GLSL-To-TGSI For Mesa Is Still Not Merged

[oibaf]

It would be interesting to test it with shader-db: http://lists.freedesktop.org/archive...ay/007694.html

[Plombo]

There is nothing to discuss, they will not merge it until mesa 7.11 gets released

Not necessarily, Ian's message just said it should wait until after 7.11 if the changes would make it harder to cherry-pick changes to 7.11. I'm not convinced that's the case - there are only 2 existing files with major changes in core Mesa, and even those wouldn't make it difficult to backport fixes to 7.11.

I sent a reply to Ian's message to the list yesterday; it'll probably be discussed more if I get it ready before Mesa 7.11 is released on July 22. Which is likely, since all I really have to do is rebase against master and rewrite some of the commit messages. So we'll see.

It might be obvious by now, but I should mention that I've changed my mind - I decided I'd rather just wait until after Mesa 7.11 is released to try and get it merged to master.

[Plombo]

Have you run any tests yourself yet? What kind of differences are you expecting? So far when comparing the glsl-to-tgsi branch to the respective mainline commit that appears to be where you last pulled from, I am not seeing any major differences. On R600g I am seeing a frame or two difference, Nouveau is nearly identical in frame-rates. This is for Nexuiz, Lightsmark, VDrift. There are some CPU usage differences though.

Sorry, I didn't see this post before.

I've run some tests with the nv50 driver on my NVA5 card, and like you said, there's no real difference. This is probably because unlike r600g, the nv50 driver has an optimizing shader compiler.

The performance being similar is the reason that I emphasized GLSL 1.30 support instead of performance in my original message to the list - performace-wise, there are maybe a few gains to be made from GLSL->TGSI over GLSL->Mesa->TGSI, but not incredibly.

I didn't think to test the CPU usage in my tests, though.

[Plombo]

My plan for now is at least r300g, r600g, and nv50. Also to be monitoring CPU and GPU usage during so.

Okay, sounds good. Hopefully your benchmarks will be more interesting than the few I've done on my NVA5.

[Michael]

Not necessarily, Ian's message just said it should wait until after 7.11 if the changes would make it harder to cherry-pick changes to 7.11. I'm not convinced that's the case - there are only 2 existing files with major changes in core Mesa, and even those wouldn't make it difficult to backport fixes to 7.11.

I sent a reply to Ian's message to the list yesterday; it'll probably be discussed more if I get it ready before Mesa 7.11 is released on July 22. Which is likely, since all I really have to do is rebase against master and rewrite some of the commit messages. So we'll see.



If you do, please be sure to test several different Gallium drivers (at least r300g/r600g/nvfx/nv50), since each driver compiles and optimizes the TGSI input differently (and they use different flags to make the compiler run different lowering passes and generate different TGSI).

Have you run any tests yourself yet? What kind of differences are you expecting? So far when comparing the glsl-to-tgsi branch to the respective mainline commit that appears to be where you last pulled from, I am not seeing any major differences. On R600g I am seeing a frame or two difference, Nouveau is nearly identical in frame-rates. This is for Nexuiz, Lightsmark, VDrift. There are some CPU usage differences though.

[Michael]

How do you monitor gpu usage? Does dynpm expose such information through debugfs?

For the Radeon driver, the Phoronix Test Suite can monitor the fence count as a basic way to judge load (unfortunately, doesn't have a percentage-based load indicator like is exposed through Catalyst).

[darkbasic]

How do you monitor gpu usage? Does dynpm expose such information through debugfs?

[Michael]

If you do, please be sure to test several different Gallium drivers (at least r300g/r600g/nvfx/nv50), since each driver compiles and optimizes the TGSI input differently (and they use different flags to make the compiler run different lowering passes and generate different TGSI).

My plan for now is at least r300g, r600g, and nv50. Also to be monitoring CPU and GPU usage during so.

[darkbasic]

I might be running some.

Awesome

[Plombo]

There is nothing to discuss, they will not merge it until mesa 7.11 gets released

Not necessarily, Ian's message just said it should wait until after 7.11 if the changes would make it harder to cherry-pick changes to 7.11. I'm not convinced that's the case - there are only 2 existing files with major changes in core Mesa, and even those wouldn't make it difficult to backport fixes to 7.11.

I sent a reply to Ian's message to the list yesterday; it'll probably be discussed more if I get it ready before Mesa 7.11 is released on July 22. Which is likely, since all I really have to do is rebase against master and rewrite some of the commit messages. So we'll see.

If you have some benchmarks with and without GLSL-To-TGSI it would be much more interesting

I might be running some.

If you do, please be sure to test several different Gallium drivers (at least r300g/r600g/nvfx/nv50), since each driver compiles and optimizes the TGSI input differently (and they use different flags to make the compiler run different lowering passes and generate different TGSI).