aclnnCummin

aclnnCtcloss

接口原型

每个算子有两段接口，必须先调用“aclnnXxxGetWorkspaceSize”接口获取入参并根据计算流程计算所需workspace大小，再调用“aclnnXxx”接口执行计算。两段式接口如下：

**第一段接口：**aclnnStatus aclnnCtcLossGetWorkspaceSize(const aclTensor *logProbs, const aclTensor *targets, const aclIntArray *inputLengths, const aclIntArray *targetLengths, int64_t blank, bool zeroInfinity, aclTensor *negLoglikelihoodOut, aclTensor *logAlphaOut, uint64_t *workspaceSize, aclOpExecutor **executor)
**第二段接口：**aclnnStatus aclnnCtcLoss(void *workspace, uint64_t workspaceSize, aclOpExecutor *executor, const aclrtStream stream)

功能描述

算子功能：计算连接时序分类损失值。
计算公式：

定义ytk表示在时刻t时真实字符为k的概率。(一般地，ytk是经过softmax之后输出矩阵中的一个元素)。将字符集L′可以构成的所有序列的集合称为L′T，将L′T中的任意一个序列称为路径，并标记为π。π的分布为公式为：

定义多对一（many to one）映射B：L′T→L≤T，通过映射B计算得到 l∈L≤T的条件概率，等于对应于l的所有可能路径概率之和，公式如下：

将找到使p(l|x)值最大的l的路径的任务称为解码，公式如下：

aclnnCtcLossGetWorkspaceSize

接口定义：

aclnnStatus aclnnCtcLossGetWorkspaceSize(const aclTensor *logProbs, const aclTensor *targets, const aclIntArray *inputLengths, const aclIntArray *targetLengths, int64_t blank, bool zeroInfinity, aclTensor *negLoglikelihoodOut, aclTensor *logAlphaOut, uint64_t *workspaceSize, aclOpExecutor **executor)
参数说明：
- logProbs：Device侧的aclTensor，表示输出的对数概率。数据类型支持FLOAT、DOUBLE。当shape为(T, N, C)，其中T为输入长度，N为批处理大小，C为类别数。支持非连续的Tensor，数据格式支持ND。
- targets：Device侧的aclTensor，数据类型支持INT64、INT32、FLOAT、FLOAT16、BOOL。当shape为(N, S)，S为不小于targetLengths中的最大值。当targets是未填充的且在1维内级联的，shape为(SUM(targetLengths))。支持非连续的Tensor，数据格式支持ND。
- inputLengths：Host侧的aclIntArray，数据类型支持UINT8、INT8、INT16、INT32、INT64，数组长度为N，数组中的每个值必须≤T。
- targetLengths：Host侧的aclIntArray，数据类型支持UINT8、INT8、INT16、INT32、INT64，数组长度为N，当targets的shape为(N, S)时，数组中的每个值必须≤S。
- blank：int整型，空白标识，默认为0，取值范围为[0, C），C为类别数。
- zeroInfinity：bool类型，表示是否将无限损耗和相关梯度归零，默认值为False。
- negLogLikelihoodOut：Device侧的aclTensor，表示输出的损失值，数据类型支持FLOAT、DOUBLE（数据类型必须和logProbs一致），支持非连续的Tensor，数据格式支持ND。
- logAlphaOut：Device侧的aclTensor，表示输入到目标的可能跟踪的概率，shape必须为3维。数据类型支持FLOAT、DOUBLE（数据类型必须和logProbs一致），支持非连续的Tensor，数据格式支持ND。
- workspaceSize：返回用户需要在Device侧申请的workspace大小。
- executor：返回op执行器，包含了算子计算流程。
返回值：

返回aclnnStatus状态码，具体参见aclnn返回码。

:::note 说明第一段接口完成入参校验，出现以下场景时报错：
- 返回161001（ACLNN_ERR_PARAM_NULLPTR）：传入的logProbs、targets、inputLengths、targetLengths、negLogLikelihoodOut、logAlphaOut是空指针。
- 返回161002（ACLNN_ERR_PARAM_INVALID）：
  - logProbs、targets、inputLengths、targetLengths的数据类型不在支持的范围内。
  - logProbs、targets、inputLengths、targetLengths、negLogLikelihoodOut、logAlphaOut的shape不满足要求，或者inputLengths、targetLengths的ArrayList长度不满足要求。
  - blank不满足取值范围。 :::

aclnnCtcLoss

接口定义：

aclnnStatus aclnnCtcLoss(void *workspace, uint64_t workspaceSize, aclOpExecutor *executor, const aclrtStream stream)
参数说明：
- workspace：在Device侧申请的workspace内存起址。
- workspaceSize：在Device侧申请的workspace大小，由第一段接口aclnnCtcLossGetWorkspaceSize获取。
- executor：op执行器，包含了算子计算流程。
- stream：指定执行任务的AscendCL stream流。
返回值：

返回aclnnStatus状态码，具体参见aclnn返回码。

调用示例

#include <iostream>
#include <vector>
#include "acl/acl.h"
#include "aclnnop/aclnn_ctc_loss.h"

#define CHECK_RET(cond, return_expr) \
  do &#123;                               \
    if (!(cond)) &#123;                   \
      return_expr;                   \
    &#125;                                \
  &#125; while (0)

#define LOG_PRINT(message, ...)     \
  do &#123;                              \
    printf(message, ##__VA_ARGS__); \
  &#125; while (0)

int64_t GetShapeSize(const std::vector<int64_t>& shape) &#123;
  int64_t shape_size = 1;
  for (auto i : shape) &#123;
    shape_size *= i;
  &#125;
  return shape_size;
&#125;

int Init(int32_t deviceId, aclrtContext* context, aclrtStream* stream) &#123;
  // 固定写法，AscendCL初始化
  auto ret = aclInit(nullptr);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclInit failed. ERROR: %d\n", ret); return ret);
  ret = aclrtSetDevice(deviceId);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtSetDevice failed. ERROR: %d\n", ret); return ret);
  ret = aclrtCreateContext(context, deviceId);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtCreateContext failed. ERROR: %d\n", ret); return ret);
  ret = aclrtSetCurrentContext(*context);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtSetCurrentContext failed. ERROR: %d\n", ret); return ret);
  ret = aclrtCreateStream(stream);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtCreateStream failed. ERROR: %d\n", ret); return ret);
  return 0;
&#125;

template <typename T>
int CreateAclTensor(const std::vector<T>& hostData, const std::vector<int64_t>& shape, void** deviceAddr,
                    aclDataType dataType, aclTensor** tensor) &#123;
  auto size = GetShapeSize(shape) * sizeof(T);
  // 调用aclrtMalloc申请device侧内存
  auto ret = aclrtMalloc(deviceAddr, size, ACL_MEM_MALLOC_HUGE_FIRST);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtMalloc failed. ERROR: %d\n", ret); return ret);

  // 调用aclrtMemcpy将Host侧数据拷贝到device侧内存上
  ret = aclrtMemcpy(*deviceAddr, size, hostData.data(), size, ACL_MEMCPY_HOST_TO_DEVICE);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtMemcpy failed. ERROR: %d\n", ret); return ret);

  // 计算连续tensor的strides
  std::vector<int64_t> strides(shape.size(), 1);
  for (int64_t i = shape.size() - 2; i >= 0; i--) &#123;
    strides[i] = shape[i + 1] * strides[i + 1];
  &#125;

  // 调用aclCreateTensor接口创建aclTensor
  *tensor = aclCreateTensor(shape.data(), shape.size(), dataType, strides.data(), 0, aclFormat::ACL_FORMAT_ND, shape.data(), shape.size(), *deviceAddr);
  return 0;
&#125;

int main() &#123;
  // 1. （固定写法）device/context/stream初始化, 参考AscendCL对外接口列表
  // 根据自己的实际device填写deviceId
  int32_t deviceId = 0;
  aclrtContext context;
  aclrtStream stream;
  auto ret = Init(deviceId, &context, &stream);
  // check根据自己的需要处理
  CHECK_RET(ret == 0, LOG_PRINT("Init acl failed. ERROR: %d\n", ret); return ret);

  // 2. 构造输入与输出，需要根据API的接口自定义构造
  std::vector<int64_t> logProbsShape = &#123;12, 4, 5&#125;;
  std::vector<int64_t> targetsShape = &#123;4, 7&#125;;
  std::vector<int64_t> negLoglikelihoodOutShape = &#123;4&#125;;
  std::vector<int64_t> logAlphaOutShape = &#123;4, 12, 16&#125;;
  void* logProbsDeviceAddr = nullptr;
  void* targetsDeviceAddr = nullptr;
  void* negLoglikelihoodOutDeviceAddr = nullptr;
  void* logAlphaOutDeviceAddr = nullptr;
  aclTensor* logProbs = nullptr;
  aclTensor* targets = nullptr;
  aclIntArray* inputLengths = nullptr;
  aclIntArray* targetLengths = nullptr;
  aclTensor* negLoglikelihoodOut = nullptr;
  aclTensor* logAlphaOut = nullptr;
  std::vector<float> logProbsHostData = &#123;
    -1.0894, -2.7162, -0.9764, -1.9126, -2.6162,
    -2.0684, -2.4871, -2.0866, -1.7205, -0.7187,
    -2.4423, -1.2017, -1.4653, -1.1821, -2.5942,
    -2.4670, -2.7257, -1.4135, -2.1042, -0.7248,

    -3.7759, -1.3742, -1.2549, -1.5807, -1.4562,
    -1.3826, -1.8995, -1.8527, -0.9493, -2.8895,
    -1.6316, -2.6603, -2.5014, -0.6992, -1.8609,
    -1.9269, -2.2350, -0.8073, -1.8906, -1.8947,

    -0.3468, -2.5855, -2.0723, -2.7147, -3.6668,
    -0.9541, -1.7258, -2.0693, -1.6378, -2.1531,
    -3.5386, -3.4830, -0.2532, -2.0557, -3.3261,
    -1.1480, -1.8080, -0.8244, -3.2414, -3.1909,

    -0.8866, -0.7540, -4.4312, -3.4634, -2.6000,
    -1.2785, -1.8347, -3.3122, -0.7620, -2.8349,
    -1.4975, -1.3865, -0.9645, -3.8171, -2.0939,
    -2.3536, -2.0773, -1.4981, -0.8372, -2.0938,

    -1.2186, -0.8285, -2.9399, -2.1159, -2.3620,
    -2.3139, -0.6503, -2.7249, -1.2340, -3.7927,
    -0.7143, -2.5084, -3.2826, -2.6651, -1.1334,
    -1.6965, -1.9728, -2.3849, -1.6052, -0.9554,

    -1.6384, -1.2596, -2.1680, -1.8476, -1.3866,
    -3.0455, -0.5737, -2.5339, -2.1118, -1.6681,
    -2.4675, -2.8842, -0.4329, -3.6266, -1.6925,
    -3.1023, -2.7696, -1.2755, -0.6470, -2.4143,

    -2.0107, -2.0912, -1.3053, -0.8557, -3.0683,
    -1.2872, -3.6523, -1.6703, -2.7596, -0.8063,
    -2.4633, -1.2959, -1.6153, -2.3072, -1.0705,
    -3.0543, -0.6473, -1.1650, -2.9025, -2.7710,

    -3.5519, -2.0400, -1.8667, -1.4289, -0.8050,
    -1.4602, -0.7452, -1.5754, -3.1624, -3.1247,
    -1.4677, -1.2725, -2.9575, -1.8883, -1.2513,
    -1.2164, -1.5894, -2.2217, -2.3714, -1.2110,

    -2.0843, -0.6515, -1.4252, -2.9402, -2.7964,
    -1.5261, -2.5471, -1.7167, -1.9846, -0.9488,
    -1.4847, -1.7093, -1.4095, -1.7293, -1.7675,
    -0.9203, -4.2299, -1.8740, -1.4076, -1.6671,

    -1.9052, -0.8330, -2.1839, -2.2459, -1.6193,
    -2.9108, -1.2114, -1.4616, -1.7297, -1.4330,
    -2.2656, -0.7878, -1.8533, -1.8711, -2.0349,
    -2.2457, -2.1395, -1.4509, -0.7538, -2.6381,

    -0.8078, -2.1054, -2.6703, -1.1108, -3.3867,
    -1.7774, -1.8426, -1.9473, -1.3293, -1.3273,
    -1.3490, -1.9842, -2.5357, -2.2161, -0.8800,
    -1.5412, -1.8003, -2.7603, -0.8606, -2.0066,

    -1.8342, -2.2741, -1.8348, -1.5833, -0.9877,
    -3.5196, -2.3361, -0.9124, -0.9307, -2.5531,
    -1.4862, -1.2153, -1.4453, -3.4462, -1.5625,
    -2.6455, -1.4153, -1.3079, -1.1568, -2.2897&#125;;
  std::vector<int64_t> targetsHostData = &#123;
    1, 2, 1, 1, 2, 4, 1,
    2, 2, 2, 2, 2, 2, 3,
    4, 2, 1, 4, 3, 1, 4,
    4, 1, 4, 2, 2, 2, 3&#125;;

  std::vector<float> negLoglikelihoodOutHostData = &#123;0, 0, 0, 0&#125;;
  std::vector<float> logAlphaOutHostData = &#123;
    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,

