Sarek_gemm.HostHost-side launch configuration for sgemm_tiled_kernel. Kept element-type- agnostic so an f64/df64 or L15b mma kernel reuses it by construction.
Tile edge. MUST equal the kernel's shared extent (tile*tile) and tile_size literal; 16 -> 256 threads/block, 2 KiB shared.
val block : unit -> Spoc_framework.Framework_sig.dimsBlock dimensions: one thread per output element of a tile.
val grid : m:int -> n:int -> Spoc_framework.Framework_sig.dimsGrid dimensions covering an MxN output (ceil-div, so partial edge blocks are launched and masked by the kernel's boundary guards).
Shared-memory footprint of the kernel's two float32 tiles, for occupancy estimation only. The kernel declares its tiles with let%shared (STATIC shared memory), so launches do NOT need to pass ~shared_mem - passing it merely reserves the same bytes again as an unused dynamic region (audit finding L5).