[maleadt]

AMDGPU's precompilation workload doesn't use the GPU, which isn't even discovered while AMDGPU precompiles, so the first reduction, scan or predicate of a session compiles AcceleratedKernels' host code and kernels.

This runs them in the workload of a new AMDGPU extension, which precompiles with AMDGPU loaded and ROCm discovered, when there is a device. AMDGPU.functional() only checks that ROCm's libraries were found, so the guard also counts the devices (with the unchecked HIP call; older AMDGPU versions without it skip the workload). The arrays are just large enough to select the kernels that a vector of 10^6 elements and a 1000×1000 matrix do; at AMDGPU's default tunings that takes longer matrix columns than on CUDA. Like AMDGPU's own workload it needs Julia 1.12.

On a gfx1036 with Julia 1.13, the first sum of 10^6 elements takes 17 ms instead of 1.1 s, cumsum 0.21 instead of 1.4 s, any and all 70 ms instead of 1.1 s, and sum along either dimension of a 1000×1000 matrix 0.45 instead of 1.6 s. The extension's image is 9 MB and takes 4 s to precompile; load time doesn't change measurably.