RCCL Replayer update (#1603)

RCCL recorder w/ suggested change and UT



[ROCm/rccl commit: 9a55ff60a9]

projects/rccl/test/CMakeLists.txt

@@ -46,6 +46,7 @@ if(BUILD_TESTS)
     ScatterTests.cpp
     SendRecvTests.cpp
     StandaloneTests.cpp
+    _RecorderTests.cpp
     common/main.cpp
     common/CallCollectiveForked.cpp
     common/CollectiveArgs.cpp
@@ -55,6 +56,7 @@ if(BUILD_TESTS)
     common/TestBed.cpp
     common/TestBedChild.cpp
     common/StandaloneUtils.cpp
+    ../src/misc/recorder.cc
     )
 
   add_executable(rccl-UnitTests ${TEST_SOURCE_FILES})

projects/rccl/test/_RecorderTests.cpp

@@ -0,0 +1,72 @@
+/*************************************************************************
+ * Copyright (c) 2025 Advanced Micro Devices, Inc. All rights reserved.
+ *
+ * See LICENSE.txt for license information
+ ************************************************************************/
+
+#include <gtest/gtest.h>
+#include <rccl/rccl.h>
+
+#include "RcclMockFuncs.hpp"
+//#include "TestBed.hpp"
+
+namespace RcclUnitTesting
+{
+  /**
+   * \brief Verify correctness of Recorder record() correctness in binary mode
+   * ******************************************************************************************/
+  TEST(Recorder, ParseBinary)
+  {
+    // to add after binary export of logging is supported
+  }
+
+  /**
+   * \brief Verify correctness of Recorder record() correctness in json mode
+   * ******************************************************************************************/
+  TEST(Recorder, ParseJson)
+  {
+    setenv("RCCL_REPLAY_FILE", "test.json", 1);
+
+    int pid = getpid();
+    hipStream_t stream;
+    hipStreamCreate(&stream);
+
+    int array[] = {2, 3, 5};
+    ncclComm comm{.nRanks = 1, .localRank = 1, .localRankToRank = array, .opCount = 8, .planner = {.nTasksColl = 13, .nTasksP2p = 21}};  
+    rccl::rcclApiCall call(rccl::rrAllToAllv, {.sendbuff = (void*)0x7f22f9600000, .recvbuff = (void*)0x7f22f9601000, .count = 0, .datatype = ncclFloat32, .comm = &comm, .stream = stream});
+    rccl::Recorder::instance().record(call);
+
+    std::vector<rccl::rcclApiCall> calls;
+    char entry[4096];
+    //parse the outfile
+    std::string filename = "test" + std::to_string(pid) + ".json";
+    std::ifstream fp("test" + std::to_string(pid) + ".json");
+    fp.getline(entry, 4096);
+    fp.getline(entry, 4096);
+    fp.getline(entry, 4096);
+    parseJsonEntry(entry, calls);
+    int result = memcmp((char*)&calls[0]+4, (char*)&call+4, sizeof(rccl::rcclApiCall)-4);
+    fp.close(); // care that recorder is not designed to anticipate fp closing before destructor
+    remove(filename.c_str());
+    unsetenv("RCCL_REPLAY_FILE");
+    assert(!result);
+  }
+
+  /**
+   * \brief Verify RCCL Recorder's integrity in multithread context by comparing Recorder
+   * instance across different threads.
+   * ******************************************************************************************/
+  static void recorderCmp(void** recorder)
+  {
+    *recorder = &(rccl::Recorder::instance());
+  }
+  TEST(Recorder, VerifyMultithread)
+  {
+    void *p1, *p2;
+    std::thread t1(recorderCmp, &p1);
+    std::thread t2(recorderCmp, &p2);
+    t1.join();
+    t2.join();
+    assert(p1 == p2);
+  }
+}

projects/rccl/test/common/RcclMockFuncs.hpp

@@ -0,0 +1,7 @@
+#include "info.h"
+#include "comm.h"
+
+void ncclDebugLog(ncclDebugLogLevel, unsigned long, char const*, int, char const*, ...) {};
+ncclResult_t getHostName(char* hostname, int maxlen, const char delim) {
+  return ncclSuccess;
+}