diff --git a/examples/exec_tag_sync.cu b/examples/exec_tag_sync.cu index 0ef4ee7..1366931 100644 --- a/examples/exec_tag_sync.cu +++ b/examples/exec_tag_sync.cu @@ -27,6 +27,9 @@ // Used to initialize input data: #include +// Used to run the benchmark on a CUDA stream +#include + // `sequence_bench` measures the execution time of `thrust::sequence`. Since // algorithms in `thrust::` implicitly sync the CUDA device, the // `nvbench::exec_tag::sync` must be passed to `state.exec(...)`. diff --git a/examples/exec_tag_timer.cu b/examples/exec_tag_timer.cu index 6aab858..812dda8 100644 --- a/examples/exec_tag_timer.cu +++ b/examples/exec_tag_timer.cu @@ -23,6 +23,7 @@ // Thrust simplifies memory management, etc: #include +#include #include #include