下载
中文
注册

aclnnRoiAlignV2Backward

支持的产品型号

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

接口原型

每个算子分为两段式接口,必须先调用“aclnnRoiAlignV2BackwardGetWorkspaceSize”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“aclnnRoiAlignV2Backward”接口执行计算。

  • aclnnStatus aclnnRoiAlignV2BackwardGetWorkspaceSize(const aclTensor* gradOutput, const aclTensor* boxes, const aclIntArray* inputShape, int64_t pooledHeight, int64_t pooledWidth, float spatialScale, int64_t samplingRatio, bool aligned, aclTensor* gradInput, uint64_t* workspaceSize, aclOpExecutor** executor)
  • aclnnStatus aclnnRoiAlignV2Backward(void* workspace, uint64_t workspaceSize, aclOpExecutor* executor, aclrtStream stream)

功能描述

算子功能:aclnnRoiAlignV2的反向传播,RoiAlign是一种池化层,用于非均匀输入尺寸的特征图,并输出固定尺寸的特征图。

aclnnRoiAlignV2BackwardGetWorkspaceSize

  • 参数说明:

    • gradOutput(aclTensor*,计算输入):反向传播的输入。Device侧的aclTensor,数据类型支持FLOAT,必须与boxes、gradInput数据类型一致。支持非连续的Tensor数据格式支持NCHW。维度为4维,shape为(K,C,pooledHeight,pooledWidth),表示反向传播的输入梯度张量一个batch内有K个元素,每个元素有C个尺寸为pooledHeight * pooledWidth的特征图,K需要与boxes第0维保持一致。

    • boxes(aclTensor*,计算输入):感兴趣区域box坐标。Device侧的aclTensor,数据类型支持FLOAT,必须与gradOutput、gradInput数据类型一致。支持非连续的Tensor数据格式支持ND。维度为2维,shape为(K,5),5代表box相关信息(image_id,x1,y1,x2,y2)。

    • inputShape(aclIntArray*,计算输入):正向输入的shape,用来指定反向传播的输出shape。Host侧的aclIntArray,支持的数据类型为INT32、INT64。size大小为4,值为(B,C,inputHeight,inputWidth),表示正向RoiAlign的输入张量一个batch内有B张图像,每个图像有C个尺寸为inputHeight * inputWidth的特征图。

    • pooledHeight(int64_t,计算输入):正向RoiAlign池化后输出图像的高度。Host侧的输入参数。

    • pooledWidth(int64_t,计算输入):正向RoiAlign池化后输出图像的宽度。Host侧的输入参数。

    • spatialScale(float,计算输入):乘法空间尺度因子,将ROI坐标从其输入空间尺度转换为池化时使用的尺度,即输入特征图X相对于输入图像的空间尺度。Host侧的输入参数,需大于0。

    • samplingRatio(int64_t,计算输入):RoiAlign中用于计算每个输出元素的和W上的bin数。Host侧的输入参数,需大于等于0。

    • aligned(bool,计算输入):如果为false,则对齐aclnnRoiAlign版本实现;如果为true,则box坐标像素偏移-0.5来使相邻像素索引更好对齐。Host侧的输入参数。

    • gradInput(aclTensor*,计算输出):反向传播的输出。Device侧的aclTensor,数据类型支持FLOAT,必须与gradOutput、boxes数据类型一致。支持非连续的Tensor数据格式支持NCHW。维度为4维,shape为(B,C,inputHeight,inputWidth)。

    • 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。

  • 返回码:

    aclnnStatus:返回状态码,具体参见aclnn返回码

返回 161001(ACLNN_ERR_PARAM_NULLPTR):传入的gradOutput、boxes、inputShape、gradInput是空指针。
返回 161002(ACLNN_ERR_PARAM_INVALID):1. gradOutput和gradInput的数据类型和数据格式不在支持的范围内。
                                      2. gradOutput、boxes、inputShape和gradInput的shape不满足约束限制。
                                      3. spatialScale需大于0,samplingRatio需大于等于0。

aclnnRoiAlignV2Backward

  • 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aclnnRoiAlignV2BackwardGetWorkspaceSize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。

