[SYCL][Graph] Skip Graph tests based on sycl-ls output (#12812)

The graph extension tests are currently skipped during execution for devices which don't support the graphs extension. However, this early return causes the tests to be reported as passed and makes it hard from looking at the results to know if the tests actually stressed the graphs code or not. Improved this situation by changing the SYCL-Graph device info query to an aspect such that `sycl-ls --verbose` will output `ext_oneapi_graph` for supported devices. This can then be used to inform the LIT config and set a requirement for tests, enabling the tests to be obviously skipped for devices that don't support graphs. To enable setting this requirement in `lit.local.cfg` files some extra directories have been created, in particular `UnsupportedDevice` which contains tests that don't have a requirement as the tests verify expected errors are thrown when using the graphs API with unsupported devices. The removal of the device info query means that we can no longer report if a device emulates support for SYCL-Graph, however we currently have no such implementations as they haven't yet deemed to provide enough value. This is technically an ABI breaking change however due to the removal of symbols, but SYCL-Graph is currently an experimental extension so such changes may be permitted.
intel · Mar 11, 2024 · a6301e9 · a6301e9
1 parent 6b0066a
commit a6301e9
Show file tree

Hide file tree

Showing 155 changed files with 173 additions and 674 deletions.
diff --git a/llvm/include/llvm/SYCLLowerIR/DeviceConfigFile.td b/llvm/include/llvm/SYCLLowerIR/DeviceConfigFile.td
@@ -69,6 +69,7 @@ def AspectExt_oneapi_tangle_group : Aspect<"ext_oneapi_tangle_group">;
 def AspectExt_intel_matrix : Aspect<"ext_intel_matrix">;
 def AspectExt_oneapi_is_composite : Aspect<"ext_oneapi_is_composite">;
 def AspectExt_oneapi_is_component : Aspect<"ext_oneapi_is_component">;
+def AspectExt_oneapi_graph : Aspect<"ext_oneapi_graph">;
 // Deprecated aspects
 def AspectInt64_base_atomics : Aspect<"int64_base_atomics">;
 def AspectInt64_extended_atomics : Aspect<"int64_extended_atomics">;
@@ -119,7 +120,7 @@ def : TargetInfo<"__TestAspectList",
     AspectExt_oneapi_interop_semaphore_import, AspectExt_oneapi_interop_semaphore_export,
     AspectExt_oneapi_mipmap, AspectExt_oneapi_mipmap_anisotropy, AspectExt_oneapi_mipmap_level_reference, AspectExt_intel_esimd,
     AspectExt_oneapi_ballot_group, AspectExt_oneapi_fixed_size_group, AspectExt_oneapi_opportunistic_group,
-    AspectExt_oneapi_tangle_group, AspectExt_intel_matrix, AspectExt_oneapi_is_composite, AspectExt_oneapi_is_component],
+    AspectExt_oneapi_tangle_group, AspectExt_intel_matrix, AspectExt_oneapi_is_composite, AspectExt_oneapi_is_component, AspectExt_oneapi_graph],
     []>;
 // This definition serves the only purpose of testing whether the deprecated aspect list defined in here and in SYCL RT
 // match.

diff --git a/sycl/doc/extensions/experimental/sycl_ext_oneapi_graph.asciidoc b/sycl/doc/extensions/experimental/sycl_ext_oneapi_graph.asciidoc
@@ -43,6 +43,7 @@ Dan Holmes, Intel +
 Greg Lueck, Intel +
 Steffen Larsen, Intel +
 Jaime Arteaga Molina, Intel +
+Andrei Elovikov, Intel +
 Ewan Crawford, Codeplay +
 Ben Tracy, Codeplay +
 Duncan McBain, Codeplay +
@@ -296,37 +297,18 @@ Adding an executable graph as a sub-graph does not affect its existing node
 dependencies, such that it could be submitted in future without any side
 effects of prior uses as a sub-graph.
 
-=== Device Info Query
-
-[source, c++]
-----
-namespace sycl::ext::oneapi::experimental {
-enum class graph_support_level {
-  unsupported,
-  native,
-  emulated
-};
-}
-----
+=== Querying Device Support
 
 Due to the experimental nature of the extension, support is not available across
-all devices. The following device support query is added to the
-`sycl::ext::oneapi::experimental` namespace for reporting devices which are
-are currently supported, and how that support is implemented.
+all devices.
 
