aclnnFusedCrossEntropyLossWithMaxSum

Applicable Products

Product	Supported
Ascend 950PR/Ascend 950DT	×
Atlas A3 training products/Atlas A3 inference products	√
Atlas A2 training products/Atlas A2 inference products	√
Atlas 200I/500 A2 inference products	×
Atlas inference products	×
Atlas training products	×

Function

Description: This operator is part of the cross entropy computation module in vocabulary parallelism scenarios, addressing memory and computational efficiency issues for extremely large vocabularies. This part computes the loss and SoftMax results.
Formula:
$lossOut = log(sum\_exp\_logits) - predicted\_logits$ $softMaxOutOptional = exp(vocab\_parallel\_logits -logits\_max.unsqueeze(dim = -1)) \ sum\_exp\_logits.unsqueeze(dim = -1)$

Function Prototype

Each operator uses a two-phase API. You must first call the "aclnnFusedCrossEntropyLossWithMaxSumGetWorkspaceSize" API to obtain the workspace size required for computation and the executor that encapsulates the operator computation process, and then call the "aclnnFusedCrossEntropyLossWithMaxSum" API to perform the computation.

aclnnStatus aclnnFusedCrossEntropyLossWithMaxSumGetWorkspaceSize(
    const aclTensor* logitsMax,
    const aclTensor* sumExpLogits,
    const aclTensor* predictedLogits,
    float            labelSmoothing,
    const aclTensor* inputOptional,
    const aclTensor* weightOptional,
    const aclTensor* vocabParallelLogitsOptional,
    aclTensor*       lossOut,
    aclTensor*       softMaxOutOptional,
    uint64_t*        workspaceSize,
    aclOpExecutor**  executor)

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

aclnnFusedCrossEntropyLossWithMaxSumGetWorkspaceSize

Parameters

Parameter	Input/Output	Description	Instruction	Data Type	Data Format	Dimension (Shape)	Non-contiguous Tensor
logitsMax (aclTensor*)	Input	Maximum value of each row after the matmul computation, corresponding to logitsMax in the formula.	-	FLOAT	ND	1	√
sumExpLogits (aclTensor*)	Input	The result of applying exp to the difference between the matmul result and the maximum value of each row, i.e., sumExpLogits in the formula.	Same as the shape of logitsMax.	FLOAT	ND	1	√
predictedLogits (aclTensor*)	Input	Represents the result after subtracting the maximum value of each row from the matmul computation result and then filtering with maskedTargetOut, i.e., predictedLogits in the formula.	Has the same shape as logitsMax.	FLOAT	ND	1	√
labelSmoothing (float)	Input	Label smoothing coefficient, used to mitigate overfitting.	Currently only 0 is supported.	FLOAT	-	-	-
inputOptional (aclTensor*)	Input	Left matrix input for matmul.	Currently only supports null pointer input.	FLOAT	-	-	√
weightOptional (aclTensor*)	Input	Right matrix input for matmul, i.e., the weight matrix.	Currently only supports a null pointer as input.	FLOAT	-	-	√
vocabParallelLogitsOptional (aclTensor*)	Output	Result of the matmul computation, i.e., vocabParallelLogits in the formula.	The first dimension of shape must be consistent with the first dimension of logitsMax.	FLOAT, FLOAT16, BFLOAT16	ND	2	√
lossOut (aclTensor*)	Output	Intermediate variable, representing the loss in the formula.	The shape is consistent with logitsMax.	FLOAT	ND	1	√
softMaxOutOptional（aclTensor*）	Output	Intermediate variable, corresponding to vocabParallelLogits in the formula.	Shape is the same as vocabParallelLogitsOptional.	FLOAT	ND	2	√
workspaceSize (uint64_t*)	Output	Returns the workspace size to be allocated on the Device side.	-	-	-	-	-
executor (aclOpExecutor**)	Output	Returns the operator executor, which contains the operator computation flow.	-	-	-	-	-

