[SYCL][L0v2] Enable verified Level Zero adapter tests

sycl/test-e2e/Adapters/level_zero/batch_barrier.cpp

@@ -1,6 +1,7 @@
 // REQUIRES: gpu, level_zero, level_zero_dev_kit
 // UNSUPPORTED: level_zero_v2_adapter
-// UNSUPPORTED-INTENDED: v2 adapter does not support regular cmd lists
+// UNSUPPORTED-INTENDED: this test relies on L0v1 regular-command-list
+// batching behavior.
 
 // UNSUPPORTED: windows && gpu-intel-gen12
 // UNSUPPORTED-TRACKER: https://github.com/intel/llvm/issues/21556

sycl/test-e2e/Adapters/level_zero/batch_event_status.cpp

@@ -5,7 +5,8 @@
 // UNSUPPORTED-TRACKER: https://github.com/intel/llvm/issues/21556
 
 // UNSUPPORTED: level_zero_v2_adapter
-// UNSUPPORTED-INTENDED: v2 adapter does not support regular cmd lists
+// UNSUPPORTED-INTENDED: this test relies on L0v1 regular-command-list
+// batching behavior.
 
 // UNSUPPORTED: ze_debug
 // UNSUPPORTED-TRACKER: https://github.com/intel/llvm/issues/22347

sycl/test-e2e/Adapters/level_zero/batch_test.cpp

@@ -5,8 +5,8 @@
 // UNSUPPORTED: ze_debug
 // UNSUPPORTED-TRACKER: https://github.com/intel/llvm/issues/22347
 // UNSUPPORTED: level_zero_v2_adapter
-// UNSUPPORTED-INTENDED: V1-only; V2 does not preserve zeCommandListClose/
-// zeCommandQueueExecuteCommandLists flush points.
+// UNSUPPORTED-INTENDED: this test relies on L0v1 regular-command-list
+// batching behavior.
 
 // UNSUPPORTED: windows && gpu-intel-gen12
 // UNSUPPORTED-TRACKER: https://github.com/intel/llvm/issues/21556

sycl/test-e2e/Adapters/level_zero/batch_test_copy_with_compute.cpp

@@ -2,8 +2,8 @@
 // UNSUPPORTED: ze_debug
 // UNSUPPORTED-TRACKER: https://github.com/intel/llvm/issues/22347
 // UNSUPPORTED: level_zero_v2_adapter
-// UNSUPPORTED-INTENDED: V1-only; V2 does not preserve zeCommandListClose/
-// zeCommandQueueExecuteCommandLists flush points.
+// UNSUPPORTED-INTENDED: this test relies on L0v1 regular-command-list
+// batching behavior.
 
 // UNSUPPORTED: windows && (gpu-intel-gen12 || gpu-intel-dg2)
 // UNSUPPORTED-TRACKER: https://github.com/intel/llvm/issues/21556

sycl/test-e2e/Adapters/level_zero/device_scope_events.cpp

@@ -2,7 +2,7 @@
 // UNSUPPORTED: ze_debug
 // UNSUPPORTED-TRACKER: https://github.com/intel/llvm/issues/22347
 // UNSUPPORTED: level_zero_v2_adapter
-// UNSUPPORTED-INTENDED: V1-only
+// UNSUPPORTED-INTENDED: UR_L0_DEVICE_SCOPE_EVENTS is V1-only.
 
 // UNSUPPORTED: windows && (gpu-intel-gen12 || gpu-intel-dg2)
 // UNSUPPORTED-TRACKER: https://github.com/intel/llvm/issues/21766

sycl/test-e2e/Adapters/level_zero/eager_init.cpp

@@ -4,8 +4,8 @@
 // function names that we match in the test.
 
 // UNSUPPORTED: level_zero_v2_adapter
-// UNSUPPORTED-INTENDED: Level Zero v2 adapter performs lazy initialization
-// and does not support SYCL_EAGER_INIT=1, which this test requires.
+// UNSUPPORTED-INTENDED: L0v2 does not consume SYCL_EAGER_INIT/UR_L0_EAGER_INIT;
+// queue execution resources are initialized eagerly by design.
 
