2.16.2-1

Add support for CUDA 12.0, drop Kepler (sm_35).
Support for H100 features.
Make socket code more robust and protected. Solves #555.
Improve performance on large CUDA graphs, reducing dependencies.
Reduce inter-socket bandwidth on AMD CPUs to favor better paths.
Various fixes to ncclCommAbort.
Make service thread polling resistant to EINTR.
Compile with profiling API by default.
Extend NVTX instrumentation with call arguments.

src/include/nvtx3.hpp

@@ -92,6 +92,7 @@
 /* clang-format on */
 
 #include <nvtx3/nvToolsExt.h>
+#include <nvtx3/nvToolsExtPayload.h>
 
 #include <memory>
 #include <string>
@@ -1732,6 +1733,22 @@ class event_attributes {
     attributes_.messageType = m.get_type();
   }
 
+  /**
+   * @brief Variadic constructor where the first argument is a binary payload.
+   *
+   * Sets the value of the `EventAttribute`s message based on `m` and forwards
+   * the remaining variadic parameter pack to the next constructor.
+   *
+   */
+  template <typename... Args>
+  NVTX3_RELAXED_CONSTEXPR explicit event_attributes(nvtxPayloadData_t const* bpl, Args const&... args) noexcept
+    : event_attributes(args...)
+  {
+    attributes_.payloadType = NVTX_PAYLOAD_TYPE_BINARY;
+    attributes_.reserved0 = 1; // NCCL uses only a single binary payload per event.
+    attributes_.payload.ullValue = NVTX_POINTER_AS_PAYLOAD_ULLVALUE(bpl);
+  }
+
   ~event_attributes()                       = default;
   event_attributes(event_attributes const&) = default;
   event_attributes& operator=(event_attributes const&) = default;