Digital Media Processors based on DaVinci™ Technology

Re: Video processing algorithm in DSP

liangzhu chen — Wed, 30 Sep 2009 12:12:37 GMT

[root@OMAP3EVM example]# ./app_remote.xv5T
@0x0008bff1:[T:0x4001cfb0] ti.sdo.ce.examples.apps.video_copy.singlecpu - main> ti.sdo.ce.exampl<1>Existing entry's end address is covered by given entry's start & end address]
Existing entry's end address is covered by given entry's start & end address., can not create TLB entry for address: [0x86000000] size: [0x1000000]
es.apps.video_copy.singlecpu
@0<1> DSP_init status [0x80008050]
DSP_init status [0x80008050]
x0008c272:[T:0x4001cfb0] ti.sdo.ce.examples.apps.video_copy.singlecpu - App-> Application started.
app: error: can't open engine video_copy
@0x0009787a:[T:0x4001cfb0] ti.sdo.ce.examples.apps.video_copy.singlecpu - app done.

in the xdcpaths.mak ,I set PROGRAMs = DSP_SERVER APP_CLIENT

Re: Video processing algorithm in DSP

Juan Gonzales — Fri, 24 Apr 2009 15:41:48 GMT

Hi Jonners,

So first, I just want to clarify what you mean by image, since 'DSP image' is a term we often use to refer to the DSP binary which is loaded and executed on the DSP. From the context of your inquiry, it appears we are referring to sharing memory buffers (which can contain a video image or anything else) between the ARM and DSP. Just wanted to clarify this for anyone else that may be reading this post.

If we are aligned on the above, then the simplest form of this process are demos (encode, decode, encodedecode) included with the DVSDK; this will give you the Linux application level interface (what is commonly referred to VISA APIs) to allow ARM to pass video buffer to DSP (say for encoding), and DSP pass the processed video buffer back to ARM. In this scenerio, all you would need to learn is VISA APIs (they use DSPLINK underneath without application developer having to learn it)

If however, you are writing not only the consumer Linux application, but also the DSP algorithm, and packaging your algorithm into a DSP server image that ARM Linux application can consume, then you will need to learn a bit more than the VISA APIs. In this case the best place to look are the codec engine examples, which include examples with source code for DSP algorithms, DSP servers and applications. Again, if you stick to our codec engine framework, eventually this would lead to using VISA APIs or extension of them at the Linux application level. Let me know if this helps.

Re: Video processing algorithm in DSP

Jonners — Fri, 24 Apr 2009 14:54:32 GMT

Juan Gonzales:

BMillikan:

Is there an example of this somewhere so we can kind of get a feel for how this works? I understand the CMEM pointer sharing, but what does the "DSP" driver look like in linux? How does the DSP side know that that it received a new image? Can it be interrrupt or event driven (instead of using a flag?)

Hopefully my previous post will help you too. The best place to see how it all fits together is to look at the codec engine examples. FYI, only one DSP image can be loaded on DSP at an one time, but this image can have multiple instances of dsp algorithms. Let us know if you need any further assistance.

Hi Juan,

I have similar a question. I am using DM6446. I have done all my image processing on the DSP side. Now, I would like to send an image from the ARM to the DSP for processing and send the results back to the arm. I have read how CE, DSP link works but, I am a bit confused.

Can you help, please give or direct me to a simplest example on how to pass an image from the ARM to the DSP for processing and back to the ARM.

Thanks in advance.

Jonners

Re: Video processing algorithm in DSP

Jonners — Fri, 24 Apr 2009 14:50:53 GMT

Hi Juan,

Can you help, please give or direct me to a simplest example on how to pass an image from the ARM to the DSP for processing and back to the ARM.

Thanks in advance.

Jonners

Re: Video processing algorithm in DSP

Miao — Tue, 10 Mar 2009 12:49:36 GMT

I also want to develop a video processing algorithm of XDAIS(non-XDM), in my opinion:

a) in ccs, write XDAIS algorithm and complie to .lib file, i.e. a64p file in linux;

b) in linux. produce from .a64p to .x64p(codec server);

c)in linux, write application program thar run in arn, as well as stub and skeleton;

d)in linux, install the executable program and the .x64p to workdir and work

am i right?

question:

1.in b) below, how to produce .x64p file (dsp server or codec server) from .a64p(changed suffix from .lib in ccs) in linux ?

i've done the same as Lorry

LorryAstra:

I've tried to compile an achive file in CCS (for example: videnc_copy.lib) in windows, and then I rename it (videnc_copy.a64P) in lib folder in Linux. And then I compile an codec server and copy it to "/home/usraccount/workdir/filesys/opt/dvevm". I want to know whether my codec can work well, but I failed.

Sequence:

a) I compile videnc_copy.lib in CCS, and I'm sure that all the interfaces have been realized.

b) I copy videnc_copy.lib to /home/usracc/dvevm_x_xx/codec_engine_x_xx/examples/codecs/videnc_copy/lib , and I rename this lib file to videnc_copy.a64P.

c) In .../examples/servers/video_copy/ folder, I recompile codec server file and copy it to .../workdir/filesys/opt/dvevm.

d) I want to test whether videnc_copy can work well.

After checking, I find that the videnc works, but it is not the code that I compile in CCS, it works as the original videnc_copy.a64P in videnc_copy example in Linux. It sounds that the videnc_copy.lib does not affect codec server (video_copy.x64P).

So I guess, the new videnc_copy.a64P( renamed from videnc_copy.lib) is not compiled when generating video_copy.x64P.

then how can i produce 'my' video_copy.x64p from 'my' videnc_copy.a64P(renamed from videnc_copy.lib)?

2. if i write an XDAIS algorithm in ccs and compile it to .lib, then write a test.c in ccs and produce .out, can this .out file be codec server in linux after renamed as .x64p?

if not, then the .out file is just used to test the .lib, and after testing we integrate the .lib to .x64p (codec server) in linux , am i right ?

Re: Video processing algorithm in DSP

mursel yildiz — Mon, 09 Mar 2009 13:58:48 GMT

I think it has.. couse anytime I try to define a function (arbitrary) inside videnc_copy.c file, I get errors like "the function is not defined" types...

Re: Video processing algorithm in DSP

LorryAstra — Mon, 09 Mar 2009 07:00:02 GMT

Of course, it does not have special rules, it is c language.

Re: Video processing algorithm in DSP

mursel yildiz — Mon, 09 Mar 2009 06:30:40 GMT

Hi Lorry...

Is that possible to use a function (that is predefined somewhere) inside a codec engine process function? Or let me ask how we may define a functions that are planned to be used inside a process function ?

Re: Video processing algorithm in DSP

LorryAstra — Thu, 05 Mar 2009 06:06:21 GMT

That's really a problem. for example: If I want to blur a image, I don't want to use convolution in ImgLib, so I have to use the code like " x = input [ i ]; y = x * a; "... continuousely, If this is a main reason for a slow program, then my blur algorithm will be very slow. I guess if there are two or more thread in an algorithm, the collision you've said maybe occur in that scene.

Re: Video processing algorithm in DSP

mursel yildiz — Wed, 04 Mar 2009 21:15:55 GMT

hi Lorry...

I actually have another question for the problem...( I should clear it more :) ) ... Having experimented many algorithms on DSP core ( just for exploring DSP performance ) I realized that no matter what kind of statement inside the code I write, may be because of a possible pipeline architecture, once I write statements like:

x = input;

y = x * a;

i.e. , y waits for read of x, the DSP algorithm becomes slower, I think second statement waits after "instruction decode" phase for the first statement and delaying other instructions ... However once I try

x = input;

....(unrelated stm1)....

....(unrelated stm2)....

....(unrelated stm3)...

y = x * a

the DSP algorithm becomes faster... Well actually this is just a suggestion, I am not sure.... Do you think this may be the problem???

Re: Video processing algorithm in DSP

LorryAstra — Wed, 04 Mar 2009 02:27:13 GMT

Hi, mursel , I just tried your code described above, but I did not find your question ( I use evm6446) , I believe to read from memory is not the main reason related to your question, because this kind of operation is very common in algorithm I believe. So I think maybe .... there are some other code operations collide with the reading. That's just what I guess. :)

Re: Video processing algorithm in DSP

mursel yildiz — Tue, 03 Mar 2009 17:03:23 GMT

sorry for symbols they are off course [ i ] :D

Re: Video processing algorithm in DSP

mursel yildiz — Tue, 03 Mar 2009 17:00:25 GMT

Hi all,

I have written my algorithm inside videnc_process() function and saw that it worked well... However it is unfortunately slow. I have a for statement inside the function like:

for(i = 0; i < 100000; i ++) {

// x = myinbuf;

......

}

when I dont perform this memory read statement ( x = myinbuf ) the algorithm runs very fast and I get 25 fps image frames. However, when I write this statement inside for, I get only 1 fps image frame. Actually there are lots of MAC operations and divisions inside this for statement and they are performed very fast as expected... The given myinbuf is generated by a memcpy() statement from CMEM area. I could not actually understand why a memory read statement costed so much.... Should I try something different??? or what may be the reason for such a case ??

Re: Video processing algorithm in DSP

Chris Ring — Thu, 26 Feb 2009 17:25:59 GMT

Juan Gonzales:

Inside codec engine, you have three types of examples.

1) The DSP codec (e.g. videnc_copy); this normally produces a library (a64P or lib file) that runs on DSP.

2) The DSP server which produces an x64P file. This is the actual DSP image that you will load and run on DSP.

3) ARM side Linux application which uses codec engine APIs to load DSP server and access codecs within a server.

FWIW, the CE examples reflect the user roles described here:

http://wiki.davincidsp.com/index.php?title=Codec_Engine_Roles

That article may also help with understanding what docs/tools are available as you march through the different roles.

Chris

Re: Video processing algorithm in DSP

Juan Gonzales — Thu, 26 Feb 2009 14:49:54 GMT

BMillikan:

Is there an example of this somewhere so we can kind of get a feel for how this works? I understand the CMEM pointer sharing, but what does the "DSP" driver look like in linux? How does the DSP side know that that it received a new image? Can it be interrrupt or event driven (instead of using a flag?)

Hopefully my previous post will help you too. The best place to see how it all fits together is to look at the codec engine examples. FYI, only one DSP image can be loaded on DSP at an one time, but this image can have multiple instances of dsp algorithms. Let us know if you need any further assistance.