POCL: fuse muladd into fma on devices with a fused multiply-add

[vchuravy]

Fixes the PoCL half of #865 with today's pocl_standalone_jll. Draft until JuliaGPU/GPUCompiler.jl#1010 is released; this PR should then raise the GPUCompiler compat to that release.

SPIR-V can't carry the contract flag without SPV_KHR_float_controls2, so muladd reaches PoCL as a separate multiply and add, which it never fuses. JuliaGPU/GPUCompiler.jl#1010 can instead fuse what may be contracted into OpenCL's fma before emitting SPIR-V (SPIRVCompilerTarget(; fma = true)). Since OpenCL's fma must be correctly rounded, that is only worth it on a device with a fused multiply-add. This PR:

• adds a single_fp_config device property to nanoOpenCL (CL_DEVICE_SINGLE_FP_CONFIG);
• defaults fma in compiler_config to whether that config has CL_FP_FMA. PoCL sets the bit on x86 only when the CPU has an FMA. The double-precision config isn't used, because OpenCL mandates CL_FP_FMA there, and PoCL's generic fallback always sets it;
• accepts fma as a keyword of compiler_config and @opencl, to override it.

Results

On an AMD Ryzen 9 5950X (AVX2 + FMA), y[i] = muladd(a[i], x[i], y[i]) in a @kernel compiles to a vectorized vfmadd231ps, and the result matches fma.(a, x, y) bit for bit. Without this PR it compiles to vmulps + vaddps. Plain a * x + y stays unfused, and @simd reductions still vectorize (see the GPUCompiler PR for details).

Testing

• New testset "POCL fused multiply-add" checks:
  • that the default follows CL_FP_FMA;
  • that a muladd kernel is fused exactly when it does;
  • that fma = false turns fusion off.
• The full test suite passes locally on PoCL with GPUCompiler#1010 dev'ed.

Relation to #876: this route works with today's PoCL (LLVM 22). #876 (SPV_KHR_float_controls2) needs PoCL with LLVM 23. Once both are in, a device that accepts the extension could leave the contraction to the driver instead.

🤖 Generated with Claude Code

[codecov]

Codecov Report

❌ Patch coverage is 0% with 4 lines in your changes missing coverage. Please review.
✅ Project coverage is 13.27%. Comparing base (b8b5e1a) to head (6149f32).

Files with missing lines	Patch %	Lines
src/pocl/nanoOpenCL.jl	0.00%	4 Missing ⚠️

❗ There is a different number of reports uploaded between BASE (b8b5e1a) and HEAD (6149f32). Click for more details.

HEAD has 60 uploads less than BASE

Flag	BASE (b8b5e1a)	HEAD (6149f32)
	65	5

Additional details and impacted files

@@             Coverage Diff             @@
##             main     #892       +/-   ##
===========================================
- Coverage   82.59%   13.27%   -69.33%     
===========================================
  Files          28       26        -2     
  Lines        2454     2342      -112     
===========================================
- Hits         2027      311     -1716     
- Misses        427     2031     +1604     

☔ 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.