Enable TFLite model parsing with FlatBuffer support and comprehensive TFLite enhancements by tdarote · Pull Request #146 · qualcomm-linux/meta-qcom-distro

tdarote · 2026-01-30T10:00:37Z

This pull request enables TFLite model parsing capabilities by integrating FlatBuffer support and implements comprehensive enhancements to the TFLite recipe.

Key Changes:
-> FlatBuffer Integration for TFLite:
- Added flatbuffer bbappend file to enable TFLite's schema handling capabilities for model parsing

-> Comprehensive TFLite Enhancements:
- Added benchmark_model config option
- Fixed protobuf dependency in benchmark tools
- Added dynamic OpenCL library loading support
- Excluded subdirectories from all builds
- Force delegate symbols from shared library
- Add version support to C API
- Fix label_image dependencies
- Add install rule for C interface shared library

recipes-ml/tflite/tensorflow-lite_2.20.0.qcom.bb

recipes-ml/tflite/files/0001-lite-Add-config-option-to-enable-benchmark_model.patch

recipes-ml/tflite/files/0002-cmake-lite-tools-benchmark-require-protobuf-through-.patch

recipes-ml/tflite/files/0003-feat-tflite-Add-dynamic-OpenCL-library-loading-suppo.patch

recipes-ml/tflite/files/0004-cmake-Exclude-subdirectories-from-all-builds.patch

lumag · 2026-01-30T12:17:39Z

Waiting for the patches to be posted upstream

recipes-ml/tflite/files/0005-c-Force-delegate-symbols-from-shared-library.patch

lumag

Please fix commit subject for the flatbuffers patch

recipes-ml/tflite/files/0002-cmake-lite-tools-benchmark-require-protobuf-through-.patch

tdarote · 2026-02-02T11:04:25Z

Waiting for the patches to be posted upstream

-updated upstream status to submitted

recipes-ml/tflite/files/0006-c-Add-version-support-to-C-API.patch

recipes-ml/tflite/files/0007-cmake-Fix-label-image-dependencies.patch

recipes-ml/tflite/files/0008-cmake-Add-install-rule-for-c-interface-shared-librar.patch

recipes-ml/tflite/files/0002-cmake-lite-tools-benchmark-require-protobuf-through-.patch

tdarote · 2026-02-02T12:53:27Z

Please fix commit subject for the flatbuffers patch

DONE

lumag · 2026-02-02T13:02:22Z

Okay. Upstream (meta-oe) uses flatbuffers 25.12.19. To prevent possible issues with other packages which might depend on that version, we need to provide a separate version of the recipe (flatbuffers-tflite.bb, require recipes-devtools/flatbuffers/flatbuffers.bb) and use it for building TFLite.

tdarote · 2026-02-02T13:39:34Z

Okay. Upstream (meta-oe) uses flatbuffers 25.12.19. To prevent possible issues with other packages which might depend on that version, we need to provide a separate version of the recipe (flatbuffers-tflite.bb, require recipes-devtools/flatbuffers/flatbuffers.bb) and use it for building TFLite.

--done

lumag · 2026-02-02T18:43:18Z

This doesn't build in my test system:

| DEBUG: Python function extend_recipe_sysroot finished
| DEBUG: Executing shell function do_configure
| CMake Error at CMakeLists.txt:23 (project):
|   VERSION ".." format invalid.
|
|
| -- Configuring incomplete, errors occurred!
| WARNING: exit code 1 from a shell command.
ERROR: Task (/home/lumag/Projects/RPB/build-rpb/conf/../../layers/meta-qcom-distro/recipes-ml/tflite/tensorflow-lite_2.20.0.qcom.bb:do_configure) failed with exit code '1'
NOTE: Tasks Summary: Attempted 5751 tasks of which 5578 didn't need to be rerun and 1 failed.

recipes-support/flatbuffers/flatbuffers-tflite.inc

ricardosalveti · 2026-02-03T01:46:04Z

Please explain why this flatbuffer recipe is needed, why the one from meta-oe is not enough, differences, etc, as part of your git commit message.

tdarote · 2026-02-03T06:51:49Z

Please explain why this flatbuffer recipe is needed, why the one from meta-oe is not enough, differences, etc, as part of your git commit message.

If we are using flatbuffers version from meta-oe then we are facing below config issue as :

| /local/mnt/workspace/tushar/kas-tflite-build/build/tmp/work/armv8-2a-qcom-linux/tensorflow-lite/2.20.0.qcom/sources/tensorflow-lite-2.20.0.qcom/tensorflow/compiler/mlir/lite/schema/schema_generated.h:25:41: error: static assertion failed: Non-compatible flatbuffers version included
| 25 | static_assert(FLATBUFFERS_VERSION_MAJOR == 24 &&
| | ^
| /local/mnt/workspace/tushar/kas-tflite-build/build/tmp/work/armv8-2a-qcom-linux/tensorflow-lite/2.20.0.qcom/sources/tensorflow-lite-2.20.0.qcom/tensorflow/compiler/mlir/lite/schema/schema_generated.h:25:41: note: the comparison reduces to '(25 == 24)'
| ninja: build stopped: subcommand failed.
|
| WARNING: /local/mnt/workspace/tushar/kas-tflite-build/build/tmp/work/armv8-2a-qcom-linux/tensorflow-lite/2.20.0.qcom/temp/run.do_compile.3752035:153 exit 1 from 'eval ${DESTDIR:+DESTDIR=${DESTDIR} }VERBOSE=1 cmake --build '/local/mnt/workspace/tushar/kas-tflite-build/build/tmp/work/armv8-2a-qcom-linux/tensorflow-lite/2.20.0.qcom/build' "$@" -- ${EXTRA_OECMAKE_BUILD}'

ricardosalveti · 2026-02-12T20:20:10Z

Didn't really build for me, do_configure is still trying to reach network:


|           --- LOG END ---
|           error: downloading 'https://github.com/Maratyszcza/FP16/archive/0a92994d729ff76a58f692d3028ca1b64b145d91.zip' failed
|           status_code: 6
|           status_string: "Could not resolve hostname"
|           log:
|           --- LOG BEGIN ---
|           timeout on name lookup is not supported
|
|   getaddrinfo(3) failed for github.com:443
|
|   Store negative name resolve for github.com:443
|
|   Could not resolve host: github.com
|
|   closing connection #0
|
|
|
|           --- LOG END ---
|
|
|
|
| ninja: build stopped: subcommand failed.
| CMake Warning at /home/rsalveti/build/qualcomm-linux/build/tmp/work/armv8-2a-qcom-linux/tensorflow-lite/2.20.0/recipe-sysroot-native/usr/share/cmake-4.2/Modules/FetchContent.cmake:2121 (message):
...
| WARNING: Backtrace (BB generated script):
|       #1: cmake_do_configure, /home/rsalveti/build/qualcomm-linux/build/tmp/work/armv8-2a-qcom-linux/tensorflow-lite/2.20.0/temp/run.do_configure.2827760, line 173
|       #2: do_configure, /home/rsalveti/build/qualcomm-linux/build/tmp/work/armv8-2a-qcom-linux/tensorflow-lite/2.20.0/temp/run.do_configure.2827760, line 152
|       #3: main, /home/rsalveti/build/qualcomm-linux/build/tmp/work/armv8-2a-qcom-linux/tensorflow-lite/2.20.0/temp/run.do_configure.2827760, line 191
ERROR: Task (/home/rsalveti/build/qualcomm-linux/build/../meta-qcom-distro/recipes-ml/tflite/tensorflow-lite_2.20.0.bb:do_configure) failed with exit code '1'

tdarote · 2026-02-14T06:20:57Z

Didn't really build for me, do_configure is still trying to reach network:


|           --- LOG END ---
|           error: downloading 'https://github.com/Maratyszcza/FP16/archive/0a92994d729ff76a58f692d3028ca1b64b145d91.zip' failed
|           status_code: 6
|           status_string: "Could not resolve hostname"
|           log:
|           --- LOG BEGIN ---
|           timeout on name lookup is not supported
|
|   getaddrinfo(3) failed for github.com:443
|
|   Store negative name resolve for github.com:443
|
|   Could not resolve host: github.com
|
|   closing connection #0
|
|
|
|           --- LOG END ---
|
|
|
|
| ninja: build stopped: subcommand failed.
| CMake Warning at /home/rsalveti/build/qualcomm-linux/build/tmp/work/armv8-2a-qcom-linux/tensorflow-lite/2.20.0/recipe-sysroot-native/usr/share/cmake-4.2/Modules/FetchContent.cmake:2121 (message):
...
| WARNING: Backtrace (BB generated script):
|       #1: cmake_do_configure, /home/rsalveti/build/qualcomm-linux/build/tmp/work/armv8-2a-qcom-linux/tensorflow-lite/2.20.0/temp/run.do_configure.2827760, line 173
|       #2: do_configure, /home/rsalveti/build/qualcomm-linux/build/tmp/work/armv8-2a-qcom-linux/tensorflow-lite/2.20.0/temp/run.do_configure.2827760, line 152
|       #3: main, /home/rsalveti/build/qualcomm-linux/build/tmp/work/armv8-2a-qcom-linux/tensorflow-lite/2.20.0/temp/run.do_configure.2827760, line 191
ERROR: Task (/home/rsalveti/build/qualcomm-linux/build/../meta-qcom-distro/recipes-ml/tflite/tensorflow-lite_2.20.0.bb:do_configure) failed with exit code '1'

