[Mlir-commits] [mlir] [mlir][Bufferization] Accelerate bufferization pass (PR #160655)

[Ziliang Zhang]

ziliangzl wrote:

> How much compilation time improvement are you seeing due to this? Can you also measure the memory overhead? You can measure memory usage with `time -v`.

Hi @matthias-springer  , we tried using this optimization to compile the UNet model (in our downstream custom model compiler) and reduced the compilation time from hours to around 30 minutes. We haven’t measured the memory overhead yet, but maybe @mingzheTerapines and I can take a look at that.

Also, I have another optimization that can reduce the YOLO8x model compilation time from 12 minutes to 6 minutes in debug builds. I might submit it to this PR or create a new PR a bit later.


https://github.com/llvm/llvm-project/pull/160655