下载
中文
注册

aclnnEinsum

支持的产品型号

  • Atlas 推理系列产品。
  • 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aclnnEinsumGetWorkspaceSize;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aclnnEinsum接口执行计算

  • aclnnStatus aclnnEinsumGetWorkspaceSize(const aclTensorList *tensors, const char * equation, aclTensor *output, uint64_t *workspaceSize, aclOpExecutor **executor)
  • aclnnStatus aclnnEinsum(void *workspace, uint64_t workspaceSize, aclOpExecutor *executor, aclrtStream stream)

功能描述

  • 算子功能:使用爱因斯坦求和约定对张量序列计算代数张量运算, 形式为term1, term2 -> output-term使用以下等式产生输出张量, 其中reduce-sum对出现在输入项(term1, term2)中但未出现在输出项中的所有索引执行求和。
  • 计算公式: output[output-term] = reduce-sum( input1[term1] * input2[term2] )

aclnnEinsumGetWorkspaceSize

  • 参数说明

    • tensors(aclTensorList*, 计算输入):Device侧的aclTensorList。支持非连续的Tensor,不支持空tensor,数据格式支持ND。
      • Atlas A2训练系列产品/Atlas 800I A2推理产品:数据类型支持FLOAT16、FLOAT、INT16、UINT16、INT32、UINT32、INT64、UINT64
      • Atlas 推理系列产品:数据类型支持FLOAT16、INT16、UINT16、INT32、UINT32、INT64、UINT64
    • equation(char*, 计算输入):Host侧 表达式字符串。表示爱因斯坦求和约定的简写公式
    • output(aclTensor*, 计算输出):Device侧的输出tensor。支持非连续的Tensor,不支持空tensor,数据格式支持ND。
      • Atlas A2训练系列产品/Atlas 800I A2推理产品:数据类型支持FLOAT16、FLOAT、INT16、UINT16、INT32、UINT32、INT64、UINT64
      • Atlas 推理系列产品:数据类型支持FLOAT16、INT16、UINT16、INT32、UINT32、INT64、UINT64
    • 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):1. 传入的tensors、output是空指针
    返回161002(ACLNN_ERR_PARAM_INVALID):1. tensors和output的数据类型不在支持的范围内:
                                     2. equation(可扩充) 不在注册表内
                                     3. 当equation=='abcd,abced->abce'
                                        1) tensors 中包含2个Tensor (i.e. tensors[0] & tensors[1])
                                        2) tensors[0] 、tensors[1]、out三者数据类型需保持一致
                                        3) tensors[0] 必须为4维;
                                        4) tensors[1] 必须为5维;
                                        5) tensors[0] 前3维 必须等于 tensors[1] 前3维度;
                                        6) tensors[0] 第4维 必须等于 tensors[1] 第5维度;
                                     4. 当equation=='a,b->ab'
                                        1) tensorList 中包含2个Tensor (i.e. tensors[0] & tensors[1])
                                        2) tensors[0] 、tensors[1]、out三者数据类型需保持一致
                                        3) tensors[0] 必须为1维;
                                        4) tensors[1] 必须为1维;

aclnnEinsum

  • 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aclnnEinsumGetWorkspaceSize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

目前equation需完全匹配, 才能找到对应函数

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_einsum.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)

template <typename T>
int64_t GetShapeSize(const std::vector<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_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> selfShape1 = {1, 2, 3, 4};
  std::vector<int64_t> selfShape2 = {1, 2, 3, 5, 4};
  std::vector<int64_t> outShape = {1, 2, 3, 5};
  void* input1DeviceAddr = nullptr;
  void* input2DeviceAddr = nullptr;
  void* outDeviceAddr = nullptr;
  aclTensor* input1 = nullptr;
  aclTensor* input2 = nullptr;
  aclTensor* out = nullptr;
  std::vector<int32_t> input1HostData = {0, 1, 2, 6, 5, 1, 6, 4, 4, 8, 0, 3, 5, 2, 2, 6, 9, 9, 9, 2, 0, 8, 0, 9};
  std::vector<int32_t> input2HostData = {4, 7, 1, 6, 9, 6, 6, 1, 3, 7, 1, 3, 5, 0, 0, 7, 6, 3, 3, 7, 2, 0, 5, 0,
                                       0, 7, 9, 3, 7, 2, 3, 3, 5, 1, 9, 0, 0, 9, 8, 9, 4, 3, 1, 2, 8, 3, 0, 5,
                                       5, 0, 1, 5, 4, 6, 6, 0, 5, 5, 2, 6, 4, 8, 2, 1, 7, 7, 9, 8, 9, 3, 9, 9,
                                       5, 5, 8, 1, 5, 8, 9, 1, 8, 6, 6, 9, 9, 6, 7, 9, 1, 8, 5, 2, 0, 2, 3, 1,
                                       5, 3, 7, 9, 6, 2, 5, 3, 6, 6, 4, 9, 8, 7, 6, 5, 0, 0, 9, 2, 6, 1, 0, 6};
  std::vector<int32_t> outHostData(30, 0);

  // 创建input1 aclTensor
  ret = CreateAclTensor(input1HostData, selfShape1, &input1DeviceAddr, aclDataType::ACL_INT32, &input1);
  CHECK_RET(ret == ACL_SUCCESS, return ret);

  // 创建input2 aclTensor
  ret = CreateAclTensor(input2HostData, selfShape2, &input2DeviceAddr, aclDataType::ACL_INT32, &input2);
  CHECK_RET(ret == ACL_SUCCESS, return ret);

  // 创建out aclTensor
  ret = CreateAclTensor(outHostData, outShape, &outDeviceAddr, aclDataType::ACL_INT32, &out);
  CHECK_RET(ret == ACL_SUCCESS, return ret);

  std::vector<aclTensor*> tmp{input1, input2};
  aclTensorList* tensorList = aclCreateTensorList(tmp.data(), tmp.size());

  const char equation[] = "abcd,abced->abce";

  // 3. 调用CANN算子库API
  uint64_t workspaceSize = 0;
  aclOpExecutor* executor;
  // 调用aclnnNonMaxuuppression第一段接口
  ret = aclnnEinsumGetWorkspaceSize(tensorList, equation, out, &workspaceSize, &executor);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnEinsumGetWorkspaceSize 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aclnnNonMaxSuppression第二段接口
  ret = aclnnEinsum(workspaceAddr, workspaceSize, executor, stream);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("allnnEinsum 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);
  ret = aclrtMemcpy(outHostData.data(), outHostData.size() * sizeof(outHostData[0]),
                    outDeviceAddr, size * sizeof(outHostData[0]), ACL_MEMCPY_DEVICE_TO_HOST);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("copy outHostData from device to host failed. ERROR: %d\n", ret); return ret);
  for (int64_t i = 0; i < size; i++) {
    LOG_PRINT("relult[%ld] is: %i\n", i, outHostData[i]);
  }

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


  // 7. 释放Device资源, 需要根据具体API的接口定义修改
  aclrtFree(input1DeviceAddr);
  aclrtFree(input2DeviceAddr);
  aclrtFree(outDeviceAddr);
  if (workspaceSize > 0) {
    aclrtFree(workspaceAddr);
  }
  aclrtDestroyStream(stream);
  aclrtResetDevice(deviceId);
  aclFinalize();
  return 0;
}