tflite: add patches for GPU/OpenCL fixes and new recipes #1385

tdarote · 2026-01-13T09:48:31Z

Added patches for GPU work-group sizing, OpenCL loading, CMake fixes, and integration of multi-model tools.
Introduced pkg-config template and new recipes for flatbuffers and TensorFlow Lite 2.20.0.

lumag

Please provide a sensible commit message. It should be describing the design decisions, the issues that you faced, etc. rather than simply stating the patch contents.

More importantly, please send patches upstream before contributing them here.

lumag · 2026-01-13T10:21:58Z

recipes-ml/tflite/files/0001-fix-gpu-adjust-work-group-size-and-remove-Adreno-spe.patch

+- Improves compatibility across GPUs.
+- Prevents oversized work-groups and incorrect buffer alignment on Adreno devices.
+
+Upstream-Status: Pending


Please submit upstream.

lumag · 2026-01-13T10:24:18Z

recipes-ml/tflite/files/0007-tensorflow-lite-Major-version-dlopen-for-OpenCL-libs.patch

+Date: Thu, 9 Oct 2025 22:42:17 +0200
+Subject: [PATCH 07/11] tensorflow-lite: Major version dlopen for OpenCL libs
+
+Upstream-Status: Inappropriate [OE specific]


Missing commit message, it's impossible to judge whether it's really OE-specific or not

Thanks for the feedback. I’ll update the commit message to include the rationale. The patch makes TensorFlow Lite use dlopen with major versioned OpenCL libraries (e.g., libOpenCL.so.1) instead of unversioned names. This is required in OE because the unversioned symlink (libOpenCL.so) is often in -dev packages and not present on target images, causing runtime failures. Upstream doesn’t enforce this because they assume full development environments, so this is OE-specific.

On my Debian system I also don't have libOpenCL.so, unless I install the dev package. Please work with upstream in order to implement a generic fix for the issue.

lumag · 2026-01-13T10:25:56Z

recipes-ml/tflite/flatbuffers-native_24.3.25.bb

@@ -0,0 +1,29 @@
+SUMMARY = "Memory Efficient Serialization Library"


Why do we need a separate -native recipe?

We need flatbuffers-native because TFLite’s build requires the flatc tool at build time, and it must match FlatBuffers v24.3.25 to avoid ABI/schema mismatches. This version isn’t available on the build host, and we didn’t find a native provider in our current layers. Providing a native recipe ensures deterministic builds and the exact tool version required by TFLite 2.20.0.

It should be better to carry a full downgrade version of the recipe in meta-oe/recipes-devtools/flatbuffers and select this one using with the PREFERRED_VERSION

Why do we need separate -native recipe? Can't we use BBCLASSEXTEND?

I agree with @quaresmajose , the separate version should be provided in meta-oe.

We can handle this using the bbappend file in the flatbuffer layer we shared. I’ll include the update in the next patch.

I think you missed the point. The flatbuffers recipe is a part of meta-oe layer. Please keep it there.

lumag · 2026-01-13T10:29:24Z

recipes-ml/tflite/flatbuffers.bbappend

@@ -0,0 +1,4 @@
+# Tensorflow-lite needs an extremely specific version, so lock it to that


The base path is different

lumag · 2026-01-13T10:29:49Z

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

@@ -0,0 +1,136 @@
+inherit cmake pkgconfig


Why are we packaging it instead of using meta-tensorflow?

The bulk of meta-tensorflow is openjdk and bazel support, but we can't use bazel for 2.20, so we opt for the cmake option. That means we aren't reusing anything from meta-tensorflow anymore.

This should be mentioned in the PR description at the minimum.

We evaluated meta-tensorflow, which currently builds TensorFlow 2.19.0 with full C++ APIs and additional components. Our requirement is to upgrade to TensorFlow Lite 2.20.0 and build only the C APIs for a minimal footprint. The full TensorFlow build in meta-tensorflow introduces unnecessary dependencies, increases image size, and build time, which is not acceptable for our target.
Therefore, we introduced a dedicated TFLite recipe that:

Pins to 2.20.0 (latest stable for our BSP).
Compiles only the C API (no C++ interpreter, Python bindings, or extra tooling).
Applies OE-specific adjustments (e.g., OpenCL major-version dlopen for runtime compatibility).
This approach ensures a lean build optimized for embedded targets while meeting version and feature requirements.

So, please work with meta-tensorflow maintainers to update tflite (the layer also provides tflite) to 2.20.
I'd really prefer to avoid fragmentation here: there is an established layer which provides TF and TF Lite.

Before we decide to integrate entire recipes which are known to also be available as part of other common layers, we should at least have a proper discussion with upstream, and only bring these recipes here when really required (something specific to our own BSP).

