更多样例

样例一

下面的样例展示了数学库kernel侧API和Tiling API GetXxxMaxMinTmpSize的配套使用方法，具体流程如下：

Host侧调用Tiling接口，获取所需临时空间的大小，并将其写入tiling data中；kernel侧再读取tiling data，获取相应的临时空间大小，并根据此分配临时空间。

Host侧Tiling API 使用样例:

#include <vector>

#include "register/op_def_registry.h"
#include "register/tilingdata_base.h"
#include "tiling/tiling_api.h"

namespace optiling {

BEGIN_TILING_DATA_DEF(AsinCustomTilingData)
  TILING_DATA_FIELD_DEF(uint32_t, srcSize);
  TILING_DATA_FIELD_DEF(uint32_t, tmpBufferSize);
END_TILING_DATA_DEF;

static ge::graphStatus TilingFunc(gert::TilingContext* context)
{
    // Input source shapes.
    std::vector<int64_t> srcDims = {16, 128};
    uint32_t srcSize = 1;
    for (auto dim : srcDims) {
        srcSize *= dim;
    }
    uint32_t typeSize = 2;
    ge::Shape shape(srcDims);
    uint32_t minValue = 0;
    uint32_t maxValue = 0;
    AscendC::GetAsinMaxMinTmpSize(shape, typeSize, false, maxValue, minValue);

    auto platformInfo = context->GetPlatformInfo();
    auto ascendcPlatform = platform_ascendc::PlatformAscendC(platformInfo);
    uint64_t tailSize = 0; // ub剩余空间大小
    ascendcPlatform.GetCoreMemSize(platform_ascendc::CoreMemType::UB, tailSize); // 本样例中使用完整的ub空间，实际情况下tailSize需要减掉用户已使用的ub空间
    auto tmpSize = tailSize >= maxValue ? maxValue : tailSize;

    AsinCustomTilingData tiling;
    tiling.set_srcSize(srcSize);
    tiling.set_tmpBufferSize(tmpSize);
    context->SetBlockDim(1);
    tiling.SaveToBuffer(context->GetRawTilingData()->GetData(), context->GetRawTilingData()->GetCapacity());
    context->GetRawTilingData()->SetDataSize(tiling.GetDataSize());
    context->SetTilingKey(1);

    return ge::GRAPH_SUCCESS;
}
} // namespace optiling

kernel侧读取tiling data，获取相应的临时空间大小，并根据此分配临时空间:

#include "kernel_operator.h"
template <typename srcType>
class KernelAsin {
public:
    __aicore__ inline KernelAsin()
    {}
    __aicore__ inline void Init(GM_ADDR src_gm, GM_ADDR dst_gm, uint32_t srcSize, uint32_t tmpBufferSize)
    {
        src_global.SetGlobalBuffer(reinterpret_cast<__gm__ srcType *>(src_gm), srcSize);
        dst_global.SetGlobalBuffer(reinterpret_cast<__gm__ srcType *>(dst_gm), srcSize);

        pipe.InitBuffer(inQueue, 1, srcSize * sizeof(srcType));
        pipe.InitBuffer(outQueue, 1, srcSize * sizeof(srcType));
        pipe.InitBuffer(tmpQueue, 1, tmpBufferSize);
        bufferSize = srcSize;
    }
    __aicore__ inline void Process()
    {
        CopyIn();
        Compute();
        CopyOut();
    }

private:
    __aicore__ inline void CopyIn()
    {
        LocalTensor<srcType> srcLocal = inQueue.AllocTensor<srcType>();
        DataCopy(srcLocal, src_global, bufferSize);
        inQueue.EnQue(srcLocal);
    }
    __aicore__ inline void Compute()
    {
        LocalTensor<srcType> dstLocal = outQueue.AllocTensor<srcType>();

        LocalTensor<srcType> srcLocal = inQueue.DeQue<srcType>();
        LocalTensor<uint8_t> sharedTmpBuffer = tmpQueue.AllocTensor<uint8_t>();
        Asin<srcType, false>(dstLocal, srcLocal, sharedTmpBuffer, bufferSize);

        outQueue.EnQue<srcType>(dstLocal);
        inQueue.FreeTensor(srcLocal);
        tmpQueue.FreeTensor(sharedTmpBuffer);
    }
    __aicore__ inline void CopyOut()
    {
        LocalTensor<srcType> dstLocal = outQueue.DeQue<srcType>();
        DataCopy(dst_global, dstLocal, bufferSize);
        outQueue.FreeTensor(dstLocal);
    }

private:
    GlobalTensor<srcType> src_global;
    GlobalTensor<srcType> dst_global;

