pto.mte_ub_gm

pto.mte_ub_gm is part of the DMA Copy instruction set.

PTO ISA v0.6 surface

The v0.6 PTO micro-instruction surface replaces the earlier pto.copy_ubuf_to_gm plus standalone set_loop_size_* / set_loop_stride_* configuration ops with a single grouped instruction: pto.mte_ub_gm with inline nburst(...) and optional loop(...) clauses.

Summary

Execute a grouped UB→GM DMA transfer. nburst(...) defines the innermost repeated burst transfer, and optional loop(...) groups add outer repetition levels.

Mechanism

The MTE3 engine reads %n_burst source rows from %ub_src and writes them to %gm_dst. Each row transfers %len_burst contiguous bytes, and the source / destination stride operands give the start-to-start byte distance from one row to the next. Optional outer loop(...) groups wrap nburst(...) to express multi-level repetition without external loop-config state. Padding bytes added during a previous GM→UB load are stripped: MTE3 reads %len_burst bytes from each 32B-aligned UB row and writes only valid data to GM.

Syntax

pto.mte_ub_gm %ub_src, %gm_dst, %len_burst
  nburst(%n_burst, %src_stride, %dst_stride)
  [loop(%loop_count, %loop_src_stride, %loop_dst_stride)]*
  : !pto.ptr<T, ub>, !pto.ptr<T, gm>, i64, i64, i64, i64,
    [loop i64, i64, i64,]*

Inputs

Parameter Width Description
%ub_src ptr UB source pointer (!pto.ptr<T, ub>, 32B-aligned)
%gm_dst ptr GM destination pointer (!pto.ptr<T, gm>)
%len_burst 16 bits Contiguous bytes transferred per burst row
nburst(%n_burst, %src_stride, %dst_stride) 16 bits / 21 bits / 40 bits Required innermost burst group: count, UB source stride, GM destination stride
loop(%loop_count, %loop_src_stride, %loop_dst_stride) 21 bits / 21 bits / 40 bits Optional outer repetition group: count, UB source stride, GM destination stride

Expected Outputs

Result Type Description
None — This form does not return SSA values; it writes data into Global Memory.

Side Effects

Reads UB-visible storage and writes GM-visible storage. The MTE3 pipe is engaged for the duration of the transfer; downstream consumers (and ordering against further GM writes) must synchronize through pto.set_flag / pto.wait_flag (PIPE_V → PIPE_MTE3, and pto.mem_bar between back-to-back stores to the same GM address).

Constraints

Constraints

  • nburst(...) is always required.
  • Each loop(...) group must be provided as a complete triple when present.
  • nburst(...) is the innermost group.
  • loop(...) groups are ordered from inner to outer; the first loop(...) group wraps nburst(...), and each additional loop(...) group wraps all earlier groups.
  • %ub_src MUST be 32-byte aligned.

Exceptions

Exceptions

  • The verifier rejects illegal operand shapes, malformed clause groups, and attribute combinations not valid for the selected target profile.
  • Any additional illegality stated in the constraints section is also part of the contract.

Target-Profile Restrictions

Target-Profile Restrictions
  • CPU simulation preserves the visible copy contract but may not expose all DMA overlap hazards.
  • A2/A3 and A5 may narrow supported element sizes, row widths, or cache-control semantics.

Examples

// Two-level outbound transfer: rows x tiles.
pto.mte_ub_gm %ub_in, %gm_out, %len_burst
  nburst(%rows, %ub_row_stride, %gm_row_stride)
  loop(%tiles, %ub_tile_stride, %gm_tile_stride)
  : !pto.ptr<f16, ub>, !pto.ptr<f16, gm>, i64, i64, i64, i64,
    loop i64, i64, i64
// Three-level outbound transfer: rows x tiles x batches.
pto.mte_ub_gm %ub_in, %gm_out, %len_burst
  nburst(%rows, %ub_row_stride, %gm_row_stride)
  loop(%tiles, %ub_tile_stride, %gm_tile_stride)
  loop(%batches, %ub_batch_stride, %gm_batch_stride)
  : !pto.ptr<f16, ub>, !pto.ptr<f16, gm>, i64, i64, i64, i64,
    loop i64, i64, i64, loop i64, i64, i64