-Table {counter: tableNumber}. Device Info Queries.
+Table {counter: tableNumber}. Device Support Aspect.
 [%header]
 |===
-| Device Descriptors | Return Type | Description
-
-|`info::device::graph_support`
-|`graph_support_level`
-|When passed to `device::get_info<...>()`, the function returns `native`
-if there is an underlying SYCL backend command-buffer construct which is used
-to propagate the graph to the backend. If no backend construct exists, or
-building on top of it has not yet been implemented, then `emulated` is
-returned. Otherwise `unsupported` is returned if the SYCL device doesn't
-support using this graph extension.
+| Device Descriptor | Description
+
+|`aspect::ext_oneapi_graph`
+| Indicates that the device supports the APIs described in this extension.
 
 |===
 

diff --git a/sycl/include/sycl/device_aspect_macros.hpp b/sycl/include/sycl/device_aspect_macros.hpp
@@ -313,6 +313,11 @@
 #define __SYCL_ALL_DEVICES_HAVE_ext_oneapi_is_component__ 0
 #endif
 
+#ifndef __SYCL_ALL_DEVICES_HAVE_ext_oneapi_graph__
+// __SYCL_ASPECT(ext_oneapi_graph, 61)
+#define __SYCL_ALL_DEVICES_HAVE_ext_oneapi_graph__ 0
+#endif
+
 #ifndef __SYCL_ANY_DEVICE_HAS_host__
 // __SYCL_ASPECT(host, 0)
 #define __SYCL_ANY_DEVICE_HAS_host__ 0
@@ -617,3 +622,8 @@
 // __SYCL_ASPECT(ext_oneapi_is_component, 60)
 #define __SYCL_ANY_DEVICE_HAS_ext_oneapi_is_component__ 0
 #endif
+
+#ifndef __SYCL_ANY_DEVICE_HAS_ext_oneapi_graph__
+// __SYCL_ASPECT(ext_oneapi_graph, 61)
+#define __SYCL_ANY_DEVICE_HAS_ext_oneapi_graph__ 0
+#endif
diff --git a/sycl/include/sycl/info/aspects.def b/sycl/include/sycl/info/aspects.def
@@ -55,3 +55,4 @@ __SYCL_ASPECT(ext_oneapi_tangle_group, 57)
 __SYCL_ASPECT(ext_intel_matrix, 58)
 __SYCL_ASPECT(ext_oneapi_is_composite, 59)
 __SYCL_ASPECT(ext_oneapi_is_component, 60)
+__SYCL_ASPECT(ext_oneapi_graph, 61)
diff --git a/sycl/include/sycl/info/ext_oneapi_device_traits.def b/sycl/include/sycl/info/ext_oneapi_device_traits.def
@@ -13,11 +13,6 @@ __SYCL_PARAM_TRAITS_SPEC(ext::oneapi::experimental, device, matrix_combinations,
                          std::vector<ext::oneapi::experimental::matrix::combination>,
                          PI_EXT_ONEAPI_DEVICE_INFO_MATRIX_COMBINATIONS)
 
-__SYCL_PARAM_TRAITS_SPEC(
-    ext::oneapi::experimental, device, graph_support,
-    ext::oneapi::experimental::graph_support_level,
-    0 /* No PI device code needed */)
-
 // Bindless images pitched allocation
 __SYCL_PARAM_TRAITS_SPEC(ext::oneapi::experimental, device,
                          image_row_pitch_align, uint32_t,

diff --git a/sycl/include/sycl/info/info_desc.hpp b/sycl/include/sycl/info/info_desc.hpp
@@ -193,14 +193,9 @@ template <typename T, T param> struct compatibility_param_traits {};
   } /*namespace info */                                                        \
   } /*namespace Namespace */
 
