Describe the bug
In Windows, when optimizing a shader group in parallel, the performance is almost as the optimization is performed sequentially.
I tracked down the problem, in src\liboslexec\llvm_instance.cpp.
In the block inside if (use_rs_bitcode()) {
the call to ll.validate_global_mappings(names_of_unmapped_globals); triggers a global lock inside llvm.
commenting out the validation solves the problem. A possible solution would be to have an option to turn off validation, so I can turn it on only during debugging.
OSL version and dependencies
- OSL branch/version: 1.14.5.1
- OS: Windows
- C++ compiler: MSVC 2022
- LLVM version: 19.1.7
- OIIO version: 3.0.2
Describe the bug
In Windows, when optimizing a shader group in parallel, the performance is almost as the optimization is performed sequentially.
I tracked down the problem, in src\liboslexec\llvm_instance.cpp.
In the block inside if (use_rs_bitcode()) {
the call to ll.validate_global_mappings(names_of_unmapped_globals); triggers a global lock inside llvm.
commenting out the validation solves the problem. A possible solution would be to have an option to turn off validation, so I can turn it on only during debugging.
OSL version and dependencies