[pocl] Enable SPV_EXT_shader_atomic_float_add where supported

[vchuravy]

The SPIR-V backend refuses to translate a module containing an atomic fadd unless the extension guarding it has been listed:

LLVM ERROR: The atomic float instruction requires the following SPIR-V extension: SPV_EXT_shader_atomic_float_add

Enzyme's reverse mode runs into this because it accumulates gradients with atomic fadd, so reverse-mode AD over a POCL kernel fails to compile outright.

Change

Derive the extension list from the device, the same way supports_fp16/supports_fp64 already are, keyed off the corresponding OpenCL extension cl_ext_float_atomics (which PoCL advertises on this machine).

Listing an extension only permits it — nothing is emitted unless a module actually needs those instructions — so kernels that don't use float atomics are unaffected. An explicit extensions= keyword still wins, and the new default_spirv_extensions helper gives later extensions an obvious home.

Verification

extensions = "+SPV_EXT_shader_atomic_float_add"     # default, from the device
override   = "+SPV_KHR_expect_assume"               # explicit kwarg still wins
PLAIN KERNEL: PASS
KA POCL FORWARD-MODE ENZYME: PASS  (dA[1:4] = [2.0, 4.0, 6.0, 8.0])

End-to-end, reverse-mode Enzyme over a POCL kernel now runs and produces correct gradients:

KA POCL REVERSE-MODE ENZYME: PASS  (dA[1:4] = [2.0, 4.0, 6.0, 8.0])

That run needed a SPIRV_LLVM_Backend_jll newer than the currently pinned 22.1.8, because reverse mode also trips two LLVM SPIR-V backend bugs:

1. A segfault in SPIRVInstructionSelector::selectExtractVal, already fixed upstream by [SPIR-V] Rewrite extractvalue over aggregate spv_extractv result llvm/llvm-project#200065 (fd1f9225458a, 2026-06-07) — not in 22.1.8. Bumping the JLL past that commit is all that is needed.
2. An OpCompositeExtract pointee-type mismatch on pointers extracted from aggregates, still present on trunk; fix pending upstream.

I verified this by overriding the JLL to point at a patched trunk llc; with the shipped 22.1.8 the same test fails with Failed to translate LLVM code to SPIR-V. This PR is independent of those and is needed regardless — without it the module does not compile even with a fixed backend.

x-ref #583, EnzymeAD/Enzyme.jl#3405, EnzymeAD/Enzyme.jl#3309.

🤖 Generated with Claude Code

[github-actions]

