Update on "[ET-VK][Ops] dequantize_per_channel reference impl and testing"

# Context

In order to properly enable dynamic quantization, we create the dequantize_per_channel operator as its seemingly useful to have for the pipeline.

# Changes

This creates the wrapper for the cpu reference implementation, and also a dummy reference implementation I created just to test against it.

Differential Revision: [D77746138](https://our.internmc.facebook.com/intern/diff/D77746138/)

[ghstack-poisoned]

Author: morelos

.ci/scripts/test_yolo12.sh

@@ -0,0 +1,197 @@
+#!/bin/bash
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+set -ex
+# shellcheck source=/dev/null
+source "$(dirname "${BASH_SOURCE[0]}")/utils.sh"
+
+while [[ $# -gt 0 ]]; do
+  case "$1" in
+    -model)
+      MODEL_NAME="$2" # stories110M
+      shift 2
+      ;;
+    -mode)
+      MODE="$2" # portable or xnnpack+custom or xnnpack+custom+qe
+      shift 2
+      ;;
+    -pt2e_quantize)
+      PT2E_QUANTIZE="$2"
+      shift 2
+      ;;
+    -upload)
+      UPLOAD_DIR="$2"
+      shift 2
+      ;;
+    -video_path)
+      VIDEO_PATH="$2" # portable or xnnpack+custom or xnnpack+custom+qe
+      shift 2
+      ;;
+    *)
+      echo "Unknown option: $1"
+      usage
+      ;;
+  esac
+done
+
+# Default mode to xnnpack+custom if not set
+MODE=${MODE:-"openvino"}
+
+# Default UPLOAD_DIR to empty string if not set
+UPLOAD_DIR="${UPLOAD_DIR:-}"
+
+# Default PT2E_QUANTIZE to empty string if not set
+PT2E_QUANTIZE="${PT2E_QUANTIZE:-}"
+
+# Default CMake Build Type to release mode
+CMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE:-Release}
+
+if [[ $# -lt 5 ]]; then # Assuming 4 mandatory args
+    echo "Expecting atleast 5 positional arguments"
+    echo "Usage: [...]"
+fi
+if [[ -z "${MODEL_NAME:-}" ]]; then
+  echo "Missing model name, exiting..."
+  exit 1
+fi
+
+
+if [[ -z "${MODE:-}" ]]; then
+  echo "Missing mode, choose openvino or xnnpack, exiting..."
+  exit 1
+fi
+
+if [[ -z "${PYTHON_EXECUTABLE:-}" ]]; then
+  PYTHON_EXECUTABLE=python3
+fi
+
+TARGET_LIBS=""
+
+if [[ "${MODE}" =~ .*openvino.* ]]; then
+  OPENVINO=ON
+  TARGET_LIBS="$TARGET_LIBS openvino_backend "
+
+  git clone https://github.com/openvinotoolkit/openvino.git
+  cd openvino && git b16b776ac119dafda51f69a80f1e6b7376d02c3b
+  git submodule update --init --recursive
+  sudo ./install_build_dependencies.sh
+  mkdir build && cd build
+  cmake .. -DCMAKE_BUILD_TYPE=Release -DENABLE_PYTHON=ON
+  make -j$(nproc)
+
+  cd ..
+  cmake --install build --prefix dist
+
+  source dist/setupvars.sh
+  cd ../backends/openvino
+  pip install -r requirements.txt
+  cd ../../
+else
+  OPENVINO=OFF
+fi
+
+if [[ "${MODE}" =~ .*xnnpack.* ]]; then
+  XNNPACK=ON
+  TARGET_LIBS="$TARGET_LIBS xnnpack_backend "
+else
+  XNNPACK=OFF
+fi
+
+which "${PYTHON_EXECUTABLE}"
+
+
+DIR="examples/models/yolo12"
+$PYTHON_EXECUTABLE -m pip install -r ${DIR}/requirements.txt
+
+cmake_install_executorch_libraries() {
+    rm -rf cmake-out
+    build_dir=cmake-out
+    mkdir $build_dir
+
+
+    retry cmake -DCMAKE_INSTALL_PREFIX="${build_dir}" \
+          -DCMAKE_BUILD_TYPE="${CMAKE_BUILD_TYPE}" \
+          -DEXECUTORCH_BUILD_OPENVINO="$OPENVINO" \
+          -DEXECUTORCH_BUILD_XNNPACK="$XNNPACK" \
+          -DEXECUTORCH_BUILD_EXTENSION_DATA_LOADER=ON \
+          -DEXECUTORCH_BUILD_EXTENSION_MODULE=ON \
+          -DEXECUTORCH_BUILD_EXTENSION_RUNNER_UTIL=ON \
+          -DEXECUTORCH_BUILD_EXTENSION_TENSOR=ON \
+          -B"${build_dir}"
+
+    # Build the project
+    cmake --build ${build_dir} --target install --config ${CMAKE_BUILD_TYPE} -j$(nproc)
+
+    export CMAKE_ARGS="
+                       -DEXECUTORCH_BUILD_OPENVINO="$OPENVINO" \
+                       -DEXECUTORCH_BUILD_XNNPACK="$XNNPACK" \
+                       -DEXECUTORCH_BUILD_EXTENSION_DATA_LOADER=ON \
+                       -DEXECUTORCH_BUILD_EXTENSION_MODULE=ON \
+                       -DEXECUTORCH_BUILD_EXTENSION_RUNNER_UTIL=ON \
+                       -DEXECUTORCH_ENABLE_LOGGING=ON \
+                       -DEXECUTORCH_BUILD_EXTENSION_TENSOR=ON \
+                       -DEXECUTORCH_BUILD_PYBIND=ON"
+
+    echo $TARGET_LIBS
+    export CMAKE_BUILD_ARGS="--target $TARGET_LIBS"
+    pip install . --no-build-isolation
+}
+
+cmake_build_demo() {
+    echo "Building yolo12 runner"
+    retry cmake \
+        -DCMAKE_BUILD_TYPE="$CMAKE_BUILD_TYPE" \
+        -DUSE_OPENVINO_BACKEND="$OPENVINO" \
+        -DUSE_XNNPACK_BACKEND="$XNNPACK" \
+        -Bcmake-out/${DIR} \
+        ${DIR}
+    cmake --build cmake-out/${DIR} -j9 --config "$CMAKE_BUILD_TYPE"
+
+}
+
+cleanup_files() {
+ rm $EXPORTED_MODEL_NAME
+}
+
+prepare_artifacts_upload() {
+  if [ -n "${UPLOAD_DIR}" ]; then
+    echo "Preparing for uploading generated artifacs"
+    zip -j model.zip "${EXPORTED_MODEL_NAME}"
+    mkdir -p "${UPLOAD_DIR}"
+    mv model.zip "${UPLOAD_DIR}"
+    mv result.txt "${UPLOAD_DIR}"
+
+  fi
+}
+
+
+# Export model.
+EXPORTED_MODEL_NAME="${MODEL_NAME}_fp32_${MODE}.pte"
+echo "Exporting ${EXPORTED_MODEL_NAME}"
+EXPORT_ARGS="--model_name=${MODEL_NAME} --backend=${MODE}"
+
+# Add dynamically linked library location
+cmake_install_executorch_libraries
+
+$PYTHON_EXECUTABLE -m examples.models.yolo12.export_and_validate ${EXPORT_ARGS}
+
+
+RUNTIME_ARGS="--model_path=${EXPORTED_MODEL_NAME} --input_path=${VIDEO_PATH}"
+# Check build tool.
+cmake_build_demo
+# Run yolo12 runner
+NOW=$(date +"%H:%M:%S")
+echo "Starting to run yolo12 runner at ${NOW}"
+# shellcheck source=/dev/null
+cmake-out/examples/models/yolo12/Yolo12DetectionDemo ${RUNTIME_ARGS} > result.txt
+NOW=$(date +"%H:%M:%S")
+echo "Finished at ${NOW}"
+
+RESULT=$(cat result.txt)
+
+prepare_artifacts_upload
+cleanup_files

