RadeonSI May Eventually Switch To NIR Completely

[geearf]

Why are they not translating SPIRV to LLVM IR directly? Weird...

It's already been explained in this topic once, and in about every other IR-related topic.

[duby229]

Why are they not translating SPIRV to LLVM IR directly? Weird...

Optimization pass. Better binary performance. Besides, the driver devs have claimed for years that translating between IR's is very close to free., it doesn't take much time at all. So it's better to take a very slight hit to compile times in order to get better runtime.

At least that is more or less the gist of it.

[log0]

Optimization pass. Better binary performance. Besides, the driver devs have claimed for years that translating between IR's is very close to free., it doesn't take much time at all. So it's better to take a very slight hit to compile times in order to get better runtime.

At least that is more or less the gist of it.

So LLVM own optimization passes suck so much. that it is better to run let NIR optimize the code?

[log0]

It's already been explained in this topic once, and in about every other IR-related topic.

I don't see any explanations, only claims that it is somehow better than using LLVM IR directly.

[duby229]

So LLVM own optimization passes suck so much. that it is better to run let NIR optimize the code?

Not that it sucks, it just is a different format. NIR is structured differently , so it allows types of optimization that LLVM isnt suited to.

[artivision]

I don't see any explanations, only claims that it is somehow better than using LLVM IR directly.

If you want you can create an OGL or D3D state tracker that directly targets HW. You may gain a 3-5% because of no IR but you cannot reduce the job that must be done to each layer, you will just done it at once. If you do that you will have a big minus: there is no more unification. So if you correct something it doesn't correct for every api or driver implementation. Mesa is faster than windowz in native apps because all apis have a unified lower part and all drivers a unified upper part.

[smitty3268]

So LLVM own optimization passes suck so much. that it is better to run let NIR optimize the code?

The explanation that has been given repeatedly is that llvm ir is very low-level. With NIR, they have more information about the program that's being run, and can optimize things that wouldn't be possible in LLVM for that reason. For example, they can optimize across different shader stages, which is something llvm ir has no concept of.

[difron]

The explanation that has been given repeatedly is that llvm ir is very low-level. With NIR, they have more information about the program that's being run, and can optimize things that wouldn't be possible in LLVM for that reason. For example, they can optimize across different shader stages, which is something llvm ir has no concept of.

Does amd isa has concept of shader stages? If not then such optimization is not needed. You just can optimize code as a whole, cant you?
I belive there is not other reason then: "Hey we have already SPIRV TO NIR code wirtten, lets use it."

[duby229]

Does amd isa has concept of shader stages? If not then such optimization is not needed. You just can optimize code as a whole, cant you?
I belive there is not other reason then: "Hey we have already SPIRV TO NIR code wirtten, lets use it."

No doubt that code is already written is part of the reason they want to use it. No sense in starting something new that already exists. AMD GPU's are pipelined so is that what you mean by "stages"? The reason NIR is able to do optimizations that LLVM can't is because it preserves more information about the application running. It doesn't really have anything to do with the GPU hardware. Just the information and the format of that information.

[difron]

No doubt that code is already written is part of the reason they want to use it. No sense in starting something new that already exists. AMD GPU's are pipelined so is that what you mean by "stages"? The reason NIR is able to do optimizations that LLVM can't is because it preserves more information about the application running. It doesn't really have anything to do with the GPU hardware. Just the information and the format of that information.

The question is do we need to preserve this information until hw level. If not then optimization can be done regardless this information.

Edit: Well, to sum up my concern. In real radeonsi driver are there any optimizations in tgsi form which would benefit from switching to nir? Or this optimization benefits are purely hypothetical?