Return Value

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

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	logitsMax, sumExpLogits, predictedLogits, and lossOut are null pointers.
ACLNN_ERR_PARAM_INVALID	161002	The data types of the input and output are not within the supported range.
		The shapes of the input and output do not meet the constraints.
		The dimensions of the input and output do not meet the constraints.
		labelSmoothing is not equal to 0.

aclnnFusedCrossEntropyLossWithMaxSum

Parameters:

Parameter	Input/Output	Description
workspace	Input	Workspace memory address applied for on the Device side.
workspaceSize	Input	Workspace size applied for on the Device side, obtained through the first-phase API aclnnFusedCrossEntropyLossWithMaxSumGetWorkspaceSize.
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 Code.

Constraints

Deterministic computation:
- aclnnFusedCrossEntropyLossWithMaxSum 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_fused_cross_entropy_loss_with_max_sum.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 resources.
  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 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);

  // Calculate 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. For details, see the ACL API manual.
  // Set deviceId based on the 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 inputs and outputs based on the API interface.
  std::vector<int64_t> logitsMaxShape = {2};
  std::vector<int64_t> sumExpLogitsShape = {2};
  std::vector<int64_t> predictedLogitsShape = {2};
  std::vector<int64_t> inputShape = {2};
  std::vector<int64_t> weightShape = {2};
  std::vector<int64_t> vocabParallelLogitsOptionalShape = {2, 2};
  std::vector<int64_t> lossOutShape = {2};
  std::vector<int64_t> softMaxOutOptionalShape = {2, 2};

  float labelSmoothing = 0;

  void* logitsMaxDeviceAddr = nullptr;
  void* sumExpLogitsDeviceAddr = nullptr;
  void* predictedLogitsDeviceAddr = nullptr;
  void* inputDeviceAddr = nullptr;
  void* weightDeviceAddr = nullptr;
  void* vocabParallelLogitsOptionalDeviceAddr = nullptr;
  void* lossOutDeviceAddr = nullptr;
  void* softMaxOutOptionalDeviceAddr = nullptr;

  aclTensor* logitsMax = nullptr;
  aclTensor* sumExpLogits = nullptr;
  aclTensor* predictedLogits = nullptr;
  aclTensor* input = nullptr;
  aclTensor* weight = nullptr;
  aclTensor* vocabParallelLogitsOptional = nullptr;
  aclTensor* lossOut = nullptr;
  aclTensor* softMaxOutOptional = nullptr;

  std::vector<float> logitsMaxHostData = {0.5, 1};
  std::vector<float> sumExpLogitsHostData = {0.5, 1};
  std::vector<float> predictedLogitsHostData = {0.5, 1};
  std::vector<float> inputHostData = {0, 1};
  std::vector<float> weightHostData = {0, 1};
  std::vector<float> vocabParallelLogitsOptionalHostData = {1, 0.5, 0.5, 1};
  std::vector<float> lossOutHostData = {0, 0};
  std::vector<float> softMaxOutOptionalHostData = {0, 0, 0, 0};
  // Create an aclTensor.
  ret = CreateAclTensor(logitsMaxHostData, logitsMaxShape, &logitsMaxDeviceAddr, aclDataType::ACL_FLOAT, &logitsMax);
  CHECK_RET(ret == ACL_SUCCESS, return ret);

  ret = CreateAclTensor(sumExpLogitsHostData, sumExpLogitsShape, &sumExpLogitsDeviceAddr, aclDataType::ACL_FLOAT, &sumExpLogits);
  CHECK_RET(ret == ACL_SUCCESS, return ret);

  ret = CreateAclTensor(predictedLogitsHostData, predictedLogitsShape, &predictedLogitsDeviceAddr, aclDataType::ACL_FLOAT, &predictedLogits);
  CHECK_RET(ret == ACL_SUCCESS, return ret);

  ret = CreateAclTensor(inputHostData, inputShape, &inputDeviceAddr, aclDataType::ACL_FLOAT, &input);
  CHECK_RET(ret == ACL_SUCCESS, return ret);

