[cadence] Add PR-triggered hifi op-test stage on the Xtensa ISS

.ci/scripts/test-cadence-xtensa.sh

@@ -0,0 +1,84 @@
+#!/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.
+#
+# Build the Cadence Xtensa op-level gtest tests for the configured backend and
+# run them on the Instruction Set Simulator (xt-run).
+#
+# Requires the Xtensa toolchain env to already be set (run
+# .ci/scripts/setup-xtensa-tools.sh <backend> first): XTENSA_TOOLCHAIN,
+# TOOLCHAIN_VER, XTENSA_CORE, CADENCE_OPT_FLAG, and xt-clang/xt-run on PATH.
+#
+# Unlike build-cadence-xtensa.sh (the runner, built -fno-exceptions -fno-rtti),
+# the gtest tests need exceptions + RTTI, so those flags are NOT set here.
+
+set -euo pipefail
+
+: "${XTENSA_TOOLCHAIN:?run setup-xtensa-tools.sh first}"
+: "${TOOLCHAIN_VER:?run setup-xtensa-tools.sh first}"
+: "${XTENSA_CORE:?run setup-xtensa-tools.sh first}"
+: "${CADENCE_OPT_FLAG:?run setup-xtensa-tools.sh first}"
+
+# Map the optimized-kernel flag to the backend dir + gtest target name.
+case "${CADENCE_OPT_FLAG}" in
+  EXECUTORCH_NNLIB_OPT) TARGET_DIR=hifi ;;
+  EXECUTORCH_VISION_OPT) TARGET_DIR=vision ;;
+  EXECUTORCH_FUSION_G3_OPT) TARGET_DIR=fusion_g3 ;;
+  *)
+    echo "ERROR: unknown CADENCE_OPT_FLAG='${CADENCE_OPT_FLAG}'" >&2
+    exit 1
+    ;;
+esac
+TEST_TARGET="cadence_${TARGET_DIR}_op_tests"
+TEST_ELF="cmake-out/backends/cadence/${TARGET_DIR}/operators/tests/${TEST_TARGET}"
+
+NPROC=$(nproc)
+echo "=== building ${TEST_TARGET} for ${XTENSA_CORE} (${CADENCE_OPT_FLAG}) ==="
+xt-clang --version | head -1
+
+rm -rf cmake-out
+cmake \
+  -DCMAKE_TOOLCHAIN_FILE=./backends/cadence/cadence.cmake \
+  -DCMAKE_INSTALL_PREFIX=cmake-out \
+  -DCMAKE_BUILD_TYPE=Release \
+  -DEXECUTORCH_BUILD_CADENCE=ON \
+  "-D${CADENCE_OPT_FLAG}=ON" \
+  -DEXECUTORCH_BUILD_PORTABLE_OPS=ON \
+  -DEXECUTORCH_BUILD_CADENCE_OP_TESTS=ON \
+  -DEXECUTORCH_BUILD_EXECUTOR_RUNNER=OFF \
+  -DEXECUTORCH_BUILD_EXTENSION_RUNNER_UTIL=ON \
+  -DEXECUTORCH_ENABLE_LOGGING=ON \
+  -DEXECUTORCH_BUILD_PTHREADPOOL=OFF \
+  -DEXECUTORCH_BUILD_CPUINFO=OFF \
+  -DEXECUTORCH_USE_DL=OFF \
+  -DEXECUTORCH_BUILD_KERNELS_LLM=OFF \
+  -DEXECUTORCH_BUILD_DEVTOOLS=OFF \
+  -DHAVE_FNMATCH_H=OFF \
+  -DFLATCC_ALLOW_WERROR=OFF \
+  -DPYTHON_EXECUTABLE="$(which python3)" \
+  -Bcmake-out .
+
+cmake --build cmake-out --target "${TEST_TARGET}" -j"${NPROC}"
+
+if [[ ! -f "${TEST_ELF}" ]]; then
+  echo "ERROR: ${TEST_ELF} was not produced" >&2
+  exit 1
+fi
+
+echo "=== running ${TEST_TARGET} on xt-run ==="
+LOG=$(mktemp)
+# --exit_with_target_code propagates gtest_main's exit code, so a failing test
+# fails this step; also assert on the gtest summary lines as a backstop.
+xt-run --turbo --exit_with_target_code "${TEST_ELF}" 2>&1 | tee "${LOG}"
+if grep -q "\[  FAILED  \]" "${LOG}"; then
+  echo "ERROR: gtest reported failures for ${TEST_TARGET}" >&2
+  exit 1
+fi
+if ! grep -q "\[  PASSED  \]" "${LOG}"; then
+  echo "ERROR: ${TEST_TARGET} did not report a gtest PASSED summary" >&2
+  exit 1
+fi
+echo "Cadence ${TARGET_DIR} op tests passed on ${XTENSA_CORE}."

