"Wide or tall" and "sparse matrix dense matrix" multiplications
نویسنده
چکیده
This note explores sparse matrix dense matrix (SMDM) multiplications, useful in block Krylov or block Lanczos methods. SMDM computations are AU , and V A, multiplication of a large sparse matrix m × n matrix A by a matrix V of k rows of length m or a matrix U of k columns of length k, k << m, k << n . In a block Lanczos or Krylov algorithm, matrix matrix multiplications with the ”tall” U and ”wide” V are also needed. This note relates some experience in efficiently computing SMDM and ”Wide or Tall” computations on multi-core architectures. 1
منابع مشابه
An Efficient Phone N-Gram Forward-Backward Computation Using Dense Matrix Multiplication
The forward-backward algorithm is commonly used to train neural network acoustic models when optimizing a sequence objective like MMI and sMBR. Recent work on lattice-free MMI training of neural network acoustic models shows that the forward-backward algorithm can be computed efficiently in the probability domain as a series of sparse matrix multiplications using GPUs. In this paper, we present...
متن کاملLazy Householder Decomposition of Sparse Matrices
This paper describes Householder reduction of a rectangular sparse matrix to small band upper triangular form Bk+1. Bk+1 is upper triangular with nonzero entries only on the diagonal and on the nearest k superdiagonals. The algorithm is similar to the Householder reduction used as part of the standard dense SVD computation. For the sparse “lazy” algorithm, matrix updates are deferred until a ro...
متن کاملUBk+1V Block Sparse Householder Decomposition
This paper describes Householder reduction of a rectangular sparse matrix to small band upper triangular form. Using block Householder transformations gives good orthogonality, is computationally efficient, and has good potential for parallelization. The algorithm is similar to the standard dense Householder reduction used as part of the usual dense SVD computation. For the sparse algorithm, th...
متن کاملBlock Algorithms for Sparse Matrix by Dense Matrix Multiplication
Sparse matrix computations appear in many linear algebra kernels of scienti c applications. The study, evaluation and optimization of sparse matrix codes is more complex than the dense case. Moreover, the irregularity of some memory accesses and the a-priory lack of knowledge of the number of iterations to be perfomed in some loops (both depending on the sparsity pettern) limit the succes of pr...
متن کاملTrace-Penalty Minimization for Large-Scale Eigenspace Computation
The Rayleigh-Ritz (RR) procedure, including orthogonalization, constitutes a major bottleneck in computing relatively high-dimensional eigenspaces of large sparse matrices. Although operations involved in RR steps can be parallelized to a certain level, their parallel scalability, which is limited by some inherent sequential steps, is lower than dense matrix-matrix multiplications. The primary ...
متن کامل