[jira] [Commented] (SINGA-41) Support single node single GPU training

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

    [ https://issues.apache.org/jira/browse/SINGA-41?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14743047#comment-14743047 ] 

Haibo Chen commented on SINGA-41:
---------------------------------

Support single node single GPU training for latest Singa version

> Support single node single GPU  training
> ----------------------------------------
>
>                 Key: SINGA-41
>                 URL: https://issues.apache.org/jira/browse/SINGA-41
>             Project: Singa
>          Issue Type: New Feature
>            Reporter: Haibo Chen
>
> For new feature: support single node single GPU training.
> The blob class:
> 1) add a member function
> Dtype* mutable_xpu_data();  // make a decision(GPU or CPU) at compile time
> 2) add a include file: #include <cuda_runtime.h> //support cuda code
> The param class:
> 1) add three member function
> float* mutable_xpu_data() 
> float* mutable_xpu_grad() 
> float* mutable_xpu_history() 
> The layer class:
> 1) we use xpu instead of cpu in some layers of subclass,eg.ConvolutionLayer、Dropout...
> Makefile.GPU:
> 1) For CPU mode,set CUDA_DIR :=
> step 1:
> make
> 2) For GPU mode,set CUDA_DIR := your cuda installation path
> step 1:
> make gpu         // rename layer file : layer.cc ->  layer.cu
> step 2:
> make



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