[Mesa-dev] 16-bit comparisons in NIR

[Giuseppe Bilotta]

On Sat, Apr 21, 2018 at 2:16 AM, Jason Ekstrand <jason at jlekstrand.net> wrote:
>
> I was thinking about this a bit this morning and it gets even more sticky.
> What happens if you have
>
> bool e = (a < b) && (c < d);
>
> where a and b are 16-bit and c and d are 32-bit?  In this case, one
> comprison has a 32-bit value and one has a 16-bit value and you have to pick
> one for the &&.

In OpenCL C, relational operators don't return a bool, but a signed
integer type with the same size and number of components as the
operands; so, comparing scalar 16-bit data to scalar 16-bit data
(doesn't matter if short, ushort or half) would produce a short, and
comparing scalar 64-bit data to scalar 64-bit data would produce a
long (which in OpenCL C is always 64 bit); the same also holds for
vector types (so e.g. comparing a half8 to a half8 produces a short8).

So, in your example above, what would happen in OpenCL C would be this:

(a < b) would return a short (which is 16-bit in OpenCL C);
(c < d) would return an int (which is 32-bit in OpenCL C);

their logical combination would return a bool, with appropriate type promotion.

I don't suppose something like this could be implemented in NIR?

-- 
Giuseppe "Oblomov" Bilotta