AMD/RTG Announces GPUOpen For Offering More Open-Source Code

[blackiwid]

btw all the performance patches we see in new mesa like hyperZ and so on all are for older gens too. its harder to find improvements that are only for newer amd cards/apus, thats not only catching up to the state of the older driver than the opposite.

did my last comment get eaten by the comment system I wrote about that amd released 2014 a vdpau feature for the free driver for 7 year old hardware to answer to the claim that you get bad support for amd hardware from the opensource guys after 2 years.

there was a "approved" button, but this comment displayed now immidietly, so I am a bit confused

[bridgman]

Yeah, I'm surprised by a lot of the comments I hear about AMD not working on the open source drivers for older chips.

The commit histories suggest something different.

[profoundWHALE]

Wait, so if GCN is open sourced, could we have standard compilers like LLVM or GCC make code that execute on it?

[duby229]

Wait, so if GCN is open sourced, could we have standard compilers like LLVM or GCC make code that execute on it?

I'm confused by this question. All modern graphics drivers need a compiler. Radeonsi has been using LLVM for that purpose for a long time. Of course there are certain things a GPU can't do without access to the CPU's memory, and even then it isn't the right type of architecture for most loads. Once Vulkan gets released and SPIR-V compilers become available, at that point GCN should be able to execute basically anything you can code. It's only a question of whether or not the load you coded is the right fit for the GPU's architecture.

EDIT: OpenGL is particularly high level. The thing is Vulkan is particularly low level for that reason. OpenGL implements functionality, where-as Vulkan only implements resources. In this sense OpenGL dictates functionality, but Vulkan lets game engines implement whatever functionality they can code.

[Nille_kungen]

Yeah, I'm surprised by a lot of the comments I hear about AMD not working on the open source drivers for older chips.

The commit histories suggest something different.

I think some people doesn't understand that development slows down when the driver reaches parity with the blob driver, after that it's mostly hardware limits.
The UVD+ code drop was awesome because it was something that made that old hardware "sing again".
And i don't think that amdgpu hardware will be dropped in the same manner as previous generations because even if it's dropped then that old catalyst version might still install and work on later kernels and xorg.
But personally i don't think i will care because by the time amdgpu supported hardware gets legacy i can't see my self still running catalyst on that hardware generation.

[bridgman]

I'm confused by this question. All modern graphics drivers need a compiler. Radeonsi has been using LLVM for that purpose for a long time. Of course there are certain things a GPU can't do without access to the CPU's memory, and even then it isn't the right type of architecture for most loads.

Right. We've had an open source r600/radeonsi LLVM back-end for a while - the open source graphics stack converts TGSI to LLVM IR and the open source compute stack uses Clang to convert OpenCL C to LLVM IR, then the LLVM stack and back-end generate optimized ISA.

What's new is that we are going to be using that direct-to-ISA back end in more places rather than going through an intermediate representation like AMDIL or HSAIL.

[cbxbiker61]

Wait, so if GCN is open sourced, could we have standard compilers like LLVM or GCC make code that execute on it?

HSA support is coming with GCC6, due in 2016.

[artivision]

Yeah, I'm surprised by a lot of the comments I hear about AMD not working on the open source drivers for older chips.

The commit histories suggest something different.

Mr. Bridgman we are not on Linux to save money, we can pirate Windowz for this reason if we want. We are on free software for other reasons. So first AMD should understand those reasons and then follow the right policy. I prefer to not save money and upgrade from HD5000 to GCN if i can run every game available on Linux, so Linux gets fast an 25% PC share and i then hope for the first AAA open source game. So i think that we all expect is "all games for Vulcan to, or at least D3D11-12 state trackers". All the rest are bubbles!!

[bridgman]

Um... is it possible you were responding to someone else ? The discussion was about the r600 vs radeonsi OpenGL drivers where most people seemed to be criticizing us for "not working on the older open source drivers" (which I thought was untrue).

EDIT - I guess you're saying that OpenGL doesn't really matter at this point ? If so, that's probably true from a "big game developer" POV but I'm not sure it's true from an end-user POV yet -- nor will it be for a while, since I don't expect to see much porting of existing OpenGL games to Vulkan except where the underlying game engine is being used for enough new games to justify porting it.

My feeling was that we need at least a bit more OpenGL (4.3-ish seems to be the sweet spot for latest games) and from that point Vulkan probably becomes more important. I think there's a good chance that the most recent major Linux games *will* appear on Vulkan as well, but what's less clear is whether they will be the pipe-cleaners or whether the initial focus will be on new games.

[duby229]

Um... is it possible you were responding to someone else ? The discussion was about the r600 vs radeonsi OpenGL drivers where most people seemed to be criticizing us for "not working on the older open source drivers" (which I thought was untrue).

EDIT - I guess you're saying that OpenGL doesn't really matter at this point ? If so, that's probably true from a "big game developer" POV but I'm not sure it's true from an end-user POV yet -- nor will it be for a while, since I don't expect to see much porting of existing OpenGL games to Vulkan except where the underlying game engine is being used for enough new games to justify porting it.

My feeling was that we need at least a bit more OpenGL (4.3-ish seems to be the sweet spot for latest games) and from that point Vulkan probably becomes more important. I think there's a good chance that the most recent major Linux games *will* appear on Vulkan as well, but what's less clear is whether they will be the pipe-cleaners or whether the initial focus will be on new games.

Oh, man.... That sounds so awful. I wouldn't want any game I develop to be a pipe cleaner. Hopefully the Vulkan drivers won't be as bad as you make them sound.