From 897501a78b4aa61ebb17a61acf8d78957c4effd0 Mon Sep 17 00:00:00 2001
From: Hongrui Chen <chraac@gmail.com>
Date: Fri, 26 Dec 2025 21:17:53 +0800
Subject: [PATCH] Refactor CMake configuration for Hexagon NPU: consolidate
 device target setup and improve build structure

---
 ggml/src/ggml-qnn/npu/CMakeLists.txt        | 307 ++++++--------------
 ggml/src/ggml-qnn/npu/device/CMakeLists.txt | 135 +++++++++
 2 files changed, 223 insertions(+), 219 deletions(-)
 create mode 100644 ggml/src/ggml-qnn/npu/device/CMakeLists.txt

diff --git a/ggml/src/ggml-qnn/npu/CMakeLists.txt b/ggml/src/ggml-qnn/npu/CMakeLists.txt
index a5b6d18e34..471364c199 100644
--- a/ggml/src/ggml-qnn/npu/CMakeLists.txt
+++ b/ggml/src/ggml-qnn/npu/CMakeLists.txt
@@ -37,229 +37,98 @@ set(common_incs
 
 include_directories(${common_incs})
 
-function(add_device_target target_name DSP_ARCH IS_SIMULATOR BUILD_CPU_COUNT)
-    if(${CMAKE_BUILD_TYPE} MATCHES "Debug|Dbg")
-        set(HEXAGON_BUILD_CONFIG "Debug")
-        set(EXTRA_BUILD_FLAGS
-            VERBOSE=1
-            TREE=1
-        )
-    else()
-        set(HEXAGON_BUILD_CONFIG "Release")
-        set(EXTRA_BUILD_FLAGS)
-    endif()
+# host build
+file(GLOB common_srcs "${CMAKE_CURRENT_LIST_DIR}/common/*.cpp")
+file(GLOB host_srcs "${CMAKE_CURRENT_LIST_DIR}/host/*.cpp")
+set(stub_srcs "${CMAKE_CURRENT_BINARY_DIR}/npu_device_stub.c")
+add_library(hexagon-npu-host STATIC
+    ${common_srcs}
+    ${host_srcs}
+    ${stub_srcs}
+)
 
-    if(${GGML_SANITIZE_ADDRESS} OR ${LLAMA_SANITIZE_ADDRESS})
-        set(GGML_HEXAGON_NPU_SANITIZE_ADDRESS ON)
-    else()
-        set(GGML_HEXAGON_NPU_SANITIZE_ADDRESS OFF)
-    endif()
+# disable warnings for the stub
+set_source_files_properties(
+    ${stub_srcs}
+    PROPERTIES
+    COMPILE_FLAGS "-w"
+)
 
-    set(EXTRA_BUILD_FLAGS ${EXTRA_BUILD_FLAGS} GGML_HEXAGON_ENABLE_QUANTIZED_TENSORS=${GGML_HEXAGON_ENABLE_QUANTIZED_TENSORS})
-    set(EXTRA_BUILD_FLAGS ${EXTRA_BUILD_FLAGS} GGML_HEXAGON_NPU_SANITIZE_ADDRESS=${GGML_HEXAGON_NPU_SANITIZE_ADDRESS})
-    set(EXTRA_BUILD_FLAGS ${EXTRA_BUILD_FLAGS} GGML_HEXAGON_ENABLE_PERFORMANCE_TRACKING=${GGML_HEXAGON_ENABLE_PERFORMANCE_TRACKING})
+build_idl(idl/hexagon_npu.idl hexagon-npu-host)
 
-    set(HEXAGON_TOOLS_VARIANT $ENV{DEFAULT_TOOLS_VARIANT})
-    set(BUILD_DIR ${CMAKE_CURRENT_LIST_DIR}/hexagon_${HEXAGON_BUILD_CONFIG}_${HEXAGON_TOOLS_VARIANT}_${DSP_ARCH})
-    set(BUILD_BINARY_NAME ${BUILD_DIR}/libhexagon_npu_skel_${DSP_ARCH}.so)
+# Add compile definitions to the target
+target_compile_definitions(hexagon-npu-host PUBLIC
+    VERIFY_PRINT_ERROR
+    GGML_QNN_ENABLE_HEXAGON_BACKEND
+)
 
-    if(${IS_SIMULATOR})
-        set(HEXAGON_TOOLCHAIN_TYPE "hexagonsim")
-        set(OUTPUT_BINARY_NAME libhexagon_npu_skel_${DSP_ARCH}_sim.so)
-    else()
-        set(HEXAGON_TOOLCHAIN_TYPE "hexagon")
-        set(OUTPUT_BINARY_NAME libhexagon_npu_skel_${DSP_ARCH}.so)
-    endif()
-
-    add_custom_target(${target_name} ALL
-        COMMAND ${CMAKE_COMMAND} -E remove_directory ${BUILD_DIR}
-        COMMAND build_cmake ${HEXAGON_TOOLCHAIN_TYPE} DSP_ARCH=${DSP_ARCH} BUILD=${HEXAGON_BUILD_CONFIG} ${EXTRA_BUILD_FLAGS} -j${BUILD_CPU_COUNT}
-        COMMAND ${CMAKE_COMMAND} -E copy ${BUILD_BINARY_NAME} ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/${OUTPUT_BINARY_NAME}
-        BYPRODUCTS ${BUILD_BINARY_NAME}
-        WORKING_DIRECTORY ${CMAKE_CURRENT_LIST_DIR}
-    )
-endfunction()
-
-function(add_dsp_targets_for_host host_target DSP_ARCH BUILD_CPU_COUNT)
-    add_device_target(hexagon-npu-device-${DSP_ARCH} ${DSP_ARCH} FALSE ${BUILD_CPU_COUNT})
-    add_device_target(hexagon-npu-device-${DSP_ARCH}-sim ${DSP_ARCH} TRUE ${BUILD_CPU_COUNT})
-    add_dependencies(hexagon-npu-device-${DSP_ARCH}-sim hexagon-npu-device-${DSP_ARCH})
-    add_dependencies(${host_target} hexagon-npu-device-${DSP_ARCH}-sim)
-endfunction()
-
-if(${CMAKE_SYSTEM_NAME} MATCHES "Android|Linux|Windows")
-    # host build
-    file(GLOB common_srcs "${CMAKE_CURRENT_LIST_DIR}/common/*.cpp")
-    file(GLOB host_srcs "${CMAKE_CURRENT_LIST_DIR}/host/*.cpp")
-    set(stub_srcs "${CMAKE_CURRENT_BINARY_DIR}/npu_device_stub.c")
-    add_library(hexagon-npu-host STATIC
-        ${common_srcs}
-        ${host_srcs}
-        ${stub_srcs}
-    )
-
-    # disable warnings for the stub
-    set_source_files_properties(
-        ${stub_srcs}
-        PROPERTIES
-        COMPILE_FLAGS "-w"
-    )
-
-    build_idl(idl/hexagon_npu.idl hexagon-npu-host)
-
-    # Add compile definitions to the target
+if(GGML_HEXAGON_ENABLE_QUANTIZED_TENSORS)
     target_compile_definitions(hexagon-npu-host PUBLIC
-        VERIFY_PRINT_ERROR
-        GGML_QNN_ENABLE_HEXAGON_BACKEND
+        GGML_HEXAGON_ENABLE_QUANTIZED_TENSORS
     )
-
-    if(GGML_HEXAGON_ENABLE_QUANTIZED_TENSORS)
-        target_compile_definitions(hexagon-npu-host PUBLIC
-            GGML_HEXAGON_ENABLE_QUANTIZED_TENSORS
-        )
-    endif()
-
-    target_include_directories(hexagon-npu-host PRIVATE
