...
-- Performing Test NVBench_CXX_FLAG__Wall
-- Performing Test NVBench_CXX_FLAG__Wall - Success
-- Performing Test NVBench_CXX_FLAG__Wextra
-- Performing Test NVBench_CXX_FLAG__Wextra - Success
-- Performing Test NVBench_CXX_FLAG__Wconversion
-- Performing Test NVBench_CXX_FLAG__Wconversion - Success
-- Performing Test NVBench_CXX_FLAG__Woverloaded_virtual
-- Performing Test NVBench_CXX_FLAG__Woverloaded_virtual - Success
-- Performing Test NVBench_CXX_FLAG__Wcast_qual
-- Performing Test NVBench_CXX_FLAG__Wcast_qual - Success
-- Performing Test NVBench_CXX_FLAG__Wpointer_arith
-- Performing Test NVBench_CXX_FLAG__Wpointer_arith - Success
-- Performing Test NVBench_CXX_FLAG__Wunused_local_typedefs
-- Performing Test NVBench_CXX_FLAG__Wunused_local_typedefs - Success
-- Performing Test NVBench_CXX_FLAG__Wunused_parameter
-- Performing Test NVBench_CXX_FLAG__Wunused_parameter - Success
-- Performing Test NVBench_CXX_FLAG__Wvla
-- Performing Test NVBench_CXX_FLAG__Wvla - Success
-- Performing Test NVBench_CXX_FLAG__Wgnu
-- Performing Test NVBench_CXX_FLAG__Wgnu - Failed
-- Performing Test NVBench_CXX_FLAG__Wno_gnu_line_marker
-- Performing Test NVBench_CXX_FLAG__Wno_gnu_line_marker - Success
-- Performing Test NVBench_CXX_FLAG__Werror
-- Performing Test NVBench_CXX_FLAG__Werror - Success
-- CPM: Using local package fmt@12.1.0
-- CPM: Adding package nlohmann_json@3.12.0 (3.12.0)
-- NVBench CUDA architectures: all-major
-- Configuring done (4.0s)
-- Generating done (0.1s)
-- Build files have been written to: /localhome/local-opavlyk/repos/nvbench/build
(dev) local-opavlyk@2u1g-b650-1311:~/repos/nvbench$

The build fails with linking errors, due to -lfmt linker option upsetting the linker.

[189/231] : && /usr/bin/g++ -Wl,--dependency-file=testing/CMakeFiles/nvbench.test.create.dir/link.d testing/CMakeFiles/nvbench.test.create.dir/create.cu.o -o bin/nvbench.test.create  -Wl,-rpath,/localhome/local-opavlyk/repos/nvbench/build/lib:/usr/local/cuda/extras/CUPTI/lib64:/localhome/local-opavlyk/miniforge/envs/dev/lib  lib/libnvbench.so  -lfmt  /usr/local/cuda/targets/x86_64-linux/lib/stubs/libnvidia-ml.so  /usr/local/cuda/targets/x86_64-linux/lib/stubs/libcuda.so  -ldl  /usr/lib/x86_64-linux-gnu/librt.a  /usr/local/cuda/extras/CUPTI/lib64/libcupti.so  /usr/local/cuda/extras/CUPTI/lib64/libnvperf_target.so  /usr/local/cuda/extras/CUPTI/lib64/libnvperf_host.so  /localhome/local-opavlyk/miniforge/envs/dev/lib/libfmt.so.12.1.0  -lstdc++fs  -lcudadevrt  -lcudart_static  -lrt  -lpthread  -ldl -L"/usr/local/cuda/targets/x86_64-linux/lib/stubs" -L"/usr/local/cuda/targets/x86_64-linux/lib" && :
FAILED: [code=1] bin/nvbench.test.create
: && /usr/bin/g++ -Wl,--dependency-file=testing/CMakeFiles/nvbench.test.create.dir/link.d testing/CMakeFiles/nvbench.test.create.dir/create.cu.o -o bin/nvbench.test.create  -Wl,-rpath,/localhome/local-opavlyk/repos/nvbench/build/lib:/usr/local/cuda/extras/CUPTI/lib64:/localhome/local-opavlyk/miniforge/envs/dev/lib  lib/libnvbench.so  -lfmt  /usr/local/cuda/targets/x86_64-linux/lib/stubs/libnvidia-ml.so  /usr/local/cuda/targets/x86_64-linux/lib/stubs/libcuda.so  -ldl  /usr/lib/x86_64-linux-gnu/librt.a  /usr/local/cuda/extras/CUPTI/lib64/libcupti.so  /usr/local/cuda/extras/CUPTI/lib64/libnvperf_target.so  /usr/local/cuda/extras/CUPTI/lib64/libnvperf_host.so  /localhome/local-opavlyk/miniforge/envs/dev/lib/libfmt.so.12.1.0  -lstdc++fs  -lcudadevrt  -lcudart_static  -lrt  -lpthread  -ldl -L"/usr/local/cuda/targets/x86_64-linux/lib/stubs" -L"/usr/local/cuda/targets/x86_64-linux/lib" && :
/usr/bin/ld: cannot find -lfmt: No such file or directory
collect2: error: ld returned 1 exit status

The issue appears to be caused by fmt library being fetched by CPM as well as installed into Python environment:

(dev) local-opavlyk@2u1g-b650-1311:~/repos/nvbench$ conda list fmt
# packages in environment at /localhome/local-opavlyk/miniforge/envs/dev:
#
# Name                     Version          Build            Channel
fmt                        12.1.0           hff5e90c_0       conda-forge

[orpheezt]

I have an similar problem where fmt headers are not recognized. As the system fmt was installed and was used by cmake.