.github/workflows/lint.yml

@@ -83,8 +83,13 @@ jobs:
       script: |
         FILES_NEEDS_FORMAT=$(/opt/google-java-format -n \
           extension/android/executorch_android/src/main/java/org/pytorch/executorch/*.java \
+          extension/android/executorch_android/src/main/java/org/pytorch/executorch/extension/llm/*.java \
+          extension/android/executorch_android/src/main/java/org/pytorch/executorch/annotations/*.java \
+          extension/android/executorch_android/src/androidTest/java/org/pytorch/executorch/*.java \
           examples/demo-apps/android/LlamaDemo/app/src/main/java/com/example/executorchllamademo/*.java \
-          extension/benchmark/android/benchmark/app/src/main/java/org/pytorch/minibench/*.java)
+          examples/demo-apps/android/LlamaDemo/app/src/androidTest/java/com/example/executorchllamademo/*.java \
+          extension/benchmark/android/benchmark/app/src/main/java/org/pytorch/minibench/*.java \
+          extension/benchmark/android/benchmark/app/src/androidTest/java/org/pytorch/minibench/*.java)
         if [ -n "$FILES_NEEDS_FORMAT" ]; then
           echo "Warning: The following files need formatting. Please use google-java-format."
           echo "Use a binary from https://github.com/google/google-java-format/releases/"

