Merge pull request iree-org#6949 from not-jenni:main-to-google

PiperOrigin-RevId: 394551308

benchmarks/README.md

@@ -0,0 +1,62 @@
+# IREE Benchmarks
+
+This directory contains configuration definition for IREE's continuous
+benchmarks suite. Benchmark results are posted to https://perf.iree.dev.
+
+The https://buildkite.com/iree/iree-benchmark Buildkite pipeline runs on each
+commit to the `main` branch and posts those results to the dashboard. The
+pipeline also runs on pull requests with the `buildkite:benchmark` label,
+posting results compared against their base commit as comments.
+
+## Types of benchmarks
+
+```
+├── TensorFlow
+│     * models authored in TensorFlow and imported with `iree-import-tf`
+└── TFLite
+      * models converted to TensorFlow Lite and imported with `iree-import-tflite`
+```
+
+## Adding new benchmarks
+
+### Machine learning model latency
+
+1. Pick the model you want to benchmark and find its source, which could be
+   a Python script, TensorFlow SavedModel from https://tfhub.dev/, TensorFlow
+   Lite FlatBuffer, or some other format with a supported path into IREE. The
+   model can optionally include trained weights if those are important for
+   benchmarking.
+
+2. Import the model into an MLIR file that IREE can compile using the core
+   `iree-translate` tool. For TensorFlow models use `iree-import-tf`, for
+   TensorFlow Lite models use `iree-import-tflite`, etc. Take notes for where
+   the model came from and how it was imported in case the MLIR file needs to
+   be regenerated in the future.
+
+   We may further automate this over time, such as by importing from Python
+   sources as part of the benchmarks pipeline directly (see
+   https://github.com/google/iree/issues/6942). For now, here are some
+   references:
+
+   * https://gist.github.com/antiagainst/35b0989bd0188dd9df4630bb0cf778f2
+   * https://colab.research.google.com/gist/ScottTodd/10838c0ccc87fa6d1b1c72e0fabea064/iree-keyword_spotting_streaming-benchmarks.ipynb
+
+3. Package the imported .mlir model file(s) for storage (see
+   [iree_mlir_benchmark_suite.cmake](../build_tools/cmake/iree_mlir_benchmark_suite.cmake)
+   and [download_file.py](../scripts/download_file.py)), then upload them to the
+   `iree-model-artifacts` Google Cloud Storage bucket with the help of a team
+   member. Files currently hosted in that bucket can be viewed at
+   https://storage.googleapis.com/iree-model-artifacts/index.html.
+
+4. Edit the appropriate `CMakeLists.txt` file under this directory to include
+   your desired benchmark configuration with the `iree_mlir_benchmark_suite`
+   function. You can test your change by running the
+   https://buildkite.com/iree/iree-benchmark pipeline on a GitHub pull request
+   with the `buildkite:benchmark` label.
+
+5. Once your changes are merged to the `main` branch, results will start to
+   appear on the benchmarks dashboard at https://perf.iree.dev.
+
+### Other project metrics
+
+TODO(#6161): Collect metrics for miscellaneous IREE system states

docs/website/docs/building-from-source/android.md

@@ -42,7 +42,7 @@ devices from the command line. Install it following the
 Build and install on your host machine:
 
 ``` shell
-cmake -B ../iree-build/ \
+cmake -GNinja -B ../iree-build/ \
   -DCMAKE_INSTALL_PREFIX=../iree-build/install \
   -DCMAKE_BUILD_TYPE=RelWithDebInfo \
   .
@@ -56,7 +56,7 @@ Build the runtime using the Android NDK toolchain:
 === "Linux and MacOS"
 
     ``` shell
-    cmake -B ../iree-build-android/ \
+    cmake -GNinja -B ../iree-build-android/ \
       -DCMAKE_TOOLCHAIN_FILE="${ANDROID_NDK?}/build/cmake/android.toolchain.cmake" \
       -DIREE_HOST_BINARY_ROOT="$PWD/../iree-build/install" \
       -DANDROID_ABI="arm64-v8a" \
@@ -69,7 +69,7 @@ Build the runtime using the Android NDK toolchain:
 === "Windows"
 
     ``` shell
-    cmake -B ../iree-build-android/ \
+    cmake -GNinja -B ../iree-build-android/ \
       -DCMAKE_TOOLCHAIN_FILE="%ANDROID_NDK%/build/cmake/android.toolchain.cmake" \
       -DIREE_HOST_BINARY_ROOT="%CD%/../iree-build/install" \
       -DANDROID_ABI="arm64-v8a" \

docs/website/docs/building-from-source/getting-started.md

@@ -9,21 +9,29 @@ compiler:
 
 === "Linux and MacOS"
 
-    <!-- TODO(scotttodd): annotation about gcc vs clang -->
+    1. Install a compiler/linker (typically "clang" and "lld" package).
 
-    ``` shell
-    sudo apt install cmake clang
-    export CC=clang
-    export CXX=clang++
-    ```
+    2. Install [CMake](https://cmake.org/download/) (typically "cmake" package).
+
+    3. Install [Ninja](https://ninja-build.org/) (typically "ninja-build"
+       package).
+
+On a relatively recent Debian/Ubuntu:
+
+``` shell
+sudo apt install cmake ninja-build clang lld
+```
 
 === "Windows"
 
-    1. Install CMake from the
+    1. Install MSVC from Visual Studio or "Tools for Visual Studio" on the
+       [official downloads page](https://visualstudio.microsoft.com/downloads/)
+
+    2. Install CMake from the
        [official downloads page](https://cmake.org/download/)
 
-    2. Install MSVC from Visual Studio or "Tools for Visual Studio" on the
-       [official downloads page](https://visualstudio.microsoft.com/downloads/)
+    3. Install Ninja either from the
+       [official site](https://ninja-build.org/).
 
     !!! note
         You will need to initialize MSVC by running `vcvarsall.bat` to use it
@@ -44,8 +52,38 @@ git submodule update --init
 
 Configure then build all targets using CMake:
 
+Configure CMake:
+
+=== "Linux and MacOS"
+
+    ``` shell
+    # Recommended for simple development using clang and lld:
+    cmake -GNinja -B ../iree-build/ -S . \
+        -DCMAKE_BUILD_TYPE=RelWithDebInfo \
+        -DIREE_ENABLE_ASSERTIONS=ON \
+        -DCMAKE_C_COMPILER=clang \
+        -DCMAKE_CXX_COMPILER=clang++ \
+        -DIREE_ENABLE_LLD=ON
+
+    # Alternately, with system compiler and your choice of CMake generator:
+    # cmake -B ../iree-build/ -S .
+
+    # Additional quality of life CMake flags:
+    # Enable ccache:
+    #   -DCMAKE_C_COMPILER_LAUNCHER=ccache -DCMAKE_CXX_COMPILER_LAUNCHER=ccache
+    ```
+
+=== "Windows"
+
+    ``` shell
+    cmake -GNinja -B ../iree-build/ -S . \
+        -DCMAKE_BUILD_TYPE=RelWithDebInfo \
+        -DIREE_ENABLE_ASSERTIONS=ON
+    ```
+
+Build:
+
 ``` shell
-cmake -B ../iree-build/ -DCMAKE_BUILD_TYPE=RelWithDebInfo .
 cmake --build ../iree-build/
 ```
 
@@ -59,12 +97,8 @@ cmake --build ../iree-build/
     for general details.
 
 ???+ Tip
-    Most IREE Core devs use [Ninja](https://ninja-build.org/) as the CMake
-    generator. The benefit is that it works the same across all platforms and
-    automatically takes advantage of parallelism. to use it, add a `-GNinja`
-    argument to your initial cmake command (and make sure to install
-    `ninja-build` from either your favorite OS package manager, or generically
-    via `python -m pip install ninja`).
+    You are welcome to try different CMake generators, but IREE devs and CIs
+    exclusively use [Ninja](https://ninja-build.org/).
 
 
 ## What's next?

docs/website/docs/building-from-source/optional-features.md

@@ -87,6 +87,7 @@ Also see [instructions for installing pre-built binaries](../bindings/python.md)
 
         ``` shell
         cmake \
+          -GNinja \
           -DCMAKE_BUILD_TYPE=RelWithDebInfo \
           -DIREE_BUILD_PYTHON_BINDINGS=ON \
           -DPython3_EXECUTABLE="$(which python)" \
@@ -102,7 +103,7 @@ Also see [instructions for installing pre-built binaries](../bindings/python.md)
     === "Windows"
 
         ``` powershell
-        cmake -DCMAKE_BUILD_TYPE=RelWithDebInfo -DIREE_BUILD_PYTHON_BINDINGS=ON .
+        cmake -GNinja -DCMAKE_BUILD_TYPE=RelWithDebInfo -DIREE_BUILD_PYTHON_BINDINGS=ON .
         cmake --build .
 
         # Add bindings\python to PYTHONPATH and use the API.

docs/website/docs/building-from-source/riscv.md

@@ -55,7 +55,7 @@ For RISC-V vector extensions support, see
 Build and install on your host machine:
 
