Skip to main content

/

DocsCode

Mojo struct

BlockwiseFP8Smem

struct BlockwiseFP8Smem[a_type: DType, b_type: DType, c_type: DType, a_scales_type: DType, transpose_b: Bool, *, config: MatmulConfig[a_type, b_type, c_type, transpose_b]]

SMEM struct for blockwise FP8 matmul: A/B tiles, A-scales, C output, barriers.

Key differences from BlockScaledSmem:

  • A-scales stored in SMEM (1D: 1 x BM per pipeline stage)
  • No B-scales in SMEM (read from global memory during epilogue)
  • Used with register-based accumulation pattern

Fields​

  • ​a_tiles_storage (BlockwiseFP8Smem[a_type, b_type, c_type, a_scales_type, transpose_b, config=config].ATileArray.Storage):
  • ​b_tiles_storage (BlockwiseFP8Smem[a_type, b_type, c_type, a_scales_type, transpose_b, config=config].BTileArray.Storage):
  • ​c_tiles_storage (BlockwiseFP8Smem[a_type, b_type, c_type, a_scales_type, transpose_b, config=config].CTileArray.Storage):
  • ​a_scales_tiles_storage (BlockwiseFP8Smem[a_type, b_type, c_type, a_scales_type, transpose_b, config=config].AScalesTileArray.Storage):
  • ​tma_mma_mbars_storage (BlockwiseFP8Smem[a_type, b_type, c_type, a_scales_type, transpose_b, config=config].InputBarriers.Storage):
  • ​accum_mbars_storage (BlockwiseFP8Smem[a_type, b_type, c_type, a_scales_type, transpose_b, config=config].AccumBarriers.Storage):
  • ​clc_mbars_full_storage (BlockwiseFP8Smem[a_type, b_type, c_type, a_scales_type, transpose_b, config=config].ClcBarriers.Storage):
  • ​clc_mbars_empty_storage (BlockwiseFP8Smem[a_type, b_type, c_type, a_scales_type, transpose_b, config=config].ClcBarriers.Storage):
  • ​clc_throttle_mbars_storage (BlockwiseFP8Smem[a_type, b_type, c_type, a_scales_type, transpose_b, config=config].ClcThrottleBarriers.Storage):
  • ​clc_response_storage (BlockwiseFP8Smem[a_type, b_type, c_type, a_scales_type, transpose_b, config=config].ClcResponse.Storage):
  • ​tmem_dealloc_mbar_storage (BlockwiseFP8Smem[a_type, b_type, c_type, a_scales_type, transpose_b, config=config].TmemDealloc.Storage):
  • ​tmem_addr_storage (BlockwiseFP8Smem[a_type, b_type, c_type, a_scales_type, transpose_b, config=config].TmemAddr.Storage):

Implemented traits​

AnyType, ImplicitlyDestructible

comptime members​

__del__is_trivial​

comptime __del__is_trivial = True

a_scales_smem_layout​

comptime a_scales_smem_layout = Layout.row_major(1, BlockwiseFP8Smem[a_type, b_type, c_type, a_scales_type, transpose_b, config=config].BM)

a_smem_layout​

comptime a_smem_layout = tile_layout_k_major[a_type, BlockwiseFP8Smem[a_type, b_type, c_type, a_scales_type, transpose_b, config=config].BM, BlockwiseFP8Smem[a_type, b_type, c_type, a_scales_type, transpose_b, config=config].BK, config.a_swizzle]()

AccumBarriers​

comptime AccumBarriers = SMemArray[SharedMemBarrier, (BlockwiseFP8Smem[a_type, b_type, c_type, a_scales_type, transpose_b, config=config].num_accum_pipeline_stages * 2)]

AScalesTileArray​

comptime AScalesTileArray = SMemTileArray[a_scales_type, BlockwiseFP8Smem[a_type, b_type, c_type, a_scales_type, transpose_b, config=config].a_scales_smem_layout, BlockwiseFP8Smem[a_type, b_type, c_type, a_scales_type, transpose_b, config=config].num_pipeline_stages, 128]

