AMD R600g Tessellation Support Lands In Mesa Git

[Nille_kungen]

rm ~/.drirc made a difference but i need to update mesa because i now saw the aruba hang bug.
I will report how it turns out.

[Nille_kungen]

Tried mesa-master_432a798 and it's not hanging anymore and renders as it should and the sliders also works as they should.
I get very low fps from my ARUBA r600 but that's expected i get a lot more from my HAINAN radeonsi.
The low fps r600 moves thru the demo a lot faster then the radeonsi with much higher fps, it's like the i forgot to release the handbrake
Hainan seams to have some dri3 bug that shows to many fps and is slow, with dri2 it's more as expected both fps and speed thru demo, heading over to bugzilla for that.
Thanks a lot for this Dave.
r600 ARUBA https://imgur.com/9WIUwla
radeonsi HAINAN dri3 https://imgur.com/dRpqwjK
radeonsi HAINAN dri2 https://imgur.com/IN8aQLY
I think the fps is to high with the hainan at least much higher then i expected.

[Raven3x7]

That wasn't my problem but
rm ~/.drirc
fixed it for me.

Same here. Had some Unigine specific settings in there, i had completely forgotten about. /facepalm

[Serafean]

Hi, I just compiled mesa git, and tried the Heaven benchmark. It brought my 6950 to its knees... (average of 12 FPS)

I'm wondering : according to official docs, this card has 88 texture units, do I understand correctly that 8 of those are used as "texture renders"?

GLRender::GLRender(): Unknown GPU
OpenGL vendor:   X.Org
OpenGL renderer: Gallium 0.4 on AMD CAYMAN (DRM 2.43.0, LLVM 3.7.0)
OpenGL version:  4.1 (Core Profile) Mesa 11.2.0-devel (git-369afdb)
OpenGL flags:    Core Profile
Found required GL_ARB_map_buffer_range
Found required GL_ARB_vertex_array_object
Found required GL_ARB_draw_instanced
Found required GL_ARB_draw_elements_base_vertex
Found required GL_ARB_transform_feedback
Found required GL_ARB_half_float_vertex
Found required GL_ARB_half_float_pixel
Found required GL_ARB_framebuffer_object
Found required GL_ARB_texture_multisample
Found required GL_ARB_uniform_buffer_object
Found required GL_ARB_geometry_shader4
Found optional GL_ARB_blend_func_extended
Found optional GL_ARB_tessellation_shader
Found optional GL_ARB_shader_bit_encoding
Found optional GL_ARB_sample_shading
Found optional GL_ARB_gpu_shader5
Found optional GL_EXT_texture_compression_s3tc
Found optional GL_ARB_texture_compression_rgtc
Shading language:        4.10
Maximum texture size:    16384
Maximum texture units:   80
Maximum texture renders: 8
...

Benchmark results:
Time:   260.494
Frames: 3123
FPS:    11.9887
Min FPS:        4.89852
Max FPS:        30.2377
Score:  301.997

Second run with DRI3 enabled :

Benchmark results:
Time:   260.222
Frames: 3602
FPS:    13.842
Min FPS:        5.1362
Max FPS:        50.2413
Score:  348.681

[bridgman]

Don't think any of the texture units are "special".

The 6970 has 24 SIMDs with 4 texture units each for a total of 96, 6950 has 22 SIMDs with 4 texture units each for a total of 88.

Don't know why the 80/8 split appears -- for anyone else wondering where the question came from, scroll down to see the following :

Maximum texture units: 80

Maximum texture renders: 8

[Serafean]

bridgman
Thanks for the reply, on a random thread on the internet, someone commented along the lines that the "Maximum texture renders" number refers to the number of ROPs, of which the 6950 should have 32...
to be honest I have no idea what I'm talking about here, just grabbing at straws trying to figure out if something I can see slows the card down...

[bridgman]

Yep, understood. Can you turn tesselation off in the benchmark (IIRC you have off/normal/extreme choices) and see if that makes a difference ?

IIRC tesselation can sometimes spill intermediate results out to memory so there might be some tuning required, not sure.

BTW what resolution / AA / AF are you running at ?

[bridgman]

OK, "maximum texture renders" seems to be 8 on pretty much all hardware out there, which suggests it refers to something like maximum number of textures that can be used to render a specific pixel (remember "multitexturing" ?).

That said, I don't know why you're seeing 80 texture units rather than 88, unless you actually have a 6930 rather than a 6950.

[Serafean]

bridgman thanks for looking into this.
Well, unless lspci is lying, and Sapphire cheated me all those years ago, it is a 6950.

01:00.0 VGA compatible controller: Advanced Micro Devices, Inc. [AMD/ATI] Cayman PRO [Radeon HD 6950]

The original benchmark was no AA, tesselation moderate. All benchmarks are at 1600x1200. All subsequent runs all report the 80 / 8 numbers. Benchmark results follow :

AA=off, no TESS
texture units/renders : 80/8
Time: 260.634
Frames: 8944
FPS: 34.3164
Min FPS: 7.79765
Max FPS: 67.9404


AA = 2x, no TESS
texture units/renders : 80/8
Time: 260.547
Frames: 7970
FPS: 30.5894
Min FPS: 7.64186
Max FPS: 53.3346

AA=off, TESS=moderate
texture units/renders : 80/8
Time: 260.83
Frames: 4301
FPS: 16.4897
Min FPS: 5.54318
Max FPS: 52.8657

AA=2x, TESS = moderate
texture units/renders: 80/8
Time: 260.19
Frames: 3996
FPS: 15.358
Min FPS: 5.46067
Max FPS: 44.0718

The results seem pretty consistent...
I also tried different tesselation modes without benchmarking, 80/8 always appeared.