Questions about the distribution of the threads over the tiles

[timespacce]

Hello,

I have questions about the cuda core sgemm.

1. Each thread block loads per iteration a 128x8 A-tile and a 8x128 B-tile from global into shared memory. By having 256 threads pro thread block each thread would compute a 8x8 matrix multiplication?

2. How are the threads distributed over the A-tile * B-tile?

Greetings,
James.

[timespacce]

https://devblogs.nvidia.com/cutlass-linear-algebra-cuda/