文档
下载
文档中心
TINSERT

TINSERT

Instruction Diagram

Introduction

Inserts a source sub-tile into a destination tile at the (indexRow, indexCol) offset. Conceptually, it is the inverse operation of TEXTRACT.

TINSERT is used for:

  • Acc → Mat insertion (with optional relu, scalar quantization, or vector quantization)
  • Acc → Vec insertion (with optional AccToVecMode, relu, scalar quantization, or vector quantization) (Ascend 950PR/Ascend 950DT)
  • Vec → Mat insertion (ND and NZ layouts) (Ascend 950PR/Ascend 950DT)
  • Vec → Vec insertion (ND and NZ layouts) (Ascend 950PR/Ascend 950DT)
  • NZ split insertion (SPLIT2, SPLIT4) (Ascend 950PR/Ascend 950DT)

Mathematical Semantics

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

dstindexRow+i,  indexCol+j=srci,j\mathrm{dst}_{\mathrm{indexRow}+i,\;\mathrm{indexCol}+j} = \mathrm{src}_{i,j}

Assembly Syntax

Synchronous form:

%dst = tinsert %src[%r0, %r1] : !pto.tile<...> -> !pto.tile<...>

AS Level 1 (SSA)

%dst = pto.tinsert %src, %idxrow, %idxcol : (!pto.tile<...>, dtype, dtype) -> !pto.tile<...>

AS Level 2 (DPS)

pto.tinsert ins(%src, %idxrow, %idxcol : !pto.tile_buf<...>, dtype, dtype) outs(%dst : !pto.tile_buf<...>)

C++ Built-in APIs

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

The public include header is <pto/pto-inst.hpp>, and the internal declaration is located in pto/common/pto_instr.hpp.

template <typename DstTileData, typename SrcTileData, typename... WaitEvents>
PTO_INST RecordEvent TINSERT(DstTileData &dst, SrcTileData &src,
                             uint16_t indexRow, uint16_t indexCol,
                             WaitEvents &... events);

template <typename DstTileData, typename SrcTileData, ReluPreMode reluMode,
          typename... WaitEvents>
PTO_INST RecordEvent TINSERT(DstTileData &dst, SrcTileData &src,
                             uint16_t indexRow, uint16_t indexCol,
                             WaitEvents &... events);

template <typename DstTileData, typename SrcTileData, AccToVecMode mode,
          ReluPreMode reluMode = ReluPreMode::NoRelu, typename... WaitEvents>
PTO_INST RecordEvent TINSERT(DstTileData &dst, SrcTileData &src,
                             uint16_t indexRow, uint16_t indexCol,
                             WaitEvents &... events);

template <typename DstTileData, typename SrcTileData,
          ReluPreMode reluMode = ReluPreMode::NoRelu, typename... WaitEvents>
PTO_INST RecordEvent TINSERT(DstTileData &dst, SrcTileData &src,
                             uint64_t preQuantScalar,
                             uint16_t indexRow, uint16_t indexCol,
                             WaitEvents &... events);

template <typename DstTileData, typename SrcTileData, AccToVecMode mode,
          ReluPreMode reluMode = ReluPreMode::NoRelu, typename... WaitEvents>
PTO_INST RecordEvent TINSERT(DstTileData &dst, SrcTileData &src,
                             uint64_t preQuantScalar,
                             uint16_t indexRow, uint16_t indexCol,
                             WaitEvents &... events);

template <typename DstTileData, typename SrcTileData, typename FpTileData,
          ReluPreMode reluMode = ReluPreMode::NoRelu, typename... WaitEvents>
PTO_INST RecordEvent TINSERT_FP(DstTileData &dst, SrcTileData &src,
                                FpTileData &fp,
                                uint16_t indexRow, uint16_t indexCol,
                                WaitEvents &... events);

template <typename DstTileData, typename SrcTileData, typename FpTileData,
          AccToVecMode mode, ReluPreMode reluMode = ReluPreMode::NoRelu,
          typename... WaitEvents>
PTO_INST RecordEvent TINSERT(DstTileData &dst, SrcTileData &src,
                             FpTileData &fp,
                             uint16_t indexRow, uint16_t indexCol,
                             WaitEvents &... events);

#if defined(PTO_NPU_ARCH_A5) || defined(PTO_NPU_ARCH_KIRIN9030) || defined(PTO_NPU_ARCH_KIRINX90)
template <TInsertMode mode, typename DstTileData, typename SrcTileData,
          typename... WaitEvents>
PTO_INST RecordEvent TINSERT(DstTileData &dst, SrcTileData &src,
                             uint16_t indexRow = 0, uint16_t indexCol = 0,
                             WaitEvents &... events);
#endif

