[Backend Tester] Add ARM TOSA flow

.ci/scripts/test_backend_linux.sh

@@ -41,12 +41,17 @@ if [[ "$FLOW" == *qnn* ]]; then
 fi
 
 if [[ "$FLOW" == *vulkan* ]]; then
-    # Setup swiftshader and Vulkan SDK which are required to build the Vulkan delegate
+    # Setup swiftshader and Vulkan SDK which are required to build the Vulkan delegate.
     source .ci/scripts/setup-vulkan-linux-deps.sh
 
     EXTRA_BUILD_ARGS+=" -DEXECUTORCH_BUILD_VULKAN=ON"
 fi
 
+if [[ "$FLOW" == *arm* ]]; then
+    # Setup ARM deps.
+    .ci/scripts/setup-arm-baremetal-tools.sh
+fi
+
 # We need the runner to test the built library.
 PYTHON_EXECUTABLE=python CMAKE_ARGS="$EXTRA_BUILD_ARGS" .ci/scripts/setup-linux.sh --build-tool cmake --build-mode Release --editable true
 

.github/workflows/_test_backend.yml

@@ -0,0 +1,79 @@
+name: Test Backend
+
+on:
+  workflow_call:
+    inputs:
+      backend:
+        description: 'Backend to test (xnnpack, coreml, vulkan, qnn)'
+        required: true
+        type: string
+      flows:
+        description: 'JSON array of flows to test'
+        required: true
+        type: string
+      ref:
+        description: 'Git ref to checkout'
+        required: false
+        type: string
+        default: ${{ github.sha }}
+      timeout:
+        description: 'Job timeout in minutes'
+        required: false
+        type: number
+        default: 120
+      run-linux:
+        description: 'Whether to run Linux tests'
+        required: false
+        type: boolean
+        default: false
+      run-macos:
+        description: 'Whether to run macOS tests'
+        required: false
+        type: boolean
+        default: false
+
+jobs:
+  test-backend-linux:
+    if: ${{ inputs.run-linux }}
+    strategy:
+      fail-fast: false
+      matrix:
+        flow: ${{ fromJSON(inputs.flows) }}
+        suite: [models, operators]
+
+    uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
+    with:
+      ref: ${{ inputs.ref }}
+      runner: linux.4xlarge.memory
+      docker-image: ci-image:executorch-ubuntu-22.04-clang12
+      submodules: recursive
+      timeout: ${{ inputs.timeout }}
+      upload-artifact: test-report-${{ matrix.flow }}-${{ matrix.suite }}
+      script: |
+        set -eux
+
+        source .ci/scripts/test_backend_linux.sh "${{ matrix.suite }}" "${{ matrix.flow }}" "${RUNNER_ARTIFACT_DIR}"
+
+  test-backend-macos:
+    if: ${{ inputs.run-macos }}
+    strategy:
+      fail-fast: false
+      matrix:
+        flow: ${{ fromJSON(inputs.flows) }}
+        suite: [models, operators]
+
+    uses: pytorch/test-infra/.github/workflows/macos_job.yml@main
+    with:
+      ref: ${{ inputs.ref }}
+      runner: macos-m1-stable
+      python-version: "3.12"
+      submodules: recursive
+      timeout: ${{ inputs.timeout }}
+      upload-artifact: test-report-${{ matrix.flow }}-${{ matrix.suite }}
+      script: |
+        set -eux
+
+        # This is needed to get the prebuilt PyTorch wheel from S3
+        ${CONDA_RUN} --no-capture-output pip install awscli==1.37.21
+
+        source .ci/scripts/test_backend_macos.sh "${{ matrix.suite }}" "${{ matrix.flow }}" "${RUNNER_ARTIFACT_DIR}"

