[libcxx-commits] [libcxx] Libcxx nth element small ranges (PR #215698)

[Louis Dionne via libcxx-commits]

================
@@ -84,6 +84,55 @@ int main(int argc, char** argv) {
     register_bm(support::single_element_data<int>, "repeated");
   }
 
+  // Benchmark std::nth_element on many independent fixed-size windows. This models
+  // the median-of-3 / median-of-5 filter use case from #214238, where the window
+  // size is a compile-time constant and the specialized sorting-network path is
+  // expected to constant-fold away the general algorithm.
+  {
+    auto bm = []<std::size_t Window>(std::string name, auto generate_data) {
+      benchmark::RegisterBenchmark(
+          name,
+          [generate_data](auto& st) {
+            std::size_t const num_windows   = st.range(0);
+            constexpr std::size_t BatchSize = 32;
+            std::vector<int> flat           = generate_data(num_windows * Window);
+            std::array<std::vector<std::array<int, Window>>, BatchSize> batches;
+            for (std::size_t b = 0; b != BatchSize; ++b) {
+              batches[b].resize(num_windows);
+              for (std::size_t i = 0; i != num_windows; ++i) {
+                std::copy_n(flat.begin() + static_cast<std::ptrdiff_t>(i * Window), Window, batches[b][i].begin());
+              }
+            }
----------------
ldionne wrote:

This seems simpler:

```suggestion
            for (std::size_t b = 0; b != BatchSize; ++b) {
              batches[b].resize(num_windows);
              for (std::size_t i = 0; i != num_windows; ++i) {
                auto data = generate_data(Window);
                std::fill_n(data.begin(), Window, batches[b][i].begin())
              }
            }
```

WDYT? This is setup code so efficiency doesn't matter too much.

https://github.com/llvm/llvm-project/pull/215698