Make sure the helpers can be used in multiple tranlsation units.

cpp/include/raft/common/detail/nvtx.cuh

@@ -16,42 +16,43 @@
 
 #pragma once
 
-#include <stdint.h>
-#include <stdlib.h>
-#include <mutex>
 #include <rmm/cuda_stream_view.hpp>
-#include <string>
-#include <unordered_map>
 
 namespace raft {
 namespace common {
 namespace detail {
 
+#ifdef NVTX_ENABLED
+
+#include <nvToolsExt.h>
+#include <stdint.h>
+#include <stdlib.h>
+#include <mutex>
+#include <string>
+#include <unordered_map>
+
 /**
  * @brief An internal struct to store associated state with the color
  * generator
  */
 struct ColorGenState {
   /** collection of all tagged colors generated so far */
-  static std::unordered_map<std::string, uint32_t> allColors;
+  static inline std::unordered_map<std::string, uint32_t> allColors;
   /** mutex for accessing the above map */
-  static std::mutex mapMutex;
+  static inline std::mutex mapMutex;
   /** saturation */
-  static constexpr float S = 0.9f;
+  static inline constexpr float S = 0.9f;
   /** value */
-  static constexpr float V = 0.85f;
+  static inline constexpr float V = 0.85f;
   /** golden ratio */
-  static constexpr float Phi = 1.61803f;
+  static inline constexpr float Phi = 1.61803f;
   /** inverse golden ratio */
-  static constexpr float InvPhi = 1.f / Phi;
+  static inline constexpr float InvPhi = 1.f / Phi;
 };
 
-std::unordered_map<std::string, uint32_t> ColorGenState::allColors;
-std::mutex ColorGenState::mapMutex;
-
 // all h, s, v are in range [0, 1]
 // Ref: http://en.wikipedia.org/wiki/HSL_and_HSV#Converting_to_RGB
-uint32_t hsv2rgb(float h, float s, float v)
+inline uint32_t hsv2rgb(float h, float s, float v)
 {
   uint32_t out = 0xff000000u;
   if (s <= 0.0f) { return out; }
@@ -117,8 +118,11 @@ uint32_t hsv2rgb(float h, float s, float v)
  * associate the currently generated color with it
  * @return returns 32b RGB integer with alpha channel set of 0xff
  */
-uint32_t generateNextColor(const std::string& tag)
+inline uint32_t generateNextColor(const std::string& tag)
 {
+  // std::unordered_map<std::string, uint32_t> ColorGenState::allColors;
+  // std::mutex ColorGenState::mapMutex;
+
   std::lock_guard<std::mutex> guard(ColorGenState::mapMutex);
   if (!tag.empty()) {
     auto itr = ColorGenState::allColors.find(tag);
@@ -132,13 +136,9 @@ uint32_t generateNextColor(const std::string& tag)
   return rgb;
 }
 
-#ifdef NVTX_ENABLED
-
-#include <nvToolsExt.h>
-
-nvtxDomainHandle_t domain = nvtxDomainCreateA("raft");
+static inline nvtxDomainHandle_t domain = nvtxDomainCreateA("raft");
 
-void pushRange_name(const char* name)
+inline void pushRange_name(const char* name)
 {
   nvtxEventAttributes_t eventAttrib = {0};
   eventAttrib.version               = NVTX_VERSION;
@@ -151,7 +151,7 @@ void pushRange_name(const char* name)
 }
 
 template <typename... Args>
-void pushRange(const char* format, Args... args)
+inline void pushRange(const char* format, Args... args)
 {
   if constexpr (sizeof...(args) > 0) {
     int length = std::snprintf(nullptr, 0, format, args...);
@@ -165,15 +165,15 @@ void pushRange(const char* format, Args... args)
 }
 
 template <typename... Args>
-void pushRange(rmm::cuda_stream_view stream, const char* format, Args... args)
+inline void pushRange(rmm::cuda_stream_view stream, const char* format, Args... args)
 {
   stream.synchronize();
   pushRange(format, args...);
 }
 
