Some questions about iree cuda backend #17904
              
                Unanswered
              
          
                  
                    
                      xiaoxuan-yu
                    
                  
                
                  asked this question in
                Q&A
              
            Replies: 0 comments
  
    Sign up for free
    to join this conversation on GitHub.
    Already have an account?
    Sign in to comment
  
        
    
Uh oh!
There was an error while loading. Please reload this page.
-
I'm trying to run StableHLO with cpp interface provided by IREE. I refered to https://github.com/iree-org/iree-template-runtime-cmake as an entry point. It runs perfectly when using llvm-cpu and local-sync. But when I tried to use CUDA as backend, I failed to get the things done correctly. The mlir I used is
which is exported from JAX and can work well when using llvm-cpu backend. And the C code is as below
I simply changed the compile command to
iree-compile --iree-hal-target-backends=cuda --iree-hal-cuda-llvm-target-arch=sm_75 mul.mlir -o build/mul.vmfbfirst but the output shows that "PERMISSION_DENIED; buffer memory type is not compatible with the requested operation; buffer has DEVICE_LOCAL, operation requires HOST_VISIBLE". I searched the issues and noticed that non-mappable buffers are used by default, so I tried to add--iree-stream-external-resources-mappable=trueflag and the program gave out the right result, but without printing of lhs and rhs buffer views.Does anyone knows what happened? Besides, due to the lack of doc, I do not understand how to use non-mappable buffers correctly without adding the flag
--iree-stream-external-resources-mappable=true. Looking forward to your reply.Beta Was this translation helpful? Give feedback.
All reactions