Benchmark Results

Show table

	main	4cc0e6e...	main / 4cc0e6e...
saxpy/default/Float32/1024	0.0619 ± 0.032 ms	0.063 ± 0.03 ms	0.982 ± 0.69
saxpy/default/Float32/1048576	0.515 ± 0.033 ms	0.512 ± 0.02 ms	1.01 ± 0.075
saxpy/default/Float32/16384	0.0545 ± 0.03 ms	0.0569 ± 0.03 ms	0.959 ± 0.73
saxpy/default/Float32/2048	0.0642 ± 0.028 ms	0.0652 ± 0.028 ms	0.986 ± 0.6
saxpy/default/Float32/256	0.0548 ± 0.036 ms	0.0619 ± 0.031 ms	0.885 ± 0.73
saxpy/default/Float32/262144	0.183 ± 0.026 ms	0.183 ± 0.013 ms	0.999 ± 0.16
saxpy/default/Float32/32768	0.0617 ± 0.031 ms	0.0619 ± 0.031 ms	0.997 ± 0.71
saxpy/default/Float32/4096	0.067 ± 0.029 ms	0.0657 ± 0.025 ms	1.02 ± 0.58
saxpy/default/Float32/512	0.0517 ± 0.036 ms	0.0603 ± 0.034 ms	0.856 ± 0.77
saxpy/default/Float32/64	0.0605 ± 0.034 ms	0.061 ± 0.033 ms	0.991 ± 0.78
saxpy/default/Float32/65536	0.0842 ± 0.031 ms	0.0804 ± 0.03 ms	1.05 ± 0.55
saxpy/default/Float64/1024	0.0595 ± 0.033 ms	0.0636 ± 0.03 ms	0.935 ± 0.69
saxpy/default/Float64/1048576	0.67 ± 0.11 ms	0.549 ± 0.16 ms	1.22 ± 0.41
saxpy/default/Float64/16384	0.0583 ± 0.031 ms	0.0671 ± 0.03 ms	0.868 ± 0.6
saxpy/default/Float64/2048	0.0634 ± 0.027 ms	0.0651 ± 0.026 ms	0.973 ± 0.57
saxpy/default/Float64/256	0.0517 ± 0.036 ms	0.0624 ± 0.032 ms	0.829 ± 0.72
saxpy/default/Float64/262144	0.183 ± 0.018 ms	0.193 ± 0.067 ms	0.951 ± 0.34
saxpy/default/Float64/32768	0.0737 ± 0.03 ms	0.0668 ± 0.031 ms	1.1 ± 0.68
saxpy/default/Float64/4096	0.0665 ± 0.027 ms	0.0649 ± 0.027 ms	1.02 ± 0.6
saxpy/default/Float64/512	0.0553 ± 0.036 ms	0.0632 ± 0.031 ms	0.874 ± 0.71
saxpy/default/Float64/64	0.0501 ± 0.036 ms	0.0627 ± 0.032 ms	0.799 ± 0.7
saxpy/default/Float64/65536	0.0971 ± 0.028 ms	0.0867 ± 0.029 ms	1.12 ± 0.5
saxpy/static workgroup=(1024,)/Float32/1024	0.0516 ± 0.036 ms	0.0515 ± 0.036 ms	1 ± 0.98
saxpy/static workgroup=(1024,)/Float32/1048576	0.46 ± 0.024 ms	0.459 ± 0.027 ms	1 ± 0.08
saxpy/static workgroup=(1024,)/Float32/16384	0.0523 ± 0.028 ms	0.052 ± 0.028 ms	1.01 ± 0.77
saxpy/static workgroup=(1024,)/Float32/2048	0.0635 ± 0.027 ms	0.0595 ± 0.033 ms	1.07 ± 0.75
saxpy/static workgroup=(1024,)/Float32/256	0.0479 ± 0.036 ms	0.0611 ± 0.034 ms	0.784 ± 0.73
saxpy/static workgroup=(1024,)/Float32/262144	0.169 ± 0.031 ms	0.163 ± 0.031 ms	1.04 ± 0.27
saxpy/static workgroup=(1024,)/Float32/32768	0.0576 ± 0.029 ms	0.0576 ± 0.029 ms	1 ± 0.72
saxpy/static workgroup=(1024,)/Float32/4096	0.0663 ± 0.026 ms	0.0646 ± 0.024 ms	1.03 ± 0.56
saxpy/static workgroup=(1024,)/Float32/512	0.0441 ± 0.036 ms	0.0578 ± 0.036 ms	0.763 ± 0.79
saxpy/static workgroup=(1024,)/Float32/64	0.0524 ± 0.036 ms	0.0565 ± 0.036 ms	0.927 ± 0.88
saxpy/static workgroup=(1024,)/Float32/65536	0.0754 ± 0.031 ms	0.0732 ± 0.03 ms	1.03 ± 0.6
saxpy/static workgroup=(1024,)/Float64/1024	0.059 ± 0.032 ms	0.0573 ± 0.035 ms	1.03 ± 0.84
saxpy/static workgroup=(1024,)/Float64/1048576	0.486 ± 0.02 ms	0.486 ± 0.019 ms	1 ± 0.058
saxpy/static workgroup=(1024,)/Float64/16384	0.0565 ± 0.028 ms	0.0563 ± 0.028 ms	1 ± 0.71
saxpy/static workgroup=(1024,)/Float64/2048	0.063 ± 0.027 ms	0.0619 ± 0.028 ms	1.02 ± 0.63
saxpy/static workgroup=(1024,)/Float64/256	0.0545 ± 0.036 ms	0.0568 ± 0.036 ms	0.96 ± 0.88
saxpy/static workgroup=(1024,)/Float64/262144	0.172 ± 0.029 ms	0.171 ± 0.03 ms	1.01 ± 0.25
saxpy/static workgroup=(1024,)/Float64/32768	0.0632 ± 0.03 ms	0.0635 ± 0.03 ms	0.995 ± 0.67
saxpy/static workgroup=(1024,)/Float64/4096	0.0651 ± 0.026 ms	0.0637 ± 0.025 ms	1.02 ± 0.57
saxpy/static workgroup=(1024,)/Float64/512	0.0533 ± 0.036 ms	0.0553 ± 0.036 ms	0.964 ± 0.9
saxpy/static workgroup=(1024,)/Float64/64	0.0457 ± 0.036 ms	0.0552 ± 0.036 ms	0.828 ± 0.85
saxpy/static workgroup=(1024,)/Float64/65536	0.0879 ± 0.03 ms	0.0831 ± 0.03 ms	1.06 ± 0.53
time_to_load	0.95 ± 0.0026 s	1.04 ± 0.0049 s	0.909 ± 0.0049

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).

[simeonschaub]

Have you seen JuliaGPU/OpenCL.jl#453? Might want to copy some of the logic from there. SPV_EXT_shader_atomic_float_min_max might be useful too

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 64.89%. Comparing base (65690fd) to head (4cc0e6e).

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #737      +/-   ##
==========================================
+ Coverage   64.77%   64.89%   +0.12%     
==========================================
  Files          24       24              
  Lines        2024     2031       +7     
==========================================
+ Hits         1311     1318       +7     
  Misses        713      713              

☔ 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.