I'm curious about how the shader code gets compiled with Mantle. I'd assume they'd still need to use the DX or GL shader compiler to do the front-end compilation, which means if you use Microsoft's shader language, you'd still need to use Microsoft's front end compiler. Either that or AMD has had to duplicate the compiler front end?
AMD Mantle API [updating]
- Thread starter MarkoIt
- Start date
Yes, I'm aware of that. But HLSL is a MS proprietary language. Does Mantle call DX APIs to compile HLSL? Or did AMD license the DX compiler to embed it in the Mantle runtime? Or did AMD clone the HLSL compiler to create their own, duplicate implementation? That's what I'm curious about.
They can compile HLSL also to their ISA directly without going through the DX ASM. I mean, there is no copyright stopping AMD from doing that. Some AMD tools (their shader/kernel analyzer jump to my mind) do that already, iirc.
Are you sure the tool isn't calling DX in the background to compile HLSL to DX ASM first?
I'm not an expert, but if you think about it, this may be a bit silly discussion. The way I understand it, AMD writes the drivers that implement the DirectX standards that Microsoft is basically only the curator of. So instead of talking to Its own DirectX compatible layer they can talk to mantle, which probably even shares a subset of the code for the DirectX HLSL implementation that AMD writes themselves, but with less overhead and more hardware specific features.
Some thoughts on Mantle and low level graphics APIs
When AMD announced Mantle, my first reaction to the idea of yet another graphics API, and a low level one at that, was negative. Like many others, I had visions of the problems in the early GPU days like happened with Glide.
After giving it some thought, I now feel very differently, and I am glad that AMD is heading in this direction (and perhaps NVidia as well in the not too distant future). These are my reasons:
The current gaming and 3d graphics market is completely different from the early days of 3dFX and Glide. In those days, there was one primary platform for 3d graphics and that was the PC (consoles were in their very early stages). There were more 3d hardware vendors than I could count (or remember) and no common graphics API. There was, however, one primary hardware platform and one primary operating system. Under these circumstances, dozens of vendor specific APIs all trying to operate on the same hardware and OS platform does not make good sense, and Glide was an example. Also, games were created from the ground up and the graphics engine was a custom part of each game, which would make supporting dozens of APIs, especially for the same platform unthinkable.
Today, the situation is reversed. There are only a few 3d graphics hardware vendors remaining (NVidia, AMD, Intel, Imagination, etc). There are now many different OS and hardware gaming platforms (Windows, Linux, IOS, Android, Playstation, and XBox for example). Also most games use commercial game engines written by just a few companies and used for many different games by many different game studios (CryEngine, UnReal Engine, Source Engine, etc). In such an environment, the game studios are largely isolated from the 3d graphics API, and its mostly the game engine companies that need to deal with it. As it is, they need to port their engines to a broad variety of OSs and hardware platforms. Having a few common low level APIs, specific to each GPU vendor, would not only greatly speed up the graphics, and add new features, but would probably actually lower the porting work by the engine companies, since the GPU companies are putting their graphics hardware on multiple OS/hardware platforms.
So it very likely will be a win, win, win, win for the GPU vendors, the gaming platforms, the game studio/engine companies, and the game customer. Better performance, more features, fewer ports, and more platforms.
When AMD announced Mantle, my first reaction to the idea of yet another graphics API, and a low level one at that, was negative. Like many others, I had visions of the problems in the early GPU days like happened with Glide.
After giving it some thought, I now feel very differently, and I am glad that AMD is heading in this direction (and perhaps NVidia as well in the not too distant future). These are my reasons:
The current gaming and 3d graphics market is completely different from the early days of 3dFX and Glide. In those days, there was one primary platform for 3d graphics and that was the PC (consoles were in their very early stages). There were more 3d hardware vendors than I could count (or remember) and no common graphics API. There was, however, one primary hardware platform and one primary operating system. Under these circumstances, dozens of vendor specific APIs all trying to operate on the same hardware and OS platform does not make good sense, and Glide was an example. Also, games were created from the ground up and the graphics engine was a custom part of each game, which would make supporting dozens of APIs, especially for the same platform unthinkable.
Today, the situation is reversed. There are only a few 3d graphics hardware vendors remaining (NVidia, AMD, Intel, Imagination, etc). There are now many different OS and hardware gaming platforms (Windows, Linux, IOS, Android, Playstation, and XBox for example). Also most games use commercial game engines written by just a few companies and used for many different games by many different game studios (CryEngine, UnReal Engine, Source Engine, etc). In such an environment, the game studios are largely isolated from the 3d graphics API, and its mostly the game engine companies that need to deal with it. As it is, they need to port their engines to a broad variety of OSs and hardware platforms. Having a few common low level APIs, specific to each GPU vendor, would not only greatly speed up the graphics, and add new features, but would probably actually lower the porting work by the engine companies, since the GPU companies are putting their graphics hardware on multiple OS/hardware platforms.
