Performs the matrix multiply-add operation, where the input left matrix A is a sparse matrix and the right matrix B is a dense matrix. For matrix A, densification is completed during asc_mmad_sparse computation; for matrix B, densification is completed by the user during input data preparation before computation (following the densification algorithm described below), so the matrix B input to this interface is a dense matrix. The dense matrix B must be loaded through , and the index matrix is loaded at the same time. The index matrix is generated during the densification of matrix B and then used for the densification of matrix A.
- Regular computation
- Synchronous computation
None
PIPE_MTE1
- Each group of 4 elements in the original sparse matrix B should contain at most 2 non-zero elements. If there are 3 or more non-zero elements, only the first 2 non-zero elements are used.
- When any of m, k, or n is 0, this instruction is not executed.
Assume that each group of 4 elements in the original sparse matrix B contains at least 2 zeros. The densified matrix B is a dense matrix with 2 zeros filtered out from each group of 4 elements. During the densification of matrix B, an index matrix is generated as follows: for each group of 4 elements in sparse matrix B, 2 two-bit indices are generated in the index matrix and encoded according to the following rules. The indices must be in the range of {0, 1, 2}. [object Object]• The first index indicates the relative position of the first non-zero element among the first 3 elements. [object Object]• The second index indicates the relative position of the second non-zero element among the last 3 elements. [object Object]For details, refer to the following table. "-" indicates that the algorithm does not care about the value at that position, because it will be filtered.
[object Object][object Object][object Object][object Object][object Object][object Object][object Object]