Title: | Fast 'match()' Function |
---|---|
Description: | Package providing a fast match() replacement for cases that require repeated look-ups. It is slightly faster that R's built-in match() function on first match against a table, but extremely fast on any subsequent lookup as it keeps the hash table in memory. |
Authors: | Simon Urbanek [aut, cre, cph] (https://urbanek.org, <https://orcid.org/0000-0003-2297-1732>) |
Maintainer: | Simon Urbanek <[email protected]> |
License: | GPL-2 |
Version: | 1.1-6 |
Built: | 2024-12-26 12:45:30 UTC |
Source: | CRAN |
coalesce
makes sure that a given index vector is coalesced,
i.e., identical values are grouped into contiguous blocks. This can be
used as a much faster alternative to sort.list
where the
goal is to group identical values, but not necessarily in a
pre-defined order. The algorithm is linear in the length of the vector.
coalesce(x)
coalesce(x)
x |
character, integer or real vector to coalesce |
The current implementation takes two passes through the vector. In the
first pass it creates a hash table for the values of x
counting
the occurrences in the process. In the second pass it assigns indices
for every element based on the index stored in the hash table.
The order of the groups of unique values is defined by the first
occurence of each unique value, hence it is identical to the order of
unique
.
One common use of coalesce
is to allow the use of arbitrary
vectors in ctapply
via
ctapply(x[coalesce(x)], ...)
.
Integer vector with the resulting permutation. x[coalesce(x)]
gives x
with contiguous unique values.
Simon Urbanek
i = rnorm(2e6) names(i) = as.integer(rnorm(2e6)) ## compare sorting and coalesce system.time(o <- i[order(names(i))]) system.time(o <- i[coalesce(names(i))]) ## more fair comparison taking the coalesce time (and copy) into account system.time(tapply(i, names(i), sum)) system.time({ o <- i[coalesce(names(i))]; ctapply(o, names(o), sum) }) ## in fact, using ctapply() on a dummy vector is faster than table() ... ## believe it or not ... (that that is actually wasteful, since coalesce ## already computed the table internally anyway ...) ftable <- function(x) { t <- ctapply(rep(0L, length(x)), x[coalesce(x)], length) t[sort.list(names(t))] } system.time(table(names(i))) system.time(ftable(names(i)))
i = rnorm(2e6) names(i) = as.integer(rnorm(2e6)) ## compare sorting and coalesce system.time(o <- i[order(names(i))]) system.time(o <- i[coalesce(names(i))]) ## more fair comparison taking the coalesce time (and copy) into account system.time(tapply(i, names(i), sum)) system.time({ o <- i[coalesce(names(i))]; ctapply(o, names(o), sum) }) ## in fact, using ctapply() on a dummy vector is faster than table() ... ## believe it or not ... (that that is actually wasteful, since coalesce ## already computed the table internally anyway ...) ftable <- function(x) { t <- ctapply(rep(0L, length(x)), x[coalesce(x)], length) t[sort.list(names(t))] } system.time(table(names(i))) system.time(ftable(names(i)))
ctapply
is a fast replacement of tapply
that assumes
contiguous input, i.e. unique values in the index are never speparated
by any other values. This avoids an expensive split
step since
both value and the index chungs can be created on the fly. It also
cuts a few corners to allow very efficient copying of values. This
makes it many orders of magnitude faster than the classical
lapply(split(), ...)
implementation.
ctapply(X, INDEX, FUN, ..., MERGE=c)
ctapply(X, INDEX, FUN, ..., MERGE=c)
X |
an atomic object, typically a vector |
INDEX |
numeric or character vector of the same length as |
FUN |
the function to be applied |
... |
additional arguments to |
MERGE |
function to merge the resulting vector or |
Note that ctapply
supports either integer, real or character
vectors as indices (note that factors are integer vectors and thus
supported, but you do not need to convert character vectors). Unlike
tapply
it does not take a list of factors - if you want to use
a cross-product of factors, create the product first, e.g. using
paste(i1, i2, i3, sep='\01')
or multiplication - whetever
method is convenient for the input types.
ctapply
requires the INDEX
to contiguous. One (slow) way
to achieve that is to use sort
or order
.
Simon Urbanek
i = rnorm(4e6) names(i) = as.integer(rnorm(1e6)) i = i[order(names(i))] system.time(tapply(i, names(i), sum)) system.time(ctapply(i, names(i), sum))
i = rnorm(4e6) names(i) = as.integer(rnorm(1e6)) i = i[order(names(i))] system.time(tapply(i, names(i), sum)) system.time(ctapply(i, names(i), sum))
fmatch
is a faster version of the built-in match
()
function. It is slightly faster than the built-in version because it
uses more specialized code, but in addition it retains the hash table
within the table object such that it can be re-used, dramatically reducing
the look-up time especially for large tables.
Although fmatch
can be used separately, in general it is also
safe to use: match <- fmatch
since it is a drop-in
replacement. Any cases not directly handled by fmatch
are passed
to match
with a warning.
fmatch.hash
is identical to fmatch
but it returns the table
object with the hash table attached instead of the result, so it can be
used to create a table object in cases where direct modification is
not possible.
