Switch to POCL

[vchuravy]

The outstanding

• POCL@v7 Switch to pocl_jll@v7 #599
• printf support [nanoPOCL] Re-enable error reporting in 1.12+ #681 Use Printf to report errors from POCL #592
• Switch to pocl_standalone PoCL: Add standalone (OpenCL-less) variant JuliaPackaging/Yggdrasil#11166
  • Renaming issue in build
• Distentagle SPIRVIntrinsics: Currently we inherit quirks from OpenCL https://github.com/JuliaGPU/KernelAbstractions.jl/blob/main/src/pocl/device/quirks.jl but this uses the same method table as OpenCL
• Test OpenCL and KA together to ensure that we have no hidden conflicts.
  • Downstream CI ala Enable downstream CI for 0.10 #608
• [nanoOpenCL] illegal addrspace cast in Julia 1.12 #624
• [pocl] Device side RNG #641 ([pocl] RNG support #659)
• float atomics
• Float16 support
  • Disable Float16 on the CPU backend #609
  • PoCL: Float16 support OpenCL.jl#312
• Windows compilation failure

cc: @maleadt

[JamesWrigley]

Gentle bump, just checking in to see if there's anything a relative noob could help with :) #542 is hitting me again so it'd be cool to get this over the line.

[christiangnrd]

#624 should probably be added to this list

[christiangnrd]

Seems like I missed method table stuff in #635 as uncovered by the buildkite OpenCL run.

[vchuravy]

Huh? That should be fixed post #615 and JuliaGPU/OpenCL.jl#325 each package should use it's own method_table and then stack that on top of the SPIRV method table.

[vchuravy]

@simeonschaub could I bother you to take a look at what it would take to implement device side rand for pocl?

[simeonschaub]

Sure, I can look into it! Shouldn't be too difficult to adapt JuliaGPU/OpenCL.jl#365. It's a little unfortunate to have two copies of the same code, I wonder whether we should eventually move some of the implementation to SPIRVIntrinsics? Not really sure that makes sense though, since the parts are all intertwined.