aclnnMatmul

Applicable Products

Function

• Description: Performs matrix multiplication between the tensor self and the tensor mat2 (supports 1D to 6D tensors as input for matrix multiplication).

  Similar APIs include aclnnMm (supports 2D tensors as input for matrix multiplication) and aclnnBatchMatmul (supports only 3D matrix multiplication, where the first dimension is the batch).

• Formula:

  $$result=self @ mat2$$

Function Prototype

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

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

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

aclnnMatmulGetWorkspaceSize

• Parameters:

  Parameter	Input/Output	Description	Instruction	Data Type	Data Format	Dimension (Shape)	Non-Contiguous Tensor
  self	Input	The first matrix in the matrix multiplication, corresponding to self in the formula.	The data type must satisfy the data type deduction rules with mat2 (see Deduction Relationship and Constraints). Must satisfy the broadcast relationship with mat2.	BFLOAT16, FLOAT16, FLOAT32	ND	1–6	√
  mat2	Input	Represents the second matrix in 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 have the same size as the reduce dimension of self. Must satisfy the broadcast relationship with self.	BFLOAT16, FLOAT16, FLOAT32	ND	1~6	√
  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).	BFLOAT16, FLOAT16, FLOAT32	ND	1~6	-
  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. 2: USE_FP16. Supports down-precision of the input to FLOAT16 for computation. 3: USE_HF32. Supports down-precision of the input to the HFLOAT32 data type for computation. 4: USE_FP32_ADD. Supports computation using a high-precision method.	INT8	-	-	-
  workspaceSize	Output	Returns the workspace size to be allocated on the Device side.	-	-	-	-	-
  executor	Output Parameter	Returns the operator executor, which contains the operator computation flow.	-	-	-	-	-

  • Atlas training products, Atlas inference products:

    • BFLOAT16 data type is not supported.

    • When the input data type is FLOAT32, cubeMathType=0 is not supported.

    • For cubeMathType=1, when the input data type is FLOAT32, it is converted to FLOAT16 for computation. When the input is of other data types, no processing is performed.

    • cubeMathType=3 is not supported.

    • No processing is performed when cubeMathType is set to 4.

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

    • cubeMathType=1: 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.

    • cubeMathType=2: This option is not supported when the input data type is BFLOAT16.

    • cubeMathType=3: When the input data type is FLOAT32, it will be converted to HFLOAT32 for computation. This option is not supported when the input is of other data types.

    • cubeMathType=4: When the input data type is FLOAT32 and the k-axis is greater than 2048, grouped accumulation is used for computation. No processing is performed when the input is of other data types or the k-axis is less than 2048.

  • Ascend 950PR/Ascend 950DT:

    • cubeMathType=1: 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.

    • cubeMathType=2: This option is not supported when the input data type is BFLOAT16.

    • cubeMathType=3: When the input data type is FLOAT32, it will be converted to HFLOAT32 for computation. This option is not supported when the input is of other data types.

    • No processing is performed when cubeMathType is set to 4.

• 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 passed 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.
  		Data type deduction cannot be performed for self and mat2.
  		The deduced data type cannot be converted to the type of the specified output out.

aclnnMatmul

• Parameters:

  Parameter	Input/Output	Description
  workspace	Input	Specifies the workspace memory address applied for on the Device side.
  workspaceSize	Input	Size of the workspace applied for on the Device side, obtained through the first-phase API aclnnMatmulGetWorkspaceSize.
  executor	Input	Operator executor, which contains the operator computation process.
  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: aclnnMatmul defaults to a deterministic implementation.

  • Ascend 950PR/Ascend 950DT: aclnnMatmul 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.

    • aclnnMatmul defaults to a non-consistent implementation. Consistent computation can be enabled through 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.

• Atlas A2 training products/Atlas A2 inference products, Atlas A3 training products/Atlas A3 inference products: Data type deduction is not supported when the two inputs are BFLOAT16 and FLOAT16 respectively. Data type deduction is not supported when the two inputs are BFLOAT16 and FLOAT32 respectively.

• When both self and mat2 are 1-dimensional, cubeMathType does not take effect.

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_matmul.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 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 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;
}

int main() {
  // 1. (Boilerplate) Initialize the device/stream. See the ACL API manual.
  // Fill in the 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. 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<float> selfHostData(512, 1);
  std::vector<float> mat2HostData(512, 1);
  std::vector<float> outHostData(256, 0);
  // Create the self aclTensor.
  ret = CreateAclTensor(selfHostData, selfShape, &selfDeviceAddr, aclDataType::ACL_FLOAT, &self);
  CHECK_RET(ret == ACL_SUCCESS, return ret);
  // Create the mat2 aclTensor.
  ret = CreateAclTensor(mat2HostData, mat2Shape, &mat2DeviceAddr, aclDataType::ACL_FLOAT, &mat2);
  CHECK_RET(ret == ACL_SUCCESS, return ret);
  // Create the out aclTensor.
  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.
  int8_t cubeMathType = 1;
  uint64_t workspaceSize = 0;
  aclOpExecutor* executor = nullptr;
  // Call the first-phase API of aclnnMatmul.
  ret = aclnnMatmulGetWorkspaceSize(self, mat2, out, cubeMathType, &workspaceSize, &executor);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnMatmulGetWorkspaceSize 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 aclnnMatmul.
  ret = aclnnMatmul(workspaceAddr, workspaceSize, executor, stream);
  CHECK_RET(ret == ACL_SUCCESS, LOG_PRINT("aclnnMatmul 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 side memory to the host side. 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. Destroy 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);
  }
  aclrtDestroyStream(stream);
  aclrtResetDevice(deviceId);
  aclFinalize();
  return 0;
}