AM5728: GStreamer appsink performance issue

xixi

Expert 2990 points

Other Parts Discussed in Thread: AM5728

HI all:

Just now we are using am5728 to develop an h264 decode application.

Our gstreamer pipeline is like below:

h264 file -> h264parse -> ducatih264dec -> vpe -> appsink.

At last we get the yuv buffer from the appsink.

Now we found the appsink’s performance is very low. The situation is

When we copy a 3MB buffer from the appsink, it will cost about 50ms.

We think the time is too long.

The below is the appsink callback function

sample = gst_app_sink_pull_sample(GST_APP_SINK(object));

buffer = gst_sample_get_buffer(sample);

gst_buffer_map(buffer, &map, GST_MAP_READ);

gettimeofday(&tv1 ,&tz);

memcpy(pConv,map.data,map.size);

gettimeofday(&tv2 ,&tz);

The tv2 – tv1 = 50ms

So we want to know how can I decrease the copy time

Our am5728 is an custom board.

The Processor SDK is 03.03.00.04

over 8 years ago

0 Biser Gatchev-XID over 8 years ago

TI__Guru**** 393215 points

Hi,

Memcpy is slow. Try using DMA.

0 xixi over 8 years ago in reply to Biser Gatchev-XID

Expert 2990 points

Thanks for your reply
And How can I use the DMA, after all We don't know the buffer's(from the appsink) physical address.

0 Biser Gatchev-XID over 8 years ago in reply to xixi

TI__Guru**** 393215 points

I have asked the software team to comment. They will respond here.

0 xixi over 8 years ago in reply to Biser Gatchev-XID

Expert 2990 points

OK , Thank you

0 Manisha Agrawal over 8 years ago in reply to xixi

TI__Mastermind 22946 points

Buffers for vpe and codec are getting allocated from omap_bo. The omapdrm driver doesn't support cache operations on it's allocated buffer and hence in absence of read cache enable, the memcpy is slow.

0 xixi over 8 years ago in reply to Manisha Agrawal

Expert 2990 points

Hi manisha

Firstly thanks for your reply.

So How can we solve this issue

If we use EDMA to transfer the data and How can we get the physical address from appsink.

Or Do you have any other advice for us, we just want to copy the data from the gstreamer pipeline.

Best regards

0 Manisha Agrawal over 8 years ago in reply to xixi

TI__Mastermind 22946 points

If you are able to modify the code, you can allocate the buffers from CMEM driver (it supports cache management) rather from omapdrm driver. Check dual camera demo loopback.c file (follow code under #ifdef USE_CMEM_BUF) to understand how to allocate buffers from CMEM driver.
git.ti.com/.../master

0 xixi over 8 years ago in reply to Manisha Agrawal

Expert 2990 points

HI manisha

Yes we can use CMEM to alloc the buffer, such as alloc_cmem_buffer()

But we are so confusing about how can we use the CMEM buffer in the gstreamer pipeline.

Best regards

wang

0 xixi over 8 years ago in reply to xixi

Expert 2990 points

Hi all

Anyone can help me about this issue

regards

Wang

0 Manisha Agrawal over 8 years ago in reply to xixi

TI__Mastermind 22946 points

You will need to modify the gstreamer plugins to allocate the buffer from cmem.

0 xixi over 8 years ago in reply to Manisha Agrawal

Expert 2990 points

So ，How can I get the gstreamer plugin source code?

Thanks

regards

0 Manisha Agrawal over 8 years ago in reply to xixi

TI__Mastermind 22946 points

You can get it from here -

https://git.ti.com/glsdk/gst-plugin-ducati

0 xixi over 7 years ago

Expert 2990 points

Hi all

Months ago, we found the am5728 appsink performance is very low and we send an post to e2e, the below is the website.

https://e2e.ti.com/support/arm/sitara_arm/f/791/t/589689

The e2e have told us the solution， which is modify the appsink’s source code and change memory allocation from omap_bo to CMEM.

Just now we don’t know how to modify the source code because we can’t find the memory allocation code.

Can you give us some advice.

Our Processor SDK RT3.3

Thanks

Regards

Processors

Processors forum

AM5728: GStreamer appsink performance issue