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 struct

HopperMatmulSM90Kernel_SMem

struct HopperMatmulSM90Kernel_SMem[a_type: DType, b_type: DType, c_type: DType, BM: Int, BN: Int, BK: Int, WG_BM: Int, WG_BN: Int, num_pipeline_stages: Int, k_group_size: Int, swizzle_bytes: Int = Int(128)]

Shared memory layout for Hopper SM90 matrix multiplication kernel.

This struct manages the shared memory allocation for:

• Input tiles (A and B matrices) with multi-stage pipelining
• Output tile (C matrix) for accumulation
• Synchronization barriers for producer-consumer coordination

The memory is organized to support asynchronous loads and efficient bank-conflict-free access patterns for tensor core operations.

All tiles use TileTensor-based types from tile_types.mojo. At TMA/WGMMA boundaries, pass {tile._storage} to construct the tile view.

Fields

• ​a_tiles_storage (HopperMatmulSM90Kernel_SMem[a_type, b_type, c_type, BM, BN, BK, WG_BM, WG_BN, num_pipeline_stages, k_group_size, swizzle_bytes].ATileArray.Storage):
• ​b_tiles_storage (HopperMatmulSM90Kernel_SMem[a_type, b_type, c_type, BM, BN, BK, WG_BM, WG_BN, num_pipeline_stages, k_group_size, swizzle_bytes].BTileArray.Storage):
• ​c_tile_storage (HopperMatmulSM90Kernel_SMem[a_type, b_type, c_type, BM, BN, BK, WG_BM, WG_BN, num_pipeline_stages, k_group_size, swizzle_bytes].CTileArray.Storage):
• ​barriers (BarrierPair[(num_pipeline_stages // k_group_size)]):

Implemented traits

AnyType, Deinitable, Movable

comptime members

ATileArray

comptime ATileArray = SMemTileArrayWithLayout[a_type, Layout(Coord(Coord(ComptimeInt(), ComptimeInt()), Coord(ComptimeInt(), ComptimeInt())), Coord(Coord(ComptimeInt(), ComptimeInt()), Coord(ComptimeInt(), ComptimeInt()))), num_pipeline_stages]

BTileArray

comptime BTileArray = SMemTileArrayWithLayout[b_type, Layout(Coord(Coord(ComptimeInt(), ComptimeInt()), Coord(ComptimeInt(), ComptimeInt())), Coord(Coord(ComptimeInt(), ComptimeInt()), Coord(ComptimeInt(), ComptimeInt()))), num_pipeline_stages]

CTile

comptime CTile = HopperMatmulSM90Kernel_SMem[a_type, b_type, c_type, BM, BN, BK, WG_BM, WG_BN, num_pipeline_stages, k_group_size, swizzle_bytes].CTileArray.Tile

CTileArray

comptime CTileArray = SMemTileArray2DRowMajor[c_type, WG_BM, WG_BN, Int(1)]

Methods

a_tiles

def a_tiles(ref[AddressSpace._value] self) -> Self.ATileArray

Get A tile array accessor (TileTensor-based).

Returns:

Self.ATileArray

b_tiles

def b_tiles(ref[AddressSpace._value] self) -> Self.BTileArray

Get B tile array accessor (TileTensor-based).

Returns:

Self.BTileArray

c_tile

def c_tile(ref[AddressSpace._value] self) -> Self.CTile

Get C tile accessor (TileTensor-based).

Returns:

Self.CTile

create_pipeline

def create_pipeline(ref[AddressSpace._value] self) -> ProducerConsumerPipeline[(num_pipeline_stages // k_group_size)]

Create producer-consumer pipeline from barrier storage.

Returns:

ProducerConsumerPipeline[(num_pipeline_stages // k_group_size)]

pipeline_storage_size

static def pipeline_storage_size() -> Int

Calculate the memory size for all pipeline stages.

Returns:

Int

output_storage_size

static def output_storage_size() -> Int

Calculate the memory size for output tile.

Returns:

Int

storage_size

static def storage_size() -> Int

Calculate the total storage size.

Returns:

Int