aclnnMatmulWeightNz

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: Performs matrix multiplication between tensor self and tensor mat2. mat2 only supports the NZ format. Only supports self with 2 dimensions and mat2 with 4 dimensions.

Similar APIs include aclnnMatmul (mat2 only supports ND), aclnnMm (supports matrix multiplication with 2D tensors as input), and aclnnBatchMatmul (only supports 3D matrix multiplication, where the first dimension is batch).
Formula:
$result=self @ mat2$

Function Prototype

Each operator is divided into a two-phase API. The "aclnnMatmulWeightNzGetWorkspaceSize" API must be called first to obtain the workspace size required for computation and the executor that encapsulates the operator computation flow. Then, the "aclnnMatmulWeightNz" API is called to perform the computation.

aclnnStatus aclnnMatmulWeightNzGetWorkspaceSize(
  const aclTensor *self,
  const aclTensor *mat2,
  aclTensor       *out,
  int8_t          cubeMathType,
  uint64_t        *workspaceSize,
  aclOpExecutor   **executor)

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

aclnnMatmulWeightNzGetWorkspaceSize

Parameters

Parameter	Input/Output	Description	Instruction	Data Type	Data Format	Shape	Non-contiguous Tensor
self	Input	Represents the first matrix in matrix multiplication, i.e., self in the formula.	The data type must satisfy the data type deduction rules with mat2 (see Deduction Relationship and Constraints). - When self is not transposed, each dimension represents: (m, k) - When self is transposed, each dimension represents: (k, m)	BFLOAT16, FLOAT16	ND	2	√
mat2	Input	Represents the second matrix of the matrix multiplication, i.e., mat2 in the formula.	The data type must satisfy the data type deduction rules with self (see Deduction Relationship and Constraints). The reduce dimension of mat2 must be equal in size to the reduce dimension of self. When matrix B is not transposed, the NZ format dimensions are: (n1, k1, k0, n0), where k0 = 16 and n0 = 16. The k in self shape and k1 in mat2 shape must satisfy the following relationship: ceil(k, k0) = k1. The n1 in mat2 shape and n in out must satisfy the following relationship: ceil(n, n0) = n1. When matrix B is transposed, the NZ format dimensions are: (k1, n1, n0, k0), where n0 = 16 and k0 = 16. The k in self shape and k1 in mat2 shape must satisfy the following relationship: ceil(k, k0) = k1. The n1 in mat2 shape and n in out must satisfy the following relationship: ceil(n, n0) = n1.	BFLOAT16, FLOAT16, FLOAT32	NZ	4	-
out	Output	Represents the output matrix of the matrix multiplication, i.e., out in the formula.	The data type must be consistent with the data type derived from self and mat2 (see Deduction Relationship and Constraints). Dimension representation: (m, n), where m is consistent with the m of self, and n satisfies the relationship ceil(n / n0) = n1 with n1 and n0 of mat2.	BFLOAT16, FLOAT16, FLOAT32	ND	2	-
cubeMathType	Input	Specifies the computation logic of the Cube unit.	If the input data types have a deduction relationship, this parameter processes the data type after deduction by default. The supported enumeration values are as follows: 0: KEEP_DTYPE, keeps the input data type for computation. 1: ALLOW_FP32_DOWN_PRECISION, allows down-precision computation of the input data. When the input data type is FLOAT32, it is converted to HFLOAT32 for computation. When the input is of other data types, no processing is performed. 2: USE_FP16, allows down-precision of the input to FLOAT16 for computation. This option is not supported when the input data type is BFLOAT16. 3: USE_HF32, allows down-precision of the input to HFLOAT32 for computation. When the input data type is FLOAT32, it is converted to HFLOAT32 for computation. This option is not supported when the input is of other data types. 4: USE_FP32_ADD, uses a high-precision method for computation.	INT8	-	-	-
workspaceSize	Output	Returns the workspace size to be applied for on the Device.	-	-	-	-	-
executor	Output	Returns the operator executor, which contains the operator computation flow.	-	-	-	-	-

Atlas A2 training products/Atlas A2 inference products, Atlas A3 training products/Atlas A3 inference products:
- Before calling this API, you must use the aclnnTransMatmulWeight API to convert the original input Format of mat2 from ND to NZ.
- cubeMathType=4: When the input data type is FLOAT32 and the k-axis is greater than 2048, grouped accumulation is used for computation. When the input is of other data types or the k-axis is less than 2048, no processing is performed.
Ascend 950PR/Ascend 950DT:
- Before calling this API, you must use aclnnNpuFormatCast to convert the original input format of mat2 from ND to NZ.
- The option cubeMathType = 1: ALLOW_FP32_DOWN_PRECISION is not supported.
- The option cubeMathType = 3: USE_HF32 is not supported.
- The option cubeMathType = 4: USE_FP32_ADD is not supported.

