[Mesa-dev] [PATCH] nir: add option to lower slt/sge/seq/sne

[Rob Clark]

On Tue, Mar 31, 2015 at 2:20 PM, Matt Turner <mattst88 at gmail.com> wrote:
> On Tue, Mar 31, 2015 at 11:11 AM, Rob Clark <robdclark at gmail.com> wrote:
>> On Tue, Mar 31, 2015 at 2:03 PM, Kenneth Graunke <kenneth at whitecape.org> wrote:
>>> I'm pretty sure you want b2f here, not u2f...the slt/sge/seq/sne opcodes
>>> are defined to return either 0.0 or 1.0.  flt and friends return 0 or
>>> 0xFFFFFFFF.  u2f converts the numerical value of the unsigned source to
>>> float, so this would return 0.0 or 4294967295.0.
>>>
>>
>> hmm, that is a bit sad (since on the flt/etc cases I'd have to
>> multiply by 0xffffffff, which would in turn require a mov for the
>> 0xffffffff or perhaps emitting a driver uniform/const), and since it
>> makes the b2f more complicated..
>
> Are you saying your hardware returns 0/1 for false/true?

yup

> If you can do integer negation as a source mod, you can just negate its uses.

nop, only float negate..  I guess I could do the minus-one thing.  But
in cases where the value is just consumed by another instruction as a
boolean, then the 0/1 thing is good enough.

(but having b2u_recast type op seems like it would give me a good way
to fix things up only if they are needed by an instruction consume
uint/int, so we could abstract better whether bool is 0/1 or 0/~0)

> Alternatively, Jason's implemented a pass to "resolve" booleans to
> these values, which is necessary for us on i965 Gen4 and Gen5. You
> could probably use that.

ok, is that already merged or on list somewhere?

BR,
-R