aclnnScaledMaskedSoftmaxBackward

Applicable Products

Product	Supported
Atlas A3 training products/Atlas A3 inference products	√
Atlas A2 training products/Atlas A2 inference products	√

Function

Operator function: backward propagation of softmax, with scaling and masking applied to the result.
Formula:
$out = gradOutput \cdot output - sum(gradOutput \cdot output)\cdot output \\ out = \begin{cases} out * scale, &\text { mask is 0 } \\ 0, &\text { mask is 1 } \end{cases}$

Function Prototype

Each operator is divided into a two-phase API. First call the "aclnnScaledMaskedSoftmaxBackwardGetWorkspaceSize" API to obtain the workspace size required for computation and the executor that contains the operator computation process, and then call the "aclnnScaledMaskedSoftmaxBackward" API to perform the computation.

aclnnStatus aclnnScaledMaskedSoftmaxBackwardGetWorkspaceSize(
    const aclTensor* gradOutput,
    const aclTensor* y,
    const aclTensor* mask,
    double scale,
    bool fixTriuMask,
    aclTensor*       out,
    uint64_t*        workspaceSize,
    aclOpExecutor**  executor)

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

aclnnScaledMaskedSoftmaxBackwardGetWorkspaceSize

Parameters:

Parameter	Input/Output	Description	Instruction	Data Type	Data Format	Dimension (Shape)	Non-contiguous Tensor
gradOutput	Input	Gradient value of backward propagation, i.e., the output gradient of the previous layer, which is gradOutput in the formula.	The data type and shape must be consistent with y. The shape must satisfy the broadcast relationship with mask.	FLOAT, FLOAT16, BFLOAT16	ND	[B,N,S1,S2]	√
y	Input	Output value of the softmax function, i.e., y in the formula.	The data type and shape must be consistent with gradOutput The shape must satisfy the broadcast relationship with mask.	FLOAT, FLOAT16, BFLOAT16	ND	[B,N,S1,S2]	√
mask	Input	Mask applied to the computation result, i.e., mask in the formula.	The shape must satisfy the broadcast relationship with y.	BOOL	ND	[B,N,S1,S2], [B,1,S1,S2], [1,N,S1,S2], [1,1,S1,S2]	√
scale	Input	Scale factor used to scale the output, corresponding to scale in the formula.	-	DOUBLE	-	-	-
fixTriuMask	Input	Indicates whether an upper-triangular mask tensor needs to be generated inside the operator.	Only false is supported.	BOOL	-	-	-
out	Output	Represents the gradient of y, corresponding to "out" in the formula.	The data type and shape must be consistent with gradOutput and y.	FLOAT, FLOAT16, BFLOAT16	ND	[B,N,S1,S2]	×
workspaceSize	Output	Returns the workspace size to be allocated on the Device side.	-	-	-	-	-
executor	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	gradOutput, y, mask, and output out are null pointers.
ACLNN_ERR_PARAM_INVALID	161002	The data type, shape, and data format of gradOutput, y, mask, and out are not within the supported range.
		The shape and dtype of gradOutput, y, and out are inconsistent.
		gradOutput, y, and mask do not satisfy the broadcast relationship.
		fixTriuMask is not false.

aclnnScaledMaskedSoftmaxBackward

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 through the first-phase API aclnnScaledMaskedSoftmaxBackwardGetWorkspaceSize.
executor	Input	Operator executor, which 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
- aclnnScaledMaskedSoftmaxBackward defaults to a non-deterministic implementation. You can enable deterministic computation by calling aclrtCtxSetSysParamOpt.
The range of the last axis S2 is (0, 4096].
The first two dimensions of the mask shape can differ from those of gradOutput and y, but must satisfy the broadcast relationship.

