For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /get-started.md).

Mojo function

store_p_quadrant

def store_p_quadrant[cols: Int, p_type: DType, //, *, BN: Int, p_tile_rows: Int](p_smem: Pointer[Scalar[p_type], address_space=AddressSpace.SHARED], p: Array[Float32, cols], row: UInt32, warp_in_wg: UInt32)

Row-local P quadrant writer (Layout-G register map -> SWIZZLE_NONE SMEM).

Lane identity = row (< 32); warp w holds columns [w*BN/4, (w+1)*BN/4). Casts the fp32 quadrant to p_type and stores it in the k-major layout the SS P@V A-descriptor reads: element (r, k) sits in block k // sw_K at block base (k // sw_K) * p_tile_rows * sw_K, at within-block offset r * sw_K. 16 B stores; the caller owns the fence_async_view_proxy + barrier before the MMA reads P.

Parameters:

• ​cols (Int): Quadrant width in elements; must equal BN // 4 (inferred).
• ​p_type (DType): Element dtype of the P SMEM tile (inferred).
• ​BN (Int): Key-block width of the full P tile.
• ​p_tile_rows (Int): Physical row count of the P SMEM tile.

Args:

• ​p_smem (Pointer[Scalar[p_type], address_space=AddressSpace.SHARED]): Base of this warpgroup's P tile in shared memory.
• ​p (Array[Float32, cols]): This thread's fp32 P quadrant register array (one row, cols wide).
• ​row (UInt32): This thread's row within the tile (lane identity, < p_tile_rows).
• ​warp_in_wg (UInt32): This warp's index within the warpgroup (0..3).