Hi @ricardosalveti,

Could you please share the detailed setup steps and your testing procedure? We are not able to reproduce the fetcher issue on our side. From the do_configure logs, we can see that CMake is still performing a download during configuration.

We also tried with BB_NO_NETWORK = "1", but the download still occurs at configure time. It appears that CMake is attempting to fetch a ZIP file, and we are unable to prevent this network access.

Right now, we are working on avoiding this download based on the references found in the do_configure logs. However, we are running into issues because the ZIP path used by CMake is not a proper source location. Even if we add the GitHub source link in SRC_URI, the configure step still tries to download files.

Could you please clarify:

How you are testing this scenario?
What additional steps or setups are you using that trigger this issue?
How you are preventing the configure-time download on your environment?

Your inputs will help us reproduce and address the problem more effectively.

tdarote · 2026-02-14T14:52:03Z

With the latest patch, we were able to avoid the source download during the configuration stage by using the details from the do_configure logs.

recipes-ml/tflite/tensorflow-lite_2.20.0.bb

anujm1 · 2026-02-16T01:17:39Z

recipes-ml/tflite/tensorflow-lite_2.20.0.bb

+OECMAKE_TARGET_COMPILE += "benchmark_model label_image"
+
+EXTRA_OECMAKE += " \
+    -DFETCHCONTENT_FULLY_DISCONNECTED=ON \


This is set by cmake.bbclass: https://git.openembedded.org/openembedded-core/tree/meta/classes-recipe/cmake.bbclass#n258

Thanks, noted. Since cmake.bbclass already sets this flag, I’ll remove it from the recipe

lumag · 2026-02-16T11:53:07Z

recipes-ml/tflite/tensorflow-lite_2.20.0.bb

+# NOTE: Dependencies are managed manually
+# To update dependencies:
+# 1. For each repo, run: git ls-remote <url> HEAD
+# 2. Get the latest commit hash


This comment is still misleading and far from being true.

is this okay to update ?
--># 2. Read dependency revisions from TFLite's Bazel configuration files

And what would that mean? If you don't want to implement a script, provide the steps for the developer to follow.

updated the steps for updating dependency hash values, can you please review once.

koenkooi · 2026-02-17T08:29:12Z

recipes-ml/tflite/tensorflow-lite_2.20.0.bb

+# Source archive checksum for the main TensorFlow fetch
+SRC_URI[sha256sum] = "cfc7749b96f63bd31c3c42b5c471bf756814053e847c10f3eb003417bc523d30"
+
+REQUIRED_DISTRO_FEATURES += "opengl"


Nit pick: don't put anything else in between SRC_URI entries, depends go on top, just below SUMMARY/LICENSE/etc , since they could include a custom fetcher.

We try to order variables 'chronologically', which some freedom to make it more understandable. For example, REQUIRED_DISTRO_FEATURES should go near the top, since it's one of the first checks, but for readability it makes more sense to group it with the either the DEPENDS that need the check or the PACKAGE_CONFIG option that has the guarded options.

updated order, can you check once ?

lumag

Implementing a script that provides SRC_URI fragment and corresponding SRCREV would be a much preferred approach.

lumag · 2026-02-17T11:20:10Z

recipes-ml/tflite/tensorflow-lite_2.20.0.bb

+#
+# 1) In the TensorFlow repo at the exact branch/tag you build (e.g. v2.20.0),
+#    read the pinned dependency revisions in:
+#       - WORKSPACE


I don't see any pins here.

please refer : https://github.com/tensorflow/tensorflow/blob/v2.20.0/tensorflow/workspace2.bzl

Okay. I really want a script to extract all that info.

provided script and its uses as comments in tflite recipes, can you please check once

lumag · 2026-02-17T11:25:08Z

recipes-ml/tflite/tensorflow-lite_2.20.0.bb

+SRCREV_fp16 = "4dfe081cf6bcd15db339cf2680b9281b8451eeb3"
+SRCREV_kleidiai = "dc69e899945c412a8ce39ccafd25139f743c60b1"
+SRCREV_pthreadpool = "c2ba5c50bb58d1397b693740cf75fad836a0d1bf"
+SRCREV_fxdiv = "63058eff77e11aa15bf531df5dd34395ec3017c8"


Please sort the entries

We maintain the dependency sequence consistent with how TFLite fetches them upstream, please confirm still you need in sorted order ?

sorted as per alphanumeric, can you check once

lumag · 2026-02-17T11:25:47Z

recipes-ml/tflite/tensorflow-lite_2.20.0.bb

+#
+# 3) For each third‑party dependency used by TFLite (xnnpack, cpuinfo,
+#    pthreadpool, ruy, fp16, fxdiv, gemmlowp, farmhash, CL/Vulkan headers, etc.),
+#    copy the commit hash (and sha256 if present) exactly as declared upstream,


Which sha256? How to identify deps which are actually used by TFLite?

can you please review once, updated in recipe comments

…nd C API Apply multiple patches to enhance TensorFlow Lite functionality: - Add benchmark_model config option - Fix protobuf dependency in benchmark tools - Add dynamic OpenCL library loading support - Exclude subdirectories from all builds - Force delegate symbols from shared library - Add version support to C API - Fix label_image dependencies - Add install rule for C interface shared library Signed-off-by: Tushar Darote <tdarote@qti.qualcomm.com>

lumag · 2026-02-17T14:59:49Z

recipes-ml/tflite/tensorflow-lite_2.20.0.bb

+#
+# How to update these SRCREV_* values:
+#
+# 1) Use the automated script 'extract_tflite_srcrevs_from_github.py' to generate


What prevents you from integrating it as a task? I pointed you to vulkan-cts several times.

The reason we cannot automatically integrate this as a task is due to the nature of dependency management in TensorFlow Lite builds. While many dependencies are pinned to specific commits from the TensorFlow repository, some dependencies like fft2d come from external repositories (Android's external/fft2d) and may not follow the same pinning strategy.

Specifically:

TensorFlow-pinned dependencies (farmhash, gemmlowp, cpuinfo, etc.) can be reliably updated using the automated script since they're all sourced from TensorFlow's workspace files
External dependencies like fft2d come from Android's external repository and may not be pinned to the same commit that TensorFlow expects
Testing requirement: We need to validate that the specific commit works correctly with our target platform and build configurations
Failure risk: If we automatically update all dependencies, we risk introducing breaking changes that wouldn't be caught until later testing phases
We currently test fft2d with its main branch tip and validate that it works correctly. This manual approach ensures stability while still allowing us to benefit from the automated script for the majority of dependencies that are properly pinned by TensorFlow.

The script is designed to be used as a maintenance tool that developers can run when updating to new TensorFlow versions, but the final validation and selection of the exact commits is done manually to ensure compatibility and stability.

External dependencies like fft2d come from Android's external repository and may not be pinned to the same commit that TensorFlow expects

How comes?

URL https://storage.googleapis.com/mirror.tensorflow.org/github.com/petewarden/OouraFFT/archive/v1.0.tar.gz # Sync with tensorflow/workspace2.bzl URL_HASH SHA256=5f4dabc2ae21e1f537425d58a49cdca1c49ea11db0d6271e2a4b27e9697548eb

I think this exactly defines the location and the way to fetch fft2d.

I converted to use git protocol instead of giving tar since I was facing below issue with tar download as :

ERROR: tensorflow-lite-2.20.0-r0 do_recipe_qa: QA Issue: tensorflow-lite: SRC_URI uses unstable GitHub/GitLab archives, convert recipe to use git protocol [src-uri-bad]
ERROR: tensorflow-lite-2.20.0-r0 do_recipe_qa: Fatal QA errors were found, failing task.
ERROR: Logfile of failure stored in: /local/mnt/workspace/tushar/5-2-26/build/tmp/work/armv8-2a-qcom-linux/tensorflow-lite/2.20.0/temp/log.do_recipe_qa.2045484
ERROR: Task (/local/mnt/workspace/tushar/5-2-26/build/../meta-qcom-distro/recipes-ml/tflite/tensorflow-lite_2.20.0.bb:do_recipe_qa) failed with exit code '1'

tdarote requested review from lumag, ndechesne, ricardosalveti, sbanerjee-quic and vkraleti as code owners January 30, 2026 10:00

lumag requested changes Jan 30, 2026

View reviewed changes

tdarote force-pushed the tflite branch from 051230b to 86b8a3d Compare January 30, 2026 11:29

tdarote force-pushed the tflite branch from 86b8a3d to b8e0224 Compare February 2, 2026 08:30

