NEWS

---

SparseArray 1.12.0

NEW FEATURES

• Subassignment of an SVT_SparseArray object with a right value that is also an SVT_SparseArray object is finally supported. For example this allows things like x[3:6, , 1] <- y x[ , 11:15, 1] <- x[ , 1:5, 3] (where 'x' and 'y' are both SVT_SparseArray objects) to operate natively on the sparse representation and therefore to be very fast and very memory-efficient.
• NaArray objects now support subassignment by an N-index, not just by an L-index or M-index. Like SVT_SparseArray objects, the replacement value (a.k.a. right value) can be a "short" vector, an ordinary array, or another NaArray object.

SIGNIFICANT USER-VISIBLE CHANGES

• A significant refactoring of the code behind subsetting and subassignment of an SVT_SparseArray object results in code that is simpler, cleaner, and faster. In particular, the following speed improvements were obtained compared to SparseArray 1.10.0: - N-dimensional subsetting of an SVT_SparseArray object is 2 to 5 times faster. - Subassignment of an SVT_SparseArray object by a matrix subscript (a.k.a. M-index) is twice faster and uses about 20% less memory. - Subassignment of an SVT_SparseArray object with a "short" vector on the right e.g. svt[i, j, k] <- 0 or svt[i, j, k] <- c(4, 0, -1, 1) where each of i, j, k is either missing or a vector of subscripts is faster.
• Revisited and optimized implementation of the extract_array() method for SVT_SparseArray arrays. Old implementation was doing: as.array(.subset_SVT_as_SVT(x, index)) New implementation is based on new internal helper .subset_SVT_as_Rarray() which avoids the intermediate SVT_SparseArray representation. With these improvements, extract_array() is now 3 to 4 times faster than in SparseArray 1.10.0. As a consequence extracting a single row from an SVT_SparseMatrix object 'x' with x[i, ] is now 15% to 20% faster.
• The extract_array() methods for SVT_SparseArray and NaArray objects are now multithreaded when the number of rows to extract is "small".
• Speed up nzvals() on SVT_SparseMatrix objects and nnavals() on NaArray objects by overriding the default methods with specific methods implemented in C. This yields a 15x to 20x speedup.

BUG FIXES

• Fix col/rowMedians() methods for SparseArray objects.
• Fix mishandling of NaN elements in is_nonna(), nnacount(), nnawhich(), and nnavals() default methods. Clarify the meaning of non-NA element in the man page for these functions.

SparseArray 1.10.0

NEW FEATURES

• Add the two following capabilities to subassignment of SVT_SparseArray objects: - The right value (a.k.a. replacement value) now can be an ordinary array. - Using a long linear subscript (i.e. long integer or numeric vector) is now supported. Note that this is also enabled for NaArray objects.

SparseArray 1.8.0

NEW FEATURES

• Make rowsum() and colsum() work on any dsparseMatrix derivative and not just on a dgCMatrix object. For example now they also accept a dgRMatrix or dgTMatrix object.
• Make crossprod(), tcrossprod(), and %*% work between a SparseMatrix object and an ordinary vector.
• Arith operations between a SparseArray (or NaArray) object 'x' and an atomic vector 'y' are no longer restricted to the latter being a single value: now the length of the latter can also be 'dim(x)[[1]]' or a divisor of it.

BUG FIXES

• Make sure that SparseArray(x) and SVT_SparseArray(x) always propagate the dimnames.
• Fix bug in Math and Math2 operations on SVT_SparseArray and NaArray objects.

SparseArray 1.6.0

NEW FEATURES

• Linear subsetting of a SVT_SparseArray object e.g. svt[11:13].
• Add the 'dim' and 'dimnames' arguments to SVT_SparseArray() constructor function. Main use case is to make it easy and efficient to construct a zero-only SVT_SparseArray object of arbitrary dimensions.
• Add is_nonzero() generic with a default method and fast methods for SVT_SparseArray, COO_SparseArray, and sparseMatrix objects.
• The Sparse Vector Tree of a SVT_SparseArray can have "lacunar leaves", that is, leaves where the nzvals are missing. These leaves are semantically equivalent to leaves where all the values in nzvals are ones. This reduces the memory footprint of a "logical" SVT_SparseArray with no NAs by half. Note that a "logical" SVT_SparseMatrix object with no NAs is similar to an ngCMatrix object from the Matrix package. ngCMatrix objects don't store any non-zero values either (they're implicitly considered to be TRUEs), only their offsets.
• col/row summarization methods (a.k.a. "matrixStats methods") now work on COO_SparseArray objects (in addition to SVT_SparseArray objects).
• Add row/colSums2() and row/colMeans2() methods for SparseArray objects.
• rowsum(), crossprod(), tcrossprod(), and %*% methods now work on COO_SparseMatrix objects (in addition to SVT_SparseMatrix objects).
• Add fast colsum() methods for SparseMatrix and dgCMatrix objects.
• Add is.na(), is.nan(), and is.infinite() methods for SVT_SparseArray objects.
• Add pmin() and pmax() methods for SparseArray objects.
• aperm(SVT_SparseArray) now supports S4Arrays::aperm2() extended semantic. See '?S4Arrays::aperm2' for details.
• Add 'dimnames' argument to randomSparseArray(), poissonSparseArray(), randomSparseMatrix(), and poissonSparseMatrix().
• Coercions from dgTMatrix, lgTMatrix, or ngTMatrix, to COO_SparseMatrix/Array or SVT_SparseMatrix/Array, and vice versa.
• Coercions from ng[C|R|T]Matrix to COO_SparseMatrix, and vice versa.
• Coercion from ngCMatrix to SVT_SparseMatrix, and vice versa.
• NaArray objects: WORK-IN-PROGRESS! New objects that use the same internal representation as SVT_SparseArray objects (Sparse Vector Tree), but background value is NA instead of zero. See https://github.com/fmicompbio/footprintR/issues/7 for the motivating use case. They support most operations supported by SVT_SparseArray objects: [, [<-, t(), aperm(), cbind(), rbind(), abind(), matrixStats operations (col*(), row*()), etc... What's missing: - a dedicated vignette; - some row*() functions are not ready yet.
• Add new generics is_nonna(), nnacount(), nnawhich(), nnavals(), and 'nnavals<-'(), with default methods and methods for NaArray objects.

