Google Git
Sign in
chromium / chromiumos / platform / tflite / refs/heads/main
commit961f27c5e8695e4a941bade8bdfa0a5bb3a516e1[log] [tgz]
authorTommy Chiang <ototot@google.com>Wed Jul 09 16:32:01 2025
committerChromeos LUCI <chromeos-scoped@luci-project-accounts.iam.gserviceaccount.com>Thu Jul 10 04:34:47 2025
treedb63377fc06a8bf46f979e15e9cea27f17dc41b7
parenta964217d52f816b7c55e4222f11ce5d38e21f248 [diff]
tflite: Compute numbers in float16 to match the NPU behavior

If we compute numbers in float32, the error will accumulate, so there
are chances that it eventually exceed the threshold.

For example, when computing 0.375732 + 0.298828 - 0.641113 in float32,
the result is 0.0334473, but it will be 0.0336914 in float16.
Note that 0.375732 + 0.298828 = 0.674561(fp32) or 0.674805(fp16).

To address the issue, we compute numbers in float16 to mimic the same
calculation in NPU.

BUG=b:430326201
TEST=`./async_delegate_test --stable_delegate_settings_file=/etc/ml_core/stable_delegate_settings.json;` on screebo and hylia passes 3000 times.

Change-Id: Iea8c46ce83150cc592ddb8c114287b00f63ab67f
Reviewed-on: https://chromium-review.googlesource.com/c/chromiumos/platform/tflite/+/6719772
Auto-Submit: Tommy Chiang <ototot@google.com>
Commit-Queue: Shik Chen <shik@chromium.org>
Tested-by: Tommy Chiang <ototot@google.com>
Reviewed-by: Shik Chen <shik@chromium.org>
2 files changed
tree: db63377fc06a8bf46f979e15e9cea27f17dc41b7
  1. android/
  2. common/
  3. delegate/
  4. patch/
  5. script/
  6. tool/
  7. udev/
  8. .bazelignore
  9. .bazelrc
  10. .bazelversion
  11. .clang-format
  12. .gitignore
  13. BUILD.bazel
  14. CPPLINT.cfg
  15. DIR_METADATA
  16. OWNERS
  17. PRESUBMIT.cfg
  18. README.md
  19. WORKSPACE.bazel
README.md

ChromeOS TFLite

This repository hosts the core ChromeOS TFLite components, enabling on-device machine learning (ODML) workloads accelerated by NPU.

The corresponding ebuild can be found at: tensorflow-9999.ebuild

TensorFlow Patch Management

Patches are stored in the patch/ directory and explicitly listed in WORKSPACE.bazel. A helper script, ./script/patcher.py, is included to facilitate patch management within a TFLite workspace.

The typical workflow:

  1. Eject (Download) TensorFlow Source Code

    Download the TensorFlow source code into a local git repository with patches applied as individual commits:

    ./script/patcher.py eject

    This creates a new local git repository at tensorflow/.

  2. Modify the TensorFlow Repository

    Make changes to the tensorflow/ repository as needed, following standard git workflows. Optionally, include a PATCH_NAME= tag in commit messages to specify the filename of the corresponding patch.

  3. Seal the Repository

    Regenerate the patch files and update the WORKSPACE.bazel file:

    ./script/patcher.py seal

    This updates the patches in the patch/ directory and reflects the changes in WORKSPACE.bazel.

It's preferred to submit changes to upstream TensorFlow first and cherry-pick them as patches. This helps minimize divergence and makes TensorFlow updates easier.