PairReduce

函数功能

PairReduceSum：相邻两个（奇偶）元素求和，例如（a1, a2, a3, a4, a5, a6...），相邻两个数据求和为（a1+a2, a3+a4, a5+a6, ......）。

函数原型

• mask参数使用逐bit模式，该模式的具体介绍请参考参数说明中的mask参数说明：

  template <typename T, bool isSetMask = true>

  __aicore__ inline void PairReduceSum(const LocalTensor<T>& dstLocal, const LocalTensor<T>& srcLocal,const int32_t repeat, const uint64_t mask[2], const int32_t dstRepStride, const int32_t srcBlkStride,const int32_t srcRepStride)

• mask参数使用连续模式，该模式的具体介绍请参考参数说明中的mask参数说明：

  template <typename T, bool isSetMask = true>

  __aicore__ inline void PairReduceSum(const LocalTensor<T>& dstLocal, const LocalTensor<T>& srcLocal,const int32_t repeat, const int32_t maskCount, const int32_t dstRepStride, const int32_t srcBlkStride,const int32_t srcRepStride)

参数说明

返回值

无

支持的型号

Atlas 训练系列产品

Atlas推理系列产品（Ascend 310P处理器）AI Core

Atlas A2训练系列产品/Atlas 800I A2推理产品

Atlas 200/500 A2推理产品

注意事项

• 操作数地址偏移对齐要求请参见通用约束。
• 如果两两相加的两个元素mask位未配置（即当前两个元素不参与运算），对于Atlas 200/500 A2推理产品，对应的目的操作数中的值会置为0，对于其他产品型号，对应的目的操作数中的值不会变化。比如float场景下对64个数使用当前指令，maskCount配置为62，表示最后两个元素不参与运算，对于Atlas 200/500 A2推理产品，目的操作数中最后一个值会返回0；对于其他产品型号，目的操作数中最后一个值不会变化。

调用示例

本样例中展示Compute流程中的部分代码。如果您需要运行样例代码，请将该代码段拷贝并替换下方主体代码的Compute函数指令执行部分即可。

• PairReduceSum-tensor高维切分计算样例-mask连续模式

  uint64_t mask = 256/sizeof(half);
  int repeat = 1;
  // repeat = 1, 128 elements one repeat, 128 elements total
  // srcBlkStride = 1, no gap between blocks in one repeat
  // dstRepStride = 1, srcRepStride = 8, no gap between repeats
  PairReduceSum<half>(dstLocal, srcLocal, repeat, mask, 1, 1, 8);

• PairReduceSum-tensor高维切分计算样例-mask逐bit模式

  uint64_t mask[2] = { UINT64_MAX, UINT64_MAX };
  int repeat = 1;
  // repeat = 1, 128 elements one repeat, 128 elements total
  // srcBlkStride = 1, no gap between blocks in one repeat
  // dstRepStride = 1, srcRepStride = 8, no gap between repeats
  PairReduceSum<half>(dstLocal, srcLocal, repeat, mask, 1, 1, 8);

