Slicing causes GPUifyLoops failures

[ChrisRackauckas]

using OrdinaryDiffEq, DiffEqGPU
function lorenz2(du,u,p,t)
 @inbounds begin
     a = u[1:2]
     du[1] = u[1]
 end
 nothing
end

u0 = Float32[1.0;0.0;0.0]
tspan = (0.0f0,100.0f0)
p = (10.0f0,28.0f0,8/3f0)
prob = ODEProblem(lorenz,u0,tspan,p)
prob_func = (prob,i,repeat) -> remake(prob,p=rand(Float32,3).*p)
monteprob = EnsembleProblem(prob, prob_func = prob_func)
@time sol = solve(monteprob,Tsit5(),EnsembleGPUArray(),trajectories=10_000,saveat=1.0f0)

@vchuravy do you know about this one?

[ArnoStrouwen]

I think I have similar issues when directly using CUDAnative instead of GPUifyLoops.

using GPUifyLoops, CUDAnative, CuArrays
using BenchmarkTools
#slicing does not work
function main1!(x_out, x_in)
    index = (blockIdx().x - 1) * blockDim().x + threadIdx().x
    stride = blockDim().x * gridDim().x
    for i = index:stride:size(x_in,2)
        @inbounds @views begin
            x_out[:,i] = x_in[:,i] # .= also does not work
        end
    end
    return nothing
end
x_in = cu(rand(2,100_000_000))
x_out = similar(x_in)
@launch CUDA() threads=1024 blocks=16 main1!(x_out,x_in)

Similarly vchuravy/GPUifyLoops.jl#87 also fails without using GPUifyLoops. And I am assuming #41 also.

# @inbounds has to be used twice
function dynamics!(x_out,x_in)
    @inbounds begin
    x_out[1] = x_in[1] + x_in[2]
    x_out[2] = 1.1*x_in[2]
    end
    return nothing
end
function dynamics_no_inbounds!(x_out,x_in)
    x_out[1] = x_in[1] + x_in[2]
    x_out[2] = 1.1*x_in[2]
    return nothing
end
function main2!(x_out, x_in)
    index = (blockIdx().x - 1) * blockDim().x + threadIdx().x
    stride = blockDim().x * gridDim().x
    for i = index:stride:size(x_in,2)
        @inbounds @views begin
            dynamics!(x_out[:,i],x_in[:,i])
        end
    end
    return nothing
end
function main3!(x_out, x_in)
    index = (blockIdx().x - 1) * blockDim().x + threadIdx().x
    stride = blockDim().x * gridDim().x
    for i = index:stride:size(x_in,2)
        @inbounds @views begin
            dynamics_no_inbounds!(x_out[:,i],x_in[:,i])
        end
    end
    return nothing
end
x_in = cu(rand(2,100_000_000))
x_out = similar(x_in)
@launch CUDA() threads=1024 blocks=16 main2!(x_out,x_in)
# @btime CuArrays.@sync begin @launch CUDA() threads=1024 blocks=16 main2!(x_out,x_in) end #20 ms (GPUifyloops is 110ms)
@launch CUDA() threads=1024 blocks=16 main3!(x_out,x_in)

[ChrisRackauckas]

Fixed with KernelAbstractions.jl