Hi Andrew, On 24/01/25 03:44, Andrew Davis wrote:
On 1/23/25 11:25 AM, Chirag wrote:From: Chirag Shilwant <[email protected]>- TensorFlow Lite is TensorFlow's lightweight solution for mobile and embedded devices. It enables low-latency inference of on-device machine learning models with a small binary size. - The tensorflow-lite v2.16 recipe in meta-arago currently supports, - CPU delegate (with RUY enabled), XNNPACK delegate & ArmNN via external delegate for ARMv8 based Cortex-A platforms. - CPU delegate (with RUY enabled) & ArmNN via external delegate for ARMv7 based Cortex-A platforms. Signed-off-by: Chirag Shilwant <[email protected]> --- ...eLists-for-building-_pywrap_tensorfl.patch | 37 ++++ ...eLists-for-building-shared-object-li.patch | 55 +++++ ...uild-issues-observed-in-TFlite-v2.16.patch | 36 ++++ .../tensorflow-lite/tensorflow-lite_2.16.bb | 203 ++++++++++++++++++ 4 files changed, 331 insertions(+)create mode 100644 meta-arago-extras/recipes-framework/tensorflow-lite/tensorflow-lite/0001-lite-Update-CMakeLists-for-building-_pywrap_tensorfl.patch create mode 100644 meta-arago-extras/recipes-framework/tensorflow-lite/tensorflow-lite/0002-lite-Update-CMakeLists-for-building-shared-object-li.patch create mode 100644 meta-arago-extras/recipes-framework/tensorflow-lite/tensorflow-lite/0003-CMakeLists-Fix-build-issues-observed-in-TFlite-v2.16.patch create mode 100644 meta-arago-extras/recipes-framework/tensorflow-lite/tensorflow-lite_2.16.bbdiff --git a/meta-arago-extras/recipes-framework/tensorflow-lite/tensorflow-lite/0001-lite-Update-CMakeLists-for-building-_pywrap_tensorfl.patch b/meta-arago-extras/recipes-framework/tensorflow-lite/tensorflow-lite/0001-lite-Update-CMakeLists-for-building-_pywrap_tensorfl.patchnew file mode 100644 index 00000000..0eed77f9 --- /dev/null+++ b/meta-arago-extras/recipes-framework/tensorflow-lite/tensorflow-lite/0001-lite-Update-CMakeLists-for-building-_pywrap_tensorfl.patch@@ -0,0 +1,37 @@ +From 64b8e731309f1ee9ad41bd65d7661854d5c3106c Mon Sep 17 00:00:00 2001 +From: Chirag Shilwant <[email protected]> +Date: Sun, 19 Jan 2025 14:54:25 +0530 +Subject: [PATCH 1/2] lite: Update CMakeLists for building + _pywrap_tensorflow_interpreter_wrapper ++- Update the CMakeLists.txt file for TensorFlow Lite to include the paths for+PYTHON, Numpy, and PYBIND11 headers. ++- This ensures that these headers are available during the build process, +which is necessary for building the _pywrap_tensorflow_interpreter_wrapper+target that relies on these dependencies. + +Upstream-Status: Inappropriate [configuration] + +Signed-off-by: Chirag Shilwant <[email protected]> +--- + tensorflow/lite/CMakeLists.txt | 3 +++ + 1 file changed, 3 insertions(+) ++diff --git a/tensorflow/lite/CMakeLists.txt b/tensorflow/lite/CMakeLists.txt+index a61f537365a..f41e57d9155 100644 +--- a/tensorflow/lite/CMakeLists.txt ++++ b/tensorflow/lite/CMakeLists.txt+@@ -727,6 +727,9 @@ set_target_properties(_pywrap_tensorflow_interpreter_wrapper PROPERTIES PREFIX "+ target_include_directories(_pywrap_tensorflow_interpreter_wrapper + PUBLIC + ${TENSORFLOW_SOURCE_DIR} ++ ${PYTHON_TARGET_INCLUDE} ++ ${NUMPY_TARGET_INCLUDE} ++ ${PYBIND11_TARGET_INCLUDE} + ) + + target_link_libraries(_pywrap_tensorflow_interpreter_wrapper +-- +2.34.1 +diff --git a/meta-arago-extras/recipes-framework/tensorflow-lite/tensorflow-lite/0002-lite-Update-CMakeLists-for-building-shared-object-li.patch