[jira] [Assigned] (SINGA-112) Clean GNU configure and make code

["wangwei (JIRA)" <ji...@apache.org>]

     [ https://issues.apache.org/jira/browse/SINGA-112?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]

wangwei reassigned SINGA-112:
-----------------------------

    Assignee: Xie Zhongle

Some files are compiled multiple times, e.g., the cudnn code.
After compilation, there would be libsinga_la-cudnn_xxx.lo and singatool-cudnn_xxxx.o.
It would be better to remove the repeated compilation to make the compilation step faster.

> Clean GNU configure and make code
> ---------------------------------
>
>                 Key: SINGA-112
>                 URL: https://issues.apache.org/jira/browse/SINGA-112
>             Project: Singa
>          Issue Type: Improvement
>            Reporter: wangwei
>            Assignee: Xie Zhongle
>
> 1. The current compilation via `configure && make` would generate some extra files under src/neuralnet/neuron_layer/. It would be better to clean the make code to remove them.
> 2. The singatool now depends on some unnecessary files, e.g., cudnn related source files. It should only depends on libsinga.so. May using this flag to remove the dependency, `-Wl,-unresolved-symbols=ignore-in-shared-libs`
> 3. If there is a cudnn.so (e.g., from cudnn v1 or v2) and inside the cuda folder, then the running (or even compilation) would have errors. One solution is putting user defined cudnn prefix before cuda prefix in the compilation command.
> 4. current we put almost all files into libsinga.so. We may need to separate some files from libsinga.so.



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)