So it very likely will be a win, win, win, win for the GPU vendors, the gaming platforms, the game studio/engine companies, and the game customer. Better performance, more features, fewer ports, and more platforms.
Beyond Mantle
Taking it a step further, if GPU vendors start providing their own low level graphics APIs that are optimized for their hardware and available across multiple gaming platforms, then it is only a small step further for the GPU companies to add a gaming specific CPU on the GPU die and run the entire game on the GPU with it's ultra high speed memory.
This is almost the inverse of the APU, where the GPU is an add on to the CPU and is crippled by the CPUs memory and other restrictions.
In this way, a game (and entire game engine) would port completely from one hardware/OS plaform to another, with no code changes. In addition, the performance improvements and optimizations would be dramatic. The GPU specific API (Mantle, etc) would be adapted to provide a complete game production environment. It would be best, of course, if the GPU manufacturers selected the same open CPU (ARM almost certainly) to imbed in their GPUs and an agreed upon common set of graphics/game development tools.
Taking it a step further, if GPU vendors start providing their own low level graphics APIs that are optimized for their hardware and available across multiple gaming platforms, then it is only a small step further for the GPU companies to add a gaming specific CPU on the GPU die and run the entire game on the GPU with it's ultra high speed memory.
This is almost the inverse of the APU, where the GPU is an add on to the CPU and is crippled by the CPUs memory and other restrictions.
In this way, a game (and entire game engine) would port completely from one hardware/OS plaform to another, with no code changes. In addition, the performance improvements and optimizations would be dramatic. The GPU specific API (Mantle, etc) would be adapted to provide a complete game production environment. It would be best, of course, if the GPU manufacturers selected the same open CPU (ARM almost certainly) to imbed in their GPUs and an agreed upon common set of graphics/game development tools.
Recent rulings from courts in the US and EU has establish that APIs and computer languages are not copyrightable.
Which is definitely a good thing. I'm less interested in the legal aspects, more in the political and technical aspects:
If Mantle calls the DX API behind the scenes to do compilation, it ties Mantle to MS' DX implementation.
If AMD wrote their own HLSL front end, it's potentially quite a bit of work to maintain it. Compilers can be rather subtle, especially if you're tasked with reimplementing one and ensuring the behavior of your clone is identical to the original.
I could be wrong, but I thought DX drivers do not process HLSL, only MS' IR after the DX runtime compiles it.
I don't get why you think this is some difficult feat. This is essentially what happens with GLSL (OpenGL) already. I don't think skipping MS' IR is a very hard problem to solve (and as Gipsel has noted, they probably already "solved it" for their tools).
Who knows? But it's still not rocket science even then. Even the Vita has Shader Model 3.0'+', which is afaik pretty much the same thing.
That part of the interview from the Digital Foundry had with two lead architect behind the xb1, makes me think that Mantle may not last for too long:
Though it still sounds like AMD would be to benefit from its work with MSFT on XB1 (time to market whereas Nvidia would have to play catch up really fast).
Last edited by a moderator:
I dont think shader compilation is a bottleneck so it doent matter if the dx runtime has to compile them
plus they can be pre compiled like battlefield 2 does.
plus they can be pre compiled like battlefield 2 does.
soz for the ot but how come you need a x86 license, its only an instruction set after all ?
If you want to provide a design compatible with an x86 20 years ago, it might be technically possible, although you might not be allowed to link any marketing to existing product trademarks.
If you want to make a performant and modern x86, the many implementation details can be patented or certain design features protected for a period of time.
Back on topic, there is no requirement that you get permission to code in x86 assembly.
Well, Cyrix for example didn't need Intel's blessing, but they had to do all the legwork to make a x86-compatible architecture from the ground-up, carefully avoiding specific patent-protected parts.
Of course, that was during the 80s and 90s. Now the technology is so much more complicated that square-one start up like that is a pure madness.
Looks like AMD will do roughly the same with the OpenGL extensions.
But I admit, it's made up from a few tweets.
Jawed
Legend
AMD's older tools definitely called the D3D compiler for HLSL. There were options for the optimisation flags to use when calling the D3D compiler.
But that's all rather by the by, since there's essentially no meaningful difference between: HLSL, GLSL and OpenCL.
That isn't to say AMD is no longer capable of screwing up compilation. But no just-in-time compiler produces reliably good results.
It really would be nice to stop prioritising JIT compilation. I've been moaning about this for years now and there's still far too much junk coming out of them.
In fact I would go so far as to say that after 2 years, the "easy to build compiler tools" GCN is failing miserably at producing lean code. Simple concepts like holding register allocation below the magic figure of 128 still seem to be utterly alien. My mind boggles: "uh yeah, we don't give a damn if suddenly your code runs at half speed because we tripped over the 128-line".
Yeah it makes you wonder why that wasn't always the delivery mechanism, as I asked earlier in the thread. My guess is - like with all big companies - differing teams/opinions internally. In any case, exposing stuff in GL is clearly a good path, so looking forward to seeing the extensions.
