[Mesa-dev] [RFC] ARB_gl_spirv and NIR backend for radeonsi

[Marek Olšák]

On Tue, May 23, 2017 at 12:07 AM, Timothy Arceri <tarceri at itsqueeze.com> wrote:
> On 23/05/17 05:02, Marek Olšák wrote:
>>
>> On Sun, May 21, 2017 at 12:48 PM, Nicolai Hähnle <nhaehnle at gmail.com>
>> wrote:
>>>
>>> Hi all,
>>>
>>> I've been looking into ARB_gl_spirv for radeonsi. I don't fancy
>>> re-inventing
>>> the ~8k LOC of src/compiler/spirv, and there's already a perfectly fine
>>> SPIR-V -> NIR -> LLVM compiler pipeline in radv, so I looked into
>>> re-using
>>> that.
>>>
>>> It's not entirely straightforward because radeonsi and radv use different
>>> "ABIs" for their shaders, i.e. prolog/epilog shader parts, different user
>>> SGPR allocations, descriptor loads work differently (obviously), and so
>>> on.
>>>
>>> Still, it's possible to separate the ABI from the meat of the NIR -> LLVM
>>> translation. So here goes...
>>>
>>>
>>> The Step-by-Step Plan
>>> =====================
>>>
>>> 1. Add an optional GLSL-to-NIR path (controlled by R600_DEBUG=nir) for
>>> very
>>> simple VS-PS pipelines.
>>>
>>> 2. Add GL_ARB_gl_spirv support to Mesa and test it on simple VS-PS
>>> pipelines.
>>>
>>> 3. Fill in all the rest:
>>> 3a. GL 4.x shader extensions (SSBOs, images, atomics, ...)
>>> 3b. Geometry and tessellation shaders
>>> 3c. Compute shaders
>>> 3d. Tests
>>>
>>> I've started with step 1 and got basic GLSL 1.30-level vertex shaders
>>> working via NIR. The code is here:
>>> https://cgit.freedesktop.org/~nh/mesa/log/?h=nir
>>>
>>> The basic approach is to introduce `struct ac_shader_abi' to capture the
>>> differences between radeonsi and radv. In the end, the entry point for
>>> NIR
>>> -> LLVM translation will simply be:
>>>
>>>     void ac_nir_translate(struct ac_llvm_context *ac,
>>>                           struct ac_shader_abi *abi,
>>>                           struct nir_shader *nir);
>>>
>>> Setting up the LLVM function with its parameters is still considered part
>>> of
>>> the driver.
>>
>>
>> This sounds good.
>>
>>>
>>>
>>> Questions
>>> =========
>>>
>>> 1. How do we get good test coverage?
>>> ------------------------------------
>>> A natural candidate would be to add a SPIR-V execution mode for the
>>> piglit
>>> shader_runner. That is, use build scripts to extract shaders from
>>> shader_test files and feed them through glslang to get spv files, and
>>> then
>>> load those from shader_runner if a `-spirv' flag is passed on the command
>>> line.
>>>
>>> This immediately runs into the difficulty that GL_ARB_gl_spirv wants SSO
>>> linking semantics, and I'm pretty sure the majority of shader_test files
>>> don't support that -- if only because they don't set a location on the
>>> fragment shader color output.
>>>
>>> Some ideas:
>>> 1. Add a GL_MESA_spirv_link_by_name extension
>>> 2. Have glslang add the locations for us (probably difficult because
>>> glslang
>>> seems to be focused on one shader stage at a time.)
>>> 3. Hack something together in the shader_test-to-spv build scripts via
>>> regular expressions (and now we have two problems? :-) )
>>> 4. Other ideas?
>>
>>
>> We have plenty of GLSL SSO shader tests in shader-db, but we can only
>> compile-test them.
>>
>> Initially I think we can convert a few shader tests to SSO manually
>> and use those.
>>
>>>
>>>
>>> 2. What's the Gallium interface?
>>> --------------------------------
>>> Specifically, does it pass SPIR-V or NIR?
>>>
>>> I'm leaning towards NIR, because then specialization, mapping of uniform
>>> locations, atomics, etc. can be done entirely in st/mesa.
>>>
>>> On the other hand, Pierre Moreau's work passes SPIR-V directly. On the
>>> third
>>> hand, it wouldn't be the first time that clover does things differently.
>>
>>
>> If you passed SPIR-V to radeonsi and let radeonsi do SPIR-V -> NIR ->
>> LLVM, you wouldn't need the serialization capability in NIR. You can
>> just use SPIR-V as the shader binary and the major NIR disadvantage is
>> gone. Also, you won't have to touch GLSL-to-NIR, and the radeonsi
>> shader cache will continue working as-is.
>>
>> However, I don't know how much GL awareness is required for doing
>> SPIR-V -> NIR in radeonsi. Additional GL-specific information might
>> have to be added to SPIR-V by st/mesa for the conversion to be doable.
>> You probably know better.
>>
>> st/mesa or core Mesa just needs to fill gl_program, gl_shader, and
>> gl_shader_program by parsing SPIR-V and not relying on NIR. I don't
>> know how feasible that is, but it seems to be the only thing needed in
>> shared code.
>>
>> That also answers the NIR vs TGSI debate for the shader cache. The
>> answer is: Neither.
>>
>
> Just to list some downsides to this approach, not switching the the GLSL
> path to also use NIR has the following negatives:
>
> 1. We don't get to leverage the large GL test suits and app ecosystem for
> testing the NIR -> LLVM pass both during development and afterwards.
>
> 2. Jason has already said it best so to quote his reply:
> "There have been a variety of different discussions over the last few years
> about compiler design choices but we've lacked the ability to get any good
> apples-to-apples comparisons.  This may provide some opportunities to do
> so."
>
> 3. The GLSL IR opts are both slow and not always optimal (possibly
> transforming the code to something that's harder to opt later), but due to
> uniform/varying optimisation requirements some optimisations are required
> *before* we can do validation. With NIR we have an opportunity to do these
> optimisations in NIR either by building a nir based linker for the final
> linking stage (uniform/varying validation/location assignment) or by a
> little bit of back and forth of information between NIR and GLSL IR. This is
> something that can't really be done with LLVM/Gallium. I was working towards
> this while at Collabora.
>
> 4. We don't get to drop the glsl_to_tgsi pass which is not the most
> maintenance friendly piece of code. Also currently 10% of cpu is spent in
> the slow tgsi optimisations during start-up of Deus EX which equals around
> 50 seconds on my machine. Most of this optimisation is clean-up simply due
> to how sloppy the glsl_to_tgsi pass is.
>
> 5. It's probably arguable but using GLSL -> NIR should result in more shared
> code paths both between radeonsi/radv and the drivers for other gpus
> anv/freedreno/vc4.
>
> Anyway just a few things to think about.

Using GLSL -> NIR for radeonsi won't really change the GLSL linker
situation, because there are 12 other drivers consuming only TGSI. I
guess it's OK to switch only radeonsi to NIR if it improves compile
times, but we also have the shader cache, so I don't know if it's
worth it just for the faster compilation that takes place only on the
first run. It's very hard to justify the massive development effort
here.

Marek