This repository was archived by the owner on Mar 30, 2022. It is now read-only.
This repository was archived by the owner on Mar 30, 2022. It is now read-only.
Prebuilt tensorflow library doesn't contain all symbols from C++ tensorflow API #436
Open
Description
Imported from JIRA: https://bugs.swift.org/browse/TF-1177
Type: | Bug | Priority: | Minor |
---|---|---|---|
Reporter: | Andrey Volodin | Assignee: | Unassigned |
Resolution: | Unresolved | Votes: | 0 |
Labels: | None |
I'm trying to implement a custom op with a S4TF. I'm successfully compile an .so file with custom kernel representation, but when I try to load it via TF_LoadLibrary, I get this message:
TF_Code(rawValue: 5)
dlopen(custom-op-master/tensorflow_zero_out/cc/zero_out.so, 6): Symbol not found: __ZN10tensorflow10DEVICE_CPUE
Referenced from: custom-op-master/tensorflow_zero_out/cc/zero_out.so
Expected in: flat namespace
in custom-op-master/tensorflow_zero_out/cc/zero_out.so
After an investigation I think that prebuilt artifact that comes with toolchain simply doesn't contain that symbol which I think comes from Eigen. Is it possible to ship a broader binary with toolchain?
In order to compile a 3rd party I did the following:
Installed tensorflow with the version from Swift repo configuration json
Obtain headers from there
Compiled custom op example with both regular g++ and clang that comes with toolchain
linked zero_out.so to libtensorflow that comes with toolchain
Metadata
Metadata
Assignees
Labels
No labels