[RAS] Add support for RAS client (#1748)

Enable RAS client binary `rcclras`

CMakeLists.txt

@@ -1231,6 +1231,9 @@ License: See LICENSE.txt for license information\n")
   Optimized primitives for collective multi-GPU communication")
 endif()
 
+## Building RCCL RAS
+include(cmake/rcclRAS.cmake)
+
 if(BUILD_TESTS)
   rocm_package_setup_component(clients)
   rocm_package_setup_client_component(tests PACKAGE_NAME unittests)

cmake/rcclRAS.cmake

@@ -0,0 +1,25 @@
+# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
+
+cmake_minimum_required(VERSION 3.16)
+
+message("Building rccl RAS client executable")
+
+set(CMAKE_IGNORE_PATH "${ROCM_PATH}/lib" "${ROCM_PATH}/include")
+
+add_executable(rcclras "${PROJECT_BINARY_DIR}/hipify/src/ras/client.cc")
+
+target_include_directories(rcclras PRIVATE ${PROJECT_BINARY_DIR}/include)
+target_include_directories(rcclras PRIVATE ${HIPIFY_DIR}/src)
+target_include_directories(rcclras PRIVATE ${HIPIFY_DIR}/src/include)
+
+if(BUILD_SHARED_LIBS)
+  target_link_libraries(rcclras PRIVATE rccl)
+  set_property(TARGET rcclras PROPERTY INSTALL_RPATH "${CMAKE_INSTALL_PREFIX}/lib")
+else()
+  add_dependencies(rccl-UnitTests rccl)
+  target_link_libraries(rcclras PRIVATE dl rt -lrccl -L${CMAKE_BINARY_DIR} -lamdhip64 -L${ROCM_PATH}/lib)
+endif()
+
+set_target_properties(rcclras PROPERTIES BUILD_RPATH "${CMAKE_BINARY_DIR}")
+
+rocm_install(TARGETS rcclras)

src/ras/client.cc

@@ -84,7 +84,7 @@ static void parseArgs(int argc, char** argv) {
         printUsage(argv[0]);
         exit(0);
       case 'r':
-        fprintf(stderr, "NCCL RAS client version " STR(NCCL_MAJOR) "." STR(NCCL_MINOR) "."
+        fprintf(stderr, "RCCL RAS client version " STR(NCCL_MAJOR) "." STR(NCCL_MINOR) "."
                 STR(NCCL_PATCH) NCCL_SUFFIX "\n");
         exit(0);
       default:

src/ras/client_support.cc

@@ -487,18 +487,20 @@ static ncclResult_t rasClientRunInit(struct rasClient* client) {
   int firstIdx, nPeers;
   struct rasValCount valCounts[NCCL_MAX_LOCAL_RANKS];
   int nValCounts;
-  static int cudaDriver = -1, cudaRuntime = -1;
+  static int hipRuntime = -1, amdgpuDriver = -1;
 
   TRACE(NCCL_RAS, "RAS: rasClientRunInit: starting");
 
   rasOutReset();
-  rasOutAppend("NCCL version " STR(NCCL_MAJOR) "." STR(NCCL_MINOR) "." STR(NCCL_PATCH) NCCL_SUFFIX
-               " compiled with CUDA " STR(CUDA_MAJOR) "." STR(CUDA_MINOR) "\n");
-  if (cudaRuntime == -1)
-    cudaRuntimeGetVersion(&cudaRuntime);
-  if (cudaDriver == -1)
-    cudaDriverGetVersion(&cudaDriver);
-  rasOutAppend("CUDA runtime version %d, driver version %d\n\n", cudaRuntime, cudaDriver);
+  rasOutAppend("RCCL version " STR(NCCL_MAJOR) "." STR(NCCL_MINOR) "." STR(NCCL_PATCH) NCCL_SUFFIX
+               " compiled with ROCm " STR(ROCM_BUILD_INFO) "\n");
+  if (hipRuntime == -1)
+    hipRuntimeGetVersion(&hipRuntime);
+  if (amdgpuDriver == -1)
+    hipDriverGetVersion(&amdgpuDriver);
+    //Find a better way to query amdgpu driver version, as hipDriverGetVersion() reports the same as hipRuntimeGetVersion()
+    //Else, cudaRuntimeGetVersion() and cudaDriverGetVersion() are anyways hipified, so no need of this mod
+  rasOutAppend("HIP runtime version %d, amdgpu driver version %d\n\n", hipRuntime, amdgpuDriver);
   msgLen = rasOutLength();
   NCCLCHECKGOTO(rasClientAllocMsg(&msg, msgLen), ret, fail);
   rasOutExtract(msg);

src/ras/ras_internal.h

@@ -22,6 +22,7 @@
 #include "ras.h"
 #include "socket.h"
 #include "utils.h"
+#include "hip_rocm_version_info.h"
 
 // Type of a RAS network or client message.
 typedef enum {