b/meta-arago-extras/recipes-framework/tensorflow-lite/tensorflow-lite/0002-lite-Update-CMakeLists-for-building-shared-object-li.patchnew file mode 100644 index 00000000..a8c4f2ba --- /dev/null+++ b/meta-arago-extras/recipes-framework/tensorflow-lite/tensorflow-lite/0002-lite-Update-CMakeLists-for-building-shared-object-li.patch@@ -0,0 +1,55 @@ +From 92adea5bb011ed36c9b63bd442bf147fad96a5bb Mon Sep 17 00:00:00 2001 +From: Chirag Shilwant <[email protected]> +Date: Sun, 19 Jan 2025 15:08:59 +0530 +Subject: [PATCH 2/2] lite: Update CMakeLists for building shared object + library + +- When building TFLite with CMake, it builds static library +(i.e. libtensorflow-lite.a) by default which isn't self-contained +since all the transitive dependencies are not included [1] + +- As documented in [1], there's an update needed in CMake build steps+inorder to generate a shared object library (i.e. libtensorflow-lite.so)+ +- This patch adds schema_conversion_utils.cc to TFLite source files+& updates CMakeLists.txt to build TFLite as a shared library. Additionally,+it ensures that the properties are set correctly to manage the shared +object version. + +[1]: https://ai.google.dev/edge/litert/build/cmake#step_5_build_litert + +Upstream-Status: Inappropriate [enable feature] + +Signed-off-by: Chirag Shilwant <[email protected]> +--- + tensorflow/lite/CMakeLists.txt | 6 +++++- + 1 file changed, 5 insertions(+), 1 deletion(-) ++diff --git a/tensorflow/lite/CMakeLists.txt b/tensorflow/lite/CMakeLists.txt+index f41e57d9155..923c673d623 100644 +--- a/tensorflow/lite/CMakeLists.txt ++++ b/tensorflow/lite/CMakeLists.txt +@@ -622,8 +622,9 @@ set(_ALL_TFLITE_SRCS + ${TFLITE_SOURCE_DIR}/schema/conversion_metadata_generated.h + ${TFLITE_SOURCE_DIR}/schema/schema_utils.cc + ${TFLITE_SOURCE_DIR}/schema/schema_generated.h ++ ${TFLITE_SOURCE_DIR}/schema/schema_conversion_utils.cc + ) +-add_library(tensorflow-lite ++add_library(tensorflow-lite SHARED EXCLUDE_FROM_ALL + ${_ALL_TFLITE_SRCS} + ) + set(_ALL_TFLITE_HDRS ${_ALL_TFLITE_SRCS}) +@@ -700,6 +701,9 @@ if(TFLITE_ENABLE_INSTALL) + ) + endif() + ++# Manage version for shared object (dynamic) library++set_target_properties(tensorflow-lite PROPERTIES VERSION "${TFLITE_VERSION_MAJOR}")++ + # The kernel tests. + if(TFLITE_KERNEL_TEST) + enable_testing() +-- +2.34.1 +diff --git a/meta-arago-extras/recipes-framework/tensorflow-lite/tensorflow-lite/0003-CMakeLists-Fix-build-issues-observed-in-TFlite-v2.16.patch b/meta-arago-extras/recipes-framework/tensorflow-lite/tensorflow-lite/0003-CMakeLists-Fix-build-issues-observed-in-TFlite-v2.16.patchnew file mode 100644 index 00000000..d38a3cdd --- /dev/null+++ b/meta-arago-extras/recipes-framework/tensorflow-lite/tensorflow-lite/0003-CMakeLists-Fix-build-issues-observed-in-TFlite-v2.16.patch@@ -0,0 +1,36 @@ +From 1b93fb907b3436edf2b64db94f74b1db724943ae Mon Sep 17 00:00:00 2001 +From: Chirag Shilwant <[email protected]> +Date: Mon, 20 Jan 2025 19:51:20 +0530 +Subject: [PATCH] CMakeLists: Fix build issues observed in TFlite v2.16 + +- Improve regular expression for filtering neon and sse +related sources. The improved expression avoids missing files +in case the absolute path contains the terms neon or sse. + +- Fixes [1] + +[1]: https://github.com/tensorflow/tensorflow/issues/70730 + +Upstream-Status: Backport from master + +Signed-off-by: Chirag Shilwant <[email protected]> +--- + tensorflow/lite/CMakeLists.txt | 2 +- + 1 file changed, 1 insertion(+), 1 deletion(-) ++diff --git a/tensorflow/lite/CMakeLists.txt b/tensorflow/lite/CMakeLists.txt+index 923c673d623..3e2fb501bde 100644 +--- a/tensorflow/lite/CMakeLists.txt ++++ b/tensorflow/lite/CMakeLists.txt+@@ -554,7 +554,7 @@ populate_tflite_source_vars("kernels/internal/reference/sparse_ops"+ ) + populate_tflite_source_vars("kernels/internal/optimized/4bit" + TFLITE_KERNEL_INTERNAL_OPT_4BIT_SRCS +- FILTER "(.*neon.*|.*sse.*)\\.(cc|h)" ++ FILTER "(.*neon_.*|.*sse_.*)\\.(cc|h)" + ) + set(TFLITE_PROFILER_SRCS + ${TFLITE_SOURCE_DIR}/profiling/platform_profiler.cc +-- +2.34.1 +diff --git a/meta-arago-extras/recipes-framework/tensorflow-lite/tensorflow-lite_2.16.bb b/meta-arago-extras/recipes-framework/tensorflow-lite/tensorflow-lite_2.16.bbnew file mode 100644 index 00000000..ca51d6dc --- /dev/null+++ b/meta-arago-extras/recipes-framework/tensorflow-lite/tensorflow-lite_2.16.bb@@ -0,0 +1,203 @@+DESCRIPTION = "TensorFlow Lite is an open source deep learning framework for \+on-device inference." +AUTHOR = "Google Inc. and Yuan Tang"+SUMMARY = "TensorFlow Lite C++ Library, Python interpreter & Benchmark Model"+HOMEPAGE = "https://www.tensorflow.org/lite"; + +LICENSE = "Apache-2.0"+LIC_FILES_CHKSUM = "file://LICENSE;md5=4158a261ca7f2525513e31ba9c50ae98"+ +PV = "2.16.2" +Is this not automatically set now from the file name version?
Ahh, yes! Will rename the recipe name to 2.16.2 in my v2 patch
+# Get major of the PV variable +MAJOR = "${@d.getVar('PV').split('.')[0]}" + +# Matches v${PV} tag +SRCREV = "810f233968cec850915324948bbbc338c97cf57f" + +SRC_URI = " \ + git://github.com/tensorflow/tensorflow;protocol=https;branch=r2.16 \+ file://0001-lite-Update-CMakeLists-for-building-_pywrap_tensorfl.patch \ + file://0002-lite-Update-CMakeLists-for-building-shared-object-li.patch \ + file://0003-CMakeLists-Fix-build-issues-observed-in-TFlite-v2.16.patch \+" + +S = "${WORKDIR}/git" + +inherit setuptools3 cmake + +DEPENDS = " \ + zlib \ + python3-numpy-native \ + python3-pybind11-native \ + python3-wheel-native \ + python3-pybind11 \ + python3-numpy \ + swig-native \ + python3 \ + gzip-native \ +" + +# Set building environment variables +TENSORFLOW_TARGET="linux" +TENSORFLOW_TARGET_ARCH:aarch64="aarch64" + +OECMAKE_SOURCEPATH = "${S}/tensorflow/lite" + +# The -O3 flag enables high-level optimizations for performance +# and the -DNDEBUG flag disables debugging code, such as assertions +# to further optimize the build for production use. +# Hence, activate -O3 optimization and disable debug symbols. +OECMAKE_C_FLAGS_RELEASE = "-O3 -DNDEBUG" +OECMAKE_CXX_FLAGS_RELEASE = "-O3 -DNDEBUG" + +OECMAKE_C_FLAGS_RELEASE:append:arm = " -march=armv7-a -mfpu=neon" +OECMAKE_CXX_FLAGS_RELEASE:append:arm = " -march=armv7-a -mfpu=neon" ++# Build tensorflow-lite.so library, _pywrap_tensorflow_interpreter_wrapper.so library and the benchmark_model application +OECMAKE_TARGET_COMPILE = "tensorflow-lite _pywrap_tensorflow_interpreter_wrapper benchmark_model"+ +do_generate_toolchain_file:append:arm() {+ # XNNPACK does not recognize this when CMAKE_SYSTEM_PROCESSOR is arm.+ # Instead, you need to change it to armv7.If ARMv7 is not supported with XNNPACK, are these lines still needed?