 ``` shell
-cmake -B ../iree-build/ \
+cmake -GNinja -B ../iree-build/ \
   -DCMAKE_INSTALL_PREFIX=../iree-build/install \
   -DCMAKE_BUILD_TYPE=RelWithDebInfo \
   .
@@ -72,7 +72,7 @@ as a reference of how to set up the cmake configuration.
 #### RISC-V 64-bit Linux target
 
 ```shell
-cmake -B ../iree-build-riscv/ \
+cmake -GNinja -B ../iree-build-riscv/ \
   -DCMAKE_TOOLCHAIN_FILE="./build_tools/cmake/riscv.toolchain.cmake" \
   -DIREE_HOST_BINARY_ROOT=$(realpath ../iree-build-host/install) \
   -DRISCV_CPU=rv64 \

iree/compiler/Codegen/Common/BUILD

@@ -40,7 +40,6 @@ cc_library(
         "LinalgBufferizePass.cpp",
         "OptimizeVectorTransferPass.cpp",
         "SetNumWorkgroupsPass.cpp",
-        "ShapeToLLVMConversion.cpp",
         "VectorizeConv.cpp",
         "VectorizeMMT4d.cpp",
     ],

iree/compiler/Codegen/Common/CMakeLists.txt

@@ -31,7 +31,6 @@ iree_cc_library(
     "LinalgBufferizePass.cpp"
     "OptimizeVectorTransferPass.cpp"
     "SetNumWorkgroupsPass.cpp"
-    "ShapeToLLVMConversion.cpp"
     "VectorizeConv.cpp"
     "VectorizeMMT4d.cpp"
   DEPS

iree/compiler/Codegen/Common/CleanupBufferAllocViewPass.cpp

@@ -18,6 +18,7 @@
 #include "iree/compiler/Dialect/HAL/IR/HALOps.h"
 #include "mlir/Dialect/Linalg/IR/LinalgOps.h"
 #include "mlir/IR/BuiltinAttributes.h"
+#include "mlir/IR/BuiltinTypes.h"
 #include "mlir/IR/Matchers.h"
 #include "mlir/IR/PatternMatch.h"
 #include "mlir/Transforms/GreedyPatternRewriteDriver.h"
@@ -51,6 +52,13 @@ struct FoldReshapeIntoInterfaceTensorLoad : OpRewritePattern<TensorReshapeOp> {
 
   LogicalResult matchAndRewrite(TensorReshapeOp reshapeOp,
                                 PatternRewriter &rewriter) const override {
+    // TODO(antigainst): enable dynamic shape support once they are needed.
+    auto reshapeSrcType = reshapeOp.src().getType().template cast<ShapedType>();
+    auto reshapeDstType = reshapeOp.getType().template cast<ShapedType>();
+    if (!reshapeSrcType.hasStaticShape() || !reshapeDstType.hasStaticShape()) {
+      return failure();
+    }
+
     auto loadOp =
         reshapeOp.src()
             .template getDefiningOp<IREE::Flow::DispatchTensorLoadOp>();
@@ -66,16 +74,18 @@ struct FoldReshapeIntoInterfaceTensorLoad : OpRewritePattern<TensorReshapeOp> {
         loadOp.source()
             .template getDefiningOp<IREE::HAL::InterfaceBindingSubspanOp>();
     if (!subspanOp) return failure();
+    assert(subspanOp.dynamic_dims().empty());
 
+    auto tensorAccess = subspanOp.getType()
+                            .template cast<IREE::Flow::DispatchTensorType>()
+                            .getAccess();
     auto newSubspanType = IREE::Flow::DispatchTensorType::get(
-        subspanOp.getType()
-            .template cast<IREE::Flow::DispatchTensorType>()
-            .getAccess(),
-        reshapeOp.getResultType());
+        tensorAccess, reshapeOp.getResultType());
 
     Value newSubspanOp = rewriter.create<IREE::HAL::InterfaceBindingSubspanOp>(
         subspanOp.getLoc(), newSubspanType, subspanOp.binding(),
-        subspanOp.byte_offset(), subspanOp.byte_length());
+        subspanOp.byte_offset(), subspanOp.byte_length(),
+        subspanOp.dynamic_dims());
 
     rewriter.replaceOpWithNewOp<IREE::Flow::DispatchTensorLoadOp>(
         reshapeOp, reshapeOp.getResultType(), newSubspanOp);