根据 Region Proposals 中的 score 域对其进行排序（score大的排前面），每次排 16 个Region Proposals。
template <typename T> __aicore__ inline void RpSort16(const LocalTensor<T>& dstLocal, const LocalTensor<T>& srcLocal, const int32_t repeatTimes)
参数名称 |
输入/输出 |
含义 |
|---|---|---|
dstLocal |
输出 |
目的操作数，存储经过排序后的 Region Proposals，类型为LocalTensor。LocalTensor数据结构的定义请参考LocalTensor。 Atlas 训练系列产品，支持的数据类型为：half Atlas推理系列产品AI Core，支持的数据类型为：half/float |
srcLocal |
输入 |
源操作数，存储未经过排序的 Region Proposals，类型为LocalTensor。LocalTensor数据结构的定义请参考LocalTensor。 Atlas 训练系列产品，支持的数据类型为：half Atlas推理系列产品AI Core，支持的数据类型为：half/float |
repeatTimes |
输入 |
重复迭代次数，int32_t类型，每次排 16 个proposals。取值范围：repeatTimes∈[0,255]。 |
Atlas 训练系列产品
Atlas推理系列产品AI Core
// repeatTimes = 2, 对2个Region Proposal进行排序 RpSort16(dstLocal, dstLocal, 2);
#include "kernel_operator.h"
namespace AscendC {
class KernelVecProposal {
public:
__aicore__ inline KernelVecProposal() {}
__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();
PreProcess();
Compute();
CopyOut();
}
private:
__aicore__ inline void CopyIn()
{
LocalTensor<half> srcLocal = inQueueSrc.AllocTensor<half>();
DataCopy(srcLocal, srcGlobal, srcDataSize);
inQueueSrc.EnQue(srcLocal);
}
__aicore__ inline void PreProcess()
{
LocalTensor<half> srcLocal = inQueueSrc.DeQue<half>();
LocalTensor<half> dstLocal = outQueueDst.AllocTensor<half>();
Concat(dstLocal, srcLocal, repeat, mode); // sort排序是基于score的，此处先创建一个有score数据的proposal，需要注意的是，非score处的数据可能是随机值
outQueueDst.EnQue<half>(dstLocal);
inQueueSrc.FreeTensor(srcLocal);
}
__aicore__ inline void Compute()
{
LocalTensor<half> dstLocal = outQueueDst.DeQue<half>();
RpSort16(dstLocal, dstLocal, repeat);
outQueueDst.EnQue<half>(dstLocal);
}
__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 = 32;
int dstDataSize = 256;
int repeat = srcDataSize / 16;
int mode = 4;
};
} // namespace AscendC
extern "C" __global__ __aicore__ void vec_proposal_kernel(__gm__ uint8_t* src, __gm__ uint8_t* dstGm)
{
AscendC::KernelVecProposal op;
op.Init(src, dstGm);
op.Process();
}
示例结果
输入数据(src_gm):
[ -1.624 -42.3 -54.12 91.25 -99.4 36.72 67.44 -66.3 -52.53
3.377 -62.47 -15.85 -31.47 3.143 58.47 -83.75 21.58 63.47
7.234 35.16 -39.72 37.8 73.06 -98.7 44.1 -77.2 67.2
19.62 -87.9 -14.875 15.86 -77.75]
输出数据(dst_gm):
[ 0. 0. 0. 0. 91.25 0. 0. 0. 0.
0. 0. 0. 67.44 0. 0. 0. 0. 0.
0. 0. 58.47 0. 0. 0. 0. 0. 0.
0. 36.72 0. 0. 0. 0. 0. 0. 0.
3.377 0. 0. 0. 0. 0. 0. 0. 3.143
0. 0. 0. 0. 0. 0. 0. -1.624 0.
0. 0. 0. 0. 0. 0. -15.85 0. 0.
0. 0. 0. 0. 0. -31.47 0. 0. 0.
0. 0. 0. 0. -42.3 0. 0. 0. 0.
0. 0. 0. -52.53 0. 0. 0. 0. 0.
0. 0. -54.12 0. 0. 0. 0. 0. 0.
0. -62.47 0. 0. 0. 0. 0. 0. 0.
-66.3 0. 0. 0. 0. 0. 0. 0. -83.75
0. 0. 0. 0. 0. 0. 0. -99.4 0.
0. 0. 0. 0. 0. 0. 73.06 0. 0.
0. 0. 0. 0. 0. 67.2 0. 0. 0.
0. 0. 0. 0. 63.47 0. 0. 0. 0.
0. 0. 0. 44.1 0. 0. 0. 0. 0.
0. 0. 37.8 0. 0. 0. 0. 0. 0.
0. 35.16 0. 0. 0. 0. 0. 0. 0.
21.58 0. 0. 0. 0. 0. 0. 0. 19.62
0. 0. 0. 0. 0. 0. 0. 15.86 0.
0. 0. 0. 0. 0. 0. 7.234 0. 0.
0. 0. 0. 0. 0. -14.875 0. 0. 0.
0. 0. 0. 0. -39.72 0. 0. 0. 0.
0. 0. 0. -77.2 0. 0. 0. 0. 0.
0. 0. -77.75 0. 0. 0. 0. 0. 0.
0. -87.9 0. 0. 0. 0. 0. 0. 0.
-98.7 0. 0. 0. ]