I was looking at the benchmarks in this answer, and wanted to compare them with diag (used in a different answer). Unfortunately, it seems that diag takes ages:
nc <- 1e4
set.seed(1)
m <- matrix(sample(letters,nc^2,replace=TRUE), ncol = nc)
microbenchmark(
diag = diag(m),
cond = m[row(m)==col(m)],
vec = m[(1:nc-1L)*nc+1:nc],
mat = m[cbind(1:nc,1:nc)],
times=10)
Comments: I tested these with identical. I took "cond" from one of the answers to this homework question. Results are similar with a matrix of integers, 1:26 instead of letters.
Results:
Unit: microseconds
expr min lq mean median uq max neval
diag 604343.469 629819.260 710371.3320 706842.3890 793144.019 837115.504 10
cond 3862039.512 3985784.025 4175724.0390 4186317.5260 4312493.742 4617117.706 10
vec 317.088 329.017 432.9099 350.1005 629.460 651.376 10
mat 272.147 292.953 441.7045 345.9400 637.506 706.860 10
It is just a matrix-subsetting operation, so I don't know why there's so much overhead. Looking inside the function, I see a few checks and then c(m)[v], where v is the same vector used in the "vec" benchmark. Timing these two...
v <- (1:nc-1L)*nc+1:nc
microbenchmark(diaglike=c(m)[v],vec=m[v])
# Unit: microseconds
# expr min lq mean median uq max neval
# diaglike 579224.436 664853.7450 720372.8105 712649.706 767281.5070 931976.707 100
# vec 334.843 339.8365 568.7808 646.799 663.5825 1445.067 100
...it seems I have found my culprit. So, the new variation on my question is: Why is there a seemingly unnecessary and very time-consuming c in diag?
Summary
As of R version 3.2.1 (World-Famous Astronaut) diag() has received an update. The discussion moved to r-devel where it was noted that c() strips non-name attributes and may have been why it was placed there. While some people worried that removing c() would cause unknown issues on matrix-like objects, Peter Dalgaard found that, "The only case where the c() inside diag() has an effect is where M[i,j] != M[(i-1)*m+j] AND c(M) will stringize M in column-major order, so that M[i,j] == c(M)[(i-1)*m+j]."
Luke Tierney tested @Frank 's removal of c(), finding it did not effect anything on CRAN or BIOC and so was implemented to replace c(x)[...] with x[...] on line 27. This leads to relatively large speedups in diag(). Below is a speed test showing the improvement with R 3.2.1's version of diag().
library(microbenchmark)
nc <- 1e4
set.seed(1)
m <- matrix(sample(letters,nc^2,replace=TRUE), ncol = nc)
microbenchmark(diagOld(m),diag(m))
Unit: microseconds
expr min lq mean median uq max neval
diagOld(m) 451189.242 526622.2775 545116.5668 531905.5635 540008.704 682223.733 100
diag(m) 222.563 646.8675 644.7444 714.4575 740.701 1015.459 100
来源：https://stackoverflow.com/questions/30035939/why-is-the-diag-function-so-slow-in-r-3-2-0-or-earlier