[pocl] Enable float add and min/max atomics via cl_ext_float_atomics

[vchuravy]

Part of #600 (the last open item: float atomics) — fixes #671.

Problem

Atomix lowers @atomic A[i] += x and @atomic max(A[i], x) on floats to LLVM atomicrmw fadd / fmin / fmax. The SPIR-V backend refuses to translate a module containing them unless the guarding extension has been permitted, so on the CPU backend every float atomic fails to compile:

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

Enzyme's reverse mode hits the first one too, because it accumulates gradients with atomic fadd.

Change

• nanoOpenCL: add the cl_ext_float_atomics constants and expose the per-precision capability bitfields as dev.single_fp_atomic_capabilities, dev.double_fp_atomic_capabilities, dev.half_fp_atomic_capabilities (zero when the extension is absent).
• default_spirv_extensions(dev) permits SPV_EXT_shader_atomic_float_add and SPV_EXT_shader_atomic_float_min_max when the device natively supports the operation for Float32 or Float64, in both global and local memory (same rule as OpenCL.jl's has_fp_atomics). Listing an extension only permits it, so kernels without float atomics are unaffected. An explicit extensions= keyword still wins.
• Testsuite: new Atomics entry running @atomic add/max/min for Int32, UInt32, Float32 and, where supports_float64, Float64 on every backend (skipped when supports_atomics is false). runtests.jl additionally checks that pocl's CPU device gets both extensions and that the override works.

pocl's CPU device reports native add and min/max atomics for single and double precision (global and local), and none for half.

Not in this PR

• Devices without native float atomics still fail with the same LLVM error, since Atomix has no POCL hook to fall back to a compare-and-swap loop (SPIRVIntrinsics 1.1 ships atomic_*_fallback! for that). Not needed for pocl's CPU device.
• Add Atomix and UnsafeAtomics atomics tests #726 adds the broader Atomix/UnsafeAtomics coverage (orderings, fences, swap/replace); it will be rebased on top of this and extend the same test/atomics.jl.
• Enzyme reverse mode additionally needs a SPIRV_LLVM_Backend_jll newer than 22.1.x ([SPIR-V] Rewrite extractvalue over aggregate spv_extractv result llvm/llvm-project#200065 plus one pending fix).

🤖 Generated with Claude Code

https://claude.ai/code/session_01DnYMbgSmZQr7KtRUwei63a

[github-actions]

