Skip to main content

/

DocsCode

Mojo struct

BlockScaledTileCore

struct BlockScaledTileCore[a_type: DType, b_type: DType, c_type: DType, sfa_dtype: DType, sfb_dtype: DType, transpose_b: Bool, *, config: BlockScaledMatmulConfig[a_type, b_type, c_type, sfa_dtype, sfb_dtype, transpose_b]]

Core tile storage for block-scaled matmul SMEM structs.

Contains derived constants, layouts, tile storage, tile accessors, and size utilities. Shared between BlockScaledSmem (CLC), GroupedBlockScaledSmem (CLC + TMA descriptors), and Grouped1D1DSmem (no CLC).

Fields​

  • ​tiles (BlockScaledTileCore[a_type, b_type, c_type, sfa_dtype, sfb_dtype, transpose_b, config=config].Tiles):

Implemented traits​

AnyType, ImplicitlyDestructible

comptime members​

ATileArray​

comptime ATileArray = BlockScaledTileCore[a_type, b_type, c_type, sfa_dtype, sfb_dtype, transpose_b, config=config].Tiles.ATileArray

BK​

comptime BK = config.block_tile_shape[2]

BM​

comptime BM = config.block_tile_shape[0]

BN​

comptime BN = config.block_tile_shape[1]

BTileArray​

comptime BTileArray = BlockScaledTileCore[a_type, b_type, c_type, sfa_dtype, sfb_dtype, transpose_b, config=config].Tiles.BTileArray

CTileArray​

comptime CTileArray = BlockScaledTileCore[a_type, b_type, c_type, sfa_dtype, sfb_dtype, transpose_b, config=config].Tiles.CTileArray

MMA_M​

comptime MMA_M = config.mma_shape[0]

MMA_N​

comptime MMA_N = config.mma_shape[1]

num_accum_pipeline_stages​

comptime num_accum_pipeline_stages = config.num_accum_pipeline_stages

num_group_pipeline_stages​

comptime num_group_pipeline_stages = (BlockScaledTileCore[a_type, b_type, c_type, sfa_dtype, sfb_dtype, transpose_b, config=config].num_pipeline_stages // config)

num_output_stages​

comptime num_output_stages = config.num_output_stages

num_pipeline_stages​

comptime num_pipeline_stages = config.num_pipeline_stages

OutputM​

comptime OutputM = config.output_tile_shape[0]

OutputN​

comptime OutputN = config.output_tile_shape[1]

Payload​

comptime Payload = BlockScaledTilePayload[a_type, b_type, sfa_dtype, sfb_dtype, IndexList(BlockScaledTileCore[a_type, b_type, c_type, sfa_dtype, sfb_dtype, transpose_b, config=config].BM, BlockScaledTileCore[a_type, b_type, c_type, sfa_dtype, sfb_dtype, transpose_b, config=config].BK, __list_literal__=Tuple()), IndexList(BlockScaledTileCore[a_type, b_type, c_type, sfa_dtype, sfb_dtype, transpose_b, config=config].BN, BlockScaledTileCore[a_type, b_type, c_type, sfa_dtype, sfb_dtype, transpose_b, config=config].BK, __list_literal__=Tuple()), IndexList(BlockScaledTileCore[a_type, b_type, c_type, sfa_dtype, sfb_dtype, transpose_b, config=config].SFA_DIM0, BlockScaledTileCore[a_type, b_type, c_type, sfa_dtype, sfb_dtype, transpose_b, config=config].SFA_DIM1, __list_literal__=Tuple()), IndexList(BlockScaledTileCore[a_type, b_type, c_type, sfa_dtype, sfb_dtype, transpose_b, config=config].SFB_DIM0, BlockScaledTileCore[a_type, b_type, c_type, sfa_dtype, sfb_dtype, transpose_b, config=config].SFB_DIM1, __list_literal__=Tuple()), BlockScaledTileCore[a_type, b_type, c_type, sfa_dtype, sfb_dtype, transpose_b, config=config].num_pipeline_stages]

SF_BK​

comptime SF_BK = sf_bk[config]()

SF_K_GROUP_SIZE​

comptime SF_K_GROUP_SIZE = sf_k_group_size[config]()

SFA_DIM0​

comptime SFA_DIM0 = sfa_dim0[config]()

SFA_DIM1​

comptime SFA_DIM1 = sfa_dim1[config]()

sfa_smem_layout​

comptime sfa_smem_layout = tile_sf_layout_k_major[BlockScaledTileCore[a_type, b_type, c_type, sfa_dtype, sfb_dtype, transpose_b, config=config].BM, (BlockScaledTileCore[a_type, b_type, c_type, sfa_dtype, sfb_dtype, transpose_b, config=config].SF_K_GROUP_SIZE * config), config.vec_sf_size]()

SFATileArray​

comptime SFATileArray = BlockScaledTileCore[a_type, b_type, c_type, sfa_dtype, sfb_dtype, transpose_b, config=config].Tiles.SFATileArray

