[Bench] add SubmitKernel benchmark variant wiht longer kernel

To test OOO queues

Author: igchor

devops/scripts/benchmarks/benches/compute.py

@@ -144,15 +144,17 @@ def benchmarks(self) -> list[Benchmark]:
             for in_order_queue in [0, 1]:
                 for measure_completion in [0, 1]:
                     for use_events in [0, 1]:
-                        benches.append(
-                            SubmitKernel(
-                                self,
-                                runtime,
-                                in_order_queue,
-                                measure_completion,
-                                use_events,
+                        for kernel_exec_time in [1, 20]:
+                            benches.append(
+                                SubmitKernel(
+                                    self,
+                                    runtime,
+                                    in_order_queue,
+                                    measure_completion,
+                                    use_events,
+                                    kernel_exec_time,
+                                )
                             )
-                        )
 
         # Add SinKernelGraph benchmarks
         for runtime in self.enabled_runtimes():
@@ -312,11 +314,20 @@ def teardown(self):
 
 
 class SubmitKernel(ComputeBenchmark):
-    def __init__(self, bench, runtime: RUNTIMES, ioq, MeasureCompletion=0, UseEvents=0):
+    def __init__(
+        self,
+        bench,
+        runtime: RUNTIMES,
+        ioq,
+        MeasureCompletion=0,
+        UseEvents=0,
+        KernelExecTime=1,
+    ):
         self.ioq = ioq
         self.runtime = runtime
         self.MeasureCompletion = MeasureCompletion
         self.UseEvents = UseEvents
+        self.KernelExecTime = KernelExecTime
         super().__init__(
             bench, f"api_overhead_benchmark_{runtime.value}", "SubmitKernel"
         )
@@ -332,7 +343,11 @@ def name(self):
         # to match the existing already stored results
         events_str = " not using events" if not self.UseEvents else ""
 
-        return f"api_overhead_benchmark_{self.runtime.value} SubmitKernel {order}{completion_str}{events_str}"
+        kernel_exec_time_str = (
+            f" KernelExecTime=self.KernelExecTime" if self.KernelExecTime != 1 else ""
+        )
+
+        return f"api_overhead_benchmark_{self.runtime.value} SubmitKernel {order}{completion_str}{events_str}{kernel_exec_time_str}"
 
     def explicit_group(self):
         order = "In Order" if self.ioq else "Out Of Order"
@@ -342,7 +357,11 @@ def explicit_group(self):
         # to match the existing already stored results
         events_str = " not using events" if not self.UseEvents else ""
 
-        return f"SubmitKernel {order}{completion_str}{events_str}"
+        kernel_exec_time_str = (
+            f" KernelExecTime=self.KernelExecTime" if self.KernelExecTime != 1 else ""
+        )
+
+        return f"SubmitKernel {order}{completion_str}{events_str}{kernel_exec_time_str}"
 
     def description(self) -> str:
         order = "in-order" if self.ioq else "out-of-order"
@@ -355,6 +374,7 @@ def description(self) -> str:
         return (
             f"Measures CPU time overhead of submitting {order} kernels through {runtime_name} API{completion_desc}. "
             f"Runs 10 simple kernels with minimal execution time to isolate API overhead from kernel execution time."
+            f"Each kernel exutes for approximately {self.KernelExecTime} micro seconds."
         )
 
     def range(self) -> tuple[float, float]:
@@ -367,7 +387,7 @@ def bin_args(self) -> list[str]:
             "--iterations=100000",
             "--Profiling=0",
             "--NumKernels=10",
-            "--KernelExecTime=1",
+            f"--KernelExecTime={self.KernelExecTime}",
             f"--UseEvents={self.UseEvents}",
         ]