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	The input self, mat2, or out is a null pointer.
ACLNN_ERR_PARAM_INVALID	161002	The data type and data format of self and mat2 are not within the supported range.
		The data type of self and mat2 cannot be deduced.
		The deduced data type cannot be converted to the specified output out type.

aclnnMatmulWeightNz

Parameters

Parameter	Input/Output	Description
workspace	Input	Workspace memory address allocated on the Device side.
workspaceSize	Input	Size of the workspace requested on the Device side, obtained by the first-phase API aclnnMatmulWeightNzGetWorkspaceSize.
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 description:
- Atlas training products, Atlas inference products, Ascend 950PR/Ascend 950DT: aclnnMatmulWeightNz defaults to a deterministic implementation.
Computation consistency description
- Atlas training products, Atlas inference products:
  - When the strong consistency computation feature is enabled, the computation result is deterministic, and multiple executions produce the same output. In addition, the computation result is independent of the data location.
  - aclnnMatmulWeightNz defaults to a non-deterministic implementation and supports enabling deterministic computation via aclrtCtxSetSysParamOpt.
  - For example, during Matrix Multiplication, the accumulation order of different basic blocks may vary, which can cause slight differences in the computation results of the same data across different rows. However, when deterministic computation is enabled, the computation results will be identical as long as the inputs are the same, even across different rows.
Data type deduction between BFLOAT16 and FLOAT16 inputs is not supported.
self only supports 2 dimensions, and mat2 only supports the Ascend private format. Before calling this API, you must complete the conversion of mat2 from ND to the Ascend private format.
The case where one of the last two axes of mat2 is 1, i.e., k=1 or n=1, is not supported.
Recommended use case: It is recommended to use this API in mte2 bound scenarios, for example, when the M axis is small, matrix A involves repeated data movement, and matrix B does not.

Example

Atlas A2 training products/Atlas A2 inference products, Atlas A3 training products/Atlas A3 inference products:

The following is sample code for the scenario where self and mat2 have the float16 data type and mat2 is in NZ format. This is for reference only. For details about compilation and execution, see Compile and Run Samples.

#include <iostream>
#include <vector>
#include <cmath>
#include "acl/acl.h"
#include "aclnnop/aclnn_matmul.h"
#include "aclnnop/aclnn_trans_matmul_weight.h"
#include "aclnnop/aclnn_cast.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;
}

// Convert the uint16_t representation of FP16 to a float representation.
float Fp16ToFloat(uint16_t h) {
  int s = (h >> 15) & 0x1;              // sign
  int e = (h >> 10) & 0x1F;             // exponent
  int f =  h        & 0x3FF;            // fraction
  if (e == 0) {
    // Zero or Denormal
    if (f == 0) {
      return s ? -0.0f : 0.0f;
    }
    // Denormals
    float sig = f / 1024.0f;
    float result = sig * pow(2, -24);
    return s ? -result : result;
  } else if (e == 31) {
      // Infinity or NaN
      return f == 0 ? (s ? -INFINITY : INFINITY) : NAN;
  }
  // Normalized
  float result = (1.0f + f / 1024.0f) * pow(2, e - 15);
  return s ? -result : result;
}

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 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 data to 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 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;
}

template <typename T>
int CreateAclTensorWeight(const std::vector<T>& hostData, const std::vector<int64_t>& shape, void** deviceAddr,
                          aclDataType dataType, aclTensor** tensor) {
  auto size = static_cast<uint64_t>(GetShapeSize(shape));

  const aclIntArray* mat2Size = aclCreateIntArray(shape.data(), shape.size());
  auto ret = aclnnCalculateMatmulWeightSize(mat2Size, &size);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnCalculateMatmulWeightSize failed. ERROR: %d\n", ret); return ret);
  size *= sizeof(T);