• 示例结果

  输入数据(src_gm): 
  [-3.441, 7.246, -0.02759, -6.324, 3.693, -7.984, -4.246, 6.332, -3.734, -2.699, -6.91, 7.887, -3.631, 5.219, 6.539, 8.688, 6.523, -6.789, -8.547, 4.258, 1.344, -8.469, -0.9253, -3.914, 3.293, -9.828, 7.082, 5.961, 2.133, 1.959, 3.928, -1.062, 9.18, -1.725, -3.645, 1.457, -2.328, -0.9487, -0.2849, -2.998, -9.281, 3.137, 0.4028, 5.961, -6.25, 2.406, -6.203, -2.699, 4.914, 1.653, -6.383, 6.855, 9.164, 0.6646, -2.854, 3.18, -0.5884, 0.4258, -5.773, -2.152, 4.258, 4.129, -8.719, -8.828, 6.145, 7.387, 1.386, -4.684, 6.324, -1.275, -1.816, 3.357, 6.832, -1.059, -9.852, -8.539, 2.938, -2.002, 9.625, -4.387, -1.309, 8.289, 2.906, -1.035, 7.723, 4.727, -6.477, 2.389, 6.75, -6.688, -0.04248, -6.613, -3.424, 7.145, 4.836, -5.617, -5.855, -5.234, -9.422, -9.852, -8.531, 2.115, 5.109, -8.094, -6.238, 9.898, -6.848, -6.051, 7.109, 4.227, -0.6187, -3.492, -4.352, 1.344, 1.526, 2.572, 2.16, -1.135, 9.812, 1.426, -8, 3.291, -2.039, 5.93, -5.52, -5.156, -9.422, 0.2236]  
  输出数据(dst_gm): 
  [3.805, -6.352, -4.289, 2.086, -6.434, 0.9766, 1.588, 15.23, -0.2656, -4.289, -7.125, -4.84, -6.535, 13.05, 4.094, 2.865, 7.453, -2.188, -3.277, -3.283, -6.145, 6.363, -3.844, -8.906, 6.566, 0.4727, 9.828, 0.3262, -0.1626, -7.926, 8.391, -17.55, 13.53, -3.297, 5.047, 1.541, 5.773, -18.39, 0.9355, 5.238, 6.98, 1.871, 12.45, -4.086, 0.0625, -6.656, 3.721, -0.7812, -11.09, -19.28, -6.414, -2.984, 3.66, -12.9, 11.34, -4.109, -3.008, 4.098, 1.025, 11.23, -4.711, 3.891, -10.67, -9.195]

• 主体代码

#include "kernel_operator.h"

namespace AscendC {
class KernelReduce {
public:
    __aicore__ inline KernelReduce() {}
    __aicore__ inline void Init(__gm__ uint8_t* src, __gm__ uint8_t* dstGm)
    {
        srcGlobal.SetGlobalBuffer((__gm__ half*)src);
        dstGlobal.SetGlobalBuffer((__gm__ half*)dstGm);

        pipe.InitBuffer(inQueueSrc, 1, srcDataSize * sizeof(half));
        pipe.InitBuffer(outQueueDst, 1, dstDataSize * sizeof(half));
    }
    __aicore__ inline void Process()
    {
        CopyIn();
        Compute();
        CopyOut();
    }

private:
    __aicore__ inline void CopyIn()
    {
        LocalTensor<half> srcLocal = inQueueSrc.AllocTensor<half>();
        DataCopy(srcLocal, srcGlobal, srcDataSize);
        inQueueSrc.EnQue(srcLocal);
    }
    __aicore__ inline void Compute()
    {
        LocalTensor<half> srcLocal = inQueueSrc.DeQue<half>();
        LocalTensor<half> dstLocal = outQueueDst.AllocTensor<half>();

        half zero(0);
        Duplicate(dstLocal, zero, dstDataSize);

        //指令执行部分（替换成上述代码）

        outQueueDst.EnQue<half>(dstLocal);
        inQueueSrc.FreeTensor(srcLocal);
    }
    __aicore__ inline void CopyOut()
    {
        LocalTensor<half> dstLocal = outQueueDst.DeQue<half>();
        DataCopy(dstGlobal, dstLocal, dstDataSize);
        outQueueDst.FreeTensor(dstLocal);
    }

private:
    TPipe pipe;
    TQue<QuePosition::VECIN, 1> inQueueSrc;
    TQue<QuePosition::VECOUT, 1> outQueueDst;
    GlobalTensor<half> srcGlobal, dstGlobal;
    int srcDataSize = 128;
    int dstDataSize = 64;
};
} // namespace AscendC

extern "C" __global__ __aicore__ void reduce_simple_kernel(__gm__ uint8_t* src, __gm__ uint8_t* dstGm)
{
    AscendC::KernelReduce op;
    op.Init(src, dstGm);
    op.Process();
}