[pocl] Back @private Scratchpad with GPUCompiler.alloca

[vchuravy]

Summary

Replaces the POCL back-end's MArray-backed @private scratchpad with a direct per-workitem stack allocation via GPUCompiler.alloca. The returned Ptr is wrapped in a CLDeviceArray over OpenCL "Function" storage (LLVM addrspace 0), which is where the SPIR-V target places allocas.

@device_override @inline function KA.Scratchpad(ctx, ::Type{T}, ::Val{Dims}) where {T, Dims}
    ptr = POCL.GPUCompiler.alloca(T, Val(prod(Dims)))
    CLDeviceArray(Dims, reinterpret(POCL.LLVMPtr{T, POCL.AS.Function}, ptr))
end

This drops the StaticArrays dependency from the POCL back-end (StaticArrays is still used by the CPU back-end).

Why

GPUCompiler.alloca emits a real entry-block alloca that the optimizer can promote, in the target's alloca address space — avoiding the unsoundness of llvmcall + alloca and the overhead/semantics of MArray. See the motivation in the companion GPUCompiler PR.

Alignment

The alloca is aligned to Base.datatype_alignment(T), which is exactly the alignment CLDeviceArray uses for its element loads/stores (alignment(::CLDeviceArray{T})), so accesses are consistent. isbits-union element types are intentionally unsupported (GPUCompiler.alloca guards on isbitstype(T)).

Status

Draft — depends on JuliaGPU/GPUCompiler.jl#859 (adds the alloca intrinsic). Project.toml compat is bumped to GPUCompiler = "1.23"; this can be un-drafted once that is merged and released.

Testing

Verified end-to-end against the local GPUCompiler branch: @private Float32 (4,) lowers to alloca [16 x i8], align 4 in addrspace 0 with no surviving julia.gpu.alloca, and the kernel runs correctly on the POCL CPU device.

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 64.73%. Comparing base (e0e969a) to head (92135c8).
⚠️ Report is 3 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #714      +/-   ##
==========================================
- Coverage   64.76%   64.73%   -0.04%     
==========================================
  Files          23       23              
  Lines        2021     2022       +1     
==========================================
  Hits         1309     1309              
- Misses        712      713       +1     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[github-actions]

