32-Bit Floats



Drew Schmidt Wei-Chen Chen

See all contributors



float is a single precision (aka float) matrix framework for R. Base R has no single precision type. Its "numeric" vectors/matrices are double precision (or possibly integer, but you know what I mean). Floats have half the precision of double precision data, for a pretty obvious performance vs accuracy tradeoff.

A matrix of floats should use about half as much memory as a matrix of doubles, and your favorite matrix routines will generally compute about twice as fast on them as well. However, the results will not be as accurate, and are much more prone to roundoff error/mass cancellation issues. Statisticians have a habit of over-hyping the dangers of roundoff error in this author's opinion. If your data is well-conditioned, then using floats is "probably" fine for many applications.

Type promotion always defaults to the higher precision. So if a float matrix operates with an integer matrix, the integer matrix will be cast to a float first. Likewise if a float matrix operates with a double matrix, the float will be cast to a double first. Similarly, any float matrix that is explicitly converted to a "regular" matrix will be stored in double precision.


The package requires the single precision BLAS/LAPACK routines which are not included in the default libRblas and libRlapack shipped from CRAN. If your BLAS/LAPACK libraries do not have what is needed, then they will be built (note that a fortran compiler is required in this case). However, these can take a very long time to compile, and will have much worse performance than optimized libraries. The topic of which BLAS/LAPACK to use and how to use them has been written about many times. If this is the first you're hearing of it, I would recommend you use Microsoft R Open.

To install the R package, run:


The development version is maintained on GitHub, and can easily be installed by any of the packages that offer installations from GitHub:

### Pick your preference


If you are installing on Windows and wish to get the best performance, then you will need to install from source after editing some files. After installing high-performance BLAS and LAPACK libraries, delete the text $(LAPACK_OBJS) from line in src/Makevars.win beginning with OBJECTS =. You will also need to add the appropriate link line. This will ensure that on building, the package links with your high-performance libraries instead of compiling the reference versions. This is especially important for 32-bit Windows where the internal LAPACK and BLAS libraries are built without compiler optimization because of a compiler bug.

Also, if you are using Windows on big endian hardware (I'm not even sure if this is possible), then you will need to change the 0 in src/windows/endianness.h to a 1. Failure to do so will cause very bizarre things to happen with the NA handlers.


The goal of the package is to recreate the matrix algebra facilities of the base package, but with floats. So we do not include higher statistical methods (like lm() and prcomp()).

Is something missing? Please let me know.

Basic utilities

Method Status
[ done
c() done
cbind() and rbind() done
diag() done
is.na() done
is.float() done
min() and max() done
na.omit(), na.exclude() done
nrow(), ncol(), dim() done
object.size() done
print() done
scale() Available for logical center and scale
str() done
sweep() Available for FUN's "+", "-", "*", and "/". Others impossible(?)
typeof() and storage.mode() No storage.mode<- method.
which.min() and which.max() done

Binary Operations

Method Status
+ done
* done
- done
/ done
^ done
> done
>= done
== done
< done
<= done

Casters and Converters

Method Status
dbl() done
int() done
fl() done
as.vector() and as.matrix() done

Linear algebra

Method Status
%*% done
backsolve() and forwardsolve() done
chol(), chol2inv() done
crossprod() and tcrossprod() done
eigen() only for symmetric inputs
isSymmetric() done
La.svd() and svd() done
norm() done
qr(), qr.Q(), qr.R() done
rcond() done
solve() done
t() done

Math functions

Method Status
abs(), sqrt() done
ceiling(), floor(), trunc(), round() done
exp(), exp1m() done
gamma(), lgamma() done
is.finite(), is.infinite(), is.nan() done
log(), log10(), log2() done
sin(), cos(), tan(), asin(), acos(), atan() done
sinh(), cosh(), tanh(), asinh(), acosh(), atanh() done


Method Status
.Machine_float float analogue of .Machine. everything you'd actually want is there

Sums and Means

Method Status
colMeans() done
colSums() done
rowMeans() done
rowSums() done
sum() done

Package Use

Memory consumption is roughly half when using floats:


m = 10000
n = 2500

memuse::howbig(m, n)
## 190.735 MiB

x = matrix(rnorm(m*n), m, n)
## 200000200 bytes

s = fl(x)
## 100000784 bytes

And the runtime performance is (generally) roughly 2x better:

cols <- cols <- c("test", "replications", "elapsed", "relative")
reps <- 5

benchmark(crossprod(x), crossprod(s), replications=reps, columns=cols)
##           test replications elapsed relative
## 2 crossprod(s)            5   3.185    1.000
## 1 crossprod(x)            5   7.163    2.249

However, the accuracy is better in the double precision version:

cpx = crossprod(x)
cps = crossprod(s)
all.equal(cpx, dbl(cps))
## [1] "Mean relative difference: 3.478718e-07"

For this particular example, the difference is fairly small; but for some operations/data, the difference could be significantly larger due to roundoff error.

A Note About Memory Consumption

Because of the use of S4 for the nice syntax, there is some memory overhead which is noticeable for small vectors/matrices. This cost is amortized quickly for reasonably large vectors/matrices. But storing many very small float vectors/matrices can be surprisingly costly.

For example, consider the cost for a single float vector vs a double precision vector:

## 632 bytes
## 48 bytes

However once we get to 147 elements, the storage is identical:

## 1216 bytes
## 1216 bytes

And for vectors/matrices with many elements, the size of the double precision data is roughly twice that of the float data:

## 40624 bytes
## 80040 bytes

The above analysis assumes that your float and double values are conforming to the IEEE-754 standard (which is required to build this package). It specifies that a float requires 4 bytes, and a double requires 8. The size of an int is actually system dependent, but is probably 4 bytes. This means that for most, a float matrix should always be larger than a similarly sized integer matrix, because the overhead for our float matrix is simply larger. However, for objects with many elements, the sizes will be roughly equal:

## 40624 bytes
## 40040 bytes


Why would I want to do arithmetic in single precision?

It's (generally) twice as fast and uses half the RAM compared to double precision. For a some data analysis tasks, that's more important than having (roughly) twice as many decimal digits.

Why does floatmat + 1 produce a numeric (double) matrix but floatmat + 1L produce a float matrix?

Type promotion always defaults to the highest type available. If you want the arithmetic to be carried out in single precision, cast the 1 with fl(1) first.

Doesn't that make R's type system even more of a mess?


How would I create my own methods?

If you can formulate the method in terms of existing functionality from the float package, then you're good. If not, you will likely have to write your own C/C++ code. See the For Developers section of the package vignette.