grid_dim.y boundary check

[zhaolianshuizls]

Is it possible num_tiles is so large that it exceeds the grid_dim.y boundary?
https://github.com/NVIDIA/cub/blob/b2e8bccb8c0cd15279974fe4b9b8d6fcd1842b57/cub/device/dispatch/dispatch_select_if.cuh#L422

[gevtushenko]

Hello, @zhaolianshuizls! The code you mentioned is written in terms of tiles. Tile size depends on the data type, but let's say it's 128 threads, each processing ten items, so the tile size is 1280 (~2^10) items. To calculate the number of tiles, select first checks the maximum x-dimension of a grid of thread blocks. It should be around 2^31 blocks. This already converts to 2^41 items without increasing the number of blocks along the Y dimension. If you are using our device API, the num_items type is currently limited to be an int, so it should be impossible to get scan_grid_size.y > 1. If you are using the dispatch layer directly, providing OffsetT different from int, it should be possible to get scan_grid_size.y > 1. In this case, the maximal grid size along the y dimension is 2^16, giving us the num_items limit of about 2^57 items.