Linking and search path

[thewh1teagle]

I'm confused with linking and search path while creating bindings to sherpa-onnx in sherpa-rs/build.rs

Do I always need to link kaldi-native-fbank-core, sherpa-onnx-core etc? What if shared libs is on eg. with Cuda or DirectML?

As for espeak-ng, piper_phonemize. always link static? only when tts enabled?
I'm confuse with what to link and how.
Also regarding DLLs, they should always be available inside build/lib folder after compile?

Also, I like to link statically as far as possible. When it's not possible?

[csukuangfj]

You have to distinguish between a shared build and a static build of sherpa-onnx.

For a static build with TTS, please refer to

sherpa-onnx/cmake/sherpa-onnx-static.pc.in

Line 25 in 35c1b4a

Libs: -L"${libdir}" -lsherpa-onnx-c-api -lsherpa-onnx-core -lkaldi-decoder-core -lsherpa-onnx-kaldifst-core -lsherpa-onnx-fstfar -lsherpa-onnx-fst -lkaldi-native-fbank-core -lpiper_phonemize -lespeak-ng -lucd -lonnxruntime -lssentencepiece_core -Wl,-rpath,${libdir} @SHERPA_ONNX_PKG_WITH_CARGS@ @SHERPA_ONNX_PKG_CONFIG_EXTRA_LIBS@

For a static build without TTS, please refer to

sherpa-onnx/cmake/sherpa-onnx-static-no-tts.pc.in

Line 25 in 35c1b4a

Libs: -L"${libdir}" -lsherpa-onnx-c-api -lsherpa-onnx-core -lkaldi-decoder-core -lsherpa-onnx-kaldifst-core -lsherpa-onnx-fst -lkaldi-native-fbank-core -lonnxruntime -lssentencepiece_core -Wl,-rpath,${libdir} @SHERPA_ONNX_PKG_WITH_CARGS@ @SHERPA_ONNX_PKG_CONFIG_EXTRA_LIBS@

For a shared build (regardless of with or without TTS), please refer to

sherpa-onnx/cmake/sherpa-onnx-shared.pc.in

Line 25 in 35c1b4a

Libs: -L"${libdir}" -lsherpa-onnx-c-api -lonnxruntime -Wl,-rpath,${libdir} @SHERPA_ONNX_PKG_WITH_CARGS@ @SHERPA_ONNX_PKG_CONFIG_EXTRA_LIBS@

---

What if shared libs is on eg. with Cuda or DirectML?

In that case, only shared lib should be used.

---

Also regarding DLLs, they should always be available inside build/lib folder after compile?

Yes, as long as you run make install. Otherwise, the onnxruntime library won't be copied to build/lib.

---

Also, I like to link statically as far as possible. When it's not possible?

Whe you use CUDA or directml, you have to use a shared lib of sherpa-onnx.

---

By the way, it you want to always use a static build of sherpa-onnx, you need to change CMakeLists.txt so that
it supports static building even if you use CUDA or directML.

The source code of sherpa-onnx can be built from the source, so it is possible to always use a static build.

[thewh1teagle]

Thank you! it really helps.
Now I use these 3 .pc files to know exactly what to link.
I understand that regarding the DLLs files they should be places under lib in the end and I don't need to search in another places.
Regarding when using directml or cuda - how can I know which lib names to link dynamic / static? I can't see .pc file of cuda.

Also, the search path with cuda/directml etc - should I set it to something like build/_deps/onnxruntime-src/lib?

[csukuangfj]

how can I know which lib names to link dynamic / static?

Please see what lib files are in build/_deps/onnxruntime-src/lib

---

Also, the search path with cuda/directml etc - should I set it to something like build/_deps/onnxruntime-src/lib?

Yes, I think that is reasonable.