How Valve Made L4D2 Faster On Linux Than Windows

[bridgman]

The GLSL compiler compiles to assembly-level (MAD, MUL, TEX, LOG, FRC, LIT, and other assembly-level commands). Then the GPU hardware execute them internally with smaller instructions, atomic operations, possibly micro-instructions (if the GPU has microcode), and others. The compiler doesn't have access to the entire instruction set like you have on a CPU with a software-rasterizer.

If you're saying "the GPU hardware may execute hardware instructions by breaking them down internally to even simpler operations, in the same way that modern x86 CPUs execute x86 instructions by breaking them down to simpler operations" then I guess I agree. That doesn't mean there is an internal instruction set you could potentially program in though.

Even if Nvidia gives you access on the entire instruction-set you still can't use it with any compiler or anything else. Thats because GPUs missing various control and execution units that only a CPU has. For a GPU, programs must be pre-controled, and thats a higher level sub-set.

I don't work for NVidia

That said, I think you'll find that both NVidia and AMD GPUs today are more capable than what you are describing. I know that SI is, and I imagine the corresponding NVidia parts aren't too different.

[bridgman]

Again, the key point here is that GPU vendors can potentially deliver larger performance gains if they're not locked into a fixed instruction set, so there is a strong bias towards maintaining portability by not programming to the (constantly changing) hardware ISA.

That doesn't mean you can't, just that it's not necessarily a good idea in many cases.

[artivision]

Again, the key point here is that GPU vendors can potentially deliver larger performance gains if they're not locked into a fixed instruction set, so there is a strong bias towards maintaining portability by not programming to the (constantly changing) hardware ISA.

That doesn't mean you can't, just that it's not necessarily a good idea in many cases.

If there was that possibility then where is a GCC target for Nvidia or AMD GPUs (for C++ I mean)? Any way to be fair, I read the Nvidia's Docs and Kepler has some Atomic-operations exposure. Probably newer GPU hardware can run even part of the compiler on shaders. Any way Fusion its the next state for GPUs, but not there yet.

[bridgman]

If there was that possibility then where is a GCC target for Nvidia or AMD GPUs (for C++ I mean)?

Most of the GPU-related work is using the LLVM backend rather than the GCC backend, and both NVidia and AMD have released code for that (not sure if the NVidia code targets IL or ISA though; we released code for both and are working to get them upstream into the LLVM project).

I imagine there will be a GCC target at some point (we're already being asked for it) but so far LLVM has been the priority.