NEWS.md
The hidden R options for deescalating the error for using useNames = NA to a warning has been removed; useNames = NA is now always an error.
Calling colRanks() and rowRanks() without explicitly specifying argument ties.method is deprecated since version 1.3.0 [2024-04-10]. If not explicitly specified, a deprecation warning is now produced every 10:th call not specifying the ties.method argument.
colTabulates() and rowTabulates() asserting that double values are passed, reported on the class of the input data, not the storage type.rowSums2() is now significantly faster for larger matrices.None of the error messages use a trailing period.
Addressing changes in the C API of R-devel resulted in compiler errors such as error: implicit declaration of function 'Calloc'; did you mean 'calloc'? [-Wimplicit-function-declaration].
Addressing changes in stricter compiler flags of R-devel resulted in compiler warning embedding a directive within macro arguments has undefined behavior [-Wembedded-directive].
colRanks() and rowRanks() without explicitly specifying argument ties.method is deprecated since version 1.3.0 [2024-04-10]. If not explicitly specified, a deprecation warning is now produced every 25:th call not specifying the ties.method argument.validateIndices() has been removed. It had been defunct since version 0.63.0 (2022-11-14).Calling colRanks() and rowRanks() without explicitly specifying argument ties.method will be deprecated when using R (>= 4.4.0). The reason is that the current default is ties.method = "max", but we want to change that to ties.method = "average" to align it with base::rank(). In order to minimize the risk for sudden changes in results, we ask everyone to explicitly specify their intent. The first notice will be through deprecation warnings, which will only occur every 50:th call to keep the noise level down. We will make it more noisy in future releases, and eventually also escalated to defunct errors.
Using a scalar value for argument center of colSds(), rowSds(), colVars(), rowVars(), colMads(), rowMads(), colWeightedMads(), and rowWeightedMads() is now defunct.
useNames = TRUE is the new default for all functions. For backward compatibility, it used to be useNames = NA.
colQuantiles() and rowQuantiles() gained argument digits, just like stats::quantile() gained that argument in R 4.1.0.
colQuantiles() and rowQuantiles() only sets quantile percentage names when useNames = TRUE, to align with how argument names of stats::quantile() works in base R.
colMeans2() and rowMeans2() gained argument refine. If refine = TRUE, then the sample average for numeric matrices are calculated using a two-pass scan, resulting in higher precision. The default is refine = TRUE to align it with colMeans(), but also mean2() in this package. If the higher precision is not needed, using refine = FALSE will be almost twice as fast.
colSds(), rowSds(), colVars(), and rowVars() gained argument refine. If refine = TRUE, then the sample average for numeric matrices are calculated using a two-pass scan, resulting in higher precision for the estimate of the center and therefore also the variance.
Unnecessary checks for missing indices are eliminated, yielding better performance. This change does not affect user-facing API.
Made colQuantiles() and rowQuantiles() a bit faster for type != 7L, by making sure percentage names are only generated once, instead of once per column or row.
Contrary to other functions in the package, and how it works in base R, functions colCumsums(), colCumprods(), colCummins(), colCummaxs(), colRanges(), colRanks(), and colDiffs(), plus the corresponding row-based versions, did not drop the names attribute when both row and column names were NULL. Now also these functions behaves the same as the case when neither row or column names are set.
colQuantiles() and rowQuantiles() did not generate quantile percentage names exactly the same way as stats::quantile(), which would reveal itself for certain combinations of probs and digits.
Package compiles again with older compilers not supporting the C99 standard (e.g. GCC 4.8.5 (2015), which is the default on RHEL / CentOS 7.9). This was the case also for matrixStats (<= 0.54.0).
Added more information to the error message produced when argument center for col- and rowVars() holds an invalid value.
Fix two compilation warnings on a function declaration without a prototype is deprecated in all versions of C [-Wstrict-prototypes].
colCummins(), colCummaxs(), rowCummins(), and rowCummaxs() now support also logical input.which for colOrderStats() and rowOrderStats() is out of range, the error message now reports on the value of which. Similarly, when argument probs for colQuantiles() and rowQuantiles() is out of range, the error message reports on its value too.Handling of the useNames argument is now done in the native code.