    TPipe pipe;
    TQue<QuePosition::VECIN, 1> inQueue;
    TQue<QuePosition::VECCALC, 1> tmpQueue;
    TQue<QuePosition::VECOUT, 1> outQueue;
    uint32_t bufferSize = 0;
};

extern "C" __global__ __aicore__ void kernel_asin_operator(GM_ADDR src_gm, GM_ADDR dst_gm, GM_ADDR tiling) 
{                                                                                                         
    GET_TILING_DATA(tilingData, tiling);
    KernelAsin<half> op;
    op.Init(src_gm, dst_gm, tilingData.srcSize, tilingData.tmpBufferSize);
    if (TILING_KEY_IS(1)) {
        op.Process();
    }
}

样例二

下面的样例展示了数学库kernel侧API和Tiling API GetXxxTmpBufferFactorSize的配套使用方法，具体流程如下：

Host侧调用Tiling接口，获取maxLiveNodeCnt和extraBuf，并推算算子单次最大计算元素数量，将其写入tiling data中；kernel侧再读取tiling data，获取该值，基于该值分配临时空间。

Host侧Tiling API 使用样例:

#include <vector>
#include <cassert>
#include "register/op_def_registry.h"
#include "register/tilingdata_base.h"
#include "tiling/tiling_api.h"

namespace optiling {
BEGIN_TILING_DATA_DEF(AsinCustomTilingData)
TILING_DATA_FIELD_DEF(uint32_t, srcSize);
TILING_DATA_FIELD_DEF(uint32_t, tmpBufferSize);
END_TILING_DATA_DEF;

static ge::graphStatus TilingFunc(gert::TilingContext *context)
{
    // Input source shapes.
    std::vector<int64_t> srcDims = { 16, 128 };
    uint32_t srcSize = 1;
    uint32_t srcCurSize = 1;
    for (auto dim : srcDims) {
        srcSize *= dim;
    }
    uint32_t typeSize = 2;

    auto platformInfo = context->GetPlatformInfo();
    auto ascendcPlatform = platform_ascendc::PlatformAscendC(platformInfo);
    uint64_t tailSize = 0; // ub剩余空间大小
    ascendcPlatform.GetCoreMemSize(platform_ascendc::CoreMemType::UB, tailSize);

    uint32_t asinMaxLiveNodeCnt = 0;
    uint32_t asinExtraBuf = 0;

    uint32_t acosMaxLiveNodeCnt = 0;
    uint32_t acosExtraBuf = 0;

    AscendC::GetAsinTmpBufferFactorSize(typeSize, asinMaxLiveNodeCnt, asinExtraBuf);
    AscendC::GetAcosTmpBufferFactorSize(typeSize, acosMaxLiveNodeCnt, acosExtraBuf);
    // tmp的大小需要减去UB上调用api接口输入和输出占用的大小
    // 该示例中包括Asin接口的输入输出，以及Acos的输入输出，其中Asin接口的输出作为Acos的输入，因此一共需要3份src的空间大小
    auto tmpSize = tailSize - srcSize * typeSize * 3;
    assert(tmpSize >= asinExtraBuf);
    assert(tmpSize >= acosExtraBuf);
    // 计算Asin算子单次最大计算元素数量
    if (asinMaliveNodeCnt != 0) {
        srcAsinCurSize = (tmpSize - asinExtraBuf) / asinMaxLiveNodeCnt / typeSize;
    } else {
        srcAsinCurSize = srcSize;
    }
    // 计算Acos算子单次最大计算元素数量
    if (acosMaxLiveNodeCnt != 0) {
        srcAcosCurSize = (tmpSize - acosExtraBuf) / acosMaxLiveNodeCnt / typeSize; 
    } else {
        srcAcosCurSize = srcSize;
    }
    srcCurSize = min(srcAsinCurSize, srcAcosCurSize);

