Testing Out Mesa's GLSL-To-TGSI Translator

[BenderRodriguez]

On-Topic:

Either Mesa's IR is doing good job or the alternative GLSL is doing a bad job There is almost no difference between them. Also i must admit the CPU usage is a nice thing, it should be used at most tests be it disk tests, filesystem or graphics. It shows us how "efficient" the code is.

Off-Topic:
No matter what is written on this forum Michael get's abused every time, if i were him i'd just tell everyone to f*** *** if they don't like it, i actually feel sorry for all the abuses he is taking. I am with you Michael, you are not perfect always but you don't deserve definitely the abuses from some trolls...

[Qaridarium]

i allready use : GLSL-To-TGSI Translator

the biggest bugs are allready fixed.

the only ugly bug i can tell is a flash window mode black or white flickering bug.

[pingufunkybeat]

I don't know why people are complaining about performance. This work was about reducing complexity and adding features, not about performance.

The extra step through the Mesa IR really isn't such a big deal in the grand scheme of things. But it made things unnecessarily complex, and that's a bad thing.

[BenderRodriguez]

I don't know why people are complaining about performance. This work was about reducing complexity and adding features, not about performance.

The extra step through the Mesa IR really isn't such a big deal in the grand scheme of things. But it made things unnecessarily complex, and that's a bad thing.

Is there any wiki explaining pros and cons of this solution?

[Plombo]

Either Mesa's IR is doing good job or the alternative GLSL is doing a bad job There is almost no difference between them. Also i must admit the CPU usage is a nice thing, it should be used at most tests be it disk tests, filesystem or graphics. It shows us how "efficient" the code is.

It's not really who's doing a "good job" or "bad job". It's that Mesa IR and TGSI are so similar that translating from GLSL->TGSI produces almost the same code as GLSL->Mesa->TGSI.

Also, the CPU usage is actually irrelevant in this particular test, since the translator is only run once for each shader and isn't running during the whole test.

i allready use : GLSL-To-TGSI Translator

the biggest bugs are allready fixed.

the only ugly bug i can tell is a flash window mode black or white flickering bug.

Are you sure that's a problem with glsl_to_tgsi?

I don't know why people are complaining about performance. This work was about reducing complexity and adding features, not about performance.

The extra step through the Mesa IR really isn't such a big deal in the grand scheme of things. But it made things unnecessarily complex, and that's a bad thing.

It's good to know that there's at least one well-informed person in these forums.

Is there any wiki explaining pros and cons of this solution?

No.

[pingufunkybeat]

Plombo, how is this going to affect the work intel is doing on the GLSL compiler?

AFAIK, they don't use TGSI, and have no intention of doing it, and are using Mesa IR directly like before.

[Plombo]

Plombo, how is this going to affect the work intel is doing on the GLSL compiler?

AFAIK, they don't use TGSI, and have no intention of doing it, and are using Mesa IR directly like before.

It won't affect it at all. ir_to_mesa is still there, and it isn't going away any time soon.

By the way, Intel has actually moved away from ir_to_mesa for i965 fragment shaders; they wrote a custom IR backend to code generate directly for the GPU without going through Mesa IR. They still use it for i915 and i965 vertex shaders as far as I know, though.

[smitty3268]

Is there any wiki explaining pros and cons of this solution?

Pros vs Cons:

GLSL IR -> MESA IR -> TGSI

becomes

GLSL IR -> TGSI

So it simplifies gallium by removing an unnecessary step.

The other benefit is that GLSL IR and TGSI (mostly) support GLSL 1.3, while MESA IR is stuck at the 1.2 level. And no one really seemed interested in adding 1.3 support to it.

[Qaridarium]

Are you sure that's a problem with glsl_to_tgsi?

no I'm not-

[oibaf]

It won't affect it at all. ir_to_mesa is still there, and it isn't going away any time soon.

Will mesa_to_tgsi be removed, then?

EDIT: Is glsl-to-tgsi still missing the optimizations done in mesa IR? Would porting them to glsl-to-tgsi improve generated shaders?