Xor

功能说明

按元素执行Xor运算，Xor（异或）的概念和运算规则如下：

• 概念：参加运算的两个数据，按⼆进制位进⾏“异或”运算。
• 运算规则：0^0=0；0^1=1；1^0=1；1^1=0；即：参加运算的两个对象，如果两个相应位为“异”（值不同），则该位结果为1，否则为 0【同0异1】。

例如：3^5=6，即0000 0011^0000 0101 = 0000 0110

函数原型

• 通过sharedTmpBuffer入参传入临时空间
  • 源操作数Tensor全部/部分参与计算

    1 2	template <typename T, bool isReuseSource = false> __aicore__ inline void Xor(const LocalTensor<T>& dstTensor, const LocalTensor<T>& src0Tensor, const LocalTensor<T>& src1Tensor, const LocalTensor<uint8_t>& sharedTmpBuffer, const uint32_t calCount)

  • 源操作数Tensor全部参与计算

    1 2	template <typename T, bool isReuseSource = false> __aicore__ inline void Xor(const LocalTensor<T>& dstTensor, const LocalTensor<T>& src0Tensor, const LocalTensor<T>& src1Tensor, const LocalTensor<uint8_t>& sharedTmpBuffer)

• 接口框架申请临时空间
  • 源操作数Tensor全部/部分参与计算

    1 2	template <typename T, bool isReuseSource = false> __aicore__ inline void Xor(const LocalTensor<T> &dstTensor, const LocalTensor<T> &src0Tensor, const LocalTensor<T> &src1Tensor, const uint32_t calCount)

  • 源操作数Tensor全部参与计算

    1 2	template <typename T, bool isReuseSource = false> __aicore__ inline void Xor(const LocalTensor<T> &dstTensor, const LocalTensor<T> &src0Tensor, const LocalTensor<T> &src1Tensor)

由于该接口的内部实现中涉及复杂的数学计算，需要额外的临时空间来存储计算过程中的中间变量。临时空间支持开发者通过sharedTmpBuffer入参传入和接口框架申请两种方式。

• 通过sharedTmpBuffer入参传入，使用该tensor作为临时空间进行处理，接口框架不再申请。该方式开发者可以自行管理sharedTmpBuffer内存空间，并在接口调用完成后，复用该部分内存，内存不会反复申请释放，灵活性较高，内存利用率也较高。
• 接口框架申请临时空间，开发者无需申请，但是需要预留临时空间的大小。

通过sharedTmpBuffer传入的情况，开发者需要为tensor申请空间；接口框架申请的方式，开发者需要预留临时空间。临时空间大小BufferSize的获取方式如下：通过GetXorMaxMinTmpSize中提供的接口获取需要预留空间范围的大小。

参数说明

返回值

无

支持的型号

Atlas A2训练系列产品/Atlas 800I A2推理产品

Atlas推理系列产品（Ascend 310P处理器）AI Core

Atlas 200/500 A2推理产品

约束说明

• 不支持源操作数与目的操作数地址重叠。
• 当前仅支持ND格式的输入，不支持其他格式。
• calCount需要保证小于或等于src0Tensor和src1Tensor和dstTensor存储的元素范围。
• 对于不带calCount参数的接口，需要保证src0Tensor和src1Tensor的shape大小相等。
• 不支持sharedTmpBuffer与源操作数和目的操作数地址重叠。
• 操作数地址偏移对齐要求请参见通用约束。

调用示例

调用样例 kernel侧xor_custom.cpp

#include "kernel_operator.h"
using namespace AscendC;
constexpr int32_t BUFFER_NUM = 1;
class KernelXor {
public:
    __aicore__ inline KernelXor() {}
    __aicore__ inline void Init(GM_ADDR x, GM_ADDR y, GM_ADDR z, uint32_t totalLength, uint32_t totalLength2, uint32_t tilenum, uint32_t tmpSize, uint32_t isProf,uint32_t mcount)
    {
        AscendCUtils::SetOverflow(1);
        ASSERT(GetBlockNum() != 0 && "block dim can not be zero!");
        this->totalLength = totalLength;
        this->blockLength = totalLength / GetBlockNum();
        this->blockLength2 = totalLength2 / GetBlockNum();
        this->tilenum = tilenum;
        this->isProf = isProf;
        this->tmpSize = tmpSize;
        this->mcount = mcount;
        ASSERT(tilenum != 0 && "tile num can not be zero!");
        this->tileLength = this->blockLength / tilenum / BUFFER_NUM;
        this->tileLength2 = this->blockLength2 / tilenum / BUFFER_NUM;
        
        xGm.SetGlobalBuffer((__gm__ int16_t*)x + this->blockLength * GetBlockIdx(), this->blockLength);
        yGm.SetGlobalBuffer((__gm__ int16_t*)y + this->blockLength2 * GetBlockIdx(), this->blockLength2);
        zGm.SetGlobalBuffer((__gm__ int16_t*)z + this->blockLength * GetBlockIdx(), this->blockLength);
        