    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,

    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,

    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0&#125;;
  // 创建logProbs aclTensor
  ret = CreateAclTensor(logProbsHostData, logProbsShape, &logProbsDeviceAddr, aclDataType::ACL_FLOAT, &logProbs);
  CHECK_RET(ret == ACL_SUCCESS, return ret);
  // 创建targets aclTensor
  ret = CreateAclTensor(targetsHostData, targetsShape, &targetsDeviceAddr, aclDataType::ACL_INT64, &targets);
  CHECK_RET(ret == ACL_SUCCESS, return ret);
  std::vector<int64_t> inputLengthsSizeData = &#123;10,10,10,10&#125;;
  inputLengths = aclCreateIntArray(inputLengthsSizeData.data(), 4);
  CHECK_RET(inputLengths != nullptr, return ACL_ERROR_BAD_ALLOC);
  std::vector<int64_t> targetLengthsSizeData = &#123;2, 3, 1, 5&#125;;
  targetLengths = aclCreateIntArray(targetLengthsSizeData.data(), 4);
  CHECK_RET(targetLengths != nullptr, return ACL_ERROR_BAD_ALLOC);
  // 创建negLoglikelihoodOut aclTensor
  ret = CreateAclTensor(negLoglikelihoodOutHostData, negLoglikelihoodOutShape, &negLoglikelihoodOutDeviceAddr, aclDataType::ACL_FLOAT, &negLoglikelihoodOut);
  CHECK_RET(ret == ACL_SUCCESS, return ret);
  // 创建logAlphaOut aclTensor
  ret = CreateAclTensor(logAlphaOutHostData, logAlphaOutShape, &logAlphaOutDeviceAddr, aclDataType::ACL_FLOAT, &logAlphaOut);
  CHECK_RET(ret == ACL_SUCCESS, return ret);

  // 3.调用CANN算子库API，需要修改为具体的算子接口
  uint64_t workspaceSize = 0;
  aclOpExecutor* executor;
  // 调用aclnnCtcLoss第一段接口
  ret = aclnnCtcLossGetWorkspaceSize(logProbs, targets, inputLengths, targetLengths, 0, false, negLoglikelihoodOut, logAlphaOut, &workspaceSize, &executor);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnCtcLossGetWorkspaceSize failed. ERROR: %d\n", ret); return ret);
  // 根据第一段接口计算出的workspaceSize申请device内存
  void* workspaceAddr = nullptr;
  if (workspaceSize > 0) &#123;
    ret = aclrtMalloc(&workspaceAddr, workspaceSize, ACL_MEM_MALLOC_HUGE_FIRST);
    CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("allocate workspace failed. ERROR: %d\n", ret); return ret;);
  &#125;
  // 调用aclnnCtcLoss第二段接口
  ret = aclnnCtcLoss(workspaceAddr, workspaceSize, executor, stream);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnCtcLoss failed. ERROR: %d\n", ret); return ret);

  // 4. （固定写法）同步等待任务执行结束
  ret = aclrtSynchronizeStream(stream);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtSynchronizeStream failed. ERROR: %d\n", ret); return ret);

  // 5. 获取输出的negLoglikelihoodOut值，将device侧内存上的结果拷贝至Host侧，需要根据具体API的接口定义修改
  auto size = GetShapeSize(negLoglikelihoodOutShape);
  std::vector<float> resultData(size, 0);
  ret = aclrtMemcpy(resultData.data(), resultData.size() * sizeof(resultData[0]), negLoglikelihoodOutDeviceAddr, size * sizeof(float),ACL_MEMCPY_DEVICE_TO_HOST);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("copy result from device to host failed. ERROR: %d\n", ret); return ret);
  for (int64_t i = 0; i < size; i++) &#123;
    LOG_PRINT("negLoglikelihoodOut result[%ld] is: %f\n", i, resultData[i]);
  &#125;

  // 6. 获取输出的logAlphaOut值，将device侧内存上的结果拷贝至Host侧，需要根据具体API的接口定义修改
  auto size1 = GetShapeSize(logAlphaOutShape);
  std::vector<float> resultData1(size1, 0);
  ret = aclrtMemcpy(resultData1.data(), resultData1.size() * sizeof(resultData1[0]), logAlphaOutDeviceAddr, size1 * sizeof(float),
                    ACL_MEMCPY_DEVICE_TO_HOST);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("copy result from device to host failed. ERROR: %d\n", ret); return ret);
  for (int64_t i = 0; i < size1; i++) &#123;
    LOG_PRINT("logAlphaOut result[%ld] is: %f\n", i, resultData1[i]);
  &#125;

  // 7. 释放aclTensor和aclScalar，需要根据具体API的接口定义修改
  aclDestroyTensor(logProbs);
  aclDestroyTensor(targets);
  aclDestroyIntArray(inputLengths);
  aclDestroyIntArray(targetLengths);
  aclDestroyTensor(negLoglikelihoodOut);
  aclDestroyTensor(logAlphaOut);
  return 0;
&#125;

父主题： NN类算子接口

aclnnCtcLossBackward

接口原型

第一段接口：aclnnStatus aclnnCtcLossBackwardGetWorkspaceSize(const aclTensor *gradOut, const aclTensor *logProbs, const aclTensor *targets, const aclIntArray *inputLengths, const aclIntArray *targetLengths, const aclTensor *negLogLikelihood, const aclTensor *logAlpha, int64_t blank, bool zeroInfinity, aclTensor *out, uint64_t *workspaceSize, aclOpExecutor **executor)
第二段接口：aclnnStatus aclnnCtcLossBackward(void *workspace, uint64_t workspaceSize, aclOpExecutor *executor, const aclrtStream stream)

功能描述

算子功能：损失函数aclnnCtcloss的反向传播。

aclnnCtcLossBackwardGetWorkspaceSize

接口定义：

aclnnStatus aclnnCtcLossBackwardGetWorkspaceSize(const aclTensor *gradOut, const aclTensor *logProbs, const aclTensor *targets, const aclIntArray *inputLengths, const aclIntArray *targetLengths, const aclTensor *negLogLikelihood, const aclTensor *logAlpha, int64_t blank, bool zeroInfinity, aclTensor *out, uint64_t *workspaceSize, aclOpExecutor **executor)
参数说明：
- gradOut：Device侧的aclTensor，表示梯度更新系数，必须为1维非空Tensor，数据类型支持FLOAT、DOUBLE（数据类型必须和logProbs一致），支持非连续的Tensor，数据格式支持ND。
- logProbs：Device侧的aclTensor，表示输出的对数概率，数据类型支持FLOAT、DOUBLE。当shape为(T, N, C)，其中T为输入长度，N为批处理大小，C为类别数。支持非连续的Tensor，数据格式支持ND。
- targets：Device侧的aclTensor，数据类型支持INT64、INT32、FLOAT、FLOAT16、BOOL。当shape为(N, S)，S为不小于targetLengths中的最大值。当targets是未填充的且在1维内级联的，shape为(SUM(targetLengths))。支持非连续的Tensor，数据格式支持ND。
- inputLengths：Host侧int数组，数据类型支持UINT8、INT8、INT16、INT32、INT64，数组长度为N，数组中的每个值必须≤T。
- targetLengths：Host侧int数组，数据类型支持UINT8、INT8、INT16、INT32、INT64，数组长度为N，当targets的shape为(N, S)时，数组中的每个值必须≤S。
- negLogLikelihood：Device侧的aclTensor，表示相对于每个输入节点可微分的损失值，必须为1维非空Tensor。数据类型支持FLOAT、DOUBLE（数据类型必须和logProbs一致），支持非连续的Tensor，数据格式支持ND。
- logAlpha：Device侧的aclTensor，表示输入到目标的可能跟踪的概率，必须为3维非空Tensor。数据类型支持FLOAT、DOUBLE（数据类型必须和logProbs一致），支持非连续的Tensor，数据格式支持ND。
- blank：int整型，空白标识，默认为0，取值范围为[0, C），C为类别数。
- zeroInfinity：bool类型，表示是否将无限损耗和相关梯度归零，默认值为False。
- out：Device侧的aclTensor，表示CTC的损失梯度，数据类型支持FLOAT、DOUBLE（数据类型必须和gradOut一致），shape为(T, N, C)，支持非连续的Tensor，数据格式支持ND。
- workspaceSize：返回用户需要在Device侧申请的workspace大小。
- executor：返回op执行器，包含了算子计算流程。
返回值：

返回aclnnStatus状态码，具体参见aclnn返回码。

:::note 说明第一段接口完成入参校验，出现以下场景时报错：
- 返回161001（ACLNN_ERR_PARAM_NULLPTR）：传入的gradOut、logProbs、targets、inputLengths、targetLengths、negLogLikelihood、logAlpha、out是空指针。
- 返回161002（ACLNN_ERR_PARAM_INVALID）：
  - gradOut、logProbs、targets、inputLengths、targetLengths、negLogLikelihood、logAlpha、out的数据类型不在支持的范围内。
  - gradOut、negLogLikelihood、logAlpha、out和logProbs的数据类型不一致。
  - gradOut、logProbs、targets、negLogLikelihood、logAlpha、out的shape不满足要求，或者inputLengths、targetLengths的ArrayList的长度不满足要求。
  - blank不满足取值范围。 :::

aclnnCtcLossBackward

接口定义：

aclnnStatus aclnnCtcLossBackward(void *workspace, uint64_t workspaceSize, aclOpExecutor *executor, const aclrtStream stream)
参数说明：
- workspace：在Device侧申请的workspace内存起址。
- workspaceSize：在Device侧申请的workspace大小，由第一段接口aclnnCtcLossBackwardGetWorkspaceSize获取。
- executor：op执行器，包含了算子计算流程。
- stream：指定执行任务的AscendCL stream流。
返回值：

返回aclnnStatus状态码，具体参见aclnn返回码。

调用示例

#include <iostream>
#include <vector>
#include "acl/acl.h"
#include "aclnnop/aclnn_ctc_loss_backward.h"

#define CHECK_RET(cond, return_expr) \
  do &#123;                               \
    if (!(cond)) &#123;                   \
      return_expr;                   \
    &#125;                                \
  &#125; while (0)

#define LOG_PRINT(message, ...)     \
  do &#123;                              \
    printf(message, ##__VA_ARGS__); \
  &#125; while (0)

int64_t GetShapeSize(const std::vector<int64_t>& shape) &#123;
  int64_t shape_size = 1;
  for (auto i : shape) &#123;
    shape_size *= i;
  &#125;
  return shape_size;
&#125;

int Init(int32_t deviceId, aclrtContext* context, aclrtStream* stream) &#123;
  // 固定写法，AscendCL初始化
  auto ret = aclInit(nullptr);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclInit failed. ERROR: %d\n", ret); return ret);
  ret = aclrtSetDevice(deviceId);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtSetDevice failed. ERROR: %d\n", ret); return ret);
  ret = aclrtCreateContext(context, deviceId);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtCreateContext failed. ERROR: %d\n", ret); return ret);
  ret = aclrtSetCurrentContext(*context);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtSetCurrentContext failed. ERROR: %d\n", ret); return ret);
  ret = aclrtCreateStream(stream);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtCreateStream failed. ERROR: %d\n", ret); return ret);
  return 0;
&#125;

template <typename T>
int CreateAclTensor(const std::vector<T>& hostData, const std::vector<int64_t>& shape, void** deviceAddr,
                    aclDataType dataType, aclTensor** tensor) &#123;
  auto size = GetShapeSize(shape) * sizeof(T);
  // 调用aclrtMalloc申请device侧内存
  auto ret = aclrtMalloc(deviceAddr, size, ACL_MEM_MALLOC_HUGE_FIRST);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtMalloc failed. ERROR: %d\n", ret); return ret);

  // 调用aclrtMemcpy将Host侧数据拷贝到device侧内存上
  ret = aclrtMemcpy(*deviceAddr, size, hostData.data(), size, ACL_MEMCPY_HOST_TO_DEVICE);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtMemcpy failed. ERROR: %d\n", ret); return ret);

  // 计算连续tensor的strides
  std::vector<int64_t> strides(shape.size(), 1);
  for (int64_t i = shape.size() - 2; i >= 0; i--) &#123;
    strides[i] = shape[i + 1] * strides[i + 1];
  &#125;

  // 调用aclCreateTensor接口创建aclTensor
  *tensor = aclCreateTensor(shape.data(), shape.size(), dataType, strides.data(), 0, aclFormat::ACL_FORMAT_ND, shape.data(), shape.size(), *deviceAddr);
  return 0;
&#125;

int main() &#123;
  // 1. （固定写法）device/context/stream初始化, 参考AscendCL对外接口列表
  // 根据自己的实际device填写deviceId
  int32_t deviceId = 0;
  aclrtContext context;
  aclrtStream stream;
  auto ret = Init(deviceId, &context, &stream);
  // check根据自己的需要处理
  CHECK_RET(ret == 0, LOG_PRINT("Init acl failed. ERROR: %d\n", ret); return ret);