-        ${HEXAGON_SDK_ROOT}/ipc/fastrpc/rpcmem/inc/
-        ${QNN_SDK_ROOT}/include/QNN/
-        ${CMAKE_CURRENT_LIST_DIR}/host/
-        ${CMAKE_CURRENT_LIST_DIR}/
-    )
-
-    target_include_directories(hexagon-npu-host PUBLIC
-        ${HEXAGON_SDK_ROOT}/incs/ # TODO: this is for rpc-mem
-    )
-
-    if(NOT ${CMAKE_SYSTEM_NAME} MATCHES "Windows")
-        set_target_properties(hexagon-npu-host PROPERTIES OUTPUT_NAME "hexagon_npu")
-    endif()
-
-    if(${CMAKE_SYSTEM_NAME} MATCHES "Android|Linux")
-        target_link_options(hexagon-npu-host PUBLIC -pie)
-    endif()
-
-    if(GGML_HEXAGON_ENABLE_PERFORMANCE_TRACKING)
-        message("GGML_HEXAGON_ENABLE_PERFORMANCE_TRACKING is enabled")
-        target_compile_definitions(hexagon-npu-host PUBLIC GGML_HEXAGON_ENABLE_PERFORMANCE_TRACKING)
-    else()
-        message("GGML_HEXAGON_ENABLE_PERFORMANCE_TRACKING is disabled")
-    endif()
-
-    link_options(hexagon-npu-host)
-
-    choose_dsprpc("3" dsprpc) # cdsprpc
-    link_custom_library(hexagon-npu-host ${dsprpc})
-
-    cmake_host_system_information(RESULT BUILD_CPU_COUNT QUERY NUMBER_OF_PHYSICAL_CORES)
-    add_dsp_targets_for_host(hexagon-npu-host "v73" ${BUILD_CPU_COUNT})
-    add_dsp_targets_for_host(hexagon-npu-host "v75" ${BUILD_CPU_COUNT})
-    add_dsp_targets_for_host(hexagon-npu-host "v79" ${BUILD_CPU_COUNT})
-
-    list(APPEND NPU_RUNTIME_LIBS "${HEXAGON_SDK_ROOT}/tools/utils/sysmon/sysMonApp")
-    list(APPEND NPU_RUNTIME_LIBS "${HEXAGON_SDK_ROOT}/tools/utils/sysmon/sysMonAppLE")
-
-    foreach(RUNTIME_LIB ${NPU_RUNTIME_LIBS})
-        message("Copy: ${RUNTIME_LIB} -> ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}")
-        add_custom_command(
-            TARGET hexagon-npu-host POST_BUILD
-            COMMAND ${CMAKE_COMMAND} -E copy
-            ${RUNTIME_LIB}
-            ${CMAKE_RUNTIME_OUTPUT_DIRECTORY})
-    endforeach()
-else()
-    # hexagon npu build, this section will run inside the `build_cmake` script
-    cmake_minimum_required(VERSION 3.14.3)
-    project(hexagon_npu C CXX ASM)
-
-    # check if QNN_SDK_ROOT is set
-    if(NOT DEFINED ENV{QNN_SDK_ROOT})
-        message(FATAL_ERROR "QNN_SDK_ROOT not defined")
-    endif()
-
-    set(QNN_SDK_ROOT $ENV{QNN_SDK_ROOT})
-    message("QNN_SDK_ROOT: ${QNN_SDK_ROOT}")
-    message("GGML_HEXAGON_ENABLE_QUANTIZED_TENSORS: ${GGML_HEXAGON_ENABLE_QUANTIZED_TENSORS}")
-
-    include_directories(
-        ${QNN_SDK_ROOT}/include/QNN/
-    )
-    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -std=c++17")
