Grok-1 model compilation slows with performance optimizing flags

[archana-ramalingam]

What happened?

Grok-1 model compilation is 7 mins slower and the vmfb file is 1 MB larger when using the following flags:
--iree-dispatch-creation-enable-aggressive-fusion=true --iree-preprocessing-pass-pipeline="builtin.module(util.func(iree-preprocessing-generalize-linalg-matmul-experimental))"

With flags:
vmfb size: 3 MB
compile time: 8 mins

Without flags:
vmfb size: 2 MB
compile time: under 1 min

Steps to reproduce your issue

1. Fetch latest grok mlir from here: https://github.com/nod-ai/shark-ai/blob/grok-2/grok-1-q4_1.mlir
2. On latest IREE build, run the following command and compile takes ~8 mins with the flags vs without them (~1 min):

../iree-build/tools/iree-compile grok-1-q4_1.mlir --iree-hal-target-backends=rocm --iree-hip-target=gfx942 -o grok-1-q4_1.vmfb --iree-dispatch-creation-enable-aggressive-fusion=true --iree-preprocessing-pass-pipeline="builtin.module(util.func(iree-preprocessing-generalize-linalg-matmul-experimental))"

What component(s) does this issue relate to?

Compiler

Version information

No response

Additional context

No response

[IanWood1]

I can take a look at this

[IanWood1]

It seems to be coming from a single dispatch: https://gist.githubusercontent.com/IanWood1/e1cedac8fd6692ca2d25c108f0aff890/raw/4abd53eee6a8e60e231cd276445ddfef6d8d8110/grok-slow-dispatch.mlir

It can be reproduced with iree-compile grok-slow-dispatch. The majority of the time is spent inside LLVM's SLPVectorizerPass