aclnnMaskedSoftmaxWithRelPosBias

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: Replaces the softmax computation using window attention in swinTransformer.
Formula:
$out = \operatorname{softmax}(scaleValue * x + attenMaskOptional + relativePosBias)$

Function Prototype

Each operator is divided into a two-phase API. You must first call the "aclnnMaskedSoftmaxWithRelPosBiasGetWorkspaceSize" API to obtain the workspace size required for computation and the executor that contains the operator computation process, and then call the "aclnnMaskedSoftmaxWithRelPosBias" API to perform the computation.

aclnnStatus aclnnMaskedSoftmaxWithRelPosBiasGetWorkspaceSize(
  const aclTensor *x,
  const aclTensor *attenMaskOptional,
  const aclTensor *relativePosBias,
  double           scaleValue,
  int64_t          innerPrecisionMode,
  const aclTensor *out,
  uint64_t        *workspaceSize,
  aclOpExecutor  **executor)

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

aclnnMaskedSoftmaxWithRelPosBiasGetWorkspaceSize

Parameters

Parameter	Input/Output	Description	Instruction	Data Type	Data Format	Non-contiguous Tensor
x	Input	Input to be Computed, which corresponds to x in the formula.	Shape is 4D (B*W, N, S1, S2) or 5D (B, W, N, S1, S2).	FLOAT, FLOAT16, BFLOAT16	ND	-
attenMaskOptional	Input	Input to be computed, corresponding to attenMaskOptional in the formula.	Shape is 3D (W, S1, S2), 4D (W, 1, S1, S2), or 5D (1, W, 1, S1, S2).	FLOAT, FLOAT16, BFLOAT16	ND	-
relativePosBias	Input	Input to be computed, corresponding to relativePosBias in the formula.	The shape is 3D (N, S1, S2), 4D (1, N, S1, S2), or 5D (1, 1, N, S1, S2).	FLOAT, FLOAT16, BFLOAT16	ND	-
scaleValue	Input	Input to be computed, corresponding to scaleValue in the formula.	None	DOUBLE	ND	-
innerPrecisionMode	Input	Precision mode parameter.	None.	INT64	ND	-
out	Output	Output to be computed, corresponding to "out" in the formula.	Shape is the same as x.	FLOAT, FLOAT16, BFLOAT16	ND	-
workspaceSize	Output	Returns the workspace size to be requested on the Device side.	-	-	-	-
executor	Output	Returns the operator executor, which contains the operator computation flow.	-	-	-	-

Atlas inference products: BFLOAT16 is not supported.

Return Value

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

The first-phase API performs input parameter validation and reports errors in the following scenarios:

Return Code	Error Code	Description
ACLNN_ERR_PARAM_NULLPTR	161001	The passed x, attenMaskOptional, relativePosBias, or out is a null pointer.
ACLNN_ERR_INNER_TILING_ERROR	561002	The data type, data format, or shape of the input or output is not within the supported range.

aclnnMaskedSoftmaxWithRelPosBias

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 from the first-phase API aclnnMaskedSoftmaxWithRelPosBiasGetWorkspaceSize.
executor	Input	Operator executor, which contains the operator computation flow.
stream	Input	Specifies the stream on which the task is executed.

Return Value

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

Constraints

Deterministic computation:
- aclnnMaskedSoftmaxWithRelPosBias defaults to a deterministic implementation.
Atlas inference products: The scenario where the last dimension S2 of the input parameter x is not aligned to 32 bytes is not supported.

Ensure that the UB space required by the shape passed to the operator is smaller than the total UB size of the AI processor version. The total UB space size minComputeSize required by this operator is as follows, where s2AlignedSize represents the result after aligning S2 to 32 bytes.

When attenMaskOptional exists:

For the FLOAT type, the formula is as follows:
dtypeSize = 4；
xSize = s2AlignedSize * dtypeSize;
softMaskMinTmpSize = 288；
minComputeSize = xSize * 8 + softMaskMinTmpSize;
For the FLOAT16 type, the formula is as follows:
dtypeSize = 2；
xSize = s2AlignedSize * dtypeSize;
softMaskMinTmpSize = 288；
minComputeSize = xSize * 16 + softMaskMinTmpSize;

When attenMaskOptional does not exist:

For the FLOAT type, the formula is as follows:
dtypeSize = 4；
xSize = s2AlignedSize * dtypeSize;
softMaskMinTmpSize = 288；
minComputeSize = xSize * 6 + softMaskMinTmpSize；
For the FLOAT16 type, the formula is as follows:
dtypeSize = 2；
xSize = s2AlignedSize * dtypeSize;
softMaskMinTmpSize = 288；
minComputeSize = xSize* 12 + softMaskMinTmpSize;

Atlas A2 training products/Atlas A2 inference products, Atlas A3 training products/Atlas A3 inference products: For the BFLOAT16 type, the formula is consistent with that of the FLOAT16 type.