-
-    file(GLOB common_srcs "${CMAKE_CURRENT_LIST_DIR}/common/*.cpp")
-    file(GLOB device_srcs "${CMAKE_CURRENT_LIST_DIR}/device/*.cpp")
-    file(GLOB device_op_srcs "${CMAKE_CURRENT_LIST_DIR}/device/op/*.cpp")
-    set(skel_srcs "${CMAKE_CURRENT_BINARY_DIR}/npu_device_skel.c")
-    add_library(hexagon_npu_skel_OBJS OBJECT
-        ${common_srcs}
-        ${device_srcs}
-        ${device_op_srcs}
-        ${skel_srcs}
-    )
-
-    if(CMAKE_BUILD_TYPE MATCHES "Debug|Dbg")
-        message("Debug build, enable all logging")
-        target_compile_definitions(hexagon_npu_skel_OBJS PUBLIC
-            _DEBUG
-            DEBUG_LOGGING
-        )
-    else()
-        message("Release build, disable debug logging")
-        target_compile_definitions(hexagon_npu_skel_OBJS PUBLIC
-            NDEBUG
-            RELEASE_LOGGING
-        )
-    endif()
-
-    if(GGML_HEXAGON_ENABLE_QUANTIZED_TENSORS)
-        message("GGML_HEXAGON_ENABLE_QUANTIZED_TENSORS is enabled")
-        target_compile_definitions(hexagon_npu_skel_OBJS PUBLIC
-            GGML_HEXAGON_ENABLE_QUANTIZED_TENSORS
-        )
-    endif()
-
-    if(GGML_HEXAGON_NPU_SANITIZE_ADDRESS)
-        message("GGML_HEXAGON_NPU_SANITIZE_ADDRESS is enabled")
-        target_compile_options(hexagon_npu_skel_OBJS PUBLIC
-            -fsanitize=address -fno-omit-frame-pointer
-        )
-        target_link_options(hexagon_npu_skel_OBJS PUBLIC
-            -fsanitize=address
-        )
-    endif()
-
-    if(GGML_HEXAGON_ENABLE_PERFORMANCE_TRACKING)
-        message("GGML_HEXAGON_ENABLE_PERFORMANCE_TRACKING is enabled")
-        target_compile_definitions(hexagon_npu_skel_OBJS PUBLIC
-            GGML_HEXAGON_ENABLE_PERFORMANCE_TRACKING
-        )
-    endif()
-
-    build_idl(idl/hexagon_npu.idl hexagon_npu_skel_OBJS)
-
-    add_subdirectory(${HEXAGON_SDK_ROOT}/libs/qprintf qprintf_dir)
-    target_include_directories(hexagon_npu_skel_OBJS PUBLIC
-        ${HEXAGON_SDK_ROOT}/libs/qprintf/inc/
-
-        # TODO: find a better way to include these
-        ${CMAKE_CURRENT_LIST_DIR}/device/
-        ${CMAKE_CURRENT_LIST_DIR}/device/op/
-    )
-
-    # disable warnings for the skel
-    set_source_files_properties(
-        ${skel_srcs}
-        PROPERTIES
-        COMPILE_FLAGS "-w"
-    )
-
-    add_library(hexagon_npu_skel SHARED $<TARGET_OBJECTS:hexagon_npu_skel_OBJS>)
-    target_link_libraries(hexagon_npu_skel
-        ${HEXAGON_LIB_DIR}/${HEXAGON_ARCH}/G0/pic/libc++abi.so.1
