CUDA: Fallback to UnsafeAtomics

[vchuravy]

No description provided.

[vchuravy]

This of course does not work:

using Atomix
using UnsafeAtomics

using CUDA

function kernel(a, b)
    x = UnsafeAtomics.load(pointer(a, 1))
    UnsafeAtomics.store!(pointer(b, 1), x)
    return nothing
end


b = CUDA.zeros(Int, 1)
a = CUDA.ones(Int, 1)

@cuda kernel(a, b)

Yields

ERROR: LLVM error: Cannot select: 0x1e7845f0: ch = AtomicStore<(store seq_cst (s64) into %ir.4, addrspace 1)> 0x1e784580:1, 0x1e783a90, 0x1e784580, /home/vchuravy/.julia/packages/LLVM/xTJfF/src/interop/base.jl:39 @[ /home/vchuravy/.julia/packages/UnsafeAtomics/vpyYB/ext/UnsafeAtomicsLLVM/atomics.jl:178 @[ /home/vchuravy/.julia/packages/UnsafeAtomics/vpyYB/ext/UnsafeAtomicsLLVM/atomics.jl:131 @[ /home/vchuravy/.julia/packages/UnsafeAtomics/vpyYB/ext/UnsafeAtomicsLLVM/internal.jl:16 @[ /home/vchuravy/.julia/packages/UnsafeAtomics/vpyYB/src/core.jl:8 @[ /home/vchuravy/.julia/packages/UnsafeAtomics/vpyYB/src/core.jl:2 @[ /home/vchuravy/niklas.jl:8 ] ] ] ] ] ]
  0x1e783a90: i64,ch = load<(dereferenceable invariant load (s64) from `i64 addrspace(101)* null`, addrspace 101)> 0x1da29370, TargetExternalSymbol:i64'_Z6kernel13CuDeviceArrayI5Int64Li1ELi1EES1__param_2', undef:i64
    0x1e783f60: i64 = TargetExternalSymbol'_Z6kernel13CuDeviceArrayI5Int64Li1ELi1EES1__param_2'
    0x1e783780: i64 = undef
  0x1e784580: i64,ch = AtomicLoad<(load seq_cst (s64) from %ir.2, addrspace 1)> 0x1da29370, 0x1e783a20, /home/vchuravy/.julia/packages/LLVM/xTJfF/src/interop/base.jl:39 @[ /home/vchuravy/.julia/packages/UnsafeAtomics/vpyYB/ext/UnsafeAtomicsLLVM/atomics.jl:94 @[ /home/vchuravy/.julia/packages/UnsafeAtomics/vpyYB/ext/UnsafeAtomicsLLVM/atomics.jl:94 @[ /home/vchuravy/.julia/packages/UnsafeAtomics/vpyYB/ext/UnsafeAtomicsLLVM/internal.jl:12 @[ /home/vchuravy/.julia/packages/UnsafeAtomics/vpyYB/src/core.jl:7 @[ /home/vchuravy/.julia/packages/UnsafeAtomics/vpyYB/src/core.jl:1 @[ /home/vchuravy/niklas.jl:7 ] ] ] ] ] ]
    0x1e783a20: i64,ch = load<(dereferenceable invariant load (s64) from `i64 addrspace(101)* null`, addrspace 101)> 0x1da29370, TargetExternalSymbol:i64'_Z6kernel13CuDeviceArrayI5Int64Li1ELi1EES1__param_1', undef:i64
      0x1e783860: i64 = TargetExternalSymbol'_Z6kernel13CuDeviceArrayI5Int64Li1ELi1EES1__param_1'
      0x1e783780: i64 = undef
In function: _Z6kernel13CuDeviceArrayI5Int64Li1ELi1EES1_

So seq_cst is not supported...

function kernel(a, b)
    x = UnsafeAtomics.load(pointer(a, 1), UnsafeAtomics.monotonic)
    UnsafeAtomics.store!(pointer(b, 1), x, UnsafeAtomics.monotonic)
    return nothing
end

b = CUDA.zeros(Int, 1)
a = CUDA.ones(Int, 1)

@cuda kernel(a, b)

Does work

[maleadt]

PTX does not support atom.seq_cst. Instead, cmpxchg seq_cst is emulated as "fence.sc; atom.cas"
https://docs.nvidia.com/cuda/parallel-thread-execution/index.html#parallel-synchronization-and-communication-instructions-atom

[vchuravy]

Yeah, the whole order business is a mess.