Benchmark Results

Show table

	main	92135c8...	main / 92135c8...
saxpy/default/Float32/1024	0.0343 ± 0.027 ms	0.0342 ± 0.027 ms	1 ± 1.1
saxpy/default/Float32/1048576	0.411 ± 0.047 ms	0.424 ± 0.062 ms	0.969 ± 0.18
saxpy/default/Float32/16384	0.0591 ± 0.028 ms	0.0564 ± 0.026 ms	1.05 ± 0.69
saxpy/default/Float32/2048	0.0472 ± 0.029 ms	0.043 ± 0.029 ms	1.1 ± 0.99
saxpy/default/Float32/256	31.3 ± 26 μs	30.6 ± 25 μs	1.02 ± 1.2
saxpy/default/Float32/262144	0.125 ± 0.018 ms	0.122 ± 0.018 ms	1.02 ± 0.21
saxpy/default/Float32/32768	0.0636 ± 0.029 ms	0.0637 ± 0.028 ms	0.998 ± 0.64
saxpy/default/Float32/4096	0.0594 ± 0.023 ms	0.0521 ± 0.029 ms	1.14 ± 0.77
saxpy/default/Float32/512	0.0335 ± 0.026 ms	31.5 ± 25 μs	1.07 ± 1.2
saxpy/default/Float32/64	30.8 ± 25 μs	29.8 ± 22 μs	1.03 ± 1.1
saxpy/default/Float32/65536	0.0732 ± 0.029 ms	0.0683 ± 0.029 ms	1.07 ± 0.63
saxpy/default/Float64/1024	0.0495 ± 0.027 ms	0.0348 ± 0.027 ms	1.42 ± 1.4
saxpy/default/Float64/1048576	0.764 ± 0.072 ms	0.79 ± 0.084 ms	0.967 ± 0.14
saxpy/default/Float64/16384	0.0586 ± 0.029 ms	0.0556 ± 0.027 ms	1.05 ± 0.72
saxpy/default/Float64/2048	0.0507 ± 0.029 ms	0.0428 ± 0.03 ms	1.18 ± 1.1
saxpy/default/Float64/256	0.0321 ± 0.026 ms	29.3 ± 20 μs	1.09 ± 1.1
saxpy/default/Float64/262144	0.21 ± 0.027 ms	0.227 ± 0.033 ms	0.925 ± 0.18
saxpy/default/Float64/32768	0.0658 ± 0.031 ms	0.0642 ± 0.03 ms	1.02 ± 0.68
saxpy/default/Float64/4096	0.0573 ± 0.029 ms	0.0511 ± 0.03 ms	1.12 ± 0.86
saxpy/default/Float64/512	0.0327 ± 0.024 ms	30.1 ± 24 μs	1.09 ± 1.2
saxpy/default/Float64/64	30.7 ± 22 μs	29.4 ± 15 μs	1.05 ± 0.92
saxpy/default/Float64/65536	0.0837 ± 0.03 ms	0.0795 ± 0.029 ms	1.05 ± 0.54
saxpy/static workgroup=(1024,)/Float32/1024	0.0367 ± 0.024 ms	0.0346 ± 0.024 ms	1.06 ± 1
saxpy/static workgroup=(1024,)/Float32/1048576	0.476 ± 0.041 ms	0.505 ± 0.048 ms	0.943 ± 0.12
saxpy/static workgroup=(1024,)/Float32/16384	0.062 ± 0.026 ms	0.0597 ± 0.027 ms	1.04 ± 0.65
saxpy/static workgroup=(1024,)/Float32/2048	0.0546 ± 0.021 ms	0.045 ± 0.028 ms	1.21 ± 0.88
saxpy/static workgroup=(1024,)/Float32/256	0.0376 ± 0.024 ms	0.0356 ± 0.022 ms	1.06 ± 0.94
saxpy/static workgroup=(1024,)/Float32/262144	0.159 ± 0.014 ms	0.157 ± 0.015 ms	1.01 ± 0.13
saxpy/static workgroup=(1024,)/Float32/32768	0.0674 ± 0.027 ms	0.0646 ± 0.027 ms	1.04 ± 0.61
saxpy/static workgroup=(1024,)/Float32/4096	0.0619 ± 0.024 ms	0.0505 ± 0.027 ms	1.23 ± 0.8
saxpy/static workgroup=(1024,)/Float32/512	0.0393 ± 0.023 ms	0.0345 ± 0.021 ms	1.14 ± 0.97
saxpy/static workgroup=(1024,)/Float32/64	0.0375 ± 0.023 ms	0.0364 ± 0.023 ms	1.03 ± 0.92
saxpy/static workgroup=(1024,)/Float32/65536	0.0805 ± 0.026 ms	0.0758 ± 0.025 ms	1.06 ± 0.48
saxpy/static workgroup=(1024,)/Float64/1024	0.0491 ± 0.024 ms	0.0474 ± 0.024 ms	1.04 ± 0.74
saxpy/static workgroup=(1024,)/Float64/1048576	0.818 ± 0.089 ms	0.862 ± 0.1 ms	0.949 ± 0.15
saxpy/static workgroup=(1024,)/Float64/16384	0.0594 ± 0.027 ms	0.0575 ± 0.024 ms	1.03 ± 0.64
saxpy/static workgroup=(1024,)/Float64/2048	0.0489 ± 0.028 ms	0.0433 ± 0.027 ms	1.13 ± 0.96
saxpy/static workgroup=(1024,)/Float64/256	0.039 ± 0.023 ms	0.038 ± 0.023 ms	1.03 ± 0.87
saxpy/static workgroup=(1024,)/Float64/262144	0.223 ± 0.026 ms	0.251 ± 0.032 ms	0.887 ± 0.15
saxpy/static workgroup=(1024,)/Float64/32768	0.0701 ± 0.028 ms	0.0665 ± 0.026 ms	1.05 ± 0.59
saxpy/static workgroup=(1024,)/Float64/4096	0.0571 ± 0.025 ms	0.0573 ± 0.025 ms	0.996 ± 0.62
saxpy/static workgroup=(1024,)/Float64/512	0.0392 ± 0.024 ms	0.0342 ± 0.02 ms	1.15 ± 0.97
saxpy/static workgroup=(1024,)/Float64/64	0.0375 ± 0.024 ms	0.0386 ± 0.023 ms	0.972 ± 0.84
saxpy/static workgroup=(1024,)/Float64/65536	0.0886 ± 0.024 ms	0.0815 ± 0.022 ms	1.09 ± 0.42
time_to_load	1.15 ± 0.013 s	0.95 ± 0.013 s	1.21 ± 0.021

Benchmark Plots

A plot of the benchmark results have been uploaded as an artifact to the workflow run for this PR.
Go to "Actions"->"Benchmark a pull request"->[the most recent run]->"Artifacts" (at the bottom).

[christiangnrd]

Julia 1.11/1.13 on linux arm test failures unrelated