Thanks for catching, will drop the irrelevant XNNPACK comment But yeah, the logic of setting CMAKE_SYSTEM_PROCESSOR to armv7 is still needed though!
+ sed -i 's:CMAKE_SYSTEM_PROCESSOR arm:CMAKE_SYSTEM_PROCESSOR armv7:g' ${WORKDIR}/toolchain.cmake+} + +EXTRA_OECMAKE:append = " \ + -DFETCHCONTENT_FULLY_DISCONNECTED=OFF \ + -DTFLITE_ENABLE_RUY=ON \ + -DPYTHON_TARGET_INCLUDE=${RECIPE_SYSROOT}${includedir}/${PYTHON_DIR} \+ -DNUMPY_TARGET_INCLUDE=${RECIPE_SYSROOT}${PYTHON_SITEPACKAGES_DIR}/numpy/core/include \ + -DPYBIND11_TARGET_INCLUDE=${RECIPE_SYSROOT}${PYTHON_SITEPACKAGES_DIR}/pybind11/include \+ -DCMAKE_SYSTEM_NAME=${TENSORFLOW_TARGET} \ + -DTFLITE_VERSION_MAJOR=${MAJOR} \ +" + +EXTRA_OECMAKE:append:aarch64 = " \ + -DTFLITE_ENABLE_XNNPACK=ON \ + -DCMAKE_SYSTEM_PROCESSOR=${TENSORFLOW_TARGET_ARCH} \ +" ++# XNNPACK is not supported in Linux armv7-a [https://github.com/tensorflow/tensorflow/issues/64358]+# Hence, turning it off +EXTRA_OECMAKE:append:arm = " \ + -DTFLITE_ENABLE_XNNPACK=OFF \ +" + +do_configure[network] = "1" + +do_configure:prepend() { + if [ -n "${http_proxy}" ]; then + export HTTP_PROXY=${http_proxy} + export http_proxy=${http_proxy} + fi + + if [ -n "${https_proxy}" ]; then + export HTTPS_PROXY=${https_proxy} + export https_proxy=${https_proxy} + fiAre these checks needed?
Yeah, I was just being more cautious. Will drop the "if checks" in v2
If it is unset then exporting doesn't hurt, could replace all this with: export http_proxy export https_proxy do_configure[network] = "1"++ # There is no Fortran compiler in the toolchain, but bitbake sets this variable anyway with unavailable + # binary & it leads to "CMake Error: Could not find compiler set in environment variable FC:"+ unset FC +} + +do_compile[network] = "1" + +do_compile:prepend() { + # Used to download cmake dependencies when behind a proxy + if [ -n "${http_proxy}" ]; then + export HTTP_PROXY=${http_proxy} + export http_proxy=${http_proxy} + fi + + if [ -n "${https_proxy}" ]; then + export HTTPS_PROXY=${https_proxy} + export https_proxy=${https_proxy} + fiDo you need network during compile? Seems to only be used in configure.
yeah, thanks for catching! apparently it's not needed. So, it was needed earlier when I was trying to build the python wheel package directly using build_pip_package_with_cmake.sh provided by tensorflow But later I decided to revamp the logic & switched to setuptools3 afterwhich we don't neednetwork access in do_compile & even the proxy setting stuff!
