sandy4321 / matrixcsr

Additional functionality for dgRMatrix matrices (aka CSR format) in Matrix package

NOW PART OF THE RECO PACKAGE

Faster sparse-dense matrix products

Default format for sparse matrices in R is CSC - compressed column sparse. However CSR is more mainstream in many algorithms and modern C/C++ libraries.

MatrixCSR package brings faster CSR * dense matrix products to R. Noticeably CSR * dense matrix products are multithreaded (thanks to Eigen and RcppEigen libraries).

In example below on my laptop with 4-core/8-thread Intel Core i7 CPU CSR * dense martix product is more than 6 times faster then default CSC * dense product.

library(Matrix)
library(MatrixCSR)
library(microbenchmark)

set.seed(1)
nnz_share = 0.0001
n_row = 1e6
n_col = 1e5
n_row_dense = n_col
n_col_dense = 1e2

nnz = n_row * n_col * nnz_share
sp_m_coo = sparseMatrix(i = sample(n_row, nnz, TRUE), j = sample(n_col, nnz, TRUE), 
                        x = runif(nnz, TRUE), dims = c(n_row, n_col), giveCsparse = FALSE)

object.size(sp_m_coo)/1e6
# 160 Mb

sp_m_csc = as(sp_m_coo, "CsparseMatrix")
sp_m_csr = as(sp_m_coo, "RsparseMatrix")

dense_m = matrix(runif(n_row_dense * n_col_dense), nrow = n_row_dense, ncol = n_col_dense)
object.size(dense_m)/1e6
# 80 Mb


microbenchmark(
  csc_prod = sp_m_csc %*% dense_m, 
  csr_prod = sp_m_csr %*% dense_m, 
  times = 10
)
#Unit: seconds
#     expr       min        lq     mean    median        uq       max neval
# csc_prod 10.550470 10.617285 10.85617 10.713557 11.224757 11.362016    10
# csr_prod  1.627799  1.717712  1.75765  1.729939  1.774248  1.905975    10

Faster truncated SVD

irlba package provides two of state-of the art implementations of truncated SVD solvers:

1. Implicitly-restarted Lanczos method
2. Randomized SVD

With MatrixCSR package second approach can be significantly accelerated for free (without any additional manual adjustments). For example on matrix from example aboce it is more than 2 times faster than default CSC.

library(irlba)
microbenchmark(
  # randomized SVD on CSC matrix
  SVD_CSC = svdr(x = sp_m_csc, k = 10),
  # randomized SVD on CSR matrix
  SVD_CSR = svdr(x = sp_m_csr, k = 10),
  times = 1
)
#Unit: seconds
#    expr      min       lq     mean   median       uq      max neval
# SVD_CSC 25.70179 25.70179 25.70179 25.70179 25.70179 25.70179     1
# SVD_CSR 11.61360 11.61360 11.61360 11.61360 11.61360 11.61360     1