pto.mad_mx_bias

pto.mad_mx_bias is part of the Cube MAD Ops.

Summary

Bias-init MX (microscaled) cube matrix multiply: dst[m, n] = mx_product[m, n] + bias[n].

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

Mechanism

Combines the MX scaling of pto.mad_mx with the bias-init seed of pto.mad_bias. The accumulator starts from bias[n] instead of zero.

Syntax

pto.mad_mx_bias %lhs, %rhs, %dst, %bias, %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>, !pto.ptr<C, bt>, i64, i64, i64

Inputs

Same parameter shape as pto.mad_bias, with MX %lhs / %rhs scale payload requirements from pto.mad_mx.

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

Expected Outputs

Result Type Description
None — Writes the produced M x N MX-scaled tile to L0C with bias-init seed.

Side Effects

Engages the CUBE pipe; reads %bias from BT and MX scale payloads associated with %lhs / %rhs; writes to L0C.

Constraints

Constraints

  • All constraints from pto.mad_mx (MX dtype combination, scale payload prerequisites, K grouping rule).
  • All %bias constraints from pto.mad_bias: %bias must be in bt space with element type matching %dst; only N values are consumed.

Examples

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