CuArrays.CURAND.curand missing methods

[mohdibntarek]

There seems to be a few missing methods for curand. The no argument method returns an array, but should it? And the range method is not defined:

julia> x = CuArrays.CURAND.curand()
0-dimensional CuArray{Float32,0}:
0.15776986

julia> x = CuArrays.CURAND.curand(-1:2:1)
ERROR: MethodError: no method matching CuArray{Float32,N} where N(::UndefInitializer, ::StepRange{Int64,Int64})
Closest candidates are:
  CuArray{Float32,N} where N(::UndefInitializer, ::Tuple{Vararg{Int64,N}}) where {T, N} at C:\Users\user\.julia\dev\CuArrays\src\array.jl:45
  CuArray{Float32,N} where N(::UndefInitializer, ::Integer...) where T at C:\Users\user\.julia\dev\CuArrays\src\array.jl:46
  CuArray{Float32,N} where N(::Function, ::Any...) where T at C:\Users\user\.julia\dev\CuArrays\src\array.jl:56
Stacktrace:
 [1] #curand#22(::Base.Iterators.Pairs{Union{},Union{},Tuple{},NamedTuple{(),Tuple{}}}, ::Function, ::Type{Float32}, ::StepRange{Int64,Int64}, ::Vararg{Any,N} where N) at C:\Users\user\.julia\dev\CuArrays\src\rand\highlevel.jl:81
 [2] curand(::Type{Float32}, ::StepRange{Int64,Int64}) at C:\Users\user\.julia\dev\CuArrays\src\rand\highlevel.jl:81
 [3] #curand#26(::Base.Iterators.Pairs{Union{},Union{},Tuple{},NamedTuple{(),Tuple{}}}, ::Function, ::StepRange{Int64,Int64}) at C:\Users\user\.julia\dev\CuArrays\src\rand\highlevel.jl:87
 [4] curand(::StepRange{Int64,Int64}) at C:\Users\user\.julia\dev\CuArrays\src\rand\highlevel.jl:87
 [5] top-level scope at none:0

I am sure there are more, but these are the ones that came up in the context of https://github.com/luca-aki/TraceEstimation.jl, where an input rand function is to be given that is assumed to have the above methods defined.

CC: @luca-aki

[maleadt]

We aim to be identical to Base (see eg. these tests: https://github.com/JuliaGPU/GPUArrays.jl/blob/71a666f27707558d0af6cceb88409d6234e8e74a/src/testsuite/construction.jl#L64-L91), so these should be fixed indeed.

[maleadt]

Upon closer inspection, I'm not sure about this. Scalars are inherently GPU-"incompatible", so I don't think it makes much sense to support that API. CuArrays is all about arrays, I don't think we should implement a GPU-version of a non array API, especially if it's not working efficiently on GPUs at all.

Ranges seem more sensible, but tricky since we don't have any of the Sampler machinery.

[maleadt]

