topic Hi Steve, in Intel® Integrated Performance Primitives

Field copy with Alpha

BatterseaSteve — Wed, 10 Jul 2013 12:37:10 GMT

I have a question

I have a source buffer of interlaced, interleaved YUV with alpha

i.e. UYAVYA - 1920x1080

I need to strip the alpha, extract one field, resize it to half and convert to 420planer

as a single buffer (i.e not 3 separate planes but a single buffer of concatenated Y,U,V

Can anyone recommend the fastest set of methods to do this?

Cheers

Steve

The main problem I have with

BatterseaSteve — Wed, 10 Jul 2013 14:19:00 GMT

The main problem I have with this is actually the alpha strip. i.e. getting the UYAVYA into UYVY.

I tried the following:

IppiSize uyvyRoi = {1,1920*1080};

this->uyvyBuffer = ippsMalloc_8u(1920*1080*sizeof(Ipp8u)*2);

ippiCopy_16u_C1R((Ipp8u *)uyavyaBuffer, 3, this->uyvyBuffer, 2, uyvyRoi);

but this is incredibly slow (15msec!)

I am struggling to find anything else that will strip a single channel from a 3 channel image, or copy a 3 channel image to a 2 channel image.

Any help appreciated.

Steve

Hi Steve,

Sergey_K_Intel — Wed, 10 Jul 2013 14:36:38 GMT

Hi Steve,

For stripping alpha channel values you can probably use ippiSwapChannels_8u_C4C3R function. There's dstOrder[3] array, specifying the order of output channels.

Regards,
Sergey

Hi Steve,

Igor_A_Intel — Wed, 10 Jul 2013 14:49:17 GMT

Hi Steve,

as finally you need planar format - you can try to use ippiCopy_8u_C4C1R 3 times - for Y, U and V

regards, Igor

Hi Sergey

BatterseaSteve — Wed, 10 Jul 2013 14:54:24 GMT

Hi Sergey

Thanks for responding. As far as I can tell, the ippiSwapChannels_8u_C4C3R will only swap 4 component to 3 component - I need 3 component to 2 component.

Cheers

Steve

OK

BatterseaSteve — Wed, 10 Jul 2013 23:58:52 GMT

I have most of this worked out - except the alpha strip

As I said above, i can use ippiCopy_16u_C1R - but this is taking a long time - 15msecs

Can anyone tell me why?

Cheers

>>As I said above, i can use

SergeyKostrov — Thu, 11 Jul 2013 03:06:45 GMT

>>As I said above, i can use ippiCopy_16u_C1R - but this is taking a long time - 15msecs >> >>Can anyone tell me why? You didn't tell us anything about hardware, that is CPU, and my question is do you use a right CPU Dispatching DLL for Image Processing domain?

Hi Sergey

BatterseaSteve — Thu, 11 Jul 2013 08:23:55 GMT

Hi Sergey

Thanks for replying

CPU is E2660, machine is z820 dual CPU 32 cores. So no slouch! Running g windows 7 64 bi

Compiled on vs2010 64 bit

As far as I know I am using auto dispatching.

I will check though.

Take a look at Intel®

SergeyKostrov — Thu, 11 Jul 2013 13:17:54 GMT

Take a look at Intel® Integrated Performance Primitives for Windows* OS User's Guide: ... Page 19 Intel® Integrated Performance Primitives Theory of Operation Dispatching ... There is a table Identification of Codes Associated with Processor-Specific Libraries.

Hi Sergey

BatterseaSteve — Thu, 11 Jul 2013 15:31:31 GMT

Hi Sergey

I run ippInit since I am using static Libs

I check the cpuType from ippGetCpuType and it returns ippCpuAVX which is correct for my CPU I think.

Cheers

Steve

Hi

BatterseaSteve — Fri, 12 Jul 2013 16:04:00 GMT

Hi Digging a bit deeper reveals the problem to be due to the stepSize I am using.

If I run:

ippiSize uyvyRoi = {1,1920*1080};

uyavyaBuffer = ippsMalloc_8u(1920*1080*sizeof(Ipp8u)*3);

uyvyBuffer = ippsMalloc_8u(1920*1080*sizeof(Ipp8u)*2);

ippiCopy_16u_C1R((Ipp8u *)uyavyaBuffer, 3, uyvyBuffer, 2, uyvyRoi);

I get anything from 7-15msecs.

If I run:

ippiSize uyvyRoi = {1920,1080};

uyavyaBuffer = ippsMalloc_8u(1920*1080*sizeof(Ipp8u)*3);

uyvyBuffer = ippsMalloc_8u(1920*1080*sizeof(Ipp8u)*2);

ippiCopy_16u_C1R((Ipp8u *)uyavyaBuffer, 1920*2, uyvyBuffer, 1920*2, uyvyRoi);

I get 0.2msecs......

So the point is that trying to strip the alpha by treating the source buffer as a 3x(1920x1080) buffer and the dest as a 2x(1920x1080) buffer is bad. Which puts me back to square one...... Anyone got any other ideas - 'cause I can write a c loop thats quicker.

Cheers Steve

Hi Steve,

Igor_A_Intel — Mon, 15 Jul 2013 15:37:38 GMT

Hi Steve,

as I've already answered - you should use ippiCopy_8u_C3C1R 2 times (as finaly you need planar format) - this F extracts any channel you need from C3 image (so C4C1 - the same but for C4 image)

regards, Igor

Hi Igor

BatterseaSteve — Mon, 15 Jul 2013 17:49:39 GMT

Hi Igor

Thanks for replying and sorry - I did see your comment before and thought I had replied.

The problem with ippiCopy_8u_C3C1R is that it extracts the UYA VYA image into Y's and UV's.

I tried this as you suggested but I need the Y's the U's and the V's in separate buffers.

So the question remains as to how I extract the UV's into separate U and V buffers?

I have looked at the format conversions to see if there was one that would take NV12 422 and convert to YUV420p but I could not see one.

I've been on this a week now and am now delving into trying to use SSE!

Any thoughts gratefully received?

Cheers

Steve

Hi

BatterseaSteve — Mon, 22 Jul 2013 15:59:11 GMT

I solved this using a bit of bespoke SSE to split the UV buffer into separate U and V buffers.

Total time seems acceptable.

Cheers

Steve