POCL: contract fast-math flags lost through SPIR-V, no FMAs emitted #865

Activity

1. vchuravy commented on Oct 11, 2026

   

   vchuravy

   on Oct 11, 2026

   MemberAuthor

   More actions

   MWE:

   using KernelAbstractions
   import KernelAbstractions.POCL as P
   
   @kernel function axpy!(y, @Const(a), @Const(x))
       i = @index(Global)
       @inbounds y[i] = muladd(a[i], x[i], y[i])
   end
   y = rand(1024); a = rand(1024); x = rand(1024)
   k = axpy!(POCLBackend(), 64)
   P.@device_code_llvm   k(y, a, x; ndrange=1024)   # fmul contract + fadd contract
   P.@device_code_native k(y, a, x; ndrange=1024)   # OpFMul + OpFAdd, entry point has ContractionOff

   With POCL_CACHE_DIR set, PoCL's .so.o contains vmulsd/vaddsd (and vmulpd/vaddpd), no vfmadd.

   There are two causes:

   1. Base SPIR-V can't carry the per-instruction contract flag. AllowContract only exists in SPV_KHR_float_controls2.
   2. The LLVM SPIR-V back-end adds ContractionOff to every kernel entry point unless the module has !opencl.enable.FP_CONTRACT (SPIRVAsmPrinter::outputExecutionMode). Clang emits that metadata for OpenCL C; GPUCompiler doesn't.

   I tested these variants with a hook in POCL's compiler job (SPIRV_LLVM_Backend_jll 23.1.1, pocl_standalone_jll, x86 with FMA):

   Variant	SPIR-V	PoCL x86
   baseline	FMul + FAdd, ContractionOff	vmulsd/vaddsd
   add !opencl.enable.FP_CONTRACT	FMul + FAdd, no ContractionOff	still vmulsd/vaddsd
   fuse fmul contract + fadd contract into llvm.fmuladd before codegen	OpExtInst fma	vfmadd213sd, vfmadd231pd

   The metadata alone doesn't help. PoCL won't contract plain fmul/fadd, and it would also be the wrong semantics, since it allows contracting any fmul/fadd.

   Proposed fix: run an IR pass in GPUCompiler's SPIR-V finish_ir! (next to lower_minimum_maximum!) that fuses single-use contract-flagged fmul + fadd/fsub into llvm.fmuladd. This fixes all SPIR-V consumers (OpenCL.jl, oneAPI.jl, POCL). One caveat: the back-end lowers fmuladd to OpenCL.std fma, not mad, which is slow on devices without hardware FMA.

   Longer term, emit SPV_KHR_float_controls2 (FPFastMathMode AllowContract) and have PoCL accept it. I'm looking into the PoCL side.
2. vchuravy commented on Oct 11, 2026

   

   vchuravy

   on Oct 11, 2026

   MemberAuthor

   More actions

   PoCL side: pocl/pocl#2393 (draft) enables SPV_KHR_float_controls2 for the CPU devices with LLVM >= 23. With it, FPFastMathMode AllowContract reaches PoCL as contract and a test kernel compiles to vfmadd. To use this from KA, GPUCompiler still needs to pass +SPV_KHR_float_controls2 to the SPIR-V back-end when the device reports it. I haven't yet checked that the back-end then emits AllowContract on each instruction. Until then, the llvm.fmuladd fusion above is the fix that works today.
3. vchuravy

   added this to the 0.10.1 milestone on Oct 11, 2026