ARM reads captured data from shared memory is too slow!

-Perry

I am working on DVRRDK_02_00_00.23 code base with dm8168evm + vs card. I have sucessfully intercepted the capture video data frame that VPSS-M3 has placed into the shared memory. So that I can process the capture video frames in A8 using my own algorithm and place the processed data into another buffer. But I noticed that reading YUV data from shared memory is much slower than accessing data from the local memory that was allocated by A8 itself.

I basically I got the frame pointer from ipcFramesSendRecvFxn() and then use mmap() to obtain the virtual address of the frame pointer on A8. Is there any specific thing I need to be aware of when applying an algorithm directly on the captured frame data which sit in the shared memory region?

Thanks a lot,

-Perry

over 11 years ago

0 Badri Narayanan over 11 years ago

TI__Guru 59700 points

The frame buffers are not cached on A8 and that will result in much slower access time compared to local memory that was allocated by A8 itself.The buffer should be mapped as Cacheable+WriteBack_WriteAllocate+Not Strongly Ordered to get good performance on CPU access from A8

0 guohua wei over 8 years ago in reply to Badri Narayanan

Intellectual 350 points

Hi Badri Narayanan
I encountered a similar problem,and I use dvrrdk4.0 and dm8168. What do you mean "The buffer should be mapped as Cacheable+WriteBack_WriteAllocate+Not Strongly Ordered". How to do that?

Processors

Processors forum

ARM reads captured data from shared memory is too slow!