dsp: which is the fastest way to transmit the image data?

Liang,

Copying the data across processors is not a good idea (e.g., w/ Proc_write()). The standard practice for transferring large data sets to the DSP for crunching is to use DDR memory, which is visible to both the DSP and ARM.  The ARM will map this memory to its address space and give the virtual address of this memory to the "capture" device so that the DDR memory is directly filled by the source driver.  You will then use DSPLINK to communicate the physical address to the DSP, most likely using a small MSGQ message, and the DSP will access this memory directly, possibly allowing it to be cached.

The TI product Linux Utils has a module named CMEM that can be used to get a physically contiguous [large] data buffer, and also to get the physical address (that the DSP needs). CMEM handles memory that is assigned to it by the user, and typically this memory is set aside outside of Linux control. Or, you can choose to put your DSP-side code into an IUNIVERSAL alg (wizards available!) and use Codec Engine to do some of the stuff above.

Here are some links with more information on some of the topics above:

·         http://processors.wiki.ti.com/index.php/Cache_Management

·         http://processors.wiki.ti.com/index.php/CMEM_Overview

·         http://processors.wiki.ti.com/index.php/Getting_started_with_IUNIVERSAL

·         http://processors.wiki.ti.com/index.php/Codec_Engine_Application_Developers_Guide

·         http://processors.wiki.ti.com/index.php/Codec_Engine_Overhead