By: Michael S (already5chosen.delete@this.yahoo.com), November 17, 2012 6:56 pm
Room: Moderated Discussions
Felid (Felid.delete@this.mailinator.com) on November 17, 2012 5:49 pm wrote:
> Michael S (already5chosen.delete@this.yahoo.com) on November 17, 2012 2:45 pm wrote:
>
> > So, reference manual is lying?
>
> Apart from 1-2 clocks difference, AIDA64 confirmes official numbers. There are 2 instructions (div, sqrt),
> 2 types (int, float) and several sizes of elements (8/16/32/64 bits, SP/DP) and 3 sizes of operands (scalar,
> 128 bit and 256 bit). Not all combinatinos are possible, but all of them are confirmed.
I was asking about integer division, where official manual claims latency of 35~45 for IvyB.
I just measured it myself and my measurements approximately agree with what you said in your previous post - when rdx is non-zero then unsigned division takes ~92 clocks. So, official number is bogus.
> Michael S (already5chosen.delete@this.yahoo.com) on November 17, 2012 2:45 pm wrote:
>
> > So, reference manual is lying?
>
> Apart from 1-2 clocks difference, AIDA64 confirmes official numbers. There are 2 instructions (div, sqrt),
> 2 types (int, float) and several sizes of elements (8/16/32/64 bits, SP/DP) and 3 sizes of operands (scalar,
> 128 bit and 256 bit). Not all combinatinos are possible, but all of them are confirmed.
I was asking about integer division, where official manual claims latency of 35~45 for IvyB.
I just measured it myself and my measurements approximately agree with what you said in your previous post - when rdx is non-zero then unsigned division takes ~92 clocks. So, official number is bogus.
