GPUCommonAlgorithm: partitionPoint

Device code has no std::partition_point, and kernels write the binary
search by hand (e.g. the last ROF at or before a cluster in
ITS/tracking/GPU/cuda/TrackingKernels.cu). Add
GPUCommonAlgorithm::partitionPoint(first, count, pred), std::partition_point
over the indices [first, first + count), callable from host and device
code; lower and upper bounds are its predicates key(i) < value and
!(value < key(i)).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Felix Schlepper <felix.schlepper@cern.ch>

Author: 2 people

GPU/Common/GPUCommonAlgorithm.h

@@ -49,6 +49,8 @@ class GPUCommonAlgorithm
 #endif
   template <class T>
   GPUd() static void swap(T& a, T& b);
+  template <class I, class P>
+  GPUhd() static I partitionPoint(I first, I count, const P& pred);
 
  private:
   template <typename I>
@@ -330,6 +332,24 @@ GPUdi() void GPUCommonAlgorithm::swap(T& a, T& b)
 }
 #endif
 
+// std::partition_point over the indices [first, first + count): the first index for which pred is false, pred being true on a prefix.
+// A lower bound of value in a sorted key(i) is pred(i) = key(i) < value, an upper bound pred(i) = !(value < key(i)).
+template <class I, class P>
+GPUhdi() I GPUCommonAlgorithm::partitionPoint(I first, I count, const P& pred)
+{
+  while (count > 0) {
+    const I step = count / 2;
+    const I middle = first + step;
+    if (pred(middle)) {
+      first = middle + 1;
+      count -= step + 1;
+    } else {
+      count = step;
+    }
+  }
+  return first;
+}
+
 } // namespace o2::gpu
 
 // ----------------------------- WORK GROUP FUNCTIONS -----------------------------