 // UNSUPPORTED: windows && gpu-intel-gen12
 // UNSUPPORTED-TRACKER: https://github.com/intel/llvm/issues/21556

sycl/test-e2e/Adapters/level_zero/events_caching.cpp

@@ -1,32 +1,44 @@
 // REQUIRES: gpu, level_zero
-// UNSUPPORTED: level_zero_v2_adapter
-// UNSUPPORTED-INTENDED: v2 adapter does not allow disabling caching
 
 // RUN: %{build}  -o %t.out
 
-// RUN: %{l0_leak_check} %{run} %t.out 2>&1 | FileCheck --check-prefixes=CACHING-ENABLED %s
-// RUN: env SYCL_PI_LEVEL_ZERO_DISABLE_EVENTS_CACHING=0 %{l0_leak_check} %{run} %t.out 2>&1 | FileCheck --check-prefixes=CACHING-ENABLED %s
-// RUN: env SYCL_PI_LEVEL_ZERO_DISABLE_EVENTS_CACHING=1 %{l0_leak_check} %{run} %t.out 2>&1 | FileCheck --check-prefixes=CACHING-DISABLED %s
+// RUN: %{l0_leak_check} %{run} %t.out 2>&1 | FileCheck --check-prefixes=%if level_zero_v2_adapter %{CHECK-V2%} %else %{CACHING-ENABLED%} %s
+// RUN: %if !level_zero_v2_adapter %{env SYCL_PI_LEVEL_ZERO_DISABLE_EVENTS_CACHING=0 %{l0_leak_check} %{run} %t.out 2>&1 | FileCheck --check-prefixes=CACHING-ENABLED %s%}
+// RUN: %if !level_zero_v2_adapter %{env SYCL_PI_LEVEL_ZERO_DISABLE_EVENTS_CACHING=1 %{l0_leak_check} %{run} %t.out 2>&1 | FileCheck --check-prefixes=CACHING-DISABLED %s%}
 //
 // With events caching we should be reusing them and 9 should be enough.
 // Might require more than one if previous one hasn't been released by the time
 // we need a new one.
+//
+// L0v2 has no on/off switch for event caching, so check reuse directly.
+// Handle reuse alone may result from driver address recycling.
+// A low destroy count confirms resource reuse.
 
 // CACHING-ENABLED: zeEventCreate = {{[1-9]}}
 // CACHING-DISABLED: zeEventCreate = 256
+// CHECK-V2-DAG: Distinct native event handles: {{[1-9][0-9]?[[:space:]]*$}}
+// CHECK-V2-DAG: zeEventDestroy = {{[1-9][0-9]?([[:space:]]|$)}}
 
 // Check event caching modes in the L0 adapter.
 
+#include <iostream>
+#include <set>
 #include <sycl/detail/core.hpp>
+#include <sycl/ext/oneapi/backend/level_zero.hpp>
 
 int main() {
   sycl::queue deviceQueue;
+  std::set<ze_event_handle_t> distinctEvents;
 
   for (int i = 0; i < 256; i++) {
     auto Event = deviceQueue.submit([&](sycl::handler &cgh) {
       cgh.single_task<class SimpleKernel>([=]() {});
     });
     Event.wait();
+    distinctEvents.insert(
+        sycl::get_native<sycl::backend::ext_oneapi_level_zero>(Event));
   }
+  std::cout << "Distinct native event handles: " << distinctEvents.size()
+            << "\n";
   return 0;
 }

sycl/test-e2e/Adapters/level_zero/interop-direct.cpp

@@ -1,13 +1,14 @@
 // REQUIRES: level_zero, level_zero_dev_kit
 // UNSUPPORTED: ze_debug
 // UNSUPPORTED-TRACKER: https://github.com/intel/llvm/issues/22347
-// UNSUPPORTED: level_zero_v2_adapter
-// UNSUPPORTED-INTENDED: V1-only behavior for
-// UR_L0_USE_IMMEDIATE_COMMANDLISTS=0/1
 
 // RUN: %{build} %level_zero_options -o %t.out
-// RUN: env UR_L0_USE_IMMEDIATE_COMMANDLISTS=0 %{run} %t.out
-// RUN: env UR_L0_USE_IMMEDIATE_COMMANDLISTS=1 %{run} %t.out
+// RUN: %if !level_zero_v2_adapter %{env UR_L0_USE_IMMEDIATE_COMMANDLISTS=0 %{run} %t.out%}
+// RUN: %if !level_zero_v2_adapter %{env UR_L0_USE_IMMEDIATE_COMMANDLISTS=1 %{run} %t.out%}
+//
+// L0v2 make_queue() supports wrapping an in-order immediate command list,
+// but not a bare command queue.
+// RUN: %if level_zero_v2_adapter %{%{build} %level_zero_options -DTEST_LEVEL_ZERO_V2_NATIVE_INTEROP -o %t_v2.out && %{run} %t_v2.out%}
 
 #include <iostream>
 #include <level_zero/ze_api.h>
@@ -31,26 +32,14 @@ int main() {
     return 1;
   }
 