-void popRange() { nvtxDomainRangePop(domain); }
+inline void popRange() { nvtxDomainRangePop(domain); }
 
-void popRange(rmm::cuda_stream_view stream)
+inline void popRange(rmm::cuda_stream_view stream)
 {
   stream.synchronize();
   popRange();
@@ -182,18 +182,18 @@ void popRange(rmm::cuda_stream_view stream)
 #else  // NVTX_ENABLED
 
 template <typename... Args>
-void pushRange(const char* format, Args... args)
+inline void pushRange(const char* format, Args... args)
 {
 }
 
 template <typename... Args>
-void pushRange(rmm::cuda_stream_view stream, const char* format, Args... args)
+inline void pushRange(rmm::cuda_stream_view stream, const char* format, Args... args)
 {
 }
 
-void popRange() {}
+inline void popRange() {}
 
-void popRange(rmm::cuda_stream_view stream) {}
+inline void popRange(rmm::cuda_stream_view stream) {}
 
 #endif  // NVTX_ENABLED
 

cpp/include/raft/common/nvtx.hpp

@@ -27,7 +27,7 @@ namespace common {
  * @param args the arguments for the printf-style formatting
  */
 template <typename... Args>
-void PUSH_RANGE(const char* format, Args... args)
+inline void PUSH_RANGE(const char* format, Args... args)
 {
   detail::pushRange(format, args...);
 }
@@ -39,19 +39,19 @@ void PUSH_RANGE(const char* format, Args... args)
  * @param stream stream to synchronize
  */
 template <typename... Args>
-void PUSH_RANGE(rmm::cuda_stream_view stream, const char* format, Args... args)
+inline void PUSH_RANGE(rmm::cuda_stream_view stream, const char* format, Args... args)
 {
   detail::pushRange(stream, format, args...);
 }
 
 /** Pop the latest range */
-void POP_RANGE() { detail::popRange(); }
+inline void POP_RANGE() { detail::popRange(); }
 
 /**
  * @brief Synchronize CUDA stream and pop the latest nvtx range
  * @param stream stream to synchronize
  */
-void POP_RANGE(rmm::cuda_stream_view stream) { detail::popRange(stream); }
+inline void POP_RANGE(rmm::cuda_stream_view stream) { detail::popRange(stream); }
 
 /** Push a named nvtx range that would be popped at the end of the object lifetime. */
 class AUTO_RANGE {

cpp/test/distance/distance_base.cuh

@@ -16,6 +16,7 @@
 
 #include <gtest/gtest.h>
 #include <raft/cudart_utils.h>
+#include <raft/common/nvtx.hpp>
 #include <raft/cuda_utils.cuh>
 #include <raft/distance/distance.hpp>
 #include <raft/random/rng.hpp>
@@ -410,6 +411,9 @@ class DistanceTest : public ::testing::TestWithParam<DistanceInputs<DataType>> {
 
   void SetUp() override
   {
+    auto testInfo = testing::UnitTest::GetInstance()->current_test_info();
+    RAFT_USING_RANGE("test::%s/%s", testInfo->test_suite_name(), testInfo->name());
+
     raft::random::Rng r(params.seed);
     int m               = params.m;
     int n               = params.n;

cpp/test/eigen_solvers.cu

@@ -14,6 +14,7 @@
  * limitations under the License.
  */
 
+#include <raft/common/nvtx.hpp>
 #include <raft/handle.hpp>
 #include <raft/spectral/partition.hpp>
 
@@ -27,6 +28,7 @@ namespace raft {
 
 TEST(Raft, EigenSolvers)
 {
+  RAFT_USING_RANGE("test::EigenSolvers");
   using namespace matrix;
   using index_type = int;
   using value_type = double;
@@ -67,6 +69,7 @@ TEST(Raft, EigenSolvers)
 
 TEST(Raft, SpectralSolvers)
 {
+  RAFT_USING_RANGE("test::SpectralSolvers");
   using namespace matrix;
   using index_type = int;
   using value_type = double;