Arithmetic bug with 64-bit integers

[maleadt]

Computing 2^19 + 2 * 1 fails when working with 64-bit integers, but works with 32-bit ones. Consistently reproducible with a GeForce GTX TITAN on drivers 378.13 (current short-lived) and 381.09 (current beta), but works properly on 375.39 (current long-lived). SASS code is identical across driver versions, so this looks to be an even lower-level bug than #4.

Reproduced from a bogus OOB (64-bit indices...) when working on Rodiina/needle.
Looked something like:

function oob(reference)
    index    = 4130784 - 32784 * blockIdx().x + 16 * blockIdx().x + threadIdx().x + 2051

    ref = @cuStaticSharedMem(Int32, (16, 16))
    for ty = 0:15
        i = index + 2049 * ty + 1
        @inbounds ref[threadIdx().x, ty + 1] = reference[i]
    end

    return nothing
end

array = CuArray{Int32}(2049, 2049)
@cuda (1,1) oob(array)

Full repro:

using CUDAdrv, CUDAnative

function kernel{T}(one::T, ptr::Ptr{T})
    val = T(524288) + T(2) * one
    Base.pointerset(ptr, val, 1, 8)
    return nothing
end

dev = CuDevice(0)
ctx = CuContext(dev)

function test(name, T)
    ref = CuArray{T}(1)
    @cuda (1,1) kernel(T(1), pointer(ref))
    println("$name: ", Array(ref)[1])

    if !isfile("$name.ll")
        open("$name.ll", "w") do io
            CUDAnative.code_llvm(io, kernel, Tuple{T, Ptr{T}};
                                 dump_module=true, cap=capability(dev))
        end
    end
end

test("32bit", Int32)
test("64bit", Int64)

destroy(ctx)

32bit: 524290
64bit: 2

Bug has been filed with NVIDIA.

cc @cfoket

[maleadt]

SASS code is identical across driver versions, so this looks to be an even lower-level bug than #4.

Unless of course the ptxas I use (part of the CUDA toolkit, not the driver) to get a hold of the SASS code contains its own copy of the assembler.

[maleadt]

Filed with NVIDIA as bug 200303640. Not that I have access to the bug tracker...

[maleadt]

Fixed on 384.47 (current beta), 381.22 (current short-lived) and 375.66 (current long-lived).

Not sure how we should deal with this. Warn when detecting a bad driver version? But we lack libNVML on eg. macOS, so that'd be annoying to detect.

[maleadt]

Seeing how this is fixes on current long-lived, and NVIDIA hasn't provided my with any details about the versions that contain this bug, I guess there's nothing much we can do about this except for advising users to be running the latest long-lived kernel...