Copying Data in external memory

Even if cache is enabled, if you have never accessed buffer A before (or if you have accessed enough other data for buffer A to be evicted from the cache) than there will still have to be an initial cache miss for all of the data to be read in from DDR, so that is likely the delay you are seeing, or at least part of it.

If your goal is to get buffer A to buffer B in external memory than DMA will be the most efficient way to do it, since DMA will allow more efficient bursting memory accesses than a multitude of individual CPU transfers, in addition to allowing the CPU to do other work while the transfer is in progress. Of course with DMA you also have to worry about cache coherency, so you need to do an invalidate for the buffer B region before reading it from the CPU.

Questions for you:

1. Which parts of the device will access buffer A and when?  For example, is it written there by the video capture interface?  Does the CPU directly operate on that data?
2. Same question but for buffer B.
   

There are lots of cache coherence issues "lurking" in this simple scenario so if you can provide me specific usage info I can provide you specific cache operations to perform.

No matter how you write your code you'll never be able to move 8 bytes per cpu cycle.  There are several factors affecting the speed at which you can transfer data in external memory:

• External memory runs at a slower clock speed than CPU.
• The width of the external bus is less than the internal bus widths so you can transfer fewer bytes per DDR cycle.  (i.e. a STDW instruction will translate to multiple DDR stores)
  
• DDR is a type of DRAM which inherently requires operations like refreshes, page open, page close, etc. that further degrades overall throughput.

Brad

Hi Bernie and Brad,

Thanks for your replies.

Considering the Video Decoder process,Buffer A is my reference buffer which holds the

reference frame (I or P frames) and are operated only by the CPU.Buffer B is my output

buffer, in case of B frames it is directly filled by the process while for I and P frames

it is copied from the corresponding reference buffers(Buffer A).This buffer later on,is

required by the peripheral to display.

Waiting for your replies,

Regards,

Sandeep

Optimally you would want to avoid the copy all together and just point the display driver to the buffer that currently contains the frame you want to display, i.e. if buffer A contains the frame you want to display give a pointer to buffer A to the video driver and have your video decoder process fill a different buffer.

If you did want to keep the copy than the fastest way to get from A to B would be to use a DMA transfer, something like:

1. Perform cache writeback on buffer A space (maybe BCACHE_wb(...), this ensures that the latest data is in external memory)
2. Configure a DMA transfer to go from A to B and initiate it with CPU with a completion interrupt (or you could poll on it, but this would waste CPU cycles)
3. Wait for the DMA to complete (block the thread and do something else or poll)
4. In the completion interrupt unblock the thread and than pass buffer B to the display driver

This should get the data from A to B in the fastest way possible since the DMA allows for bursting memory accesses in addition to allowing your CPU to keep processing while the transfer goes on.

EDIT: This is all assuming that all the data is being generated by the CPU, since you mention this is a video decoder (h.264?). It also assumes that buffer B is only used to display video, such that the CPU never needs to read back from buffer B (if you did you would need to perform a cache invalidate on buffer B).