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_masked_softmax_with_rel_pos_bias.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 for 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;
}

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 data from the host to the device 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 the 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.
  // 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 the input and output. Customize the construction based on the API interface.
  std::vector<int64_t> xShape = {1, 1, 1, 2, 16};
  std::vector<int64_t> attenMaskOptionalShape = {1, 2, 16};
  std::vector<int64_t> relativePosBiasShape = {1, 2, 16};
  std::vector<int64_t> outShape = {1, 1, 1, 2, 16};

  void* xDeviceAddr = nullptr;
  void* attenMaskOptionalDeviceAddr = nullptr;
  void* relativePosBiasDeviceAddr = nullptr;
  void* outDeviceAddr = nullptr;

  aclTensor* x = nullptr;
  aclTensor* attenMaskOptional = nullptr;
  aclTensor* relativePosBias = nullptr;
  aclTensor* out = nullptr;

  std::vector<float> xHostData = {1.08, -1.56, -1.3, -2.01, 2.18, -2.23, -3.58, 3.22, 1.25, -0.56, -0.3, -1.01, 1.08, -1.13, -3.08, -2.22, -0.08, -2.56, 1.35, 1.01, 0.35, -1.03, -1.28, 1.22, 0.08, -2.56, -1.01, -1.01, -0.18, -6.23, 4.55, -1.82};
  std::vector<float> attenMaskOptionalHostData = {2, 2, 2, 2, 3, 3, 3, 3, 2, 2, 2, 2, 3, 3, 3, 3, 2, 2, 2, 2, 3, 3, 3, 3, 2, 2, 2, 2, 3, 3, 3, 3,};
  std::vector<float> relativePosBiasHostData = {1, 1, 1, 1, 4, 4, 4, 4, 1, 1, 1, 1, 4, 4, 4, 4, 1, 1, 1, 1, 4, 4, 4, 4, 1, 1, 1, 1, 4, 4, 4, 4};
  std::vector<float> outHostData = {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0};

  ret = CreateAclTensor(xHostData, xShape, &xDeviceAddr, aclDataType::ACL_FLOAT, &x);
  CHECK_RET(ret == ACL_SUCCESS, return ret);
  ret = CreateAclTensor(attenMaskOptionalHostData, attenMaskOptionalShape, &attenMaskOptionalDeviceAddr, aclDataType::ACL_FLOAT, &attenMaskOptional);
  CHECK_RET(ret == ACL_SUCCESS, return ret);
  ret = CreateAclTensor(relativePosBiasHostData, relativePosBiasShape, &relativePosBiasDeviceAddr, aclDataType::ACL_FLOAT, &relativePosBias);
  CHECK_RET(ret == ACL_SUCCESS, return ret);
  ret = CreateAclTensor(outHostData, outShape, &outDeviceAddr, aclDataType::ACL_FLOAT, &out);
  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 aclnnMaskedSoftmaxWithRelPosBias.
  ret = aclnnMaskedSoftmaxWithRelPosBiasGetWorkspaceSize(x, attenMaskOptional, relativePosBias, 1.0, 0, out, &workspaceSize, &executor);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnMaskedSoftmaxWithRelPosBiasGetWorkspaceSize failed. ERROR: %d\n", ret); return ret);

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

  // 4. (Boilerplate) Synchronously wait for the 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 the device memory to the host memory. Modify based on the specific API definition.
  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);
  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]);
  }

  // 6. Release aclTensor and aclScalar. Modify based on the specific API definition.
  aclDestroyTensor(x);
  aclDestroyTensor(attenMaskOptional);
  aclDestroyTensor(relativePosBias);
  aclDestroyTensor(out);

  // 7. Release device resources.
  aclrtFree(xDeviceAddr);
  aclrtFree(attenMaskOptionalDeviceAddr);
  aclrtFree(relativePosBiasDeviceAddr);
  aclrtFree(outDeviceAddr);
  if (workspaceSize > 0) {
    aclrtFree(workspaceAddr);
  }
  aclrtDestroyStream(stream);
  aclrtResetDevice(deviceId);
  aclFinalize();
  return 0;
}