Another one that doesn't seem to work

julia> X = CuArrays.rand(Bool, 5)
[ Info: Building the CUDAnative run-time library for your sm_70 device, this might take a while...
ERROR: InvalidIRError: compiling JuliaGPU/CuArrays.jl#68(CuArrays.CuKernelState, CUDAnative.CuDeviceArray{NTuple{4,UInt32},1,CUDAnative.AS.Global}, CUDAnative.CuDeviceArray{Bool,1,CUDAnative.AS.Global}) resulted in invalid LLVM IR
Reason: unsupported dynamic function invocation (call to floattype)
Stacktrace:
 [1] gpu_rand at /home/tim/Julia/pkg/GPUArrays/src/random.jl:47
 [2] JuliaGPU/CuArrays.jl#68 at /home/tim/Julia/pkg/GPUArrays/src/random.jl:81
Stacktrace:
 [1] check_ir(::CUDAnative.CompilerJob, ::LLVM.Module) at /home/tim/Julia/pkg/CUDAnative/src/compiler/validation.jl:114
 [2] macro expansion at /home/tim/Julia/pkg/CUDAnative/src/compiler/driver.jl:188 [inlined]
 [3] macro expansion at /home/tim/Julia/depot/packages/TimerOutputs/7zSea/src/TimerOutput.jl:216 [inlined]
 [4] codegen(::Symbol, ::CUDAnative.CompilerJob; libraries::Bool, dynamic_parallelism::Bool, optimize::Bool, strip::Bool, strict::Bool) at /home/tim/Julia/pkg/CUDAnative/src/compiler/driver.jl:186
 [5] compile(::Symbol, ::CUDAnative.CompilerJob; libraries::Bool, dynamic_parallelism::Bool, optimize::Bool, strip::Bool, strict::Bool) at /home/tim/Julia/pkg/CUDAnative/src/compiler/driver.jl:47
 [6] #compile#134 at /home/tim/Julia/pkg/CUDAnative/src/compiler/driver.jl:36 [inlined]
 [7] macro expansion at /home/tim/Julia/pkg/CUDAnative/src/execution.jl:389 [inlined]
 [8] cufunction(::GPUArrays.var"#68#69"{Bool}, ::Type{Tuple{CuArrays.CuKernelState,CUDAnative.CuDeviceArray{NTuple{4,UInt32},1,CUDAnative.AS.Global},CUDAnative.CuDeviceArray{Bool,1,CUDAnative.AS.Global}}}; name::Nothing, kwargs::Base.Iterators.Pairs{Union{},Union{},Tuple{},NamedTuple{(),Tuple{}}}) at /home/tim/Julia/pkg/CUDAnative/src/execution.jl:357
 [9] cufunction(::Function, ::Type) at /home/tim/Julia/pkg/CUDAnative/src/execution.jl:357
 [10] macro expansion at /home/tim/Julia/pkg/CUDAnative/src/execution.jl:174 [inlined]
 [11] macro expansion at ./gcutils.jl:105 [inlined]
 [12] macro expansion at /home/tim/Julia/pkg/CUDAnative/src/execution.jl:171 [inlined]
 [13] _gpu_call(::CuArrays.CuArrayBackend, ::Function, ::CuArray{Bool,1}, ::Tuple{CuArray{NTuple{4,UInt32},1},CuArray{Bool,1}}, ::Tuple{Tuple{Int64},Tuple{Int64}}) at /home/tim/Julia/pkg/CuArrays/src/gpuarray_interface.jl:60
 [14] gpu_call at /home/tim/Julia/pkg/GPUArrays/src/abstract_gpu_interface.jl:151 [inlined]
 [15] gpu_call(::Function, ::CuArray{Bool,1}, ::Tuple{CuArray{NTuple{4,UInt32},1},CuArray{Bool,1}}) at /home/tim/Julia/pkg/GPUArrays/src/abstract_gpu_interface.jl:128
 [16] rand!(::GPUArrays.RNG, ::CuArray{Bool,1}) at /home/tim/Julia/pkg/GPUArrays/src/random.jl:78
 [17] rand!(::CuArray{Bool,1}) at /home/tim/Julia/pkg/CuArrays/src/rand/random.jl:176
 [18] rand(::Type, ::Int64) at /home/tim/Julia/pkg/CuArrays/src/rand/random.jl:191
 [19] top-level scope at REPL[3]:1

[kshyatt]

Just tried this again, seems to still be an issue:

julia> CUDA.rand(Bool, 10)
ERROR: StackOverflowError:
Stacktrace:
     [1] macro expansion
       @ ~/.julia/dev/CUDA/lib/cudadrv/libcuda.jl:84 [inlined]
     [2] macro expansion
       @ ~/.julia/dev/CUDA/lib/cudadrv/error.jl:97 [inlined]
     [3] cuDevicePrimaryCtxRetain
       @ ~/.julia/dev/CUDA/lib/utils/call.jl:26 [inlined]
     [4] CuContext
       @ ~/.julia/dev/CUDA/lib/cudadrv/context.jl:57 [inlined]
     [5] context(dev::CuDevice)
       @ CUDA ~/.julia/dev/CUDA/lib/cudadrv/state.jl:222
     [6] CUDA.TaskLocalState(dev::CuDevice) (repeats 2 times)
       @ CUDA ~/.julia/dev/CUDA/lib/cudadrv/state.jl:50
     [7] task_local_state!()
       @ CUDA ~/.julia/dev/CUDA/lib/cudadrv/state.jl:73
     [8] prepare_cuda_state()
       @ CUDA ~/.julia/dev/CUDA/lib/cudadrv/state.jl:88
     [9] initialize_context()
       @ CUDA ~/.julia/dev/CUDA/lib/cudadrv/error.jl:80
    [10] macro expansion
       @ ~/.julia/dev/CUDA/lib/cudadrv/libcuda.jl:344 [inlined]
    [11] macro expansion
       @ ~/.julia/dev/CUDA/lib/cudadrv/error.jl:97 [inlined]
    [12] cuMemGetInfo_v2(free::Base.RefValue{UInt64}, total::Base.RefValue{UInt64})
       @ CUDA ~/.julia/dev/CUDA/lib/utils/call.jl:26
    [13] info()
       @ CUDA.Mem ~/.julia/dev/CUDA/lib/cudadrv/memory.jl:754
    [14] CUDA.MemoryInfo()
       @ CUDA ~/.julia/dev/CUDA/src/pool.jl:145
    [15] OutOfGPUMemoryError(sz::Int64) (repeats 2 times)
       @ CUDA ~/.julia/dev/CUDA/src/pool.jl:198
    [16] throw_api_error(res::CUDA.cudaError_enum)
       @ CUDA ~/.julia/dev/CUDA/lib/cudadrv/error.jl:89
    [17] macro expansion
       @ ~/.julia/dev/CUDA/lib/cudadrv/error.jl:101 [inlined]

[kshyatt]

CUDA.rand(Bool, 10) does work now! Should we close this?