CE memcpy performance in dm816x evm

Gaston Schelotto

Hello,

I've observed in CE examples that memcpy implemented in the codec takes approximately 280ms to perform an in.dat->out.dat copy of 8Mb. Next, I've replaced this function by a custom one as follows:

for(i = 0; i < numInBytes; i++)
*(outBufs->descs[0].buf+i) = *(inBufs->descs[0].buf+i);

and it takes almost 2400ms to perform the same copy process.

I've tried to build the codec server as a release profile but I've got no improvements. Also I've included in the codec some pragmas like LOOP_COUNT_INFO with no success.

The questions are predictable:

What could be the cause of such a difference? How the memcpy is implemented in EZDSK 5.04.00.11 and how I could achieve a similar performance than the provided memcpy? Currently I've not intention to use the DMA.

Regards,
gaston

over 14 years ago

0 GAnthony over 14 years ago

TI__Intellectual 2790 points

Gaston:

The memcpy function is part of the DSP compiler's run time support (RTS) library, and is typically highly optimized, usually in assembly.

The assignment instruction in the loop above:

*(outBufs->descs[0].buf+i) = *(inBufs->descs[0].buf+i);

can be optimized by compiler options, and rewriting the code to take the any redundant operations out of the for loop: the field de-reference, array indexing, and simplify the addition to a pointer increment.

Regards,

- Gil

0 Gaston Schelotto over 13 years ago in reply to GAnthony

Expert 2880 points

Gil, thank you for the clarification. Now let's assume that I want to perform a 1920x1080x3 image analysis using CE. This involves a reading of almost 6MB, a little process and finally writing the result as another image of the same resolution (again 6MB). I know is a vague question but how I could achieve this (RD->process->WR) in real time (i.e 10ms)? The 280ms obtained from the first post, using a highly optimized function as memcpy, tells me that's not possible. I hope I'm wrong and you can give me a hint.

Regards,
gaston

0 GAnthony over 13 years ago in reply to Gaston Schelotto

TI__Intellectual 2790 points

Gaston:

Most video algorithms (and I suppose Image too) use DMA to move data from external memory into internal memory where highly optimized kernels (code) process chucks of data (macroblocks). The DMA is then used to transfer those processed chunks (or lines) back to external shared memory.

memcpy, no matter how highly optimized in software, generally cannot be beat by a DMA (hardware copy).

If you want to evaluate the time taken for a DMA copy, note the videnc1_copy codec comes with a useDMA configuration option which can be enabled in the server.cfg file:

VIDENC1_COPY.alg.useDMA = false;

There are some forum links which discuss this subject:

http://e2e.ti.com/support/embedded/linux/f/354/t/129914.aspx?pi25291=1

http://e2e.ti.com/support/embedded/bios/f/355/p/70940/258741.aspx#258741

http://e2e.ti.com/support/embedded/linux/f/354/p/105920/374369.aspx#374369

Regards,

- Gil

0 Gaston Schelotto over 13 years ago in reply to GAnthony

Expert 2880 points

GAnthony said:
If you want to evaluate the time taken for a DMA copy, note the videnc1_copy codec comes with a useDMA configuration option

Okay, I'd also like to evaluate the time taken for a process running in 'scratch' internal memory of C674x. To do this, I've taken a look at C64x+ iUniversal Codec Creation - from memcpy to Canny Edge Detector which seems to be close to my purpose. Basically I want to use my UNIVERSAL VISA API rather than VIDENC or VIDDEC. However I find this wiki example quite outdated with references to CCSv3 and CE 2.21, which is hard to understand using my current framework and platform. Is there a similar example using CE 3.xx on a TI816x platform?

Regards,
gaston

0 Ramsey over 13 years ago in reply to Gaston Schelotto

TI__Genius 12025 points

Gaston,

I could not find any articles specific to CodecEngine 3.x, TI816x platform, and iUniversal codec. Here is a Getting Started guide for iUniversal, but you are probably way beyond that already.

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

Maybe you can post specific questions as you work your way through the Canny Edge Detector article.

~Ramsey

Processors

Processors forum

CE memcpy performance in dm816x evm