README.md

• PI
• Installing
• Torch-MLIR

PI

Early days of a lightweight MLIR Python frontend with support for PyTorch (through Torch-MLIR but without a true dependency on PyTorch itself).

Installing

Just

pip install -r requirements.txt
pip install . -v

and you're good to go.

PyTorch

examples/minimal.py lowers

class MyConv2d(nn.Module):
    def __init__(self):
        super().__init__()
        self.conv = nn.Conv2d(3, 1, 3)

    def forward(self, x):
        y = self.conv(x)
        z = y + y
        w = z * z
        return w

to

module attributes {pi.module_name = "MyConv2d"} {
  func.func @forward(%arg0: !torch.vtensor<[?,?,?,?],f32>) -> !torch.vtensor<[?,1,?,?],f32> {
    %false = torch.constant.bool false
    %none = torch.constant.none
    %int3 = torch.constant.int 3
    %int0 = torch.constant.int 0
    %int1 = torch.constant.int 1
    %0 = torch.prim.ListConstruct %int1 : (!torch.int) -> !torch.list<int>
    %1 = torch.aten.empty.memory_format %0, %none, %none, %none, %none, %none : !torch.list<int>, !torch.none, !torch.none, !torch.none, !torch.none, !torch.none -> !torch.vtensor<[1],f32>
    %2 = torch.prim.ListConstruct %int1, %int3, %int3, %int3 : (!torch.int, !torch.int, !torch.int, !torch.int) -> !torch.list<int>
    %3 = torch.aten.empty.memory_format %2, %none, %none, %none, %none, %none : !torch.list<int>, !torch.none, !torch.none, !torch.none, !torch.none, !torch.none -> !torch.vtensor<[1,3,3,3],f32>
    %4 = torch.prim.ListConstruct %int1, %int1 : (!torch.int, !torch.int) -> !torch.list<int>
    %5 = torch.prim.ListConstruct %int0, %int0 : (!torch.int, !torch.int) -> !torch.list<int>
    %6 = torch.prim.ListConstruct  : () -> !torch.list<int>
    %7 = torch.aten.convolution %arg0, %3, %1, %4, %5, %4, %false, %6, %int1 : 
        !torch.vtensor<[?,?,?,?],f32>, 
        !torch.vtensor<[1,3,3,3],f32>, 
        !torch.vtensor<[1],f32>, 
        !torch.list<int>, 
        !torch.list<int>, 
        !torch.list<int>, 
        !torch.bool, 
        !torch.list<int>, 
        !torch.int -> !torch.vtensor<[?,1,?,?],f32>
    %8 = torch.aten.add.Tensor %7, %7, %int1 : !torch.vtensor<[?,1,?,?],f32>, !torch.vtensor<[?,1,?,?],f32>, !torch.int -> !torch.vtensor<[?,1,?,?],f32>
    %9 = torch.aten.mul.Tensor %8, %8 : !torch.vtensor<[?,1,?,?],f32>, !torch.vtensor<[?,1,?,?],f32> -> !torch.vtensor<[?,1,?,?],f32>
    return %9 : !torch.vtensor<[?,1,?,?],f32>
  }
}

In addition, we have several full end-to-end model examples, including ResNet18, InceptionV3, MobileNetV3.

In general, PI is very alpha; to get a rough idea of the current status check the latest tests.

Currently, we're passing ~650 out of 786 of Torch-MLIR's test-suite (torch-mlir==20230127.731).

Development

Spin up a venv (or conda environment) with pip install -r requirements.txt and configure CMake with

cmake \
    -DCMAKE_INSTALL_PREFIX=$PI_SRC_DIR/pi \
    -DPython3_EXECUTABLE=$(which python) \
    -S $PI_SRC_DIR \
    -B $PI_BUILD_DIR

where $PI_SRC_DIR is the path to the checkout of this repo and $PI_BUILD_DIR is where you want to build into. Then

cmake --build $PI_BUILD_DIR --target install

which will install _mlir_libs, dialects, and runtime underneath $PI_SRC_DIR/pi/mlir. Then add $PI_SRC_DIR to your PYTHONPATH and you're good to go. E.g.,

PYTHONPATH=$PI_SRC_DIR pytest $PI_SRC_DIR/tests/unit/*

Why build the install target? Because you can't do a pip install . -e (editable install) because of the pybind/C-extension so this is the next best thing.

Note, if you're using CLion and you're getting something like

Process finished with exit code 127

you need to disable Add content roots to PYTHONPATH and Add source roots to PYTHONPATH in Run/Debug Configurations.

If you're fancy you can add these CMake flags:

-DCMAKE_EXE_LINKER_FLAGS_INIT="-fuse-ld=lld"
-DCMAKE_MODULE_LINKER_FLAGS_INIT="-fuse-ld=lld"
-DCMAKE_SHARED_LINKER_FLAGS_INIT="-fuse-ld=lld"
-DCMAKE_C_COMPILER_LAUNCHER=ccache
-DCMAKE_CXX_COMPILER_LAUNCHER=ccache