https://hgpu.org/?p=8327
Architectural explorations for streaming accelerators with customized memory layouts