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
ssd_chunk_output
def ssd_chunk_output[kernel_dtype: DType, SLT: TensorLayout, SEngine: TensorEngine, A_LT: TensorLayout, D_LT: TensorLayout, dt_bias_LT: TensorLayout, y_LT: TensorLayout, Engine: TensorEngine, XFn: def[width: Int, alignment: Int](Int, Int, Int) -> SIMD[kernel_dtype, width] & RegisterPassable & ImplicitlyCopyable, BFn: def[width: Int, alignment: Int](Int, Int, Int) -> SIMD[kernel_dtype, width] & RegisterPassable & ImplicitlyCopyable, CFn: def[width: Int, alignment: Int](Int, Int, Int) -> SIMD[kernel_dtype, width] & RegisterPassable & ImplicitlyCopyable, DSTATE: Int, use_grouped_cb: Bool = False](slots_i: Int32, nheads_i: Int32, head_dim_i: Int32, ngroups_i: Int32, D: TileTensor[kernel_dtype, D_LT, ImmUntrackedOrigin, Engine=Engine], y: TileTensor[kernel_dtype, y_LT, MutUntrackedOrigin, Engine=Engine], info: TileTensor[.int32, SLT, MutUntrackedOrigin, Engine=SEngine], cs: TileTensor[.float32, SLT, ImmUntrackedOrigin, Engine=SEngine], delta_in: TileTensor[.float32, SLT, MutUntrackedOrigin, Engine=SEngine], states: TileTensor[.float32, SLT, MutUntrackedOrigin, Engine=SEngine], cb_in: TileTensor[.float32, SLT, ImmUntrackedOrigin, Engine=SEngine], x_fn: XFn, b_fn: BFn, c_fn: CFn)
Computes y of one 64-row tile of a chunk for one head on tensor cores.
Grid (slots, nheads, (CHUNK / TILE) * head_dim / TILE), block
OUT_THREADS (4 warps, 16 rows of the tile each). Operands are bf16 and
accumulation is fp32 (m16n8k16). Each warp keeps its C fragments in
registers: the carry-in term C S'^T and, per 16 columns, C B^T are
MMAs, the decay mask is applied to the accumulator fragment, which is
already the A-operand layout of the next MMA M X.