-        ${HEXAGON_LIB_DIR}/${HEXAGON_ARCH}/G0/pic/libc++.so.1
-        ${HEXAGON_LIB_DIR}/${HEXAGON_ARCH}/G0/pic/libc.so
-    )
-    set_target_properties(hexagon_npu_skel PROPERTIES OUTPUT_NAME "hexagon_npu_skel_${HEXAGON_ARCH}")
-    target_link_libraries(hexagon_npu_skel qprintf_static)
-
-    copy_binaries(hexagon_npu_skel)
 endif()
+
+target_include_directories(hexagon-npu-host PRIVATE
+    ${HEXAGON_SDK_ROOT}/ipc/fastrpc/rpcmem/inc/
+    ${QNN_SDK_ROOT}/include/QNN/
+    ${CMAKE_CURRENT_LIST_DIR}/host/
+    ${CMAKE_CURRENT_LIST_DIR}/
+)
+
+target_include_directories(hexagon-npu-host PUBLIC
+    ${HEXAGON_SDK_ROOT}/incs/ # TODO: this is for rpc-mem
+)
+
+if(NOT ${CMAKE_SYSTEM_NAME} MATCHES "Windows")
+    set_target_properties(hexagon-npu-host PROPERTIES OUTPUT_NAME "hexagon_npu")
+endif()
+
+if(${CMAKE_SYSTEM_NAME} MATCHES "Android|Linux")
+    target_link_options(hexagon-npu-host PUBLIC -pie)
+endif()
+
+if(GGML_HEXAGON_ENABLE_PERFORMANCE_TRACKING)
+    message("GGML_HEXAGON_ENABLE_PERFORMANCE_TRACKING is enabled")
+    target_compile_definitions(hexagon-npu-host PUBLIC GGML_HEXAGON_ENABLE_PERFORMANCE_TRACKING)
+else()
+    message("GGML_HEXAGON_ENABLE_PERFORMANCE_TRACKING is disabled")
+endif()
+
+link_options(hexagon-npu-host)
+
+choose_dsprpc("3" dsprpc) # cdsprpc
+link_custom_library(hexagon-npu-host ${dsprpc})
+
+cmake_host_system_information(RESULT BUILD_CPU_COUNT QUERY NUMBER_OF_PHYSICAL_CORES)
+
+# Build HTP bits
+set(HTP_CMAKE_ARGS
+    -DCMAKE_TOOLCHAIN_FILE=${CMAKE_CURRENT_SOURCE_DIR}/../../ggml-hexagon/htp/cmake-toolchain.cmake
+    -DCMAKE_BUILD_TYPE=Release
+    -DCMAKE_INSTALL_LIBDIR=${CMAKE_CURRENT_BINARY_DIR}
+    -DHEXAGON_SDK_ROOT=$ENV{HEXAGON_SDK_ROOT}
+    -DHEXAGON_TOOLS_ROOT=$ENV{HEXAGON_TOOLS_ROOT}
+    -DHEXAGON_HTP_DEBUG=${GGML_HEXAGON_HTP_DEBUG}
+    -DGGML_HEXAGON_FP32_QUANTIZE_GROUP_SIZE=${GGML_HEXAGON_FP32_QUANTIZE_GROUP_SIZE})
+
+ExternalProject_Add(hexagon_npu_skel_v73
+    SOURCE_DIR ${CMAKE_CURRENT_SOURCE_DIR}/device BUILD_ALWAYS ON
+    CMAKE_ARGS ${HTP_CMAKE_ARGS} -DDSP_VERSION=v73 -DPREBUILT_LIB_DIR="toolv19_v73")
+ExternalProject_Add(hexagon_npu_skel_v75
+    SOURCE_DIR ${CMAKE_CURRENT_SOURCE_DIR}/device BUILD_ALWAYS ON
+    CMAKE_ARGS ${HTP_CMAKE_ARGS} -DDSP_VERSION=v75 -DPREBUILT_LIB_DIR="toolv19_v75")
+ExternalProject_Add(hexagon_npu_skel_v79
+    SOURCE_DIR ${CMAKE_CURRENT_SOURCE_DIR}/device BUILD_ALWAYS ON
+    CMAKE_ARGS ${HTP_CMAKE_ARGS} -DDSP_VERSION=v79 -DPREBUILT_LIB_DIR="toolv19_v79")
+
