Improve QNN backend build-from-source user experience

backends/qualcomm/CMakeLists.txt

@@ -33,53 +33,40 @@ if(NOT DEFINED QNN_SDK_ROOT AND DEFINED ENV{QNN_SDK_ROOT})
   )
 endif()
 
-# Last-resort fallback: download during cmake configure when building wheels and
-# QNN_SDK_ROOT was not provided externally.
-if(NOT DEFINED QNN_SDK_ROOT AND EXECUTORCH_BUILD_WHEEL_DO_NOT_USE)
-  set(_qnn_default_sdk_dir "${CMAKE_CURRENT_BINARY_DIR}/sdk/qnn")
-
-  if(EXISTS "${_qnn_default_sdk_dir}" AND EXISTS "${_qnn_default_sdk_dir}/lib")
-    message(STATUS "Found cached Qualcomm SDK at ${_qnn_default_sdk_dir}")
-    set(QNN_SDK_ROOT
-        ${_qnn_default_sdk_dir}
-        CACHE PATH "Qualcomm SDK root directory" FORCE
-    )
-  else()
-    message(STATUS "Downloading Qualcomm SDK (fallback)")
-    execute_process(
-      COMMAND
-        ${PYTHON_EXECUTABLE}
-        ${EXECUTORCH_SOURCE_DIR}/backends/qualcomm/scripts/download_qnn_sdk.py
-        --dst-folder ${_qnn_default_sdk_dir} --print-sdk-path
-      WORKING_DIRECTORY ${EXECUTORCH_SOURCE_DIR}
-      RESULT_VARIABLE _qnn_sdk_download_result
-      OUTPUT_VARIABLE _qnn_sdk_download_output
-      ERROR_VARIABLE _qnn_sdk_download_error
-      OUTPUT_STRIP_TRAILING_WHITESPACE
-    )
-    if(NOT _qnn_sdk_download_result EQUAL 0 OR _qnn_sdk_download_output
-                                               STREQUAL ""
-    )
-      message(
-        FATAL_ERROR
-          "Failed to download Qualcomm SDK. stdout: ${_qnn_sdk_download_output}\n"
-          "stderr: ${_qnn_sdk_download_error}"
-      )
-    endif()
-    set(QNN_SDK_ROOT
-        ${_qnn_sdk_download_output}
-        CACHE PATH "Qualcomm SDK root directory" FORCE
+# Auto-download QNN SDK when QNN_SDK_ROOT was not provided externally. The SDK
+# is cached in ~/.cache/executorch/qnn/ so it survives build dir cleans and is
+# shared across editable installs, build.sh, and wheel builds.
+if(NOT DEFINED QNN_SDK_ROOT)
+  message(STATUS "QNN_SDK_ROOT not set. Auto-downloading QNN SDK...")
+  execute_process(
+    COMMAND
+      ${PYTHON_EXECUTABLE}
+      ${EXECUTORCH_SOURCE_DIR}/backends/qualcomm/scripts/download_qnn_sdk.py
+      --print-sdk-path
+    WORKING_DIRECTORY ${EXECUTORCH_SOURCE_DIR}
+    RESULT_VARIABLE _qnn_sdk_download_result
+    OUTPUT_VARIABLE _qnn_sdk_download_output
+    ERROR_VARIABLE _qnn_sdk_download_error
+    OUTPUT_STRIP_TRAILING_WHITESPACE
+  )
+  if(NOT _qnn_sdk_download_result EQUAL 0 OR _qnn_sdk_download_output STREQUAL
+                                             ""
+  )
+    message(
+      FATAL_ERROR
+        "Failed to download Qualcomm SDK.\n"
+        "stderr: ${_qnn_sdk_download_error}\n"
+        "Set QNN_SDK_ROOT manually, e.g. cmake <..> -DQNN_SDK_ROOT=<...>"
     )
   endif()
+  set(QNN_SDK_ROOT
+      ${_qnn_sdk_download_output}
+      CACHE PATH "Qualcomm SDK root directory" FORCE
+  )
   set(ENV{QNN_SDK_ROOT} ${QNN_SDK_ROOT})
 endif()
 