backends/apple/coreml/test/tester.py

@@ -0,0 +1,61 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+from typing import Any, List, Optional, Tuple
+
+import executorch
+import executorch.backends.test.harness.stages as BaseStages
+
+import torch
+from executorch.backends.apple.coreml.partition import CoreMLPartitioner
+from executorch.backends.test.harness import Tester as TesterBase
+from executorch.backends.test.harness.stages import StageType
+from executorch.exir import EdgeCompileConfig
+from executorch.exir.backend.partitioner import Partitioner
+
+
+class Partition(BaseStages.Partition):
+    def __init__(self, partitioner: Optional[Partitioner] = None):
+        super().__init__(
+            partitioner=partitioner or CoreMLPartitioner,
+        )
+
+
+class ToEdgeTransformAndLower(BaseStages.ToEdgeTransformAndLower):
+    def __init__(
+        self,
+        partitioners: Optional[List[Partitioner]] = None,
+        edge_compile_config: Optional[EdgeCompileConfig] = None,
+    ):
+        super().__init__(
+            default_partitioner_cls=CoreMLPartitioner,
+            partitioners=partitioners,
+            edge_compile_config=edge_compile_config,
+        )
+
+
+class CoreMLTester(TesterBase):
+    def __init__(
+        self,
+        module: torch.nn.Module,
+        example_inputs: Tuple[torch.Tensor],
+        dynamic_shapes: Optional[Tuple[Any]] = None,
+    ):
+        # Specialize for XNNPACK
+        stage_classes = (
+            executorch.backends.test.harness.Tester.default_stage_classes()
+            | {
+                StageType.PARTITION: Partition,
+                StageType.TO_EDGE_TRANSFORM_AND_LOWER: ToEdgeTransformAndLower,
+            }
+        )
+
+        super().__init__(
+            module=module,
+            stage_classes=stage_classes,
+            example_inputs=example_inputs,
+            dynamic_shapes=dynamic_shapes,
+        )

backends/arm/_passes/TARGETS

@@ -6,6 +6,7 @@ python_library(
     deps = [
         "//executorch/backends/arm:tosa_quant_utils",
         "//executorch/backends/arm:tosa_utils",
+        "//executorch/backends/arm/tosa/dialect:lib",
         "//executorch/backends/transforms:fuse_view_copy",
         "//executorch/backends/transforms:remove_getitem_op",
         "//executorch/backends/transforms:replace_scalar_with_tensor",

backends/arm/_passes/__init__.py

@@ -24,6 +24,7 @@
 from .convert_to_clamp import ConvertToClampPass  # noqa
 from .decompose_acosh_pass import DecomposeAcoshPass  # noqa
 from .decompose_adaptive_avg_pool2d_pass import DecomposeAdaptiveAvgPool2dPass  # noqa
+from .decompose_asin_pass import DecomposeAsinPass  # noqa
 from .decompose_atan_pass import DecomposeAtanPass  # noqa
 from .decompose_avg_pool2d import DecomposeAvgPool2d  # noqa
 from .decompose_batch_norm_no_stats import DecomposeBatchNormNoStatsPass  # noqa
@@ -50,6 +51,7 @@
 from .decompose_sqrt_pass import DecomposeSqrtPass  # noqa
 from .decompose_sum_pass import DecomposeSumPass  # noqa
 from .decompose_var_pass import DecomposeVarPass  # noqa
+from .decorate_fp32_to_int32_casting_pass import DecorateFp32toInt32CastingPass  # noqa
 from .fold_qdq_with_annotated_qparams_pass import (  # noqa
     FoldAndAnnotateQParamsPass,
     QuantizeOperatorArguments,

backends/arm/_passes/arm_pass_manager.py

@@ -29,6 +29,7 @@
     ConvertToClampPass,
     DecomposeAcoshPass,
     DecomposeAdaptiveAvgPool2dPass,
+    DecomposeAsinPass,
     DecomposeAtanPass,
     DecomposeAvgPool2d,
     DecomposeBatchNormNoStatsPass,
@@ -55,6 +56,7 @@
     DecomposeSqrtPass,
     DecomposeSumPass,
     DecomposeVarPass,
+    DecorateFp32toInt32CastingPass,
     FoldAndAnnotateQParamsPass,
     FuseBatchnorm2DPass,
     FuseConstantArgsPass,
@@ -158,6 +160,7 @@ def _tosa_080_BI_pipeline(self, exported_program: ExportedProgram) -> GraphModul
     def _tosa_080_MI_pipeline(self, exported_program: ExportedProgram) -> GraphModule:
         self.add_pass(DecomposeRoundPass())
         self.add_pass(DecomposeAcoshPass())
+        self.add_pass(DecomposeAsinPass())
         self.add_pass(DecomposeSqrtPass())
         self.add_pass(DecomposeAtanPass())
         self.add_pass(ConvertIntPowToMuls())
@@ -198,6 +201,9 @@ def _tosa_080_MI_pipeline(self, exported_program: ExportedProgram) -> GraphModul
         self.add_pass(MatchArgRanksPass(exported_program))
         self.add_pass(DecomposeAdaptiveAvgPool2dPass())
         self.add_pass(DecomposeAvgPool2d())
+        self.add_pass(
+            DecorateFp32toInt32CastingPass()
+        )  # Require that no new fp32->int32 is introduced after this pass
         self.add_pass(ComputeConstantOpsAOT(exported_program))
 
         self.add_pass(DecomposeGroupedConv())