  // 2. 构造输入与输出，需要根据API的接口自定义构造
  std::vector<int64_t> gradOutShape = &#123;4&#125;;
  // logProbsShape (T, N, C)
  std::vector<int64_t> logProbsShape = &#123;12, 4, 5&#125;;
  std::vector<int64_t> targetsShape = &#123;4, 7&#125;;
  std::vector<int64_t> negLoglikelihoodShape = &#123;4&#125;;
  // logAlphaShape (N, T, X)  X = ((max(targetLengths) * 2 + 1) + 7) / 8 * 8;
  std::vector<int64_t> logAlphaShape = &#123;4, 12, 16&#125;;
  std::vector<int64_t> outShape = &#123;12, 4, 5&#125;;
  void* gradOutDeviceAddr = nullptr;
  void* logProbsDeviceAddr = nullptr;
  void* targetsDeviceAddr = nullptr;
  void* negLoglikelihoodDeviceAddr = nullptr;
  void* logAlphaDeviceAddr = nullptr;
  void* outDeviceAddr = nullptr;
  aclTensor* gradOut = nullptr;
  aclTensor* logProbs = nullptr;
  aclTensor* targets = nullptr;
  aclIntArray* inputLengths = nullptr;
  aclIntArray* targetLengths = nullptr;
  aclTensor* negLoglikelihood = nullptr;
  aclTensor* logAlpha = nullptr;
  aclTensor* out = nullptr;
  std::vector<float> gradOutHostData = &#123;1, 1, 1, 1&#125;;
  std::vector<float> logProbsHostData = &#123;
    -1.0894, -2.7162, -0.9764, -1.9126, -2.6162,
    -2.0684, -2.4871, -2.0866, -1.7205, -0.7187,
    -2.4423, -1.2017, -1.4653, -1.1821, -2.5942,
    -2.4670, -2.7257, -1.4135, -2.1042, -0.7248,

    -3.7759, -1.3742, -1.2549, -1.5807, -1.4562,
    -1.3826, -1.8995, -1.8527, -0.9493, -2.8895,
    -1.6316, -2.6603, -2.5014, -0.6992, -1.8609,
    -1.9269, -2.2350, -0.8073, -1.8906, -1.8947,

    -0.3468, -2.5855, -2.0723, -2.7147, -3.6668,
    -0.9541, -1.7258, -2.0693, -1.6378, -2.1531,
    -3.5386, -3.4830, -0.2532, -2.0557, -3.3261,
    -1.1480, -1.8080, -0.8244, -3.2414, -3.1909,

    -0.8866, -0.7540, -4.4312, -3.4634, -2.6000,
    -1.2785, -1.8347, -3.3122, -0.7620, -2.8349,
    -1.4975, -1.3865, -0.9645, -3.8171, -2.0939,
    -2.3536, -2.0773, -1.4981, -0.8372, -2.0938,

    -1.2186, -0.8285, -2.9399, -2.1159, -2.3620,
    -2.3139, -0.6503, -2.7249, -1.2340, -3.7927,
    -0.7143, -2.5084, -3.2826, -2.6651, -1.1334,
    -1.6965, -1.9728, -2.3849, -1.6052, -0.9554,

    -1.6384, -1.2596, -2.1680, -1.8476, -1.3866,
    -3.0455, -0.5737, -2.5339, -2.1118, -1.6681,
    -2.4675, -2.8842, -0.4329, -3.6266, -1.6925,
    -3.1023, -2.7696, -1.2755, -0.6470, -2.4143,

    -2.0107, -2.0912, -1.3053, -0.8557, -3.0683,
    -1.2872, -3.6523, -1.6703, -2.7596, -0.8063,
    -2.4633, -1.2959, -1.6153, -2.3072, -1.0705,
    -3.0543, -0.6473, -1.1650, -2.9025, -2.7710,

    -3.5519, -2.0400, -1.8667, -1.4289, -0.8050,
    -1.4602, -0.7452, -1.5754, -3.1624, -3.1247,
    -1.4677, -1.2725, -2.9575, -1.8883, -1.2513,
    -1.2164, -1.5894, -2.2217, -2.3714, -1.2110,

    -2.0843, -0.6515, -1.4252, -2.9402, -2.7964,
    -1.5261, -2.5471, -1.7167, -1.9846, -0.9488,
    -1.4847, -1.7093, -1.4095, -1.7293, -1.7675,
    -0.9203, -4.2299, -1.8740, -1.4076, -1.6671,

    -1.9052, -0.8330, -2.1839, -2.2459, -1.6193,
    -2.9108, -1.2114, -1.4616, -1.7297, -1.4330,
    -2.2656, -0.7878, -1.8533, -1.8711, -2.0349,
    -2.2457, -2.1395, -1.4509, -0.7538, -2.6381,

    -0.8078, -2.1054, -2.6703, -1.1108, -3.3867,
    -1.7774, -1.8426, -1.9473, -1.3293, -1.3273,
    -1.3490, -1.9842, -2.5357, -2.2161, -0.8800,
    -1.5412, -1.8003, -2.7603, -0.8606, -2.0066,

    -1.8342, -2.2741, -1.8348, -1.5833, -0.9877,
    -3.5196, -2.3361, -0.9124, -0.9307, -2.5531,
    -1.4862, -1.2153, -1.4453, -3.4462, -1.5625,
    -2.6455, -1.4153, -1.3079, -1.1568, -2.2897&#125;;
  std::vector<int64_t> targetsHostData = &#123;
    1, 2, 1, 1, 2, 4, 1,
    2, 2, 2, 2, 2, 2, 3,
    4, 2, 1, 4, 3, 1, 4,
    4, 1, 4, 2, 2, 2, 3&#125;;

  std::vector<float> negLoglikelihoodHostData = &#123;10.1999, 16.1340, 14.9006,  9.3596&#125;;
  std::vector<float> logAlphaHostData = &#123;
    -1.0894, -2.7162, -99999, -99999, -99999, -99999, -99999, -99999, -99999, -99999, -99999, -99999, -99999, -99999, -99999, -99999,
    -4.8653, -2.2842, -6.4921, -3.9711, -99999, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,
    -5.2121, -4.7967, -2.6162, -4.1742, -4.3179, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,
    -6.0987, -5.0438, -3.3957, -6.7671, -4.4369, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,
    -7.3173, -5.5735, -4.4384, -6.1313, -5.5627, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,
    -8.9557, -6.6720, -5.7981, -6.1973, -6.7523, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,
    -10.9664, -8.6661, -7.4600, -6.3671, -7.7544, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,
    -14.5183, -10.6106, -10.7501, -7.8722, -9.6961, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,
    -16.6026, -11.2422, -12.0691, -9.1833, -9.8069, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,
    -18.5078, -12.0705, -12.7846, -11.1988, -10.6593, 0.0000, 0.0000,0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,0.0000, 0.0000,
    0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,0.0000, 0.0000,
    0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,0.0000, 0.0000,

    -2.0684, -2.0866, -99999, -99999, -99999, -99999, -99999,-99999, -99999, -99999, -99999, -99999, -99999, -99999,-99999, -99999,
    -3.4510, -3.2370, -3.4692, -99999, -99999, -99999, -99999,-99999, -99999, -99999, -99999, 0.0000, 0.0000, 0.0000,0.0000, 0.0000,
    -4.4051, -4.7144, -3.6073, -5.5385, -99999, -99999, -99999,-99999, -99999, -99999, -99999, 0.0000, 0.0000, 0.0000,0.0000, 0.0000,
    -5.6836, -7.1669, -4.6003, -6.7841, -6.8170, -99999, -99999,-99999, -99999, -99999, -99999, 0.0000, 0.0000, 0.0000,0.0000, 0.0000,
    -7.9975, -8.2040, -6.8402, -7.2185, -8.4212, -9.5419, -99999,-99999, -99999, -99999, -99999, 0.0000, 0.0000, 0.0000,0.0000, 0.0000,
    -11.0430, -9.9362, -9.6580, -8.8523, -10.0013, -10.6729, -12.5874,-99999, -99999, -99999, -99999, 0.0000, 0.0000, 0.0000,0.0000, 0.0000,
    -12.3302, -11.3209, -10.3815, -10.1533, -9.8642, -11.2589, -11.8226,-14.2577, -99999, -99999, -99999, 0.0000, 0.0000, 0.0000,0.0000, 0.0000,
    -13.7904, -12.5855, -11.5118, -11.1431, -10.7654, -11.2181, -12.2686,-13.3140, -15.7179, -99999, -99999, 0.0000, 0.0000, 0.0000,0.0000, 0.0000,
    -15.3165, -14.0400, -12.7439, -12.3341, -11.7695, -11.9899, -12.4443,-13.6840, -14.7536, -17.4346, -99999, 0.0000, 0.0000, 0.0000,0.0000, 0.0000,
    -18.2273, -15.2555, -15.4129, -13.2866, -14.2301, -12.6421, -14.4091,-13.6517, -16.2998, -16.1490, -20.3454, 0.0000, 0.0000, 0.0000,0.0000, 0.0000,
    0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,0.0000, 0.0000,
    0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,0.0000, 0.0000,

    -2.4423, -2.5942, -99999, -99999, -99999, -99999, -99999,-99999, -99999, -99999, -99999, -99999, -99999, -99999,-99999, -99999,
    -4.0739, -3.6831, -4.2258, 0.0000, 0.0000, 0.0000, 0.0000,0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,0.0000, 0.0000,
    -7.6125, -6.4925, -6.7635, 0.0000, 0.0000, 0.0000, 0.0000,0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,0.0000, 0.0000,
    -9.1100, -8.3040, -7.4232, 0.0000, 0.0000, 0.0000, 0.0000,0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,0.0000, 0.0000,
    -9.8243, -9.0682, -7.7908, 0.0000, 0.0000, 0.0000, 0.0000,0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,0.0000, 0.0000,
    -12.2918, -10.3758, -10.0124, 0.0000, 0.0000, 0.0000, 0.0000,0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,0.0000, 0.0000,
    -14.7551, -11.3089, -11.9478, 0.0000, 0.0000, 0.0000, 0.0000,0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,0.0000, 0.0000,
    -16.2228, -12.5289, -12.3528, 0.0000, 0.0000, 0.0000, 0.0000,0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,0.0000, 0.0000,
    -17.7075, -14.2718, -13.2285, 0.0000, 0.0000, 0.0000, 0.0000,0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,0.0000, 0.0000,
    -19.9731, -16.2750, -15.1923, 0.0000, 0.0000, 0.0000, 0.0000,0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,0.0000, 0.0000,
    0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,0.0000, 0.0000,
    0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,0.0000, 0.0000,

    -2.4670, -0.7248, -99999, -99999, -99999, -99999, -99999,-99999, -99999, -99999, -99999, -99999, -99999, -99999,-99999, -99999,
    -4.3939, -2.4581, -2.6517, -2.9598, -99999, -99999, -99999,-99999, -99999, -99999, -99999, 0.0000, 0.0000, 0.0000,0.0000, 0.0000,
    -5.5419, -5.5142, -3.0051, -3.3784, -4.1078, -6.1507, -99999,-99999, -99999, -99999, -99999, 0.0000, 0.0000, 0.0000,0.0000, 0.0000,
    -7.8955, -6.9286, -5.2805, -4.5115, -5.3385, -5.0374, -8.5043,-7.6488, -99999, -99999, -99999, 0.0000, 0.0000, 0.0000,0.0000, 0.0000,
    -9.5920, -7.5617, -6.8010, -6.0444, -5.8452, -4.7597, -6.7031,-7.3228, -9.3453, -99999, -99999, 0.0000, 0.0000, 0.0000,0.0000, 0.0000,
    -12.6943, -9.8527, -9.5199, -8.2901, -8.3490, -6.6950, -7.7281,-5.8361, -10.3008, -10.6208, -99999, 0.0000, 0.0000, 0.0000,0.0000, 0.0000,
    -15.7486, -12.5670, -12.0336, -8.5306, -10.6803, -9.1337, -9.4448,-6.5472, -8.8790, -10.9199, -13.6751, 0.0000, 0.0000, 0.0000,0.0000, 0.0000,
    -16.9650, -13.7373, -12.7884, -10.0734, -9.6368, -9.2326, -9.8004,-8.6463, -7.6709, -10.9785, -12.0746, 0.0000, 0.0000, 0.0000,0.0000, 0.0000,
    -17.8853, -15.3655, -13.3814, -14.2154, -10.0586, -10.1583, -9.7039,-9.8935, -8.2713, -9.5090, -11.6105, 0.0000, 0.0000, 0.0000,0.0000, 0.0000,
    -20.1310, -17.9262, -15.4983, -15.0687, -12.2888, -12.0440, -11.4581,-10.2538, -10.3368, -9.4675, -11.6393, 0.0000, 0.0000, 0.0000,0.0000, 0.0000,
    0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,0.0000, 0.0000,
    0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,0.0000, 0.0000&#125;;

  std::vector<float> outHostData = &#123;
    0, 0, 0, 0, 0,
    0, 0, 0, 0, 0,
    0, 0, 0, 0, 0,
    0, 0, 0, 0, 0,