  // Call aclrtMalloc to apply for device memory.
  ret = aclrtMalloc(deviceAddr, size, ACL_MEM_MALLOC_HUGE_FIRST);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclrtMalloc failed. ERROR: %d\n", ret); return ret);
  // Copy host-side data to device-side memory by calling aclrtMemcpy.
  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];
  }

  std::vector<int64_t> storageShape;
  storageShape.push_back(GetShapeSize(shape));

  // Call aclCreateTensor to create an aclTensor.
  *tensor = aclCreateTensor(shape.data(), shape.size(), dataType, strides.data(), 0, aclFormat::ACL_FORMAT_ND,
                            storageShape.data(), storageShape.size(), *deviceAddr);
  return 0;
}

int main() {
  // 1. (Boilerplate) Initialize the device/stream. See the ACL API manual.
  // Enter the 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> selfShape = {16, 32};
  std::vector<int64_t> mat2Shape = {32, 16};
  std::vector<int64_t> outShape = {16, 16};
  void* selfDeviceAddr = nullptr;
  void* mat2DeviceAddr = nullptr;
  void* outDeviceAddr = nullptr;
  aclTensor* self = nullptr;
  aclTensor* mat2 = nullptr;
  aclTensor* out = nullptr;
  std::vector<uint16_t> selfHostData(512, 0x3C00); // float16_t uses 0x3C00 to represent int_16 value 1.
  std::vector<uint16_t> mat2HostData(512, 0x3C00); // float16_t uses 0x3C00 to represent int_16 value 1.
  std::vector<uint16_t> outHostData(256, 0);
  // Create the self aclTensor.
  ret = CreateAclTensor(selfHostData, selfShape, &selfDeviceAddr, aclDataType::ACL_FLOAT16, &self);
  CHECK_RET(ret == ACL_SUCCESS, return ret);
  // Create the other aclTensor.
  ret = CreateAclTensorWeight(mat2HostData, mat2Shape, &mat2DeviceAddr, aclDataType::ACL_FLOAT16, &mat2);
  CHECK_RET(ret == ACL_SUCCESS, return ret);
  // Create the out aclTensor.
  ret = CreateAclTensor(outHostData, outShape, &outDeviceAddr, aclDataType::ACL_FLOAT16, &out);
  CHECK_RET(ret == ACL_SUCCESS, return ret);

  // 3. Call the CANN operator library API. Replace with the specific API name.
  int8_t cubeMathType = 1;
  uint64_t workspaceSize = 0;
  aclOpExecutor* executor;
  // Call TransWeight.
  ret = aclnnTransMatmulWeightGetWorkspaceSize(mat2, &workspaceSize, &executor);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnTransMatmulWeightGetWorkspaceSize 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 aclnnTransMatmulWeight.
  ret = aclnnTransMatmulWeight(workspaceAddr, workspaceSize, executor, stream);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnTransMatmulWeight failed. ERROR: %d\n", ret); return ret);

  // Call the first-phase API of aclnnMatmulWeightNz.
  uint64_t workspaceSizeMm = 0;
  ret = aclnnMatmulWeightNzGetWorkspaceSize(self, mat2, out, cubeMathType, &workspaceSizeMm, &executor);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnMatmulWeightNzGetWorkspaceSize failed. ERROR: %d\n", ret); return ret);

  // Apply for device memory based on the workspaceSize calculated by the first-phase API.
  void* workspaceAddrMm = nullptr;
  if (workspaceSizeMm > 0) {
    ret = aclrtMalloc(&workspaceAddrMm, workspaceSizeMm, 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 aclnnMatmulWeightNz.
  ret = aclnnMatmulWeightNz(workspaceAddrMm, workspaceSizeMm, executor, stream);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnMatmulWeightNz failed. ERROR: %d\n", ret); return ret);

  // 4. (Boilerplate) Wait for the task 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.
  auto size = GetShapeSize(outShape);
  std::vector<uint16_t> 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);
  // fp16 data cannot be directly printed in C. Read it as uint16 and manually convert the binary representation to float-based fp16.
  for (int64_t i = 0; i < size; i++) {
    float fp16Float = Fp16ToFloat(resultData[i]);
    LOG_PRINT("result[%ld] is: %f\n", i, fp16Float);
  }

