Add a bit more context info to the benchmark output and make the dlop…

…en isolated
rapidsai · Aug 3, 2023 · 228b9cc · 228b9cc
1 parent 1af0342
commit 228b9cc
Show file tree

Hide file tree

Showing 3 changed files with 21 additions and 2 deletions.
diff --git a/cpp/bench/ann/src/common/benchmark.cpp b/cpp/bench/ann/src/common/benchmark.cpp
@@ -30,7 +30,7 @@ struct lib_handle {
   void* handle{nullptr};
   explicit lib_handle(const std::string& name)
   {
-    handle = dlopen(name.c_str(), RTLD_LAZY | RTLD_GLOBAL | RTLD_DEEPBIND | RTLD_NODELETE);
+    handle = dlopen(name.c_str(), RTLD_LAZY | RTLD_LOCAL);
     if (handle == nullptr) {
       auto error_msg = "Failed to load " + name;
       auto err       = dlerror();

diff --git a/cpp/bench/ann/src/common/cuda_stub.hpp b/cpp/bench/ann/src/common/cuda_stub.hpp
@@ -101,6 +101,16 @@ namespace stub {
   *device = 0;
   return cudaSuccess;
 };
+[[gnu::weak, gnu::noinline]] cudaError_t cudaDriverGetVersion(int* driver)
+{
+  *driver = 0;
+  return cudaSuccess;
+};
+[[gnu::weak, gnu::noinline]] cudaError_t cudaRuntimeGetVersion(int* runtime)
+{
+  *runtime = 0;
+  return cudaSuccess;
+};
 [[gnu::weak, gnu::noinline]] cudaError_t cudaGetDeviceProperties(struct cudaDeviceProp* prop,
                                                                  int device)
 {
@@ -129,6 +139,8 @@ RAFT_DECLARE_CUDART(cudaEventSynchronize);
 RAFT_DECLARE_CUDART(cudaEventElapsedTime);
 RAFT_DECLARE_CUDART(cudaEventDestroy);
 RAFT_DECLARE_CUDART(cudaGetDevice);
+RAFT_DECLARE_CUDART(cudaDriverGetVersion);
+RAFT_DECLARE_CUDART(cudaRuntimeGetVersion);
 RAFT_DECLARE_CUDART(cudaGetDeviceProperties);
 
 #undef RAFT_DECLARE_CUDART

diff --git a/cpp/bench/ann/src/common/util.hpp b/cpp/bench/ann/src/common/util.hpp
@@ -139,7 +139,10 @@ struct cuda_timer {
 
 inline auto cuda_info()
 {
-  int dev;
+  int dev, driver = 0, runtime = 0;
+  cudaDriverGetVersion(&driver);
+  cudaRuntimeGetVersion(&runtime);
+
   cudaDeviceProp device_prop;
   cudaGetDevice(&dev);
   cudaGetDeviceProperties(&device_prop, dev);
@@ -151,6 +154,10 @@ inline auto cuda_info()
   props.emplace_back("gpu_mem_bus_width", std::to_string(device_prop.memoryBusWidth));
   props.emplace_back("gpu_mem_global_size", std::to_string(device_prop.totalGlobalMem));
   props.emplace_back("gpu_mem_shared_size", std::to_string(device_prop.sharedMemPerMultiprocessor));
+  props.emplace_back("gpu_driver_version",
+                     std::to_string(driver / 1000) + "." + std::to_string((driver % 100) / 10));
+  props.emplace_back("gpu_runtime_version",
+                     std::to_string(runtime / 1000) + "." + std::to_string((runtime % 100) / 10));
   return props;
 }