-if(NOT DEFINED QNN_SDK_ROOT)
-  message(
-    FATAL_ERROR
-      "Please define QNN_SDK_ROOT, e.g. cmake <..> -DQNN_SDK_ROOT=<...>"
-  )
-elseif(CMAKE_TOOLCHAIN_FILE MATCHES ".*(iOS|ios\.toolchain)\.cmake$")
+if(CMAKE_TOOLCHAIN_FILE MATCHES ".*(iOS|ios\.toolchain)\.cmake$")
   message(FATAL_ERROR "ios is not supported by Qualcomm AI Engine Direct")
 endif()
 

backends/qualcomm/scripts/build.sh

@@ -16,21 +16,34 @@ if [[ "$(uname -s)" == "Darwin" ]]; then
     exit 1
 fi
 
-if [[ -z ${QNN_SDK_ROOT} ]]; then
-    echo "Please export QNN_SDK_ROOT=/path/to/qnn_sdk"
-    exit -1
-fi
+SCRIPT_DIR="$( cd "$(dirname "$0")" ; pwd -P)"
+
+# If QNN_SDK_ROOT is set, pass it to cmake. Otherwise cmake will
+# auto-download the SDK via download_qnn_sdk.py during configure.
+if [[ -n ${QNN_SDK_ROOT} ]]; then
+    QNN_SDK_CMAKE_FLAG="-DQNN_SDK_ROOT=${QNN_SDK_ROOT}"
 
+    # Ensure LD_LIBRARY_PATH includes QNN SDK libs
+    QNN_LIB_DIR="${QNN_SDK_ROOT}/lib/x86_64-linux-clang"
+    if [[ -d "${QNN_LIB_DIR}" ]] && [[ ":${LD_LIBRARY_PATH:-}:" != *":${QNN_LIB_DIR}:"* ]]; then
+        export LD_LIBRARY_PATH="${QNN_LIB_DIR}:${LD_LIBRARY_PATH:-}"
+    fi
+else
+    QNN_SDK_CMAKE_FLAG=""
+    echo "[QNN] QNN_SDK_ROOT not set. SDK will be auto-downloaded during cmake configure."
+fi
 
 set -o xtrace
 
 usage() {
   echo "Usage: Build the aarch64 version of executor runner or the python interface of Qnn Manager"
-  echo "First, you need to set the environment variable for QNN_SDK_ROOT"
-  echo ", and if you want to build the android version of executor runner"
-  echo ", you need to export ANDROID_NDK_ROOT=/path/to/android_ndkXX"
+  echo ""
+  echo "QNN SDK and Android NDK will be auto-downloaded if not set."
+  echo "To use a custom SDK, export QNN_SDK_ROOT=/path/to/qnn_sdk"
+  echo "To use a custom NDK, export ANDROID_NDK_ROOT=/path/to/android_ndkXX"
   echo "(or export TOOLCHAIN_ROOT_HOST=/path/to/sysroots/xx_host, "
   echo "TOOLCHAIN_ROOT_TARGET=/path/to/sysroots/xx_target for linux embedded with --enable_linux_embedded)"
+  echo ""
   echo "e.g.: executorch$ ./backends/qualcomm/scripts/build.sh --skip_x86_64"
   exit 1
 }
@@ -51,13 +64,13 @@ BUILD_TYPE="RelWithDebInfo"
 BUILD_JOB_NUMBER="16"
 
 # Default to use CDSP for now
-DSP_TYPE=3 
+DSP_TYPE=3
 
-if [ -z PYTHON_EXECUTABLE ]; then
+if [ -z "$PYTHON_EXECUTABLE" ]; then
   PYTHON_EXECUTABLE="python3"
 fi
 
-if [ -z BUCK2 ]; then
+if [ -z "$BUCK2" ]; then
   BUCK2="buck2"
 fi
 
@@ -85,8 +98,14 @@ PRJ_ROOT="$( cd "$(dirname "$0")/../../.." ; pwd -P)"
 
 if [ "$BUILD_ANDROID" = true ]; then
     if [[ -z ${ANDROID_NDK_ROOT} ]]; then
-        echo "Please export ANDROID_NDK_ROOT=/path/to/android_ndkXX"
-        exit -1
+        echo "[QNN] ANDROID_NDK_ROOT not set. Auto-downloading Android NDK..."
+        source "${SCRIPT_DIR}/install_qnn_sdk.sh"
+        setup_android_ndk
+        if [[ -z ${ANDROID_NDK_ROOT} ]]; then
+            echo "[QNN] Error: Failed to download Android NDK."
+            echo "[QNN] Set ANDROID_NDK_ROOT manually."
+            exit 1
+        fi
     fi
 
     BUILD_ROOT=$PRJ_ROOT/$CMAKE_ANDROID
@@ -112,7 +131,7 @@ if [ "$BUILD_ANDROID" = true ]; then
         -DEXECUTORCH_BUILD_EXTENSION_TENSOR=ON \
         -DEXECUTORCH_ENABLE_EVENT_TRACER=ON \
         -DEXECUTORCH_ENABLE_LOGGING=ON \
-        -DQNN_SDK_ROOT=$QNN_SDK_ROOT \
+        ${QNN_SDK_CMAKE_FLAG} \
         -DCMAKE_TOOLCHAIN_FILE=$ANDROID_NDK_ROOT/build/cmake/android.toolchain.cmake \
         -DANDROID_ABI='arm64-v8a' \
         -DEXECUTORCH_BUILD_KERNELS_QUANTIZED=ON \
@@ -126,7 +145,7 @@ if [ "$BUILD_ANDROID" = true ]; then
     CMAKE_PREFIX_PATH="${BUILD_ROOT};${BUILD_ROOT}/third-party/gflags;"
 