%fin%
is a version of the built-in %in%
function
that uses fmatch
instead of match
().
fmatch(x, table, nomatch = NA_integer_, incomparables = NULL) fmatch.hash(x, table, nomatch = NA_integer_, incomparables = NULL) x %fin% table
fmatch(x, table, nomatch = NA_integer_, incomparables = NULL) fmatch.hash(x, table, nomatch = NA_integer_, incomparables = NULL) x %fin% table
x |
values to be matched |
table |
values to be matched against |
nomatch |
the value to be returned in the case when no match is
found. It is coerced to |
incomparables |
a vector of values that cannot be matched. Any
value other than |
See match
for the purpose and details of the
match
function. fmatch
is a drop-in replacement for
the match
function with the focus on
performance. incomparables
are not supported by fmatch
and will be passed down to match
.
The first match against a table results in a hash table to be computed
from the table. This table is then attached as the ".match.hash"
attribute of the table so that it can be re-used on subsequent calls
to fmatch
with the same table.
The hashing algorithm used is the same as the match
function in
R, but it is re-implemented in a slightly different way to improve its
performance at the cost of supporting only a subset of types (integer,
real and character). For any other types fmatch
falls back to
match
(with a warning).
fmatch
: A vector of the same length as x
- see
match
for details.
fmatch.hash
: table
, possibly coerced to match the type
of x
, with the hash table attached.
%fin%
: A logical vector the same length as x
- see
%in%
for details.
fmatch
modifies the table
by attaching an attribute to
it. It is expected that the values will not change unless that
attribute is dropped. Under normal circumstances this should not have
any effect from user's point of view, but there is a theoretical
chance of the cache being out of sync with the table in case the table
is modified directly (e.g. by some C code) without removing
attributes.
In cases where the table
object cannot be modified (or such
modification would not survive) fmatch.hash
can be used to build
the hash table and return table
object including the hash
table. In that case no lookup is done and x
is only used to
determine the type into which table
needs to be coerced.
Also fmatch
does not convert to a common encoding so strings
with different representation in two encodings don't match.
Simon Urbanek
# some random speed comparison examples: # first use integer matching x = as.integer(rnorm(1e6) * 1000000) s = 1:100 # the first call to fmatch is comparable to match system.time(fmatch(s,x)) # but the subsequent calls take no time! system.time(fmatch(s,x)) system.time(fmatch(-50:50,x)) system.time(fmatch(-5000:5000,x)) # here is the speed of match for comparison system.time(base::match(s, x)) # the results should be identical identical(base::match(s, x), fmatch(s, x)) # next, match a factor against the table # this will require both x and the factor # to be cast to strings s = factor(c("1","1","2","foo","3",NA)) # because the casting will have to allocate a string # cache in R, we run a dummy conversion to take # that out of the equation dummy = as.character(x) # now we can run the speed tests system.time(fmatch(s, x)) system.time(fmatch(s, x)) # the cache is still valid for string matches as well system.time(fmatch(c("foo","bar","1","2"),x)) # now back to match system.time(base::match(s, x)) identical(base::match(s, x), fmatch(s, x)) # finally, some reals to match y = rnorm(1e6) s = c(y[sample(length(y), 100)], 123.567, NA, NaN) system.time(fmatch(s, y)) system.time(fmatch(s, y)) system.time(fmatch(s, y)) system.time(base::match(s, y)) identical(base::match(s, y), fmatch(s, y)) # this used to fail before 0.1-2 since nomatch was ignored identical(base::match(4L, 1:3, nomatch=0), fmatch(4L, 1:3, nomatch=0))
# some random speed comparison examples: # first use integer matching x = as.integer(rnorm(1e6) * 1000000) s = 1:100 # the first call to fmatch is comparable to match system.time(fmatch(s,x)) # but the subsequent calls take no time! system.time(fmatch(s,x)) system.time(fmatch(-50:50,x)) system.time(fmatch(-5000:5000,x)) # here is the speed of match for comparison system.time(base::match(s, x)) # the results should be identical identical(base::match(s, x), fmatch(s, x)) # next, match a factor against the table # this will require both x and the factor # to be cast to strings s = factor(c("1","1","2","foo","3",NA)) # because the casting will have to allocate a string # cache in R, we run a dummy conversion to take # that out of the equation dummy = as.character(x) # now we can run the speed tests system.time(fmatch(s, x)) system.time(fmatch(s, x)) # the cache is still valid for string matches as well system.time(fmatch(c("foo","bar","1","2"),x)) # now back to match system.time(base::match(s, x)) identical(base::match(s, x), fmatch(s, x)) # finally, some reals to match y = rnorm(1e6) s = c(y[sample(length(y), 100)], 123.567, NA, NaN) system.time(fmatch(s, y)) system.time(fmatch(s, y)) system.time(fmatch(s, y)) system.time(base::match(s, y)) identical(base::match(s, y), fmatch(s, y)) # this used to fail before 0.1-2 since nomatch was ignored identical(base::match(4L, 1:3, nomatch=0), fmatch(4L, 1:3, nomatch=0))