Added cudaGetLastError() calls to reset benchmarking kernel errors (issue 88). (#173)

* Create and use NVBENCH_CUDA_CALL_RESET_ERROR.

* Moved cudaGetLastError() call to NVBENCH_CUDA_CALL macro

---------

Co-authored-by: Sergey Pavlov <psvvsp89@gmail.com>

testing/CMakeLists.txt

@@ -16,6 +16,7 @@ set(test_srcs
   named_values.cu
   option_parser.cu
   range.cu
+  reset_error.cu
   ring_buffer.cu
   runner.cu
   state.cu

testing/reset_error.cu

@@ -0,0 +1,30 @@
+#include <nvbench/cuda_call.cuh>
+
+#include "test_asserts.cuh"
+
+
+namespace
+{
+    __global__ void multiply5(const int32_t* __restrict__ a, int32_t* __restrict__ b)
+    {
+      const auto id = blockIdx.x * blockDim.x + threadIdx.x;
+      b[id] = 5 * a[id];
+    }
+}
+
+int main()
+{ 
+  multiply5<<<256, 256>>>(nullptr, nullptr);
+
+  try
+  {
+    NVBENCH_CUDA_CALL(cudaStreamSynchronize(0));
+    ASSERT(false);
+  }
+  catch (const std::runtime_error &)
+  {
+    ASSERT(cudaGetLastError() == cudaError_t::cudaSuccess);
+  }
+
+  return 0;
+}