TROWARGMAX¶

Tile Operation Diagram¶

Introduction¶

Get the column index of the maximum element, or both value and column index of the maximum element for each row.

Math Interpretation¶

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

Assembly Syntax¶

Synchronous form:

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

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

IR Level 1 (SSA)¶

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

IR Level 2 (DPS)¶

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

C++ Intrinsic¶

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

Output index only:

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

Output both value and index:

template <typename TileDataOutVal, typename TileDataOutIdx, typename TileDataIn, typename TileDataTmp,
          typename... WaitEvents>
PTO_INST RecordEvent TROWARGMAX(TileDataOutVal &dstVal, TileDataOutIdx &dstIdx, TileDataIn &src, TileDataTmp &tmp,
                                WaitEvents &... events)

Constraints¶

General constraints / checks¶

• Supported source element types: half, float, int32_t, int16_t (A2A3). On A5 any 2-/4-byte source is accepted (half, bfloat16_t, int16_t, uint16_t, float, int32_t, uint32_t).
• src must use standard ND layout: row-major and non-fractal (BLayout::RowMajor, SLayout::NoneBox).
• When output index only:
  • dst and src must be TileType::Vec.
  • Supported destination element types: uint32_t, int32_t.
  • Runtime checks follow the shared row-reduce check path:
    • src.GetValidRow() != 0
    • src.GetValidCol() != 0
    • src.GetValidRow() == dst.GetValidRow()
  • dst is checked through the shared row-reduce-index path and may use either of these non-fractal layouts:
    • DN layout with one column (BLayout::ColMajor, Cols == 1), or
    • ND layout whose valid column count is 1.
• When output both value and index:
  • dstVal, dstIdx, src must be TileType::Vec.
  • Supported destination element types: uint32_t, int32_t.
  • Runtime checks follow the shared row-reduce check path:
    • src.GetValidRow() != 0
    • src.GetValidCol() != 0
    • src.GetValidRow() == dstIdx.GetValidRow()
    • src.GetValidRow() == dstVal.GetValidRow()
  • dstVal, dstIdx are checked through the shared row-reduce-index path and may use either of these non-fractal layouts:
    • DN layout with one column (BLayout::ColMajor, Cols == 1), or
    • ND layout whose valid column count is 1.

About temporary tile tmp¶

• Temporary tile is only used by A2A3, A5 accepts tmp tile but leaves it unused.
• The A2A3 implementation selects one of three code paths based on srcValidCol relative to elementPerRepeat (abbreviated elemPerRpt below):

Case 1: srcValidCol <= elemPerRpt¶

• Index-only mode: tmp is not used. The hardware vcmax instruction writes directly to dst.
• Value + Index mode: tmp is used as a small buffer (2 elements per row: one value + one index). tmp may use either of these non-fractal layouts:
  • DN layout with one column (BLayout::ColMajor, Cols == 1), rows is twice of src.
  • ND layout whose valid column count is 2, rows is the same as src.

Case 2: elemPerRpt < srcValidCol <= elemPerRpt² (Stage 1 reduction)¶

• tmp is used for a single-stage reduction.
• Rows of tmp tile is equal to src.
• tmp tile's stride per row can be calculated using:

R1 = ceil(validCol / elemPerRpt)
stride = (ceil(R1 * 2 / elemPerBlock) + ceil(R1 / elemPerBlock)) * elemPerBlock

Case 3: srcValidCol > elemPerRpt² (Stage 2 reduction)¶

• tmp is used for a two-stage reduction, requiring more space than Stage 1.
• Rows of tmp tile is equal to src.
• tmp tile's stride per row can be calculated using:

R1 = ceil(validCol / elemPerRpt)
R2 = ceil(R1 / elemPerRpt)
stage1_size = ceil(R1 * 2 / elemPerBlock) * elemPerBlock
stage2_end  = ceil(R1 / elemPerBlock) * elemPerBlock + ceil(R2 * 2 / elemPerBlock) * elemPerBlock
stride = max(stage1_size, stage2_end) + 2

• The + 2 accounts for the final value + index result stored at the end of each row's tmp region.

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, uint32_t, 16, 1, BLayout::ColMajor>;
  using DstValT = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor>;
  using TmpT = Tile<TileType::Vec, float, 16, 16>;
  SrcT src;
  DstT dst;
  DstValT dst;
  TmpT tmp;
  TROWARGMAX(dst, src, tmp);
  TROWARGMAX(dstVal, 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, uint32_t, 16, 1, BLayout::ColMajor>;
  using DstValT = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor>;
  using TmpT = Tile<TileType::Vec, float, 16, 16>;
  SrcT src;
  DstT dst;
  DstValT dst;
  TmpT tmp;
  TASSIGN(src, 0x1000);
  TASSIGN(dst, 0x2000);
  TASSIGN(dstVal, 0x3000);
  TASSIGN(tmp, 0x4000);
  TROWARGMAX(dst, src, tmp);
  TROWARGMAX(dstVal, dst, src, tmp);
}

ASM Form Examples¶

Auto Mode¶

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

PTO Assembly Form¶

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