Conversation
Signed-off-by: Ken Museth <ken.museth@gmail.com>
| target_link_libraries(ex_make_mgpu_nanovdb PRIVATE CUDA::curand) | ||
| # ex_make_mgpu_nanovdb uses TF32 WMMA which requires SM_80+ (Ampere or newer) | ||
| execute_process( | ||
| COMMAND bash -c "nvidia-smi --query-gpu=compute_cap --format=csv,noheader 2>/dev/null | sort -n | tail -1 | tr -d '.'" |
There was a problem hiding this comment.
What if you're building on a machine that doesn't have a GPU and going to run it on an SM >=8.0 machine? Or are we assuming all examples are built/run on the same machine and/or they're not used as CI tests (which might have this behaviour)?
| target_link_libraries(ex_make_mgpu_nanovdb PRIVATE CUDA::curand) | ||
| # ex_make_mgpu_nanovdb uses TF32 WMMA which requires SM_80+ (Ampere or newer) | ||
| execute_process( | ||
| COMMAND bash -c "nvidia-smi --query-gpu=compute_cap --format=csv,noheader 2>/dev/null | sort -n | tail -1 | tr -d '.'" |
There was a problem hiding this comment.
For Windows, assuming availability of bash is a hard requirement. Can we rely on something like CMAKE_CUDA_ARCHITECTURES (see, e.g. this) instead of probing the local machine with bash and nvidia-smi?
There was a problem hiding this comment.
I asked Codex to scan CMAKE_CUDA_ARCHITECTURES and only add target if at least one architecture is >=80
if(CUDAToolkit_FOUND)
set(_nanovdb_has_sm80 OFF)
foreach(_cuda_arch IN LISTS CMAKE_CUDA_ARCHITECTURES)
string(REGEX MATCH "^[0-9]+" _cuda_arch_number "${_cuda_arch}")
if(_cuda_arch_number GREATER_EQUAL 80)
set(_nanovdb_has_sm80 ON)
endif()
endforeach()
if(_nanovdb_has_sm80)
nanovdb_example(NAME "ex_make_mgpu_nanovdb") # requires cuRAND and SM_80+ (TF32 WMMA)
target_link_libraries(ex_make_mgpu_nanovdb PRIVATE CUDA::curand)
else()
message(STATUS "Skipping ex_make_mgpu_nanovdb: requires CUDA architecture 80 or newer")
endif()
unset(_nanovdb_has_sm80)
unset(_cuda_arch_number)
unset(_cuda_arch)
endif()
There was a problem hiding this comment.
Good catch, both of you — the nvidia-smi/bash probe was checking the wrong thing (the configure-time machine's GPU) in the wrong way (not portable to Windows). Replaced it with a check against CMAKE_CUDA_ARCHITECTURES instead, which already governs what SM architectures this build actually targets (and is what the rest of nanovdb/CMakeLists.txt uses for this purpose too, defaulting to 75 when unset) — so it correctly handles cross-compiling and GPU-less CI runners, and has no bash/nvidia-smi dependency at all.
@apradhana — used your Codex-suggested loop over CMAKE_CUDA_ARCHITECTURES as-is. Verified the arch-matching regex/comparison standalone against 75, 80, 86, the -real/-virtual suffix forms, native, and empty/multi-value lists — all resolve as expected.
The unit test provides its own main() and uses only gtest (no gmock), so it needs just GTest::gtest. Listing gmock / gtest_main / gmock_main as well pulled libgtest.a and libgmock.a onto the link line multiple times, which the macOS (Xcode 15+) linker flags as "ignoring duplicate libraries". Linking only GTest::gtest removes the warning and also works against gtest-only GoogleTest installs (BUILD_GMOCK=OFF). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Signed-off-by: Ken Museth <ken.museth@gmail.com>
Signed-off-by: Ken Museth <ken.museth@gmail.com>
9711fe7 to
4bd415f
Compare
swahtz
left a comment
swahtz
left a comment
There was a problem hiding this comment.
Thanks for addressing my notes
apradhana
left a comment
apradhana
left a comment
There was a problem hiding this comment.
Thanks for addressing the comment regarding bash+nvidia-smi approach, Ken. I tested the CMake with CUDA architecture: 75 skips, while 80, 75;80, and suffixed values like 80-real build as expected.
3 participants
ex_make_mgpu_nanovdb uses TF32 WMMA which requires SM_80+ (Ampere or newer)
changed cmake to address this issue