        if (this->tmpSize != 0){
            pipe.InitBuffer(tmpQueue, BUFFER_NUM, this->tmpSize);
        }
        pipe.InitBuffer(inQueueX, BUFFER_NUM, this->tileLength * sizeof(int16_t));
        pipe.InitBuffer(inQueueY, BUFFER_NUM, this->tileLength2 * sizeof(int16_t));
        pipe.InitBuffer(outQueueZ, BUFFER_NUM, this->tileLength * sizeof(int16_t));
    }
    __aicore__ inline void Process()
    {
        int32_t loopCount = this->tilenum * BUFFER_NUM;
        for (int32_t i = 0; i < loopCount; i++) {
            CopyIn(i);
            Compute(i); 
            CopyOut(i);
        }
    }
private:
    __aicore__ inline void CopyIn(int32_t progress)
    {
        LocalTensor<int16_t> xLocal = inQueueX.AllocTensor<int16_t>();
        DataCopy(xLocal, xGm[progress * this->tileLength], this->tileLength);
        inQueueX.EnQue(xLocal);
        LocalTensor<int16_t> yLocal = inQueueY.AllocTensor<int16_t>();
        DataCopy(yLocal, yGm[progress * this->tileLength2], this->tileLength2);
        inQueueY.EnQue(yLocal);
    }
    __aicore__ inline void Compute(int32_t progress)
    {
        LocalTensor<int16_t> xLocal = inQueueX.DeQue<int16_t>();
        LocalTensor<int16_t> yLocal = inQueueY.DeQue<int16_t>();
        LocalTensor<int16_t> zLocal = outQueueZ.AllocTensor<int16_t>();
        if (this->tmpSize != 0) {  // 使用temTensor
            LocalTensor<uint8_t> tmpLocal = tmpQueue.AllocTensor<uint8_t>();
            if (this->mcount != this->totalLength) {
               Xor(zLocal, xLocal, yLocal, tmpLocal, this->mcount);
            } else {  // 全部计算不用传入长度
               Xor(zLocal, xLocal, yLocal, tmpLocal);
            }
            tmpQueue.FreeTensor(tmpLocal);
        } else {  // 不使用temTensor
            if (this->mcount != this->totalLength) {
               Xor(zLocal, xLocal, yLocal, this->mcount);
            } else {
               Xor(zLocal, xLocal, yLocal);
            }
        }
        outQueueZ.EnQue<int16_t>(zLocal);
        inQueueX.FreeTensor(xLocal);
        inQueueY.FreeTensor(yLocal);
    }
    __aicore__ inline void CopyOut(int32_t progress)
    {
        LocalTensor<int16_t> zLocal = outQueueZ.DeQue<int16_t>();
        DataCopy(zGm[progress * this->tileLength], zLocal, this->tileLength);
        outQueueZ.FreeTensor(zLocal);
    }
private:
    TPipe pipe;
    TQue<QuePosition::VECIN, BUFFER_NUM> inQueueX;
    TQue<QuePosition::VECIN, BUFFER_NUM> inQueueY;
    TQue<QuePosition::VECIN, BUFFER_NUM> tmpQueue;
    TQue<QuePosition::VECOUT, BUFFER_NUM> outQueueZ;
    GlobalTensor<int16_t> xGm;
    GlobalTensor<int16_t> yGm;
    GlobalTensor<int16_t> zGm;
    uint32_t blockLength;
    uint32_t blockLength2;
    uint32_t tilenum;
    uint32_t isProf;
    uint32_t tileLength;
    uint32_t tileLength2;   
    uint32_t tmpSize; 
    uint32_t mcount;
    uint32_t totalLength;
};
extern "C" __global__ __aicore__ void xor_custom(GM_ADDR x, GM_ADDR y, GM_ADDR z, GM_ADDR workspace, GM_ADDR tiling) {
    GET_TILING_DATA(tilingData, tiling);
    // TODO: user kernel impl
    KernelXor op;
    op.Init(x, y, z, tilingData.totalLength, tilingData.totalLength2, tilingData.tilenum, tilingData.tmpSize, tilingData.isProf,tilingData.mcount);
    if (TILING_KEY_IS(1)) {
        op.Process();
    }
}

host侧xor_custom_tiling.h

#include "register/tilingdata_base.h"
namespace optiling {
BEGIN_TILING_DATA_DEF(XorCustomTilingData)
  TILING_DATA_FIELD_DEF(uint32_t, totalLength);
  TILING_DATA_FIELD_DEF(uint32_t, totalLength2);
  TILING_DATA_FIELD_DEF(uint32_t, tmpSize);
  TILING_DATA_FIELD_DEF(uint32_t, tilenum);
  TILING_DATA_FIELD_DEF(uint32_t, isProf);
  TILING_DATA_FIELD_DEF(uint32_t, mcount);
END_TILING_DATA_DEF;
REGISTER_TILING_DATA_CLASS(XorCustom, XorCustomTilingData)
}

host侧xor_custom.cpp

#include "xor_custom_tiling.h"
#include "register/op_def_registry.h"
#include "tiling/tiling_api.h"
namespace optiling {
static ge::graphStatus TilingFunc(gert::TilingContext* context)
{
  XorCustomTilingData tiling;
  const gert::RuntimeAttrs* cosattrs = context->GetAttrs();
  const uint32_t tilenum = *(cosattrs->GetAttrPointer<uint32_t>(0));
  const uint32_t blockdim = *(cosattrs->GetAttrPointer<uint32_t>(1));
  const uint32_t sizeflag = *(cosattrs->GetAttrPointer<uint32_t>(2));
  const uint32_t isProf = *(cosattrs->GetAttrPointer<uint32_t>(3));
  const uint32_t countflag = *(cosattrs->GetAttrPointer<uint32_t>(4));
  uint32_t totalLength = context->GetInputTensor(0)->GetShapeSize();
  uint32_t totalLength2 = context->GetInputTensor(1)->GetShapeSize();
  context->SetBlockDim(blockdim);
  tiling.set_totalLength(totalLength);
  tiling.set_totalLength2(totalLength2);
  tiling.set_tilenum(tilenum);
  tiling.set_isProf(isProf);
  
  if (countflag == 0) {
      tiling.set_mcount(totalLength2);
  } else if (countflag == 1) {
      tiling.set_mcount(totalLength);
  }
  
  std::vector<int64_t> shape_vec = {totalLength};
  ge::Shape srcShape(shape_vec);
  uint32_t typeSize = sizeof(int16_t);
  uint32_t maxValue = 0;
  uint32_t minValue = 0;
  bool isReuseSource = false;
  AscendC::GetXorMaxMinTmpSize(srcShape, typeSize, isReuseSource, maxValue, minValue);
  // sizeflag 0：代表取最小的tempBuffer 1：取最大的tempBuffer
  if (sizeflag == 0) {
      tiling.set_tmpSize(minValue);
  } else if (sizeflag == 1) {
      tiling.set_tmpSize(maxValue);
  } else if (sizeflag == 2) {
      tiling.set_tmpSize(0);
  }
  tiling.SaveToBuffer(context->GetRawTilingData()->GetData(), context->GetRawTilingData()->GetCapacity());
  context->GetRawTilingData()->SetDataSize(tiling.GetDataSize());
  context->SetTilingKey(1);
  size_t *currentWorkspace = context->GetWorkspaceSizes(1);
  currentWorkspace[0] = 0;
  return ge::GRAPH_SUCCESS;
}
}
namespace ge {
static ge::graphStatus InferShape(gert::InferShapeContext* context)
{
    const gert::Shape* x1_shape = context->GetInputShape(0);
    gert::Shape* y_shape = context->GetOutputShape(0);
    *y_shape = *x1_shape;
    return GRAPH_SUCCESS;
}
}
namespace ops {
class XorCustom : public OpDef {
public:
    explicit XorCustom(const char* name) : OpDef(name)
    {
        this->Input("x")
            .ParamType(REQUIRED)
            .DataType({ge::DT_INT16})
            .Format({ge::FORMAT_ND});
        this->Input("y")
            .ParamType(REQUIRED)
            .DataType({ge::DT_INT16})
            .Format({ge::FORMAT_ND});
        this->Output("z")
            .ParamType(REQUIRED)
            .DataType({ge::DT_INT16})
            .Format({ge::FORMAT_ND});
        this->SetInferShape(ge::InferShape);
        this->Attr("tilenum")
            .AttrType(REQUIRED)            
            .Int(0);
        this->Attr("blockdim")
            .AttrType(REQUIRED)            
            .Int(0);
        this->Attr("sizeflag")
            .AttrType(REQUIRED)            
            .Int(0);
        this->Attr("isProf")
            .AttrType(REQUIRED)            
            .Int(0);
        this->Attr("countflag")
            .AttrType(REQUIRED)
            .Int(0);
        this->AICore()
            .SetTiling(optiling::TilingFunc);
        this->AICore().AddConfig("ascendxxx");
    }
};

输入输出的数据类型为int16_t，一维向量包含32个数。例如向量中第一个数据进行异或：(-5753) xor 18745 = -24386
输入数据(src0Local): [-5753 28501 20334 -5845  ... -20817 3403 21261 22241]
输入数据(src1Local): [18745 -24448 20873 10759 ... 21940 -26342 9251 31019]
输出数据(dstLocal): [-24386 -12331 7911 -15572 ... -1253 -27567 30510 12234]