-  // Create Driver
-  uint32_t driver_handle_count = 0;
-  result = zeDriverGet(&driver_handle_count, nullptr);
-  if (result != ZE_RESULT_SUCCESS) {
-    std::cout << "zeDriverGet failed\n";
-    return 1;
-  }
-  std::cout << "Found " << driver_handle_count << " driver(s)\n";
-  if (driver_handle_count == 0)
-    return 1;
-
-  std::vector<ze_driver_handle_t> driver_handles(driver_handle_count);
-  result = zeDriverGet(&driver_handle_count, driver_handles.data());
-  if (result != ZE_RESULT_SUCCESS) {
-    std::cout << "zeDriverGet failed\n";
-    return 1;
-  }
-
-  ze_driver_handle_t ZeDriver = driver_handles[0];
-  std::cout << "Using default driver, index 0\n";
+  // Respects ONEAPI_DEVICE_SELECTOR, unlike raw zeDriverGet/zeDeviceGet.
+  device SyclSelectedDevice;
+  platform SyclSelectedPlatform = SyclSelectedDevice.get_platform();
+  ze_driver_handle_t ZeDriver =
+      get_native<backend::ext_oneapi_level_zero>(SyclSelectedPlatform);
+  ze_device_handle_t ZeDevice0 =
+      get_native<backend::ext_oneapi_level_zero>(SyclSelectedDevice);
+  std::cout << "Using SYCL-selected driver and device\n";
 
   // Everything below is created with ownership::keep, so this test owns the
   // Level Zero handles and must destroy them. Declared before the SYCL interop
@@ -79,42 +68,31 @@ int main() {
   }
   ZeOwned.Context = ZeContext;
 
-  // Create Devices
-  uint32_t device_count = 0;
-  result = zeDeviceGet(ZeDriver, &device_count, nullptr);
-  if (result != ZE_RESULT_SUCCESS) {
-    std::cout << "zeDeviceGet failed to get count of devices\n";
-    return 1;
-  }
-
-  std::vector<ze_device_handle_t> ZeDevices(device_count);
-  result = zeDeviceGet(ZeDriver, &device_count, ZeDevices.data());
-  if (result != ZE_RESULT_SUCCESS) {
-    std::cout << "zeDeviceGet failed to get device handles\n";
-    return 1;
-  }
-  std::cout << "Using default device, index 0\n";
-
   // Create Command Queue
   ze_command_queue_desc_t Qdescriptor = {};
   Qdescriptor.stype = ZE_STRUCTURE_TYPE_COMMAND_QUEUE_DESC;
   Qdescriptor.mode = ZE_COMMAND_QUEUE_MODE_ASYNCHRONOUS;
   Qdescriptor.ordinal = 0;
   Qdescriptor.index = 0;
+#ifdef TEST_LEVEL_ZERO_V2_NATIVE_INTEROP
+  Qdescriptor.flags = ZE_COMMAND_QUEUE_FLAG_IN_ORDER;
+#endif
 
+#ifndef TEST_LEVEL_ZERO_V2_NATIVE_INTEROP
   ze_command_queue_handle_t ZeCommand_queue = nullptr;
-  result = zeCommandQueueCreate(ZeContext, ZeDevices[0], &Qdescriptor,
+  result = zeCommandQueueCreate(ZeContext, ZeDevice0, &Qdescriptor,
                                 &ZeCommand_queue);
   if (result != ZE_RESULT_SUCCESS) {
     std::cout << "zeCommandQueueCreate failed\n";
     return 1;
   }
   ZeOwned.Queue = ZeCommand_queue;
   std::cout << "Commandqueue created: " << ZeCommand_queue << std::endl;
+#endif
 
   // Create Command List
   ze_command_list_handle_t ZeCommand_list = nullptr;
-  result = zeCommandListCreateImmediate(ZeContext, ZeDevices[0], &Qdescriptor,
+  result = zeCommandListCreateImmediate(ZeContext, ZeDevice0, &Qdescriptor,
                                         &ZeCommand_list);
   if (result != ZE_RESULT_SUCCESS) {
     std::cout << "zeCommandListCreate failed\n";
@@ -125,7 +103,7 @@ int main() {
 
   // Interop object creation
   backend_traits<backend::ext_oneapi_level_zero>::return_type<device> ZeDevice;
-  ZeDevice = ZeDevices[0];
+  ZeDevice = ZeDevice0;
 
   backend_input_t<backend::ext_oneapi_level_zero, platform>
       InteropPlatformInput{ZeDriver};
@@ -146,6 +124,7 @@ int main() {
       make_context<backend::ext_oneapi_level_zero>(InteropContextInput);
   std::cout << "Made context\n";
 