+list(APPEND NPU_RUNTIME_LIBS "${HEXAGON_SDK_ROOT}/tools/utils/sysmon/sysMonApp")
+list(APPEND NPU_RUNTIME_LIBS "${HEXAGON_SDK_ROOT}/tools/utils/sysmon/sysMonAppLE")
+
+# Install Hexagon skels required at runtime
+install(FILES
+    ${CMAKE_CURRENT_BINARY_DIR}/libhexagon_npu_skel_v73.so
+    ${CMAKE_CURRENT_BINARY_DIR}/libhexagon_npu_skel_v75.so
+    ${CMAKE_CURRENT_BINARY_DIR}/libhexagon_npu_skel_v79.so
+    ${HEXAGON_SDK_ROOT}/tools/utils/sysmon/sysMonApp
+    ${HEXAGON_SDK_ROOT}/tools/utils/sysmon/sysMonAppLE
+    TYPE LIB)
diff --git a/ggml/src/ggml-qnn/npu/device/CMakeLists.txt b/ggml/src/ggml-qnn/npu/device/CMakeLists.txt
new file mode 100644
index 0000000000..167c0bb1fe
--- /dev/null
+++ b/ggml/src/ggml-qnn/npu/device/CMakeLists.txt
@@ -0,0 +1,135 @@
+
+# hexagon npu build, this section will run inside the `build_cmake` script
+cmake_minimum_required(VERSION 3.14.3)
+project(hexagon_npu C CXX ASM)
+
+enable_language(ASM)
+cmake_policy(SET CMP0115 OLD)
+
+if(DEFINED ENV{HEXAGON_SDK_ROOT})
+    set(HEXAGON_SDK_ROOT $ENV{HEXAGON_SDK_ROOT})
+    message("HEXAGON_SDK_ROOT (from environment): ${HEXAGON_SDK_ROOT}")
+elseif(DEFINED HEXAGON_SDK_ROOT)
+    message("HEXAGON_SDK_ROOT: ${HEXAGON_SDK_ROOT}")
+else()
+    message(FATAL_ERROR "HEXAGON_SDK_ROOT not defined")
+endif()
+
+if(${CMAKE_SYSTEM_NAME} MATCHES "Android")
+    set(PREBUILT_LIB_DIR "android_aarch64")
+elseif(${CMAKE_SYSTEM_NAME} MATCHES "Linux")
+    set(PREBUILT_LIB_DIR "UbuntuARM_aarch64")
+elseif(${CMAKE_SYSTEM_NAME} MATCHES "Windows")
+    # Windows
+    set(PREBUILT_LIB_DIR "windows_aarch64")
+endif()
+
+if(HEXAGON_SDK_ROOT)
+    include(${HEXAGON_SDK_ROOT}/build/cmake/hexagon_fun.cmake)
+else()
+    include(${HEXAGON_CMAKE_ROOT}/hexagon_fun.cmake)
+endif()
+
+# Base Include dirs for the Project
+set(common_incs
+    ${CMAKE_CURRENT_BINARY_DIR}/
+    ${HEXAGON_SDK_ROOT}/incs/
+    ${HEXAGON_SDK_ROOT}/incs/stddef/
+    ${HEXAGON_SDK_ROOT}/incs/HAP/
+    ${HEXAGON_SDK_ROOT}/rtos/qurt/
+    ${HEXAGON_SDK_ROOT}/utils/examples/
+)
+
+include_directories(${common_incs})
+
+# check if QNN_SDK_ROOT is set
+if(NOT DEFINED ENV{QNN_SDK_ROOT})
+    message(FATAL_ERROR "QNN_SDK_ROOT not defined")
+endif()
+
+set(QNN_SDK_ROOT $ENV{QNN_SDK_ROOT})
+message("QNN_SDK_ROOT: ${QNN_SDK_ROOT}")
+message("GGML_HEXAGON_ENABLE_QUANTIZED_TENSORS: ${GGML_HEXAGON_ENABLE_QUANTIZED_TENSORS}")
+
+include_directories(
+    ${QNN_SDK_ROOT}/include/QNN/
+)
+set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -std=c++17")
+
