[PATCH] D152766: AMDGPU: Use correct lowering for llvm.log2.f32

[Matt Arsenault via Phabricator via llvm-commits]

arsenm added inline comments.


================
Comment at: llvm/docs/AMDGPUUsage.rst:963
+                                             half). Not implemented for double. Hardware provides
+                                             1ULP accuracy for float, and 0.51ULP for half. Float
+                                             instruction does not natively support denormal
----------------
arsenm wrote:
> rampitec wrote:
> > foad wrote:
> > > rampitec wrote:
> > > > rampitec wrote:
> > > > > foad wrote:
> > > > > > rampitec wrote:
> > > > > > > rampitec wrote:
> > > > > > > > arsenm wrote:
> > > > > > > > > rampitec wrote:
> > > > > > > > > > Typo 0.51ULP?
> > > > > > > > > That's what the last 3 public ISA docs say
> > > > > > > > For what I know this does not make any sense. Someones finger must be mistyped. @b-sumner can you comment please?
> > > > > > > I.e. I understand 0.5ULP. This is half of a bit, when you do not known how to round it: up or down. But what is 1/100 of a bit? Is that like your computation is 100 bits more accurate, but you do not know how to round it?
> > > > > > It's when the true result is something.51 but the hardware rounds it down, or the true result is something.49 but the hardware rounds it up.
> > > > > > 
> > > > > > I assume for f16 someone exhaustively tested all inputs, measured the largest error from the true result, and decided 0.51 bits was a nice neat upper bound that they could publish.
> > > > > But the representation is in binary, not in decimal. There can be no true result 0.51 decimal to have this rounding problem, it must be 0.51 of a bit. This is the definition of the ULP: 1 bit. We have 0.51 of a bit here.
> > > > I.e. anything above 0.5ULP is 1ULP to me. If your previous mantissa bit is 1 you round it up, if it is 0 you round it down, or whatever rules you have you can only change a single bit. Your 100th bit which you cannot represent doesn't matter. If you can have a single bit error in the lowest mantissa bit, this is 1ULP.
> > > There is a true infinitely precise mathematical result. The documented error bound is the difference from the true mathematical result, not the difference from the correctly rounded result.
> > We certainly do not have a common ground here, so I am calling for @b-sumner again.
> I can also just delete the note about the ULP. It's not exactly great to have to repeat the ISA manual on every instruction 
But yes, the ULP definition is in terms of the true result, not the expected rounded result


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D152766/new/

https://reviews.llvm.org/D152766