Skip to content

mesa/i965: ARB_gl_spirv and ARB_spirv_extensions implementation, plus 4.6

Alejandro Piñeiro requested to merge apinheiro/mesa:review/arb_gl_spirv-arb_spirv_extensions into master

TL; DR; This series implements all the mesa bits needed to enable ARB_gl_spirv and ARB_spirv_extensions on the i965 driver, and with that, expose OpenGL 4.6 on that driver.

Detailed explanation: we reached a point where we consider our development branch good enough to enable both extensions on i965, so we preferred to send the full series, instead of keeping sending subseries of specific sub-features. As a collateral effect of enabling both extensions we can also expose OpenGL 4.6 on i965.

It is worth to note that some of those patches could be easily squashed with some others, but we preferred to keep them, to make easier the review process.

Although on this MR the patch numbering will get quickly obsolete, here an overview of the the patches at the moment this MR got sent:

  • Patches 1-5 provides improvements over current ARB_gl_spirv xfb support. It basically extends the xfb nir gathering pass recently added for Vulkan needs, and then replaces the ARB_gl_spirv custom one and uses it instead.

  • Patches 6-24 implemens UBO/SSBO support, and calls to link them on the i965 driver. It is worth to note that it is using the new deref based UBO/SSBO path, so we could get arrays of arrays properly supported on the spirv to nir pass.

  • Patches 25-33 adds the different resources already supported by the ARB_gl_spirv linker to the resource list (input/output, ubo/ssbo, etc), and fill up properly some data for such resources, in preparation to the program interface queries support.

  • Patches 34-41 fixes and extends several program interfaces queries. In most of the cases they are just to handling the fact that now resource name can be NULL, and applies the value expected for such case defined on the ARB_gl_spirv spec.

  • Patches 42-45 gets serialization working when using SPIR-V shaders. As the previous case, most patches are about handling the fact that names can be NULL, plus initializing UniformDataDefaults when using the ARB_gl_spirv nir linker. All this got interaction with ARB_get_program_binary working (at least with the tests we had an hand).

  • Patch 46 prevents to load the shader cache. Even if we handle the serialization with previous patches, shader cache is still not working with SPIR-V shaders, so we prevent the cache to be used on that case. Taking into account that the ARB_gl_spirv spec doesn't require the cache to be working, we thought that we shouldn't wait to enable the extension until finishing it, and clearly we shouldn't wait for it to send to review.

  • Patches 47-48 add some validations. It is worth to note that ARB_gl_spirv doesn't require too much validation. That includes the validations added with those patches. As with Vulkan, it is assumed that the SPIR-V shaders should be correct. Having said so, the spec also allows to raise them as link errors if you want. We implemented those validations during development, and we think that they are fine and simple enough validations to be included.

  • Patch 49: enable ARB_gl_spirv on i965.

  • Patches 50-55 adds the support for ARB_spirv_extensions, and enable it on i965. FWIW, we sent this several months ago to review. It didn't change anything since then (except minor changes due rebases).

  • Patch 56: exposes OpenGL 4.6 on i965.

So about how this series got tested. First, this series gets all the ARB_gl_spirv and ARB_spirv_extensions CTS tests passing. So in theory that should be enough to enable both extensions. In the practice, those tests were not enough, so we tested this series with piglit patches. For that we had two types of tests:

  • Barebone tests: tests we were adding as we implemented the extension. Piglit master already have some of them.

  • Borrowed tests: recycled tests from other specs. For this we use a heavily modified version of a script wrote by Nicolai Hähnle, that do some fixes on shader_test tests, and then uses glslang to convert them to SPIR-V.

Unfourtunately, right now the previous two are not ready for review (except those that we already sent), as we were focusing on getting the Mesa support done. We are cleaning both this week, so we will send to review all those on the following weeks.

Having said so, right now we get the following outcome when run all the shader piglit tests on SPIR-V mode:

[34596/34596] skip: 6020, pass: 28559, fail: 17, crash: 0

FWIW, this is the outcome on GLSL mode:

[34596/34596] skip: 4877, pass: 29714, fail: 3, crash: 2

Probably the main difference is the greater amount of skipped tests on the SPIR-V run (for example, those crashed tests on the GLSL run are skipped on the SPIR-V run). This increase comes from:

  • Tests that doesn't make sense on ARB_gl_spirv, like old GLSL functionality that is not supported, or name-based validations.

  • Glslang limitations when dealing with some shaders, usually when auto-assigning locations/bindings through stages. Note that for the latter, it is assumed that your GLSL shaders would set explicitly the location/bindings. glslang provide that functionality somewhat experimental, as a way to support already written old shaders (like we have).

For both we try to automatize the skipping as much as possible (on the mentioned script) but we also maintain a exclude list for them.

Edited by Alejandro Piñeiro

Merge request reports