GPU RTC: Check that current compiler versions are the same as those used to create RTC code in cache

Author: davidrohr

GPU/GPUTracking/Base/GPUReconstruction.h

@@ -275,6 +275,7 @@ class GPUReconstruction
 
   // Private helpers for library loading
   static std::shared_ptr<LibraryLoader>* GetLibraryInstance(DeviceType type, bool verbose);
+  static std::string getBackendVersions();
 
   // Private helper functions for memory management
   size_t AllocateRegisteredMemoryHelper(GPUMemoryResource* res, void*& ptr, void*& memorypool, void* memorybase, size_t memorysize, void* (GPUMemoryResource::*SetPointers)(void*), void*& memorypoolend, const char* device);

GPU/GPUTracking/Base/GPUReconstructionAvailableBackends.template.h

@@ -12,6 +12,9 @@
 /// \file GPUReconstructionAvailableBackends.h
 /// \author David Rohr
 
+// clang-format off
 #cmakedefine CUDA_ENABLED
 #cmakedefine HIP_ENABLED
 #cmakedefine OPENCL_ENABLED
+#cmakedefine GPUCA_COMPILER_VERSIONS @GPUCA_COMPILER_VERSIONS@
+// clang-format on

GPU/GPUTracking/Base/GPUReconstructionLibrary.cxx

@@ -184,3 +184,8 @@ int32_t GPUReconstruction::LibraryLoader::CloseLibrary()
   mGPUEntry = nullptr;
   return 0;
 }
+
+std::string GPUReconstruction::getBackendVersions()
+{
+  return GPUCA_M_STR(GPUCA_COMPILER_VERSIONS);
+}

GPU/GPUTracking/Base/cuda/GPUReconstructionCUDAGenRTC.cxx

@@ -80,13 +80,16 @@ int32_t GPUReconstructionCUDA::genRTC(std::string& filename, uint32_t& nCompile)
     GPUInfo("RTC Launch Bounds:\n%s", launchBounds.c_str());
   }
 
-  char shasource[21], shaparam[21], shacmd[21], shakernels[21], shabounds[21];
+  const std::string compilerVersions = getBackendVersions();
+
+  char shasource[21], shaparam[21], shacmd[21], shakernels[21], shabounds[21], shaversion[21];
   if (GetProcessingSettings().rtc.cacheOutput) {
     o2::framework::internal::SHA1(shasource, _binary_GPUReconstructionCUDArtc_src_start, _binary_GPUReconstructionCUDArtc_src_len);
     o2::framework::internal::SHA1(shaparam, rtcparam.c_str(), rtcparam.size());
     o2::framework::internal::SHA1(shacmd, baseCommand.c_str(), baseCommand.size());
     o2::framework::internal::SHA1(shakernels, kernelsall.c_str(), kernelsall.size());
     o2::framework::internal::SHA1(shabounds, launchBounds.c_str(), launchBounds.size());
+    o2::framework::internal::SHA1(shaversion, compilerVersions.c_str(), compilerVersions.size());
   }
 
   nCompile = GetProcessingSettings().rtc.compilePerKernel ? kernels.size() : 1;
@@ -131,11 +134,12 @@ int32_t GPUReconstructionCUDA::genRTC(std::string& filename, uint32_t& nCompile)
           }
           return 0;
         };
-        if (checkSHA(shasource, "source") || // TODO: CHECK VERSION!
+        if (checkSHA(shasource, "source") ||
             checkSHA(shaparam, "param") ||
             checkSHA(shacmd, "command line") ||
             checkSHA(shakernels, "kernel definitions") ||
-            checkSHA(shabounds, "launch bounds")) {
+            checkSHA(shabounds, "launch bounds") ||
+            checkSHA(shaversion, "compiler versions")) {
           break;
         }
         GPUSettingsProcessingRTC cachedSettings;
@@ -237,6 +241,7 @@ int32_t GPUReconstructionCUDA::genRTC(std::string& filename, uint32_t& nCompile)
           fwrite(shacmd, 1, 20, fp) != 20 ||
           fwrite(shakernels, 1, 20, fp) != 20 ||
           fwrite(shabounds, 1, 20, fp) != 20 ||
+          fwrite(shaversion, 1, 20, fp) != 20 ||
           fwrite(&GetProcessingSettings().rtc, sizeof(GetProcessingSettings().rtc), 1, fp) != 1) {
         throw std::runtime_error("Error writing cache file");
       }

GPU/GPUTracking/CMakeLists.txt

@@ -396,6 +396,7 @@ install(DIRECTORY utils
         DESTINATION ${CMAKE_INSTALL_INCLUDEDIR}/GPU
         FILES_MATCHING PATTERN *.h)
 
+set(GPUCA_COMPILER_VERSIONS "CUDA ${CMAKE_CUDA_COMPILER_VERSION} HIP ${hip_VERSION} ${CMAKE_HIP_COMPILER_VERSION} ROCM ${CMAKE_HIP_COMPILER_ROCM_ROOT} LLVM ${LLVM_PACKAGE_VERSION}")
 
 # GPUReconstructionLibrary needs to know which GPU backends are enabled for proper error messages
 configure_file(Base/GPUReconstructionAvailableBackends.template.h ${CMAKE_CURRENT_BINARY_DIR}/GPUReconstructionAvailableBackends.h)

dependencies/FindO2GPU.cmake

@@ -139,8 +139,8 @@ if(ENABLE_CUDA)
       message(${FAILURE_SEVERITY} "CUDA found but thrust not available")
       set(CMAKE_CUDA_COMPILER OFF)
     endif()
-    if (NOT CMAKE_CUDA_COMPILER_VERSION VERSION_GREATER_EQUAL "12.6")
-      message(${FAILURE_SEVERITY} "CUDA Version too old: ${CMAKE_CUDA_COMPILER_VERSION}, 12.6 required")
+    if (NOT CMAKE_CUDA_COMPILER_VERSION VERSION_GREATER_EQUAL "12.8")
+      message(${FAILURE_SEVERITY} "CUDA Version too old: ${CMAKE_CUDA_COMPILER_VERSION}, 12.8 required")
       set(CMAKE_CUDA_COMPILER OFF)
     endif()
   endif()
@@ -279,7 +279,7 @@ if(ENABLE_HIP)
   elseif(NOT ENABLE_HIP STREQUAL "AUTO")
     message(FATAL_ERROR "HIP requested, but CMAKE_PREFIX_PATH env variable does not contain rocm folder!")
   endif()
-  if(hip_FOUND AND NOT hip_VERSION VERSION_GREATER_EQUAL "5.5")
+  if(hip_FOUND AND NOT hip_VERSION VERSION_GREATER_EQUAL "6.3")
     set(hip_FOUND 0)
   endif()
   if(hip_FOUND AND hipcub_FOUND AND rocthrust_FOUND AND rocprim_FOUND AND hip_HIPCC_EXECUTABLE AND hip_HIPIFY_PERL_EXECUTABLE)