Passing idxs, rows, and cols arguments of type integer is now less efficient than it used to, because the new code re-design (see below) requires an internal allocation of an equally long R_xlen_t vector that is populated by indices coerced from R_len_t to R_xlen_t integers.
R CMD check would produce a NOTE on the package installation size being large, which no longer is the case. The downside is that extra overhead when passing integer indices (see above comment).useNames = NA in the previous release, colQuantiles() and rowQuantiles() got useNames = TRUE.useNames = TRUE. To drop them, set useNames = FALSE. To preserve the current, inconsistent behavior, set useNames = NA, which, for backward compatibility reasons, remains the default for now.meanOver() and sumOver(), and argument method from weightedVar(), that have been defunct since January 2018.colVars() and rowVars() with argument center now calculates the sample variance using the n/(n-1)*avg((x-center)^2) formula rather than the n/(n-1)*(avg(x^2)-center^2) formula that was used in the past. Both give the same result when center is the correct sample mean estimate. The main reason for this change is that, if an incorrect center is provided, in contrast to the old approach, the new approach is guaranteed to give at least non-negative results, despite being incorrect. BACKWARD COMPATIBILITY: Out of all 314 reverse dependencies on CRAN and Bioconductor, only four called these functions with argument center. All of them pass their package checks also after this update. To further protect against a negative impact in existing user scripts, colVars() and rowVars() will calculate both versions and assert that the result is the same. If not, an informative error is produced. To limit the performance impact, this validation is run only once every 50:th call, a frequency that can be controlled by R option matrixStats.vars.formula.freq. Setting it to 0 or NULL will disable the validation. The default can also be controlled by environment variable R_MATRIXSTATS_VARS_FORMULA_FREQ. This validation framework will be removed in a future version of the package after it has been established that this change has no negative impact.Now colWeightedMads() and rowWeightedMads() accept center of the same length as the number of columns and rows, respectively.
colAvgsPerRowSet() and rowAvgsPerRowSet() gained argument na.rm.
Now weightedMean() and weightedMedian() and the corresponding row- and column-based functions accept logical x, where FALSE is treated as integer 0 and TRUE as 1.
Now x_OP_y() and t_tx_OP_y() accept logical x and y, where FALSE is treated as integer 0 and TRUE as 1.
colQuantiles() and rowQuantiles() on a logical matrix should return a numeric vector for type = 7. However, when there were only missing values (= NA) in the matrix, then it would return a “logical” vector instead.
colAvgsPerRowSet() on a single-column matrix would produce an error on non-matching dimensions. Analogously, for rowAvgsPerRowSet() and single- row matrices.
colVars(x) and rowVars(x) with x being an array would give the wrong value if both argument dim. and center would be specified.
The documentation was unclear on what the center argument should be. They would not detect when an incorrect specification was used, notably when the length of center did not match the matrix dimensions. Now these functions give an informative error message when center is of the incorrect length.
center of colSds(), rowSds(), colVars(), rowVars(), colMads(), rowMads(), colWeightedMads(), and rowWeightedMads() is now deprecated.colCumprods() and rowCumprods() now support also logical input. Thanks to Constantin Ahlmann-Eltze at EMBL Heidelberg for the patch.colCollapse() and rowCollapse() did not expand idxs argument before subsetting by cols and rows, respectively. Thanks to Constantin Ahlmann-Eltze for reporting on this.
colAnys(), rowAnys(), anyValue(), colAlls(), rowAlls(), and allValue() with value=FALSE and numeric input would incorrectly consider all values different from one as FALSE. Now it is only values that are zero that are considered FALSE. Thanks to Constantin Ahlmann-Eltze for the bug fix.
colQuantiles() and rowQuantiles() now supports only integer, numeric and logical input. Previously, it was also possible to pass, for instance, character input, but that was a mistake. The restriction on input allows for further optimization of these functions.
