Add cuda_timer, cuda_stream.

testing/CMakeLists.txt

@@ -1,8 +1,9 @@
 set(test_srcs
   axes_metadata.cu
   benchmark.cu
-  int64_axis.cu
+  cuda_timer.cu
   float64_axis.cu
+  int64_axis.cu
   params.cu
   state.cu
   state_generator.cu

testing/cuda_timer.cu

@@ -0,0 +1,58 @@
+#include <nvbench/cuda_timer.cuh>
+
+#include <nvbench/cuda_stream.cuh>
+#include <nvbench/types.cuh>
+
+#include "test_asserts.cuh"
+
+#include <fmt/format.h>
+
+#include <cuda/std/chrono>
+
+#include <cuda_runtime.h>
+
+__global__ void sleep_kernel(double seconds)
+{
+  const auto start = cuda::std::chrono::high_resolution_clock::now();
+  const auto ns    = cuda::std::chrono::nanoseconds(
+    static_cast<nvbench::int64_t>(seconds * 1000 * 1000 * 1000));
+  const auto finish = start + ns;
+
+  auto now = cuda::std::chrono::high_resolution_clock::now();
+  while (now < finish)
+  {
+    now = cuda::std::chrono::high_resolution_clock::now();
+  }
+}
+
+void test_basic(cudaStream_t time_stream,
+                cudaStream_t exec_stream,
+                bool expected)
+{
+  nvbench::cuda_timer timer;
+
+  NVBENCH_CUDA_CALL(cudaDeviceSynchronize());
+
+  timer.start(time_stream);
+  sleep_kernel<<<1, 1, 0, exec_stream>>>(0.25);
+  timer.stop(time_stream);
+
+  NVBENCH_CUDA_CALL(cudaDeviceSynchronize());
+  const bool captured = timer.get_duration() > 0.25;
+  ASSERT_MSG(captured == expected,
+             fmt::format("Unexpected result from timer: {} seconds "
+                         " (expected {})",
+                         timer.get_duration(),
+                         (expected ? "> 0.25s" : "< 0.25s")));
+}
+
+void test_basic()
+{
+  nvbench::cuda_stream stream1;
+  nvbench::cuda_stream stream2;
+
+  test_basic(stream1, stream1, true);
+  test_basic(stream1, stream2, false);
+}
+
+int main() { test_basic(); }