aclnnSinkhorn

Applicable Products

Function

• Operator function:

  Computes the Sinkhorn distance, which can be used for expert routing in MoE models.

• Formula:

  $$p=Sinkhorn(cost, tol)$$

  Input: cost(R, C): 2D cost matrix tol: tolerance

  Initialization:

  $$cost = exp(cost) \\ d0 = ones(R) \\ d1 = ones(C) \\ eps = 0.00000001 \\ error = 1e9 \\ d1\_old= d1 \\$$

  Repeat:

  $$d0 = \frac{1}{R * (sum(d1 * cost, 1) + eps)} \\ d1 = \frac{1}{C * (sum(d0.unsqueeze(1) * cost, 0) + eps)} \\ error = mean(abs(d1\_old - d1)) \\ d1\_old = d1$$

  Until:

  $$error <= tol$$

  Output:

  $$p = d1 * cost * d0.unsqueeze(1)$$

Function Prototype

Each operator uses a two-phase API. You must first call the "aclnnSinkhornGetWorkspaceSize" API to obtain the input parameters and calculate the required workspace size based on the computation process, and then call the "aclnnSinkhorn" API to perform the computation.

• aclnnStatus aclnnSinkhornGetWorkspaceSize(const aclTensor *cost, const aclScalar *tol, aclTensor *p, uint64_t *workspaceSize, aclOpExecutor** executor)

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

aclnnSinkhornGetWorkspaceSize

• Parameters:

  • cost (aclTensor*, computation input): Represents the cost tensor, i.e., cost in the formula, an aclTensor on the Device side. Supported data types: BFLOAT16, FLOAT16, FLOAT. Values must be in the range [0, 1] and can be normalized. The Data Format supports ND. The input is a two-dimensional matrix with no more than 10,000 rows and no more than 1,024 columns. Supports Non-contiguous Tensor.

  • tol (aclScalar*, input parameter): Represents the error tolerance for Sinkhorn computation. The data type supports FLOAT. If a null pointer is passed, tol defaults to 0.0001.

  • p (aclTensor*, computation output): Represents the optimal transport tensor, i.e., p in the formula, an aclTensor on the Device side. Supported data types: BFLOAT16, FLOAT16, FLOAT. The Data Format supports ND. The shape dimension is 2. Non-contiguous tensor is not supported. The data type and shape are consistent with those of the input parameter cost.

  • workspaceSize (uint64_t*, output parameter): Returns the workspace size to be allocated on the Device side.

  • executor (aclOpExecutor**, output parameter): Returns the op executor, which includes computation.

• Return value:

  aclnnStatus: return code. For details, see aclnn Return Codes.

  The first-phase API performs input parameter validation and returns an error in the following scenarios:

  Return Value	Error Code	Description
  ACLNN_ERR_PARAM_NULLPTR	161001	The cost or p passed in is a null pointer.
  ACLNN_ERR_PARAM_INVALID	161002	The data types of cost and p are not within the supported range.
  		The data type deduction cannot be performed between cost and p.

aclnnSinkhorn

• Parameters:

  Parameter	Input/Output	Description
  workspace	Input	Address of the workspace memory applied on the device side.
  workspaceSize	Input	Size of the workspace applied on the device side, obtained from the first-phase API aclnnSinkhornGetWorkspaceSize.
  executor	Input	Operator executor that contains the operator computation flow.
  stream	Input	Specifies the stream for task execution.

• Return value:

  aclnnStatus: return code. For details, see aclnn Return Codes.

Constraints

• Deterministic computation:

  • aclnnSinkhorn defaults to a deterministic implementation.

Example

The following provides example code for reference only. For details about compilation and running, see Compile and Run Samples.

#include <iostream>
#include <vector>
#include "acl/acl.h"
#include "aclnnop/aclnn_sinkhorn.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 shapeSize = 1;
  for (auto i : shape) {
    shapeSize *= i;
  }
  return shapeSize;
}

int Init(int32_t deviceId, aclrtStream *stream) {
  // Boilerplate: Initialize ACL.
  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);
  // Call aclrtMalloc to apply for device-side memory.
  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);
  // Call aclrtMemcpy to copy host-side data to the device-side memory.
  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);

  // Compute the strides of a contiguous tensor.
  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];
  }

  // Call the aclCreateTensor API to create an 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. (Boilerplate) Initialize the device/stream. See the AscendCL external API list.
  // Set deviceId based on your actual device.
  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. Construct the input and output. Custom construction is required based on the API.
  std::vector<int64_t> costShape = {3, 2};
  std::vector<int64_t> pShape = {3, 2};
  void* costDeviceAddr = nullptr;
  void* pDeviceAddr = nullptr;
  aclTensor* cost = nullptr;
  aclScalar* tol = nullptr;
  aclTensor* p = nullptr;
  std::vector<float> costHostData = {45, 48, 65, 68, 68, 10};
  std::vector<float> pHostData(6, 0);

  float tolValue = 0.0001;

  // Create the cost aclTensor.
  ret = CreateAclTensor(costHostData, costShape, &costDeviceAddr, aclDataType::ACL_FLOAT, &cost);
  CHECK_RET(ret == ACL_SUCCESS, return ret);

  // Create the p aclTensor.
  ret = CreateAclTensor(pHostData, pShape, &pDeviceAddr, aclDataType::ACL_FLOAT, &p);
  CHECK_RET(ret == ACL_SUCCESS, return ret);

  // Create the tol aclScalar.
  tol = aclCreateScalar(&tolValue, aclDataType::ACL_FLOAT);
  CHECK_RET(tol != nullptr, return ret);

  // 3. Call the CANN operator library API. Replace with the specific API name.
  uint64_t workspaceSize = 0;
  aclOpExecutor* executor;
  // Call the first-phase API of aclnnSinkhorn.
  ret = aclnnSinkhornGetWorkspaceSize(cost, tol, p, &workspaceSize, &executor);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnSinkhornGetWorkspaceSize failed. ERROR: %d\n", ret); return ret);
  // Allocate device memory based on the workspaceSize computed by the first-phase API.
  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);
  }
  // Call the second-phase API of aclnnSinkhorn.
  ret = aclnnSinkhorn(workspaceAddr, workspaceSize, executor, stream);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnSinkhorn failed. ERROR: %d\n", ret); return ret);

  // 4. (boilerplate) Synchronously wait for task execution to complete.
  ret = aclrtSynchronizeStream(stream);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtSynchronizeStream failed. ERROR: %d\n", ret); return ret);

  // 5. Obtain the output value and copy the result from device memory to host memory. Modify based on the specific API definition.
  auto size = GetShapeSize(pShape);
  std::vector<float> pData(size, 0);
  ret = aclrtMemcpy(pData.data(), pData.size() * sizeof(pData[0]), pDeviceAddr,
                    size * sizeof(pData[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++) {
    LOG_PRINT("p result[%ld] is: %e\n", i, pData[i]);
  }

  // 6. Release aclTensor and aclScalar. Modify based on the specific API definition.
  aclDestroyTensor(cost);
  aclDestroyTensor(p);
  aclDestroyScalar(tol);

  // Release device resources. Modify based on the specific API definition as needed.
  aclrtFree(costDeviceAddr);
  aclrtFree(pDeviceAddr);
  if (workspaceSize > 0) {
    aclrtFree(workspaceAddr);
  }
  aclrtDestroyStream(stream);
  aclrtResetDevice(deviceId);
  aclFinalize();

  return 0;
}