I am currently working on implementing the decoupled-lookback in Futhark, as my master thesis. The development is under the branch `onepassscan`. It is still missing some features, like it cannot do map'o'scan, meaning include the map function in the scan optimization. It is likely that the implementation will only be used for Nvidia, since the cache guarentees only are prommised to work for Nvidia.
The current version of Futhark is using the reduce-then-scan strategy.
I'll note that these "cache guarantees" in the Vulkan world are the Vulkan 1.2 memory model, and are supported by the latest drivers for AMD, Intel, and NVidia [1]. This recent change is one big reason I'm saying Vulkan + SPIR-V is getting ready for real compute workloads, and this wasn't the case even months ago.
The current version of Futhark is using the reduce-then-scan strategy.