On Wed, 2017-01-18 at 09:50 +1100, Timothy Arceri wrote:
> Previously the last stage would not get optimised until the backend
> did
> its GLSL IR opt loop.
Wait that would be all stages would not get optimised until the backend
called the glsl ir opts. Forgot we worked on a stage at a time for a
Previously the last stage would not get optimised until the backend did
its GLSL IR opt loop. I plan on removing that from i965 shortly which
caused huge regressions in Deus-ex and Tomb Raider which have large
constant arrays. Moving lowering before the opt loop in the GLSL linker
fixes this and un
I would think that the LLVM backend should be able recognize that it's
spilling an array of literals.
Marek
On Thu, Oct 30, 2014 at 1:43 PM, Marek Olšák wrote:
> Hi Tom,
>
> Could you please elaborate on how we should change the TGSI->LLVM translation?
>
> Marek
>
> On Fri, Oct 17, 2014 at 4:49
Hi Tom,
Could you please elaborate on how we should change the TGSI->LLVM translation?
Marek
On Fri, Oct 17, 2014 at 4:49 PM, Tom Stellard wrote:
> On Wed, Oct 15, 2014 at 05:32:11PM -0700, Kenneth Graunke wrote:
>> Consider GLSL code such as:
>>
>>const ivec2 offsets[] =
>> ivec2[](i
On Wed, Oct 15, 2014 at 05:32:11PM -0700, Kenneth Graunke wrote:
> Consider GLSL code such as:
>
>const ivec2 offsets[] =
> ivec2[](ivec2(-1, -1), ivec2(-1, 0), ivec2(-1, 1),
> ivec2(0, -1), ivec2(0, 0), ivec2(0, 1),
> ivec2(1, -1), ivec2(1, 0), ivec2(1, 1
The radeonsi LLVM backend handles this very poorly too. It seems to do
the same thing that i965 does.
Marek
On Thu, Oct 16, 2014 at 2:32 AM, Kenneth Graunke wrote:
> Consider GLSL code such as:
>
>const ivec2 offsets[] =
> ivec2[](ivec2(-1, -1), ivec2(-1, 0), ivec2(-1, 1),
>
Kenneth Graunke writes:
> Consider GLSL code such as:
>
>const ivec2 offsets[] =
> ivec2[](ivec2(-1, -1), ivec2(-1, 0), ivec2(-1, 1),
> ivec2(0, -1), ivec2(0, 0), ivec2(0, 1),
> ivec2(1, -1), ivec2(1, 0), ivec2(1, 1));
>
>ivec2 offset = offsets[];
>
>
I like the idea (when you suggested it in April), but I think this
implementation has some issues. I believe making the constant array
into a uniform will cause it to be shown by the GL API (e.g., returned
by glGetActiveUniform) and counted against uniform limits. Since I
doubt we have any piglit
Consider GLSL code such as:
const ivec2 offsets[] =
ivec2[](ivec2(-1, -1), ivec2(-1, 0), ivec2(-1, 1),
ivec2(0, -1), ivec2(0, 0), ivec2(0, 1),
ivec2(1, -1), ivec2(1, 0), ivec2(1, 1));
ivec2 offset = offsets[];
Both i965 and nv50 currently handle this v