ATileArray​

comptime ATileArray = SMemTileArray[a_type, BlockwiseFP8Smem[a_type, b_type, c_type, a_scales_type, transpose_b, config=config].a_smem_layout, BlockwiseFP8Smem[a_type, b_type, c_type, a_scales_type, transpose_b, config=config].num_pipeline_stages, 128]

b_smem_layout​

comptime b_smem_layout = tile_layout_k_major[b_type, BlockwiseFP8Smem[a_type, b_type, c_type, a_scales_type, transpose_b, config=config].BN, BlockwiseFP8Smem[a_type, b_type, c_type, a_scales_type, transpose_b, config=config].BK, config.b_swizzle]() if transpose_b else tile_layout_mn_major[b_type, BlockwiseFP8Smem[a_type, b_type, c_type, a_scales_type, transpose_b, config=config].BN, BlockwiseFP8Smem[a_type, b_type, c_type, a_scales_type, transpose_b, config=config].BK, config.b_swizzle]()

BK​

comptime BK = config.block_tile_shape.__getitem__[3, DType.int64, Int](2)

BM​

comptime BM = config.block_tile_shape.__getitem__[3, DType.int64, Int](0)

BN​

comptime BN = config.block_tile_shape.__getitem__[3, DType.int64, Int](1)

BTileArray​

comptime BTileArray = SMemTileArray[b_type, BlockwiseFP8Smem[a_type, b_type, c_type, a_scales_type, transpose_b, config=config].b_smem_layout, BlockwiseFP8Smem[a_type, b_type, c_type, a_scales_type, transpose_b, config=config].num_pipeline_stages, 128]

c_smem_layout​

comptime c_smem_layout = Layout.row_major(BlockwiseFP8Smem[a_type, b_type, c_type, a_scales_type, transpose_b, config=config].OutputM, BlockwiseFP8Smem[a_type, b_type, c_type, a_scales_type, transpose_b, config=config].OutputN)

ClcBarriers​

comptime ClcBarriers = SMemArray[SharedMemBarrier, BlockwiseFP8Smem[a_type, b_type, c_type, a_scales_type, transpose_b, config=config].num_clc_pipeline_stages]

ClcResponse​

comptime ClcResponse = SMemArray[UInt128, BlockwiseFP8Smem[a_type, b_type, c_type, a_scales_type, transpose_b, config=config].num_clc_pipeline_stages]

ClcThrottleBarriers​

comptime ClcThrottleBarriers = SMemArray[SharedMemBarrier, (BlockwiseFP8Smem[a_type, b_type, c_type, a_scales_type, transpose_b, config=config].num_clc_pipeline_stages * 2)]

CTileArray​

comptime CTileArray = SMemTileArray[c_type, BlockwiseFP8Smem[a_type, b_type, c_type, a_scales_type, transpose_b, config=config].c_smem_layout, BlockwiseFP8Smem[a_type, b_type, c_type, a_scales_type, transpose_b, config=config].num_output_stages, 128]

InputBarriers​

comptime InputBarriers = SMemArray[SharedMemBarrier, (BlockwiseFP8Smem[a_type, b_type, c_type, a_scales_type, transpose_b, config=config].num_group_pipeline_stages * 2)]

MMA_M​

comptime MMA_M = config.mma_shape.__getitem__[3, DType.int64, Int](0)

MMA_N​

comptime MMA_N = config.mma_shape.__getitem__[3, DType.int64, Int](1)

num_accum_pipeline_stages​

comptime num_accum_pipeline_stages = config.num_accum_pipeline_stages

num_clc_pipeline_stages​

comptime num_clc_pipeline_stages = config.num_clc_pipeline_stages

num_group_pipeline_stages​

comptime num_group_pipeline_stages = (BlockwiseFP8Smem[a_type, b_type, c_type, a_scales_type, transpose_b, config=config].num_pipeline_stages // config)

num_output_stages​

comptime num_output_stages = Int.__init__[Int](config.num_output_stages)

num_pipeline_stages​

comptime num_pipeline_stages = config.num_pipeline_stages

OutputM​

comptime OutputM = config.output_tile_shape.__getitem__[2, DType.int64, Int](0)

OutputN​

comptime OutputN = config.output_tile_shape.__getitem__[2, DType.int64, Int](1)