Constraints

General Constraints/Checks

  • TINSERT has the following overload families:
    • Normal insertion: TINSERT(dst, src, indexRow, indexCol)
    • relu form: TINSERT<..., reluMode>(dst, src, indexRow, indexCol)
    • Accumulator-to-vector form: TINSERT<..., mode, reluMode>(dst, src, indexRow, indexCol)
    • Scalar quantization form: TINSERT<..., reluMode>(dst, src, preQuantScalar, indexRow, indexCol) and TINSERT<..., mode, reluMode>(dst, src, preQuantScalar, indexRow, indexCol)
    • Vector quantization form: TINSERT_FP<..., reluMode>(dst, src, fp, indexRow, indexCol) and TINSERT<..., FpTileData, mode, reluMode>(dst, src, fp, indexRow, indexCol)
    • NZ split form (Ascend 950PR/Ascend 950DT/Kirin9030/KirinX90 only): TINSERT<TInsertMode::SPLIT2>(dst, src, indexRow, indexCol) or TINSERT<TInsertMode::SPLIT4>(...)
  • reluMode takes the value of ReluPreMode::{NoRelu, NormalRelu}.
  • mode takes the value of AccToVecMode::{SingleModeVec0, SingleModeVec1, DualModeSplitM, DualModeSplitN}.
  • Runtime boundary: indexRow + src.ValidRow <= dst.Rows and indexCol + src.ValidCol <= dst.Cols.

Implementation Check for Atlas A2/A3 Training Products/Atlas A2/A3 Inference Products

  • Supported tile type pairs: only TileType::Acc → TileType::Mat.
  • The source layout must be (BFractal: ColMajor, SFractal: RowMajor).
  • The target layout must be (BFractal: ColMajor, SFractal: RowMajor), and SFractalSize == 512.
  • Dst.Cols * sizeof(DstDType) must be a non-zero multiple of 32 bytes.
  • Supported dtype pairs for normal/relu (non-quantized):
    • float Acc → half, bfloat16_t
  • Supported dtype pairs for scalar quantization:
    • float Acc → int8_t
    • int32_t Acc → int8_t, uint8_t, half, int16_t
  • Supported dtype pairs for vector quantization (TINSERT_FP):
    • float Acc → int8_t
    • int32_t Acc → int8_t, uint8_t, half, int16_t
  • Vector quantization requires providing the FpTileData scaling operand (TileType::Scaling).

Ascend 950PR/Ascend 950DT Implementation Check

In addition to the Acc → Mat path, Ascend 950PR/Ascend 950DT also supports the Acc → Vec, Vec → Vec, Vec → Mat, and NZ split paths.

  • Acc → Mat (TileType::Acc → TileType::Mat):

    • The source Acc type must be float or int32_t; the source layout must be (BFractal: ColMajor, SFractal: RowMajor).
    • The target layout must be (!isRowMajor, SFractal: RowMajor) (NZ format).
    • Non-quantized (normal/relu) target types:
      • float Acc → half, bfloat16_t, float
      • int32_t Acc → int32_t
    • Scalar quantization target types:
      • float Acc → int8_t, uint8_t, hifloat8_t, half, bfloat16_t, float8_e4m3_t
      • int32_t Acc → int8_t, uint8_t, half, bfloat16_t
    • Vector quantization (TINSERT_FP) target type: same as scalar quantization.

  • Acc → Vec (TileType::Acc → TileType::Vec):

    • The source Acc type must be float or int32_t; the source layout must be (BFractal: ColMajor, SFractal: RowMajor).
    • Non-quantized (normal/relu) target type:
      • float Acc → half, bfloat16_t, float
      • int32_t Acc → int32_t
    • Scalar quantization target type:
      • float Acc → int8_t, uint8_t, hifloat8_t, half, bfloat16_t, float8_e4m3_t
      • int32_t Acc → int8_t, uint8_t, half, bfloat16_t
    • Vector quantization (TINSERT_FP/TINSERT with FpTileData) target type: same as scalar quantization.
    • The target layout must be one of the following: NZ-to-NZ (!isRowMajor, SFractal: RowMajor), NZ-to-ND (isRowMajor, SFractal: NoneBox), or NZ-to-DN (!isRowMajor, SFractal: NoneBox).
    • AccToVecMode selects SingleModeVec0, SingleModeVec1, DualModeSplitM, or DualModeSplitN.
    • Dual-target modes (DualModeSplitM, DualModeSplitN) require QuantMode_t::NoQuant and do not support the NZ-to-DN path.
    • For 32-bit target types (float/int32_t), when using DualModeSplitN, the ValidCol before splitting must be an integer multiple of 32.
    • The target stride must be non-zero, and dstStride * sizeof(dstType) must be a multiple of 32 bytes.