    AsinCustomTilingData tiling;
    tiling.set_srcSize(srcSize);
    tiling.set_srcCurSize(srcCurSize);
    tiling.set_tmpBufferSize(tmpSize);
    context->SetBlockDim(1);
    tiling.SaveToBuffer(context->GetRawTilingData()->GetData(), context->GetRawTilingData()->GetCapacity());
    context->GetRawTilingData()->SetDataSize(tiling.GetDataSize());
    context->SetTilingKey(1);

    return ge::GRAPH_SUCCESS;
}
} // namespace optiling

kernel侧样例:

#include "kernel_operator.h"
template <typename srcType>
class KernelAsin {
public:
    __aicore__ inline KernelAsin()
    {}
    __aicore__ inline void Init(GM_ADDR src_gm, GM_ADDR dst_gm, uint32_t srcSizeIn, uint32_t srcCurSizeIn, uint32_t tmpBufferSize)
    {
        srcSize = srcSizeIn;
        srcCurSize = srcCurSizeIn;
        src_global.SetGlobalBuffer(reinterpret_cast<__gm__ srcType *>(src_gm), srcSize);
        dst_global.SetGlobalBuffer(reinterpret_cast<__gm__ srcType *>(dst_gm), srcSize);

        pipe.InitBuffer(inQueue, 1, srcSize * sizeof(srcType));
        pipe.InitBuffer(outQueue, 1, srcSize * sizeof(srcType));
        pipe.InitBuffer(tmpQueue1, 1, srcCurSize * sizeof(srcType));
        pipe.InitBuffer(tmpQueue, 1, tmpBufferSize);
    }
    __aicore__ inline void Process()
    {
        CopyIn();
        Compute();
        CopyOut();
    }

private:
    __aicore__ inline void CopyIn()
    {
        LocalTensor<srcType> srcLocal = inQueue.AllocTensor<srcType>();
        DataCopy(srcLocal, src_global, srcSize);
        inQueue.EnQue(srcLocal);
    }
    __aicore__ inline void Compute()
    {
        LocalTensor<srcType> dstLocal = outQueue.AllocTensor<srcType>();
        LocalTensor<srcType> srcLocal = inQueue.DeQue<srcType>();
        LocalTensor<uint8_t> sharedTmpBuffer = tmpQueue.AllocTensor<uint8_t>();
        LocalTensor<srcType> tmpresBuffer = tmpQueue1.AllocTensor<srcType>();

        for (int32_t offset = 0; offset < srcSize; offset += srcCurSize) {
            Asin<srcType, false>(tmpresBuffer, srcLocal[offset], sharedTmpBuffer, srcCurSize);
            PipeBarrier<PIPE_V>();
            Acos<srcType, false>(dstLocal[offset], tmpresBuffer, sharedTmpBuffer, srcCurSize);
            PipeBarrier<PIPE_V>();
        }
        outQueue.EnQue<srcType>(dstLocal);
        inQueue.FreeTensor(srcLocal);
        tmpQueue.FreeTensor(sharedTmpBuffer);
    }
    __aicore__ inline void CopyOut()
    {
        LocalTensor<srcType> dstLocal = outQueue.DeQue<srcType>();
        DataCopy(dst_global, dstLocal, srcSize);
        outQueue.FreeTensor(dstLocal);
    }

private:
    GlobalTensor<srcType> src_global;
    GlobalTensor<srcType> dst_global;

    TPipe pipe;
    TQue<QuePosition::VECIN, 1> inQueue;
    TQue<QuePosition::VECCALC, 1> tmpQueue;
    TQue<QuePosition::VECCALC, 1> tmpQueue1;
    TQue<QuePosition::VECOUT, 1> outQueue;
    uint32_t srcSize, srcCurSize;
};

extern "C" __global__ __aicore__ void kernel_asin_operator(GM_ADDR src_gm, GM_ADDR dst_gm, GM_ADDR tiling) 
{                                                                                                         
    GET_TILING_DATA(tilingData, tiling);
    KernelAsin<half> op;
    op.Init(src_gm, dst_gm, tilingData.srcSize, tilingData.srcCurSize, tilingData.tmpBufferSize);
    if (TILING_KEY_IS(1)) {
        op.Process();
    }
}

父主题： 数学库