search

tensorflow / tflite-micro

Infrastructure to enable deployment of ML models to low-power resource-constrained embedded targets (including microcontrollers and digital signal processors).
Apache License 2.0
1.91k stars 822 forks source link

build: C++17 is not enabled on all platforms #2650

Closed rkuester closed 3 months ago

rkuester commented 3 months ago

Bazel builds have been C++17 since 52c9568; however, Makefile builds remain C++11. Consistent with our webpage documentation and to match TF Lite's language support, it's our desire to support and compile as C++17 on all platforms.

rkuester commented 3 months ago

After trying to set -std=c++17 and c17 in PR #2648, the only problem appears to be the age of the Xtensa toolchain. The flatbuffers library includes <optional> when it's being compiled under the C++17 standard. The Xtensa toolchain is happy enough to compile with -std=c++17, but the standard library appears to be incomplete—it doesn't have <optional>. E.g., from the PR's CI run:

https://github.com/tensorflow/tflite-micro/actions/runs/10152485116/job/28074001989?pr=2648#step:3:367

xt-clang++ -std=c++17 [...]
[....]
.../micro/tools/make/downloads/flatbuffers/include/flatbuffers/stl_emulation.h:41:12: fatal error:
'optional' file not found
  #include <optional>
           ^~~~~~~~~~

Perhaps upgrading the Xtensa toolchain, which is quite old, will fix this.

For reference, the current toolchain is identified and invoked here: https://github.com/tensorflow/tflite-micro/blob/6b676567b79053d4d275ae834427b4f1259a710d/.github/workflows/xtensa_presubmit.yml#L35

rascani commented 3 months ago

I tested this out for at least the hifi3 builds and got it to compile with -stdlib=libc++. I think the default stdlib for Xtensa only supports C++14. The downside is that I think that pre-compiled library includes rtti & exception support, so it likely has some bloat. In newer versions of Xtensa toolchain, we can use libc++-re to disable those.

rkuester commented 3 months ago

Fixed in #2648.