  // 6. Release aclTensor and aclScalar. Modify based on the specific API definition.
  aclDestroyTensor(self);
  aclDestroyTensor(mat2);
  aclDestroyTensor(out);

  // 7. Release device resources. Modify based on the specific API definition.
  aclrtFree(selfDeviceAddr);
  aclrtFree(mat2DeviceAddr);
  aclrtFree(outDeviceAddr);

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

Ascend 950PR/Ascend 950DT:

The following is sample code for the scenario where self and mat2 have the bfloat16 data type and mat2 is in NZ format. This is for reference only. For details about compilation and execution, see Compile and Run Samples.

  #include <iostream>
  #include <vector>
  #include "acl/acl.h"
  #include "aclnnop/aclnn_matmul.h"
  #include "aclnnop/aclnn_npu_format_cast.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;
  }

  // Convert the uint16_t representation of bfloat16 to a float representation.
  float Bf16ToFloat(uint16_t h) {
    uint32_t sign = (h & 0x8000U) ? 0x80000000U : 0x00000000U;  // sign bit
    uint32_t exponent = (h >> 7) & 0x00FFU;                      // exponent bits
    uint32_t mantissa =  h & 0x007FU;                           // mantissa bits

    // The exponent bias remains unchanged.
    // Shift the mantissa left by 23 - 7, and pad the remaining bits with zeros.
    uint32_t f_bits = sign | (exponent << 23) | (mantissa << (23 - 7));
    // Cast to float.
    return *reinterpret_cast<float*>(&f_bits);
  }

  int Init(int32_t deviceId, aclrtStream* stream) {
    // Boilerplate code 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 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;
  }

  template <typename T>
  int CreateAclTensorWithFormat(const std::vector<T>& hostData, const std::vector<int64_t>& shape,
                                int64_t** storageShape, uint64_t* storageShapeSize, void** deviceAddr,
                                aclDataType dataType, aclTensor** tensor, aclFormat format) {
      auto size = hostData.size() * 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);
      // Copy the host-side data to the device-side memory by calling aclrtMemcpy.
      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];
      }

      *tensor = aclCreateTensor(shape.data(), shape.size(), dataType, strides.data(), 0,
                                  format, *storageShape, *storageShapeSize, *deviceAddr);
      return 0;
  }

  int main() {
    // 1. Initialize the device/stream. See the ACL API manual (boilerplate).
    // 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. Custom construction is required based on the API.
    int64_t m = 16;
    int64_t k = 32;
    int64_t n = 16;
    std::vector<int64_t> selfShape = {m, k};
    std::vector<int64_t> mat2Shape = {k, n};
    std::vector<int64_t> outShape = {m, n};
    void* selfDeviceAddr = nullptr;
    void* mat2DeviceAddr = nullptr;
    void* outDeviceAddr = nullptr;
    void* dstDeviceAddr = nullptr;

    aclTensor* self = nullptr;
    aclTensor* mat2 = nullptr;
    aclTensor* out = nullptr;
    aclTensor* mat2NZ = nullptr;

    std::vector<uint16_t> selfHostData(m * k, 0x3F80); // 0x3F80 represents 1 in uint16 for bfloat16_t.
    std::vector<uint16_t> mat2HostData(k * n, 0x3F80); // 0x3F80 represents 1 in uint16 for bfloat16_t.
    std::vector<uint16_t> outHostData(m * n, 0);
    // The space required by weightNz is greater than or equal to the original [k, n] matrix, and needs to be aligned to 16.
    int64_t kAlign = (k + 16 - 1) / 16 * 16;
    int64_t nAlign = (n + 16 - 1) / 16 * 16;
    std::vector<uint16_t> dstTensorHostData(kAlign * nAlign, 0x3F80);

    aclDataType srcDtype = aclDataType::ACL_BF16;
    aclDataType additionalDtype = aclDataType::ACL_BF16;
    // Create the self aclTensor.
    ret = CreateAclTensor(selfHostData, selfShape, &selfDeviceAddr, srcDtype, &self);
    CHECK_RET(ret == ACL_SUCCESS, return ret);
    // Create the mat2 aclTensor.
    ret = CreateAclTensor(mat2HostData, mat2Shape, &mat2DeviceAddr, srcDtype, &mat2);
    CHECK_RET(ret == ACL_SUCCESS, return ret);
    // Create the out aclTensor.
    ret = CreateAclTensor(outHostData, outShape, &outDeviceAddr, srcDtype, &out);
    CHECK_RET(ret == ACL_SUCCESS, return ret);