TmemAddr​

comptime TmemAddr = SMemArray[UInt32, 1]

TmemDealloc​

comptime TmemDealloc = SMemArray[SharedMemBarrier, 1]

Methods​

a_tiles​

a_tiles(ref [3] self) -> BlockwiseFP8Smem[a_type, b_type, c_type, a_scales_type, transpose_b, config=config].ATileArray

Returns:

BlockwiseFP8Smem

b_tiles​

b_tiles(ref [3] self) -> BlockwiseFP8Smem[a_type, b_type, c_type, a_scales_type, transpose_b, config=config].BTileArray

Returns:

BlockwiseFP8Smem

c_tiles​

c_tiles(ref [3] self) -> BlockwiseFP8Smem[a_type, b_type, c_type, a_scales_type, transpose_b, config=config].CTileArray

Returns:

BlockwiseFP8Smem

a_scales_tiles​

a_scales_tiles(ref [3] self) -> BlockwiseFP8Smem[a_type, b_type, c_type, a_scales_type, transpose_b, config=config].AScalesTileArray

Returns:

BlockwiseFP8Smem

tma_mma_mbars​

tma_mma_mbars(ref [3] self) -> BlockwiseFP8Smem[a_type, b_type, c_type, a_scales_type, transpose_b, config=config].InputBarriers

Returns:

BlockwiseFP8Smem

accum_mbars​

accum_mbars(ref [3] self) -> BlockwiseFP8Smem[a_type, b_type, c_type, a_scales_type, transpose_b, config=config].AccumBarriers

Returns:

BlockwiseFP8Smem

clc_mbars_full​

clc_mbars_full(ref [3] self) -> BlockwiseFP8Smem[a_type, b_type, c_type, a_scales_type, transpose_b, config=config].ClcBarriers

Returns:

BlockwiseFP8Smem

clc_mbars_empty​

clc_mbars_empty(ref [3] self) -> BlockwiseFP8Smem[a_type, b_type, c_type, a_scales_type, transpose_b, config=config].ClcBarriers

Returns:

BlockwiseFP8Smem

clc_throttle_mbars​

clc_throttle_mbars(ref [3] self) -> BlockwiseFP8Smem[a_type, b_type, c_type, a_scales_type, transpose_b, config=config].ClcThrottleBarriers

Returns:

BlockwiseFP8Smem

clc_response​

clc_response(ref [3] self) -> BlockwiseFP8Smem[a_type, b_type, c_type, a_scales_type, transpose_b, config=config].ClcResponse

Returns:

BlockwiseFP8Smem

tmem_dealloc_mbar​

tmem_dealloc_mbar(ref [3] self) -> BlockwiseFP8Smem[a_type, b_type, c_type, a_scales_type, transpose_b, config=config].TmemDealloc

Returns:

BlockwiseFP8Smem

tmem_addr​

tmem_addr(ref [3] self) -> BlockwiseFP8Smem[a_type, b_type, c_type, a_scales_type, transpose_b, config=config].TmemAddr

Returns:

BlockwiseFP8Smem

input_barriers​

input_barriers(ref [3] self) -> BlockwiseFP8Smem[a_type, b_type, c_type, a_scales_type, transpose_b, config=config].InputBarriers

Returns:

BlockwiseFP8Smem

accum_barriers​

accum_barriers(ref [3] self) -> BlockwiseFP8Smem[a_type, b_type, c_type, a_scales_type, transpose_b, config=config].AccumBarriers

Returns:

BlockwiseFP8Smem

tmem_dealloc​

tmem_dealloc(ref [3] self) -> BlockwiseFP8Smem[a_type, b_type, c_type, a_scales_type, transpose_b, config=config].TmemDealloc

Returns:

BlockwiseFP8Smem

ab_pipeline_size​

static ab_pipeline_size() -> Int

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

Returns:

Int

a_scales_pipeline_size​

static a_scales_pipeline_size() -> Int

Total size of A-scales 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+A-scales+C) in elements.

Returns:

Int

View source

View source

Was this page helpful?

Models

Cookbook

Puzzles

Blog

Community

Contact