-namespace ext::oneapi::experimental {
-
-enum class graph_support_level { unsupported = 0, native = 1, emulated = 2 };
-
-namespace info::device {
+namespace ext::oneapi::experimental::info::device {
 template <int Dimensions> struct max_work_groups;
-} // namespace info::device
-} // namespace ext::oneapi::experimental
+} // namespace ext::oneapi::experimental::info::device
 #include <sycl/info/ext_codeplay_device_traits.def>
 #include <sycl/info/ext_intel_device_traits.def>
 #include <sycl/info/ext_oneapi_device_traits.def>

diff --git a/sycl/source/detail/device_impl.cpp b/sycl/source/detail/device_impl.cpp
@@ -602,6 +602,31 @@ bool device_impl::has(aspect Aspect) const {
 
     return Result != nullptr;
   }
+  case aspect::ext_oneapi_graph: {
+    size_t ResultSize = 0;
+    bool CallSuccessful = getPlugin()->call_nocheck<PiApiKind::piDeviceGetInfo>(
+                              MDevice, PI_DEVICE_INFO_EXTENSIONS, 0, nullptr,
+                              &ResultSize) == PI_SUCCESS;
+    if (!CallSuccessful || ResultSize == 0) {
+      return PI_FALSE;
+    }
+
+    std::unique_ptr<char[]> Result(new char[ResultSize]);
+    CallSuccessful = getPlugin()->call_nocheck<PiApiKind::piDeviceGetInfo>(
+                         MDevice, PI_DEVICE_INFO_EXTENSIONS, ResultSize,
+                         Result.get(), nullptr) == PI_SUCCESS;
+
+    if (!CallSuccessful) {
+      return PI_FALSE;
+    }
+
+    std::string_view ExtensionsString(Result.get());
+    std::cout << ExtensionsString;
+    const bool Support =
+        ExtensionsString.find("ur_exp_command_buffer") != std::string::npos;
+
+    return Support;
+  }
   }
   throw runtime_error("This device aspect has not been implemented yet.",
                       PI_ERROR_INVALID_DEVICE);

diff --git a/sycl/source/detail/device_info.hpp b/sycl/source/detail/device_info.hpp
@@ -1185,34 +1185,6 @@ struct get_device_info_impl<
   }
 };
 
-// Specialization for graph extension support
-template <>
-struct get_device_info_impl<
-    ext::oneapi::experimental::graph_support_level,
-    ext::oneapi::experimental::info::device::graph_support> {
-  static ext::oneapi::experimental::graph_support_level
-  get(const DeviceImplPtr &Dev) {
-    size_t ResultSize = 0;
-    Dev->getPlugin()->call<PiApiKind::piDeviceGetInfo>(
-        Dev->getHandleRef(), PI_DEVICE_INFO_EXTENSIONS, 0, nullptr,
-        &ResultSize);
-    if (ResultSize == 0)
-      return ext::oneapi::experimental::graph_support_level::unsupported;
-
-    std::unique_ptr<char[]> Result(new char[ResultSize]);
-    Dev->getPlugin()->call<PiApiKind::piDeviceGetInfo>(
-        Dev->getHandleRef(), PI_DEVICE_INFO_EXTENSIONS, ResultSize,
-        Result.get(), nullptr);
-
-    std::string_view ExtensionsString(Result.get());
-    bool CmdBufferSupport =
-        ExtensionsString.find("ur_exp_command_buffer") != std::string::npos;
-    return CmdBufferSupport
-               ? ext::oneapi::experimental::graph_support_level::native
-               : ext::oneapi::experimental::graph_support_level::unsupported;
-  }
-};
-
 // Specialization for composite devices extension.
 template <>
 struct get_device_info_impl<
@@ -2175,13 +2147,6 @@ inline uint32_t get_device_info_host<
                       PI_ERROR_INVALID_DEVICE);
 }
 
-template <>
-inline ext::oneapi::experimental::graph_support_level
-get_device_info_host<ext::oneapi::experimental::info::device::graph_support>() {
-  // No support for graphs on the host device.
-  return ext::oneapi::experimental::graph_support_level::unsupported;
-}
-
 template <>
 inline uint32_t get_device_info_host<
     ext::oneapi::experimental::info::device::image_row_pitch_align>() {

diff --git a/sycl/source/detail/graph_impl.cpp b/sycl/source/detail/graph_impl.cpp
@@ -16,10 +16,6 @@
 #include <sycl/feature_test.hpp>
 #include <sycl/queue.hpp>
 
-// Developer switch to use emulation mode on all backends, even those that
-// report native support, this is useful for debugging.
-#define FORCE_EMULATION_MODE 0
-
 namespace sycl {
 inline namespace _V1 {
 