The returned type of colQuantiles() and rowQuantiles() is now the same as for stats::quantile(), which depends on argument type.
colQuantiles() and rowQuantiles() with the default type = 7L and when there are no missing values are now significantly faster and use significantly fewer memory allocations.colDiffs() and rowDiffs() gave an error if argument dim. was of type numeric rather than type integer.
varDiff(), sdDiff(), madDiff(), iqrDiff(), and the corresponding row- and column functions silently treated a diff less than zero as diff = 0. Now an error is produced.
Error messages on argument dim. referred to non-existing argument dim.
Error messages on negative values in argument dim. reported a garbage value instead of the negative value.
The Markdown reports produced by the internal benchmark report generator did not add a line between tables and the following text (a figure caption) causing the following text to be included in a cell on an extra row in the table (at least when rendered on GitHub Wiki pages).
weightedVar(), weightedSd(), weightedMad(), and their row- and column- specific counter parts now return a missing value if there are missing values in any of the weights w after possibly dropping (x, w) elements with missing values in x (na.rm = TRUE). Previously, na.rm = TRUE would also drop (x, w) elements where w was missing. With this change, we now have that for all functions in this package, na.rm = TRUE never applies to weights - only x values.colRanks() and rowRanks() now supports the same set of ties.method as base::rank() plus "dense" as defined by data.table::frank(). For backward compatible reasons, the default ties.method remains the same as in previous versions. Thank to Brian Montgomery for contributing this.
colCumsums() and rowCumsums() now support also logical input.
weightedVar(), weightedSd(), weightedMad(), and their row- and column- specific counter parts would produce an error instead of returning a missing value when one of the weights is a missing value.stopifnot() for internal validation, because it comes with a great overhead. This was only used in weightedMad(), col-, and rowWeightedMads(), as well as col- and rowAvgsPerColSet().colLogSumExps(lx, cols) / rowLogSumExps(lx, rows) gave an error if lx has rownames / colnames.
col-/rowQuantiles() would lose rownames of output in certain cases.
Functions sum2(x) and means2(x) now accept also logical input x, which corresponds to using as.integer(x) but without the need for neither coercion nor internal extra copies. With sum2(x, mode = "double") it is possible to count number of TRUE elements beyond 2^31-1, which base::sum() does not support.
Functions col-/rowSums2() and col-/rowMeans2() now accept also logical input x.
Function binMeans(y, x, bx) now accepts logical y, which corresponds to to using as.integer(y), but without the need for coercion to integer.
Functions col-/rowTabulates(x) now support logical input x.
Now count() can count beyond 2^31-1.
allocVector() can now allocate long vectors (longer than 2^31-1).
Now sum2(x, mode = "integer") generates a warning if typeof(x) == "double" asking if as.integer(sum2(x)) was intended.
Inspired by Hmisc::wtd.var(), when sum(w) <= 1, weightedVar(x, w) now produces an informative warning that the estimate is invalid.
colAvgsPerColSet() with that of rowAvgsPerColSet().col-/rowLogSumExp() could core dump R for “large” number of columns/rows. Thanks Brandon Stewart at Princeton University for reporting on this.
count() beyond 2^31-1 would return invalid results.
Functions col-/rowTabulates(x) did not count missing values.
indexByRow(dim, idxs) would give nonsense results if idxs had indices greater than prod(dim) or non-positive indices; now it gives an error.
indexByRow(dim) would give nonsense results when prod(dim) >= 2^31; now it gives an informative error.
col-/rowAvgsPerColSet() would return vector rather than matrix if nrow(X) <= 1. Thanks to Peter Hickey (Johns Hopkins University) for troubleshooting and providing a fix.
Previously deprecated meanOver() and sumOver() are defunct. Use mean2() and sum2() instead.
Previously deprecated weightedVar(x, w, method = "0.14.2") is defunct.
Dropped previously defunct weightedMedian(..., ties = "both").
Dropped previously defunct argument centers for col-/rowMads(). Use center instead.
Dropped previously defunct argument flavor of colRanks() and rowRanks().
Since anyNA() is a built-in function since R (>= 3.1.0), please use that instead of anyMissing() part of this package. The latter will eventually be deprecated. For consistency with the anyNA() name, colAnyNAs() and rowAnyNAs() are now also available replacing the identically colAnyMissings() and rowAnyMissings() functions, which will also be deprecated in a future release.
meanOver() was renamed to mean2() and sumOver() was renamed to sum2().