    0, 0, 0, 0, 0,
    0, 0, 0, 0, 0,
    0, 0, 0, 0, 0,
    0, 0, 0, 0, 0,

    0, 0, 0, 0, 0,
    0, 0, 0, 0, 0,
    0, 0, 0, 0, 0,
    0, 0, 0, 0, 0,

    0, 0, 0, 0, 0,
    0, 0, 0, 0, 0,
    0, 0, 0, 0, 0,
    0, 0, 0, 0, 0,

    0, 0, 0, 0, 0,
    0, 0, 0, 0, 0,
    0, 0, 0, 0, 0,
    0, 0, 0, 0, 0,

    0, 0, 0, 0, 0,
    0, 0, 0, 0, 0,
    0, 0, 0, 0, 0,
    0, 0, 0, 0, 0,

    0, 0, 0, 0, 0,
    0, 0, 0, 0, 0,
    0, 0, 0, 0, 0,
    0, 0, 0, 0, 0,

    0, 0, 0, 0, 0,
    0, 0, 0, 0, 0,
    0, 0, 0, 0, 0,
    0, 0, 0, 0, 0,

    0, 0, 0, 0, 0,
    0, 0, 0, 0, 0,
    0, 0, 0, 0, 0,
    0, 0, 0, 0, 0,

    0, 0, 0, 0, 0,
    0, 0, 0, 0, 0,
    0, 0, 0, 0, 0,
    0, 0, 0, 0, 0,

    0, 0, 0, 0, 0,
    0, 0, 0, 0, 0,
    0, 0, 0, 0, 0,
    0, 0, 0, 0, 0,

    0, 0, 0, 0, 0,
    0, 0, 0, 0, 0,
    0, 0, 0, 0, 0,
    0, 0, 0, 0, 0
  &#125;;
  // 创建gradOut aclTensor
  ret = CreateAclTensor(gradOutHostData, gradOutShape, &gradOutDeviceAddr, aclDataType::ACL_FLOAT, &gradOut);
  CHECK_RET(ret == ACL_SUCCESS, return ret);
  // 创建logProbs aclTensor
  ret = CreateAclTensor(logProbsHostData, logProbsShape, &logProbsDeviceAddr, aclDataType::ACL_FLOAT, &logProbs);
  CHECK_RET(ret == ACL_SUCCESS, return ret);
  // 创建targets aclTensor
  ret = CreateAclTensor(targetsHostData, targetsShape, &targetsDeviceAddr, aclDataType::ACL_INT64, &targets);
  CHECK_RET(ret == ACL_SUCCESS, return ret);
  std::vector<int64_t> inputLengthsSizeData = &#123;10,10,10,10&#125;;
  inputLengths = aclCreateIntArray(inputLengthsSizeData.data(), 4);
  CHECK_RET(inputLengths != nullptr, return ACL_ERROR_BAD_ALLOC);
  std::vector<int64_t> targetLengthsSizeData = &#123;2, 3, 1, 5&#125;;
  targetLengths = aclCreateIntArray(targetLengthsSizeData.data(), 4);
  CHECK_RET(targetLengths != nullptr, return ACL_ERROR_BAD_ALLOC);
  // 创建negLoglikelihood aclTensor
  ret = CreateAclTensor(negLoglikelihoodHostData, negLoglikelihoodShape, &negLoglikelihoodDeviceAddr, aclDataType::ACL_FLOAT, &negLoglikelihood);
  CHECK_RET(ret == ACL_SUCCESS, return ret);
  // 创建logAlpha aclTensor
  ret = CreateAclTensor(logAlphaHostData, logAlphaShape, &logAlphaDeviceAddr, aclDataType::ACL_FLOAT, &logAlpha);
  CHECK_RET(ret == ACL_SUCCESS, return ret);
  // 创建out aclTensor
  ret = CreateAclTensor(outHostData, outShape, &outDeviceAddr, aclDataType::ACL_FLOAT, &out);
  CHECK_RET(ret == ACL_SUCCESS, return ret);

  // 3.调用CANN算子库API，需要修改为具体的算子接口
  uint64_t workspaceSize = 0;
  aclOpExecutor* executor;
  // 调用aclnnCtcLossBackward第一段接口
  ret = aclnnCtcLossBackwardGetWorkspaceSize(gradOut, logProbs, targets, inputLengths, targetLengths, negLoglikelihood, logAlpha, 0, false, out, &workspaceSize, &executor);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnCtcLossBackwardGetWorkspaceSize failed. ERROR: %d\n", ret); return ret);
  // 根据第一段接口计算出的workspaceSize申请device内存
  void* workspaceAddr = nullptr;
  if (workspaceSize > 0) &#123;
    ret = aclrtMalloc(&workspaceAddr, workspaceSize, ACL_MEM_MALLOC_HUGE_FIRST);
    CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("allocate workspace failed. ERROR: %d\n", ret); return ret;);
  &#125;
  // 调用aclnnCtcLossBackward第二段接口
  ret = aclnnCtcLossBackward(workspaceAddr, workspaceSize, executor, stream);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnCtcLossBackward failed. ERROR: %d\n", ret); return ret);

  // 4. （固定写法）同步等待任务执行结束
  ret = aclrtSynchronizeStream(stream);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtSynchronizeStream failed. ERROR: %d\n", ret); return ret);

  // 5. 获取输出的out值，将device侧内存上的结果拷贝至Host侧，需要根据具体API的接口定义修改
  auto size = GetShapeSize(outShape);
  std::vector<float> resultData(size, 0);
  ret = aclrtMemcpy(resultData.data(), resultData.size() * sizeof(resultData[0]), outDeviceAddr, size * sizeof(float), ACL_MEMCPY_DEVICE_TO_HOST);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("copy result from device to host failed. ERROR: %d\n", ret); return ret);
  for (int64_t i = 0; i < size; i++) &#123;
    LOG_PRINT("out result[%ld] is: %f\n", i, resultData[i]);
  &#125;

  // 6. 释放aclTensor和aclScalar，需要根据具体API的接口定义修改
  aclDestroyTensor(gradOut);
  aclDestroyTensor(logProbs);
  aclDestroyTensor(targets);
  aclDestroyIntArray(inputLengths);
  aclDestroyIntArray(targetLengths);
  aclDestroyTensor(negLoglikelihood);
  aclDestroyTensor(logAlpha);
  aclDestroyTensor(out);
  return 0;
&#125;

父主题： NN类算子接口

aclnnCummax

接口原型

**第一段接口：**aclnnStatus aclnnCummaxGetWorkspaceSize(const aclTensor *self, int64_t dim, aclTensor *valuesOut, aclTensor* indicesOut, uint64_t *workspaceSize, aclOpExecutor **executor)
**第二段接口：**aclnnStatus aclnnCummax(void *workspace, uint64_t workspaceSize, aclOpExecutor *executor, aclrtStream stream)

功能描述

算子功能：计算张量self中的累积最大值，并返回该值以及对应的索引。

例如valuesOut[1]为self[0]、self[1]之间的最大值，indicesOut[1]为该最大值的索引；valueOut[2]为self[0]、self[1]、self[2]之间的最大值，indicesOut[2]为该最大值索引。
计算公式：

aclnnCummaxGetWorkspaceSize

接口定义：

aclnnStatus aclnnCummaxGetWorkspaceSize(const aclTensor *self, int64_t dim, aclTensor *valuesOut, aclTensor* indicesOut, uint64_t *workspaceSize, aclOpExecutor **executor)
参数说明：
- self：Device侧的aclTensor，输入张量，数据类型支持FLOAT、FLOAT16、DOUBLE、INT32、INT64、INT16、INT8、UINT8、BOOL，支持非连续的Tensor，数据格式支持ND，且shape需要与valuesOut、indicesOut一致。
- dim：Host侧INT64类型，指定要进行最大值计算的维度，取值范围[-self.dim(), self.dim()）。
- valuesOut：Device侧的aclTensor，输出的最大值，数据类型支持FLOAT、FLOAT16、DOUBLE、INT32、INT64、INT16、INT8、UINT8、BOOL，支持非连续的Tensor，数据格式支持ND，且shape需要与self、indicesOut一致。
- indicesOut：Device侧的aclTensor，输出的最大值索引，数据类型支持INT32、INT64，支持非连续的Tensor，数据格式支持ND，且shape需要与self、valuesOut一致。
- workspaceSize：返回用户需要在Device侧申请的workspace大小。
- executor：返回op执行器，包含了算子计算流程。
返回值：

返回aclnnStatus状态码，具体参见aclnn返回码。

:::note 说明第一段接口完成入参校验，出现以下场景时报错：
- 返回161001（ACLNN_ERR_PARAM_NULLPTR）：传入的self、valuesOut、indicesOut是空指针。
- 返回161002（ACLNN_ERR_PARAM_INVALID）：
  - self、valuesOut、indicesOut的数据类型不在支持的范围之内。
  - self、valuesOut、indicesOut的shape不在支持的范围之内。
  - 当self为0维时，不支持传入dim。
  - 输入的dim值不合法。 :::

aclnnCummax

接口定义：

aclnnStatus aclnnCummax(void *workspace, uint64_t workspaceSize, aclOpExecutor *executor, aclrtStream stream)
参数说明：
- workspace：在Device侧申请的workspace内存起址。
- workspaceSize：在Device侧申请的workspace大小，由第一段接口aclnnCummaxGetWorkspaceSize获取。
- executor：op执行器，包含了算子计算流程。
- stream：指定执行任务的AscendCL stream流。
返回值：

返回aclnnStatus状态码，具体参见aclnn返回码。

调用示例

#include <iostream>
#include <vector>
#include "acl/acl.h"
#include "aclnnop/aclnn_cummax.h"

#define CHECK_RET(cond, return_expr) \
  do &#123;                               \
    if (!(cond)) &#123;                   \
      return_expr;                   \
    &#125;                                \
  &#125; while (0)

#define LOG_PRINT(message, ...)     \
  do &#123;                              \
    printf(message, ##__VA_ARGS__); \
  &#125; while (0)

int64_t GetShapeSize(const std::vector<int64_t>& shape) &#123;
  int64_t shapeSize = 1;
  for (auto i : shape) &#123;
    shapeSize *= i;
  &#125;
  return shapeSize;
&#125;

int Init(int32_t deviceId, aclrtContext* context, aclrtStream* stream) &#123;
  // 固定写法，AscendCL初始化
  auto ret = aclInit(nullptr);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclInit failed. ERROR: %d\n", ret); return ret);
  ret = aclrtSetDevice(deviceId);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtSetDevice failed. ERROR: %d\n", ret); return ret);
  ret = aclrtCreateContext(context, deviceId);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtCreateContext failed. ERROR: %d\n", ret); return ret);
  ret = aclrtSetCurrentContext(*context);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtSetCurrentContext failed. ERROR: %d\n", ret); return ret);
  ret = aclrtCreateStream(stream);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtCreateStream failed. ERROR: %d\n", ret); return ret);
  return 0;
&#125;

template <typename T>
int CreateAclTensor(const std::vector<T>& hostData, const std::vector<int64_t>& shape, void** deviceAddr,
                    aclDataType dataType, aclTensor** tensor) &#123;
  auto size = GetShapeSize(shape) * sizeof(T);
  // 调用aclrtMalloc申请device侧内存
  auto ret = aclrtMalloc(deviceAddr, size, ACL_MEM_MALLOC_HUGE_FIRST);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtMalloc failed. ERROR: %d\n", ret); return ret);
  // 调用aclrtMemcpy将Host侧数据拷贝到device侧内存上
  ret = aclrtMemcpy(*deviceAddr, size, hostData.data(), size, ACL_MEMCPY_HOST_TO_DEVICE);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtMemcpy failed. ERROR: %d\n", ret); return ret);

  // 计算连续tensor的strides
  std::vector<int64_t> strides(shape.size(), 1);
  for (int64_t i = shape.size() - 2; i >= 0; i--) &#123;
    strides[i] = shape[i + 1] * strides[i + 1];
  &#125;

  // 调用aclCreateTensor接口创建aclTensor
  *tensor = aclCreateTensor(shape.data(), shape.size(), dataType, strides.data(), 0, aclFormat::ACL_FORMAT_ND,
                            shape.data(), shape.size(), *deviceAddr);
  return 0;
&#125;

int main() &#123;
  // 1. （固定写法）device/context/stream初始化，参考AscendCL对外接口列表
  // 根据自己的实际device填写deviceId
  int32_t deviceId = 0;
  aclrtContext context;
  aclrtStream stream;
  auto ret = Init(deviceId, &context, &stream);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("Init acl failed. ERROR: %d\n", ret); return ret);

