TROWPROD¶

Tile Operation Diagram¶

Introduction¶

Reduce each row by multiplying across columns.

Math Interpretation¶

Let R = src.GetValidRow() and C = src.GetValidCol(). For 0 <= i < R:

Assembly Syntax¶

Synchronous form:

%dst = trowprod %src : !pto.tile<...> -> !pto.tile<...>

Lowering may introduce internal scratch tiles; the C++ intrinsic requires an explicit tmp operand.

AS Level 1 (SSA)¶

%dst = pto.trowprod %src, %tmp : (!pto.tile<...>, !pto.tile<...>) -> !pto.tile<...>

AS Level 2 (DPS)¶

pto.trowprod ins(%src, %tmp : !pto.tile_buf<...>, !pto.tile_buf<...>) outs(%dst : !pto.tile_buf<...>)

C++ Intrinsic¶

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

template <typename TileDataOut, typename TileDataIn, typename TileDataTmp, typename... WaitEvents>
PTO_INST RecordEvent TROWPROD(TileDataOut &dst, TileDataIn &src, TileDataTmp &tmp, WaitEvents &... events);

Constraints¶

General constraints / checks¶

• dst and src must both be TileType::Vec.
• src must use standard ND layout: row-major and non-fractal (BLayout::RowMajor, SLayout::NoneBox).
• dst must use one of the following non-fractal layouts:
  • ND layout (BLayout::RowMajor, SLayout::NoneBox), or
  • DN layout with exactly one column (BLayout::ColMajor, SLayout::NoneBox, Cols == 1).
• dst and src must use the same element type.
• Runtime valid-region checks:
  • src.GetValidRow() != 0
  • src.GetValidCol() != 0
  • src.GetValidRow() == dst.GetValidRow()
• The intrinsic signature requires an explicit tmp operand.

A2A3 implementation checks¶

• Supported element types: half, float, int32_t, int16_t.

A5 implementation checks¶

• Supported element types: half, float, int32_t, int16_t.
• In the currently inspected implementation path, the enforced constraints are on src and dst.
• No extra shape/layout assertions on tmp are enforced in the current implementation path.

Temporary Space¶

A2A3¶

tmp is used as a per-row accumulator buffer. For each row, the implementation initializes tmp with 1.0 and then multiplies blocks of src data into tmp using vmul. After all blocks are accumulated, the scalar-mode pipeline reads tmp elements and computes the final product.

• tmp must have the same element type as src/dst.
• tmp size: at least 1 row and BLOCK_BYTE_SIZE / sizeof(T) columns (i.e., 1 block: 8 elements for float/int32_t, 16 elements for half/int16_t).
• A safe default: set tmp to the same shape as src.

A5¶

tmp is accepted by the interface but not used by the A5 implementation. The A5 backend uses vector register-based reduction (vmul + vintlv for tree reduction) and does not require scratch tile storage. tmp is retained in the C++ intrinsic signature solely for API compatibility with A2A3.

Examples¶

Auto¶

#include <pto/pto-inst.hpp>

using namespace pto;

void example_auto() {
  using SrcT = Tile<TileType::Vec, float, 16, 16>;
  using DstT = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor>;
  using TmpT = Tile<TileType::Vec, float, 16, 16>;
  SrcT src;
  DstT dst;
  TmpT tmp;
  TROWPROD(dst, src, tmp);
}

Manual¶

#include <pto/pto-inst.hpp>

using namespace pto;

void example_manual() {
  using SrcT = Tile<TileType::Vec, float, 16, 16>;
  using DstT = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor>;
  using TmpT = Tile<TileType::Vec, float, 16, 16>;
  SrcT src;
  DstT dst;
  TmpT tmp;
  TASSIGN(src, 0x1000);
  TASSIGN(dst, 0x2000);
  TASSIGN(tmp, 0x3000);
  TROWPROD(dst, src, tmp);
}

ASM Form Examples¶

Auto Mode¶

# Auto mode: compiler/runtime-managed placement and scheduling.
%dst = pto.trowprod %src, %tmp : (!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.trowprod %src, %tmp : (!pto.tile<...>, !pto.tile<...>) -> !pto.tile<...>

PTO Assembly Form¶

%dst = trowprod %src : !pto.tile<...> -> !pto.tile<...>
# AS Level 2 (DPS)
pto.trowprod ins(%src, %tmp : !pto.tile_buf<...>, !pto.tile_buf<...>) outs(%dst : !pto.tile_buf<...>)