Added colSums2() and rowSums2() which work like colSums() and rowSums() of the base package but also supports efficient subsetting via optional arguments rows and cols.
Added colMeans2() and rowMeans2() which work like colMeans() and rowMeans() of the base package but also supports efficient subsetting via optional arguments rows and cols.
Functions colDiffs() and rowDiffs() gained argument dim..
Functions colWeightedMads() and rowWeightedMads() gained arguments constant and center. The current implementation only support scalars for these arguments, which means that the same values are applied to all columns and rows, respectively. In previous version a hard-to-understand error would be produced if center was of length greater than one; now an more informative error message is given.
Package is now silent when loaded; it no longer displays a startup message.
Continuous-integration testing is now also done on macOS, in addition to Linux and Windows.
ROBUSTNESS: Package now registers the native API using also R_useDynamicSymbols().
Cleaned up native low-level API and renamed native source code files to make it easier to navigate the native API.
Now using roxygen2 for help and NAMESPACE (was R.oo::Rdoc).
rowAnys(x) on numeric matrices x would return rowAnys(x == 1) and not rowAnys(x != 0). Same for colAnys(), rowAlls(), and colAlls(). Thanks Richard Cotton for reporting on this.
sumOver(x) and meanOver(x) would incorrectly return -Inf or +Inf if the intermediate sum would have that value, even if one of the following elements would turn the intermediate sum into NaN or NA, e.g. with x as c(-Inf, NaN), c(-Inf, +Inf), or c(+Inf, NA).
WORKAROUND: Benchmark reports generated by matrixStats:::benchmark() would use any custom R prompt that is currently set in the R session, which may not render very well. Now it forces the prompt to be the built-in "> " one.
The package API is only intended for matrices and vectors of type numeric, integer and logical. However, a few functions would still return if called with a data.frame. This was never intended to work and is now an error. Specifically, functions colAlls(), colAnys(), colProds(), colQuantiles(), colIQRs(), colWeightedMeans(), colWeightedMedians(), and colCollapse() now produce warnings if called with a data.frame. Same for the corresponding row- functions. The use of a `data.frame will be produce an error in future releases.
meanOver() and sumOver() are deprecated because they were renamed to mean2() and sum2(), respectively.
Previously deprecated (and ignored) argument flavor of colRanks() and rowRanks() is now defunct.
Previously deprecated support for passing non-vector, non-matrix objects to rowAlls(), rowAnys(), rowCollapse(), and the corresponding column-based versions are now defunct. Likewise, rowProds(), rowQuantiles(), rowWeightedMeans(), rowWeightedMedians(), and the corresponding column-based versions are also defunct. The rationale for this is to tighten up the identity of the matrixStats package and what types of input it accepts. This will also help optimize the code further.
SPEEDUP / CLEANUP: rowMedians() and colMedians() are now plain functions. They were previously S4 methods (due to a Bioconductor legacy). The package no longer imports the methods package.
SPEEDUP: Now native API is formally registered allowing for faster lookup of routines from R.
Package now installs on R (>= 2.12.0) as claimed. Thanks to Mikko Korpela at Aalto University School of Science, Finland, for troubleshooting and providing a fix.
logSumExp(c(-Inf, -Inf, ...)) would return NaN rather than -Inf. Thanks to Jason Xu (University of Washington) for reporting and Brennan Vincent for troubleshooting and contributing a fix.
idxs, rows and cols were added to all functions such that the calculations are performed on the requested subset while avoiding creating a subsetted copy, i.e. rowVars(x, cols = 4:6) is a much faster and more memory efficient version than rowVars(x[, 4:6]) and even yet more efficient than apply(x, MARGIN = 1L, FUN = var). These features were added by Dongcan Jiang, Peking University, with support from the Google Summer of Code program. A great thank you to Dongcan and to Google for making this possible.w and W) default to NULL, which corresponds to uniform weights.weightedVar(x, w) used the wrong bias correction factor resulting in an estimate that was tau too large, where tau = ((sum(w) - 1) / sum(w)) / ((length(w) - 1) / length(w)). Thanks to Wolfgang Abele for reporting and troubleshooting on this.
weightedVar(x) with length(x) = 1 returned 0 - not NA. Same for weightedSd().
weightedMedian(x, w = NA_real_) returned x rather than NA_real_. This only happened for length(w) = 1.
allocArray(dim) failed for prod(dim) >= .Machine$integer.max.