+#ifndef TEST_LEVEL_ZERO_V2_NATIVE_INTEROP
   backend_input_t<backend::ext_oneapi_level_zero, queue> InteropQueueInputCQ{
       ZeCommand_queue, InteropDevice, ext::oneapi::level_zero::ownership::keep};
   queue InteropQueueCQ = make_queue<backend::ext_oneapi_level_zero>(
@@ -170,6 +149,7 @@ int main() {
       return 1;
     }
   }
+#endif
 
   backend_input_t<backend::ext_oneapi_level_zero, queue> InteropQueueInputCL{
       ZeCommand_list, InteropDevice, ext::oneapi::level_zero::ownership::keep};
@@ -195,6 +175,12 @@ int main() {
     return 1;
   }
 
+#ifdef TEST_LEVEL_ZERO_V2_NATIVE_INTEROP
+  queue &InteropQueueForOps = InteropQueueCL;
+#else
+  queue &InteropQueueForOps = InteropQueueCQ;
+#endif
+
   int dataSycl[3] = {7, 8, 0};
   int dataCQ[3] = {7, 8, 0};
   int dataCL[3] = {7, 8, 0};
@@ -225,9 +211,9 @@ int main() {
     std::cout << "GPU Result from SYCL Q = {" << hostOut[0] << ", "
               << hostOut[1] << ", " << hostOut[2] << "}" << std::endl;
 
-    // Try interop queue with standard commandlist
-    InteropQueueCQ.copy<int>(addend, deviceData, 2).wait();
-    InteropQueueCQ.submit([&](handler &cgh) {
+    // Try interop queue (command queue on v1, immediate command list on v2)
+    InteropQueueForOps.copy<int>(addend, deviceData, 2).wait();
+    InteropQueueForOps.submit([&](handler &cgh) {
       accessor numbers{bufDataCQ, cgh, read_write};
       cgh.parallel_for(dataCount,
                        [=](id<1> Id) { numbers[Id] += deviceData[0]; });

sycl/test-e2e/Adapters/level_zero/memory_fill.cpp

@@ -1,6 +1,7 @@
 // REQUIRES: gpu, level_zero, level_zero_dev_kit
 // UNSUPPORTED: level_zero_v2_adapter
-// UNSUPPORTED-INTENDED: v2 adapter does not allow specifying command queue.
+// UNSUPPORTED-INTENDED: the engine-selection control used by this test is
+// V1-only.
 
 // UNSUPPORTED: windows && (gpu-intel-gen12 || gpu-intel-dg2)
 // UNSUPPORTED-TRACKER: https://github.com/intel/llvm/issues/21556

sycl/test-e2e/Adapters/level_zero/queue_create.cpp

@@ -1,12 +1,20 @@
 // REQUIRES: level_zero, level_zero_dev_kit
-// UNSUPPORTED: level_zero_v2_adapter
-// UNSUPPORTED-INTENDED: v2 adapter does not support regular cmd lists
 // RUN: %{build} %level_zero_options -o %t.out
-// RUN: env SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=0 %{l0_leak_check} %{run} %t.out 2>&1 | FileCheck %s
+// RUN: env SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=0 %{l0_leak_check} %{run} %t.out 2>&1 | FileCheck %s --check-prefixes=%if level_zero_v2_adapter %{CHECK-V2%} %else %{CHECK%}
 //
+// The test checks that the Level Zero execution resources associated with a
+// SYCL queue are created once and reused across all 1000 submissions,
+// rather than being recreated per-submission, using the embedded
+// UR_L0_LEAKS_DEBUG=1 testing capability.
+//
+// v1 creates a single regular command queue for the whole run:
 // CHECK:  zeCommandQueueCreate = 1     \--->         zeCommandQueueDestroy = 1
-// The test is to check that there is only a single level zero queue created
-// with the embedded UR_L0_LEAKS_DEBUG=1 testing capability.
+//
+// L0v2 creates a queue's immediate command lists once, in its constructor (a
+// pool of 4 today), so the count must not grow with the 1000 submissions; 1-16
+// leaves headroom. Every list that was created must be destroyed.
+// CHECK-V2: zeCommandListCreateImmediate = [[N:[1-9]|1[0-6]]]{{([[:space:]]|$)}}
+// CHECK-V2: zeCommandListCreate = 0     \---> zeCommandListDestroy = [[N]]{{([[:space:]]|$)}}
 //
 
