regarding c66x optimization (w.r.t high register pressure)

on Aug 10 2012 02:45 AM

Hello Team,

I am having one linear assembly file(c64x+ supported instructions) for interpolation functionality for H264 encoder. Currently we are trying to optimize the functionality in C66x platform. So I am trying to use the c66x supported instructions, but most of the cases the loops are getting disqualified because of huge/high register pressure. For example the interpolation function is having MPYSU4, MPYU4 instructions in c64x+ (.sa), loops are getting disqualified because of high register pressure if I try to use DMPYSU4, DMPYU4. Is there any other way to minimize the register pressure?

I went through SPRABG7 document, they mentioned that Avoid 4-way SIMD instructions to reduce likelihood of register pressure. Then how can we use the new instructions which were introduced for c66x platform.

regards,

Narayana

Posted by George Mock

on Aug 12 2012 21:45 PM

Guru52585 points

General recommendation ... Try to find a function in the dsplib library that is similar to yours and see how it does it. Read this post for an idea of what I mean.

Please post your code so we can comment further.

Thanks and regards,

-George

TI C/C++ Compiler Forum Moderator
Please click Verify Answer on the best reply to your question.
The Compiler Wiki answers most common questions.
Track an issue with SDOWP. Enter your bug id in the "Find Record ID" box.

Posted by sankara narayana tankala

on Aug 13 2012 06:33 AM

Prodigy25 points

Hi George,

I think there is no function named with interpolation(), the function which I was looking is already optimized one in c64x+ platform. I want to optimize further by using c66x supported instructions. As I understand we can optimize well the c64x+ optimized routines by using c66x instructions, but I was getting loop disqualification because of high register pressure. I need some more information for reducing the register pressure.

regards,

Narayana

Posted by ran35366

on Aug 13 2012 10:34 AM

Intellectual2325 points

The two techniques that I know to reduce registers pressure are

Do not use SIMND instructions
Do not store constants in registers, rather read them again from the L1 memory (the cache)

I have an example in which a fixed FIR filter coefficients are used. The code reads all coefficients and values from the memory. The compiler was smart enough to put some of the coefficients in registers and not reading them multiple times.

Does it help? please give me feedback

Ran