FIXPIPE Writeback Model

FIXPIPE is the cube core's dedicated writeback / post-processing pipeline. It moves L0C accumulator results out to L1, UB, or GM while applying the layout conversion (NZ → ND) required by the destination, plus optional dequantization / scale / clip / activation post-processing.

This page describes the FIXPIPE addressing model used by the three writeback ops:

Source Layout

The L0C source tile is laid out as N1 M1 M0 N0 (FRACTAL_NZ: col-major outer, row-major inner). FIXPIPE addresses one M-row of results at a time and emits them in the destination memory's natural order.

NZ → ND Conversion at Writeback

For each cube fragment in L0C, FIXPIPE applies:

C_nz[n1][m1][m0][n0]  -->  C_nd[m1*M0 + m0][n1*N0 + n0]

The conversion is fused with the writeback — no separate explicit transpose step is required. Destination strides are expressed in ND coordinates on the FIXPIPE op.

Dual-Destination Broadcast (1 → 2 Cube-to-Vector)

When the FIXPIPE destination is a Vector block UB, the cube can simultaneously broadcast to both AIV0 and AIV1 UB regions via the dedicated on-chip data path, with the tile split either along the row axis (DualModeSplitM) or the column axis (DualModeSplitN):

Split AIV0 receives AIV1 receives
Split-M (rows) Upper [M/2, N] in ND Lower [M/2, N] in ND
Split-N (cols) Left [M, N/2] in ND Right [M, N/2] in ND

This 1→2 broadcast with in-hardware tile split is the architectural basis for 1:2 Cube-to-Vector tile distribution and is selected as an attribute on pto.mte_l0c_ub.

Burst / Loop Model

Like the scalar DMA and cube data-movement ops, FIXPIPE writeback uses the grouped nburst(...) / loop(...) clause form to express row-stride and outer-stride repetition without external configuration registers.

Post-Processing Hooks

FIXPIPE optionally applies the following post-processing along the writeback path, configured via clauses or auxiliary FB payload loaded by pto.mte_l1_fb:

  • Dequantization (per-channel scale / zero-point)
  • Clip / saturate to destination element-type range
  • Activation (ReLU / clipped linear, target-defined)

Per-op pages document which post-processing clauses each pto.mte_l0c_* variant accepts.

Synchronization Around FIXPIPE

FIXP is one of the four AIC-side issue queues (alongside MTE2, MTE1, CUBE). The standard producer / consumer chain is:

CUBE (pto.mad*)  --set_flag(CUBE -> FIXP)-->  FIXP (pto.mte_l0c_*)  -->  L1 / UB / GM

After a pto.mad* finishes a tile in L0C, the producer must set_flag from PIPE_CUBE to PIPE_FIXP (using one of the configured event IDs); the FIXPIPE consumer issues a matching wait_flag before issuing pto.mte_l0c_* against the same L0C tile. Failure to synchronize results in a read of in-flight L0C state and is a verifier error.