Variable live-time counter intuitive on the CPU

[vchuravy]

I just found myself writing a kernel like this:

@kernel function localmem(A)
    I = @index(Global, Linear)
    i = @index(Local, Linear)
    N = groupsize()
    lmem = @localmem Int (N,) # Ok iff groupsize is static 
    lmem[i] = i
    @synchronize
    A[I] = lmem[N - i + 1]
end

This works as expected on the GPU, but on the CPU there is counter-intuitive variable scoping going on. The first weird part is that after the @synchronize the lifetime of N has ended.

And so I thought I could just re-introduce the variable...

@kernel function localmem(A)
    I = @index(Global, Linear)
    i = @index(Local, Linear)
    N = groupsize()
    lmem = @localmem Int (N,) # Ok iff groupsize is static 
    lmem[i] = i
    @synchronize
    N = groupsize()
    A[I] = lmem[N - i + 1]
end

But it turns out that I made @localmem special, and its live-time begins in an outside scope, which of course is now also missing N.

I will have to think about how to fix this. But most likely this means that we need a "non-divergent value" or a group constant value. In return we might be able to get rid of @private, and do the right thing automatically for all divergent values.

[vchuravy]

Needs better docs, but #34 and #35 introduces @uniform to mitigate this issue.