TCONCAT

Tile Operation Diagram

Basic Form (3 arguments)

Indexed Form (5-6 arguments)

Introduction

Concatenate two source tiles (src0 and src1) horizontally into a destination tile (dst) along the column dimension. Each row of dst contains the concatenation of corresponding rows from src0 and src1.

TCONCAT is used for:

• Concatenating two tiles along the column axis (horizontal concatenation)
• Joining tiles in attention and transformer architectures (e.g., concatenating KV cache entries)
• Combining partial results from split operations

Math Interpretation

For each row i in the valid region:

$${\mathrm{d}\mathrm{s}\mathrm{t}}_{i,j}=\{\begin{array}{ll}{\mathrm{s}\mathrm{r}\mathrm{c}0}_{i,j}& \text{if }0\le j<\mathrm{v}\mathrm{a}\mathrm{l}\mathrm{i}\mathrm{d}\mathrm{C}\mathrm{o}\mathrm{l}\mathrm{s}0\\ {\mathrm{s}\mathrm{r}\mathrm{c}1}_{i,j-\mathrm{v}\mathrm{a}\mathrm{l}\mathrm{i}\mathrm{d}\mathrm{C}\mathrm{o}\mathrm{l}\mathrm{s}0}& \text{if }\mathrm{v}\mathrm{a}\mathrm{l}\mathrm{i}\mathrm{d}\mathrm{C}\mathrm{o}\mathrm{l}\mathrm{s}0\le j<\mathrm{v}\mathrm{a}\mathrm{l}\mathrm{i}\mathrm{d}\mathrm{C}\mathrm{o}\mathrm{l}\mathrm{s}0+\mathrm{v}\mathrm{a}\mathrm{l}\mathrm{i}\mathrm{d}\mathrm{C}\mathrm{o}\mathrm{l}\mathrm{s}1\end{array}$$

Where validCols0 = src0.GetValidCol() and validCols1 = src1.GetValidCol().

Assembly Syntax

AS Level 1 (SSA)

%dst = pto.tconcat %src0, %src1 : (!pto.tile<...>, !pto.tile<...>) -> !pto.tile<...>

AS Level 2 (DPS)

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

C++ Intrinsic

Declared in include/pto/npu/a5/TConcat.hpp:

template <typename TileDst, typename TileSrc0, typename TileSrc1>
PTO_INST void TCONCAT(TileDst &dst, TileSrc0 &src0, TileSrc1 &src1);

template <typename TileDst, typename TileSrc0, typename TileSrc1, typename TileSrc0Idx, typename TileSrc1Idx>
PTO_INST void TCONCAT(TileDst &dst, TileSrc0 &src0, TileSrc1 &src1, TileSrc0Idx &src0Idx, TileSrc1Idx &src1Idx);

template <typename TileDst, typename TileSrc0, typename TileSrc1, typename TileDstIdx, typename TileSrc0Idx, typename TileSrc1Idx>
PTO_INST void TCONCAT(TileDst &dst, TileSrc0 &src0, TileSrc1 &src1, TileDstIdx &dstIdx, TileSrc0Idx &src0Idx, TileSrc1Idx &src1Idx);

Constraints

General constraints / checks

• TCONCAT has three overload variants:
  • basic form: TCONCAT(dst, src0, src1) - concatenates full valid regions
  • indexed form (5 args): TCONCAT(dst, src0, src1, src0Idx, src1Idx) - uses per-row index tiles to specify dynamic column counts
  • indexed form (6 args): TCONCAT(dst, src0, src1, dstIdx, src0Idx, src1Idx) - also outputs the concatenated column count per row
• All tiles must have TileType::Vec (vector tiles)
• All tiles must use row-major layout (isRowMajor == true)

Shape constraints

• Basic form:
  • dst.GetValidRow() == src0.GetValidRow() == src1.GetValidRow()
  • dst.GetValidCol() == src0.GetValidCol() + src1.GetValidCol()
• Indexed form:
  • Same row count constraints as basic form
  • Column counts are determined dynamically from index tiles
  • dstIdx.GetValidRow() == 1 for 6-argument form

Data type constraints

• Supported element types: int8_t, uint8_t, int16_t, uint16_t, int32_t, uint32_t, half, bfloat16_t, float
• Source and destination tiles must have identical element type
• Index tiles must use integer types (int8_t, uint8_t, int16_t, uint16_t, int32_t, uint32_t)

A5 implementation checks

• All tiles must be TileType::Vec
• All tiles must be row-major layout
• validRows must not exceed the physical tile rows for any operand
• Index tiles (if provided) must satisfy type compatibility checks

Examples

Auto Mode

#include <pto/pto-inst.hpp>

using namespace pto;

void example_auto() {
    using TileT = Tile<TileType::Vec, float, 16, 32>;
    TileT src0(16, 16);
    TileT src1(16, 16);
    TileT dst(16, 32);

    TCONCAT(dst, src0, src1);
}

Manual Mode

#include <pto/pto-inst.hpp>

using namespace pto;

void example_manual() {
    using TileT = Tile<TileType::Vec, half, 16, 64, BLayout::RowMajor, 16, 64>;
    TileT src0, src1, dst;

    TASSIGN(src0, 0x1000);
    TASSIGN(src1, 0x2000);
    TASSIGN(dst, 0x3000);

    src0.SetValidRegion(16, 32);
    src1.SetValidRegion(16, 32);

    TCONCAT(dst, src0, src1);
}

Indexed Form Example

#include <pto/pto-inst.hpp>

using namespace pto;

void example_indexed() {
    using TileT = Tile<TileType::Vec, float, 16, 64>;
    using IdxTileT = Tile<TileType::Vec, int32_t, 16, 1>;

    TileT src0(16, 32);
    TileT src1(16, 32);
    TileT dst(16, 64);
    IdxTileT src0Idx, src1Idx;

    TCONCAT(dst, src0, src1, src0Idx, src1Idx);
}

ASM Form Examples

Auto Mode

# Auto mode: compiler/runtime-managed placement and scheduling.
%dst = pto.tconcat %src0, %src1 : (!pto.tile<16x32xf32>, !pto.tile<16x32xf32>) -> !pto.tile<16x64xf32>

Manual Mode

# Manual mode: resources must be bound explicitly before issuing the instruction.
# Optional for tile operands:
# pto.tassign %src0, @tile(0x1000)
# pto.tassign %src1, @tile(0x2000)
# pto.tassign %dst, @tile(0x3000)
%dst = pto.tconcat %src0, %src1 : (!pto.tile<...>, !pto.tile<...>) -> !pto.tile<...>

Related Instructions

• TINSERT - Insert a sub-tile into a destination tile at specified offset
• TEXTRACT - Extract a sub-tile from a source tile
• TRESHAPE - Reinterpret a tile as another tile type/shape