CLEANUP: Defunct argument centers for col-/rowMads(); use center.
weightedVar(x, w, method = "0.14.2") is deprecated.
x_OP_y() and t_tx_OP_y() would return garbage on Solaris SPARC (and possibly other architectures as well) when input was integer and had missing values.product(x, na.rm = FALSE) for integer x with both zeros and NAs returned zero rather than NA.
weightedMean(x, w, na.rm = TRUE) did not handle missing values in x properly, if it was an integer. It would also return NaN if there were weights w with missing values, whereas stats::weighted.mean() would skip such data points. Now weightedMean() does the same.
(col|row)WeightedMedians() did not handle infinite weights as weightedMedian() does.
x_OP_y(x, y, OP, na.rm = FALSE) returned garbage iff x or y had missing values of type integer.
rowQuantiles() and rowIQRs() did not work for single-row matrices. Analogously for the corresponding column functions.
rowCumsums(), rowCumprods() rowCummins(), and rowCummaxs(), accessed out-of-bound elements for Nx0 matrices where N > 0. The corresponding column methods has similar memory errors for 0xK matrices where K > 0.
anyMissing(list(NULL)) returned NULL; now FALSE.
rowCounts() resulted in garbage if a previous column had NAs (because it forgot to update index kk in such cases).
rowCumprods(x) handled missing values and zeros incorrectly for integer x (not double); a zero would trump an existing missing value causing the following cumulative products to become zero. It was only a zero that trumped NAs; any other integer would work as expected. Note, this bug was not in colCumprods().
rowAnys(x, value, na.rm = FALSE) did not handle missing values in a numeric x properly. Similarly, for non-numeric and non-logical x, row- and colAnys(), row- and colAlls(), anyValue() and allValue() did not handle when value was a missing value.
All of the above bugs were identified and fixed by Dongcan Jiang (Peking University, China), who also added corresponding unit tests.
anyMissing() is no longer an S4 generic. This was done as part of the migration of making all functions of matrixStats plain R functions, which minimizes calling overhead and it will also allow us to drop methods from the package dependencies. I’ve scanned all CRAN and Bioconductor packages depending on matrixStats and none of them relied on anyMissing() dispatching on class, so hopefully this move has little impact. The only remaining S4 methods are now colMedians() and rowMedians().SPEEDUP: The check for user interrupts at the C level is now done less frequently of the functions. It does every k:th iteration, where k = 2^20, which is tested for using (iter % k == 0). It turns out, at least with the default compiler optimization settings that I use, that this test is 3 times faster if k = 2^n where n is an integer. The following functions checks for user interrupts: logSumExp(), (col|row)LogSumExps(), (col|row)Medians(), (col|row)Mads(), (col|row)Vars(), and (col|row)Cum(Min|Max|prod|sum)s().
SPEEDUP: logSumExp(x) is now faster if x does not contain any missing values. It is also faster if all values are missing or the maximum value is +Inf - in both cases it can skip the actual summation step.
all() and any() flavored methods on non-numeric and non-logical (e.g. character) vectors and matrices with na.rm = FALSE did not give results consistent with all() and any() if there were missing values. For example, with x <- c("a", NA, "b") we have all(x == "a") == FALSE and any(x == "a") == TRUE, whereas our corresponding methods would return NA in those cases. The methods fixed are allValue(), anyValue(), col-/rowAlls(), and col-/rowAnys(). Added more package tests to cover these cases.
logSumExp(x, na.rm = TRUE) would return NA if all values were NA and length(x) > 1. Now it returns -Inf for all length(x):s.
diff2() with differences >= 3 would read spurious values beyond the allocated memory. This error, introduced in 0.13.0, was harmless in the sense that the returned value was unaffected and still correct. Thanks to Brian Ripley and the CRAN check tools for catching this. I could reproduce it locally with valgrind.anyMissing() and rowMedians().Added weightedMean(), which is ~10 times faster than stats::weighted.mean().
