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

run_producer

def run_producer[dtype: DType, layout: Layout, block_rows: Int, block_cols: Int, warp_rows: Int, warp_cols: Int, producer_warps: Int, pipeline_stages: Int, k_tile_size: Int, simd_width: Int, warps_processed_per_producer: Int, tile_count: Int, swizzle: Optional[Swizzle]](matrix: LayoutTensor[dtype, layout, MutAnyOrigin, address_space=AddressSpace.GLOBAL], mut ring_buffer: RingBuffer, warp_id: Int, block_idx_dim: Int)

Generic producer function for loading matrix tiles from global to shared memory.

Parameters:

• ​dtype (DType): Element type of the matrix being loaded.
• ​layout (Layout): Memory layout of the matrix in global memory.
• ​block_rows (Int): Number of rows in a block tile (BM for matrix A, BN for matrix B).
• ​block_cols (Int): Number of columns in a block tile (BK).
• ​warp_rows (Int): Number of rows per warp tile (WM for matrix A, WN for matrix B).
• ​warp_cols (Int): Number of columns per warp tile (WK).
• ​producer_warps (Int): Number of warps assigned to producing this matrix.
• ​pipeline_stages (Int): Number of pipeline stages in the ring buffer used to overlap loads and compute.
• ​k_tile_size (Int): Width of a single K tile loaded per inner-block row.
• ​simd_width (Int): SIMD vector width of the matrix element type.
• ​warps_processed_per_producer (Int): Number of warp tiles each producer warp iterates over.
• ​tile_count (Int): Total number of K tiles to load along the K dimension.
• ​swizzle (Optional[Swizzle]): Optional swizzle pattern applied to shared memory to avoid bank conflicts.

Args:

• ​matrix (LayoutTensor[dtype, layout, MutAnyOrigin, address_space=AddressSpace.GLOBAL]): Global memory tensor of the matrix to load tiles from.
• ​ring_buffer (RingBuffer): Ring buffer coordinating producer-consumer synchronization across pipeline stages.
• ​warp_id (Int): Warp identifier of this producer warp within its producer group.
• ​block_idx_dim (Int): Block index along the matrix's leading dimension, selecting which block of tiles this block loads.