 #include <sycl/detail/core.hpp>

sycl/test-e2e/Adapters/level_zero/queue_priority.cpp

@@ -1,15 +1,20 @@
 // REQUIRES: gpu, level_zero, level_zero_dev_kit
-// UNSUPPORTED: level_zero_v2_adapter
-// UNSUPPORTED-INTENDED: v2 adapter does not support regular cmd lists
 // RUN: %{build} %level_zero_options -o %t.out
-// RUN: env UR_L0_DEBUG=1 SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=0 %{run} %t.out 2>&1 | FileCheck %s --check-prefixes=CHECK-STD
-// RUN: env UR_L0_DEBUG=1 SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1 %{run} %t.out 2>&1 | FileCheck %s --check-prefixes=CHECK-IMM
+// RUN: env UR_L0_DEBUG=1 SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=0 %{run} %t.out 2>&1 | FileCheck %s --check-prefixes=%if level_zero_v2_adapter %{CHECK-V2%} %else %{CHECK-STD%}
+// RUN: env UR_L0_DEBUG=1 SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1 %{run} %t.out 2>&1 | FileCheck %s --check-prefixes=%if level_zero_v2_adapter %{CHECK-V2%} %else %{CHECK-IMM%}
 //
 // Check that queue priority is passed to Level Zero runtime
 // This is the last value in the ZE_STRUCTURE_TYPE_COMMAND_QUEUE_DESC
 //
 // With immediate command lists the command lists are recycled between queues in
 // a context.
+//
+// SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS is a v1-only selector between
+// regular (CHECK-STD) and immediate (CHECK-IMM) command lists; v2 always uses
+// immediate command lists and does not read this variable, so both RUN lines
+// exercise the same v2 code path (CHECK-V2). The v2 checks cover only the
+// public contract; how priorities map to Level Zero is covered by the adapter
+// tests.
 #include <iostream>
 #include <sycl/detail/core.hpp>
 #include <sycl/properties/all_properties.hpp>
@@ -45,6 +50,7 @@ int main(int Argc, const char *Argv[]) {
 
   // CHECK-STD: Queue cannot be constructed with different priorities.
   // CHECK-IMM: Queue cannot be constructed with different priorities.
+  // CHECK-V2: Queue cannot be constructed with different priorities.
   try {
     test(C, D,
          {sycl::ext::oneapi::property::queue::priority_low{},
@@ -53,6 +59,7 @@ int main(int Argc, const char *Argv[]) {
     std::cerr << E.what() << std::endl;
   }
 
+  // CHECK-V2: The test passed.
   std::cout << "The test passed." << std::endl;
   return 0;
 }

sycl/test-e2e/Adapters/level_zero/usm_residency.cpp

@@ -1,6 +1,7 @@
 // REQUIRES: gpu, level_zero
 // UNSUPPORTED: level_zero_v2_adapter
-// UNSUPPORTED-INTENDED: v2 adapter does not support changing residency.
+// UNSUPPORTED-INTENDED: this test relies on the L0v1-specific
+// SYCL_PI_LEVEL_ZERO_USM_RESIDENT configuration.
 
 // RUN: %{build} %level_zero_options -o %t.out
 // RUN: env SYCL_UR_TRACE=2 UR_L0_DEBUG=-1 %{l0_leak_check} %{run} %t.out 2>&1 | FileCheck --check-prefixes=DEVICE %s

sycl/test-e2e/Adapters/queue_submit_mode.cpp

@@ -1,9 +1,7 @@
 // REQUIRES: gpu, level_zero
-// UNSUPPORTED: level_zero_v2_adapter
-// UNSUPPORTED-INTENDED: v2 adapter does support regular cmd lists
 // RUN: %{build} %level_zero_options -o %t.out
 // RUN: env %{l0_leak_check} SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=0 %{run} %t.out 0 2>&1 | FileCheck %s --check-prefixes=CHECK-STD
-// RUN: env %{l0_leak_check} SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1 %{run} %t.out 1 2>&1 | FileCheck %s --check-prefixes=CHECK-IMM
+// RUN: env %{l0_leak_check} SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1 %{run} %t.out 1 2>&1 | FileCheck %s --check-prefixes=%if level_zero_v2_adapter %{CHECK-IMM-V2%} %else %{CHECK-IMM%}
 //
 // UNSUPPORTED: target-native_cpu
 // UNSUPPORTED-TRACKER: https://github.com/intel/llvm/issues/20142
@@ -38,6 +36,8 @@ int main(int argc, char *argv[]) {
 
   // CHECK-STD: zeCommandListCreateImmediate = 1
   // CHECK-IMM: zeCommandListCreateImmediate = 2
+  // L0v2 out-of-order queues use a fixed pool of 4 immediate command lists.
+  // CHECK-IMM-V2: zeCommandListCreateImmediate = 4
   queue Q1{P};
   queue_submit(Q1);