Introduces AnnGPU::uses_stream() to determine whether stream sync is …

…needed for a GPU algo between iterations

cpp/bench/ann/src/common/ann_types.hpp

@@ -98,7 +98,16 @@ class AnnGPU {
    * end.
    */
   [[nodiscard]] virtual auto get_sync_stream() const noexcept -> cudaStream_t = 0;
-  virtual ~AnnGPU() noexcept                                                  = default;
+  /**
+   * By default a GPU algorithm uses a fixed stream to order GPU operations.
+   * However, an algorithm may need to synchronize with the host at the end of its execution.
+   * In that case, also synchronizing with a benchmark event would put it at disadvantage.
+   *
+   * We can disable event sync by passing `false` here
+   *   - ONLY IF THE ALGORITHM HAS PRODUCED ITS OUTPUT BY THE TIME IT SYNCHRONIZES WITH CPU.
+   */
+  [[nodiscard]] virtual auto uses_stream() const noexcept -> bool { return true; }
+  virtual ~AnnGPU() noexcept = default;
 };
 
 template <typename T>

cpp/bench/ann/src/common/util.hpp

@@ -118,7 +118,9 @@ struct cuda_timer {
   static inline auto extract_stream(AnnT* algo) -> std::optional<cudaStream_t>
   {
     auto gpu_ann = dynamic_cast<AnnGPU*>(algo);
-    if (gpu_ann != nullptr) { return std::make_optional(gpu_ann->get_sync_stream()); }
+    if (gpu_ann != nullptr && gpu_ann->uses_stream()) {
+      return std::make_optional(gpu_ann->get_sync_stream());
+    }
     return std::nullopt;
   }