  // 2. 构造输入与输出，需要根据API的接口自定义构造
  std::vector<int64_t> selfShape = &#123;8&#125;;
  std::vector<int64_t> valuesOutShape = &#123;8&#125;;
  std::vector<int64_t> indicesOutShape = &#123;8&#125;;
  void* selfDeviceAddr = nullptr;
  void* valuesOutDeviceAddr = nullptr;
  void* indicesOutDeviceAddr = nullptr;
  aclTensor* self = nullptr;
  aclTensor* valuesOut = nullptr;
  aclTensor* indicesOut = nullptr;
  std::vector<float> selfHostData = &#123;3.0, 3.0, 2.0, 1.0, 3.0, 2.0, 6.0, 7.0&#125;;
  std::vector<float> valuesOutHostData(8, 0.0);
  std::vector<int64_t> indicesOutHostData(8, 0);
  int64_t dim = 0;
  // 创建self aclTensor
  ret = CreateAclTensor(selfHostData, selfShape, &selfDeviceAddr, aclDataType::ACL_FLOAT, &self);
  CHECK_RET(ret == ACL_SUCCESS, return ret);
  // 创建valuesOut aclTensor
  ret = CreateAclTensor(valuesOutHostData, valuesOutShape, &valuesOutDeviceAddr, aclDataType::ACL_FLOAT, &valuesOut);
  CHECK_RET(ret == ACL_SUCCESS, return ret);
  // 创建indicesOut aclTensor
  ret = CreateAclTensor(indicesOutHostData, indicesOutShape, &indicesOutDeviceAddr, aclDataType::ACL_INT64, &indicesOut);
  CHECK_RET(ret == ACL_SUCCESS, return ret);

  // 3. 调用CANN算子库API，需要修改为具体的API名称
  uint64_t workspaceSize = 0;
  aclOpExecutor* executor;
  // 调用aclnnCummax第一段接口
  ret = aclnnCummaxGetWorkspaceSize(self, dim, valuesOut, indicesOut, &workspaceSize, &executor);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnCummaxGetWorkspaceSize failed. ERROR: %d\n", ret); return ret);
  // 根据第一段接口计算出的workspaceSize申请device内存
  void* workspaceAddr = nullptr;
  if (workspaceSize > 0) &#123;
    ret = aclrtMalloc(&workspaceAddr, workspaceSize, ACL_MEM_MALLOC_HUGE_FIRST);
    CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("allocate workspace failed. ERROR: %d\n", ret); return ret);
  &#125;
  // 调用aclnnCummax第二段接口
  ret = aclnnCummax(workspaceAddr, workspaceSize, executor, stream);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnCummax failed. ERROR: %d\n", ret); return ret);

  // 4. （固定写法）同步等待任务执行结束
  ret = aclrtSynchronizeStream(stream);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtSynchronizeStream failed. ERROR: %d\n", ret); return ret);

  // 5. 获取输出的值，将device侧内存上的结果拷贝至Host侧，需要根据具体API的接口定义修改
  // 获取valuesOut
  auto size = GetShapeSize(valuesOutShape);
  std::vector<float> resultData(size, 0);
  ret = aclrtMemcpy(resultData.data(), resultData.size() * sizeof(resultData[0]), valuesOutDeviceAddr,
                    size * sizeof(resultData[0]), ACL_MEMCPY_DEVICE_TO_HOST);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("copy result from device to host failed. ERROR: %d\n", ret); return ret);
  for (int64_t i = 0; i < size; i++) &#123;
    LOG_PRINT("result[%ld] is: %f\n", i, resultData[i]);
  &#125;
  // 获取indicesOut
  auto indicesSize = GetShapeSize(indicesOutShape);
  std::vector<int64_t> indicesResultData(indicesSize, 0);
  ret = aclrtMemcpy(indicesResultData.data(), indicesResultData.size() * sizeof(indicesResultData[0]), indicesOutDeviceAddr,
                    indicesSize * sizeof(indicesResultData[0]), ACL_MEMCPY_DEVICE_TO_HOST);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("copy result from device to host failed. ERROR: %d\n", ret); return ret);
  for (int64_t i = 0; i < indicesSize; i++) &#123;
    LOG_PRINT("result[%ld] is: %ld\n", i, indicesResultData[i]);
  &#125;

  // 6. 释放aclTensor和aclScalar，需要根据具体API的接口定义修改
  aclDestroyTensor(self);
  aclDestroyTensor(valuesOut);
  aclDestroyTensor(indicesOut);

  // 7. 释放divice 资源，需要根据具体API的接口定义修改
  aclrtFree(selfDeviceAddr);
  aclrtFree(valuesOutDeviceAddr);
  aclrtFree(indicesOutDeviceAddr);
  if (workspaceSize > 0) &#123;
    aclrtFree(workspaceAddr);
  &#125;
  aclrtDestroyStream(stream);
  aclrtDestroyContext(context);
  aclrtResetDevice(deviceId);
  aclFinalize();
  return 0;
&#125;

父主题： NN类算子接口

aclnnCummin

接口原型

**第一段接口：**aclnnStatus aclnnCumminGetWorkspaceSize(const aclTensor *self, int64_t dim, aclTensor *valuesOut, aclTensor* indicesOut, uint64_t *workspaceSize, aclOpExecutor **executor)
**第二段接口：**aclnnStatus aclnnCummin(void *workspace, uint64_t workspaceSize, aclOpExecutor *executor, aclrtStream stream)

功能描述

算子功能：计算张量self中的累积最小值，并返回该值以及对应的索引。

例如valuesOut[1]为self[0]、self[1]之间的最小值，indicesOut[1]为该最小值的索引；valueOut[2]为self[0]、self[1]、self[2]之间的最小值，indicesOut[2]为该最小值索引。
计算公式：

aclnnCumminGetWorkspaceSize

接口定义：

aclnnStatus aclnnCumminGetWorkspaceSize(const aclTensor *self, int64_t dim, aclTensor *valuesOut, aclTensor* indicesOut, uint64_t *workspaceSize, aclOpExecutor **executor)
参数说明：
- self：Device侧的aclTensor，输入张量，数据类型支持FLOAT、FLOAT16、DOUBLE、INT32、INT64、INT16、INT8、UINT8、BOOL，支持非连续的Tensor，数据格式支持ND，且shape需要与valuesOut、indicesOut一致。
- dim：Host侧INT64类型，指定要进行最小值计算的维度，取值范围[-self.dim(), self.dim()）。
- valuesOut：Device侧的aclTensor，输出的最小值，数据类型支持FLOAT、FLOAT16、DOUBLE、INT32、INT64、INT16、INT8、UINT8、BOOL，支持非连续的Tensor，数据格式支持ND，且shape需要与self、indicesOut一致。
- indicesOut：Device侧的aclTensor，输出的最小值索引，数据类型支持INT32、INT64，支持非连续的Tensor，数据格式支持ND，且shape需要与self、valuesOut一致。
- workspaceSize：返回用户需要在Device侧申请的workspace大小。
- executor：返回op执行器，包含了算子计算流程。
返回值：

返回aclnnStatus状态码，具体参见aclnn返回码。

:::note 说明第一段接口完成入参校验，出现以下场景时报错：
- 返回161001（ACLNN_ERR_PARAM_NULLPTR）：传入的self、valuesOut、indicesOut是空指针。
- 返回161002（ACLNN_ERR_PARAM_INVALID）：
  - self、valuesOut、indicesOut的数据类型不在支持的范围之内。
  - self、valuesOut、indicesOut的shape不在支持的范围之内。
  - 当self为0维时，不支持传入dim。
  - 输入的dim值不合法。 :::

aclnnCummin

接口定义：

aclnnStatus aclnnCummin(void *workspace, uint64_t workspaceSize, aclOpExecutor *executor, aclrtStream stream)
参数说明：
- workspace：在Device侧申请的workspace内存起址。
- workspaceSize：在Device侧申请的workspace大小，由第一段接口aclnnCumminGetWorkspaceSize获取。
- executor：op执行器，包含了算子计算流程。
- stream：指定执行任务的AscendCL stream流。
返回值：

返回aclnnStatus状态码，具体参见aclnn返回码。

调用示例

#include <iostream>
#include <vector>
#include "acl/acl.h"
#include "aclnnop/aclnn_cummin.h"

#define CHECK_RET(cond, return_expr) \
  do &#123;                               \
    if (!(cond)) &#123;                   \
      return_expr;                   \
    &#125;                                \
  &#125; while (0)

#define LOG_PRINT(message, ...)     \
  do &#123;                              \
    printf(message, ##__VA_ARGS__); \
  &#125; while (0)

int64_t GetShapeSize(const std::vector<int64_t>& shape) &#123;
  int64_t shape_size = 1;
  for (auto i : shape) &#123;
    shape_size *= i;
  &#125;
  return shape_size;
&#125;

int Init(int32_t deviceId, aclrtContext* context, aclrtStream* stream) &#123;
  // 固定写法，AscendCL初始化
  auto ret = aclInit(nullptr);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclInit failed. ERROR: %d\n", ret); return ret);
  ret = aclrtSetDevice(deviceId);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtSetDevice failed. ERROR: %d\n", ret); return ret);
  ret = aclrtCreateContext(context, deviceId);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtCreateContext failed. ERROR: %d\n", ret); return ret);
  ret = aclrtSetCurrentContext(*context);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtSetCurrentContext failed. ERROR: %d\n", ret); return ret);
  ret = aclrtCreateStream(stream);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtCreateStream failed. ERROR: %d\n", ret); return ret);
  return 0;
&#125;

template <typename T>
int CreateAclTensor(const std::vector<T>& hostData, const std::vector<int64_t>& shape, void** deviceAddr,
                    aclDataType dataType, aclTensor** tensor) &#123;
  auto size = GetShapeSize(shape) * sizeof(T);
  // 调用aclrtMalloc申请device侧内存
  auto ret = aclrtMalloc(deviceAddr, size, ACL_MEM_MALLOC_HUGE_FIRST);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtMalloc failed. ERROR: %d\n", ret); return ret);

  // 调用aclrtMemcpy将Host侧数据拷贝到device侧内存上
  ret = aclrtMemcpy(*deviceAddr, size, hostData.data(), size, ACL_MEMCPY_HOST_TO_DEVICE);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtMemcpy failed. ERROR: %d\n", ret); return ret);

  // 计算连续tensor的strides
  std::vector<int64_t> strides(shape.size(), 1);
  for (int64_t i = shape.size() - 2; i >= 0; i--) &#123;
    strides[i] = shape[i + 1] * strides[i + 1];
  &#125;

  // 调用aclCreateTensor接口创建aclTensor
  *tensor = aclCreateTensor(shape.data(), shape.size(), dataType, strides.data(), 0, aclFormat::ACL_FORMAT_ND,
                            shape.data(), shape.size(), *deviceAddr);
  return 0;
&#125;

int main() &#123;
  // 1. （固定写法）device/context/stream初始化, 参考AscendCL对外接口列表
  // 根据自己的实际device填写deviceId
  int32_t deviceId = 0;
  aclrtContext context;
  aclrtStream stream;
  auto ret = Init(deviceId, &context, &stream);
  // check根据自己的需要处理
  CHECK_RET(ret == 0, LOG_PRINT("Init acl failed. ERROR: %d\n", ret); return ret);

  // 2. 构造输入与输出，需要根据API的接口自定义构造
  std::vector<int64_t> selfShape = &#123;2,4&#125;;
  std::vector<int64_t> valuesOutShape = &#123;2,4&#125;;
  std::vector<int64_t> indicesOutShape = &#123;2,4&#125;;
  void* selfDeviceAddr = nullptr;
  void* valuesOutDeviceAddr = nullptr;
  void* indicesOutDeviceAddr = nullptr;
  aclTensor* self = nullptr;
  aclTensor* valuesOut = nullptr;
  aclTensor* indicesOut = nullptr;
  std::vector<float> selfHostData = &#123;3.0, 3.0, 2.0, 1.0, 4.0, 2.0, 6.0, 7.0&#125;;
  std::vector<float> valuesOutHostData(8, 0.0);
  std::vector<int64_t> indicesOutHostData(8, 0);
  int64_t dim = 0;
  // 创建self aclTensor
  ret = CreateAclTensor(selfHostData, selfShape, &selfDeviceAddr, aclDataType::ACL_FLOAT, &self);
  CHECK_RET(ret == ACL_SUCCESS, return ret);
  // 创建valuesOut aclTensor
  ret = CreateAclTensor(valuesOutHostData, valuesOutShape, &valuesOutDeviceAddr, aclDataType::ACL_FLOAT, &valuesOut);
  // 创建indicesOut aclTensor
  ret = CreateAclTensor(indicesOutHostData, indicesOutShape, &indicesOutDeviceAddr, aclDataType::ACL_INT64, &indicesOut);
  CHECK_RET(ret == ACL_SUCCESS, return ret);

