Merge pull request #153 from Xilinx/matthias.bump_torch_mlir3

Bump torch mlir (Dec 12)

.github/workflows/bazelBuildAndTest.yml

@@ -10,7 +10,11 @@ on:
 # any in-progress jobs in the same github workflow and github
 # ref (e.g. refs/heads/main or refs/pull/<pr_number>/merge).
 concurrency:
-  group: ${{ github.workflow }}-${{ github.ref }}
+  # A PR number if a pull request and otherwise the commit hash. This cancels
+  # queued and in-progress runs for the same PR (presubmit) or commit
+  # (postsubmit). The workflow name is prepended to avoid conflicts between
+  # different workflows.
+  group: ${{ github.workflow }}-${{ github.event.number || github.sha }}
   cancel-in-progress: true
 
 

.github/workflows/buildAndTest.yml

@@ -11,7 +11,11 @@ on:
 # any in-progress jobs in the same github workflow and github
 # ref (e.g. refs/heads/main or refs/pull/<pr_number>/merge).
 concurrency:
-  group: ${{ github.workflow }}-${{ github.ref }}
+  # A PR number if a pull request and otherwise the commit hash. This cancels
+  # queued and in-progress runs for the same PR (presubmit) or commit
+  # (postsubmit). The workflow name is prepended to avoid conflicts between
+  # different workflows.
+  group: ${{ github.workflow }}-${{ github.event.number || github.sha }}
   cancel-in-progress: true
 
 
@@ -24,17 +28,14 @@ jobs:
     strategy:
       fail-fast: true
       matrix:
-        os-arch: [ubuntu-x86_64] # macos-arm64, windows-x86_64
-        llvm-build: [in-tree] # out-of-tree
-        torch-binary: [ON] # OFF
+        os-arch: [ubuntu-x86_64] #, macos-arm64, windows-x86_64]
+        llvm-build: [in-tree] #, out-of-tree]
+        torch-binary: [ON]
         torch-version: [nightly, stable]
         exclude:
-          # Exclude llvm in-tree and pytorch source
-          - llvm-build: in-tree
-            torch-binary: OFF
-          # Exclude llvm out-of-tree and pytorch binary
+          # Exclude llvm out-of-tree and pytorch stable (to save resources)
           - llvm-build: out-of-tree
-            torch-binary: ON
+            torch-version: stable
           # Exclude macos-arm64 and llvm out-of-tree altogether
           - os-arch: macos-arm64
             llvm-build: out-of-tree
@@ -44,9 +45,6 @@ jobs:
             llvm-build: out-of-tree
           - os-arch: windows-x86_64
             torch-version: stable
-          # For PyTorch stable builds, we don't build PyTorch from source
-          - torch-version: stable
-            torch-binary: OFF
         include:
           # Specify OS versions
           - os-arch: ubuntu-x86_64

.github/workflows/buildRelease.yml

@@ -23,11 +23,6 @@ jobs:
         package: [ torch-mlir ]
         py_version: [ cp38-cp38, cp310-cp310 ] # cp311-cp311
         torch-version: [stable] # nightly
-        exclude:
-          - package: torch-mlir-core
-            py_version: cp38-cp38
-          - package: torch-mlir-core
-            py_version: cp310-cp310
 
     steps:
 
@@ -99,7 +94,7 @@ jobs:
     runs-on: linux-arm64
     strategy:
       matrix:
-        package: [ torch-mlir, torch-mlir-core ]
+        package: [ torch-mlir ]
         py_version: [ cp311-cp311 ]
 
     steps:
@@ -169,7 +164,7 @@ jobs:
     runs-on: macos-latest
     strategy:
       matrix:
-        package: [ torch-mlir, torch-mlir-core ]
+        package: [ torch-mlir ]
     steps:
     - name: Get torch-mlir
       uses: actions/checkout@v3
@@ -230,7 +225,7 @@ jobs:
     runs-on: windows-latest
     strategy:
       matrix:
-        package: [ torch-mlir, torch-mlir-core ]
+        package: [ torch-mlir ]
     steps:
     - name: Get torch-mlir
       uses: actions/checkout@v3
@@ -246,14 +241,8 @@ jobs:
     - name: Build Python wheels and smoke test.
       shell: pwsh
       run: |
-        if ( "${{ matrix.package }}" -eq "torch-mlir-core" )
-        {
-          $env:TORCH_MLIR_ENABLE_JIT_IR_IMPORTER='0'
-          $env:TORCH_MLIR_ENABLE_ONLY_MLIR_PYTHON_BINDINGS='1'
-        } else {
-          $env:TORCH_MLIR_ENABLE_JIT_IR_IMPORTER='1'
-          $env:TORCH_MLIR_ENABLE_ONLY_MLIR_PYTHON_BINDINGS='0'
-        }
+        $env:TORCH_MLIR_ENABLE_JIT_IR_IMPORTER='1'
+        $env:TORCH_MLIR_ENABLE_ONLY_MLIR_PYTHON_BINDINGS='0'
         $env:TORCH_MLIR_PYTHON_PACKAGE_VERSION = '${{ github.event.inputs.python_package_version }}'
         ./build_tools/python_deploy/build_windows.ps1
 

.gitignore

@@ -26,7 +26,7 @@ __pycache__
 bazel-*
 
 # Autogenerated files
-/projects/pt1/python/torch_mlir/csrc/base_lazy_backend/generated
+/projects/ltc/csrc/base_lazy_backend/generated
 
 #Docker builds
 build_oot/

CMakeLists.txt

@@ -29,6 +29,8 @@ set(CMAKE_CXX_STANDARD 17)
 # Project options
 #-------------------------------------------------------------------------------
 
+option(TORCH_MLIR_USE_INSTALLED_PYTORCH "If depending on PyTorch use it as installed in the current Python environment" ON)
+
 option(TORCH_MLIR_ENABLE_REFBACKEND "Enable reference backend" ON)
 if(TORCH_MLIR_ENABLE_REFBACKEND)
   add_definitions(-DTORCH_MLIR_ENABLE_REFBACKEND)
@@ -149,10 +151,12 @@ endfunction()
 # Configure CMake.
 list(APPEND CMAKE_MODULE_PATH ${MLIR_MAIN_SRC_DIR}/cmake/modules)
 list(APPEND CMAKE_MODULE_PATH ${LLVM_MAIN_SRC_DIR}/cmake)
+list(APPEND CMAKE_MODULE_PATH ${CMAKE_CURRENT_SOURCE_DIR}/build_tools/cmake)
 
 include(TableGen)
 include(AddLLVM)
 include(AddMLIR)
+include(AddMLIRPython)
 
 ################################################################################
 # Setup python.
@@ -231,6 +235,4 @@ endif()
 # Sub-projects
 #-------------------------------------------------------------------------------
 
-if(TORCH_MLIR_ENABLE_PROJECT_PT1)
-  add_subdirectory(projects/pt1)
-endif()
+add_subdirectory(projects)

build_tools/autogen_ltc_backend.py

@@ -29,7 +29,6 @@
     TORCH_INCLUDE_DIR = TORCH_DIR
 TORCHGEN_DIR = Path(torchgen.__path__[0]).resolve()
 TORCH_MLIR_DIR = Path(__file__).resolve().parent.parent
-TORCH_MLIR_PT1_DIR = TORCH_MLIR_DIR / "projects" / "pt1"
 
 def reindent(text, prefix=""):
     return indent(dedent(text), prefix)
@@ -114,12 +113,12 @@ def __init__(self, binary_dir):
         self.binary_dir = Path(binary_dir)
         assert self.binary_dir.is_dir(), f"Binary directory not found: {self.binary_dir}"
         self.source_yaml = self.binary_dir.joinpath("generated_native_functions.yaml")
-        self.backend_path = TORCH_MLIR_PT1_DIR.joinpath(
-            "python", "torch_mlir", "csrc", "base_lazy_backend"
+        self.backend_path = TORCH_MLIR_DIR.joinpath(
+            "projects", "ltc", "csrc", "base_lazy_backend"
         )
         assert self.backend_path.is_dir(), f"Backend path not found: {self.backend_path}"
         self.generated_path = self.binary_dir.joinpath(
-            "projects", "pt1", "python", "torch_mlir", "csrc", "base_lazy_backend", "generated"
+            "projects", "ltc", "csrc", "base_lazy_backend", "generated"
         )
         self.generated_path.mkdir(parents=True, exist_ok=True)
 
@@ -415,7 +414,7 @@ def extract_signatures(text):
                     // for ops that dont have a corresponding structured kernel or shape definition
 
                     #include "shape_inference.h"
-                    #include "torch_mlir/csrc/base_lazy_backend/utils/exception.h"
+                    #include "base_lazy_backend/utils/exception.h"
                     namespace torch {{
                     namespace lazy {{
                     {}
@@ -467,7 +466,7 @@ def gen_fallback_code(*args, **kwargs):
             node_base="torch::lazy::TorchMlirNode",
             node_base_hdr=str(self.backend_path.joinpath("mlir_node.h")),
             tensor_class=self.tensor_class,
-            tensor_class_hdr="torch_mlir/csrc/base_lazy_backend/tensor.h",
+            tensor_class_hdr="base_lazy_backend/tensor.h",
             create_aten_from_ltc_tensor="CreateFunctionalizedAtenFromLtcTensor",
             shape_inference_hdr=str(self.generated_path.joinpath("shape_inference.h")),
             lazy_ir_generator=GenMlirLazyIr,

build_tools/python_deploy/build_linux_packages.sh

@@ -364,9 +364,9 @@ function setup_venv() {
 function build_out_of_tree() {
   local torch_from_bin="$1"
   local python_version="$2"
-  echo ":::: Build out-of-tree Torch from binary: $torch_from_bin with Python: $python_version"
-
   local torch_version="$3"
+  echo ":::: Build out-of-tree Torch from binary: $torch_from_bin with Python: $python_version ($torch_version)"
+
   local enable_ltc="ON"
   if [[ "${torch_version}" == "stable" ]]
   then

build_tools/update_abstract_interp_lib.sh

@@ -42,6 +42,6 @@ if [ ! -z ${TORCH_MLIR_EXT_MODULES} ]; then
 fi
 
 PYTHONPATH="${pypath}" python \
-  -m torch_mlir.dialects.torch.importer.jit_ir.build_tools.abstract_interp_lib_gen \
+  -m torch_mlir.jit_ir_importer.build_tools.abstract_interp_lib_gen \
   --pytorch_op_extensions=${ext_module:-""} \
   --torch_transforms_cpp_dir="${torch_transforms_cpp_dir}"

build_tools/update_torch_ods.sh

@@ -43,7 +43,7 @@ fi
 
 set +u
 PYTHONPATH="${PYTHONPATH}:${pypath}" python \
-  -m torch_mlir.dialects.torch.importer.jit_ir.build_tools.torch_ods_gen \
+  -m torch_mlir.jit_ir_importer.build_tools.torch_ods_gen \
   --torch_ir_include_dir="${torch_ir_include_dir}" \
   --pytorch_op_extensions="${ext_module}" \
   --debug_registry_dump="${torch_ir_include_dir}/JITOperatorRegistryDump.txt"

build_tools/write_env_file.sh

@@ -13,7 +13,7 @@ portable_realpath() {
 
 td="$(portable_realpath "$(dirname "$0")"/..)"
 build_dir="$(portable_realpath "${TORCH_MLIR_BUILD_DIR:-$td/build}")"
-python_packages_dir="$build_dir/tools/torch-mlir/python_packages"
+python_packages_dir="$build_dir/python_packages"
 
 write_env_file() {
   echo "Updating $build_dir/.env file"

docs/Torch-ops-E2E-implementation.md

@@ -17,7 +17,7 @@ The end-to-end test is important to check the correctness of the other steps.
 
 ### Step 2. Update ods
 
-Update [torch_ods_gen.py](https://github.com/llvm/torch-mlir/blob/main/projects/pt1/python/torch_mlir/dialects/torch/importer/jit_ir/build_tools/torch_ods_gen.py) with the new op and run [update_torch_ods.sh](https://github.com/llvm/torch-mlir/blob/main/build_tools/update_torch_ods.sh) to generate the ods. Running `update_torch_ods.sh` would dump all the operators with schema into `JITOperatorRegistryDump.txt`. It’s convenient to look for ops signatures and operands names in this file.
+Update [torch_ods_gen.py](https://github.com/llvm/torch-mlir/blob/main/projects/pt1/python/torch_mlir/jit_ir_importer/build_tools/torch_ods_gen.py) with the new op and run [update_torch_ods.sh](https://github.com/llvm/torch-mlir/blob/main/build_tools/update_torch_ods.sh) to generate the ods. Running `update_torch_ods.sh` would dump all the operators with schema into `JITOperatorRegistryDump.txt`. It’s convenient to look for ops signatures and operands names in this file.
 
 ### Step 3. Propagate types
 It’s essential to make sure the new op implements shape and dtype inference. See [abstract_interp_lib](https://github.com/llvm/torch-mlir/blob/main/docs/abstract_interp_lib.md) for information on adding shape and dtype inference.

docs/abstract_interp_lib.md

@@ -26,7 +26,7 @@ The two main use cases are:
 ## Architecture
 
 Functions are defined as TorchScript-able Python functions in
-`python/torch_mlir/dialects/torch/importer/jit_ir/build_tools/abstract_interp_lib_gen.py`.
+`python/torch_mlir/jit_ir_importer/build_tools/abstract_interp_lib_gen.py`.
 The signatures of the functions are systematically derived from Torch JIT
 operator registry. Most shape functions are expected to reuse the upstream
 helper functions

docs/adding_an_e2e_test.md

@@ -5,7 +5,7 @@
 Adding support for a Torch operator in Torch-MLIR should always be accompanied
 by at least one end-to-end test to make sure the implementation of the op
 matches the behavior of PyTorch. The tests live in the
-`torch-mlir/python/torch_mlir_e2e_test/test_suite/` directory. When adding a new
+`torch-mlir/projects/pt1/python/torch_mlir_e2e_test/test_suite` directory. When adding a new
 test, choose a file that best matches the op you're testing, and if there is no
 file that best matches add a new file for your op.
 
@@ -87,7 +87,7 @@ following order:
 
 1. Shape of input tensor. Use `-1` for dynamic dimensions
 2. Dtype of the input tensor
-3. Boolean representing whether the input tensor [has value semantics](https://github.com/llvm/torch-mlir/blob/ba17a4d6c09b4bbb4ef21b1d8d4a93cb056be109/python/torch_mlir/dialects/torch/importer/jit_ir/csrc/class_annotator.h#L54-L67). This
+3. Boolean representing whether the input tensor [has value semantics](https://github.com/llvm/torch-mlir/blob/ba17a4d6c09b4bbb4ef21b1d8d4a93cb056be109/python/torch_mlir/jit_ir_importer/csrc/class_annotator.h#L54-L67). This
    will always be true for E2E tests, since the [Torch-MLIR backend contract](architecture.md#the-backend-contract) requires all tensors in the
    IR to eventually have value semantics.
 

docs/architecture.md

@@ -55,22 +55,22 @@ factored such that we can handle this with one core import path, which is
 through the PyTorch
 "[JIT IR](https://github.com/pytorch/pytorch/blob/78c8a0d75220bdd4955415b5f81509e005af4232/torch/csrc/jit/OVERVIEW.md)",
 and lives in
-[torch-mlir/python/torch_mlir/dialects/torch/importer/jit_ir](https://github.com/llvm/torch-mlir/tree/e322f6a8784009b37aa354abfa9a40a80f30877d/python/torch_mlir/dialects/torch/importer/jit_ir).
+[torch-mlir/python/torch_mlir/jit_ir_importer](https://github.com/llvm/torch-mlir/tree/e322f6a8784009b37aa354abfa9a40a80f30877d/python/torch_mlir/dialects/torch/importer/jit_ir).
 The JIT IR is a highly principled IR that faithfully models a Python subset (+
 tensors, the PyTorch op registry, and a few other things). All the other PyTorch
 program representations can eventually bottom-out on the JIT IR via some path
 provided by PyTorch. The `torch` dialect is almost entirely in 1:1
 correspondence with the JIT IR -- this allows the importer to be extremely small
 (the core is
-[under 500 lines of code](https://github.com/llvm/torch-mlir/blob/e322f6a8784009b37aa354abfa9a40a80f30877d/python/torch_mlir/dialects/torch/importer/jit_ir/csrc/node_importer.cpp#L1)).
+[under 500 lines of code](https://github.com/llvm/torch-mlir/blob/e322f6a8784009b37aa354abfa9a40a80f30877d/python/torch_mlir/jit_ir_importer/csrc/node_importer.cpp#L1)).
 
 ### Ops
 
 See [TorchOps.td](https://github.com/llvm/torch-mlir/blob/114f48e96c578ee76a6f83b3aa4aa229a8d5b76e/include/torch-mlir/Dialect/Torch/IR/TorchOps.td#L1)
 
 The ops in the `torch` dialect are almost entirely generated based on the
 PyTorch JIT IR operator registry via the script
-[torch_ods_gen.py](https://github.com/llvm/torch-mlir/blob/e322f6a8784009b37aa354abfa9a40a80f30877d/python/torch_mlir/dialects/torch/importer/jit_ir/build_tools/torch_ods_gen.py#L1) (invoked via [update_torch_ods.sh](https://github.com/llvm/torch-mlir/blob/main/build_tools/update_torch_ods.sh)).
+[torch_ods_gen.py](https://github.com/llvm/torch-mlir/blob/e322f6a8784009b37aa354abfa9a40a80f30877d/python/torch_mlir/jit_ir_importer/build_tools/torch_ods_gen.py#L1) (invoked via [update_torch_ods.sh](https://github.com/llvm/torch-mlir/blob/main/build_tools/update_torch_ods.sh)).
 This script queries the registry and generates MLIR
 [ODS](https://mlir.llvm.org/docs/OpDefinitions/) in
 [GeneratedTorchOps.td](https://github.com/llvm/torch-mlir/blob/e322f6a8784009b37aa354abfa9a40a80f30877d/include/torch-mlir/Dialect/Torch/IR/GeneratedTorchOps.td#L1). We have a guide for [adding a new op end-to-end](https://github.com/llvm/torch-mlir/wiki/Torch-ops-E2E-implementation).
@@ -195,7 +195,7 @@ values. When one `torch.jit.script`'s a `torch.nn.Module`, the result is
 actually an `IValue` that represents the module, with a hierarchy of children
 `IValue`'s. Strictly speaking, JIT IR `torch::jit::Graph`'s are only used to
 represent the bodies of methods on the modules. So in addition to importing the
-JIT IR, we also need to import the `IValue`'s. This happens inside [ivalue_importer.cpp](https://github.com/llvm/torch-mlir/blob/fde390c7669e29362b18388448ef2b188713383f/python/torch_mlir/dialects/torch/importer/jit_ir/csrc/ivalue_importer.cpp#L1).
+JIT IR, we also need to import the `IValue`'s. This happens inside [ivalue_importer.cpp](https://github.com/llvm/torch-mlir/blob/fde390c7669e29362b18388448ef2b188713383f/python/torch_mlir/jit_ir_importer/csrc/ivalue_importer.cpp#L1).
 
 Most of the IValue modeling can reuse `torch` dialect ops that already exist
 otherwise, such as `torch.constant.int` to represent an int in the object graph.

docs/development.md

@@ -20,6 +20,7 @@ source mlir_venv/bin/activate
 python -m pip install --upgrade pip
 # Install latest PyTorch nightlies and build requirements.
 python -m pip install -r requirements.txt
+python -m pip install -r torchvision-requirements.txt
 ```
 
 ## CMake Build
@@ -108,25 +109,25 @@ cmake --build build
 ### Linux and macOS
 
 ```shell
-export PYTHONPATH=`pwd`/build/tools/torch-mlir/python_packages/torch_mlir:`pwd`/examples
+export PYTHONPATH=`pwd`/build/python_packages/torch_mlir:`pwd`/projects/pt1/examples
 ```
 
 ### Windows PowerShell
 
 ```shell
-$env:PYTHONPATH = "$PWD/build/tools/torch-mlir/python_packages/torch_mlir;$PWD/examples"
+$env:PYTHONPATH = "$PWD/build/python_packages/torch_mlir;$PWD/projects/pt1/examples"
 ```
 
 ## Testing MLIR output in various dialects
 
-To test the compiler's output to the different MLIR dialects, you can use the example `examples/torchscript_resnet18_all_output_types.py`.
+To test the compiler's output to the different MLIR dialects, you can use the example `projects/pt1/examples/torchscript_resnet18_all_output_types.py`.
 
 Make sure you have activated the virtualenv and set the `PYTHONPATH` above
 (if running on Windows, modify the environment variable as shown above):
 ```shell
 source mlir_venv/bin/activate
-export PYTHONPATH=`pwd`/build/tools/torch-mlir/python_packages/torch_mlir:`pwd`/examples
-python examples/torchscript_resnet18_all_output_types.py
+export PYTHONPATH=`pwd`/build/tpython_packages/torch_mlir:`pwd`/projects/pt1/examples
+python projects/pt1/examples/torchscript_resnet18_all_output_types.py
 ```
 
 This will display the Resnet18 network example in three dialects: TORCH, LINALG on TENSORS and TOSA.
@@ -331,8 +332,8 @@ Torch-MLIR has two types of tests:
 1. End-to-end execution tests. These compile and run a program and check the
    result against the expected output from execution on native Torch. These use
    a homegrown testing framework (see
-   `python/torch_mlir_e2e_test/torchscript/framework.py`) and the test suite
-   lives at `python/torch_mlir_e2e_test/test_suite/__init__.py`.
+   `projects/pt1/python/torch_mlir_e2e_test/framework.py`) and the test suite
+   lives at `projects/pt1/python/torch_mlir_e2e_test/test_suite/__init__.py`.
 
 2. Compiler and Python API unit tests. These use LLVM's `lit` testing framework.
    For example, these might involve using `torch-mlir-opt` to run a pass and