@@ -1288,21 +1284,9 @@ void executable_command_graph::finalizeImpl() {
   impl->makePartitions();
 
   auto Device = impl->getGraphImpl()->getDevice();
-  bool CmdBufSupport =
-      Device
-          .get_info<ext::oneapi::experimental::info::device::graph_support>() ==
-      graph_support_level::native;
-
-#if FORCE_EMULATION_MODE
-  // Above query should still succeed in emulation mode, but ignore the
-  // result and use emulation.
-  CmdBufSupport = false;
-#endif
-  if (CmdBufSupport) {
-    for (auto Partition : impl->getPartitions()) {
-      if (!Partition->isHostTask()) {
-        impl->createCommandBuffers(Device, Partition);
-      }
+  for (auto Partition : impl->getPartitions()) {
+    if (!Partition->isHostTask()) {
+      impl->createCommandBuffers(Device, Partition);
     }
   }
 }

diff --git a/sycl/source/detail/graph_impl.hpp b/sycl/source/detail/graph_impl.hpp
@@ -586,9 +586,7 @@ class graph_impl {
       MAllowBuffers = true;
     }
 
-    if (SyclDevice.get_info<
-            ext::oneapi::experimental::info::device::graph_support>() ==
-        graph_support_level::unsupported) {
+    if (!SyclDevice.has(aspect::ext_oneapi_graph)) {
       std::stringstream Stream;
       Stream << SyclDevice.get_backend();
       std::string BackendString = Stream.str();

diff --git a/...aph_exception_global_device_extension.cpp → ...aph_exception_global_device_extension.cpp b/...aph_exception_global_device_extension.cpp → ...aph_exception_global_device_extension.cpp
@@ -5,7 +5,7 @@
 // when trying to use sycl_ext_oneapi_device_global
 // along with Graph.
 
-#include "graph_common.hpp"
+#include "../graph_common.hpp"
 
 using TestProperties = decltype(sycl::ext::oneapi::experimental::properties{});
 
@@ -143,10 +143,6 @@ template <OperationPath PathKind> void test(queue Queue) {
 int main() {
   queue Queue;
 
-  if (!are_graphs_supported(Queue)) {
-    return 0;
-  }
-
   test<OperationPath::Explicit>(Queue);
   test<OperationPath::RecordReplay>(Queue);
   test<OperationPath::Shortcut>(Queue);

diff --git a/sycl/test-e2e/Graph/invalid_depends_on.cpp → ...st-e2e/Graph/Error/invalid_depends_on.cpp b/sycl/test-e2e/Graph/invalid_depends_on.cpp → ...st-e2e/Graph/Error/invalid_depends_on.cpp
@@ -4,15 +4,11 @@
 // Tests that calling handler::depends_on() for events not part of the graph
 // throws.
 
-#include "graph_common.hpp"
+#include "../graph_common.hpp"
 
 int main() {
   queue Queue{};
 
-  if (!are_graphs_supported(Queue)) {
-    return 0;
-  }
-
   ext::oneapi::experimental::command_graph Graph{Queue.get_context(),
                                                  Queue.get_device()};
   ext::oneapi::experimental::command_graph Graph2{Queue.get_context(),

diff --git a/sycl/test-e2e/Graph/invalid_event_wait.cpp → ...st-e2e/Graph/Error/invalid_event_wait.cpp b/sycl/test-e2e/Graph/invalid_event_wait.cpp → ...st-e2e/Graph/Error/invalid_event_wait.cpp
@@ -4,15 +4,11 @@
 // Tests that waiting on an event returned from a Record and Replay submission
 // throws.
 
-#include "graph_common.hpp"
+#include "../graph_common.hpp"
 
 int main() {
   queue Queue{};
 
-  if (!are_graphs_supported(Queue)) {
-    return 0;
-  }
-
   ext::oneapi::experimental::command_graph Graph{Queue.get_context(),
                                                  Queue.get_device()};
   Graph.begin_recording(Queue);

diff --git a/sycl/test-e2e/Graph/invalid_queue_wait.cpp → ...st-e2e/Graph/Error/invalid_queue_wait.cpp b/sycl/test-e2e/Graph/invalid_queue_wait.cpp → ...st-e2e/Graph/Error/invalid_queue_wait.cpp
@@ -3,15 +3,11 @@
 
 // Tests that waiting on a Queue in recording mode throws.
 
