fix Rcpp::CharacterVector copies

DESCRIPTION

@@ -1,6 +1,6 @@
 Package: narray
 Title: Subset- And Name-Aware Array Utility Functions
-Version: 0.5.0
+Version: 0.5.1
 Author: Michael Schubert <[email protected]>
 Maintainer: Michael Schubert <[email protected]>
 Description: Stacking arrays according to dimension names, subset-aware

NEWS.md

@@ -1,4 +1,6 @@
-# git head
+# narray 0.5.1
+
+ * Fixed a bug where `stack` made unnecessary vector copies
 
 # narray 0.5.0
 
@@ -27,16 +29,19 @@
   * new `lambda` syntax (#14)
 
 # narray 0.3.2
+
   * add `collect` function as opposite to `mask`
   * deprecate `summarize` in favor of `map`
   * adjust tests for `testthat>=2`
 
 # narray 0.2.2
+
   * fix bug where `split` with `NA` in `subsets` caused wrong splitting; these
     are now dropped with a warning (#5)
   * fix bug where vectors are not bound along the right dimensions (#7)
   * add `rep` functions for arrays (and `rrep` and `crep` aliases for rows and
     columns, respectively)
 
 # narray 0.1.1
+
   * Initial release on CRAN

src/stack.cpp

@@ -5,10 +5,10 @@ using namespace Rcpp;
 using namespace std;
 
 template<int RTYPE> Vector<RTYPE> cpp_stack_impl(List array_list, int along, Vector<RTYPE> fill, bool ovr) {
-    auto dimnames = vector<CharacterVector>(along); // dim: names along
+    auto dimnames = vector<vector<String>>(along); // dim: names along
     auto axmap = vector<unordered_map<string, int>>(along); // dim: element name->index
     auto ax_unnamed = vector<int>(along); // counter for unnamed dimension elements
-    auto a2r = vector<vector<vector<int>>>(array_list.size()); // array > dim > element
+    auto a2r = vector<vector<vector<int>>>(array_list.size()); // index array>dim>element
 
     // create lookup tables for all present dimension names
     for (int ai=0; ai<Rf_xlength(array_list); ai++) { // array index
@@ -46,12 +46,16 @@ template<int RTYPE> Vector<RTYPE> cpp_stack_impl(List array_list, int along, Vec
             } else {
                 auto dni = as<vector<string>>(dn[d]);
                 for (int e=0; e<da[d]; e++) { // element in dimension
-                    if (axmap[d].count(dni[e]) == 0) {
-                        axmap[d].emplace(dni[e], axmap[d].size() + ax_unnamed[d]);
+                    auto it = axmap[d].find(dni[e]);
+                    if (it == axmap[d].end()) {
+                        int val = axmap[d].size() + ax_unnamed[d];
+                        axmap[d].emplace(dni[e], val);
                         dimnames[d].push_back(dni[e]);
+                        a2r[ai][d].push_back(val);
+                    } else {
+                        a2r[ai][d].push_back(it->second);
                     }
 //                    Rprintf("array %i dim %i: %s -> %i\n", ai, d, dni[e].c_str(), axmap[d][dni[e]]);
-                    a2r[ai][d].push_back(axmap[d][dni[e]]);
                 }
             }
         }
@@ -73,10 +77,13 @@ template<int RTYPE> Vector<RTYPE> cpp_stack_impl(List array_list, int along, Vec
     auto rdnames = List(dimnames.size());
     for (int i=0; i<dimnames.size(); i++) {
         rdim[i] = dimnames[i].size();
-        if (all(is_na(dimnames[i])))
+        auto rdni = CharacterVector(dimnames[i].size());
+        for (int j=0; j<rdni.size(); j++)
+            rdni[j] = dimnames[i][j];
+        if (all(is_na(rdni)))
             rdnames[i] = R_NilValue;
         else
-            rdnames[i] = dimnames[i];
+            rdnames[i] = rdni;
     }
     auto n = accumulate(rdim.begin(), rdim.end(), 1, multiplies<int>());
     auto result = Vector<RTYPE>(n, fill[0]);

tests/testthat/test-stack.r

@@ -107,13 +107,22 @@ test_that("keep_empty arg when stacking zero-length vectors", {
 test_that("performance", {
     skip_on_cran()
 
-    size = 500 # 500x500, 500 arrays
+    # stack 500 arrays, 500x500 with overwriting
+    size = 500
     syms = c(letters, LETTERS, 0:9)
     idx = do.call(paste0, expand.grid(syms, syms))
-
     ars = replicate(size, simplify=FALSE,
                     matrix(runif(size*size), nrow=size, ncol=size,
                            dimnames=list(sample(idx, size), sample(idx, size))))
     tt = system.time(stack(ars, along=2, allow_overwrite=TRUE))
-    expect_lt(tt["user.self"], 6)
+    expect_lt(tt["user.self"], 6) # 1.5 sec locally
+
+    # stack 10 arrays, 10k rows and 1 column
+    size = 5e4
+    idx = do.call(paste0, expand.grid(syms, syms, syms))[1:size]
+    ars2 = replicate(10, simplify=FALSE,
+                     matrix(runif(size), nrow=size, ncol=1,
+                            dimnames=list(sample(idx, size), sample(idx,1))))
+    tt = system.time(stack(ars2, along=2))
+    expect_lt(tt["user.self"], 2) # 0.1 sec locally
 })