tdarote commented Feb 2, 2026

View reviewed changes

recipes-ml/tflite/files/0005-c-Force-delegate-symbols-from-shared-library.patch Show resolved Hide resolved

lumag reviewed Feb 2, 2026

View reviewed changes

tdarote force-pushed the tflite branch from b8e0224 to 395cbba Compare February 2, 2026 09:01

lumag reviewed Feb 2, 2026

View reviewed changes

recipes-ml/tflite/files/0002-cmake-lite-tools-benchmark-require-protobuf-through-.patch Outdated Show resolved Hide resolved

tdarote force-pushed the tflite branch 2 times, most recently from 175b3f6 to 3ae95c6 Compare February 2, 2026 10:31

tdarote closed this Feb 2, 2026

tdarote reopened this Feb 2, 2026

tdarote commented Feb 2, 2026

View reviewed changes

tdarote requested a review from lumag February 2, 2026 13:03

tdarote force-pushed the tflite branch from 3ae95c6 to a55be7e Compare February 2, 2026 13:38

ricardosalveti reviewed Feb 3, 2026

View reviewed changes

recipes-support/flatbuffers/flatbuffers-tflite.inc Outdated Show resolved Hide resolved

tdarote force-pushed the tflite branch from a55be7e to 8dfc302 Compare February 3, 2026 06:49

tdarote force-pushed the tflite branch from 8dfc302 to 1ba8e19 Compare February 3, 2026 08:15

tdarote force-pushed the tflite branch from 69cb73a to 46bd5ac Compare February 12, 2026 11:32

koenkooi previously approved these changes Feb 12, 2026

View reviewed changes

tdarote requested a review from quaresmajose February 12, 2026 12:58

tdarote dismissed koenkooi’s stale review via 478210f February 14, 2026 14:48

tdarote force-pushed the tflite branch from 46bd5ac to 478210f Compare February 14, 2026 14:48

tdarote requested a review from koenkooi February 14, 2026 15:09

anujm1 reviewed Feb 16, 2026

View reviewed changes

recipes-ml/tflite/tensorflow-lite_2.20.0.bb Show resolved Hide resolved

anujm1 reviewed Feb 16, 2026

View reviewed changes

tdarote force-pushed the tflite branch from 478210f to e7d708f Compare February 16, 2026 05:36

tdarote requested a review from anujm1 February 16, 2026 06:20

koenkooi previously approved these changes Feb 16, 2026

View reviewed changes

anujm1 approved these changes Feb 16, 2026

View reviewed changes

lumag requested changes Feb 16, 2026

View reviewed changes

tdarote requested a review from lumag February 16, 2026 12:08

tdarote dismissed koenkooi’s stale review via 06179a7 February 17, 2026 06:04

tdarote force-pushed the tflite branch from e7d708f to 06179a7 Compare February 17, 2026 06:04

tdarote requested a review from koenkooi February 17, 2026 06:08

koenkooi requested changes Feb 17, 2026

View reviewed changes

tdarote force-pushed the tflite branch from 06179a7 to fc92845 Compare February 17, 2026 08:56

tdarote requested a review from koenkooi February 17, 2026 08:57

koenkooi previously approved these changes Feb 17, 2026

View reviewed changes

lumag requested changes Feb 17, 2026

View reviewed changes

tdarote dismissed koenkooi’s stale review via 6c2d9e7 February 17, 2026 14:55

tdarote force-pushed the tflite branch from fc92845 to 6c2d9e7 Compare February 17, 2026 14:55

lumag reviewed Feb 17, 2026

View reviewed changes

tdarote requested a review from lumag February 17, 2026 15:01

Conversation

tdarote commented Jan 30, 2026

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

lumag commented Jan 30, 2026

Uh oh!

Uh oh!

lumag left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

tdarote commented Feb 2, 2026

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

tdarote commented Feb 2, 2026

Uh oh!

lumag commented Feb 2, 2026

Uh oh!

tdarote commented Feb 2, 2026

Uh oh!

lumag commented Feb 2, 2026

Uh oh!

Uh oh!

ricardosalveti commented Feb 3, 2026

Uh oh!

tdarote commented Feb 3, 2026

Uh oh!

ricardosalveti commented Feb 12, 2026

Uh oh!

tdarote commented Feb 14, 2026

Uh oh!

tdarote commented Feb 14, 2026

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

lumag Feb 16, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

tdarote Feb 16, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

lumag left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

lumag Feb 16, 2026 •

edited

Loading

tdarote Feb 16, 2026 •

edited

Loading