使用matlab和cuda比较
(2011-12-13 08:27:30)
标签:
it
Implementing Algprithms for Signal
and Image Reconstruction on Graphical Processing Units
Our main point of comparison is between
the runtime for the Matlab implementation running on the CPU of the
host
machine (which we refer to hereafter as the \CPU implementation"),
and
the runtime for the CUDA-based GPU implementation.
A second dierence between the implementations is the use of Matlab
code for
the CPU implementation vs. CUDA code (an extension of C++) for the
GPU imple-
mentation. We believe however that a CPU implementation via C++ and
mex les
would show little if any improvement over the Matlab
implementation. The major
computational operations are DFT, DCT, and Level 1 BLAS operations,
all of which
are implemented with high eciency in Matlab. In fact, we could
speed up the GPU
implementation further by using C++ calling code in place of the
Matlab and mex
software. Multiple data transfer requests on page-locked host
memory could be made
without having to wait for the completion of the previous requests,
making room for
other jobs in the host or GPUs. In our current Matlab / mex
interface, data are
stored in the Matlab memory, which is not page-locked, so
overlapped data transfers
require extra copies to page-locked host memory. Such transfers
only degrade the
performance, so we avoid them.
分享:
喜欢
0
赠金笔
加载中,请稍候......
评论加载中,请稍候...
发评论
登录名: 密码: 找回密码 注册记住登录状态
昵 称:
评论并转载此博文
发评论
以上网友发言只代表其个人观点,不代表新浪网的观点或立场。