pto.mad_acc

pto.mad_acc is part of the Cube MAD Ops.

Summary

Accumulating cube matrix multiply: dst[m, n] = dst[m, n] + sum_k(lhs[m, k] * rhs[k, n]).

Mechanism

Like pto.mad, but adds the freshly-computed product to the existing L0C accumulator state instead of overwriting it. Typical use is K-axis tiling, where successive MAD calls accumulate partial sums along K until the full reduction is complete.

Syntax

pto.mad_acc %lhs, %rhs, %dst, %m, %n, %k
  unit_flag(check_only | check_and_set)?
  disable_gemv?
  (sat | nosat)?
  tf32_mode(round_even | round_away)?
  n_dir?
  : !pto.ptr<A, l0a>, !pto.ptr<B, l0b>, !pto.ptr<C, l0c>, i64, i64, i64

Inputs

Same parameter shape and semantics as pto.mad. See MAD Common Clauses for the optional clauses.

Expected Outputs

Result Type Description
None — Updates the existing M x N tile in L0C in place. No SSA result.

Side Effects

Engages the CUBE pipe; reads from and writes to L0C. The caller is responsible for ensuring the L0C tile has been initialized (typically by an initial pto.mad or pto.mad_bias on the same %dst before the first pto.mad_acc).

Constraints

Same as pto.mad.

Examples

// K-axis tiling: initial pto.mad then repeated pto.mad_acc.
pto.mad %l0a_k0, %l0b_k0, %l0c, %c16_i64, %c16_i64, %c32_i64
  : !pto.ptr<f16, l0a>, !pto.ptr<f16, l0b>, !pto.ptr<f32, l0c>, i64, i64, i64

pto.mad_acc %l0a_k1, %l0b_k1, %l0c, %c16_i64, %c16_i64, %c32_i64 unit_flag(check_only)
  : !pto.ptr<f16, l0a>, !pto.ptr<f16, l0b>, !pto.ptr<f32, l0c>, i64, i64, i64