    // 3. Call the CANN operator library API. Replace with the specific API name.
    int8_t cubeMathType = 1;
    aclOpExecutor* executor = nullptr;

    // Convert the weight tensor from ND to NZ. Call the npu_format_cast API.
    int64_t* dstShape = nullptr;
    uint64_t dstShapeSize = 0;
    int actualFormat;
    uint64_t workspaceSize = 0;
    void* workspaceAddr = nullptr;

    uint64_t workspaceSizeMm = 0;
    void* workspaceAddrMm = nullptr;

    // Calculate the shape and format of the target tensor.
    ret = aclnnNpuFormatCastCalculateSizeAndFormat(mat2, 29, additionalDtype, &dstShape, &dstShapeSize, &actualFormat);
    CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnNpuFormatCastCalculateSizeAndFormat failed. ERROR: %d\n", ret); return ret);

    ret = CreateAclTensorWithFormat(dstTensorHostData, mat2Shape, &dstShape, &dstShapeSize, &dstDeviceAddr, srcDtype, &mat2NZ, static_cast<aclFormat>(actualFormat));
    CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("CreateAclTensorWithFormat failed. ERROR: %d\n", ret); return ret);

    // Call the first-phase API of aclnnNpuFormatCastGetWorkspaceSize.
    ret = aclnnNpuFormatCastGetWorkspaceSize(mat2, mat2NZ, &workspaceSize, &executor);
    CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnNpuFormatCastGetWorkspaceSize failed. ERROR: %d\n", ret); return ret);

    // Allocate device memory based on the workspaceSize calculated by the first-phase API.
    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 aclnnNpuFormatCastGetWorkspaceSize second-phase API.
    ret = aclnnNpuFormatCast(workspaceAddr, workspaceSize, executor, stream);
    CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnNpuFormatCast failed. ERROR: %d\n", ret); return ret);

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

    // Call the aclnnMatmulWeightNz first-phase API.
    ret = aclnnMatmulWeightNzGetWorkspaceSize(self, mat2NZ, out, cubeMathType, &workspaceSizeMm, &executor);
    CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnMatmulWeightNzGetWorkspaceSize failed. ERROR: %d\n", ret); return ret);

    // Allocate device memory based on the workspaceSize calculated by the first-phase API.
    if (workspaceSizeMm > 0) {
      ret = aclrtMalloc(&workspaceAddrMm, workspaceSizeMm, 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 aclnnMatmulWeightNz.
    ret = aclnnMatmulWeightNz(workspaceAddrMm, workspaceSizeMm, executor, stream);
    CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnMatmulWeightNz failed. ERROR: %d\n", ret); return ret);

    // 4. Wait for task execution to complete 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.
    auto size = GetShapeSize(outShape);
    std::vector<uint16_t> 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++) {
      float bf16Float = Bf16ToFloat(resultData[i]);
      LOG_PRINT("result[%ld] is: %f\n", i, bf16Float);
    }

    // 6. Release aclTensor and aclScalar.
    aclDestroyTensor(self);
    aclDestroyTensor(mat2);
    aclDestroyTensor(out);
    aclDestroyTensor(mat2NZ);

    // 7. Release device resources.
    aclrtFree(selfDeviceAddr);
    aclrtFree(mat2DeviceAddr);
    aclrtFree(outDeviceAddr);
    aclrtFree(dstDeviceAddr);

    if (workspaceSize > 0) {
      aclrtFree(workspaceAddr);
    }

    if (workspaceSizeMm > 0) {
      aclrtFree(workspaceAddrMm);
    }
    aclrtDestroyStream(stream);
    aclrtResetDevice(deviceId);
    aclFinalize();
    return 0;
  }