Add graph capture

[pxl-th]

• Add graph capture.
• Update profiling docs to rocprofv3.

using AMDGPU
using GPUArrays

function f(o)
    x = AMDGPU.rand(Float32, size(o))
    y = AMDGPU.rand(Float32, size(o))
    o .+= sin.(x) * cos.(y) .+ 1f0
    return
end

function main()
    cache = GPUArrays.AllocCache()
    z = AMDGPU.zeros(Float32, 256, 256)
    N = 1000

    GPUArrays.@cached cache f(z)

    # Regular launch.
    # t = AMDGPU.@elapsed for i in 1:N
    #     GPUArrays.@cached cache f(z)
    # end

    # Graph launch.
    g = GPUArrays.@cached cache AMDGPU.@captured f(z)
    t = AMDGPU.@elapsed for i in 1:N
        AMDGPU.launch(g)
    end

    AMDGPU.synchronize()
    @show t
    return
end
main()

Regular	Captured
6k kernel launches	1k graph launches
~135 ms	~64 ms

[luraess]

Maybe bits of #862 could be relevant or integrated here given there was not much recent activity on that PR ?

[pxl-th]

Maybe bits of #862 could be relevant or integrated here given there was not much recent activity on that PR ?

Not sure if we should mention older versions of rocprof, rocprofv3 seems to work fine.
And for roctx we can do it properly in a separate PR