Large operator in scan

[Shiva Kaul]

In all the CUB examples, the ScanOps are elementary operations on scalar types T. In my problem, T is a matrix type. The operation is not quite as expensive as matrix multiplication, but it's nonetheless preferable to utilize multiple threads (in the same block or perhaps even warp) to compute it in parallel. Is there an elegant way to accomplish this while utilizing CUB for the overall scan?

Thanks,

Shiva