.github/workflows/_xtensa_test.yml

@@ -0,0 +1,84 @@
+# Reusable: build + run the Cadence Xtensa op-level gtest tests for one core on
+# the Instruction Set Simulator (xt-run). Mirrors _xtensa_build.yml's native
+# OIDC + docker-run skeleton (running xt-run needs the same licensed toolchain),
+# then builds the gtest op-test ELF and runs it on the simulator. The runner
+# cross-compile and these op tests are separate build configs (the tests need
+# exceptions/RTTI that the runner build disables), so this is a self-contained
+# job rather than consuming the runner artifact.
+name: xtensa-test
+
+on:
+  workflow_call:
+    inputs:
+      backend:
+        description: "Cadence backend to test (hifi4 | vision | fusion_g3)"
+        required: true
+        type: string
+      ref:
+        description: "Git ref to check out"
+        required: false
+        type: string
+        default: ""
+
+jobs:
+  test:
+    name: ${{ inputs.backend }}
+    runs-on: linux.2xlarge
+    environment: cadence
+    permissions:
+      id-token: write
+      contents: read
+    steps:
+      - name: Checkout executorch
+        uses: actions/checkout@v4
+        with:
+          submodules: recursive
+          ref: ${{ inputs.ref }}
+
+      - name: Calculate docker image
+        id: calculate-docker-image
+        uses: pytorch/test-infra/.github/actions/calculate-docker-image@main
+        with:
+          docker-image-name: ci-image:executorch-ubuntu-22.04-clang12
+
+      - name: Pull docker image
+        run: docker pull "${{ steps.calculate-docker-image.outputs.docker-image }}"
+
+      - name: Assume Cadence artifacts role (host OIDC)
+        uses: aws-actions/configure-aws-credentials@v4
+        with:
+          role-to-assume: ${{ vars.CADENCE_CI_AWS_ROLE }}
+          aws-region: ${{ vars.CADENCE_CI_AWS_REGION }}
+
+      - name: Build and run op tests on xt-run
+        env:
+          DOCKER_IMAGE: ${{ steps.calculate-docker-image.outputs.docker-image }}
+          BACKEND: ${{ inputs.backend }}
+          XTENSA_S3_BUCKET: ${{ vars.CADENCE_CI_S3_BUCKET }}
+        shell: bash
+        run: |
+          set -eux
+          # OIDC/role assumption already happened on the host above; pass the
+          # resulting AWS creds and the store/backend into the CI image, where
+          # the toolchain download + op-test build + xt-run happen.
+          docker run --rm \
+            -e BACKEND -e XTENSA_S3_BUCKET \
+            -e AWS_ACCESS_KEY_ID -e AWS_SECRET_ACCESS_KEY -e AWS_SESSION_TOKEN \
+            -e AWS_DEFAULT_REGION -e AWS_REGION \
+            -v "${GITHUB_WORKSPACE}:/work/executorch" -w /work/executorch \
+            "${DOCKER_IMAGE}" \
+            bash -c '
+              set -exo pipefail
+              eval "$(/opt/conda/bin/conda shell.bash hook)"
+              conda activate "$(conda env list --json | jq -r ".envs | .[-1]")"
+              ./install_requirements.sh > /dev/null
+              pip install --quiet awscli
+              # hifi4/fusion_g3 optimized kernels need the foss-xtensa nnlib
+              # sources, which are not vendored in executorch; the cadence
+              # installer clones them. vision has no nnlib dependency.
+              if [ "${BACKEND}" != "vision" ]; then
+                backends/cadence/install_requirements.sh
+              fi
+              source .ci/scripts/setup-xtensa-tools.sh "${BACKEND}"
+              .ci/scripts/test-cadence-xtensa.sh
+            '

.github/workflows/build-cadence-runner.yml

@@ -87,6 +87,25 @@ jobs:
       backend: hifi4
       ref: ${{ (github.event_name == 'pull_request' || github.event_name == 'pull_request_target') && github.event.pull_request.head.sha || github.sha }}
 
+  # Op-level gtest tests on the Xtensa ISS, mirroring cpu-build -> cpu-test. The
+  # op tests are a self-contained cross-compile (the gtests need exceptions and
+  # RTTI that the runner build disables), so this does not consume hifi-build's
+  # artifact; needs: hifi-build only to fail fast when the backend cannot build.
+  hifi-op-test:
+    needs: hifi-build
+    if: >-
+      github.event_name == 'push' || github.event_name == 'schedule' || github.event_name == 'workflow_dispatch' ||
+      (github.event_name == 'pull_request' && github.event.pull_request.head.repo.full_name == github.repository) ||
+      (github.event_name == 'pull_request_target' && github.event.pull_request.head.repo.full_name != github.repository &&
+      contains(github.event.pull_request.labels.*.name, 'CLA Signed') && contains(github.event.pull_request.labels.*.name, 'meta-exported'))
+    permissions:
+      id-token: write
+      contents: read
+    uses: ./.github/workflows/_xtensa_test.yml
+    with:
+      backend: hifi4
+      ref: ${{ (github.event_name == 'pull_request' || github.event_name == 'pull_request_target') && github.event.pull_request.head.sha || github.sha }}
+
   vision-build:
     if: >-
       github.event_name == 'push' || github.event_name == 'schedule' || github.event_name == 'workflow_dispatch' ||