  • Vec → Vec (TileType::Vec → TileType::Vec):

    • DstTileData::DType must equal SrcTileData::DType.
    • Supported element types: half, bfloat16_t, float, int32_t, int8_t, hifloat8_t, float8_e4m3_t, float8_e5m2_t, float8_e8m0_t, float4_e2m1x2_t, float4_e1m2x2_t.
    • The source and target layouts must match (both ND or both NZ).
    • ND path: The source valid region must be within the target boundary. The dispatch selects copy_ubuf_to_ubuf (aligned), vlds/vsts (stride-aligned, unaligned validCol), vlds/vstus (unaligned stride or indexCol), or scalar copy (1×1 element).
    • NZ path: The number of source columns must not exceed the number of target columns. Use ComputeNZBlockParams for fractal block copy_ubuf_to_ubuf.

  • Vec → Mat (TileType::Vec → TileType::Mat, UB → L1):

    • DstTileData::DType must equal SrcTileData::DType.
    • Supported element types: half, bfloat16_t, float, int32_t, int8_t, hifloat8_t, float8_e4m3_t, float8_e5m2_t, float8_e8m0_t, float4_e2m1x2_t, float4_e1m2x2_t.
    • ND path: The source must be isRowMajor; use copy_ubuf_to_cbuf. The number of bytes per row must be aligned with BLOCK_BYTE_SIZE (32 bytes).
    • NZ path: The source must be (!isRowMajor, SFractal: RowMajor); use ComputeNZBlockParams for fractal block copy_ubuf_to_cbuf.

  • NZ Split (TInsertMode::SPLIT2 / TInsertMode::SPLIT4, only Ascend 950PR/Ascend 950DT/Kirin9030/KirinX90):

    • The target must be TileType::Mat; the source must be TileType::Vec.
    • DstTileData::DType must equal SrcTileData::DType.
    • The source must be in NZ format: (!isRowMajor, SFractal: RowMajor).
    • Supported element types: half, bfloat16_t, float, int32_t, int8_t, hifloat8_t, float8_e4m3_t, float8_e5m2_t, float8_e8m0_t, float4_e2m1x2_t, float4_e1m2x2_t.
    • validRow is aligned to FRACTAL_NZ_ROW (16) for burst calculation.
    • Split the total burst of copy_ubuf_to_cbuf into 2 or 4 sub-transfers, each handling totalBurstNum / SplitCount column blocks.

Examples

Automatic

#include <pto/pto-inst.hpp>

using namespace pto;

// Vec -> Mat insertion (NZ layout)
void example_auto() {
  using SrcT = Tile<TileType::Vec, half, 16, 32, BLayout::ColMajor, 16, 32, SLayout::RowMajor>;
  using DstT = Tile<TileType::Mat, half, 16, 32, BLayout::ColMajor, -1, -1, SLayout::RowMajor>;
  SrcT src;
  DstT dst(16, 32);
  TINSERT(dst, src, /*indexRow=*/0, /*indexCol=*/0);
}

Manual

#include <pto/pto-inst.hpp>

using namespace pto;

// Vec -> Mat insertion (NZ layout, manual buffer allocation)
void example_manual() {
  using SrcT = Tile<TileType::Vec, half, 16, 32, BLayout::ColMajor, 16, 32, SLayout::RowMajor>;
  using DstT = Tile<TileType::Mat, half, 16, 32, BLayout::ColMajor, -1, -1, SLayout::RowMajor>;
  SrcT src;
  DstT dst(16, 32);
  TASSIGN(src, 0x0);
  TASSIGN(dst, 0x0);
  TINSERT(dst, src, /*indexRow=*/0, /*indexCol=*/0);
}

ASM Examples

Automatic Mode

# Automatic mode: the compiler/runtime handles resource placement and scheduling.
%dst = pto.tinsert %src, %idxrow, %idxcol : (!pto.tile<...>, dtype, dtype) -> !pto.tile<...>

Manual Mode

# Manual mode: explicitly bind resources first, then issue the instruction.
# pto.tassign %arg0, @tile(0x1000)
# pto.tassign %arg1, @tile(0x2000)
%dst = pto.tinsert %src, %idxrow, %idxcol : (!pto.tile<...>, dtype, dtype) -> !pto.tile<...>

PTO Assembly Form

%dst = tinsert %src[%r0, %r1] : !pto.tile<...> -> !pto.tile<...>
# AS Level 2 (DPS)
pto.tinsert ins(%src, %idxrow, %idxcol : !pto.tile_buf<...>, dtype, dtype) outs(%dst : !pto.tile_buf<...>)
我们使用cookie来确保您的高速浏览体验。继续浏览本站，即表示您同意我们使用cookie。查看详情