pytorch · aliafzal · Jun 10, 2026
diff --git a/.ci/scripts/build-cadence-xtensa.sh b/.ci/scripts/build-cadence-xtensa.sh
@@ -0,0 +1,79 @@
+#!/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.
+#
+# Cross-compile cadence_executor_runner for a Cadence Xtensa core and (by
+# default) smoke-test it on the Instruction Set Simulator with a trivial model.
+#
+# Requires the Xtensa toolchain env to already be set (run
+# .ci/scripts/setup-xtensa-tools.sh <backend> first): XTENSA_TOOLCHAIN,
+# TOOLCHAIN_VER, XTENSA_SYSTEM, XTENSA_CORE, XTENSAD_LICENSE_FILE,
+# CADENCE_OPT_FLAG, and xt-clang on PATH.
+#
+# Usage:
+#   .ci/scripts/build-cadence-xtensa.sh [--no-run]
+#     --no-run : compile only, skip the ISS smoke test
+
+set -euo pipefail
+
+RUN_SMOKE=1
+[[ "${1:-}" == "--no-run" ]] && RUN_SMOKE=0
+
+: "${XTENSA_TOOLCHAIN:?run setup-xtensa-tools.sh first}"
+: "${TOOLCHAIN_VER:?run setup-xtensa-tools.sh first}"
+: "${XTENSA_CORE:?run setup-xtensa-tools.sh first}"
+: "${CADENCE_OPT_FLAG:?run setup-xtensa-tools.sh first}"
+
+NPROC=$(nproc)
+echo "=== building cadence_executor_runner for ${XTENSA_CORE} (${CADENCE_OPT_FLAG}) ==="
+xt-clang --version | head -1
+
+rm -rf cmake-out
+CXXFLAGS="-fno-exceptions -fno-rtti" cmake \
+  -DCMAKE_TOOLCHAIN_FILE=./backends/cadence/cadence.cmake \
+  -DCMAKE_INSTALL_PREFIX=cmake-out \
+  -DCMAKE_BUILD_TYPE=Release \
+  -DEXECUTORCH_BUILD_CADENCE=ON \
+  "-D${CADENCE_OPT_FLAG}=ON" \
+  -DEXECUTORCH_BUILD_PORTABLE_OPS=ON \
+  -DEXECUTORCH_BUILD_CADENCE_RUNNER=ON \
+  -DEXECUTORCH_BUILD_EXECUTOR_RUNNER=OFF \
+  -DEXECUTORCH_BUILD_EXTENSION_RUNNER_UTIL=ON \
+  -DEXECUTORCH_ENABLE_LOGGING=ON \
+  -DEXECUTORCH_BUILD_PTHREADPOOL=OFF \
+  -DEXECUTORCH_BUILD_CPUINFO=OFF \
+  -DEXECUTORCH_USE_DL=OFF \
+  -DEXECUTORCH_BUILD_KERNELS_LLM=OFF \
+  -DEXECUTORCH_BUILD_DEVTOOLS=OFF \
+  -DHAVE_FNMATCH_H=OFF \
+  -DFLATCC_ALLOW_WERROR=OFF \
+  -DPYTHON_EXECUTABLE="$(which python3)" \
+  -Bcmake-out .
+
+cmake --build cmake-out --target cadence_executor_runner -j"${NPROC}"
+
+RUNNER="cmake-out/backends/cadence/cadence_executor_runner"
+if [[ ! -f "${RUNNER}" ]]; then
+  echo "ERROR: ${RUNNER} was not produced" >&2
+  exit 1
+fi
+file "${RUNNER}"
+echo "Build OK: ${RUNNER}"
+
+if [[ "${RUN_SMOKE}" == "0" ]]; then
+  echo "Skipping ISS smoke test (--no-run)."
