I guess it might have been obvious, but I was surprised that rowSums appeared much slower slower than directly doing the matrix operation. At first I assumed it was because it has some overhead in the form of input handling/casting and other odds and ends.

set.seed(1079711697)1
n <- 1e6
m <- 3
m1 <- matrix(rnorm(n*m), n, m)
f1 <- function(m) m %*% cbind(rep(1, ncol(m)))
f2 <- function(m) rowSums(m)

microbenchmark::microbenchmark(f1(m1), f2(m1))
# Unit: milliseconds
#   expr      min        lq      mean    median        uq     max neval
# f1(m1) 5.204801  5.335951  6.070637  5.433501  5.659101 20.2690   100
# f2(m1) 9.992201 10.181401 11.014871 10.254801 10.403851 31.1915   100

Above was run on a 6700K and R 4.3.1 (x86_64-w64-mingw32/x64). When I ran this on an Apple computer with an M1 though I was again surprised the pattern was reversed, but both were much faster and closer to one another– means were 3.3 and 2.7 for f1 and f2 respectively. That’s R 4.3.2 (aarch64-apple-darwin20). I haven’t messed with Rblas, but maybe you can get additional benefits there?

It also looks like there are some slight precision differences between the two methods:

summary(f1(m1) - f2(m1))
#       V1            
# Min.   :-1.776e-15  
# 1st Qu.: 0.000e+00  
# Median : 0.000e+00  
# Mean   :-1.190e-19  
# 3rd Qu.: 0.000e+00  
# Max.   : 8.882e-16  

-10^-19 is pretty small. It happens to be about the charge of a single electron in SI units.

1. I didn’t use my orcid, because I don’t want you to know who I am. ︎