SFB_DIM0​

comptime SFB_DIM0 = sfb_dim0[config]()

SFB_DIM1​

comptime SFB_DIM1 = sfb_dim1[config]()

sfb_smem_layout​

comptime sfb_smem_layout = tile_sf_layout_k_major[align_up(BlockScaledTileCore[a_type, b_type, c_type, sfa_dtype, sfb_dtype, transpose_b, config=config].MMA_N, SF_MN_GROUP_SIZE), (BlockScaledTileCore[a_type, b_type, c_type, sfa_dtype, sfb_dtype, transpose_b, config=config].SF_K_GROUP_SIZE * config), config.vec_sf_size]()

SFBTileArray​

comptime SFBTileArray = BlockScaledTileCore[a_type, b_type, c_type, sfa_dtype, sfb_dtype, transpose_b, config=config].Tiles.SFBTileArray

Tiles​

comptime Tiles = BlockScaledTileStorage[a_type, b_type, c_type, sfa_dtype, sfb_dtype, IndexList(BlockScaledTileCore[a_type, b_type, c_type, sfa_dtype, sfb_dtype, transpose_b, config=config].BM, BlockScaledTileCore[a_type, b_type, c_type, sfa_dtype, sfb_dtype, transpose_b, config=config].BK, __list_literal__=Tuple()), IndexList(BlockScaledTileCore[a_type, b_type, c_type, sfa_dtype, sfb_dtype, transpose_b, config=config].BN, BlockScaledTileCore[a_type, b_type, c_type, sfa_dtype, sfb_dtype, transpose_b, config=config].BK, __list_literal__=Tuple()), BlockScaledTileCore[a_type, b_type, c_type, sfa_dtype, sfb_dtype, transpose_b, config=config].OutputM, BlockScaledTileCore[a_type, b_type, c_type, sfa_dtype, sfb_dtype, transpose_b, config=config].OutputN, IndexList(BlockScaledTileCore[a_type, b_type, c_type, sfa_dtype, sfb_dtype, transpose_b, config=config].SFA_DIM0, BlockScaledTileCore[a_type, b_type, c_type, sfa_dtype, sfb_dtype, transpose_b, config=config].SFA_DIM1, __list_literal__=Tuple()), IndexList(BlockScaledTileCore[a_type, b_type, c_type, sfa_dtype, sfb_dtype, transpose_b, config=config].SFB_DIM0, BlockScaledTileCore[a_type, b_type, c_type, sfa_dtype, sfb_dtype, transpose_b, config=config].SFB_DIM1, __list_literal__=Tuple()), BlockScaledTileCore[a_type, b_type, c_type, sfa_dtype, sfb_dtype, transpose_b, config=config].num_pipeline_stages, BlockScaledTileCore[a_type, b_type, c_type, sfa_dtype, sfb_dtype, transpose_b, config=config].num_output_stages]

Methods​

a_tiles​

a_tiles(ref[AddressSpace._value] self) -> BlockScaledTileCore[a_type, b_type, c_type, sfa_dtype, sfb_dtype, transpose_b, config=config].ATileArray

Get A tile array accessor.

Returns:

BlockScaledTileCore

b_tiles​

b_tiles(ref[AddressSpace._value] self) -> BlockScaledTileCore[a_type, b_type, c_type, sfa_dtype, sfb_dtype, transpose_b, config=config].BTileArray

Get B tile array accessor.

Returns:

BlockScaledTileCore

c_tiles​

c_tiles(ref[AddressSpace._value] self) -> BlockScaledTileCore[a_type, b_type, c_type, sfa_dtype, sfb_dtype, transpose_b, config=config].CTileArray

Get C tile array accessor.

Returns:

BlockScaledTileCore

sfa_tiles​

sfa_tiles(ref[AddressSpace._value] self) -> BlockScaledTileCore[a_type, b_type, c_type, sfa_dtype, sfb_dtype, transpose_b, config=config].SFATileArray

Get SFA tile array accessor.

Returns:

BlockScaledTileCore

sfb_tiles​

sfb_tiles(ref[AddressSpace._value] self) -> BlockScaledTileCore[a_type, b_type, c_type, sfa_dtype, sfb_dtype, transpose_b, config=config].SFBTileArray

Get SFB tile array accessor.

Returns:

BlockScaledTileCore

ab_pipeline_size​

static ab_pipeline_size() -> Int

Total size of A+B tiles for all pipeline stages (in elements).

Returns:

Int

sf_pipeline_size​

static sf_pipeline_size() -> Int

Total size of SFA+SFB tiles for all pipeline stages (in elements).

Returns:

Int

c_output_size​

static c_output_size() -> Int

Size of C tiles for all output stages (in elements).

Returns:

Int

total_tile_size​

static total_tile_size() -> Int

Total tile storage size (A+B+SFA+SFB+C) in elements.

Returns:

Int

View source

View source

Was this page helpful?

Models

Cookbook

Puzzles

Blog

Community

Contact