  • 返回码:

    aclnnStatus:返回状态码,具体参见aclnn返回码

约束与限制

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_roi_align_v2_backward.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;
}

template <typename T>
int CreateAclNchTensor(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_NCHW,
                            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> gradOutputShape = {1, 1, 3, 3};
  std::vector<int64_t> boxesShape = {1, 5};
  std::vector<int64_t> inputShape = {1, 1, 6, 6};

  void* gradOutputDeviceAddr = nullptr;
  void* boxesDeviceAddr = nullptr;
  void* gradInputDeviceAddr = nullptr;
  aclTensor* gradOutput = nullptr;
  aclTensor* boxes = nullptr;
  aclTensor* gradInput = nullptr;

  std::vector<float> gradOutputHostData = {4.5, 6.5, 8.5, 16.5, 18.5, 20.5, 28.5, 30.5, 32.5};
  std::vector<float> boxesHostData = {0.0, -2.0, -2.0, 22.0, 22.0};
  std::vector<float> gradInputHostData = {1.125, 1.125, 1.625, 1.625, 2.125, 2.125, 1.125, 1.125, 1.625, 1.625, 2.125, 2.125,
                                    4.125, 4.125, 4.625, 4.625, 5.125, 5.125, 4.125, 4.125, 4.625, 4.625, 5.125, 5.125,
                                    7.125, 7.125, 7.625, 7.625, 8.125, 8.125, 7.125, 7.125, 7.625, 7.625, 8.125, 8.125};

  // 创建gradOutput aclTensor
  ret = CreateAclNchTensor(gradOutputHostData, gradOutputShape, &gradOutputDeviceAddr, aclDataType::ACL_FLOAT, &gradOutput);
  CHECK_RET(ret == ACL_SUCCESS, return ret);
  // 创建boxes aclTensor
  ret = CreateAclTensor(boxesHostData, boxesShape, &boxesDeviceAddr, aclDataType::ACL_FLOAT, &boxes);
  CHECK_RET(ret == ACL_SUCCESS, return ret);
  // 创建inputShape aclIntArray
  const aclIntArray *inputShapeArray = aclCreateIntArray(inputShape.data(), inputShape.size());
  CHECK_RET(inputShapeArray != nullptr, return ACL_ERROR_INTERNAL_ERROR);
  // 创建gradInput aclTensor
  ret = CreateAclNchTensor(gradInputHostData, inputShape, &gradInputDeviceAddr, aclDataType::ACL_FLOAT, &gradInput);
  CHECK_RET(ret == ACL_SUCCESS, return ret);
  int64_t pooledHeight = 3;
  int64_t pooledWidth = 3;
  int64_t samplingRatio = 2;
  float spatialScale = 0.25f;
  bool aligned = false;

  // 3. 调用CANN算子库API
  uint64_t workspaceSize = 0;
  aclOpExecutor* executor;
  // 调用aclnnRoiAlignV2Backward第一段接口
  ret = aclnnRoiAlignV2BackwardGetWorkspaceSize(gradOutput, boxes, inputShapeArray, pooledHeight, pooledWidth, spatialScale, 
                                              samplingRatio, aligned, gradInput, &workspaceSize, &executor);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnRoiAlignV2BackwardGetWorkspaceSize 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aclnnRoiAlignV2Backward第二段接口
  ret = aclnnRoiAlignV2Backward(workspaceAddr, workspaceSize, executor, stream);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnRoiAlignV2Backward 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(inputShape);
  std::vector<float> resultData(size, 0);
  ret = aclrtMemcpy(resultData.data(), resultData.size() * sizeof(resultData[0]),
                    gradInputDeviceAddr, size * sizeof(resultData[0]), ACL_MEMCPY_DEVICE_TO_HOST);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("copy resultData from device to host failed. ERROR: %d\n", ret); return ret);
  for (int64_t i = 0; i < size; i++) {
    LOG_PRINT("result[%ld] is: %f\n", i, resultData[i]);
  }

  // 6. 释放aclTensor,需要根据具体API的接口定义修改
  aclDestroyTensor(gradOutput);
  aclDestroyTensor(boxes);
  aclDestroyIntArray(inputShapeArray);
  aclDestroyTensor(gradInput);

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