SIGNIFICANT USER-VISIBLE CHANGES

• Refactored subassignment of a SVT_SparseArray object, which resulted in significant speed improvement and memory footprint reduction.
• Speed up some row summarization methods (e.g. rowSums(x)) by implementing them natively in C rather than doing a transposition followed by a column summarization (e.g. colSums(t(x))). This avoids the costly transposition step which is very time and memory consuming. Row summarization methods with native C implementation so far: rowAnyNAs() (35x speedup), rowMins() (6x speedup), rowMaxs() (6x speedup), rowSums() (20x speedup), rowMeans() (18x speedup), rowVars() (5x speedup), and rowSds() (5x speedup). More methods will follow.
• Special-case 'lambda=0' in poissonSparseArray() so the empty SVT_SparseArray object gets returned instantaneously.

BUG FIXES

• Fix integer overflow in nzwhich() methods for CsparseMatrix and RsparseMatrix objects when the object has a length >= 2^31.
• Fix bug in coercion from COO_SparseMatrix to [d|l]gCMatrix or [d|l]gRMatrix when the 'nzcoo' slot of the COO_SparseMatrix object contains duplicates.
• Make sure that coercion from CsparseMatrix to SVT_SparseMatrix works on any [d|l|n]gCMatrix **derivative** and not just on a [d|l|n]gCMatrix **instance**.

SparseArray 1.4.0

NEW FEATURES

• The following operation are now multithreaded (via OpenMP): - SparseMatrix multiplication and crossprod() - the matrixStats methods for SparseMatrix objects See ?set_SparseArray_nthread for more information.
• Define some of the basic coercion methods that used to be defined in the Matrix package but that the lazy Matrix maintainers have decided to deprecate in Matrix 1.7-0 (e.g. coercion from matrix to dgRMatrix).

BUG FIXES

• Fix aperm(SVT_SparseArray) when inner dimensions are not permuted See https://github.com/Bioconductor/SparseArray/issues/6
• Fix long-standing bug in C function _dotprod_leaf_vector_and_double_col(). In some circumstances the function was reading beyond a C array. Observed effects varied from no observed effects (on Linux) to breaking crossprod() unit tests on arm64 Mac (consistently), on Windows (sporadically), and also probably on PowerPC (or possibly a crash sometimes on this platform. See https://github.com/Bioconductor/SparseArray/issues/2

SparseArray 1.2.0

NEW FEATURES

• Add aperm() method for SVT_SparseArray objects.
• Add abind() method for SparseArray objects.
• Add dim() setter for SVT_SparseArray objects.
• Introduce nzwhich() generic and method for SparseArray derivatives. Also provide a default method for ordinary arrays and other array-like objects.
• Implement 'Logic' ops on SVT_SparseArray objects.
• Implement 'Math'/'Math2' ops on SVT_SparseArray objects of type "double".
• 'Compare' ops now support SVT_SparseArray objects of type() "raw" or "complex".
• All matrixStats methods (except row/colMedians()) now work on multidimensional SVT_SparseArray objects and support the 'dims' argument, like the row/colSums() and row/colMeans() functions in base R.
• Add row/colAnys() + row/colAlls() + row/colAnyNAs() + row/colProds() methods for SVT_SparseArray objects.

SIGNIFICANT USER-VISIBLE CHANGES

• Rename nzvals() slot getter (for COO_SparseArray objects) -> nzdata(). Also reintroduce nzvals() as a fast way to get 'x[nzwhich(x)]' on a sparse array-like object 'x'.
• Re-implement all matrixStats methods (except row/colMedians()) for SVT_SparseArray objects in C.

SparseArray 1.0.0

• First version of the package that is ready for general use.