[Mesa-dev] [PATCH 4/5] i965/fs: Return result of image atomic in a register of the expected type.

[Ilia Mirkin]

On Mon, Feb 22, 2016 at 10:09 PM, Francisco Jerez <currojerez at riseup.net> wrote:
> Ilia Mirkin <imirkin at alum.mit.edu> writes:
>
>> On Mon, Feb 22, 2016 at 9:40 PM, Francisco Jerez <currojerez at riseup.net> wrote:
>>> So the result is of float type if we're implementing the float
>>> overload of imageAtomicExchange.  This is the only back-end change
>>> required to support OES_shader_image_atomic AFAICT.
>>> ---
>>>  src/mesa/drivers/dri/i965/brw_fs_surface_builder.cpp | 2 +-
>>>  1 file changed, 1 insertion(+), 1 deletion(-)
>>>
>>> diff --git a/src/mesa/drivers/dri/i965/brw_fs_surface_builder.cpp b/src/mesa/drivers/dri/i965/brw_fs_surface_builder.cpp
>>> index 081dbad..75734d2 100644
>>> --- a/src/mesa/drivers/dri/i965/brw_fs_surface_builder.cpp
>>> +++ b/src/mesa/drivers/dri/i965/brw_fs_surface_builder.cpp
>>> @@ -1125,7 +1125,7 @@ namespace brw {
>>>           if (rsize && pred)
>>>              set_predicate(pred, bld.SEL(tmp, tmp, brw_imm_d(0)));
>>>
>>> -         return tmp;
>>> +         return retype(tmp, src0.type);
>>
>> I had chosen to stick this into emit_typed_atomic -- the typed-ness
>> seemed to fit better there. Your call though, this is just a thought.
>>
> Yeah, I had noticed this difference.  The reason for doing it here is
> that all surface_access::emit_(un)typed_something() calls take a raw UD
> payload as argument and return a raw UD result.  They're just thin
> wrappers around the logical send opcode and they don't necessarily give
> you a return value represented in any of the native types supported by
> the EU, not even for the typed messages which usually still need some
> sort of unpacking, format conversion and bitcast to the correct type.
> The brw::surface_access functions really just deal with binary blobs of
> data represented as vectors of UDs and have no idea what format the data
> is stored in.  emit_image_* OTOH are expected to honour input types and
> preserve type consistency so this really seemed like a bug in there to
> me.

Fair enough. I got thrown off by the "typed" in the emit name :)

  -ilia