The performance of the C674x DSP in DM8127/DM8148

tianxing hou

Other Parts Discussed in Thread: SYSBIOS

Hi,

I have a question about the performance of C674x DSP in DM8127/DM8148.

I try use the IMGLIB(for the C64x+) , the performance is inefficient. The result as follow:

And, I run the a function written in C language and a function by intrinsics function(ie. _packh4()) run in the C674x DSP and the DM648, the performance as follow:

Could you give some suggestions about the above question. And some suggestions for improving the performance of C674x DSP in DM8127.

Thank you.

over 12 years ago

0 Anatoliy Sivov over 12 years ago

Prodigy 120 points

I seem to have performance issues as well. Code running on C674x (DM8148) makes 6 times bigger CPU load comparing with the same code on C64+ (DM6443).

The code mostly does memcpy() and array elements access in DDR memory.

So, it looks like I missed some very base thing building DSP server, but I can't get it.

Any help will be appreciated.

0 Marko Shink over 12 years ago in reply to Anatoliy Sivov

Genius 3475 points

Hi,

Could it be that your data is in not-cached memory - check MAR bits.
Have you set compiler -O3 option.

Regards.

0 Yuanwei over 12 years ago

Intellectual 670 points

Interesting.. let me test it and see how it goes.

0 Yuanwei over 12 years ago in reply to Marko Shink

Intellectual 670 points

if no cache enabled, you might want to do edma+L2 sram.

0 li dong over 12 years ago

Prodigy 240 points

hi,

i also met the same problem.

at the same frequency the DM6467 DSP has the high performance to the DM8148 C674 DSP.

0 tianxing hou over 12 years ago in reply to li dong

Expert 2435 points

Hi,

Could you set the memory you accessed as the cacheable?

0 Anatoliy Sivov over 12 years ago in reply to Marko Shink

Prodigy 120 points

I had misconfigured L2 cache, I guess.

Now I set the following configuration:

l1DMode: "32k",
l1PMode: "32k",
l2Mode: "256k"

-o3 is set for both server itself and library (where the source code is situated). I'm not sure about other DSP components.

For sysbios I use

BIOS.libType = BIOS.LibType_Custom;
BIOS.assertsEnabled = false;
BIOS.logsEnabled = false;

and BIOS.customCCOpts = "-mv6740 --abi=eabi -q -mi10 -mo -pdr -pden -pds=238 -pds=880 -pds1110 --program_level_compile -o3 -g --optimize_with_debug";

I could remove -g here, but not sure whether it brings any profit (since I have no BIOS calls in the executed code).

Now I have 2.5 times bigger load on 8148 compared with 6443.

Could you point me to good compiler options reference, because I'm not sure, what all these -pd* mean?

And could you advise me how to check whether asm intrinsics are used for memcpy() and so on?

0 BrandonAzbell over 12 years ago in reply to Anatoliy Sivov

TI__Guru 64510 points

Anatoliy Sivov said:

Could you point me to good compiler options reference, because I'm not sure, what all these -pd* mean?

The C Compiler User's Guide for the particular architecture will have information on what the -pd* switches mean.

http://processors.wiki.ti.com/index.php/TI_Compiler_Information#Compiler_Manuals

Anatoliy Sivov said:

And could you advise me how to check whether asm intrinsics are used for memcpy() and so on?

An inspection of the C run time support library source, found in <CODE GEN TOOLS INSTALL>\lib\rtssrc.zip, can give you a sense of what is implemented.

0 Marko Shink over 12 years ago in reply to BrandonAzbell

Genius 3475 points

You can create assembly listing *.lst output files using options : -al -k
Make sure your data is in internal DSP cached memory, or L2-SRAM.
Instead of memcpy better use much faster EDMA.

Processors

Processors forum

The performance of the C674x DSP in DM8127/DM8148