fast math support for OMAPL138 ARM core

Keith Hall

I am using an OMAP L138 and will be running some hard real time maths intensive code on the DSP core and some more general purpose control code on the ARM. I have some simple floating points maths to do (multiply and divide) but the standard rts functions are quite slow. I do not want to get the DSP to do the maths as I do not want to interrupt my real time code. I have been looking for something such as IQMath or fastRTS but cannot find anything for the ARM. The ARM compiler user guide mentions FS_MUL and DIV but I cannot find the library which contains these. It may be just me but the ARM documentation seems very poor, it seems to cover the whole family with no explanation of how things apply to variants or how to configure things in CCS. Has anyone done anything similar?

over 14 years ago

0 Jeff L over 14 years ago

TI__Expert 5960 points

Hi Keith,

I think FS_MUL is part of the C support library that comes with CCS tools and handles the basic multiply operator "*" in C for floating point. You can also find the source code along with the C support lib in rtsc.src or rtscpp.src in an archived library fomat.

There is some HW acceleration inside of the ARM926EJ-S which utilizes armv5te architecture, The E stands for DSP extensions, see this white paper:

http://cadal.cse.nsysu.edu.tw/seminar/seminar_file/2002/10/ARM%20dsp-enhanced%20extensions.pdf

Also arm.com is a great source of info, this article also talks about DSP extensions and SIMD, however SIMD is not available on ARM9:

http://www.arm.com/products/processors/technologies/dsp-simd.php

0 Keith Hall over 14 years ago in reply to Jeff L

Intellectual 385 points

Jeff,

You are correct, I looked at the assembler listing produced by the compiler and the standard rts float multiply (aeabi_fsmul) does call FS_MUL. However FS_MUL does not seem to be using any of the DSP extension instructions from the armv5te document which you mentioned. Looks like I will just have to scale and use integer maths to get the speed unless there are any more ideas, I've searched the ARM website but found nothing else.

Thanks.

Processors

Processors forum

fast math support for OMAPL138 ARM core