So, it's only the do_configure stage where we need network access for tensorflow-lite Will handle this in v2 patch. -- Chirag
Andrew+} + +SETUPTOOLS_SETUP_PATH = "${B}" + +do_compile:append() {+ # Build the python wheel (procedure extract form the build_pip_package_with_cmake.sh)+ BUILD_DIR=${B} + TENSORFLOW_DIR=${S} + TENSORFLOW_LITE_DIR="${TENSORFLOW_DIR}/tensorflow/lite"+ TENSORFLOW_VERSION=$(grep "_VERSION = " "${TENSORFLOW_DIR}/tensorflow/tools/pip_package/setup.py" | cut -d= -f2 | sed "s/[ '-]//g")+ mkdir -p "${B}/tflite_runtime" + cp -r "${TENSORFLOW_LITE_DIR}/tools/pip_package/debian" \ + "${TENSORFLOW_LITE_DIR}/tools/pip_package/MANIFEST.in" \ + "${B}" + cp -r "${TENSORFLOW_LITE_DIR}/python/interpreter_wrapper" "${B}"+ cp "${TENSORFLOW_LITE_DIR}/tools/pip_package/setup_with_binary.py" "${B}/setup.py"+ cp "${TENSORFLOW_LITE_DIR}/python/interpreter.py" \ + "${TENSORFLOW_LITE_DIR}/python/metrics/metrics_interface.py" \ + "${TENSORFLOW_LITE_DIR}/python/metrics/metrics_portable.py" \ + "${B}/tflite_runtime"+ echo "__version__ = '${TENSORFLOW_VERSION}'" >> "${BUILD_DIR}/tflite_runtime/__init__.py" + echo "__git_version__ = '$(git -C "${TENSORFLOW_DIR}" describe)'" >> "${BUILD_DIR}/tflite_runtime/__init__.py"+ + export PACKAGE_VERSION="${TENSORFLOW_VERSION}" + export PROJECT_NAME="tflite_runtime"+ cp "${B}/_pywrap_tensorflow_interpreter_wrapper.so" "tflite_runtime"+ + setuptools3_do_compile +} + +do_install() { + # Install tensorflow-lite dynamic library + install -d ${D}${libdir}+ install -m 0644 ${B}/libtensorflow-lite.so.${MAJOR} ${D}${libdir}/libtensorflow-lite.so.${PV}++ ln -sf libtensorflow-lite.so.${PV} ${D}${libdir}/libtensorflow-lite.so.${MAJOR} + ln -sf libtensorflow-lite.so.${PV} ${D}${libdir}/libtensorflow-lite.so++ # armnn expects the the dynamic library of tensorflow lite to be named as "libtensorflowlite.so" + # Eg - https://github.com/ARM-software/armnn/blob/branches/armnn_24_11/delegate/cmake/Modules/FindTfLite.cmake#L19+ # Hence, create a symbolic link to handle the same + ln -snf libtensorflow-lite.so ${D}${libdir}/libtensorflowlite.so + + # Install benchmark_model + install -d ${D}/opt/${PN}/tools/+ install -m 0755 ${B}/tools/benchmark/benchmark_model ${D}/opt/${PN}/tools/+ # chrpath -r '$ORIGIN' ${D}/opt/${PN}/tools/benchmark_model + + # Install header files + install -d ${D}${includedir}/tensorflow/lite + cd ${S}/tensorflow/lite+ cp --parents $(find . -name "*.h*" -not -path "*cmake_build/*") ${D}${includedir}/tensorflow/lite+ cd ${B}+ cp --parents $(find . -name "*.h*") ${D}${includedir}/tensorflow/lite+ + # Install Tflite python3 interpreter + install -d ${D}${PYTHON_SITEPACKAGES_DIR}/tflite_runtime + install -d ${D}${PYTHON_SITEPACKAGES_DIR}/tflite_runtime.egg-info+ install -m 0755 ${B}/tflite_runtime/* ${D}${PYTHON_SITEPACKAGES_DIR}/tflite_runtime + install -m 0755 ${B}/tflite_runtime.egg-info/* ${D}${PYTHON_SITEPACKAGES_DIR}/tflite_runtime.egg-info + # chrpath -r '$ORIGIN' ${D}${PYTHON_SITEPACKAGES_DIR}/tflite_runtime/*.so+} + +PACKAGES += "${PN}-tools python3-${PN}" + +RDEPENDS:${PN} += " \ + ${PN}-tools \ + python3-${PN} \ +" + +FILES:${PN} = " \ + ${libdir}/lib*.so* \ +" + +FILES:${PN}-tools = " \ + /opt/* \ +" + +FILES:python3-${PN} = " \ + ${PYTHON_SITEPACKAGES_DIR}/tflite_runtime \ + ${PYTHON_SITEPACKAGES_DIR}/tflite_runtime.egg-info \ +" + +RDEPENDS:python3-${PN} += " python3-ctypes python3-numpy " + +PROVIDES += "python3-${PN}"
-=-=-=-=-=-=-=-=-=-=-=- Links: You receive all messages sent to this group. View/Reply Online (#15754): https://lists.yoctoproject.org/g/meta-arago/message/15754 Mute This Topic: https://lists.yoctoproject.org/mt/110775616/21656 Group Owner: [email protected] Unsubscribe: https://lists.yoctoproject.org/g/meta-arago/unsub [[email protected]] -=-=-=-=-=-=-=-=-=-=-=-
