Use active driver benchmarker for do_bench measurements (#537)

Summary:

Sometimes the triton backend may set a profiler to be different from the standard `do_bench` implmentation. For cuda this has no functional change, i.e. `triton.runtime.driver.active.get_benchmarker()` still points at `do_bench`.

Reviewed By: xuzhao9

Differential Revision: D84215437

Author: bremerm31

tritonbench/components/do_bench/run.py

@@ -475,7 +475,7 @@ def do_bench_wrapper(
                 if latency_measure_mode == "profiler"
                 else _do_bench_inductor
                 if latency_measure_mode == "inductor_benchmarker"
-                else triton.testing.do_bench
+                else triton.runtime.driver.active.get_benchmarker()
             )
 
             return Latency(