Support CUDA 10.2 #125
Labels
Comments
|
Well, for the record I managed to make it work without recompiling clang, with a very hackish solution. In case this is useful to someone:
Feel free to close the issue now or wait until 10.2 (with ptx65, etc.) is properly supported. |
mmitche
pushed a commit
to mmitche/llvm-project
that referenced
this issue
Aug 3, 2022
…723.11 (llvm#125) [release/11.x] Update dependencies from dotnet/arcade
mjklemm
pushed a commit
to mjklemm/llvm-project
that referenced
this issue
Aug 1, 2024
We determine the type of the kernel based on the MLIR code, which changes during the lowering phase. Some kernels, such as those with multiple workshare loops, are initially classified as SPMD kernels, but are later recognized as generic kernels during PFT lowering. In such cases, we need to identify the change in type and clear the trip count if it was previously set.
skatrak
added a commit
to skatrak/llvm-project
that referenced
this issue
Aug 16, 2024
This patch improves the fix in llvm#125 to detect target SPMD kernels during Flang lowering to MLIR. It transitions from a MLIR-based check to a PFT-based check, which is a more resilient alternative since the MLIR representation is in process of being built where it's being checked.
skatrak
added a commit
to skatrak/llvm-project
that referenced
this issue
Aug 16, 2024
This patch improves the fix in llvm#125 to detect target SPMD kernels during Flang lowering to MLIR. It transitions from a MLIR-based check to a PFT-based check, which is a more resilient alternative since the MLIR representation is in process of being built where it's being checked.
skatrak
added a commit
to skatrak/llvm-project
that referenced
this issue
Aug 19, 2024
This patch improves the fix in llvm#125 to detect target SPMD kernels during Flang lowering to MLIR. It transitions from a MLIR-based check to a PFT-based check, which is a more resilient alternative since the MLIR representation is in process of being built where it's being checked.
mjklemm
pushed a commit
to mjklemm/llvm-project
that referenced
this issue
Aug 19, 2024
This patch improves the fix in llvm#125 to detect target SPMD kernels during Flang lowering to MLIR. It transitions from a MLIR-based check to a PFT-based check, which is a more resilient alternative since the MLIR representation is in process of being built where it's being checked.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
While compiling with CUDA 10.2 I get the following error:
error: cannot find libdevice for sm_52. Provide path to different CUDA installation via --cuda-path, or pass -nocudalib to build without linking with libdevice.Apparently the problem is that
CudaStringToVersionreturnsCudaVersion::UNKNOWN.Unless I'm missing anything, the addition of
CudaVersion::CUDA_102after whereverCUDA_101is found should be trivial. Of course, withPtxFeature = "+ptx64", as PTX 6.5 can wait.The text was updated successfully, but these errors were encountered: