Merge branch 'main' into cmake

sstamenova · Sep 18, 2023 · 289c1d6 · 289c1d6
2 parents 0c8ab1c + 5aca454
commit 289c1d6
Show file tree

Hide file tree

Showing 539 changed files with 29,675 additions and 26,169 deletions.
diff --git a/.azure-pipelines/Windows-CI.yml b/.azure-pipelines/Windows-CI.yml
@@ -11,7 +11,8 @@ parameters:
 
 jobs:
 - job: Build_onnx_mlir_Windows
-  timeoutInMinutes: 240
+  # 4h timeout is sometimes a tiny bit short when llvm-project is rebuilt
+  timeoutInMinutes: 270
   pool:
     vmImage: 'windows-2019'
 
@@ -142,14 +143,20 @@ jobs:
 
   - script: |
       call "%ProgramFiles(x86)%\Microsoft Visual Studio\2019\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" x64
-      call onnx-mlir\utils\check-onnx-numerical.cmd
-    displayName: Run onnx-mlir numerical tests
+      call onnx-mlir\utils\check-docs.cmd
+    displayName: Run onnx-mlir doc tests
     workingDirectory: $(Agent.BuildDirectory)
-    env:
-      CTEST_PARALLEL_LEVEL: ${{ parameters.CTEST_PARALLEL_LEVEL }}
 
   - script: |
       call "%ProgramFiles(x86)%\Microsoft Visual Studio\2019\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" x64
-      call onnx-mlir\utils\check-docs.cmd
-    displayName: Run onnx-mlir doc tests
+      call onnx-mlir\utils\check-unittest.cmd
+    displayName: Run onnx-mlir unit tests
     workingDirectory: $(Agent.BuildDirectory)
+
+  - script: |
+      call "%ProgramFiles(x86)%\Microsoft Visual Studio\2019\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" x64
+      call onnx-mlir\utils\check-onnx-numerical.cmd
+    displayName: Run onnx-mlir numerical tests
+    workingDirectory: $(Agent.BuildDirectory)
+    env:
+      CTEST_PARALLEL_LEVEL: ${{ parameters.CTEST_PARALLEL_LEVEL }}
diff --git a/.github/workflows/macos-amd64-build.yml b/.github/workflows/macos-amd64-build.yml
@@ -6,7 +6,7 @@ jobs:
   build:
     runs-on: macos-latest
     steps:
-    - uses: actions/checkout@v2
+    - uses: actions/checkout@v3
       with:
         submodules: recursive
     - uses: actions/setup-python@v4
@@ -24,12 +24,10 @@ jobs:
         sh ~/work/onnx-mlir/onnx-mlir/utils/install-protobuf.sh
     - name: cache MLIR directory
       id: cache-mlir
-      uses: actions/cache@v2
-      env:
-        cache-name: cache-mlir-directory
+      uses: actions/cache@v3
       with:
         path: ~/work/onnx-mlir/llvm-project
-        key: V8-${{ runner.os }}-build-${{ env.cache-name }}-${{ hashFiles('**/clone-mlir.sh', '**/build-mlir.sh') }}
+        key: ${{ runner.os }}-mlir-${{ hashFiles('**/clone-mlir.sh', '**/build-mlir.sh') }}
     - name: clone & build MLIR
       if: steps.cache-mlir.outputs.cache-hit != 'true'
       run: |
@@ -50,6 +48,14 @@ jobs:
       run: |
         cd ~/work/onnx-mlir
         sh ~/work/onnx-mlir/onnx-mlir/utils/install-onnx-mlir.sh
+    - name: build and run docs/doc_example tests
+      run: |
+        cd ~/work/onnx-mlir
+        sh ~/work/onnx-mlir/onnx-mlir/utils/check-doc-example.sh
+    - name: build and run unit tests
+      run: |
+        cd ~/work/onnx-mlir
+        sh ~/work/onnx-mlir/onnx-mlir/utils/check-unittest.sh
     - name: run onnx-mlir backend and numerical tests
       run: |
         cd ~/work/onnx-mlir

diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -11,10 +11,15 @@ Releases are extensively tested, including the following steps.
 
 # Current releases
 
-## Prerelease 0.4.1
+## Prerelease 0.4.2
 
 < current main branch >
 
+## Prerelease 0.4.1
+
+This prerelease was cut from commit on August 26th, 2023.
+There are no security issues that we know of.
+
 ## Prerelease 0.4.0
 
 This prerelease was cut on March 24th, 2023.

diff --git a/VERSION_NUMBER b/VERSION_NUMBER
@@ -1 +1 @@
-0.4.1
+0.4.2
diff --git a/docker/Dockerfile.llvm-project b/docker/Dockerfile.llvm-project
@@ -28,8 +28,8 @@ RUN distro=$(cat /etc/os-release|grep -Po '(?<=^ID=").*(?=")|(?<=^ID=)[^"].*[^"]
           ln -sf /usr/share/zoneinfo/${TZ} /etc/localtime && \
           dpkg-reconfigure -f noninteractive tzdata && \
           apt-get install -qq -y --no-install-recommends \
-                  autoconf automake ca-certificates cmake cppcheck curl \
-                  default-jdk-headless gcc g++ git libncurses-dev \
+                  autoconf automake ca-certificates clang cmake cppcheck \
+                  curl default-jdk-headless gcc g++ git libncurses-dev \
                   libtool make maven ninja-build openjdk-11-jdk-headless \
                   python3 python3-dev python3-distutils python3-numpy \
                   python3-pip python3-pytest-xdist python3-setuptools \
@@ -38,16 +38,17 @@ RUN distro=$(cat /etc/os-release|grep -Po '(?<=^ID=").*(?=")|(?<=^ID=)[^"].*[^"]
           ln -sf /usr/bin/pytest-3 /usr/bin/pytest; \
        elif [ "${distro}" = "rhel" ] || [ "${distro}" = "fedora" ]; then \
           ln -sf /usr/share/zoneinfo/${TZ} /etc/localtime && \
-	  ([ -x /usr/bin/microdnf ] && microdnf install -y yum) && \
+          ([ -x /usr/bin/microdnf ] && microdnf install -y yum) && \
+          RHEL_VERSION=$(grep CPE_NAME /etc/os-release | cut -d':' -f5) && \
           yum install -q -y \
-              https://dl.fedoraproject.org/pub/epel/epel-release-latest-8.noarch.rpm && \
+              https://dl.fedoraproject.org/pub/epel/epel-release-latest-${RHEL_VERSION}.noarch.rpm && \
           yum update -q -y && \
           yum install -q -y \
-              autoconf automake ca-certificates cmake diffutils \
+              autoconf automake ca-certificates clang cmake diffutils \
               file java-11-openjdk-devel java-11-openjdk-headless \
               gcc gcc-c++ git libtool make ncurses-devel \
               python39 python39-devel python39-numpy python39-pip \
-              python39-setuptools python39-wheel zlib-devel && \
+              python39-setuptools python39-wheel tzdata-java zlib-devel && \
           # Use same versions as those in ubuntu:jammy
           pip3 install -q \
                Cython pytest==6.2.5 pytest-forked==1.4.0 \
@@ -83,7 +84,9 @@ RUN git clone -n https://github.com/llvm/llvm-project.git \
     && cd llvm-project \
     && git checkout ${LLVM_PROJECT_SHA1} \
     && mkdir -p build && cd build \
-    && cmake -G Ninja ../llvm \
+# Build with clang since gcc on ppc64le doesn't support __builtin_thread_pointer
+    && CC=clang CXX=clang++ \
+       cmake -G Ninja ../llvm \
              -DLLVM_ENABLE_PROJECTS=mlir \
              -DLLVM_TARGETS_TO_BUILD="host" \
              -DCMAKE_BUILD_TYPE=Release \

diff --git a/docker/Dockerfile.onnx-mlir b/docker/Dockerfile.onnx-mlir
@@ -25,6 +25,7 @@ RUN ONNX_ROOT=${WORK_DIR}/onnx-mlir/third_party/onnx \
 
 ARG NPROC=4
 ARG ACCEL=NNPA
+ARG TEST_NOFLOAT16
 ARG TEST_MCPU
 ARG KEEPSRC
 
@@ -39,24 +40,32 @@ RUN LLVM_PROJECT_ROOT=${WORK_DIR}/llvm-project \
     && rm -rf build && mkdir -p build && cd build \
 # NNPA acclerator is built on all archs to enable lit tests
 # (dependent libzdnn is built on s390x only)
-    && cmake -DMLIR_DIR=${LLVM_PROJECT_ROOT}/build/lib/cmake/mlir \
+    && CC=clang CXX=clang++ \
+       cmake -DMLIR_DIR=${LLVM_PROJECT_ROOT}/build/lib/cmake/mlir \
              -DCMAKE_BUILD_TYPE=Release \
              -DCMAKE_INSTALL_MESSAGE=NEVER \
              -DONNX_MLIR_ACCELERATORS=${ACCEL} .. \
     && make -j${NPROC} \
     && make -j${NPROC} LIT_OPTS=-v check-onnx-lit \
+# FLOAT16 backend tests only work on ppc64le platform at the moment
+    && TEST_NOFLOAT16=${TEST_NOFLOAT16:-$([ "$(uname -m)" = "s390x" ] && echo true || \
+                                         ([ "$(uname -m)" = "x86_64" ] &&  echo true || \
+                                         ([ "$(uname -m)" = "ppc64le" ] && echo || echo)))} \
 # User image is built with SIMD (currently on s390x only)
     && TEST_MCPU=${TEST_MCPU:-$([ "$(uname -m)" = "s390x" ] && echo z14 || \
-                                [ "$(uname -m)" = "x86_64" ] &&  echo || \
-                                [ "$(uname -m)" = "ppc64le" ] && echo || echo)} \
+                               ([ "$(uname -m)" = "x86_64" ] &&  echo || \
+                               ([ "$(uname -m)" = "ppc64le" ] && echo || echo)))} \
     && TEST_ARGS="-mcpu=${TEST_MCPU}" \
+    && make check-docs \
+    && make check-unittest \
+    && make check-multiple-models \
     && make NPROC=${NPROC} \
             CTEST_PARALLEL_LEVEL=${NPROC} \
+            TEST_NOFLOAT16=${TEST_NOFLOAT16} \
             TEST_MCPU=${TEST_MCPU} \
             TEST_ARGS="${TEST_ARGS}" \
             -j${NPROC} \
             check-onnx-backend-numerical \
-    && make check-docs \
     && make -j${NPROC} install \
     && echo -e "/usr/local/lib" > \
             /etc/ld.so.conf.d/onnx-mlir.conf && ldconfig \
@@ -72,8 +81,8 @@ RUN LLVM_PROJECT_ROOT=${WORK_DIR}/llvm-project \
           pip3 uninstall -q -y Cython pybind11 pytest pytest-forked \
                pytest-xdist typing-extensions && \
           yum remove -q -y \
-              adwaita-icon-theme autoconf automake cmake diffutils file \
-              git libtool make python39 && \
+              adwaita-icon-theme autoconf automake cmake file \
+              git libtool python39 && \
           rm -rf /var/cache/dnf/* /usr/local/bin/ninja; \
        fi \
     && rm -rf /tmp/* /usr/bin/python \

diff --git a/docker/Dockerfile.onnx-mlir-dev b/docker/Dockerfile.onnx-mlir-dev
@@ -19,6 +19,7 @@ RUN ONNX_ROOT=${WORK_DIR}/onnx-mlir/third_party/onnx \
 
 ARG NPROC=4
 ARG ACCEL=NNPA
+ARG TEST_NOFLOAT16
 ARG TEST_MCPU
 
 RUN LLVM_PROJECT_ROOT=${WORK_DIR}/llvm-project \
@@ -28,7 +29,8 @@ RUN LLVM_PROJECT_ROOT=${WORK_DIR}/llvm-project \
     && rm -rf build && mkdir -p build && cd build \
 # NNPA acclerator is built on all archs to enable lit tests
 # (dependent libzdnn is built on s390x only)
-    && cmake -DMLIR_DIR=${LLVM_PROJECT_ROOT}/build/lib/cmake/mlir \
+    && CC=clang CXX=clang++ \
+       cmake -DMLIR_DIR=${LLVM_PROJECT_ROOT}/build/lib/cmake/mlir \
              -DCMAKE_BUILD_TYPE=Debug \
              -DONNX_MLIR_TEST_OPTLEVEL=0 \
              -DONNX_MLIR_ACCELERATORS=${ACCEL} .. \
@@ -44,20 +46,27 @@ RUN LLVM_PROJECT_ROOT=${WORK_DIR}/llvm-project \
        fi \
     && make -j${NPROC} \
     && make -j${NPROC} LIT_OPTS=-v check-onnx-lit \
+# FLOAT16 backend tests only work on ppc64le platform at the moment
+    && TEST_NOFLOAT16=${TEST_NOFLOAT16:-$([ "$(uname -m)" = "s390x" ] && echo true || \
+                                         ([ "$(uname -m)" = "x86_64" ] &&  echo true || \
+                                         ([ "$(uname -m)" = "ppc64le" ] && echo || echo)))} \
 # Dev image is built without SIMD, placeholder for easy SIMD enablement
     && TEST_MCPU=$([ "$(uname -m)" = "s390x" ] && echo || \
-                   [ "$(uname -m)" = "x86_64" ] &&  echo || \
-                   [ "$(uname -m)" = "ppc64le" ] && echo || echo) \
+                  ([ "$(uname -m)" = "x86_64" ] &&  echo || \
+                  ([ "$(uname -m)" = "ppc64le" ] && echo || echo))) \
     && TEST_ARGS="-mcpu=${TEST_MCPU}" \
     && TEST_OPTLEVEL=0 \
+    && make check-docs \
+    && make check-unittest \
+    && make check-multiple-models \
     && make NPROC=${NPROC} \
             CTEST_PARALLEL_LEVEL=${NPROC} \
+            TEST_NOFLOAT16=${TEST_NOFLOAT16} \
             TEST_MCPU=${TEST_MCPU} \
             TEST_ARGS="${TEST_ARGS}" \
             TEST_OPTLEVEL=${TEST_OPTLEVEL} \
             -j${NPROC} \
             check-onnx-backend-numerical \
-    && make check-docs \
     && rm -f Debug/bin/*Test Debug/bin/Perf* Debug/bin/Test* \
 # When building for push event to publish the image, unshallow and
 # rename origin to upstream to make the repo a bit more dev friendly.

diff --git a/docs/AddCustomAccelerators.md b/docs/AddCustomAccelerators.md
@@ -50,9 +50,6 @@ We provide a base class [onnx_mlir::accel::Accelerator](../src/Accelerators/Acce
 // Hooks for onnx-mlir driver
 //===--------------------------------------------------------------------===//
 
-/// Load the MLIR dialects necessary to generate code for an accelerator.
-virtual void getOrLoadDialects(mlir::MLIRContext &context) const = 0;
-
 /// Add the transformations necessary to support the accelerator.
 virtual void addPasses(mlir::OwningOpRef<mlir::ModuleOp> &module,
     mlir::PassManager &pm,
@@ -65,9 +62,9 @@ virtual void addPasses(mlir::OwningOpRef<mlir::ModuleOp> &module,
 /// Register the MLIR dialects required to support an accelerator.
 virtual void registerDialects(mlir::DialectRegistry &registry) const = 0;
 
-/// Initialize the transformation passes required to generate code for an
-/// accelerator.
-virtual void initPasses(int optLevel) const = 0;
+/// Register accelerator transformation passes to make available as
+/// command line options.
+virtual void registerPasses(int optLevel) const = 0;
 
 //===--------------------------------------------------------------------===//
 // Hooks for onnx-to-krnl pass

diff --git a/docs/BuildOnLinuxOSX.md b/docs/BuildOnLinuxOSX.md
@@ -4,7 +4,8 @@
 
 We provide here directions to install ONNX-MLIR on Linux and OSX.
 On Mac, there are a couple of commands that are different.
-These differences will be listed in the explanation below, when relevant.
+These differences will be listed in the explanation below, when relevant. Installing ONNX-MLIR on Apple silicon is natively supported and it is recommended to use brew to manage prerequisites.
+
 
 ## MLIR
 
@@ -14,7 +15,7 @@ Firstly, install MLIR (as a part of LLVM-Project):
 ``` bash
 git clone -n https://github.com/llvm/llvm-project.git
 # Check out a specific branch that is known to work with ONNX-MLIR.
-cd llvm-project && git checkout 6cf7fe4a9a715bcdf3f4913753109e22dfc9940b && cd ..
+cd llvm-project && git checkout 4acc3ffbb0af5631bc7916aeff3570f448899647 && cd ..
 ```
 
 [same-as-file]: <> (utils/build-mlir.sh)
@@ -41,7 +42,7 @@ The `MLIR_DIR` cmake variable must be set before building onnx-mlir. It should p
 
 This project uses lit ([LLVM's Integrated Tester](https://llvm.org/docs/CommandGuide/lit.html)) for unit tests. When running cmake, we can also specify the path to the lit tool from LLVM using the `LLVM_EXTERNAL_LIT` variable but it is not required as long as MLIR_DIR points to a build directory of llvm-project. If `MLIR_DIR` points to an install directory of llvm-project, `LLVM_EXTERNAL_LIT` is required.
 
-To build ONNX-MLIR, use the following commands:
+To build ONNX-MLIR, use the following commands (maybe with additional `-DCMAKE_CXX_FLAGS` argument described [below](#enable-cpu-optimizations)):
 
 [same-as-file]: <> ({"ref": "utils/install-onnx-mlir.sh", "skip-doc": 2})
 ```bash
@@ -75,9 +76,16 @@ The environment variable `$pythonLocation` may be used to specify the base direc
 
 After the above commands succeed, an `onnx-mlir` executable should appear in the `Debug/bin` or `Release/bin` directory.
 
+### Enable CPU Optimizations
+
+To make the compiler run faster (without any affect on the generated code)
+you can pass `-DCMAKE_CXX_FLAGS=-march=native` to the `cmake -G Ninja ..` build configuration step above to generate code that exploits all the features of your local CPU, at the expense of portability. Or you can enable a specific CPU feature, e.g. `-DCMAKE_CXX_FLAGS=-mf16c` to enable the F16C feature to enable native conversion between float16 and (32 bit) float. It is used in `src/Support/SmallFP.hpp`.
+
 ### Known MacOS Issues
 
-There is a known issue when building onnx-mlir. If you see a error of this sorts:
+#### jsoniter issue
+
+There is a known issue when building onnx-mlir. If you see an error of this sorts:
 
 ``` shell
 Cloning into '/home/user/onnx-mlir/build/src/Runtime/jni/jsoniter'...
@@ -91,6 +99,22 @@ make: *** [Makefile:146: all] Error 2.
 
 The suggested workaround until jsoniter is fixed is as follows: install maven (e.g. `brew install maven`) and run `alias nproc="sysctl -n hw.logicalcpu"` in your shell.
 
+#### Protobuf issue (Mac M1, specific to protobuf 3.20.3 which is currently required)
+
+On Mac M1, you may have some issues building protobuf. In particular, you may fail to install onnx (via `pip install -e third_party/onnx`) or you may fail to compile `onnx-mlir` (no arm64 symbol for `InternalMetadata::~InternalMetadata`).
+
+The first failure is likely an issue with having multiple versions of protobuf.
+Installing a version with `brew` was not helpful (version 3.20.3 because of a known bug that can be corrected with a patch below).
+Uninstall the brew version, and make sure you install the right one with pip: `pip install protobuf==3.20.3`.
+
+The second failure can be remediated by downloading protobuf source code, applying a patch, and installing it on the local machine.
+See [Dockerfile.llvm-project](../docker/Dockerfile.llvm-project) on line 66 for cloning instructions. After cloning the right version, you should apply a patch [patch](https://github.com/protocolbuffers/protobuf/commit/0574167d92a232cb8f5a9107aabda0aefbc39e8b) by downloading from the link above and applying it.
+Then you should follow the steps in the [Dockerfile.llvm-project](../docker/Dockerfile.llvm-project) file (skipped the `ldconfig` step without consequences).
+You may have to brew a couple of the tools, see the `yum install` in the `Dockerfile.llvm-project` file above.
+You should then be able to successfully install protobuf and compile `onnx-mlir`.
+As the dependences between `third_party` and `onnx-mlir` might cause issues, it is always safe to delete the `third_party` directory, reinstall using `git submodule update --init --recursive`, reinstall `onnx`, delete `onnx-mlir/build` and rebuild `onnx-mlir` from scratch.
+
+
 ### Trouble shooting build issues
 
 Check this [page](TestingHighLevel.md) for helpful hints.
diff --git a/docs/BuildOnWindows.md b/docs/BuildOnWindows.md
@@ -52,7 +52,7 @@ Install MLIR (as a part of LLVM-Project):
 ```shell
 git clone -n https://github.com/llvm/llvm-project.git
 # Check out a specific branch that is known to work with ONNX-MLIR.
-cd llvm-project && git checkout 6cf7fe4a9a715bcdf3f4913753109e22dfc9940b && cd ..
+cd llvm-project && git checkout 4acc3ffbb0af5631bc7916aeff3570f448899647 && cd ..
 ```
 
 [same-as-file]: <> (utils/build-mlir.cmd)

diff --git a/docs/ConstPropagationPass.md b/docs/ConstPropagationPass.md
@@ -73,6 +73,7 @@ class Pattern<
    dag sourcePattern,
    list<dag> resultPatterns,
    list<dag> additionalConstraints = [],
+   list<dag> supplementalPatterns = [],
    dag benefitsAdded = (addBenefit 0)
 >;
 ```

diff --git a/docs/DebuggingNumericalError.md b/docs/DebuggingNumericalError.md
@@ -62,6 +62,12 @@ optional arguments:
   --upper-bound UPPER_BOUND   Upper bound values for each data type. Used to generate random inputs. E.g. --upper-bound=int64:10,float32:0.2,uint8:9. Supported types are bool, uint8, int8, uint16, int16, uint32, int32, uint64, int64, float16, float32, float64
 ```
 
+## Helper script to compare a model under two distinct compile option.
+
+Based on the above `utils/runONNXModel.py`, the `utils/checkONNXModel.py` allows a user to run a given model twice, under two distinct compile options, and compare its results.
+This let a user simply test a new option, comparing the safe version of the compiler (e.g. `-O0` or `-O3`) with a more advanced version (e.g. `-O3` or `-O3 -march=x86-64`). Simply specify the compile options using the `--ref-compile-args` and `--test-compile-args` flags, a model using the `--model` flag, and possibly a `--shape-info` in presence of dynamic shape inputs.
+Full options are listed under the `--help` flag.
+
 ## Debugging the Code Generated for an Operator.
 
 If you know, or suspect, that a particular ONNX MLIR operator produces an incorrect result, and want to narrow down the problem, we provide a couple of useful Krnl operators that allow printing (at runtime) the value of a tensor, or a value that has a primitive data type.