[SYCL] Add support for default contexts (#2805)

A default context is now defined per platform. The default
context initially contains all devices in the platform. Queue
creation will use the device's platform's default context
instead of creating a new context with that device.

The new extension introduces ext_oneapi_get_default_context()
member function to the platform class.

DPC++ runtime allows control of this feature with
SYCL_ENABLE_DEFAULT_CONTEXTS env variable. It is enabled
by default on Linux and disabled on Windows.

Author: jbrodman

sycl/doc/EnvironmentVariables.md

@@ -47,6 +47,7 @@ subject to change. Do not rely on these variables in production code.
 | `SYCL_CACHE_THRESHOLD` | Positive integer | Cache eviction threshold in days (default value is 7 for 1 week). Set to 0 for disabling time-based cache eviction. |
 | `SYCL_CACHE_MIN_DEVICE_IMAGE_SIZE` | Positive integer | Minimum size of device code image in bytes which is reasonable to cache on disk because disk access operation may take more time than do JIT compilation for it. Default value is 0 to cache all images. |
 | `SYCL_CACHE_MAX_DEVICE_IMAGE_SIZE` | Positive integer | Maximum size of device image in bytes which is cached. Too big kernels may overload disk too fast. Default value is 1 GB. |
+| `SYCL_ENABLE_DEFAULT_CONTEXTS` | '1' or '0' | Enable ('1') or disable ('0') creation of default platform contexts in SYCL runtime. The default context for each platform contains all devices in the platform. Refer to [Platform Default Contexts](extensions/PlatformContext/PlatformContext.adoc) extension to learn more. Enabled by default on Linux and disabled on Windows. |
 
 `(*) Note: Any means this environment variable is effective when set to any non-null value.`
 

sycl/doc/extensions/PlatformContext/PlatformContext.adoc

@@ -0,0 +1,34 @@
+= SYCL(TM) Proposals: Platform Default Contexts
+James Brodman <james[email protected]>
+v0.1
+:source-highlighter: pygments
+:icons: font
+:y: icon:check[role="green"]
+:n: icon:times[role="red"]
+
+== Platform Default Contexts
+
+This extension adds the notion of a default SYCL context per SYCL platform. The default context for each platform contains all devices in the platform.
+
+The platform class gains one new method:
+
+[cols="^60a,40"]
+|===
+| Member Function | Description
+
+|
+[source,c++]
+----
+context ext_oneapi_get_default_context()
+----
+
+| Returns the current default context for this `platform`
+
+|===
+
+This extension also modifies the behavior of `queue` constructors. Queues will no longer create a new `context` upon construction. Instead, they will use the default context from the device's platform.
+
+== Feature Test Macro
+
+This extension defines the macro `SYCL_EXT_ONEAPI_DEFAULT_CONTEXT` to `1` to indicate that it is enabled.
+

sycl/doc/extensions/README.md

@@ -37,6 +37,7 @@ DPC++ extensions status:
 | [Use Pinned Memory Property](UsePinnedMemoryProperty/UsePinnedMemoryPropery.adoc)                                           | Supported                                 | |
 | [Level-Zero backend specification](LevelZeroBackend/LevelZeroBackend.md)                                           	      | Supported                                 | |
 | [ITT annotations support](ITTAnnotations/ITTAnnotations.rst) | Supported | |
+| [Platform Context](PlatformContext/PlatformContext.adoc) | Proposal | |
 | [SYCL_EXT_ONEAPI_DEVICE_IF](DeviceIf/device_if.asciidoc) | Proposal | |
 | [SYCL_INTEL_group_sort](GroupAlgorithms/SYCL_INTEL_group_sort.asciidoc)                                                     | Proposal                                  | |
 | [Invoke SIMD](InvokeSIMD/InvokeSIMD.asciidoc)                                                                               | Proposal                                  | |

sycl/include/CL/sycl/platform.hpp

@@ -7,7 +7,9 @@
 //===----------------------------------------------------------------------===//
 
