assignment changes random seed

[bjarthur]

on the CPU:

julia> using Random

julia> foo = Vector{Float64}(undef, 1000);

julia> Random.seed!(1)
TaskLocalRNG()

julia> rand(3)
3-element Vector{Float64}:
 0.07336635446929285
 0.34924148955718615
 0.6988266836914685

julia> foo[1:3] .= 4;

julia> rand(3)
3-element Vector{Float64}:
 0.6282647403425017
 0.9149290036628314
 0.19280811624587546

same on the GPU. notice the first set of random numbers is the same as above. the second are different.

julia> using Random, CUDA

julia> foo = CuVector{Float64}(undef, 1000);

julia> Random.seed!(1)
TaskLocalRNG()

julia> rand(3)
3-element Vector{Float64}:
 0.07336635446929285
 0.34924148955718615
 0.6988266836914685

julia> foo[1:3] .= 4;

julia> rand(3)
3-element Vector{Float64}:
 0.7805192636751863
 0.6702639583444937
 0.16771210647092682

i suppose this could be a bug in the CPU code. let me know if i should refile the issue elsewhere.

[bjarthur]

oh, and this is with julia 1.7.0-rc2, CUDA 3.5.0, and GPUArrays 8.1.2. i did not notice this with julia 1.6 and CUDA 3.4.2

[bjarthur]

skipping the assignment on the CPU yields the same random numbers the second time, so it seems this is a GPU problem.

julia> using Random

julia> foo = Vector{Float64}(undef, 1000);

julia> Random.seed!(1)
TaskLocalRNG()

julia> rand(3)
3-element Vector{Float64}:
 0.07336635446929285
 0.34924148955718615
 0.6988266836914685

julia> rand(3)
3-element Vector{Float64}:
 0.6282647403425017
 0.9149290036628314
 0.19280811624587546

[maleadt]

This seems caused by task switches, so better file this upstream:

julia> Random.seed!(1); rand(); rand()
0.34924148955718615

julia> Random.seed!(1); rand(); rand()
0.34924148955718615

julia> Random.seed!(1); rand(); wait(@async begin end); rand()
0.19280811624587546

[bjarthur]

using the MersenneTwister RNG instead of the default fixes it for my use case. thanks for looking into this.