TROWEXPANDMAX

Tile Operation Diagram

Mode 1 — Scalar per row (ColMajor src1)

TROWEXPANDMAX Mode 1 tile operation

Mode 2 — 32-byte block per row (RowMajor src1)

TROWEXPANDMAX Mode 2 tile operation

Introduction

Row-wise broadcast max: take max(src0, src1) where the broadcast operand provides per-row values.

The instruction supports two modes determined by the layout of the expanded operand (src1 when src0 matches dst shape, or src0 when src1 matches dst shape):

  • Mode 1: The expanded operand is ColMajor with a single column (one scalar per row). Each scalar is broadcast across the entire row.
  • Mode 2: The expanded operand is RowMajor with 32 / sizeof(T) columns per row (a 32-byte block per row).

Math Interpretation

Let R = dst.GetValidRow() and C = dst.GetValidCol().

Mode 1

Let s_i be the per-row scalar taken from the expanded operand (one value per row, ColMajor layout).

For 0 <= i < R and 0 <= j < C:

\[ \mathrm{dst}_{i,j} = \max(\mathrm{src0}_{i,j}, s_i) \]

Mode 2

Let b_i be the 32-byte block for row i taken from the expanded operand (RowMajor, 32 / sizeof(T) values per row). The block naturally repeats every elementsPerRepeat elements within a row.

For 0 <= i < R and 0 <= j < C:

\[ \mathrm{dst}_{i,j} = \max(\mathrm{src0}_{i,j}, b_i[\,j \bmod (32 / \mathit{sizeof}(T))\,]) \]

Assembly Syntax

Synchronous form:

%dst = trowexpandmax %src0, %src1 : !pto.tile<...>, !pto.tile<...> -> !pto.tile<...>

AS Level 1 (SSA)

%dst = pto.trowexpandmax %src0, %src1 : !pto.tile<...>, !pto.tile<...> -> !pto.tile<...>

AS Level 2 (DPS)

pto.trowexpandmax ins(%src0, %src1 : !pto.tile_buf<...>, !pto.tile_buf<...>) outs(%dst : !pto.tile_buf<...>)

C++ Intrinsic

Declared in include/pto/common/pto_instr.hpp:

template <typename TileDataDst, typename TileDataSrc0, typename TileDataSrc1, typename... WaitEvents>
PTO_INST RecordEvent TROWEXPANDMAX(TileDataDst &dst, TileDataSrc0 &src0, TileDataSrc1 &src1, WaitEvents &... events);

template <typename TileDataDst, typename TileDataSrc0, typename TileDataSrc1, typename TileDataTmp,
          typename... WaitEvents>
PTO_INST RecordEvent TROWEXPANDMAX(TileDataDst &dst, TileDataSrc0 &src0, TileDataSrc1 &src1, TileDataTmp &tmp, WaitEvents &... events);

Constraints

  • TileDataDst::DType == TileDataSrc0::DType == TileDataSrc1::DType
  • TileDataDst::DType, TileDataSrc0::DType, TileDataSrc1::DType must be one of: half, float, int16, int32 for A2, A3 and A5, uint16, uint32, bfloat16_t, int8, uint8 for A5.
  • TileDataDst must be RowMajor (TileDataDst::isRowMajor == true).
  • Exactly one of src0 or src1 must have the same valid shape as dst (i.e., validRow == dst.validRow and validCol == dst.validCol). That operand is the full-sized operand. The other operand is the expanded operand (row-broadcast source).
  • The full-sized operand must be RowMajor (isRowMajor == true).

Mode 1 — Expanded operand is ColMajor (scalar per row)

When the expanded operand is ColMajor (isRowMajor == false):

  • Its valid column count must be 1 (one scalar per row): srcX.GetValidCol() == 1.
  • Its valid row count must equal dst.GetValidRow(): srcX.GetValidRow() == dst.GetValidRow().

Mode 2 — Expanded operand is RowMajor (32-byte block per row)

When the expanded operand is RowMajor (isRowMajor == true):

  • Its valid column count must be 32 / sizeof(T) (a 32-byte block per row): srcX.GetValidCol() == 32 / sizeof(T).
  • For half / int16 / uint16: validCol == 16.
  • For float / int32 / uint32: validCol == 8.
  • Its valid row count must equal dst.GetValidRow(): srcX.GetValidRow() == dst.GetValidRow().

Additional target-specific constraints

Exact layout, fractal, and alignment constraints may vary by backend target. See backend headers under include/pto/npu/*/TRowExpand*.hpp.

Temporary tile

The C++ API provides an overload with an explicit TileDataTmp &tmp. This overload only supports Mode 1 (ColMajor expanded operand, scalar per row).

  • A2A3: The tmp tile is used as a broadcast buffer. The per-row scalar values from the ColMajor expanded operand are broadcast via the vbrcb instruction into the tmp buffer, creating a 32-byte block per row, which is then used as the expanded operand in the binary operation. The vbrcb instruction uses a repeat stride of 8 blocks (256 bytes) between repeat groups, processing 8 rows per repeat. Minimum tmp size calculation:
    • Common parameters:
      • R = dst.GetValidRow(), T = TileDataDst::DType.
    • For R < 256: $$ \text{tmpSize} = \left\lceil\frac{R}{8}\right\rceil \times 256 \text{ bytes} $$
    • For R >= 256:
      • The operation is looped, with at most 30 repeats (240 rows) per loop iteration. The tmp buffer is reused across loops, so the per-loop requirement is: $$ \text{tmpSize} = 30 \times 256 = 7680 \text{ bytes} $$
    • A compact shape-independent upper bound for any Mode 1 invocation is 8KB (8192 bytes).
    • The 3-arg overload (without tmp) supports both Mode 1 and Mode 2. For Mode 1, it uses an internal 8KB buffer (TMP_UB_OFFSET). For Mode 2, no broadcast buffer is needed.
  • A5: The tmp tile is accepted and ignored ([[maybe_unused]]). A5 hardware supports row-broadcast natively via the vlds instruction's broadcast modes, so no scratch buffer is required.

Examples

See related examples in docs/isa/ and docs/coding/tutorials/.

ASM Form Examples

Auto Mode

# Auto mode: compiler/runtime-managed placement and scheduling.
%dst = pto.trowexpandmax %src0, %src1 : !pto.tile<...>, !pto.tile<...> -> !pto.tile<...>

Manual Mode

# Manual mode: resources must be bound explicitly before issuing the instruction.
# Optional for tile operands:
# pto.tassign %arg0, @tile(0x1000)
# pto.tassign %arg1, @tile(0x2000)
%dst = pto.trowexpandmax %src0, %src1 : !pto.tile<...>, !pto.tile<...> -> !pto.tile<...>

PTO Assembly Form

%dst = trowexpandmax %src0, %src1 : !pto.tile<...>, !pto.tile<...> -> !pto.tile<...>
# AS Level 2 (DPS)
pto.trowexpandmax ins(%src0, %src1 : !pto.tile_buf<...>, !pto.tile_buf<...>) outs(%dst : !pto.tile_buf<...>)