Switch to using native support for CUDA instead of FindCUDA macros. #859
Conversation
|
I reopened this pull request because the original one had a bad e-mail in it. (One that I no longer have.) |
…nt variable to the top of the file. That is because there is a check added in this branch to see if CUDA support should be enabled. This check is needed to avoid probing for nvcc unless needed, because that check may fail and abort the build.
…orm arguments for --compiler-options.
…/qiskit-aer into native_CUDA_Cmake_support
There is a compatibility problem between 2 conan dependencies: urllib3 and requests. They are fixing it and I've prepared a PR with a workaround in the meantime (#1031 ). |
|
I have limited CUDA arch to just 5.2 to check that CI works. It seems we are running out of time with all the archs enabled. |
Hmm, it's been running for hours. I imagine it is stuck somehow but there is no log reported yet. |
Yeah, timed out on verify_wheels.py again. |
There is a bug in the code that makes the verify_wheel.py script run forever when we build the gpu wheel. The problem is the check of memory available on CUDA devices when there are no CUDA devices available, as is the case for our GPU wheel build action.
|
Yeah, I've already found it and hopefully fixed it (#1035). It was a change already introduced in master. |
This reverts commit 7a9b5f6.
…GDaniels/qiskit-aer into native_CUDA_Cmake_support
Looked like it worked except for a Mac build that ran out of time. Is there an automatic heuristic that retriggers the build or is there a way to force one without doing a commit? |
Yeah, that was weird, but the full gpu build wheel action worked at last!!! |
Reminder: There's still a small patch to get Windows working. It involves copying a modified pybind header file into the distribution, though. |
No problem. I think we can address Win later as we don't officially support it yet for GPUs. Pybind 2.6 is already released and fixes the error when compiling with nvcc. |
|
We probably should add a release note about this because it bumps the minimum cmake version. But we can do that after it's added to #1036 |
…iskit#859) * Switch to using native support for CUDA instead of FindCUDA macros. * Move check for AER_THRUST_BACKEND being communicated via an environment variable to the top of the file. That is because there is a check added in this branch to see if CUDA support should be enabled. This check is needed to avoid probing for nvcc unless needed, because that check may fail and abort the build. * Add x86_64 checks for issue 856, pull Qiskit#864. * Avoid compiling SIMD_SOURCE_FILE if there is not one. * Use AMD64 as a synonym for x86_64. * Add AMD64 as a synonym for x86_64. * Pull fixes from vvilpas's issue-856-build-ppc * Unconditinally use add_library even with CUDA (not cuda_add_library). * Unconditinally use add_library, not cuda_add_library for CUDA. Transform arguments for --compiler-options. * pulled from revive_CUDA_windows * no-op commit to force rebuild * note some places to refactor * Fix all but the last suggested changes by vvilpas * Still need AER_COMPILER_FLAGS initialized. * Use functions for prepending --compiler-options to pass default inherited options to NVCC. One is for when a list is prepared, and one is for when a list is not prepared. * Move include of nvcc_add_compiler_options to be adjacent to the FindCUDA call. * Likewise include nvcc_add_compiler_options inside the CUDA conditional. * Include nvcc_add_compiler_options.cmake after CMAKE environment variable assignments. * Added file. * space -> tab tweaks for local consistency * fix typo in argument name * Use an intermediate temporary variable, and an explicit set. * Remove an indirect * Follow change to remove an indirect from nvcc functions. * More quoting needed for list conversion. * More quoting needed for list conversion. * Enable gpu wheel to check that wheel building works * Do not include the whole FindCUDA, just the architecture selection bits. Assume that CUDA will be found. * force CMAKE_CUDA_COMPILER to nvcc. Trying to figure out what is wrong with qiskit-aer-gpu builds. * force CMAKE_CUDA_COMPILER to /usr/local/cuda/nvcc. Trying to figure out what is wrong with qiskit-aer-gpu builds. * remove hardcoded /usr/local/cuda/bin/nvcc in CMakeLists.txt and put a path extension in the wheels.yml workflow. * Nope, try CUDACXX instead. * Nope, try adding -- -DCMAKE_CUDA_COMPILER=/usr/local/cuda/bin/nvcc * Nope, try adding PATH in CIBW_ENVIRONMENT. * do not run verify_wheels for GPU build * return to standard file * Fix terminate call. Enable CI test * disable the CONAN patch to see if that is the cause of the build failure * reapply the CONAN patch -- there is something else wrong * add this back * Set CUDACXX for CUDA wheel build * Set AER_CUDA_ARCH to 5.2 to make a quick test on CI wheel build * Fix gpu wheel build There is a bug in the code that makes the verify_wheel.py script run forever when we build the gpu wheel. The problem is the check of memory available on CUDA devices when there are no CUDA devices available, as is the case for our GPU wheel build action. * Enable gpu wheel build temporarily to check that it works * Revert "Enable gpu wheel build temporarily to check that it works" This reverts commit 7a9b5f6. * Set cuda devices to 0 in case of error for consistency * Recover cmake-optiona-conan release note * Default CUDA archs to Auto again * disable gpu build test Co-authored-by: vvilpas <[email protected]> Co-authored-by: Victor Villar <[email protected]> Co-authored-by: Christopher J. Wood <[email protected]> (cherry picked from commit 796a12f)
Summary
This switches to native support in CMake (versions >= 3.8) so that CUDA-specific CMake macros aren't used to set up the build. It is to address issue #843 and make it possible to do per-CUDA file build options.
Details and comments