MVE arm_scale_q15 does not operate as specified about cmsis-dsp HOT 5 CLOSED

@kjbracey Thank for reporting this. It is a bug.

MVE versions are not strictly identical to the scalar versions and may give a slightly different result. It was required (sometimes) for vectorizing more easily.

But the difference should be minimal. Here, it is clearly not a small difference.

from cmsis-dsp.

@kjbracey It looks like the scalar versions are also not consistent.
q31 is first doing a >> 32 before applying the shift.

So it is not just a consistency issue with MVE versions.

from cmsis-dsp.

So it does. Yet the high-precision intermediate is very clearly specified, and would seem to be a key feature of this specific function.

Although other characteristics might be suitable for most users - eg I replaced my call to q15 with one using a float16 intermediate because always getting a consistent 11 bits of precision was fine.

Maybe there could be an API fork between the "high precision, no loss near unity gain" expensive version as currently specified and a "pretty good precision" performance version.

from cmsis-dsp.

I have changed the MVE implementations for arm_scale_q15 and arm_scale_q7. Now they behave like the scalar version but they take more cycles.

For the q31 (both scalar and MVE) : nothing has changed but I modified the documentation to make to make it clear it is doing a shift by 32 before applying the shift argument so there is a loss of accuracy.

from cmsis-dsp.

Seems reasonable to me. I haven't tested it yet though. Thanks!

from cmsis-dsp.