If we add saturating adds we should follow most architecture conventions and explicitly call it saturating.
On Sunday, March 6, 2016, Andy Walls <
address@hidden> wrote:
On Sun, 2016-03-06 at 16:33 -0500, West, Nathan wrote:
> By the way, if you choose to do this please don't be afraid to ask
> questions; this is a pretty well defined problem, but [...]
Hi Nathan,
Since you mentioned it, a question popped to my mind:
The current add_const blocks let integer overflow happen, so that
integer numbers wrap around. In DSP, saturation math is often what is
desired, not overflow. We'd rather the numbers clip at the max or min
value, rather than experience some huge jump. SIMD instruction sets
provide instructions that perform saturation math, e.g. the MMX
intrinsic __m64 _m_paddsw(__m64 a, __m64 b) for adding 4 sets of 16 bit
integers.
Should the new volk routines, meant to speed up the integer version of
add_const blocks, perform saturation math, or stick with the current
overflow behavior?
> Cheers,
>
> Nathan
Regards,
Andy