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_scaled_masked_softmax_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) {
  // Boilerplate: resource initialization.
  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;
}

void PrintOutResult(std::vector<int64_t> &shape, void** deviceAddr) {
  auto size = GetShapeSize(shape);
  std::vector<float> resultData(size, 0);
  auto ret = aclrtMemcpy(resultData.data(), resultData.size() * sizeof(resultData[0]),
                         *deviceAddr, 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);
  for (int64_t i = 0; i < size; i++) {
    LOG_PRINT("result[%ld] is: %f\n", i, resultData[i]);
  }
}

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 memory on the device side.
  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 data from the host to the device.
  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 aclCreateTensor 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. Refer to the acl API manual.
  // Fill in 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 inputs and outputs. Customize the construction based on the API interface.
  // input
  std::vector<float> gradOutputHostData = {1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16};
  std::vector<float> yHostData = {1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16};
  std::vector<char> maskHostData = {1, 1, 0, 0, 1, 1, 0, 0, 1, 1, 0, 0, 1, 1, 0, 0};
  std::vector<float> outHostData(16, 0);
  std::vector<int64_t> gradOutputShape = {2, 2, 2, 2};
  std::vector<int64_t> yShape = {2, 2, 2, 2};
  std::vector<int64_t> maskShape = {2, 2, 2, 2};
  std::vector<int64_t> outShape = {2, 2, 2, 2};
  void *gradOutputDeviceAddr = nullptr;
  void *yDeviceAddr = nullptr;
  void *maskDeviceAddr = nullptr;
  void *outDeviceAddr = nullptr;
  aclTensor *gradOutput = nullptr;
  aclTensor *y = nullptr;
  aclTensor *mask = nullptr;
  aclTensor *out = nullptr;

  ret = CreateAclTensor(gradOutputHostData, gradOutputShape, &gradOutputDeviceAddr, aclDataType::ACL_FLOAT, &gradOutput);
  CHECK_RET(ret == ACL_SUCCESS, return ret);
  ret = CreateAclTensor(yHostData, yShape, &yDeviceAddr, aclDataType::ACL_FLOAT, &y);
  CHECK_RET(ret == ACL_SUCCESS, return ret);
  ret = CreateAclTensor(maskHostData, maskShape, &maskDeviceAddr, aclDataType::ACL_BOOL, &mask);
  CHECK_RET(ret == ACL_SUCCESS, return ret);
  ret = CreateAclTensor(outHostData, outShape, &outDeviceAddr, aclDataType::ACL_FLOAT, &out);
  CHECK_RET(ret == ACL_SUCCESS, return ret);

  // attr
  double scale = 1.0f;
  bool fixTriuMask = false;

  uint64_t workspaceSize = 0;
  aclOpExecutor *executor;

  // 3. Call the CANN operator library API. Change to the specific API name.
  // aclnnScaledMaskedSoftmaxBackward
  ret = aclnnScaledMaskedSoftmaxBackwardGetWorkspaceSize(gradOutput, y, mask, scale, fixTriuMask, out, &workspaceSize, &executor);
  CHECK_RET(
      ret == ACL_SUCCESS,
      LOG_PRINT("aclnnScaledMaskedSoftmaxBackwardGetWorkspaceSize 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);
  }

  // aclnnScaledMaskedSoftmaxBackward
  ret = aclnnScaledMaskedSoftmaxBackward(workspaceAddr, workspaceSize, executor, stream);
  CHECK_RET(ret == ACL_SUCCESS,
            LOG_PRINT("aclnnScaledMaskedSoftmaxBackward failed. ERROR: %d\n", ret);
            return ret);

  // 4. (Boilerplate) Wait for the task execution to finish synchronously.
  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 the device memory to the host memory. Modify this based on the specific API definition.
  PrintOutResult(outShape, &outDeviceAddr);

  auto size = GetShapeSize(outShape);
  std::vector<float> resultData(size, 0);
  ret = aclrtMemcpy(resultData.data(), resultData.size() * sizeof(resultData[0]), outDeviceAddr,
                    size * sizeof(resultData[0]), ACL_MEMCPY_DEVICE_TO_HOST);

  // 6. Release aclTensor and aclScalar. Modify this based on the specific API definition.
  aclDestroyTensor(gradOutput);
  aclDestroyTensor(y);
  aclDestroyTensor(mask);
  aclDestroyTensor(out);

  // 7. Release device resources.
  aclrtFree(gradOutputDeviceAddr);
  aclrtFree(yDeviceAddr);
  aclrtFree(maskDeviceAddr);
  aclrtFree(outDeviceAddr);

  if (workspaceSize > 0) {
    aclrtFree(workspaceAddr);
  }
  aclrtDestroyStream(stream);
  aclrtResetDevice(deviceId);
  aclFinalize();

  return 0;
}