.github/workflows/nightly.yml

@@ -36,51 +36,3 @@ jobs:
     uses: ./.github/workflows/_link_check.yml
     with:
       ref: ${{ github.sha }}
-
-  backend-test-linux:
-    uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
-    strategy:
-      fail-fast: false
-      matrix:
-        flow: [
-          qnn, qnn_16a16w, qnn_16a8w, qnn_16a4w, qnn_16a4w_block, qnn_8a8w,
-          vulkan, vulkan_static_int8_per_channel,
-          xnnpack, xnnpack_dynamic_int8_per_channel, xnnpack_static_int8_per_channel, xnnpack_static_int8_per_tensor
-        ]
-        suite: [models, operators]
-    with:
-      ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
-      runner: linux.4xlarge.memory
-      docker-image: ci-image:executorch-ubuntu-22.04-clang12
-      submodules: recursive
-      timeout: 120
-      upload-artifact: test-report-${{ matrix.flow }}-${{ matrix.suite }}
-      script: |
-        set -eux
-
-        source .ci/scripts/test_backend_linux.sh "${{ matrix.suite }}" "${{ matrix.flow }}" "${RUNNER_ARTIFACT_DIR}"
-
-  backend-test-macos:
-    uses: pytorch/test-infra/.github/workflows/macos_job.yml@main
-    permissions:
-      id-token: write
-      contents: read
-    strategy:
-      fail-fast: false
-      matrix:
-        flow: [coreml, coreml_static_int8]
-        suite: [models, operators]
-    with:
-      ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
-      runner: macos-m1-stable
-      python-version: 3.12
-      submodules: recursive
-      timeout: 120
-      upload-artifact: test-report-${{ matrix.flow }}-${{ matrix.suite }}
-      script: |
-        set -eux
-
-        # This is needed to get the prebuilt PyTorch wheel from S3
-        ${CONDA_RUN} --no-capture-output pip install awscli==1.37.21
-
-        source .ci/scripts/test_backend_macos.sh "${{ matrix.suite }}" "${{ matrix.flow }}" "${RUNNER_ARTIFACT_DIR}"

.github/workflows/test-backend-arm.yml

@@ -0,0 +1,27 @@
+name: Test ARM Backend
+
+on:
+  schedule:
+    - cron: 0 2 * * *
+  push:
+    tags:
+      - ciflow/nightly/*
+  pull_request:
+    paths:
+      - .github/workflows/test-backend-arm.yml
+      - .github/workflows/_test_backend.yml
+  workflow_dispatch:
+
+concurrency:
+  group: ${{ github.workflow }}--${{ github.event.pull_request.number || github.sha }}-${{ github.event_name == 'workflow_dispatch' }}
+  cancel-in-progress: true
+
+jobs:
+  test-arm:
+    uses: ./.github/workflows/_test_backend.yml
+    with:
+      backend: arm
+      flows: '["arm_tosa"]'
+      ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
+      timeout: 120
+      run-linux: true

.github/workflows/test-backend-coreml.yml

@@ -0,0 +1,27 @@
+name: Test CoreML Backend
+
+on:
+  schedule:
+    - cron: 0 2 * * *
+  push:
+    tags:
+      - ciflow/nightly/*
+  pull_request:
+    paths:
+      - .github/workflows/test-backend-coreml.yml
+      - .github/workflows/_test_backend.yml
+  workflow_dispatch:
+
+concurrency:
+  group: ${{ github.workflow }}--${{ github.event.pull_request.number || github.sha }}-${{ github.event_name == 'workflow_dispatch' }}
+  cancel-in-progress: true
+
+jobs:
+  test-coreml:
+    uses: ./.github/workflows/_test_backend.yml
+    with:
+      backend: coreml
+      flows: '["coreml", "coreml_static_int8"]'
+      ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
+      timeout: 120
+      run-macos: true

.github/workflows/test-backend-qnn.yml

@@ -0,0 +1,27 @@
+name: Test QNN Backend
+
+on:
+  schedule:
+    - cron: 0 2 * * *
+  push:
+    tags:
+      - ciflow/nightly/*
+  pull_request:
+    paths:
+      - .github/workflows/test-backend-qnn.yml
+      - .github/workflows/_test_backend.yml
+  workflow_dispatch:
+
+concurrency:
+  group: ${{ github.workflow }}--${{ github.event.pull_request.number || github.sha }}-${{ github.event_name == 'workflow_dispatch' }}
+  cancel-in-progress: true
+
+jobs:
+  test-qnn:
+    uses: ./.github/workflows/_test_backend.yml
+    with:
+      backend: qnn
+      flows: '["qnn", "qnn_16a16w", "qnn_16a8w", "qnn_16a4w", "qnn_16a4w_block", "qnn_8a8w"]'
+      ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
+      timeout: 120
+      run-linux: true

.github/workflows/test-backend-vulkan.yml

@@ -0,0 +1,27 @@
+name: Test Vulkan Backend
+
+on:
+  schedule:
+    - cron: 0 2 * * *
+  push:
+    tags:
+      - ciflow/nightly/*
+  pull_request:
+    paths:
+      - .github/workflows/test-backend-vulkan.yml
+      - .github/workflows/_test_backend.yml
+  workflow_dispatch:
+
+concurrency:
+  group: ${{ github.workflow }}--${{ github.event.pull_request.number || github.sha }}-${{ github.event_name == 'workflow_dispatch' }}
+  cancel-in-progress: true
+
+jobs:
+  test-vulkan:
+    uses: ./.github/workflows/_test_backend.yml
+    with:
+      backend: vulkan
+      flows: '["vulkan", "vulkan_static_int8_per_channel"]'
+      ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
+      timeout: 120
+      run-linux: true