So please work first with upstream, see if it is possible to update the revision there and if they would also accept making the recipe more flexible in order for us to later decide how to build it (e.g. based on pkgconfig values) as part of meta-qcom-distro.

latest recipe for tflite is 2.19 https://layers.openembedded.org/layerindex/recipe/396204/
there are 20 patches which do not affect our case
on top of that basel is used instead of cmake, where it is clearly stated by tflite repo owners that cmake must be used for cross compilation
none of the timing performance optimizations are there
so bbappend whould here to fix all those challenges or to try submitting everything to meta-tensorflow?

lumag · 2026-01-13T10:30:46Z

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

+}
+
+FILES:${PN} += "${libdir} ${bindir}"
+INSANE_SKIP:${PN} += "dev-so \


Why? Your git commit message is notexisting

koenkooi · 2026-01-13T13:38:43Z

Please have a look at the recipe in #1319, that has a much better bb structure and has some comments (but not enough!) explaining the weird bits.

quaresmajose · 2026-01-13T14:24:38Z

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

+    libeigen \
+"
+
+SRCREV = "${AUTOREV}"


Please use a pinned version

quaresmajose · 2026-01-13T14:26:06Z

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

+
+TF_TARGET_EXTRA ??= ""
+
+do_configure[network] = "1"


This will break yocto check layer and is not allowed

quaresmajose · 2026-01-13T14:28:39Z

recipes-ml/tflite/flatbuffers-native_24.3.25.bb

@@ -0,0 +1,29 @@
+SUMMARY = "Memory Efficient Serialization Library"


It should be better to carry a full downgrade version of the recipe in meta-oe/recipes-devtools/flatbuffers and select this one using with the PREFERRED_VERSION

ievlogie · 2026-01-21T11:37:25Z

recipes-ml/tflite/files/0004-cmake-lite-tools-benchmark-require-protobug-through-.patch

+From c7df7a3627ef250bf7a391e3bc9e247753837e07 Mon Sep 17 00:00:00 2001
+From: Koen Kooi <koen.kooi@oss.qualcomm.com>
+Date: Thu, 9 Oct 2025 18:11:16 +0200
+Subject: [PATCH 4/8] cmake: lite/tools/benchmark: require protobug through


protobug -> protobuf

ievlogie · 2026-01-21T11:38:59Z

recipes-ml/tflite/files/0007-feat-tflite-Improve-shared-library-linking-and-build.patch

+diff --git a/tensorflow/lite/delegates/gpu/cl/opencl_wrapper.cc b/tensorflow/lite/delegates/gpu/cl/opencl_wrapper.cc
+index 49551fd372a..b8229ec1f96 100644
+--- a/tensorflow/lite/delegates/gpu/cl/opencl_wrapper.cc
+++ b/tensorflow/lite/delegates/gpu/cl/opencl_wrapper.cc


move changes in opencl_wrapper to recipes-ml/tflite/files/0006-feat-tflite-Add-dynamic-OpenCL-library-loading-suppo.patch

This commit introduces a series of patches that enhance TensorFlow Lite's GPU capabilities and build system: **GPU Optimizations:** - Fix GPU work group size adjustments and remove Adreno-specific optimizations - Improve softmax 1x1 operations to account for reported maximum threads - Optimize work group picking to ensure max_z_size doesn't exceed max work group size - Add dynamic OpenCL library loading support for better cross-platform compatibility **Build System Improvements:** - Fix protobuf dependencies in benchmark tools and label_image examples - Enhance shared library linking and build configuration - Add project versioning with VERSION and SOVERSION settings - Introduce pkg-config support with tensorflow-lite.pc.in file **New Recipes:** - Add complete TensorFlow Lite recipe (v2.20.0) with all patches applied - Append flatbuffers recipe to ensure proper protobuf dependencies These changes significantly improve GPU performance, build reliability, and cross-platform compatibility for TensorFlow Lite applications. Signed-off-by: Tushar Darote <tdarote@qti.qualcomm.com>

tdarote requested review from lumag, ndechesne, ricardosalveti, sbanerjee-quic and vkraleti as code owners January 13, 2026 09:48

lumag requested changes Jan 13, 2026

View reviewed changes

quaresmajose reviewed Jan 13, 2026

View reviewed changes

lumag marked this pull request as draft January 14, 2026 02:03

tdarote force-pushed the tflite branch from b53a503 to 2d93af3 Compare January 21, 2026 10:54

ievlogie reviewed Jan 21, 2026

View reviewed changes

Raja-Ganapathi-Busam force-pushed the tflite branch from 2d93af3 to 1d23b87 Compare January 21, 2026 13:27

		@@ -0,0 +1,29 @@
		SUMMARY = "Memory Efficient Serialization Library"

		@@ -0,0 +1,4 @@
		# Tensorflow-lite needs an extremely specific version, so lock it to that


		TF_TARGET_EXTRA ??= ""

		do_configure[network] = "1"

tflite: add patches for GPU/OpenCL fixes and new recipes #1385

Are you sure you want to change the base?

tflite: add patches for GPU/OpenCL fixes and new recipes #1385

Uh oh!

Conversation

tdarote commented Jan 13, 2026

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!

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!

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!

koenkooi commented Jan 13, 2026

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!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

6 participants