[PATCH] Refactor and enhance FMA combine

[Mehdi Amini]

> On Mar 25, 2015, at 9:41 PM, Owen Anderson <resistor at mac.com> wrote:
> 
>> 
>> On Mar 24, 2015, at 1:39 PM, Mehdi Amini <mehdi.amini at apple.com <mailto:mehdi.amini at apple.com>> wrote:
>> 
>> 
>>> On Mar 23, 2015, at 9:31 PM, Owen Anderson <resistor at mac.com <mailto:resistor at mac.com>> wrote:
>>> 
>>> 
>>>> On Mar 23, 2015, at 1:48 PM, Mehdi AMINI <mehdi.amini at apple.com <mailto:mehdi.amini at apple.com>> wrote:
>>>> 
>>>>> In principle you're right, that might not be *always* beneficial. But in general, it should be, because even when "high precision" operations are twice more expensive than "low precision" one, the transformation does not worsen things. Right now this is only enabled for PPC, for which low and high precision operations have the same cost. Tell me if this is not acceptable.
>>>> Well you can imagine having more than twice the throughput in f16 than f32 on some targets, and you can also imagine that 2 x f16 operations consume less power than one f32.
>>>> I'd rather have Owen's opinion on this.
>>> 
>>> It’s pretty standard for GPUs to have higher throughput on narrower datatypes.  For instance, if double precision is half the throughput of single precision, then the proposed optimization turns a three cycle sequence into a four cycle sequence.
>> 
>> Not exactly, I believe the proposed optimization turns two “low" and a “high” into two “high”.
>> Note that it seems to me that this optimization can apply if the two low are f16 and the high is a double precision. In pseudo IR code:
>> 
>> %mul = fmul half %u, %v
>> %fma = fma half %x, %y, %mul
>> %fmaext = fpextend half %fma to double
>> %fadd = fadd double %fmaext, %z
>> 
>> becomes:
>> 
>> %xext = fpextend half %x to double
>> %yext = fpextend half %y to double
>> %uext = fpextend half %u to double
>> %vext = fpextend half %v to double
>> 
>> %fma1 = fma double %uext, %vext, %zext
>> %fma = fma double %xext, %yext, %fma1
>> 
>> (assuming that both half and double are legal on the target)
> 
> In that case, this looks more reasonable.  The profitabilty would depend on the ratio of the processor in question, but 2x seems like a pretty common design point.

NVidia GT200 has a 1:8 fp64:fp32 ratio, and the brand new GM200 has a 1:32 ratio :)
But I agree that it is probably not the common case and I’m OK with the added comment so that if anyone has a need to fix it, it should be spottable.

What remains in this revision is the canonicalization that should be done in a specific combine and not here I think.

— 
Mehdi

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20150326/0aa58828/attachment.html>