-- CPM: Using local package fmt@11.2.0
-- CPM: Adding package nlohmann_json@3.12.0 (3.12.0)

So I had to force to fetch from source CPM

diff --git a/cmake/NVBenchDependencies.cmake b/cmake/NVBenchDependencies.cmake
index b9faf27..857ca3f 100644
--- a/cmake/NVBenchDependencies.cmake
+++ b/cmake/NVBenchDependencies.cmake
@@ -1,5 +1,11 @@
 ################################################################################
 # fmtlib/fmt
+#
+# Some environments provide a CMake fmt package without headers on the include
+# path (e.g. only shared libs are present). When that happens, CPM reports a
+# local package but builds will fail with missing fmt headers. Force CPM to
+# download fmt by triggering CPMAddPackage.
+
 set(export_set_details)
 set(install_fmt OFF)
 if(NOT BUILD_SHARED_LIBS AND NVBench_ENABLE_INSTALL_RULES)
@@ -8,6 +14,12 @@ if(NOT BUILD_SHARED_LIBS AND NVBench_ENABLE_INSTALL_RULES)
   set(install_fmt ON)
 endif()
 
+set(_nvbench_fmt_patch_args
+  # Any PATCH_COMMAND (even empty) forces CPMAddPackage, bypassing local package
+  # discovery when fmt headers are missing.
+  PATCH_COMMAND ""
+)
+
 rapids_cpm_find(fmt 11.2.0 ${export_set_details}
   CPM_ARGS
     GIT_REPOSITORY "https://github.com/fmtlib/fmt.git"
@@ -20,8 +32,17 @@ rapids_cpm_find(fmt 11.2.0 ${export_set_details}
       # Disable install rules since we're linking statically.
       "FMT_INSTALL ${install_fmt}"
       "CMAKE_POSITION_INDEPENDENT_CODE ON"
+    ${_nvbench_fmt_patch_args}
 )
 
+# Ensure fmt headers are reachable even when a local package omits include dirs.
+if(TARGET fmt::fmt)
+  get_target_property(_nvbench_fmt_includes fmt::fmt INTERFACE_INCLUDE_DIRECTORIES)
+  if((NOT _nvbench_fmt_includes) AND DEFINED fmt_SOURCE_DIR)
+    target_include_directories(fmt::fmt INTERFACE "${fmt_SOURCE_DIR}/include")
+  endif()
+endif()
+
 if(NOT fmt_ADDED)
   set(fmt_is_external TRUE)
 endif()

[domcharrier]

I could fix the test linker error in a Conda environment on my end via the following patch (fmt::fmt is the correct target, not fmt):

diff --git a/testing/CMakeLists.txt b/testing/CMakeLists.txt
index f407258..5756465 100644
--- a/testing/CMakeLists.txt
+++ b/testing/CMakeLists.txt
@@ -42,7 +42,7 @@ foreach(test_src IN LISTS test_srcs)
   string(PREPEND test_name "nvbench.test.")
   add_executable(${test_name} "${test_src}")
   target_include_directories(${test_name} PRIVATE "${CMAKE_CURRENT_LIST_DIR}")
-  target_link_libraries(${test_name} PRIVATE nvbench::nvbench fmt)
+  target_link_libraries(${test_name} PRIVATE nvbench::nvbench fmt::fmt)
   set_target_properties(${test_name} PROPERTIES COMPILE_FEATURES cuda_std_17)
   nvbench_config_target(${test_name})
   add_test(NAME ${test_name} COMMAND "$<TARGET_FILE:${test_name}>" ${NVBench_TEST_ARGS_${test_name}})
diff --git a/testing/device/CMakeLists.txt b/testing/device/CMakeLists.txt
index 2a5fdf9..b7272ee 100644
--- a/testing/device/CMakeLists.txt
+++ b/testing/device/CMakeLists.txt
@@ -1,7 +1,7 @@
 # Test that we're converging to an accurate mean + stdev without timing out:
 set(test_name nvbench.test.device.noisy_bench)
 add_executable(${test_name} noisy_bench.cu)
-target_link_libraries(${test_name} PRIVATE nvbench::main fmt)
+target_link_libraries(${test_name} PRIVATE nvbench::main fmt::fmt)
 nvbench_config_target(${test_name})
 add_dependencies(nvbench.test.all ${test_name})