MVE integer absmax functions can return -1.0 about cmsis-dsp HOT 6 CLOSED

While in there, just noticed that blkCnt is 16-bit for the q15 function, in a probable cut-and-paste error, limiting you to 512K values.

Other notes:

• Using blkCnt and having a tail section seems redundant anyway - wouldn't a simple counter and single predicated loop work better? Produces better code on armclang. And it's what the non-no_idx do.
• curExtremeValVec should be initialised via vdupq_n_u16
• maxValue would be better uint16_t, as it gets the unsigned result from vmaxvq then you can finish with *pResult = __USAT(maxValue, 15).

from cmsis-dsp.

@kjbracey Thanks for good feedback. I'll have a look today.

from cmsis-dsp.

Just to reassure you - I'm not planning on going through and critiquing all the implementations randomly!

I did happen to see this in my debug output, so ended up staring at it. (On audio data I'd just expanded to be full-range - I noticed the scale accuracy problem while doing that).

from cmsis-dsp.

@kjbracey No worry. It is good feedback and helps improve the library.

from cmsis-dsp.

It should be corrected now.

from cmsis-dsp.

Looks good by inspection. Only thought is that clip_q63_to_q31 is overkill.

Could do it with *pResult = maxValue - (maxValue >> 31); (That's a general "clip_u31_to_q31").

Although I guess the compiler might manage to optimise out most of the 64-bitness anyway, so the remaining clip shouldn't be much worse.

from cmsis-dsp.