 #pragma once
+
 #include <CL/sycl/aspects.hpp>
+#include <CL/sycl/context.hpp>
 #include <CL/sycl/detail/common.hpp>
 #include <CL/sycl/detail/export.hpp>
 #include <CL/sycl/stl.hpp>
@@ -25,6 +27,9 @@ namespace detail {
 class platform_impl;
 }
 
+// Feature test macro for Default Context
+#define SYCL_EXT_ONEAPI_DEFAULT_CONTEXT 1
+
 /// Encapsulates a SYCL platform on which kernels may be executed.
 ///
 /// \ingroup sycl_api
@@ -132,6 +137,11 @@ class __SYCL_EXPORT platform {
   /// given feature.
   bool has(aspect Aspect) const;
 
+  /// Return this platform's default context
+  ///
+  /// \return the default context
+  context ext_oneapi_get_default_context() const;
+
 private:
   pi_native_handle getNative() const;
 

sycl/plugins/cuda/pi_cuda.cpp

@@ -111,7 +111,7 @@ pi_result forLatestEvents(const pi_event *event_wait_list,
 ///
 pi_result check_error(CUresult result, const char *function, int line,
                       const char *file) {
-  if (result == CUDA_SUCCESS) {
+  if (result == CUDA_SUCCESS || result == CUDA_ERROR_DEINITIALIZED) {
     return PI_SUCCESS;
   }
 

sycl/source/detail/config.def

@@ -38,3 +38,4 @@ CONFIG(SYCL_OVERRIDE_PI_OPENCL, 1024, __SYCL_OVERRIDE_PI_OPENCL)
 CONFIG(SYCL_OVERRIDE_PI_LEVEL_ZERO, 1024, __SYCL_OVERRIDE_PI_LEVEL_ZERO)
 CONFIG(SYCL_OVERRIDE_PI_CUDA, 1024, __SYCL_OVERRIDE_PI_CUDA)
 CONFIG(SYCL_OVERRIDE_PI_ROCM, 1024, __SYCL_OVERRIDE_PI_ROCM)
+CONFIG(SYCL_ENABLE_DEFAULT_CONTEXTS, 1, __SYCL_ENABLE_DEFAULT_CONTEXTS)

sycl/source/detail/config.hpp

@@ -285,6 +285,38 @@ template <> class SYCLConfig<SYCL_DEVICE_FILTER> {
   }
 };
 
+template <> class SYCLConfig<SYCL_ENABLE_DEFAULT_CONTEXTS> {
+  using BaseT = SYCLConfigBase<SYCL_ENABLE_DEFAULT_CONTEXTS>;
+
+public:
+  static bool get() {
+#ifdef WIN32
+    constexpr bool DefaultValue = false;
+#else
+    constexpr bool DefaultValue = true;
+#endif
+
+    const char *ValStr = getCachedValue();
+
+    if (!ValStr)
+      return DefaultValue;
+
+    return ValStr[0] == '1';
+  }
+
+  static void reset() { (void)getCachedValue(/*ResetCache=*/true); }
+
+  static const char *getName() { return BaseT::MConfigName; }
+
+private:
+  static const char *getCachedValue(bool ResetCache = false) {
+    static const char *ValStr = BaseT::getRawValue();
+    if (ResetCache)
+      ValStr = BaseT::getRawValue();
+    return ValStr;
+  }
+};
+
 } // namespace detail
 } // namespace sycl
 } // __SYCL_INLINE_NAMESPACE(cl)

sycl/source/detail/global_handler.cpp

@@ -51,6 +51,15 @@ ProgramManager &GlobalHandler::getProgramManager() {
   return getOrCreate(MProgramManager);
 }
 
