The New R300 Register Allocator Is Still Being Developed

[89c51]

All you can hope for is to have enough open source developers to shadow what the binary driver developers are doing - or at least the most essential parts.

but thats exactly what causes -and will continue to cause if it doesn't get solved- all the problems with the open source stack

lack of manpower

[Qaridarium]

Maybe better get a nv card to play with nouveau instead, much faster

if you pay me the bump-gate harm ? and power consuming costs because of bad hardware..

sure you can start to pay reparations a new nv card and a monthly payment should be fine.

[Kano]

Even a lowend 9400 gt with nouveau beats a my hd 5670. It is about 3-4 times faster with trine and has correct rendering.

[pingufunkybeat]

But isn't this performance advantage only limited to a small range of cards?

I don't know how often you test, but there has been a lot of improvement in the last couple of years. My HD4550 does Prey at high quality and high definition and manages to make it playable.

It's possible that the CPU-boundedness is preventing better frame rates on more powerful cards, though.

[Kano]

As there is no trine benchmark i used openarena to figure out the expected speed with quake live at lowest settings - and had to see that not all drivers are performance tuned, therefore i repeated the tests with settings which disabled vsync (or used max performance mode for nouveau) - a bit hard to read however (2 test systems used, lots of gfx cards/drivers):

http://openbenchmarking.org/result/1...GR-1104128GR35

I also tested my Ion netbook (first 2 tests ran at 800 mhz, the last 2 at 1600 mhz):

http://openbenchmarking.org/result/1...KANO-110419300

[Prescience500]

Something that would help bring more paid developers in would be for parts of Linux open source drivers being used in other things. One of the complaints I've heard leveled at Khronos and OpenGL is that they don't provide a driver component like Microsoft does. What if Khronos used Gallium3d for this component to be provided for all drivers. The only problem is that I doubt this would ever even be considered unless Gallium could approach parity with the blobs. Google might be able to pressure video card makers for their android phones to use first rate open source drivers. If they used mesa and Gallium3d, then we would have more firms helping. Unfortunately, a similar chicken and egg conundrum exists here as my last possibility.

[Kjella]

you don't get an important point... an R700 is an r600 with extras and a southern island chip is an r600 with extras and an northern island chip is an r600 with extras. the general architecture chances are only r300,r600,r900

Then why does HDMI audio work on R600/700 but not on Evergreen or Nothern Island? Why didn't my Evergreen card work at all first, but R600/700 did? Because they're not the same with extras. They're different. Yes, some changes are smaller and some - like general architecture changes - are larger, but they all have changes. Nothing will work on a new generation without an updated driver.

[bridgman]

Then why does HDMI audio work on R600/700 but not on Evergreen or Nothern Island? Why didn't my Evergreen card work at all first, but R600/700 did? Because they're not the same with extras. They're different. Yes, some changes are smaller and some - like general architecture changes - are larger, but they all have changes. Nothing will work on a new generation without an updated driver.

Right. The fundamental difference between CPUs and GPUs here is that CPUs are expected to do the *same* thing every year (run a slowly changing set of x86 binaries) while GPUs are expected to do *different* things every year or two (new graphics features, new pipeline definitions, more programmable functionality for game developers etc..). If CPUs had a million lines of driver code between the x86 binary and the hardware you would be likely to see the same kind of evolution in the exposed CPU hardware -- instead CPUs evolve less (more execution units rather than totally different architecture) and hide that evolving functionality behind an instruction decoder rather than behind a driver.

[Qaridarium]

As there is no trine benchmark i used openarena to figure out the expected speed with quake live at lowest settings - and had to see that not all drivers are performance tuned, therefore i repeated the tests with settings which disabled vsync (or used max performance mode for nouveau) - a bit hard to read however (2 test systems used, lots of gfx cards/drivers):

http://openbenchmarking.org/result/1...GR-1104128GR35

I also tested my Ion netbook (first 2 tests ran at 800 mhz, the last 2 at 1600 mhz):

http://openbenchmarking.org/result/1...KANO-110419300

any benchmark over min-fps=60fps is invalid.

[Qaridarium]

Then why does HDMI audio work on R600/700 but not on Evergreen or Nothern Island? Why didn't my Evergreen card work at all first, but R600/700 did? Because they're not the same with extras. They're different. Yes, some changes are smaller and some - like general architecture changes - are larger, but they all have changes. Nothing will work on a new generation without an updated driver.

you are just making FAIL arguments we talk about graphic core logics and you fuck around with "HDMI audio" AUDIO===???
ooh LOL o come one.

for example hd6870 vs 6950 the architecture is the same because :they just cut of the smal-shader-co-processors (5D logic is 1 bigblog+4 smal co processors) and quating the big-blog 6870 big shader and call the new 4 pices of 6870 a 6950 4D shader block.
on the shader core the 6950 only do have the 6870 shaders (the big one without the smal co processors)

and remember the 6870 only do have the r600 hd 2900 shader architecture means the hd 6950 do have a hd2900 shader but yes they cut of the little shaders for simple shader code.

means an hd6950 is an hd2900 with cutting of the "little shaders"

yes they need to rework the drivers because the code is for bigblock+minishaders