+  exit 0
+fi
+
+echo "=== ISS smoke test: export add.pte and run on xt-run --turbo ==="
+python3 -m examples.portable.scripts.export --model_name=add >/dev/null
+LOG=$(mktemp)
+xt-run --turbo "${RUNNER}" --model_path=add.pte 2>&1 | tee "${LOG}"
+if ! grep -q "Model executed successfully" "${LOG}"; then
+  echo "ERROR: ISS smoke test did not report success for ${XTENSA_CORE}" >&2
+  exit 1
+fi
+echo "ISS smoke test passed for ${XTENSA_CORE}."
diff --git a/.ci/scripts/setup-xtensa-tools.sh b/.ci/scripts/setup-xtensa-tools.sh
@@ -0,0 +1,149 @@
+#!/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.
+#
+# Download and install the licensed Cadence Xtensa toolchain + core config for
+# a given backend, then export the environment that
+# backends/cadence/cadence.cmake and xt-run need.
+#
+# The artifacts (host tools, the core tarball, and the bundled license) cannot
+# be hosted publicly, so they are fetched at runtime from an auth-gated object
+# store. The store location is provided by the caller via XTENSA_S3_BUCKET (set
+# from a CI variable); credentials are obtained out of band before this runs.
+#
+# Usage:
+#   XTENSA_S3_BUCKET=<bucket> .ci/scripts/setup-xtensa-tools.sh <backend>
+#     backend = hifi4 | vision | fusion_g3
+#
+# In GitHub Actions this appends the toolchain env to $GITHUB_ENV so later
+# steps inherit it. Run locally to populate a workspace for manual builds.
+#
+# Modeled on .ci/scripts/setup-arm-baremetal-tools.sh.
+
+set -euo pipefail
+
+BACKEND="${1:-}"
+if [[ -z "${BACKEND}" ]]; then
+  echo "ERROR: usage: XTENSA_S3_BUCKET=<bucket> $0 <hifi4|vision|fusion_g3>" >&2
+  exit 1
+fi
+
+S3_BUCKET="${XTENSA_S3_BUCKET:-}"
+if [[ -z "${S3_BUCKET}" ]]; then
+  echo "ERROR: XTENSA_S3_BUCKET is not set (provide it from a CI variable)." >&2
+  exit 1
+fi
+S3_TOOLCHAIN_PREFIX="${XTENSA_S3_TOOLCHAIN_PREFIX:-toolchains}"
+S3_CORE_PREFIX="${XTENSA_S3_CORE_PREFIX:-cores}"
+
+# Per-backend mapping: core tarball, toolchain tarball, core name, OPT flag.
+# The toolchain's clang major must match the core's codegen plugin:
+#   hifi4 / fusion_g3 cores (RI-2022.10, clang 10) -> RI-2022.9 host tools
+#   vision core           (RJ-2025.5,  clang 15)   -> RJ-2025.5 host tools
+case "${BACKEND}" in
+  hifi4)
+    CORE_NAME="hifi4_ss_spfpu_7_et_ci2"
+    CORE_TARBALL="hifi4_ss_spfpu_7_et_ci2_linux.tgz"
+    TOOLCHAIN_TARBALL="XtensaTools_RI_2022_9_linux.tgz"
+    TOOLCHAIN_VER="RI-2022.9-linux"
+    OPT_FLAG="EXECUTORCH_NNLIB_OPT"
+    ;;
+  fusion_g3)
+    CORE_NAME="XRC_FuG3_TYP_SPVFPU_et_c2"
+    CORE_TARBALL="XRC_FuG3_TYP_SPVFPU_et_c2_linux.tgz"
+    TOOLCHAIN_TARBALL="XtensaTools_RI_2022_9_linux.tgz"
+    TOOLCHAIN_VER="RI-2022.9-linux"
+    OPT_FLAG="EXECUTORCH_FUSION_G3_OPT"
+    ;;
+  vision)
+    CORE_NAME="XRC_Vision_110_AO_et_ci2"
+    CORE_TARBALL="XRC_Vision_110_AO_et_ci2_linux.tgz"
+    TOOLCHAIN_TARBALL="XtensaTools_RJ_2025_5_linux.tgz"
+    TOOLCHAIN_VER="RJ-2025.5-linux"
+    OPT_FLAG="EXECUTORCH_VISION_OPT"
+    ;;
+  *)
+    echo "ERROR: unknown backend '${BACKEND}' (expected hifi4|vision|fusion_g3)" >&2
+    exit 1
+    ;;
+esac
+
+XTENSA_ROOT="${XTENSA_ROOT:-/tmp/xtensa}"
+TOOLS_ROOT="${XTENSA_ROOT}/tools"     # contains <ver>-linux/XtensaTools
+CORES_ROOT="${XTENSA_ROOT}/cores"     # contains <corever>-linux/<core>
+REGISTRY_ROOT="${XTENSA_ROOT}/registry/${CORE_NAME}"
+DL_DIR="${XTENSA_ROOT}/download"
+mkdir -p "${TOOLS_ROOT}" "${CORES_ROOT}" "${REGISTRY_ROOT}" "${DL_DIR}"
+
+s3_get() {
+  # $1 = s3 key, $2 = local dest
+  local key="$1" dest="$2"
+  echo "Downloading s3://${S3_BUCKET}/${key} ..."
