[Mesa-dev] [PATCH] gallium/tgsi: clarify (possibly change) TGSI_OPCODE_UCMP definition

[Roland Scheidegger]

Am 08.05.2013 12:47, schrieb Marek Olšák:
> Modifiers are actually very useful with MOV. However I don't think the
> modifiers really care what the type is. They just change the sign bit
> of float (which is a bitwise operation). Also, UCMP doesn't do
> anything with the 2nd and 3rd argument, so their types don't matter.
Well, they do matter for the input modifiers.
They might just change the most significant bit on your hardware but
that's not how they are defined in tgsi.
I don't really care that much if the 2nd and 3rd argument to ucmp are
ints or floats (it's true for d3d10 movc translation floats are the
natural choice but we can easily translate that to several ops instead,
e.g. mov's with modifiers followed by ucmp). I do care however that all
drivers treat them the same, not like now where softpipe will do two's
complement negation on src input modifiers whereas llvmpipe will flip
the sign bit.

> 
> I think the ABS and SUB opcodes should be removed in favor of the modifiers.
Those are in very widespread use and at least for SUB it definitely
avoids additional effort for drivers which can't do input modifiers
natively.

Roland


> 
> Marek
> 
> On Wed, May 8, 2013 at 12:14 PM, Christoph Bumiller
> <e0425955 at student.tuwien.ac.at> wrote:
>> On 08.05.2013 03:48, sroland at vmware.com wrote:
>>> From: Roland Scheidegger <sroland at vmware.com>
>>>
>>> UCMP while an integer opcode isn't really consistently implemented as
>>> having all integer arguments. softpipe will assume all arguments are
>>> ints, whereas gallivm has the arguments defined as untyped which
>>> means they'll get treated as floats. This means input modifiers will
>>> not work the same. Fix this by saying only first arg is an integer,
>>> which seems more useful than making all arguments integers - this would
>>> be similar to d3d10 movc opcode.
>>> ---
>>>  src/gallium/docs/source/tgsi.rst |    5 +++++
>>>  1 file changed, 5 insertions(+)
>>>
>>> diff --git a/src/gallium/docs/source/tgsi.rst b/src/gallium/docs/source/tgsi.rst
>>> index 3af1fb7..852f8a0 100644
>>> --- a/src/gallium/docs/source/tgsi.rst
>>> +++ b/src/gallium/docs/source/tgsi.rst
>>> @@ -1291,6 +1291,11 @@ Support for these opcodes indicated by PIPE_SHADER_CAP_INTEGERS (all of them?)
>>>
>>>  .. opcode:: UCMP - Integer Conditional Move
>>>
>>> +.. note::
>>> +
>>> +   Only the first source arg is an integer, the 2nd and 3rd ones are
>>> +   considered floats (for input modifier purposes).
>>> +
>>
>> As long as you patch up all the occurrences of
>> tgsi_opcode_infer_src_type and make it take an argument to identify the
>> source ...
>>
>> I'd rather just forbid modifiers on moves, i.e. MOV and UCMP, since at
>> least MOV returns TGSI_TYPE_UNTYPED and untyped values can't be operated on.
>> For the ordinary MOV we have NEG and ABS, and for UCMP the backend
>> optimizer can take care of merging modifiers into the instruction
>> (nvc0's UCMP (slct u32) doesn't support modifiers)