-#include "graph_common.hpp"
+#include "../graph_common.hpp"
 
 int main() {
   queue Queue{};
 
-  if (!are_graphs_supported(Queue)) {
-    return 0;
-  }
-
   ext::oneapi::experimental::command_graph Graph{Queue.get_context(),
                                                  Queue.get_device()};
   Graph.begin_recording(Queue);

diff --git a/sycl/test-e2e/Graph/Error/lit.local.cfg b/sycl/test-e2e/Graph/Error/lit.local.cfg
@@ -0,0 +1 @@
+config.required_features += ['aspect-ext_oneapi_graph']
diff --git a/sycl/test-e2e/Graph/Explicit/add_node_while_recording.cpp b/sycl/test-e2e/Graph/Explicit/add_node_while_recording.cpp
@@ -14,10 +14,6 @@
 int main() {
   queue Queue{};
 
-  if (!are_graphs_supported(Queue)) {
-    return 0;
-  }
-
   bool Success = false;
 
   exp_ext::command_graph Graph{Queue.get_context(), Queue.get_device()};

diff --git a/sycl/test-e2e/Graph/Explicit/basic_usm_host.cpp b/sycl/test-e2e/Graph/Explicit/basic_usm_host.cpp
@@ -4,7 +4,8 @@
 // RUN: %if level_zero %{env SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=0 %{l0_leak_check} %{run} %t.out 2>&1 | FileCheck %s --implicit-check-not=LEAK %}
 // Extra run to check for immediate-command-list in Level Zero
 // RUN: %if level_zero && linux %{env SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1 %{l0_leak_check} %{run} %t.out 2>&1 | FileCheck %s --implicit-check-not=LEAK %}
-//
+
+// REQUIRES: aspect-usm_host_allocations
 
 #define GRAPH_E2E_EXPLICIT
 

diff --git a/sycl/test-e2e/Graph/Explicit/basic_usm_mixed.cpp b/sycl/test-e2e/Graph/Explicit/basic_usm_mixed.cpp
@@ -4,7 +4,9 @@
 // RUN: %if level_zero %{env SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=0 %{l0_leak_check} %{run} %t.out 2>&1 | FileCheck %s --implicit-check-not=LEAK %}
 // Extra run to check for immediate-command-list in Level Zero
 // RUN: %if level_zero && linux %{env SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1 %{l0_leak_check} %{run} %t.out 2>&1 | FileCheck %s --implicit-check-not=LEAK %}
-//
+
+// REQUIRES: aspect-usm_host_allocations
+// REQUIRES: aspect-usm_shared_allocations
 
 #define GRAPH_E2E_EXPLICIT
 

diff --git a/sycl/test-e2e/Graph/Explicit/basic_usm_shared.cpp b/sycl/test-e2e/Graph/Explicit/basic_usm_shared.cpp
@@ -4,7 +4,8 @@
 // RUN: %if level_zero %{env SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=0 %{l0_leak_check} %{run} %t.out 2>&1 | FileCheck %s --implicit-check-not=LEAK %}
 // Extra run to check for immediate-command-list in Level Zero
 // RUN: %if level_zero && linux %{env SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1 %{l0_leak_check} %{run} %t.out 2>&1 | FileCheck %s --implicit-check-not=LEAK %}
-//
+
+// REQUIRES: aspect-usm_shared_allocations
 
 #define GRAPH_E2E_EXPLICIT
 

diff --git a/sycl/test-e2e/Graph/Explicit/basic_usm_system.cpp b/sycl/test-e2e/Graph/Explicit/basic_usm_system.cpp
@@ -4,7 +4,8 @@
 // RUN: %if level_zero %{env SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=0 %{l0_leak_check} %t.out 2>&1 | FileCheck %s --implicit-check-not=LEAK %}
 // Extra run to check for immediate-command-list in Level Zero
 // RUN: %if level_zero && linux %{env SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1 %{l0_leak_check} %{run} %t.out 2>&1 | FileCheck %s --implicit-check-not=LEAK %}
-//
+
+// REQUIRES: aspect-usm_system_allocations
 
 #define GRAPH_E2E_EXPLICIT