[UR][CUDA] Add opportunistic queue serialize prop, impl for cuda (#18443)

Makes short kernels that don't need to see the same global memory (or
user guarantees global memory writes are complete) launch faster. See
https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#programmatic-dependent-launch-and-synchronization

Makes lots of short kernels in cutlass great again. cc @FMarno who
identified this performance gap.

---------

Signed-off-by: JackAKirk <[email protected]>
Co-authored-by: Jakub Chlanda <[email protected]>

Author: JackAKirk

unified-runtime/include/ur_api.h

@@ -36,6 +36,8 @@ etors:
       desc: "work-group cluster dimensions"
     - name: WORK_GROUP_MEMORY
       desc: "Implicit work group memory allocation"
+    - name: OPPORTUNISTIC_QUEUE_SERIALIZE
+      desc: "Whether to opportunistically execute kernel launches serially on a native queue"
 --- #--------------------------------------------------------------------------
 type: union
 desc: "Specifies a launch property value"
@@ -56,6 +58,10 @@ members:
       name: workgroup_mem_size
       desc: "[in] non-zero value indicates the amount of work group memory to allocate in bytes"
       tag: $X_EXP_LAUNCH_PROPERTY_ID_WORK_GROUP_MEMORY
+    - type: int
+      name: opportunistic_queue_serialize
+      desc: "[in] non-zero value indicates an opportunistic native queue serialized kernel"
+      tag: $X_EXP_LAUNCH_PROPERTY_ID_OPPORTUNISTIC_QUEUE_SERIALIZE
 --- #--------------------------------------------------------------------------
 type: struct
 desc: "Kernel launch property"

unified-runtime/include/ur_print.hpp

@@ -560,6 +560,13 @@ UR_APIEXPORT ur_result_t UR_APICALL urEnqueueKernelLaunchCustomExp(
       attr.value.cooperative = launchPropList[i].value.cooperative;
       break;
     }
+    case UR_EXP_LAUNCH_PROPERTY_ID_OPPORTUNISTIC_QUEUE_SERIALIZE: {
+      auto &attr = launch_attribute.emplace_back();
+      attr.id = CU_LAUNCH_ATTRIBUTE_PROGRAMMATIC_STREAM_SERIALIZATION;
+      attr.value.programmaticStreamSerializationAllowed =
+          launchPropList[i].value.opportunistic_queue_serialize;
+      break;
+    }
     case UR_EXP_LAUNCH_PROPERTY_ID_WORK_GROUP_MEMORY: {
       break;
     }

unified-runtime/scripts/core/exp-launch-properties.yml

@@ -66,6 +66,15 @@ TEST_P(urEnqueueKernelLaunchCustomTest, Success) {
       props.push_back(coop_prop);
     }
 
+    if (compute_capability >= 9.0) {
+      ur_exp_launch_property_t opportunistic_queue_serialize_prop;
+      opportunistic_queue_serialize_prop.id =
+          UR_EXP_LAUNCH_PROPERTY_ID_OPPORTUNISTIC_QUEUE_SERIALIZE;
+      opportunistic_queue_serialize_prop.value.opportunistic_queue_serialize =
+          1;
+      props.push_back(opportunistic_queue_serialize_prop);
+    }
+
     ur_bool_t cluster_launch_supported = false;
     ASSERT_SUCCESS(
         urDeviceGetInfo(device, UR_DEVICE_INFO_CLUSTER_LAUNCH_SUPPORT_EXP,