Benchmarks Of AMD's Newest Gallium3D Driver

[HokTar]

I can understand Qaridarium in the sense that OpenCL would be nice, fortunately Zack restarted his work, so maybe the state tracker will be finished. After that support would be welcome in the drivers.

Porting Wayland to OpenCL is by no means the job of radeon developers, so please stop that Qaridarium.

[bridgman]

If I remember the thread correctly the context was 100 devs targetted at performance improvement, not porting Wayland to OpenCL

[Veerappan]

I can understand Qaridarium in the sense that OpenCL would be nice, fortunately Zack restarted his work, so maybe the state tracker will be finished. After that support would be welcome in the drivers.

Porting Wayland to OpenCL is by no means the job of radeon developers, so please stop that Qaridarium.

Yeah, I noticed that Zack had started committing to the clover branch again a few days ago. This definitely makes me happy, as it means that I'll eventually actually benefit from the results of my own thesis project.

[HokTar]

Yeah, I noticed that Zack had started committing to the clover branch again a few days ago. This definitely makes me happy, as it means that I'll eventually actually benefit from the results of my own thesis project.

So this means you have started working on the vp8 decoder? Would be awesome!

[Veerappan]

So this means you have started working on the vp8 decoder? Would be awesome!

Yeah, I've started work on it. Nothing to show for it yet beyond a configure option and some stub code that gets inserted into the runtime cpu detection blocks. I've basically added in the infrastructure to let me do the OpenCL execution, but haven't really fleshed out the implementation yet.

Before I can really go much further, I'll be doing some more profiling the standard C code and refactoring it a bit to make it more conducive to execution on GPUs, then the real OpenCL fun begins.

As for the implementation, I'm currently doing it with AMD's Stream/OpenCL run-time in Ubuntu, but some of the testing will also happen on Nvidia + Mac/Win.

[89c51]

since it was mentioned using opencl for graphics related stuff (even thought i suspect airlied's post was a bit sarcastic )


does this thing makes sense???

i thought opencl was a framework for tasks that are computation intensive ie decoders, math heavy stuff etc ???

[Pfanne]

since it was mentioned using opencl for graphics related stuff (even thought i suspect airlied's post was a bit sarcastic )


does this thing makes sense???

i thought opencl was a framework for tasks that are computation intensive ie decoders, math heavy stuff etc ???

opencl allows you to do massive parallelized computing.
thats pretty much what you do, whenever you try to get something onto your screen
most of the time you can split the rendering of your screen into many tiny jobs. (sounds ideal for opencl doesnt it)
opencl is probably only viable for raytracing. other stuff should probably still be handled by opengl.

[BlackStar]

it makes sence thats because openCL is the only API who can beat directX9/10/11

you can do better graphics in openCL than directX11 can do it.

thats because you can do raytracing in openCL and directX11 can not handle raytracing.

OpenCL is not a graphics API. It can be used for graphics-related tasks, but it is slower than GL/D3D because it lacks certain fixed-function features (e.g. blending).

Of course you can do raytracing in GL/D3D, what do you think people where using all those years before OpenCL was released?

[elanthis]

you don't unterstand what raytracing is thats because there are no FPS anymore you get unlimied fps on any kind of hardware with raytracing.

I think you have no idea what you're talking about. Actually, I KNOW you have no idea what you're talking about. Par for the course.

The scenes in a ray tracer are still rendered into a single frame which is displayed once fully assembled. There is a very definite FPS involved in the process.

The FPS will be dependent on a similar combination of factors that concerns triangle-mesh rendering, including the destination framebuffer size, number of objects, and complexity of the lighting equations used.

The "RPS" you mention is more or less the same idea as the "triangles per second" or "fragment fill rate" that you have on contemporary 3D rasterization hardware. All it indicates is how complex of a scene the hardware can manage while maintaining a usable FPS.

Carmack mentioned that he hoped that we'd have MIXED MODE renderers in within 3-5 years. These are not actual ray tracing engines, but rather traditional triangle rasterizers that used some extremely simplified and inaccurate ray tracing techniques to compute shadows and lighting on the GPU during rendering rather than on the CPU before rendering. That's it, nothing more.

Also, just to be clear, even if you're right about ray-tracing magically becoming feasible, DirectX is in no way being threatened by OpenCL, because DirectX has DirectCompute -- same damn thing, just a different API and syntax. More games make use of DirectCompute than OpenCL by a huge margin right now, today. (Not for rendering; for physics and such.)

[BlackStar]

you don't unterstand what raytracing is thats because there are no FPS anymore you get unlimied fps on any kind of hardware with raytracing.

on raytracing you have rays per seconds RPS

less RPS means just more black or white pixels left per frame.

The notion of "frame" is independent of the rendering method (rasterization or ray-tracing). Even ray-traced games use frames, as you yourself mention, which means "frames per second" is well defined.

There are very good reasons for using frames instead of displaying each drawing operation on the fly. Seeing each pixel update as it is traced (or rasterized) would be extremely annoying to the user - try it! Modify glxgears to turn off double-buffering (it's a 2 line change). You won't have "frames" anymore but the result won't be pretty.