Added count(x, value) which is a notably faster than sum(x == value). This can also be used to count missing values etc.
Added allValue() and anyValue() for all(x == value) and any(x == value).
Added diff2(), which is notably faster than base::diff() for vectors, which it is designed for.
Added iqrDiff() and (col|row)IqrDiffs().
CONSISTENCY: Now rowQuantiles(x, na.rm = TRUE) returns all NAs for rows with missing values. Analogously for colQuantiles(), colIQRs(), rowIQRs() and iqr(). Previously, all these functions gave an error saying missing values are not allowed.
COMPLETENESS: Added corresponding “missing” vector functions for already existing column and row functions. Similarly, added “missing” column and row functions for already existing vector functions, e.g. added iqr() and count() to complement already existing (col|row)IQRs() and (col|row)Counts() functions.
ROBUSTNESS: Now column and row methods give slightly more informative error messages if a data.frame is passed instead of a matrix.
SPEEDUP: (col|row)Diffs() are now implemented in native code and notably faster than diff() for matrices.
SPEEDUP: Made binCounts() and binMeans() a bit faster.
SPEEDUP: Implemented weightedMedian() in native code, which made it ~3-10 times faster. Dropped support for ties = "both", because it would have to return two values in case of ties, which made the API unnecessarily complicated. If really needed, then call the function twice with ties = "min" and ties = "max".
SPEEDUP: (col|row)Anys() and (col|row)Alls() is now notably faster compared to previous versions.
anyMissing() into a plain R function, the specific anyMissing() implementations for data.frame:s and and list:s were dropped and is now handled by anyMissing() for "ANY", which is the only S4 method remaining now. In a near future release, this remaining "ANY" method will turned into a plain R function and the current S4 generic will be dropped. We know of no CRAN and Bioconductor packages that rely on it being a generic function. Note also that since R (>= 3.1.0) there is a base::anyNA() function that does the exact same thing making anyMissing() obsolete.product() on integer vector incorrectly used C-level abs() on intermediate values despite those being doubles requiring fabs(). Despite this, the calculated product would still be correct (at least when validated on several local setups as well as on the CRAN servers). Again, thanks to Brian Ripley for pointing out another invalid integer-double coercion at the C level.(col|row)Cumsums(x) where x is integer would return garbage for columns (rows) containing missing values.
rowMads(x) where x is numeric (not integer) would give incorrect results for rows that had an odd number of values (no ties). Analogously issues with colMads(). Added package tests for such cases too. Thanks to Brian Ripley and the CRAN check tools for (yet again) catching another coding mistake. Details: This was because the C-level calculation of the absolute value of residuals toward the median would use integer-based abs() rather than double-based fabs(). Now it fabs() is used when the values are double and abs() when they are integers.
SPEEDUP: Implemented (col|row)Mads(), (col|row)Sds(), and (col|row)Vars() in native code.
SPEEDUP: Made (col|row)Quantiles(x) faster for x without missing values (and default type = 7L quantiles). It should still be implemented in native code though.
SPEEDUP: Made rowWeightedMeans() faster.
(col|row)Medians(x) when x is integer would give invalid median values in case (a) it was calculated as the mean of two values (“ties”), and (b) the sum of those values where greater than .Machine$integer.max. Now such ties are calculated using floating point precision. Add lots of package tests.SPEEDUP: Now (col|row)Mins(), (col|row)Maxs(), and (col|row)Ranges() are implemented in native code providing a significant speedup.
SPEEDUP: Now colOrderStats() also is implemented in native code, which indirectly makes colMins(), colMaxs() and colRanges() faster.
SPEEDUP: colTabulates(x) no longer uses rowTabulates(t(x)).
SPEEDUP: colQuantiles(x) no longer uses rowQuantiles(t(x)).
SPEEDUP: Now colCollapse(x) no longer utilizes rowCollapse(t(x)). Added package tests for (col|row)Collapse().
