Consuming a populated JIT cache with read-only permissions


I have a JIT cache that’s already been populated. And I’m seeing the expected speed-up when I load my CUDA application using this cache.

But I have restrictions that require me to make the JIT cache directory read-only. Once it’s read-only, JIT compilation and loading becomes slow again… suggesting that the cache is not being used.

Is it required to have the JIT cache directory as read+write-enabled, even though we just want to only read from it?

Does setting the Cuda environment variable “CUDA_DISABLE_PTX_JIT” help?

No, it doesn’t seem like it.

Setting CUDA_DISABLE_PTX_JIT=1 makes the load to fail.

It looks like the GPU binary code is expected to be embedded in the CUDA application if CUDA_DISABLE_PTX_JIT is set…

In my case, I want to read from the JIT cache.

You could file a bug if you wish.