+  aws s3 cp "s3://${S3_BUCKET}/${key}" "${dest}" --only-show-errors
+}
+
+# 1. Toolchain (host xt-clang/xt-run). Skip re-extract if already present.
+if [[ ! -d "${TOOLS_ROOT}/${TOOLCHAIN_VER}/XtensaTools" ]]; then
+  s3_get "${S3_TOOLCHAIN_PREFIX}/${TOOLCHAIN_TARBALL}" "${DL_DIR}/${TOOLCHAIN_TARBALL}"
+  tar xzf "${DL_DIR}/${TOOLCHAIN_TARBALL}" -C "${TOOLS_ROOT}"
+fi
+TOOLCHAIN_HOME="${TOOLS_ROOT}/${TOOLCHAIN_VER}/XtensaTools"
+if [[ ! -x "${TOOLCHAIN_HOME}/bin/xt-clang" ]]; then
+  echo "ERROR: xt-clang not found at ${TOOLCHAIN_HOME}/bin after extract" >&2
+  exit 1
+fi
+
+# 2. Core config (ISA libs, params, examples, bundled magic-key license).
+s3_get "${S3_CORE_PREFIX}/${CORE_TARBALL}" "${DL_DIR}/${CORE_TARBALL}"
+tar xzf "${DL_DIR}/${CORE_TARBALL}" -C "${CORES_ROOT}"
+CORE_DIR=$(echo "${CORES_ROOT}"/*/"${CORE_NAME}")
+if [[ ! -d "${CORE_DIR}" ]]; then
+  echo "ERROR: core dir for ${CORE_NAME} not found under ${CORES_ROOT}" >&2
+  exit 1
+fi
+
+# 3. Build a local Xtensa core registry with the XPG-internal build paths in
+#    the params file rewritten to our extracted toolchain + core locations.
+#    The vendor ships params referencing /././home/xpgcust/... build paths.
+PARAMS_SRC="${CORE_DIR}/config/${CORE_NAME}-params"
+TOOLS_PFX=$(sed -n 's/^install-prefix = //p' "${PARAMS_SRC}" | head -1)
+TOOLSUB_PFX=$(sed -n 's/^xtensa-tools = //p' "${PARAMS_SRC}" | head -1)
+CFG_PFX=$(sed -n 's/^config-prefix = //p' "${PARAMS_SRC}" | head -1)
+sed \
+  -e "s|${TOOLS_PFX}|${TOOLCHAIN_HOME}|g" \
+  -e "s|${TOOLSUB_PFX}|${TOOLCHAIN_HOME}/Tools|g" \
+  -e "s|${CFG_PFX}|${CORE_DIR}|g" \
+  "${PARAMS_SRC}" > "${REGISTRY_ROOT}/${CORE_NAME}-params"
+ln -sf "${CORE_NAME}-params" "${REGISTRY_ROOT}/default-params"
+
+LICENSE_FILE="${CORE_DIR}/misc/license.dat"
+
+# 4. Export environment. cadence.cmake reads XTENSA_TOOLCHAIN/TOOLCHAIN_VER;
+#    xt-clang/xt-run read XTENSA_SYSTEM/XTENSA_CORE; xtensad reads
+#    XTENSAD_LICENSE_FILE (the bundled uncounted magic key, no server needed).
