Error: too few arguments to function ‘CUresult cuGraphInstantiate_v2'

[sanchit-grover-intel]

I am trying to build LLVM from source in an openSUSE container and I am getting the following error:

[3393/3692] Building CXX object _deps/unified-runtime-build/source/adapters/cuda/CMakeFiles/ur_adapter_cuda.dir/command_buffer.cpp.o
FAILED: _deps/unified-runtime-build/source/adapters/cuda/CMakeFiles/ur_adapter_cuda.dir/command_buffer.cpp.o
/usr/bin/c++ -DUMF_ENABLE_POOL_TRACKING -DUR_VALIDATION_LAYER_SUPPORTED_VERSION=\"0\" -DUR_VERSION=\"0\" -D_DEBUG -D_FORTIFY_SOURCE=2 -D_GLIBCXX_ASSERTIONS -D_GNU_SOURCE -D__STDC_CONSTANT_MACROS -D__STDC_FORMAT_MACROS -D__STDC_LIMIT_MACROS -Dur_adapter_cuda_EXPORTS -I/tmp/llvm-2023-WW46/build/_deps/unified-runtime-build/source/adapters/cuda -I/tmp/llvm-2023-WW46/build/_deps/unified-runtime-src/source/adapters/cuda -I/tmp/llvm-2023-WW46/build/include -I/tmp/llvm-2023-WW46/llvm/include -I/tmp/llvm-2023-WW46/build/_deps/unified-runtime-src/source/adapters/cuda/../.. -I/tmp/llvm-2023-WW46/build/_deps/unified-runtime-src/include -I/tmp/llvm-2023-WW46/build/_deps/unified-runtime-src/source/common -I/tmp/llvm-2023-WW46/build/_deps/unified-runtime-src/source/common/unified_malloc_framework/include -I/tmp/llvm-2023-WW46/build/_deps/unified-runtime-src/source/common/umf_pools -isystem /usr/local/cuda/include -fPIC -fno-semantic-interposition -fvisibility-inlines-hidden -Werror=date-time -fno-lifetime-dse -Wall -Wextra -Wno-unused-parameter -Wwrite-strings -Wcast-qual -Wno-missing-field-initializers -pedantic -Wno-long-long -Wimplicit-fallthrough -Wno-maybe-uninitialized -Wno-nonnull -Wno-noexcept-type -Wdelete-non-virtual-dtor -Wno-comment -Wno-misleading-indentation -fdiagnostics-color -ffunction-sections -fdata-sections -Wall -Wextra -Wno-deprecated-declarations -Wno-error -O3 -DNDEBUG -std=c++1z -fPIC -UNDEBUG -fPIC -Wall -Wpedantic -fdiagnostics-color=always -MD -MT _deps/unified-runtime-build/source/adapters/cuda/CMakeFiles/ur_adapter_cuda.dir/command_buffer.cpp.o -MF _deps/unified-runtime-build/source/adapters/cuda/CMakeFiles/ur_adapter_cuda.dir/command_buffer.cpp.o.d -o _deps/unified-runtime-build/source/adapters/cuda/CMakeFiles/ur_adapter_cuda.dir/command_buffer.cpp.o -c /tmp/llvm-2023-WW46/build/_deps/unified-runtime-src/source/adapters/cuda/command_buffer.cpp
In file included from /tmp/llvm-2023-WW46/build/_deps/unified-runtime-src/source/adapters/cuda/context.hpp:20:0,
                 from /tmp/llvm-2023-WW46/build/_deps/unified-runtime-src/source/adapters/cuda/command_buffer.hpp:14,
                 from /tmp/llvm-2023-WW46/build/_deps/unified-runtime-src/source/adapters/cuda/command_buffer.cpp:11:
/tmp/llvm-2023-WW46/build/_deps/unified-runtime-src/source/adapters/cuda/command_buffer.cpp: In function ‘ur_result_t urCommandBufferFinalizeExp(ur_exp_command_buffer_handle_t)’:
/tmp/llvm-2023-WW46/build/_deps/unified-runtime-src/source/adapters/cuda/command_buffer.cpp:144:67: error: too few arguments to function ‘CUresult cuGraphInstantiate_v2(CUgraphExec_st**, CUgraph, CUgraphNode_st**, char*, size_t)’
                                       hCommandBuffer->CudaGraph, 0));
                                                                   ^
/tmp/llvm-2023-WW46/build/_deps/unified-runtime-src/source/adapters/cuda/common.hpp:31:16: note: in definition of macro ‘UR_CHECK_ERROR’
   checkErrorUR(Result, __func__, __LINE__, __FILE__)
                ^~~~~~
In file included from /tmp/llvm-2023-WW46/build/_deps/unified-runtime-src/source/adapters/cuda/context.hpp:12:0,
                 from /tmp/llvm-2023-WW46/build/_deps/unified-runtime-src/source/adapters/cuda/command_buffer.hpp:14,
                 from /tmp/llvm-2023-WW46/build/_deps/unified-runtime-src/source/adapters/cuda/command_buffer.cpp:11:
/usr/local/cuda/include/cuda.h:17028:18: note: declared here
 CUresult CUDAAPI cuGraphInstantiate(CUgraphExec *phGraphExec, CUgraph hGraph, CUgraphNode *phErrorNode, char *logBuffer, size_t bufferSize);

This is the build command that is being used:

python ./buildbot/configure.py \
    --cuda \
    --cmake-opt=-DCMAKE_PREFIX_PATH=/usr/local/cuda/lib64/stubs \
    --cmake-opt=-DCUDA_TOOLKIT_ROOT_DIR=/usr/local/cuda \
    --cmake-opt=-DCMAKE_INSTALL_PREFIX=/usr/local/llvm && \
 python ./buildbot/compile.py

When building the tag 2023-WW46, I get this error for CUDA 11.4 and 11.7 but not for 12.2. When using 2023-WW13, this error is not encountered. Does the newer LLVM tag not support older CUDA versions?

[EwanC]

The PRs that are intended to fix this that have been merged are:

• [UR][CUDA] Fix compatibility with CUDA 11.x oneapi-src/unified-runtime#1037
• [UR] Bump tag to v0.8.1 #12004

[sanchit-grover-intel]

Thank you @EwanC for the clarification. I was able to successfully build with a recent nightly build. May I know if there is going to be a release soon (this year perhaps) that will include this fix?