backends/cadence/CMakeLists.txt

@@ -132,3 +132,13 @@ if(EXECUTORCH_BUILD_CADENCE_RUNNER)
   endif()
   target_link_options(cadence_executor_runner PRIVATE -static -lm)
 endif()
+
+# Cadence op-level gtest tests, cross-compiled for the Xtensa ISS (run via
+# xt-run). Built only when EXECUTORCH_BUILD_CADENCE_OP_TESTS is ON and the
+# selected backend ships a tests dir. See .ci/scripts/test-cadence-xtensa.sh.
+if(EXECUTORCH_BUILD_CADENCE_OP_TESTS
+   AND EXISTS
+       "${CMAKE_CURRENT_SOURCE_DIR}/${TARGET_DIR}/operators/tests/CMakeLists.txt"
+)
+  add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/${TARGET_DIR}/operators/tests)
+endif()

backends/cadence/hifi/operators/tests/CMakeLists.txt

@@ -0,0 +1,83 @@
+# 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.
+#
+# gtest op-level tests for the Cadence HiFi backend, cross-compiled for Xtensa
+# and run on the Instruction Set Simulator (xt-run). Each test calls the kernel
+# directly (impl::HiFi::native::<op>) and checks with EXPECT_TENSOR_EQ, mirroring
+# the BUCK tests. Built only when EXECUTORCH_BUILD_CADENCE_OP_TESTS is ON; see
+# .ci/scripts/test-cadence-xtensa.sh.
+
+cmake_minimum_required(VERSION 3.19)
+
+# Tame the vendored googletest build on the Xtensa clang toolchain. It hardcodes
+# -Werror -Wconversion on its own sources, which clang-10 trips (e.g. an int ->
+# signed char narrowing in gtest-port.h); -Wno-error keeps those non-fatal. A
+# ccache launcher also preprocesses then recompiles the .i, where the -I dirs
+# are unused and clang warns; -Qunused-arguments silences that. These land after
+# googletest's own flags, so they win, and apply to this whole subtree.
+add_compile_options(-Qunused-arguments -Wno-error)
+
+# Let files say "include <executorch/path/to/header.h>".
+set(_common_include_directories
+    ${EXECUTORCH_ROOT}/.. ${EXECUTORCH_ROOT}/runtime/core/portable_type/c10
+)
+
+# Build googletest FOR the Xtensa target. find_package(GTest) would resolve the
+# HOST gtest, which cannot link Xtensa objects, so add the source tree instead.
+# The cadence toolchain uses -stdlib=libc++ (exceptions/RTTI/iostream
+# available); the ISS is single-threaded, so disable pthreads.
+if(NOT TARGET gtest)
+  set(gtest_disable_pthreads
+      ON
+      CACHE BOOL "" FORCE
+  )
+  set(INSTALL_GTEST
+      OFF
+      CACHE BOOL "" FORCE
+  )
+  add_subdirectory(
+    ${EXECUTORCH_ROOT}/third-party/googletest
+    ${CMAKE_CURRENT_BINARY_DIR}/googletest
+  )
+endif()
+
+# Op-level gtests, one source per operator. Each calls its kernel directly and
+# checks with EXPECT_TENSOR_EQ, mirroring the BUCK tests.
+add_executable(
+  cadence_hifi_op_tests
+  test_op_cat.cpp
+  test_op_dequantize_per_tensor_out.cpp
+  test_op_div.cpp
+  test_op_im2row_out.cpp
+  test_op_permute_copy.cpp
+  test_op_quantize_per_tensor.cpp
+  test_op_quantized_conv2d_out.cpp
+  test_op_quantized_matmul_out.cpp
+  test_op_quantized_relu_out.cpp
+  test_op_transpose_copy.cpp
+  ${EXECUTORCH_ROOT}/runtime/core/exec_aten/testing_util/tensor_util.cpp
+)
+target_compile_definitions(cadence_hifi_op_tests PRIVATE GTEST_HAS_PTHREAD=0)
+target_include_directories(
+  cadence_hifi_op_tests PRIVATE ${_common_include_directories}
+                                ${CMAKE_BINARY_DIR}
+)
+
+# Direct-call: link the libs that define the kernels under test. custom_ops has
+# the quantized ops; aten_ops_cadence has the aten-compliant ops (cat, div,
+# im2row, permute, transpose); cadence_kernels is the shared nnlib layer;
+# executorch provides the platform layer the tests' runtime_init() needs. No
+# cadence_ops_lib registry is needed for direct-call tests.
+target_link_libraries(
+  cadence_hifi_op_tests
+  PRIVATE gtest
+          gtest_main
+          gmock
+          executorch
+          custom_ops
+          aten_ops_cadence
+          cadence_kernels
+)