Mark Harris
Mark Harris
I think we probably want both a way to disable RMM NVTX completely, and possibly some control over granularity. When debugging allocation bottlenecks, it's helpful to have finer granularity of...
Another important current example: ```c++ // auto col = {1, 2, 3, 4, 5}; cudf::rolling_window(col, 0, 0, 1, ???); // null, null, null, null, null cudf::rolling_window(col, 0, 0, 1, min);...
BTW, I found the original negative window sizes request. https://github.com/rapidsai/cudf/issues/3650 Interestingly, the reason for the request was to enable windows that DO NOT include the current row/element. So I don't...
I need to correct myself. Using 5, -1, 1 should work for a window that only includes the 5 elements before the current element, and not the current element!
Before we do anything here, we need to gather further requirements. For example @mythrocks mentioned the need for offset windows. @mythrocks can you please elaborate?
I'm sorry, I don't.
I definitely won't stand in your way. :)
I think that it's important to document the increased memory usage that happens if the user passes a host_mr. It effectively doubles -- half on device_mr and half on host_mr....