HCgemmBatched

Applicable Products

Function Description

API function:

• asdBlasMakeHCgemmBatchedPlan: initializes the operator configuration corresponding to the handle.
• asdBlasHCgemmBatched: computes the product of complex matrices.

Formula:

• asdBlasHCgemmBatched formula:
• Example:

  inTensorA[i]:

  [   [ 1+i, 1+2i ],
      [ 1+3i, 1+4i ]  ]

  inTensorB[i]:

  [   [ 2+i, 2+2i ],
      [ 2+3i, 2+4i ]  ]

  inTensorC[i]:

  [   [ 3+i, 3+2i ],
      [ 3+3i, 3+4i ]  ]

  The input transa is N, and the input transb is T.

  The input m is 2, the input n is 2, the input k is 2, the input alpha is 1+i, and the input beta is 2+2i.

  The input lda is 2, the input ldb is 2, and the input ldc is 2.

  The input batchCount is 1.

  After the asdBlasHCgemmBatched operator is called, the output C is as follows:

  [   [ -15+19i, -27+19i ],
      [ -37+21i, -57+13i ]  ]

Function Prototype

• AspbStatus asdBlasMakeHCgemmBatchedPlan(asdBlasHandle handle)
• AspbStatus asdBlasHCgemmBatched(asdBlasHandle handle, asdBlasOperation_t transa, asdBlasOperation_t transb, const int64_t m,const int64_t n, const int64_t k, const std::complex<op::fp16_t> &alpha, aclTensor *A,const int64_t lda, aclTensor *B, const int64_t ldb, const std::complex<op::fp16_t> &beta,aclTensor *C, const int64_t ldc, const int64_t batchCount)

Parameter Description

• asdBlasMakeHCgemmBatchedPlan

  Parameter	Input/Output	Type	Description
  handle	Input	asdBlasHandle	Handle of the HCgemmBatched operator.

• asdBlasHCgemmBatched

  Parameter	Input/Output	Type	Description
  handle	Input	asdBlasHandle	Handle of the HCgemmBatched operator.
  transa	Input	asdBlasOperation_t	Whether matrix A should be transposed. The value must be ASDBLAS_OP_N.
  transb	Input	asdBlasOperation_t	Whether matrix B should be transposed. The value must be ASDBLAS_OP_N.
  m	Input	const int64_t	Number of rows of matrix C. The value range is {1-32}.
  n	Input	const int64_t	Number of columns of matrix C. The value range is {1-32}.
  k	Input	const int64_t	Common dimension of matrices A and B. The value range is {1-32}.
  alpha	Input	const std::complex<op::fp16_t> &	alpha in the formula, which is a complex scalar used to multiply the result of matrix multiplication. The value must be 1+0j.
  A	Input	aclTensor *	A in the formula, which is in column-major order and is a tensor on the device. The data type is COMPLEX32, the data format is ND, and the shape is [batchCount, m, k].
  lda	Input	const int64_t	Memory address offset between adjacent elements in matrix A. The value is the same as that of k.
  B	Input	aclTensor *	B in the formula, which is a tensor on the device. The data type can only be COMPLEX32, the data format can be ND, and the shape is [batchCount, k, n].
  ldb	Input	const int64_t	Memory address offset between adjacent elements in matrix B. The value is the same as that of n.
  beta	Input	const std::complex<op::fp16_t> &	beta in the formula is a complex scalar, and is used to multiply the matrix C. The value must be 0+0j.
  C	Output	aclTensor *	C in the formula, which is a tensor on the device. The data type can only be COMPLEX32, the data format can be ND, and the shape is [batchCount, m, n].
  ldc	Input	const int64_t	Memory address offset between the left and right adjacent elements of C. The value is the same as n.
  batchCount	Input	const int64_t	Number of batches. The value range is {12 - 26208}.

Return Value Description

For details about the return values, see Return Value.

Constraints

• asdBlasMakeHCgemmBatchedPlan: none.
• asdBlasHCgemmBatched
  • The supported CANN version is CANN 8.0 or later.
  • The supported input data type is COMPLEX32.
  • The supported output data type is COMPLEX32.
  • During actual operator computation, only 3D ND computation is supported.
  • The input data of the operator is in row-major order. The input shape is [batchCount, m, k], [batchCount, k, n], or [batchCount, m, n], and the output shape is [batchCount, m, n].

Calling Example

For details about the operator calling example, see HcgemmBatched.