  // 3.调用CANN算子库API，需要修改为具体的算子接口
  uint64_t workspaceSize = 0;
  aclOpExecutor* executor;
  // 调用aclnnCummin第一段接口
  ret = aclnnCumminGetWorkspaceSize(self, dim, valuesOut, indicesOut,&workspaceSize, &executor);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnCumminGetWorkspaceSize failed. ERROR: %d\n", ret); return ret);
  // 根据第一段接口计算出的workspaceSize申请device内存
  void* workspaceAddr = nullptr;
  if (workspaceSize > 0) &#123;
    ret = aclrtMalloc(&workspaceAddr, workspaceSize, ACL_MEM_MALLOC_HUGE_FIRST);
    CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("allocate workspace failed. ERROR: %d\n", ret); return ret;);
  &#125;
  // 调用aclnnCummin第二段接口
  ret = aclnnCummin(workspaceAddr, workspaceSize, executor, stream);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnCummin failed. ERROR: %d\n", ret); return ret);

  // 4. （固定写法）同步等待任务执行结束
  ret = aclrtSynchronizeStream(stream);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtSynchronizeStream failed. ERROR: %d\n", ret); return ret);

  // 5. 获取输出的值，将device侧内存上的结果拷贝至Host侧，需要根据具体API的接口定义修改
  // 获取 valuesOutShape
  auto size = GetShapeSize(valuesOutShape);
  std::vector<float> resultData(size, 0);
  ret = aclrtMemcpy(resultData.data(), resultData.size() * sizeof(resultData[0]), valuesOutDeviceAddr, size * sizeof(resultData[0]),
                    ACL_MEMCPY_DEVICE_TO_HOST);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("copy result from device to host failed. ERROR: %d\n", ret); return ret);
  for (int64_t i = 0; i < size; i++) &#123;
    LOG_PRINT("result[%ld] is: %f\n", i, resultData[i]);
  &#125;
  // 获取 valuesOutShape
  auto indicesSize = GetShapeSize(indicesOutShape);
  std::vector<int64_t> indicesResultData(indicesSize, 0);
  ret = aclrtMemcpy(indicesResultData.data(), indicesResultData.size() * sizeof(indicesResultData[0]), indicesOutDeviceAddr,
                    indicesSize * sizeof(indicesResultData[0]), ACL_MEMCPY_DEVICE_TO_HOST);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("copy result from device to host failed. ERROR: %d\n", ret); return ret);
  for (int64_t i = 0; i < indicesSize; i++) &#123;
    LOG_PRINT("result[%ld] is: %ld\n", i, indicesResultData[i]);
  &#125;

  // 6. 释放aclTensor和aclScalar，需要根据具体API的接口定义修改
  aclDestroyTensor(self);
  aclDestroyTensor(valuesOut);
  aclDestroyTensor(indicesOut);

  // 7. 释放device资源，需要根据具体API的接口定义修改
  aclrtFree(selfDeviceAddr);
  aclrtFree(valuesOutDeviceAddr);
  aclrtFree(indicesOutDeviceAddr);
  if (workspaceSize > 0) &#123;
    aclrtFree(workspaceAddr);
  &#125;
  aclrtDestroyStream(stream);
  aclrtDestroyContext(context);
  aclrtResetDevice(deviceId);
  aclFinalize();
  return 0;
&#125;

父主题： NN类算子接口

aclnnCumsum

接口原型

第一段接口：aclnnStatus aclnnCumsumGetWorkspaceSize(const aclTensor *self, int64_t dim, aclDataType dtype, aclTensor *out, uint64_t *workspaceSize, aclOpExecutor **executor)
第二段接口：aclnnStatus aclnnCumsum(void *workspace, uint64_t workspaceSize, aclOpExecutor *executor, const aclrtStream stream)

功能描述

算子功能：对输入张量self的元素，按照指定维度dim依次进行累加，并将结果保存到输出张量out中。
计算公式：xi是输入张量self中，从维度dim视角来看的某个元素（其它维度下标不变，只dim维度下标依次递增），yi是输出张量out中对应位置的元素，则

aclnnCumsumGetWorkspaceSize

接口定义：

aclnnStatus aclnnCumsumGetWorkspaceSize(const aclTensor *self, int64_t dim, aclDataType dtype, aclTensor *out, uint64_t *workspaceSize, aclOpExecutor **executor)
参数说明：
- self：Device侧的aclTensor，数据类型支持FLOAT、DOUBLE、COMPLEX64、COMPLEX128、UINT8、INT8、INT16、INT32、INT64、FLOAT16、BFLOAT16（仅Atlas A2训练系列产品支持）、BOOL，数据类型需要能转换成out的数据类型。支持非连续的Tensor，数据格式支持ND，数据维度不支持8维以上。
- dim：Host侧的整数，数据类型支持INT64。取值范围是[-self.dim(), self.dim()-1]。
- dtype：Host侧的数据类型枚举，表示输出张量所需的数据类型，支持FLOAT、FLOAT16、INT32、DOUBLE、UINT8、INT8、INT16、INT64、COMPLEX64、COMPLEX128，且需与输出张量out的数据类型一致。
- out：Device侧的aclTensor，数据类型支持FLOAT、FLOAT16、INT32、DOUBLE、UINT8、INT8、INT16、INT64、COMPLEX64、COMPLEX128，且数据类型需要与传入的dtype一致，shape需要与self一致。支持非连续的Tensor，数据格式支持ND，数据维度不支持8维以上。
- workspaceSize：返回用户需要在Device侧申请的workspace大小。
- executor：返回op执行器，包含了算子计算流程。
返回值：

返回aclnnStatus状态码，具体参见aclnn返回码。

:::note 说明第一段接口完成入参校验，出现以下场景时报错：
- 返回161001（ACLNN_ERR_PARAM_NULLPTR）：传入的self或out是空指针。
- 返回161002（ACLNN_ERR_PARAM_INVALID）：
  - self、out的数据类型或数据格式不在支持的范围内。
  - self的数据类型不能转换为out的数据类型。
  - 当self.dim()==0时，dim的取值范围不在[-1, 0]内；当self.dim()>0时，dim的取值范围不在[-self.dim(), self.dim()-1]内。
  - 参数dtype和out的数据类型不一致。
  - self和out的shape不一致。
  - self、out的维度超过8维。 :::

aclnnCumsum

接口定义：

aclnnStatus aclnnCumsum(void *workspace, uint64_t workspaceSize, aclOpExecutor *executor, const aclrtStream stream)
参数说明：
- workspace（void *）：在Device侧申请的workspace内存起址。
- workspaceSize（uint64_t）：在Device侧申请的workspace大小，由第一段接口aclnnCumsumGetWorkspaceSize获取。
- executor（aclOpExecutor *）：op执行器，包含了算子计算流程。
- stream（const aclrtStream）：指定执行任务的AscendCL stream流。
返回值：

返回aclnnStatus状态码，具体参见aclnn返回码。

调用示例

#include <iostream>
#include <vector>
#include "acl/acl.h"
#include "aclnnop/aclnn_cumsum.h"

#define CHECK_RET(cond, return_expr) \
  do &#123;                               \
    if (!(cond)) &#123;                   \
      return_expr;                   \
    &#125;                                \
  &#125; while (0)

#define LOG_PRINT(message, ...)     \
  do &#123;                              \
    printf(message, ##__VA_ARGS__); \
  &#125; while (0)

int64_t GetShapeSize(const std::vector<int64_t>& shape) &#123;
  int64_t shape_size = 1;
  for (auto i : shape) &#123;
    shape_size *= i;
  &#125;
  return shape_size;
&#125;

int Init(int32_t deviceId, aclrtContext* context, aclrtStream* stream) &#123;
  // 固定写法，AscendCL初始化
  auto ret = aclInit(nullptr);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclInit failed. ERROR: %d\n", ret); return ret);
  ret = aclrtSetDevice(deviceId);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtSetDevice failed. ERROR: %d\n", ret); return ret);
  ret = aclrtCreateContext(context, deviceId);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtCreateContext failed. ERROR: %d\n", ret); return ret);
  ret = aclrtSetCurrentContext(*context);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtSetCurrentContext failed. ERROR: %d\n", ret); return ret);
  ret = aclrtCreateStream(stream);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtCreateStream failed. ERROR: %d\n", ret); return ret);
  return 0;
&#125;

template <typename T>
int CreateAclTensor(const std::vector<T>& hostData, const std::vector<int64_t>& shape, void** deviceAddr,
                    aclDataType dataType, aclTensor** tensor) &#123;
  auto size = GetShapeSize(shape) * sizeof(T);
  // 调用aclrtMalloc申请device侧内存
  auto ret = aclrtMalloc(deviceAddr, size, ACL_MEM_MALLOC_HUGE_FIRST);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtMalloc failed. ERROR: %d\n", ret); return ret);

  // 调用aclrtMemcpy将Host侧数据拷贝到device侧内存上
  ret = aclrtMemcpy(*deviceAddr, size, hostData.data(), size, ACL_MEMCPY_HOST_TO_DEVICE);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtMemcpy failed. ERROR: %d\n", ret); return ret);

  // 计算连续tensor的strides
  std::vector<int64_t> strides(shape.size(), 1);
  for (int64_t i = shape.size() - 2; i >= 0; i--) &#123;
    strides[i] = shape[i + 1] * strides[i + 1];
  &#125;

  // 调用aclCreateTensor接口创建aclTensor
  *tensor = aclCreateTensor(shape.data(), shape.size(), dataType, strides.data(), 0, aclFormat::ACL_FORMAT_ND,
                            shape.data(), shape.size(), *deviceAddr);
  return 0;
&#125;

int main() &#123;
  // 1. （固定写法）device/context/stream初始化, 参考AscendCL对外接口列表
  // 根据自己的实际device填写deviceId
  int32_t deviceId = 0;
  aclrtContext context;
  aclrtStream stream;
  auto ret = Init(deviceId, &context, &stream);
  // check根据自己的需要处理
  CHECK_RET(ret == 0, LOG_PRINT("Init acl failed. ERROR: %d\n", ret); return ret);

  // 2. 构造输入与输出，需要根据API的接口自定义构造
  std::vector<int64_t> selfShape = &#123;2, 2&#125;;
  std::vector<int64_t> outShape = &#123;2, 2&#125;;
  void* selfDeviceAddr = nullptr;
  void* outDeviceAddr = nullptr;
  aclTensor* self = nullptr;
  aclTensor* out = nullptr;
  std::vector<float> selfHostData = &#123;0, 1, 2, 3&#125;;
  std::vector<float> outHostData = &#123;0, 0, 0, 0&#125;;
  int64_t dim = 0;
  aclDataType dtype = aclDataType::ACL_FLOAT;
  // 创建self aclTensor
  ret = CreateAclTensor(selfHostData, selfShape, &selfDeviceAddr, aclDataType::ACL_FLOAT, &self);
  CHECK_RET(ret == ACL_SUCCESS, return ret);
  // 创建out aclTensor
  ret = CreateAclTensor(outHostData, outShape, &outDeviceAddr, aclDataType::ACL_FLOAT, &out);
  CHECK_RET(ret == ACL_SUCCESS, return ret);

  // 3.调用CANN算子库API，需要修改为具体的算子接口
  uint64_t workspaceSize = 0;
  aclOpExecutor* executor;
  // 调用aclnnCumsum第一段接口
  ret = aclnnCumsumGetWorkspaceSize(self, dim, dtype, out, &workspaceSize, &executor);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnCumsumGetWorkspaceSize failed. ERROR: %d\n", ret); return ret);
  // 根据第一段接口计算出的workspaceSize申请device内存
  void* workspaceAddr = nullptr;
  if (workspaceSize > 0) &#123;
    ret = aclrtMalloc(&workspaceAddr, workspaceSize, ACL_MEM_MALLOC_HUGE_FIRST);
    CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("allocate workspace failed. ERROR: %d\n", ret); return ret;);
  &#125;
  // 调用aclnnCumsum第二段接口
  ret = aclnnCumsum(workspaceAddr, workspaceSize, executor, stream);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnCumsum failed. ERROR: %d\n", ret); return ret);

  // 4. （固定写法）同步等待任务执行结束
  ret = aclrtSynchronizeStream(stream);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtSynchronizeStream failed. ERROR: %d\n", ret); return ret);

  // 5. 获取输出的值，将device侧内存上的结果拷贝至Host侧，需要根据具体API的接口定义修改
  auto size = GetShapeSize(outShape);
  std::vector<float> resultData(size, 0);
  ret = aclrtMemcpy(resultData.data(), resultData.size() * sizeof(resultData[0]), outDeviceAddr, size * sizeof(float),
                    ACL_MEMCPY_DEVICE_TO_HOST);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("copy result from device to host failed. ERROR: %d\n", ret); return ret);
  for (int64_t i = 0; i < size; i++) &#123;
    LOG_PRINT("result[%ld] is: %f\n", i, resultData[i]);
  &#125;

  // 6. 释放aclTensor和aclScalar，需要根据具体API的接口定义修改
  aclDestroyTensor(self);
  aclDestroyTensor(out);
  return 0;
&#125;

