[PATCH] D56900: [Fixed Point Arithmetic] Fixed Point and Integer Conversions

Wed Feb 20 17:53:42 PST 2019

leonardchan added inline comments.

================
Comment at: clang/test/Frontend/fixed_point_conversions.c:437
+  // DEFAULT-NEXT: [[RES:%[a-z0-9]+]] = trunc i39 [[SATMIN]] to i16
+  // DEFAULT-NEXT: store i16 [[RES]], i16* %sat_sa, align 2
+
----------------
ebevhan wrote:
> leonardchan wrote:
> > leonardchan wrote:
> > > ebevhan wrote:
> > > > leonardchan wrote:
> > > > > ebevhan wrote:
> > > > > > Conversions like this are a bit odd. There shouldn't be a need to upsize/upscale the container before the saturation, should there?
> > > > > I see. You're saying that we can just check directly if the value exceeds 255 (or goes under -256) since the range of target values can always fit in the range of source values. Therefore we do not need to cast up since the only reason we would need to is if converting to a type with a greater source range.
> > > > > 
> > > > > In this case, we could technically have a special case for integers where I think we can perform the saturation checks without the initial sign extension. I think it might be simpler though if in `EmitFixedPointConversion`, we treat an integer as a 'zero-scale fixed point number'. I think the reason the upsizing occurs in the first place is so that we satisfy the first FX conversion rule of calculating the result with full precision of both operands.
> > > > > I see. You're saying that we can just check directly if the value exceeds 255 (or goes under -256) since the range of target values can always fit in the range of source values. Therefore we do not need to cast up since the only reason we would need to is if converting to a type with a greater source range.
> > > > 
> > > > That's right. Though, for integer to fixed-point conversion, I don't think you ever need to upcast first; either the integer is larger than or equal to the integral part of the fixed-point type, in which case we can check the magnitude in the type as is, or it's smaller, and we don't have to do any saturation.
> > > > 
> > > > > I think it might be simpler though if in `EmitFixedPointConversion`, we treat an integer as a 'zero-scale fixed point number'.
> > > > 
> > > > Isn't this already the case? The semantics for an integer type are fetched as a zero scale fixed-point type and used that way (except when the target is an integer due to the rounding requirement).
> > > > That's right. Though, for integer to fixed-point conversion, I don't think you ever need to upcast first; either the integer is larger than or equal to the integral part of the fixed-point type, in which case we can check the magnitude in the type as is, or it's smaller, and we don't have to do any saturation.
> > > 
> > > I see. I think this would be more of a matter then of when checking for saturation, we either should check against the source value after resizing and scaling (container), or the source by itself before resizing and scaling. Actually, I think that when comparing saturation against the source value, we could save an instruction since we can just generate a constant  to compare the source against instead of comparing against a potentially shifted value. I think this could work when converting from fixed point types as well.
> > > 
> > > With saturation conversion, (if the target scale >= src scale) currently we could generate up to 7 instructions:
> > > - 1 resize + 1 shift on the result if target scale > src scale
> > > - 1 cmp gt + 1 select for clamping to max
> > > - 1 cmp lt + 1 select for clamping to min
> > > - 1 resize if container width != target width
> > > 
> > > I think we don't need either the first or last resize if the constants that we check against are the respective max's/min's of the target type against the source. I think this deserves a patch on it's own though since it could change a bunch of tests that depend on `EmitFixedPointConversion`.
> > > 
> > > >Isn't this already the case? The semantics for an integer type are fetched as a zero scale fixed-point type and used that way (except when the target is an integer due to the rounding requirement).
> > > 
> > > What I meant by this was that we are already doing the right thing in that we calculate the result with the full precision of both operands.
> > Added this change in D57553 and made it a child of this patch. 
> I think the patch demonstrated to me that this emission can't be optimized in general without breaking the minmax pattern on the saturation, and that is very useful in some cases.
> 
> What I think I'm concerned about is conversions like `int -> _Sat _Fract`, where there really is no point to upscaling at all, since the resulting will either be -1, 0 or ~0.999.
I think upscaling is also necessary in this case if we decide to keep the minmax pattern for all conversion cases, otherwise we'd still be comparing against the result:

```
  %0 = load i32, i32* %i, align 4
  %1 = icmp sgt i32 %0, 1
  %satmax = select i1 %1, i32 32767, i32 %0
  %2 = icmp slt i32 %0, 1
  %satmin = select i1 %2, i32 -32768, i32 %satmax
  store i16 %resize1, i16* %sat_f, align 2
```

However, this takes only 6 instructions whereas the pattern we currently have takes 9:

```
  %0 = load i32, i32* %i, align 4
  %resize = sext i32 %0 to i47
  %upscale = shl i47 %resize, 15
  %1 = icmp sgt i47 %upscale, 32767
  %satmax = select i1 %1, i47 32767, i47 %upscale
  %2 = icmp slt i47 %satmax, -32768
  %satmin = select i1 %2, i47 -32768, i47 %satmax
  %resize1 = trunc i47 %satmin to i16
  store i16 %resize1, i16* %sat_f, align 2
```

We could say that for specific cases, we compare against the source value and others we do minmax.

Repository:
  rC Clang

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D56900/new/

https://reviews.llvm.org/D56900