+emit() {
+  # Export into the current shell (so callers that `source` this script get the
+  # vars) and append to $GITHUB_ENV (so later workflow steps inherit them too).
+  echo "$1"
+  export "${1?}"
+  if [[ -n "${GITHUB_ENV:-}" ]]; then echo "$1" >> "${GITHUB_ENV}"; fi
+}
+echo "=== Xtensa env for backend '${BACKEND}' (core ${CORE_NAME}) ==="
+emit "XTENSA_TOOLCHAIN=${TOOLS_ROOT}"
+emit "TOOLCHAIN_VER=${TOOLCHAIN_VER}"
+emit "XTENSA_SYSTEM=${REGISTRY_ROOT}"
+emit "XTENSA_CORE=${CORE_NAME}"
+emit "XTENSAD_LICENSE_FILE=${LICENSE_FILE}"
+emit "CADENCE_OPT_FLAG=${OPT_FLAG}"
+if [[ -n "${GITHUB_PATH:-}" ]]; then
+  echo "${TOOLCHAIN_HOME}/bin" >> "${GITHUB_PATH}"
+fi
+export PATH="${TOOLCHAIN_HOME}/bin:${PATH}"
+
+echo "=== sanity ==="
+xt-clang --version 2>&1 | head -1
+xt-run --show-config=cores 2>&1 | sed -n '/available/,/registry/p' | head -6
+echo "Xtensa toolchain ready for ${BACKEND}."
diff --git a/.github/workflows/build-cadence-runner.yml b/.github/workflows/build-cadence-runner.yml
@@ -50,3 +50,67 @@ jobs:
     uses: ./.github/workflows/_test_cadence.yml
     with:
       ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
+
+  # Cross-compile cadence_executor_runner for the Cadence Xtensa cores (build
+  # stage; an ISS test stage follows later, cf. cpu-build -> cpu-test). The
+  # toolchain + core configs are licensed and cannot be hosted publicly, so they
+  # are fetched at runtime from an auth-gated store using a short-lived OIDC
+  # credential; the store, role, and region are supplied via CI variables and
+  # are not committed. Fork PRs are skipped (they cannot assume the role).
+  xtensa-build:
+    name: xtensa-build-${{ matrix.backend }}
+    if: ${{ github.event_name != 'pull_request' || github.event.pull_request.head.repo.fork == false }}
+    permissions:
+      id-token: write
+      contents: read
+    strategy:
+      fail-fast: false
+      matrix:
+        # fusion_g3 is omitted until the upstream fusion_g3 operator <->
+        # nnlib-FusionG3 API skew is resolved (its op wrappers call kernels
+        # absent from the vendored nnlib, so the runner does not link).
+        backend: [hifi4, vision]
+    uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
+    with:
+      job-name: xtensa-build-${{ matrix.backend }}
+      runner: linux.2xlarge
+      docker-image: ci-image:executorch-ubuntu-22.04-clang12
+      submodules: recursive
+      ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
+      timeout: 90
+      upload-artifact: cadence-xtensa-build-${{ matrix.backend }}
+      script: |
+        set -eux
+        # The generic Linux job uses the base env, not the one set up by the image.
+        CONDA_ENV=$(conda env list --json | jq -r ".envs | .[-1]")
+        conda activate "${CONDA_ENV}"
+
+        ./install_requirements.sh > /dev/null
+        pip install --quiet awscli
+
+        # Obtain a short-lived credential for the artifact store via GitHub
+        # OIDC. The ACTIONS_ID_TOKEN_REQUEST_* vars are present because this job
+        # has id-token: write. Role and store are provided via CI variables.
