[kernel] Clean up shared memory usage in pipelines

nod-ai / SHARK-ModelDev

Unified compiler/runtime for interfacing with PyTorch Dynamo.

Apache License 2.0

95 stars 48 forks source link

Open antiagainst opened 3 months ago

antiagainst commented 3 months ago

Right now we have a few issues with shared memory planning and usage that needs to figure out ways to improve:

We rely on bufferization to allocate shared memory. Need to be explicit to control it.
There are many stages of shared memory can be updated. This makes estimation hard. Need to figure out a way to improve that.