How to make sure that all blocks are in the grid

Accelerated Computing CUDA CUDA Programming and Performance
1

I would like to use cooperative groups, and grid sync function, with maximum possible number of blocks.

As I know from documentation If block will not be currently managed by SM it will make impossible for kernel to execute.

Now number of blocks deployed depends on amount of required shared memory and required warps, but also (I suppose) othe activities on GPU.
Considering all those 3 things how to calculate (presumably using occupancy API) maximum number of blocks .

For example I want to have blocks with 512 threads and 8000 bits of shared memory each - and now calculate in runtime how many blocks can i fire to fit them in a grid.

2

There is an example of how to do this already provided in the programming guide, here.

Here is also a complete example.

1 Like