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
TiledMma
struct TiledMma[out_type: DType, in_type: DType, shape: IndexList[Int(3)], group_size: Int]
Stateless MMA computation on TileTensors.
Direct TileTensor port of TiledTensorCore.mma. Iterates group_size k-steps, indexes A/B register tiles per step, and calls gpu_mma. No register ownership, no SMEM loading: pure computation.
Parameters
- out_type (
DType): Accumulator data type (typically float32). - in_type (
DType): Input element data type (bfloat16 or float8). - shape (
IndexList[Int(3)]): MMA instruction shape [M, N, K]. - group_size (
Int): Number of k-steps per mma() call.
Implemented traits
comptime members
a_frag_size
comptime a_frag_size = (Int((mul shape[Int(0)], shape[Int(2)])) // _resolve_warp_size())
c_frag_size
comptime c_frag_size = (Int((mul shape[Int(0)], shape[Int(1)])) // _resolve_warp_size())
MMA_K
comptime MMA_K = shape[Int(2)]
MMA_M
comptime MMA_M = shape[Int(0)]
MMA_N
comptime MMA_N = shape[Int(1)]
Methods
mma
static def mma[a_layout: TensorLayout, b_layout: TensorLayout, c_layout: TensorLayout](a_reg: TileTensor[in_type, a_layout, address_space=AddressSpace.LOCAL], b_reg: TileTensor[in_type, b_layout, address_space=AddressSpace.LOCAL], c_reg: TileTensor[out_type, c_layout, MutUntrackedOrigin, address_space=AddressSpace.LOCAL])
Execute group_size MMA operations across the K dimension.
Mirrors TiledTensorCore.mma: iterates group_size k-steps, tiles A/B registers per step via vectorize, and accumulates into C.
Parameters:
- a_layout (
TensorLayout): Inferred layout of A register tile. - b_layout (
TensorLayout): Inferred layout of B register tile. - c_layout (
TensorLayout): Inferred layout of C register tile.
Args:
- a_reg (
TileTensor[in_type, a_layout, address_space=AddressSpace.LOCAL]): A fragments [num_m_mmas, group_size * a_frag_size]. - b_reg (
TileTensor[in_type, b_layout, address_space=AddressSpace.LOCAL]): B fragments [num_n_mmas, group_size * a_frag_size]. - c_reg (
TileTensor[out_type, c_layout, MutUntrackedOrigin, address_space=AddressSpace.LOCAL]): Accumulator [num_m_mmas, num_n_mmas * c_frag_size].