Integer overflow for large ROCArray A in AMDGPU.rand!(A)

[williamfgc]

The following error appears for a large ROCArray, 50Kx50K (or larger) Float32 on AMDGPU.rand! It points at this line in random.jl and might be related to using Int32.

To reproduce: Hardware: Crusher MI250X using rocm 5.1.0 and AMDGPU v0.3.7 Code:

import AMDGPU
A = AMDGPU.ROCArray{Float32,2}(undef, 50000, 50000)
AMDGPU.rand!(A)

Error message:

ERROR: LoadError: InexactError: trunc(Int32, 3600000000)
Stacktrace:
  [1] throw_inexacterror(f::Symbol, #unused#::Type{Int32}, val::Int64)
    @ Core ./boot.jl:614
  [2] checked_trunc_sint
    @ ./boot.jl:636 [inlined]
  [3] toInt32
    @ ./boot.jl:673 [inlined]
  [4] Int32
    @ ./boot.jl:763 [inlined]
  [5] convert
    @ ./number.jl:7 [inlined]
  [6] cconvert
    @ ./essentials.jl:412 [inlined]
  [7] macro expansion
    @ ~/.julia/packages/AMDGPU/f6OQx/src/rand/error.jl:44 [inlined]
  [8] rocrand_generate_uniform
    @ ~/.julia/packages/AMDGPU/f6OQx/src/rand/librocrand.jl:42 [inlined]
  [9] rand!
    @ ~/.julia/packages/AMDGPU/f6OQx/src/rand/random.jl:50 [inlined]
 [10] rand!
    @ ~/.julia/packages/AMDGPU/f6OQx/src/random.jl:43 [inlined]
 [11] macro expansion
    @ ./timing.jl:242 [inlined]

I'll be happy to help testing on this system.

[jpsamaroo]

Seems like we might have to work around this manually by checking if overflow would happen, and if so, we perform multiple calls to the generator. @williamfgc would you be willing to implement this (and add a test)?

[williamfgc]

@jpsamaroo thanks for your quick response. I might not have enough bandwidth in the next few weeks if the fix requires some dedicated effort. Happy to test if someone merges a fix during that time, though.

perform multiple calls to the generator

By generator you mean the call to the rocrand_generate_uniform here?

[jpsamaroo]

By generator you mean the call to the rocrand_generate_uniform here?

No, probably here and in similar locations as necessary. We don't want to make the low-level API more complicated, in case users want to directly access those calls for other purposes.

No rush on this, I also don't have much bandwidth while I resolve other issues plaguing AMDGPU.jl :smile: