forward-opt: MLIR project to optimize and deploy VIT forward inference on X86

=======================

Directories:

1. python: python tools for quantization and transforming.

2. tools: cpp of forward-opt

3. include/lib: c++ headers/sources for forwar-opt

4. models: torch-mlir models to run

5. bash_tools: some bash scripts

Build

Change your llvm install path in build_forwardopt.sh and CMakeList.txt, and run build_forwardopt.sh (recommemd to run it through ctrl+C/ctrl+v line-by-line).

Run

Change your llvm install path in env.sh, and source env.sh.

Then run the model transforming following scripts in ./models.(eg. ./models/simplr/scripts.sh).

Pytorch to mlir needs torch-mlir python package.

Dependencies

LLVM-17

Maybe we need a new version of llvm. Commit: 4553dc46a05ec6f1e2aebcde1ce185772a26780b

Please download it from

https://github.com/llvm/llvm-project/tree/4553dc46a05ec6f1e2aebcde1ce185772a26780b

Install it with ./bash_tools/build_llvm.sh

Visual Studio 2022 and its clang-cl component

If you want to run VIT and link .dll on windows10/11, vs2022 and the integrated clang is neccesary.

Newest torch-mlir

Please download it from

https://github.com/segmentKOBE/torch-mlir-forward