父主题： NN类算子接口

aclnnDiag

接口原型

**第一段接口：**aclnnStatus aclnnDiagGetWorkspaceSize(const aclTensor* self, int64_t diagonal, aclTensor* out, uint64_t* workspaceSize, aclOpExecutor** executor)
**第二段接口：**aclnnStatus aclnnDiag(void* workspace, uint64_t workspaceSize, aclOpExecutor* executor, aclrtStream stream)

功能描述

算子功能：如果输入张量self是一维向量，则返回二维矩阵张量，其中self元素为对角线。如果输入张量self是二维张量，则输出一维向量，取值为diagonal指定的输入矩阵的对角线元素。

aclnnDiagGetWorkspaceSize

接口定义：

aclnnStatus aclnnDiagGetWorkspaceSize(const aclTensor* self, int64_t diagonal, aclTensor* out, uint64_t* workspaceSize, aclOpExecutor** executor)
参数说明：
- self(aclTensor*, 计算输入)：输入张量，数据类型支持FLOAT、FLOAT16、DOUBLE、INT32、INT64、INT16、INT8、UINT8、BOOL、BFLOAT16（仅Atlas A2训练系列产品支持）、COMPLEX32、COMPLEX64、COMPLEX128。支持非连续的Tensor，数据格式支持ND。
- diagonal(int64_t，计算输入)：对角线输入，默认值为0，数据类型支持INT64。
- out(aclTensor*, 计算输出)：输出张量，数据类型支持FLOAT、FLOAT16、DOUBLE、INT32、INT64、INT16、INT8、UINT8、BOOL、BFLOAT16（仅Atlas A2训练系列产品支持）、COMPLEX32、COMPLEX64、COMPLEX128。支持非连续的Tensor，数据格式支持ND。
- workspaceSize(uint64_t*, 出参)：返回用户需要在Device侧申请的workspace大小。
- executor(aclOpExecutor**, 出参)：返回op执行器，包含了算子计算流程。
返回值：

返回aclnnStatus状态码，具体参见aclnn返回码。

:::note 说明第一段接口完成入参校验，出现以下场景时报错：
- 返回161001（ACLNN_ERR_PARAM_NULLPTR）：传入的self或out是空指针。
- 返回161002（ACLNN_ERR_PARAM_INVALID）：
  - self和out的数据类型和数据格式不在支持的范围内。
  - diagonal不在支持的数据类型范围内。 :::

aclnnDiag

接口定义：

aclnnStatus aclnnDiag(void* workspace, uint64_t workspaceSize, aclOpExecutor* executor, aclrtStream stream)
参数说明：
- workspace(void*, 入参)：在Device侧申请的workspace内存起址。
- workspaceSize(uint64_t, 入参)：在Device侧申请的workspace大小，由第一段接口aclnnDiagGetWorkspaceSize获取。
- executor(aclOpExecutor*, 入参)：op执行器，包含了算子计算流程。
- stream(aclrtStream, 入参)：指定执行任务的AscendCL stream流。
返回值：

返回aclnnStatus状态码，具体参见aclnn返回码。

调用示例

#include <iostream>
#include <vector>
#include "acl/acl.h"
#include "aclnnop/aclnn_diag.h"

#define CHECK_RET(cond, return_expr) \
  do &#123;                               \
    if (!(cond)) &#123;                   \
      return_expr;                   \
    &#125;                                \
  &#125; while (0)

#define LOG_PRINT(message, ...)     \
  do &#123;                              \
    printf(message, ##__VA_ARGS__); \
  &#125; while (0)

int64_t GetShapeSize(const std::vector<int64_t>& shape) &#123;
  int64_t shapeSize = 1;
  for (auto i : shape) &#123;
    shapeSize *= i;
  &#125;
  return shapeSize;
&#125;

int Init(int32_t deviceId, aclrtContext* context, aclrtStream* stream) &#123;
  // 固定写法，AscendCL初始化
  auto ret = aclInit(nullptr);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclInit failed. ERROR: %d\n", ret); return ret);
  ret = aclrtSetDevice(deviceId);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtSetDevice failed. ERROR: %d\n", ret); return ret);
  ret = aclrtCreateContext(context, deviceId);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtCreateContext failed. ERROR: %d\n", ret); return ret);
  ret = aclrtSetCurrentContext(*context);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtSetCurrentContext failed. ERROR: %d\n", ret); return ret);
  ret = aclrtCreateStream(stream);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtCreateStream failed. ERROR: %d\n", ret); return ret);
  return 0;
&#125;

template <typename T>
int CreateAclTensor(const std::vector<T>& hostData, const std::vector<int64_t>& shape, void** deviceAddr,
                    aclDataType dataType, aclTensor** tensor) &#123;
  auto size = GetShapeSize(shape) * sizeof(T);
  // 调用aclrtMalloc申请device侧内存
  auto ret = aclrtMalloc(deviceAddr, size, ACL_MEM_MALLOC_HUGE_FIRST);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtMalloc failed. ERROR: %d\n", ret); return ret);
  // 调用aclrtMemcpy将host侧数据拷贝到device侧内存上
  ret = aclrtMemcpy(*deviceAddr, size, hostData.data(), size, ACL_MEMCPY_HOST_TO_DEVICE);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtMemcpy failed. ERROR: %d\n", ret); return ret);

  // 计算连续tensor的strides
  std::vector<int64_t> strides(shape.size(), 1);
  for (int64_t i = shape.size() - 2; i >= 0; i--) &#123;
    strides[i] = shape[i + 1] * strides[i + 1];
  &#125;

  // 调用aclCreateTensor接口创建aclTensor
  *tensor = aclCreateTensor(shape.data(), shape.size(), dataType, strides.data(), 0, aclFormat::ACL_FORMAT_ND,
                            shape.data(), shape.size(), *deviceAddr);
  return 0;
&#125;

int main() &#123;
  // 1. （固定写法）device/context/stream初始化，参考AscendCL对外接口列表
  // 根据自己的实际device填写deviceId
  int32_t deviceId = 0;
  aclrtContext context;
  aclrtStream stream;
  auto ret = Init(deviceId, &context, &stream);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("Init acl failed. ERROR: %d\n", ret); return ret);

  // 2. 构造输入与输出，需要根据API的接口自定义构造
  std::vector<int64_t> selfShape = &#123;1, 2&#125;;
  std::vector<int64_t> outShape = &#123;3, 3&#125;;
  void* selfDeviceAddr = nullptr;
  void* outDeviceAddr = nullptr;
  aclTensor* self = nullptr;
  aclTensor* out = nullptr;
  std::vector<float> selfHostData = &#123;2, 5&#125;;
  std::vector<float> outHostData(0, 0, 0, 0, 0, 0, 0, 0, 0);
  int64_t diagonalVal = 1;

  // 创建self aclTensor
  ret = CreateAclTensor(selfHostData, selfShape, &selfDeviceAddr, aclDataType::ACL_FLOAT, &self);
  CHECK_RET(ret == ACL_SUCCESS, return ret);
  // 创建out aclTensor
  ret = CreateAclTensor(outHostData, outShape, &outDeviceAddr, aclDataType::ACL_FLOAT, &out);
  CHECK_RET(ret == ACL_SUCCESS, return ret);

  // 3. 调用CANN算子库API，需要修改为具体的API名称
  uint64_t workspaceSize = 0;
  aclOpExecutor* executor;
  // 调用aclnnDiag第一段接口
  ret = aclnnDiagGetWorkspaceSize(self, diagonalVal, out, &workspaceSize, &executor);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnDiagGetWorkspaceSize failed. ERROR: %d\n", ret); return ret);
  // 根据第一段接口计算出的workspaceSize申请device内存
  void* workspaceAddr = nullptr;
  if (workspaceSize > 0) &#123;
    ret = aclrtMalloc(&workspaceAddr, workspaceSize, ACL_MEM_MALLOC_HUGE_FIRST);
    CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("allocate workspace failed. ERROR: %d\n", ret); return ret);
  &#125;
  // 调用aclnnDiag第二段接口
  ret = aclnnDiag(workspaceAddr, workspaceSize, executor, stream);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnDiag failed. ERROR: %d\n", ret); return ret);

  // 4. （固定写法）同步等待任务执行结束
  ret = aclrtSynchronizeStream(stream);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtSynchronizeStream failed. ERROR: %d\n", ret); return ret);

  // 5. 获取输出的值，将device侧内存上的结果拷贝至host侧，需要根据具体API的接口定义修改
  auto size = GetShapeSize(outShape);
  std::vector<float> resultData(size, 0);
  ret = aclrtMemcpy(resultData.data(), resultData.size() * sizeof(resultData[0]), outDeviceAddr,
                    size * sizeof(resultData[0]), ACL_MEMCPY_DEVICE_TO_HOST);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("copy result from device to host failed. ERROR: %d\n", ret); return ret);
  for (int64_t i = 0; i < size; i++) &#123;
    LOG_PRINT("result[%ld] is: %f\n", i, resultData[i]);
  &#125;

  // 6. 释放aclTensor，需要根据具体API的接口定义修改
  aclDestroyTensor(self);
  aclDestroyTensor(out);
  return 0;
&#125;

父主题： NN类算子接口

aclnnDiv/aclnnInplaceDiv

接口原型

:::note 说明

aclnnDiv和aclnnInplaceDiv实现相同的功能，其使用区别如下，请根据自身实际场景选择合适的算子。
- aclnnDiv：需新建一个输出张量对象存储计算结果。
- aclnnInplaceDiv：无需新建输出张量对象，直接在输入张量的内存中存储计算结果。
每个算子分为两段接口，必须先调用“aclnnXxxGetWorkspaceSize”接口获取入参并根据计算流程计算所需workspace大小，再调用“aclnnXxx”接口执行计算。 :::
aclnnDiv两段式接口如下：
- **第一段接口：**aclnnStatus aclnnDivGetWorkspaceSize(const aclTensor *self, const aclTensor *other, const aclTensor *out, uint64_t *workspaceSize, aclOpExecutor **executor)
- **第二段接口：**aclnnStatus aclnnDiv(void *workspace, uint64_t workspaceSize, aclOpExecutor *executor, const aclrtStream stream)
aclnnInplaceDiv两段式接口如下：
- **第一段接口：**aclnnStatus aclnnInplaceDivGetWorkspaceSize(aclTensor* selfRef, const aclTensor* other, uint64_t* workspaceSize, aclOpExecutor** executor)
- **第二段接口：**aclnnStatus aclnnInplaceDiv(void* workspace, uint64_t workspaceSize, aclOpExecutor* executor, aclrtStream stream)

功能描述

算子功能：完成张量self与张量other的除法计算。
计算公式：

aclnnDivGetWorkspaceSize

接口定义：

aclnnStatus aclnnDivGetWorkspaceSize(const aclTensor *self, const aclTensor *other, const aclTensor *out, uint64_t *workspaceSize, aclOpExecutor **executor)
参数说明：
- self：Device侧的aclTensor，被除数。数据类型支持FLOAT、FLOAT16、DOUBLE、INT32、INT64、INT16、INT8、UINT8、BOOL、BFLOAT16（仅Atlas A2训练系列产品支持），且数据类型需要与other构成互推导关系，shape需要与other满足broadcast关系。支持非连续的Tensor，数据格式支持ND。
- other：Device侧的aclTensor，与self进行除法运算的一维Tensor。数据类型支持FLOAT、FLOAT16、DOUBLE、INT32、INT64、INT16、INT8、UINT8、BOOL，且数据类型需要与self构成互推导关系，shape需要与self满足broadcast关系。支持非连续的Tensor，数据格式支持ND。
- out：Device侧的aclTensor，除法计算的结果。数据类型支持FLOAT、FLOAT16、DOUBLE、BFLOAT16（仅Atlas A2训练系列产品支持），且数据类型需要是self与other推导之后可转换的数据类型，shape需要是self与other broadcast之后的shape。支持非连续的Tensor，数据格式支持ND。
- workspaceSize：返回用户需要在Device侧申请的workspace大小。
- executor：返回op执行器，包含了算子计算流程。
返回值：

返回aclnnStatus状态码，具体参见aclnn返回码。

:::note 说明第一段接口完成入参校验，出现以下场景时报错：
- 返回161001（ACLNN_ERR_PARAM_NULLPTR）：传入的self、other或out是空指针。
- 返回161002（ACLNN_ERR_PARAM_INVALID）：
  - self和other的数据类型和数据格式不在支持的范围内。
  - self和other不满足数据类型推导规则。
  - 推导出的数据类型无法转换为指定输出out的类型。
  - self和other的shape无法做broadcast。
  - self或other的维度大于8。 :::

aclnnDiv

接口定义：

aclnnStatus aclnnDiv(void *workspace, uint64_t workspaceSize, aclOpExecutor *executor, const aclrtStream stream)
参数说明：
- workspace：在Device侧申请的workspace内存起址。
- workspaceSize：在Device侧申请的workspace大小，由第一段接口aclnnDivGetWorkspaceSize获取。
- executor：op执行器，包含了算子计算流程。
- stream：指定执行任务的AscendCL stream流。
返回值：