+file(GLOB common_srcs "${CMAKE_CURRENT_LIST_DIR}/common/*.cpp")
+file(GLOB device_srcs "${CMAKE_CURRENT_LIST_DIR}/device/*.cpp")
+file(GLOB device_op_srcs "${CMAKE_CURRENT_LIST_DIR}/device/op/*.cpp")
+set(skel_srcs "${CMAKE_CURRENT_BINARY_DIR}/npu_device_skel.c")
+add_library(hexagon_npu_skel_OBJS OBJECT
+    ${common_srcs}
+    ${device_srcs}
+    ${device_op_srcs}
+    ${skel_srcs}
+)
+
+if(CMAKE_BUILD_TYPE MATCHES "Debug|Dbg")
+    message("Debug build, enable all logging")
+    target_compile_definitions(hexagon_npu_skel_OBJS PUBLIC
+        _DEBUG
+        DEBUG_LOGGING
+    )
+else()
+    message("Release build, disable debug logging")
+    target_compile_definitions(hexagon_npu_skel_OBJS PUBLIC
+        NDEBUG
+        RELEASE_LOGGING
+    )
+endif()
+
+if(GGML_HEXAGON_ENABLE_QUANTIZED_TENSORS)
+    message("GGML_HEXAGON_ENABLE_QUANTIZED_TENSORS is enabled")
+    target_compile_definitions(hexagon_npu_skel_OBJS PUBLIC
+        GGML_HEXAGON_ENABLE_QUANTIZED_TENSORS
+    )
+endif()
+
+if(GGML_HEXAGON_NPU_SANITIZE_ADDRESS)
+    message("GGML_HEXAGON_NPU_SANITIZE_ADDRESS is enabled")
+    target_compile_options(hexagon_npu_skel_OBJS PUBLIC
+        -fsanitize=address -fno-omit-frame-pointer
+    )
+    target_link_options(hexagon_npu_skel_OBJS PUBLIC
+        -fsanitize=address
+    )
+endif()
+
+if(GGML_HEXAGON_ENABLE_PERFORMANCE_TRACKING)
+    message("GGML_HEXAGON_ENABLE_PERFORMANCE_TRACKING is enabled")
+    target_compile_definitions(hexagon_npu_skel_OBJS PUBLIC
+        GGML_HEXAGON_ENABLE_PERFORMANCE_TRACKING
+    )
+endif()
+
+build_idl(../idl/hexagon_npu.idl hexagon_npu_skel_OBJS)
+
+add_subdirectory(${HEXAGON_SDK_ROOT}/libs/qprintf qprintf_dir)
+target_include_directories(hexagon_npu_skel_OBJS PUBLIC
+    ${HEXAGON_SDK_ROOT}/libs/qprintf/inc/
+
+    # TODO: find a better way to include these
+    ${CMAKE_CURRENT_LIST_DIR}
+    ${CMAKE_CURRENT_LIST_DIR}/op/
+)
+
+# disable warnings for the skel
+set_source_files_properties(
+    ${skel_srcs}
+    PROPERTIES
+    COMPILE_FLAGS "-w"
+)
+
+add_library(hexagon_npu_skel SHARED $<TARGET_OBJECTS:hexagon_npu_skel_OBJS>)
+target_link_libraries(hexagon_npu_skel
+    ${HEXAGON_LIB_DIR}/${HEXAGON_ARCH}/G0/pic/libc++abi.so.1
+    ${HEXAGON_LIB_DIR}/${HEXAGON_ARCH}/G0/pic/libc++.so.1
+    ${HEXAGON_LIB_DIR}/${HEXAGON_ARCH}/G0/pic/libc.so
+)
+set_target_properties(hexagon_npu_skel PROPERTIES OUTPUT_NAME "hexagon_npu_skel_${HEXAGON_ARCH}")
+target_link_libraries(hexagon_npu_skel qprintf_static)
+
+copy_binaries(hexagon_npu_skel)