torch-mlir/python/npcomp/compiler/pytorch/backend/iree.py

114 lines
3.4 KiB
Python

# Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
# See https://llvm.org/LICENSE.txt for license information.
# SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
import os
import torch
import numpy as np
from mlir.ir import *
from mlir.passmanager import *
from npcomp.compiler.utils import logging
import iree.runtime as ireert
import iree.compiler as ireec
from .abc import NpcompBackend
__all__ = [
"IreeNpcompBackend",
]
PREPARE_FOR_IREE_PASSES = (
"npcomp-iree-backend-lower-linkage",
)
class IreeModuleInvoker:
"""Wrapper around a native IREE module for calling functions."""
def __init__(self, iree_module):
super().__init__()
self._iree_module = iree_module
def __getattr__(self, function_name):
return self.__getitem__(function_name)
def __getitem__(self, function_name):
def invoke(*args):
results = self._iree_module[function_name](*args)
if isinstance(results, np.ndarray):
return results
if len(results) == 1:
# De-tuple.
return results[0]
else:
return tuple(results)
invoke.__isnpcomp__ = True
return invoke
class TorchIreeModuleInvoker(IreeModuleInvoker):
"""Allows torch.Tensor inputs to be passed to module invocations."""
def __getitem__(self, function_name: str):
numpy_invoke = super().__getitem__(function_name)
def invoke(*args):
args = tuple(
arg.numpy() if isinstance(arg, torch.Tensor) else arg for arg in args)
return numpy_invoke(*args)
return invoke
class IreeNpcompBackend(NpcompBackend):
"""Main entry-point for the backend."""
def __init__(self):
super().__init__()
self._debug = logging.debug_enabled()
def compile(self, imported_module: Module):
"""Compiles an imported module, with a flat list of functions.
The module is expected to conform to the npcomp backend contract.
See the VerifyBackendContract pass for more details.
Args:
imported_module: The MLIR module consisting of funcs in the torch
dialect.
Returns:
An opaque, backend specific module object that can be passed to load.
The object may actually be something more specific to the backend (i.e.
for IREE, it is a serialized VM flatbuffer) but the contract is that
it is operated on by methods on this class.
"""
with imported_module.context as context:
if self._debug:
logging.debug("IR passed to IREE compiler backend:\n{}",
imported_module)
pipeline_str = ",".join(PREPARE_FOR_IREE_PASSES)
if self._debug:
logging.debug("Running Prepare For IREE pipeline '{}'", pipeline_str)
pm = PassManager.parse(pipeline_str)
pm.run(imported_module)
if self._debug:
logging.debug(
"IREE Input IR (this is what IREE's compiler will see):\n{}",
imported_module)
# Backend.
binary = ireec.compile_str(str(imported_module),
target_backends=["dylib-llvm-aot"])
return binary
def load(self, iree_module) -> TorchIreeModuleInvoker:
"""Loads a compiled artifact into the runtime."""
vm_module = ireert.VmModule.from_flatbuffer(iree_module)
iree_config = ireert.Config(driver_name="dylib")
ctx = ireert.SystemContext(config=iree_config)
ctx.add_vm_module(vm_module)
return TorchIreeModuleInvoker(ctx.modules.module)