返回aclnnStatus状态码，具体参见aclnn返回码。

aclnnInplaceDivGetWorkspaceSize

接口定义：

aclnnStatus aclnnInplaceDivGetWorkspaceSize(aclTensor* selfRef, const aclTensor* other, uint64_t* workspaceSize, aclOpExecutor** executor)
参数说明：
- selfRef：Device侧的aclTensor，被除数。数据类型支持FLOAT、FLOAT16、DOUBLE、INT32、INT64、INT16、INT8、UINT8、BFLOAT16（仅Atlas A2训练系列产品支持），且数据类型需要与other构成互推导关系，shape需要与other满足broadcast关系。支持非连续的Tensor，数据格式支持ND。
- other：Device侧的aclTensor，与self进行除法运算的一维Tensor。数据类型支持FLOAT、FLOAT16、DOUBLE、INT32、INT64、INT16、INT8、UINT8、BOOL、BFLOAT16（仅Atlas A2训练系列产品支持），且数据类型需要与selfRef构成互推导关系，shape需要与selfRef满足broadcast关系。支持非连续的Tensor，数据格式支持ND。
- workspaceSize：返回用户需要在Device侧申请的workspace大小。
- executor：返回op执行器，包含了算子计算流程。
返回值：

返回aclnnStatus状态码，具体参见aclnn返回码。

:::note 说明第一段接口完成入参校验，出现以下场景时报错：
- 返回161001（ACLNN_ERR_PARAM_NULLPTR）：传入的selfRef、other是空指针。
- 返回161002（ACLNN_ERR_PARAM_INVALID）：
  - selfRef和other的数据类型和数据格式不在支持的范围内。
  - selfRef和other不满足数据类型推导规则。
  - selfRef与other的shape无法做broadcast。
  - selfRef或other的维度大于8。 :::

aclnnInplaceDiv

接口定义：

aclnnStatus aclnnInplaceDiv(void* workspace, uint64_t workspaceSize, aclOpExecutor* executor, aclrtStream stream)
参数说明：
- workspace：在Device侧申请的workspace内存起址。
- workspaceSize：在Device侧申请的workspace大小，由第一段接口aclnnInplaceDivGetWorkspaceSize获取。
- executor：op执行器，包含了算子计算流程。
- stream：指定执行任务的AscendCL stream流。
返回值：

返回aclnnStatus状态码，具体参见aclnn返回码。

调用示例

#include <vector>
#include "acl/acl.h"
#include "aclnnop/aclnn_div.h"

#define CHECK_RET(cond, return_expr) \
  do &#123;                               \
    if (!(cond)) &#123;                   \
      return_expr;                   \
    &#125;                                \
  &#125; while (0)

#define LOG_PRINT(message, ...)     \
  do &#123;                              \
    printf(message, ##__VA_ARGS__); \
  &#125; while (0)

int64_t GetShapeSize(const std::vector<int64_t>& shape) &#123;
  int64_t shape_size = 1;
  for (auto i : shape) &#123;
    shape_size *= i;
  &#125;
  return shape_size;
&#125;

int Init(int32_t deviceId, aclrtContext* context, aclrtStream* stream) &#123;
  // 固定写法，AscendCL初始化
  auto ret = aclInit(nullptr);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclInit failed. ERROR: %d\n", ret); return ret);
  ret = aclrtSetDevice(deviceId);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtSetDevice failed. ERROR: %d\n", ret); return ret);
  ret = aclrtCreateContext(context, deviceId);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtCreateContext failed. ERROR: %d\n", ret); return ret);
  ret = aclrtSetCurrentContext(*context);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtSetCurrentContext failed. ERROR: %d\n", ret); return ret);
  ret = aclrtCreateStream(stream);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtCreateStream failed. ERROR: %d\n", ret); return ret);
  return 0;
&#125;

template <typename T>
int CreateAclTensor(const std::vector<T>& hostData, const std::vector<int64_t>& shape, void** deviceAddr,
                    aclDataType dataType, aclTensor** tensor) &#123;
  auto size = GetShapeSize(shape) * sizeof(T);
  // 调用aclrtMalloc申请device侧内存
  auto ret = aclrtMalloc(deviceAddr, size, ACL_MEM_MALLOC_HUGE_FIRST);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtMalloc failed. ERROR: %d\n", ret); return ret);

  // 调用aclrtMemcpy将Host侧数据拷贝到device侧内存上
  ret = aclrtMemcpy(*deviceAddr, size, hostData.data(), size, ACL_MEMCPY_HOST_TO_DEVICE);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtMemcpy failed. ERROR: %d\n", ret); return ret);

  // 计算连续tensor的strides
  std::vector<int64_t> strides(shape.size(), 1);
  for (int64_t i = shape.size() - 2; i >= 0; i--) &#123;
    strides[i] = shape[i + 1] * strides[i + 1];
  &#125;

  // 调用aclCreateTensor接口创建aclTensor
  *tensor = aclCreateTensor(shape.data(), shape.size(), dataType, strides.data(), 0, aclFormat::ACL_FORMAT_ND,
                            shape.data(), shape.size(), *deviceAddr);
  return 0;
&#125;

int main() &#123;
  // 1. （固定写法）device/context/stream初始化, 参考AscendCL对外接口列表
  // 根据自己的实际device填写deviceId
  int32_t deviceId = 0;
  aclrtContext context;
  aclrtStream stream;
  auto ret = Init(deviceId, &context, &stream);
  // check根据自己的需要处理
  CHECK_RET(ret == 0, LOG_PRINT("Init acl failed. ERROR: %d\n", ret); return ret);

  // 2. 构造输入与输出，需要根据API的接口自定义构造
  std::vector<int64_t> selfShape = &#123;4, 2&#125;;
  std::vector<int64_t> otherShape = &#123;4, 2&#125;;
  std::vector<int64_t> outShape = &#123;4, 2&#125;;
  void* selfDeviceAddr = nullptr;
  void* otherDeviceAddr = nullptr;
  void* outDeviceAddr = nullptr;
  aclTensor* self = nullptr;
  aclTensor* other = nullptr;
  aclTensor* out = nullptr;
  std::vector<float> selfHostData = &#123;0, 1, 2, 3, 4, 5, 6, 7&#125;;
  std::vector<float> otherHostData = &#123;1, 1, 1, 2, 2, 2, 3, 3&#125;;
  std::vector<float> outHostData = &#123;0, 0, 0, 0, 0, 0, 0, 0&#125;;
  // 创建self aclTensor
  ret = CreateAclTensor(selfHostData, selfShape, &selfDeviceAddr, aclDataType::ACL_FLOAT, &self);
  CHECK_RET(ret == ACL_SUCCESS, return ret);
  // 创建other aclTensor
  ret = CreateAclTensor(otherHostData, otherShape, &otherDeviceAddr, aclDataType::ACL_FLOAT, &other);
  CHECK_RET(ret == ACL_SUCCESS, return ret);
  // 创建out aclTensor
  ret = CreateAclTensor(outHostData, outShape, &outDeviceAddr, aclDataType::ACL_FLOAT, &out);
  CHECK_RET(ret == ACL_SUCCESS, return ret);

  // 3.调用CANN算子库API，需要修改为具体的算子接口
  uint64_t workspaceSize = 0;
  aclOpExecutor* executor;
  // 调用aclnnDiv第一段接口
  ret = aclnnDivGetWorkspaceSize(self, other, out, &workspaceSize, &executor);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnDivGetWorkspaceSize failed. ERROR: %d\n", ret); return ret);
  // 根据第一段接口计算出的workspaceSize申请device内存
  void* workspaceAddr = nullptr;
  if (workspaceSize > 0) &#123;
    ret = aclrtMalloc(&workspaceAddr, workspaceSize, ACL_MEM_MALLOC_HUGE_FIRST);
    CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("allocate workspace failed. ERROR: %d\n", ret); return ret;);
  &#125;
  // 调用aclnnDiv第二段接口
  ret = aclnnDiv(workspaceAddr, workspaceSize, executor, stream);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnDiv failed. ERROR: %d\n", ret); return ret);
  uint64_t inplaceWorkspaceSize = 0;
  aclOpExecutor* inplaceExecutor;
  // 调用aclnnInplaceDiv第一段接口
  ret = aclnnInplaceDivGetWorkspaceSize(self, other, &inplaceWorkspaceSize, &inplaceExecutor);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnInplaceDivGetWorkspaceSize failed. ERROR: %d\n", ret); return ret);
  // 根据第一段接口计算出的workspaceSize申请device内存
  void* inplaceWorkspaceAddr = nullptr;
  if (inplaceWorkspaceSize > 0) &#123;
    ret = aclrtMalloc(&inplaceWorkspaceAddr, inplaceWorkspaceSize, ACL_MEM_MALLOC_HUGE_FIRST);
    CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("allocate workspace failed. ERROR: %d\n", ret); return ret;);
  &#125;
  // 调用aclnnInplaceDiv第二段接口
  ret = aclnnInplaceDiv(inplaceWorkspaceAddr, inplaceWorkspaceSize, inplaceExecutor, stream);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnInplaceDiv failed. ERROR: %d\n", ret); return ret);

  // 4. （固定写法）同步等待任务执行结束
  ret = aclrtSynchronizeStream(stream);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtSynchronizeStream failed. ERROR: %d\n", ret); return ret);

  // 5. 获取输出的值，将device侧内存上的结果拷贝至Host侧，需要根据具体API的接口定义修改
  auto size = GetShapeSize(outShape);
  std::vector<float> resultData(size, 0);
  ret = aclrtMemcpy(resultData.data(), resultData.size() * sizeof(resultData[0]), outDeviceAddr, size * sizeof(float),
                    ACL_MEMCPY_DEVICE_TO_HOST);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("copy result from device to host failed. ERROR: %d\n", ret); return ret);
  for (int64_t i = 0; i < size; i++) &#123;
    LOG_PRINT("result[%ld] is: %f\n", i, resultData[i]);
  &#125;

  auto inplaceSize = GetShapeSize(selfShape);
  std::vector<float> inplaceResultData(inplaceSize, 0);
  ret = aclrtMemcpy(inplaceResultData.data(), inplaceResultData.size() * sizeof(inplaceResultData[0]), selfDeviceAddr,
                    inplaceSize * sizeof(float), ACL_MEMCPY_DEVICE_TO_HOST);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("copy result from device to host failed. ERROR: %d\n", ret); return ret);
  for (int64_t i = 0; i < inplaceSize; i++) &#123;
    LOG_PRINT("inplaceResult[%ld] is: %f\n", i, inplaceResultData[i]);
  &#125;

  // 6. 释放aclTensor和aclScalar，需要根据具体API的接口定义修改
  aclDestroyTensor(self);
  aclDestroyTensor(other);
  aclDestroyTensor(out);
  return 0;
&#125;

父主题： NN类算子接口

在线提单

aclnnCtcloss​

接口原型​

功能描述​

aclnnCtcLossGetWorkspaceSize​

aclnnCtcLoss​

调用示例​

aclnnCtcLossBackward​

接口原型​

功能描述​

aclnnCtcLossBackwardGetWorkspaceSize​

aclnnCtcLossBackward​

调用示例​

aclnnCummax​

接口原型​

功能描述​

aclnnCummaxGetWorkspaceSize​

aclnnCummax​

调用示例​

aclnnCummin​

接口原型​

功能描述​

aclnnCumminGetWorkspaceSize​

aclnnCummin​

调用示例​

aclnnCumsum​

接口原型​

功能描述​

aclnnCumsumGetWorkspaceSize​

aclnnCumsum​

调用示例​

aclnnDiag​

接口原型​

功能描述​

aclnnDiagGetWorkspaceSize​

aclnnDiag​

调用示例​

aclnnDiv/aclnnInplaceDiv​

接口原型​

功能描述​

aclnnDivGetWorkspaceSize​

aclnnDiv​

aclnnInplaceDivGetWorkspaceSize​

aclnnInplaceDiv​

调用示例​

aclnnCtcloss

接口原型

功能描述

aclnnCtcLossGetWorkspaceSize

aclnnCtcLoss

调用示例

aclnnCtcLossBackward

接口原型

功能描述

aclnnCtcLossBackwardGetWorkspaceSize

aclnnCtcLossBackward

调用示例

aclnnCummax

接口原型

功能描述

aclnnCummaxGetWorkspaceSize

aclnnCummax

调用示例

aclnnCummin

接口原型

功能描述

aclnnCumminGetWorkspaceSize

aclnnCummin

调用示例

aclnnCumsum

接口原型

功能描述

aclnnCumsumGetWorkspaceSize

aclnnCumsum

调用示例

aclnnDiag

接口原型

功能描述

aclnnDiagGetWorkspaceSize

aclnnDiag

调用示例

aclnnDiv/aclnnInplaceDiv

接口原型

功能描述

aclnnDivGetWorkspaceSize

aclnnDiv

aclnnInplaceDivGetWorkspaceSize

aclnnInplaceDiv

调用示例