aclnnCtcLoss

支持的产品型号

• Atlas 推理系列产品。
• Atlas 训练系列产品。
• Atlas A2训练系列产品/Atlas 800I A2推理产品。

接口原型

每个算子分为两段式接口，必须先调用“aclnnCtcLossGetWorkspaceSize”接口获取计算所需workspace大小以及包含了算子计算流程的执行器，再调用“aclnnCtcLoss”接口执行计算。

• 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)
• aclnnStatus aclnnCtcLoss(void* workspace, uint64_t workspaceSize, aclOpExecutor* executor, aclrtStream stream)

功能描述

• 算子功能：计算连接时序分类损失值。

• 计算表达式：

  定义$y_{k}^{t}$表示在时刻$t$时真实字符为$k$的概率。(一般地，$y_{k}^{t}$是经过softmax之后的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(1)：

  $$p(π|x)=\prod_{t=1}^{T}y^{t}_{π_{t}} , \forall π \in L'^{T}. \tag{1}$$

  定义多对一(many to one)映射B: $L^{'T} \to L^{\leq T}$，通过映射B计算得到$l \in L^{\leq T}$的条件概率，等于对应于$l$的所有可能路径的概率之和，公式(2):

  $$p(l|x)=\sum_{π \in B^{-1}(l)}p(π|x).\tag{2}$$

  将找到使$p(l|x)$值最大的$l$的路径的任务称为解码，公式(3)：

  $$h(x)=^{arg \ max}_{l \in L^{ \leq T}} \ p(l|x).\tag{3}$$

aclnnCtcLossGetWorkspaceSize

• 参数说明：

  • logProbs(aclTensor*, 计算输入): 表示输出的对数概率，Device侧的aclTensor。数据类型支持FLOAT、DOUBLE。shape为($T, N, C$)，$T$为输入长度，$N$为批处理大小，$C$为类别数，必须大于0，包括空白标识。支持非连续的Tensor。数据格式支持ND。
  • targets(aclTensor*, 计算输入): 表示包含目标序列的标签，Device侧的aclTensor。数据类型支持INT64、INT32、BOOL、FLOAT、FLOAT16。当shape为($N, S$)，$S$为不小于$targetLengths$中的最大值的值；或者shape为(SUM($targetLengths$))，假设$targets$是未填充的而且在1维内级联的。支持非连续的Tensor。数据格式支持ND。
  • inputLengths(aclIntArray*, 计算输入)：表示输入序列的实际长度，Host侧的aclIntArray。数组长度为$N$，数组中的每个值必须小于等于$T$。
  • targetlengths(aclIntArray*, 计算输入)：表示目标序列的实际长度，Host侧的aclIntArray。数组长度为$N$，当targets的shape为($N, S$)时，数组中的每个值必须小于等于$S$。
  • blank(int, 计算输入)：表示空白标识，Host侧的整型。数值必须小于$C$大于等于0。
  • zeroInfinity(bool, 计算输入)：表示是否将无限损耗和相关梯度归零，Host侧的bool类型。
  • negLogLikelihoodOut(aclTensor*, 计算输出): 表示输出的损失值，Device侧的aclTensor。数据类型FLOAT、DOUBLE，且数据类型必须和logProbs一致。shape大小为($N$)的Tensor。支持非连续的Tensor。数据格式支持ND。
  • logAlphaOut(aclTensor*, 计算输出): 表示输入到目标的可能跟踪的概率，Device侧的aclTensor。数据类型支持FLOAT、DOUBLE，且数据类型必须和logProbs一致。shape为3维。支持非连续的Tensor。数据格式支持ND。
  • workspaceSize(uint64_t*, 出参): 返回需要在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：

  aclnnStatus: 返回状态码，具体参见aclnn返回码。

  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是空指针时。
  返回161002（ACLNN_ERR_PARAM_INVALID）：1. logProbs、targets、inputLengths、targetlengths的数据类型不在支持的范围之内。
                                        2. logProbs、targets、inputLengths、targetlengths、negLogLikelihoodOut、logAlphaOut的Tensor不满足对应的shape要求，或者inputLengths、targetLengths的ArrayList的长度不满足要求。
                                        3. blank不满足取值范围。

aclnnCtcLoss

• 参数说明：

  • workspace(void*, 入参): 在Device侧申请的workspace内存地址。
  • workspaceSize(uint64_t, 入参): 在Device侧申请的workspace大小，由第一段接口aclnnCtcLossGetWorkspaceSize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：

  aclnnStatus: 返回状态码，具体参见aclnn返回码。

约束与限制

无

调用示例

示例代码如下，仅供参考，具体编译和执行过程请参考编译与运行样例。

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

#define CHECK_RET(cond, return_expr) \
  do {                               \
    if (!(cond)) {                   \
      return_expr;                   \
    }                                \
  } while (0)

#define LOG_PRINT(message, ...)     \
  do {                              \
    printf(message, ##__VA_ARGS__); \
  } while (0)

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

int Init(int32_t deviceId, aclrtStream* stream) {
  // 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 = aclrtCreateStream(stream);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtCreateStream failed. ERROR: %d\n", ret); return ret);
  return 0;
}

template <typename T>
int CreateAclTensor(const std::vector<T>& hostData, const std::vector<int64_t>& shape, void** deviceAddr,
                    aclDataType dataType, aclTensor** tensor) {
  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--) {
    strides[i] = shape[i + 1] * strides[i + 1];
  }

  // 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;
}

int main() {
  // 1. （固定写法）device/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;
  aclrtStream stream;
  auto ret = Init(deviceId, &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 = {12, 4, 5};
  std::vector<int64_t> targetsShape = {4, 7};
  std::vector<int64_t> negLoglikelihoodOutShape = {4};
  std::vector<int64_t> logAlphaOutShape = {4, 12, 16};
  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 = {
    -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};
  std::vector<int64_t> targetsHostData = {
    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};

  std::vector<float> negLoglikelihoodOutHostData = {0, 0, 0, 0};
  std::vector<float> logAlphaOutHostData = {
    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,

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

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

    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0};
  // 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 = {10,10,10,10};
  inputLengths = aclCreateIntArray(inputLengthsSizeData.data(), 4);
  CHECK_RET(inputLengths != nullptr, return ACL_ERROR_BAD_ALLOC);
  std::vector<int64_t> targetLengthsSizeData = {2, 3, 1, 5};
  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API
  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) {
    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;);
  }
  // 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++) {
    LOG_PRINT("negLoglikelihoodOut result[%ld] is: %f\n", i, resultData[i]);
  }

  // 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++) {
    LOG_PRINT("logAlphaOut result[%ld] is: %f\n", i, resultData1[i]);
  }

  // 7. 释放aclTensor和IntArray，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);

  // 8. 释放device资源，需要根据具体API的接口定义修改
  aclrtFree(logProbsDeviceAddr);
  aclrtFree(targetsDeviceAddr);
  aclrtFree(negLoglikelihoodOutDeviceAddr);
  aclrtFree(logAlphaOutDeviceAddr);
  if (workspaceSize > 0) {
    aclrtFree(workspaceAddr);
  }
  aclrtDestroyStream(stream);
  aclrtResetDevice(deviceId);
  aclFinalize();
  return 0;
}