  ret = CreateAclTensor(weightHostData, weightShape, &weightDeviceAddr, aclDataType::ACL_FLOAT, &weight);
  CHECK_RET(ret == ACL_SUCCESS, return ret);

  ret = CreateAclTensor(vocabParallelLogitsOptionalHostData, vocabParallelLogitsOptionalShape, &vocabParallelLogitsOptionalDeviceAddr, aclDataType::ACL_FLOAT, &vocabParallelLogitsOptional);
  CHECK_RET(ret == ACL_SUCCESS, return ret);

  ret = CreateAclTensor(lossOutHostData, lossOutShape, &lossOutDeviceAddr, aclDataType::ACL_FLOAT, &lossOut);
  CHECK_RET(ret == ACL_SUCCESS, return ret);

  ret = CreateAclTensor(softMaxOutOptionalHostData, softMaxOutOptionalShape, &softMaxOutOptionalDeviceAddr, aclDataType::ACL_FLOAT, &softMaxOutOptional);
  CHECK_RET(ret == ACL_SUCCESS, 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 aclnnFusedCrossEntropyLossWithMaxSum.
  ret = aclnnFusedCrossEntropyLossWithMaxSumGetWorkspaceSize(logitsMax, sumExpLogits, predictedLogits, labelSmoothing, input, weight,
                                                          vocabParallelLogitsOptional, lossOut, softMaxOutOptional, &workspaceSize, &executor);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnFusedCrossEntropyLossWithMaxSumGetWorkspaceSize failed. ERROR: %d\n", ret); return ret);
  // Apply for device memory based on the workspaceSize calculated 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 aclnnFusedCrossEntropyLossWithMaxSum.
  ret = aclnnFusedCrossEntropyLossWithMaxSum(workspaceAddr, workspaceSize, executor, stream);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnFusedCrossEntropyLossWithMaxSum failed. ERROR: %d\n", ret); return ret);

  // 4. (Boilerplate) Synchronize and 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 values and copy the results from the device memory to the host memory. Modify based on the specific API definition.
  auto size = GetShapeSize(lossOutShape);
  std::vector<float> resultData(size, 0);
  ret = aclrtMemcpy(resultData.data(), resultData.size() * sizeof(resultData[0]), lossOutDeviceAddr,
                    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++) {
    LOG_PRINT("result[%ld] is: %f\n", i, resultData[i]);
  }

  size = GetShapeSize(softMaxOutOptionalShape);
  std::vector<float> secondResultData(size, 0);
  ret = aclrtMemcpy(secondResultData.data(), secondResultData.size() * sizeof(secondResultData[0]), softMaxOutOptionalDeviceAddr,
                    size * sizeof(secondResultData[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("result[%ld] is: %f\n", i, secondResultData[i]);
  }

  // 6. Release aclTensor and aclScalar. Modify based on the specific API definition.
  aclDestroyTensor(logitsMax);
  aclDestroyTensor(sumExpLogits);
  aclDestroyTensor(predictedLogits);
  aclDestroyTensor(input);
  aclDestroyTensor(weight);
  aclDestroyTensor(vocabParallelLogitsOptional);
  aclDestroyTensor(lossOut);
  aclDestroyTensor(softMaxOutOptional);

  // 7. Release device resources. Modify based on the specific API definition.
  aclrtFree(logitsMaxDeviceAddr);
  aclrtFree(sumExpLogitsDeviceAddr);
  aclrtFree(predictedLogitsDeviceAddr);
  aclrtFree(inputDeviceAddr);
  aclrtFree(weightDeviceAddr);
  aclrtFree(vocabParallelLogitsOptionalDeviceAddr);
  aclrtFree(lossOutDeviceAddr);
  aclrtFree(softMaxOutOptionalDeviceAddr);
  if (workspaceSize > 0) {
    aclrtFree(workspaceAddr);
  }
  aclrtDestroyStream(stream);
  aclrtResetDevice(deviceId);
  aclFinalize();
  return 0;
}