SPEEDUP: Now colDiffs(x) no longer uses rowDiffs(t(x)). Added package tests for (col|row)Diffs().
SPEEDUP: Package no longer utilizes match.arg() due to its overhead; methods sumOver(), (col|row)Prods() and (col|row)Ranks() were updated.
dim. For instance, rowCounts(x, dim = c(nrow, ncol)) is the same as rowCounts(matrix(x, nrow, ncol)), but more efficient since it avoids creating/allocating a temporary matrix.colCounts() is implemented in native code. Moreover, (col|row)Counts() are now also implemented in native code for logical input (previously only for integer and double input). Added more package tests and benchmarks for these functions.sdDiff(), madDiff(), varDiff(), weightedSd(), weightedVar() and weightedMad() into plain functions (were generic functions).indexByRow() in native code and it is no longer a generic function, but a regular function, which is also faster to call. The first argument of indexByRow() has been changed to dim such that one should use indexByRow(dim(X)) instead of indexByRow(X) as in the past. The latter form is still supported, but deprecated.allocVector(), allocMatrix(), and allocArray() for faster allocation numeric vectors, matrices and arrays, particularly when filled with non-missing values.Better support for long vectors.
PRECISION: Using greater floating-point precision in more internal intermediate calculations, where possible.
binCounts() and binMeans() it is possible that a bin gets a higher count than what can be represented by an R integer (.Machine$integer.max = 2^31-1). If that happens, an informative warning is generated and the bin count is set to .Machine$integer.max. If this happens for binMeans(), the corresponding mean is still properly calculated and valid..Call() and takes care of most of the argument validation and construction of the return value. This function dispatch to functions in the low-level API based on data type(s) and other arguments. The low-level API is written to work with basic C data types only.sumOver() and meanOver(), which are notably faster versions of sum(x[idxs]) and mean(x[idxs]). Moreover, instead of having to do sum(as.numeric(x)) to avoid integer overflow when x is an integer vector, one can do sumOver(x, mode = "numeric"), which avoids the extra copy created when coercing to numeric (this numeric copy is also twice as large as the integer vector). Added package tests and benchmark reports for these functions.SPEEDUP: Made anyMissing(), logSumExp(), (col|row)Medians(), and (col|row)Counts() slightly faster by making the native code assign the results directly to the native vector instead of to the R vector, e.g. ansp[i] = v where ansp = REAL(ans) instead of REAL(ans)[i] = v.
Added benchmark reports for anyMissing() and logSumExp().
binMeans() returned 0.0 instead of NA_real_ for empty bins.Added benchmark reports for also non-matrixStats functions col-/rowSums() and col-/rowMeans().
Now all colNnn() and rowNnn() methods are benchmarked in a combined report making it possible to also compare colNnn(x) with rowNnn(t(x)).
Relaxed some packages tests such that they assert numerical correctness via all.equal() rather than identical().
Submitted to CRAN.
product() incorrectly assumed that the value of prod(c(NaN, NA)) is uniquely defined. However, as documented in help("is.nan"), it may be NA or NaN depending on R system/platform.signTabulate() for tabulating the number of negatives, zeros, positives and missing values. For doubles, the number of negative and positive infinite values are also counted.SPEEDUP: Now col-/rowProds() utilizes new product() function.
SPEEDUP: Added product() for calculating the product of a numeric vector via the logarithm.
SPEEDUP: Made weightedMedian() a plain function (was an S3 method).
CLEANUP: Now only exporting plain functions and generic functions.
SPEEDUP: Turned more S4 methods into S3 methods, e.g. rowCounts(), rowAlls(), rowAnys(), rowTabulates() and rowCollapse().
method to col-/rowProds() for controlling how the product is calculated.SPEEDUP: Package is now byte compiled.
SPEEDUP: Made rowProds() and rowTabulates() notably faster.
SPEEDUP: Now rowCounts(), rowAnys(), rowAlls() and corresponding column methods can search for any value in addition to the default TRUE. The search for a matching integer or double value is done in native code, which is notably faster (and more memory efficient because it avoids creating any new objects).
SPEEDUP: Made colVars() and colSds() notably faster and rowVars() and rowSds() a slightly bit faster.