Benchmark Results

Show table

	main	0d0a4ad...	main / 0d0a4ad...
saxpy/default/Float32/1024	0.0861 ± 0.008 ms	0.085 ± 0.0097 ms	1.01 ± 0.15
saxpy/default/Float32/1048576	0.5 ± 0.025 ms	0.489 ± 0.027 ms	1.02 ± 0.076
saxpy/default/Float32/16384	0.0685 ± 0.035 ms	0.066 ± 0.035 ms	1.04 ± 0.76
saxpy/default/Float32/2048	0.0834 ± 0.027 ms	0.0835 ± 0.027 ms	0.999 ± 0.46
saxpy/default/Float32/256	0.0863 ± 0.0097 ms	0.0857 ± 0.012 ms	1.01 ± 0.18
saxpy/default/Float32/262144	0.178 ± 0.031 ms	0.173 ± 0.03 ms	1.03 ± 0.25
saxpy/default/Float32/32768	0.0739 ± 0.031 ms	0.0721 ± 0.031 ms	1.02 ± 0.61
saxpy/default/Float32/4096	0.0831 ± 0.031 ms	0.085 ± 0.031 ms	0.978 ± 0.51
saxpy/default/Float32/512	0.0853 ± 0.012 ms	0.0852 ± 0.011 ms	1 ± 0.2
saxpy/default/Float32/64	0.087 ± 0.01 ms	0.0859 ± 0.013 ms	1.01 ± 0.2
saxpy/default/Float32/65536	0.09 ± 0.031 ms	0.0858 ± 0.031 ms	1.05 ± 0.53
saxpy/default/Float64/1024	0.085 ± 0.017 ms	0.0853 ± 0.014 ms	0.996 ± 0.26
saxpy/default/Float64/1048576	0.546 ± 0.08 ms	0.557 ± 0.098 ms	0.98 ± 0.22
saxpy/default/Float64/16384	0.0715 ± 0.031 ms	0.0687 ± 0.03 ms	1.04 ± 0.64
saxpy/default/Float64/2048	0.0841 ± 0.029 ms	0.0844 ± 0.029 ms	0.996 ± 0.48
saxpy/default/Float64/256	0.0855 ± 0.0079 ms	0.0859 ± 0.0093 ms	0.995 ± 0.14
saxpy/default/Float64/262144	0.194 ± 0.036 ms	0.192 ± 0.037 ms	1.01 ± 0.27
saxpy/default/Float64/32768	0.079 ± 0.03 ms	0.0763 ± 0.029 ms	1.03 ± 0.55
saxpy/default/Float64/4096	0.0811 ± 0.028 ms	0.0805 ± 0.03 ms	1.01 ± 0.51
saxpy/default/Float64/512	0.0853 ± 0.01 ms	0.0854 ± 0.013 ms	0.999 ± 0.19
saxpy/default/Float64/64	0.0863 ± 0.011 ms	0.0864 ± 0.011 ms	0.999 ± 0.18
saxpy/default/Float64/65536	0.0964 ± 0.03 ms	0.0954 ± 0.03 ms	1.01 ± 0.45
saxpy/static workgroup=(1024,)/Float32/1024	0.0831 ± 0.0099 ms	0.0827 ± 0.012 ms	1 ± 0.19
saxpy/static workgroup=(1024,)/Float32/1048576	0.425 ± 0.03 ms	0.422 ± 0.034 ms	1.01 ± 0.11
saxpy/static workgroup=(1024,)/Float32/16384	0.065 ± 0.032 ms	0.0631 ± 0.032 ms	1.03 ± 0.73
saxpy/static workgroup=(1024,)/Float32/2048	0.0814 ± 0.028 ms	0.083 ± 0.029 ms	0.981 ± 0.48
saxpy/static workgroup=(1024,)/Float32/256	0.0838 ± 0.012 ms	0.0837 ± 0.015 ms	1 ± 0.23
saxpy/static workgroup=(1024,)/Float32/262144	0.156 ± 0.032 ms	0.152 ± 0.033 ms	1.03 ± 0.31
saxpy/static workgroup=(1024,)/Float32/32768	0.0686 ± 0.03 ms	0.0681 ± 0.029 ms	1.01 ± 0.62
saxpy/static workgroup=(1024,)/Float32/4096	0.0818 ± 0.029 ms	0.0822 ± 0.03 ms	0.995 ± 0.51
saxpy/static workgroup=(1024,)/Float32/512	0.0833 ± 0.011 ms	0.0835 ± 0.0021 ms	0.997 ± 0.13
saxpy/static workgroup=(1024,)/Float32/64	0.0837 ± 0.012 ms	0.0842 ± 0.0048 ms	0.994 ± 0.16
saxpy/static workgroup=(1024,)/Float32/65536	0.0819 ± 0.032 ms	0.081 ± 0.032 ms	1.01 ± 0.56
saxpy/static workgroup=(1024,)/Float64/1024	0.0821 ± 0.018 ms	0.0827 ± 0.0036 ms	0.993 ± 0.22
saxpy/static workgroup=(1024,)/Float64/1048576	0.496 ± 0.084 ms	0.495 ± 0.093 ms	1 ± 0.25
saxpy/static workgroup=(1024,)/Float64/16384	0.0662 ± 0.029 ms	0.0635 ± 0.028 ms	1.04 ± 0.64
saxpy/static workgroup=(1024,)/Float64/2048	0.08 ± 0.03 ms	0.0829 ± 0.03 ms	0.964 ± 0.51
saxpy/static workgroup=(1024,)/Float64/256	0.0829 ± 0.011 ms	0.0832 ± 0.0025 ms	0.996 ± 0.13
saxpy/static workgroup=(1024,)/Float64/262144	0.185 ± 0.033 ms	0.186 ± 0.038 ms	0.991 ± 0.27
saxpy/static workgroup=(1024,)/Float64/32768	0.0745 ± 0.027 ms	0.0736 ± 0.028 ms	1.01 ± 0.53
saxpy/static workgroup=(1024,)/Float64/4096	0.0787 ± 0.029 ms	0.0744 ± 0.031 ms	1.06 ± 0.58
saxpy/static workgroup=(1024,)/Float64/512	0.0832 ± 0.0091 ms	0.083 ± 0.0023 ms	1 ± 0.11
saxpy/static workgroup=(1024,)/Float64/64	0.0835 ± 0.015 ms	0.0837 ± 0.005 ms	0.997 ± 0.19
saxpy/static workgroup=(1024,)/Float64/65536	0.0934 ± 0.027 ms	0.0913 ± 0.028 ms	1.02 ± 0.43
time_to_load	0.876 ± 0.00089 s	0.896 ± 0.025 s	0.979 ± 0.027

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).
⚠️ Report is 7 commits behind head on main.

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.

[christiangnrd]

CUDA failure on 1.10/1.11 seem real?