     # DSP_TYPE variable only matters when building direct_mode.
-    # Ignore the variable for traditional mode. 
+    # Ignore the variable for traditional mode.
 
     if [ "$BUILD_HEXAGON" = "true" ]; then
         DIRECT_MODE_FLAG="-DBUILD_DIRECT_MODE=ON"
@@ -274,7 +293,7 @@ if [ "$BUILD_OE_LINUX" = true ]; then
         -DEXECUTORCH_BUILD_EXTENSION_TENSOR=ON \
         -DEXECUTORCH_ENABLE_EVENT_TRACER=ON \
         -DEXECUTORCH_ENABLE_LOGGING=ON \
-        -DQNN_SDK_ROOT=$QNN_SDK_ROOT \
+        ${QNN_SDK_CMAKE_FLAG} \
         -DEXECUTORCH_BUILD_KERNELS_QUANTIZED=ON \
         -DPYTHON_EXECUTABLE=$PYTHON_EXECUTABLE \
         -B$BUILD_ROOT
@@ -335,7 +354,7 @@ if [ "$BUILD_X86_64" = true ]; then
     cmake \
         -DCMAKE_BUILD_TYPE=$BUILD_TYPE \
         -DCMAKE_INSTALL_PREFIX=$BUILD_ROOT \
-        -DQNN_SDK_ROOT=${QNN_SDK_ROOT} \
+        ${QNN_SDK_CMAKE_FLAG} \
         -DEXECUTORCH_BUILD_QNN=ON \
         -DEXECUTORCH_BUILD_DEVTOOLS=ON \
         -DEXECUTORCH_BUILD_EXTENSION_LLM=ON \

backends/qualcomm/scripts/download_qnn_sdk.py

@@ -23,18 +23,23 @@
     _handler.setFormatter(logging.Formatter("%(message)s"))
     logger.addHandler(_handler)
     logger.setLevel(logging.INFO)
+    logger.propagate = False
 
 PKG_ROOT = pathlib.Path(__file__).parent.parent
 
+# Output stream for progress messages. Defaults to stdout, but redirected to
+# stderr when --print-sdk-path is used (so stdout only contains the path).
+_output_stream = sys.stdout
+
 
 def _progress(msg: str) -> None:
     """Print a progress line with carriage return (no newline). Not suited for logging."""
-    print(msg, end="", flush=True)
+    print(msg, end="", flush=True, file=_output_stream)
 
 
 def _progress_newline() -> None:
     """End a progress line."""
-    print(flush=True)
+    print(flush=True, file=_output_stream)
 
 
 ##########################
@@ -317,6 +322,10 @@ def _download_qnn_sdk(
         logger.info("[QNN] Skipping Qualcomm SDK (only supported on Linux x86).")
         return None
 
+    if dst_folder.exists() and any(dst_folder.iterdir()):
+        logger.info(f"[QNN] Using cached QNN SDK v{QNN_VERSION} at {dst_folder}")
+        return dst_folder
+
     logger.info(f"[QNN] Downloading Qualcomm AI Runtime SDK v{QNN_VERSION}...")
 
     dst_folder.mkdir(parents=True, exist_ok=True)
@@ -761,6 +770,14 @@ def main(argv: Optional[List[str]] = None) -> int:
     )
     args = parser.parse_args(argv)
 
+    # When --print-sdk-path is used, stdout must contain ONLY the SDK path.
+    # Redirect all logger and progress output to stderr.
+    if args.print_sdk_path:
+        global _output_stream
+        _output_stream = sys.stderr
+        for handler in logger.handlers:
+            handler.stream = sys.stderr
+
     logging.basicConfig(level=logging.INFO)
 
     dst = args.dst_folder if args.dst_folder else _get_sdk_dir()

tools/cmake/preset/pybind.cmake

@@ -37,7 +37,7 @@ elseif(CMAKE_SYSTEM_NAME STREQUAL "Linux")
   set_overridable_option(EXECUTORCH_BUILD_EXTENSION_LLM_RUNNER ON)
   set_overridable_option(EXECUTORCH_BUILD_EXTENSION_LLM ON)
   if(CMAKE_SYSTEM_PROCESSOR MATCHES "^(x86_64|amd64|i.86)$")
-    set_overridable_option(EXECUTORCH_BUILD_QNN OFF)
+    set_overridable_option(EXECUTORCH_BUILD_QNN ON)
   endif()
 elseif(CMAKE_SYSTEM_NAME STREQUAL "Windows" OR CMAKE_SYSTEM_NAME STREQUAL
                                                "WIN32"