.github/workflows/test-backend-xnnpack.yml

@@ -0,0 +1,27 @@
+name: Test XNNPACK Backend
+
+on:
+  schedule:
+    - cron: 0 2 * * *
+  push:
+    tags:
+      - ciflow/nightly/*
+  pull_request:
+    paths:
+      - .github/workflows/test-backend-xnnpack.yml
+      - .github/workflows/_test_backend.yml
+  workflow_dispatch:
+
+concurrency:
+  group: ${{ github.workflow }}--${{ github.event.pull_request.number || github.sha }}-${{ github.event_name == 'workflow_dispatch' }}
+  cancel-in-progress: true
+
+jobs:
+  test-xnnpack:
+    uses: ./.github/workflows/_test_backend.yml
+    with:
+      backend: xnnpack
+      flows: '["xnnpack", "xnnpack_dynamic_int8_per_channel", "xnnpack_static_int8_per_channel", "xnnpack_static_int8_per_tensor"]'
+      ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
+      timeout: 120
+      run-linux: true

backends/arm/test/tester/arm_tester.py

@@ -65,6 +65,7 @@
 
 from executorch.backends.arm.vgf_partitioner import VgfPartitioner
 
+from executorch.backends.test.harness.error_statistics import ErrorStatistics
 from executorch.backends.test.harness.stages import Stage, StageType
 from executorch.backends.xnnpack.test.tester import Tester
 from executorch.devtools.backend_debug import get_delegation_info
@@ -379,6 +380,7 @@ def to_edge_transform_and_lower(
         transform_passes: Optional[
             Union[Sequence[PassType], Dict[str, Sequence[PassType]]]
         ] = None,
+        generate_etrecord: bool = False,
     ):
         if to_edge_and_lower_stage is None:
             if partitioners is None:
@@ -412,7 +414,9 @@ def to_edge_transform_and_lower(
                 to_edge_and_lower_stage.partitioners = partitioners
             if edge_compile_config is not None:
                 to_edge_and_lower_stage.edge_compile_conf = edge_compile_config
-        return super().to_edge_transform_and_lower(to_edge_and_lower_stage)
+        return super().to_edge_transform_and_lower(
+            to_edge_and_lower_stage, generate_etrecord=generate_etrecord
+        )
 
     def to_executorch(self, to_executorch_stage: Optional[ToExecutorch] | None = None):
         if to_executorch_stage is None:
@@ -443,6 +447,7 @@ def run_method_and_compare_outputs(
         qtol=0,
         error_callbacks=None,
         run_eager_mode=False,
+        statistics_callback: Callable[[ErrorStatistics], None] | None = None,
     ):
         """
         Compares the run_artifact output of 'stage' with the output of a reference stage.
@@ -698,10 +703,17 @@ def _compare_outputs(
         rtol=1e-03,
         qtol=0,
         error_callbacks=None,
+        statistics_callback: Callable[[ErrorStatistics], None] | None = None,
     ):
         try:
             super()._compare_outputs(
-                reference_output, stage_output, quantization_scale, atol, rtol, qtol
+                reference_output,
+                stage_output,
+                quantization_scale,
+                atol,
+                rtol,
+                qtol,
+                statistics_callback=statistics_callback,
             )
         except AssertionError as e:
             if error_callbacks is None:

backends/test/suite/flow.py

@@ -1,6 +1,6 @@
 import logging
 
-from dataclasses import dataclass
+from dataclasses import dataclass, field
 from typing import Callable
 
 from executorch.backends.test.harness import Tester
@@ -35,6 +35,15 @@ class TestFlow:
     is_delegated: bool = True
     """ Indicates whether the flow is expected to generate CALL_DELEGATE nodes. """
 
+    skip_patterns: list[str] = field(default_factory=lambda: [])
+    """ Tests with names containing any substrings in this list are skipped. """
+
+    supports_serialize: bool = True
+    """ True if the test flow supports the Serialize stage. """
+
+    def should_skip_test(self, test_name: str) -> bool:
+        return any(pattern in test_name for pattern in self.skip_patterns)
+
 
 def all_flows() -> dict[str, TestFlow]:
     flows = []
@@ -109,4 +118,13 @@ def all_flows() -> dict[str, TestFlow]:
     except Exception as e:
         logger.info(f"Skipping QNN flow registration: {e}")
 
+    try:
+        from executorch.backends.test.suite.flows.arm import ARM_TOSA_FLOW
+
+        flows += [
+            ARM_TOSA_FLOW,
+        ]
+    except Exception as e:
+        logger.info(f"Skipping ARM flow registration: {e}")
+
     return {f.name: f for f in flows if f is not None}