+        OIDC_TOKEN=$(curl -sS \
+          -H "Authorization: bearer ${ACTIONS_ID_TOKEN_REQUEST_TOKEN}" \
+          "${ACTIONS_ID_TOKEN_REQUEST_URL}&audience=sts.amazonaws.com" | jq -r '.value')
+        CREDS=$(aws sts assume-role-with-web-identity \
+          --role-arn "${{ vars.CADENCE_CI_AWS_ROLE }}" \
+          --role-session-name "xtensa-${{ matrix.backend }}-${GITHUB_RUN_ID}" \
+          --web-identity-token "${OIDC_TOKEN}" \
+          --duration-seconds 7200 \
+          --query Credentials --output json)
+        export AWS_ACCESS_KEY_ID=$(echo "${CREDS}" | jq -r .AccessKeyId)
+        export AWS_SECRET_ACCESS_KEY=$(echo "${CREDS}" | jq -r .SecretAccessKey)
+        export AWS_SESSION_TOKEN=$(echo "${CREDS}" | jq -r .SessionToken)
+        export AWS_DEFAULT_REGION="${{ vars.CADENCE_CI_AWS_REGION }}"
+        export XTENSA_S3_BUCKET="${{ vars.CADENCE_CI_S3_BUCKET }}"
+
+        # Download + install toolchain and core, exporting the Xtensa env into
+        # this shell, then cross-compile cadence_executor_runner (build only;
+        # the ISS smoke run lives in the downstream test job).
+        source .ci/scripts/setup-xtensa-tools.sh "${{ matrix.backend }}"
+        .ci/scripts/build-cadence-xtensa.sh --no-run
+
+        # Publish the cross-compiled runner for the downstream test job.
+        cp cmake-out/backends/cadence/cadence_executor_runner "${RUNNER_ARTIFACT_DIR}/"
@@ -97,3 +97,38 @@ else()
 endif()
 
 add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/${TARGET_DIR}/operators)
+
+# Cadence executor_runner: cross-compiled ExecuTorch runner for the Xtensa ISS
+# (xt-run / xt-run --turbo). Self-contained, gflags-free argv parser, reads .pte
+# via xt-run semi-hosting.
+#
+# Usage: cmake ... -DEXECUTORCH_BUILD_CADENCE_RUNNER=ON xt-run --turbo
+# cmake-out/backends/cadence/cadence_executor_runner \ --model_path=add.pte
+if(EXECUTORCH_BUILD_CADENCE_RUNNER)
+  add_executable(cadence_executor_runner cadence_executor_runner.cpp)
+  target_compile_definitions(
+    cadence_executor_runner PRIVATE ET_ENABLE_ENUM_STRINGS=0
+  )
+  target_include_directories(
+    cadence_executor_runner
+    PRIVATE ${_common_include_directories} ${CMAKE_BINARY_DIR}
+            ${CMAKE_BINARY_DIR}/include
+  )
+  # Mirror the upstream executor_runner cadence link list (top-level
+  # CMakeLists.txt: list(APPEND _executor_runner_libs cadence_ops_lib)). Do NOT
+  # add --whole-archive: cadence_ops_lib is also pulled transitively, and
+  # forcing a second copy double-runs its static kernel-registration
+  # initializers and asserts at runtime.
+  target_link_libraries(
+    cadence_executor_runner PRIVATE executorch extension_evalue_util
+                                    extension_runner_util cadence_ops_lib
+  )
+  # Vision and Fusion-G3 ops (e.g. op_softmax) reference iDMA scheduling symbols
+  # and those cores ship libidma in their LSP. HiFi4 and generic cores do not
+  # use iDMA and their LSPs may not provide libidma, so only link it for the
+  # cores that need it.
+  if(EXECUTORCH_VISION_OPT OR EXECUTORCH_FUSION_G3_OPT)
+    target_link_options(cadence_executor_runner PRIVATE -lidma)
+  endif()
+  target_link_options(cadence_executor_runner PRIVATE -static -lm)
+endif()