pto.mad_mx_acc

pto.mad_mx_acc is part of the Cube MAD Ops.

Summary

Accumulating MX (microscaled) cube matrix multiply: dst[m, n] = dst[m, n] + mx_product[m, n].

See MX Matmul Model for the per-K-group scaled multiply-accumulate.

Mechanism

Like pto.mad_mx but adds the MX-scaled product to existing L0C state. Typical use is K-axis tiling for MX GEMM.

Syntax

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

Inputs

Same parameter shape as pto.mad_mx.

See MAD Common Clauses for the optional clauses. tf32_mode(...) is not accepted.

Expected Outputs

Result Type Description
None — Updates the existing M x N tile in L0C with MX-scaled accumulation.

Side Effects

Same as pto.mad_mx. The caller is responsible for ensuring the L0C tile has been initialized (typically by an initial pto.mad_mx or pto.mad_mx_bias on the same %dst).

Constraints

Same as pto.mad_mx.

Examples

pto.mad_mx_acc %l0a, %l0b, %l0c, %c16_i64, %c16_i64, %c64_i64
  : !pto.ptr<f8E4M3FN, l0a>, !pto.ptr<f8E4M3FN, l0b>, !pto.ptr<f32, l0c>, i64, i64, i64