Added benchmark reports, e.g. matrixStats:::benchmark("colMins").
indexByRow(), madDiff(), sdDiff() and varDiff().binMeans(x, bx) would try to access an out-of-bounds value of argument y iff x contained elements that are left of all bins in bx. This bug had no impact on the results and since no assignment was done it should also not crash/core dump R. This was discovered thanks to new memtests (ASAN and valgrind) provided by CRAN.rowProds() would throw "Error in rowSums(isNeg) :xmust be an array of at least two dimensions" on matrices where all rows contained at least one zero. Thanks to Roel Verbelen at KU Leuven for the report.weighedVar() and weightedSd().right to binCounts() and binMeans() to specify whether binning should be done by (u,v] or [u,v). Added system tests validating the correctness of the two cases.anyMissing() everywhere possible.ROBUSTNESS: Now importing loadMethod from methods package such that matrixStats S4-based methods also work when methods is not loaded, e.g. when Rscript is used, cf. Section ‘Default packages’ in ‘R Installation and Administration’.
ROBUSTNESS: Updates package system tests such that the can run with only the base package loaded.
help("rowQuantiles").(col|row)Tabulates() by replacing rm() calls with NULL assignments.binCounts() and binMeans() now uses Hoare’s Quicksort for presorting x before counting/averaging. They also no longer test in every iteration (== for every data point) whether the last bin has been reached or not, but only after completing a bin.logSumExp() used an invalid check for missing value of an integer argument. Detected by Brian Ripley upon CRAN submission.preserveShape to colRanks(). For backward compatibility the default is preserveShape = FALSE, but it may change in the future.ties.method to rowRanks() and colRanks(), but still only support for "max" (as before).anyMissing() for data type raw, which always returns FALSE.ROBUSTNESS: Added system test for anyMissing().
ROBUSTNESS: Now S3 methods are declared in the namespace.
In some cases binCounts() and binMeans() could try to go past the last bin resulting a core dump.
binCounts() and binMeans() would return random/garbage values for bins that were beyond the last data point.
Added binMeans() for fast sample-mean calculation in bins. Thanks to Martin Morgan at the Fred Hutchinson Cancer Research Center, Seattle, for contributing the core code for this.
Added binCounts() for fast element counting in bins.
Added weightedMad() from aroma.core v2.5.0.
Added weightedMedian() from aroma.light v1.25.2.
colMads() would return the incorrect estimates. This bug was introduced in matrixStats v0.4.0 (2011-11-11).rowAvgsPerColSet() and colAvgsPerRowSet().Added help pages with an example to rowIQRs() and colIQRs().
Added example to rowQuantiles().
rowRanks() and colRanks(). Thanks Hector Corrada Bravo (University of Maryland) and Harris Jaffee (John Hopkins).colMedians(x) no longer uses rowMedians(t(x)); instead there is now an optimized native-code implementation. Also, colMads() utilizes the new colMedians() directly. This improvement was kindly contributed by Harris Jaffee at Biostatistics of John Hopkins, USA.colMedians() and rowMedians().colRanges(x) would return a matrix of wrong dimension if x did not have any missing values. This would affect all functions relying on colRanges(), e.g. colMins() and colMaxs(). Added a redundancy test for this case. Thanks Pierre Neuvial at UC Berkeley for reporting this.
(col|row)Ranges() return a matrix with dimension names.
"%#x" in rowTabulates() when creating the column names of the result matrix. It gave an error OSX with R v2.9.0 devel (2009-01-13 r47593b) current the OSX server at R-forge.rowWeightedMedians() to run conditionally on aroma.light, which is only a suggested package - not a required one. This in order to prevent R CMD check to fail on CRAN, which prevents it for building binaries (as it currently happens on their OSX servers).rowOrderStats(), the stack would not become UNPROTECTED before calling error.rowCollapse() was broken and returned the wrong elements.Added (col|row)Collapse().
Added indexByRow().
Added (col|row)OrderStats().
Added (col|row)Ranges() and (col|row)(Min|Max)s().
Added colMedians().
Now anyMissing() support most data types as structures.