+std::unordered_map<PlatformImplPtr, ContextImplPtr> &
+GlobalHandler::getPlatformToDefaultContextCache() {
+  return getOrCreate(MPlatformToDefaultContextCache);
+}
+
+std::mutex &GlobalHandler::getPlatformToDefaultContextCacheMutex() {
+  return getOrCreate(MPlatformToDefaultContextCacheMutex);
+}
+
 Sync &GlobalHandler::getSync() { return getOrCreate(MSync); }
 
 std::vector<PlatformImplPtr> &GlobalHandler::getPlatformCache() {
@@ -84,6 +93,16 @@ void shutdown() {
   // First, release resources, that may access plugins.
   GlobalHandler::instance().MScheduler.Inst.reset(nullptr);
   GlobalHandler::instance().MProgramManager.Inst.reset(nullptr);
+#ifndef _WIN32
+  GlobalHandler::instance().MPlatformToDefaultContextCache.Inst.reset(nullptr);
+#else
+  // Windows does not maintain dependencies between dynamically loaded libraries
+  // and can unload SYCL runtime dependencies before sycl.dll's DllMain has
+  // finished. To avoid calls to nowhere, intentionally leak platform to device
+  // cache. This will prevent destructors from being called, thus no PI cleanup
+  // routines will be called in the end.
+  GlobalHandler::instance().MPlatformToDefaultContextCache.Inst.release();
+#endif
   GlobalHandler::instance().MPlatformCache.Inst.reset(nullptr);
 
   // Call to GlobalHandler::instance().getPlugins() initializes plugins. If

sycl/source/detail/global_handler.hpp

@@ -12,11 +12,13 @@
 #include <CL/sycl/detail/util.hpp>
 
 #include <memory>
+#include <unordered_map>
 
 __SYCL_INLINE_NAMESPACE(cl) {
 namespace sycl {
 namespace detail {
 class platform_impl;
+class context_impl;
 class Scheduler;
 class ProgramManager;
 class Sync;
@@ -25,6 +27,7 @@ class device_filter_list;
 class XPTIRegistry;
 
 using PlatformImplPtr = std::shared_ptr<platform_impl>;
+using ContextImplPtr = std::shared_ptr<context_impl>;
 
 /// Wrapper class for global data structures with non-trivial destructors.
 ///
@@ -53,6 +56,11 @@ class GlobalHandler {
   ProgramManager &getProgramManager();
   Sync &getSync();
   std::vector<PlatformImplPtr> &getPlatformCache();
+
+  std::unordered_map<PlatformImplPtr, ContextImplPtr> &
+  getPlatformToDefaultContextCache();
+
+  std::mutex &getPlatformToDefaultContextCacheMutex();
   std::mutex &getPlatformMapMutex();
   std::mutex &getFilterMutex();
   std::vector<plugin> &getPlugins();
@@ -80,6 +88,9 @@ class GlobalHandler {
   InstWithLock<ProgramManager> MProgramManager;
   InstWithLock<Sync> MSync;
   InstWithLock<std::vector<PlatformImplPtr>> MPlatformCache;
+  InstWithLock<std::unordered_map<PlatformImplPtr, ContextImplPtr>>
+      MPlatformToDefaultContextCache;
+  InstWithLock<std::mutex> MPlatformToDefaultContextCacheMutex;
   InstWithLock<std::mutex> MPlatformMapMutex;
   InstWithLock<std::mutex> MFilterMutex;
   InstWithLock<std::vector<plugin>> MPlugins;

sycl/source/detail/platform_impl.cpp

@@ -17,6 +17,7 @@
 
 #include <algorithm>
 #include <cstring>
+#include <mutex>
 #include <string>
 #include <vector>
 

sycl/source/detail/platform_impl.hpp

@@ -7,6 +7,7 @@
 //===----------------------------------------------------------------------===//
 
 #pragma once
+
 #include <CL/sycl/detail/common.hpp>
 #include <CL/sycl/detail/pi.hpp>
 #include <CL/sycl/info/info_desc.hpp>

sycl/source/detail/queue_impl.hpp

@@ -20,6 +20,7 @@
 #include <CL/sycl/properties/queue_properties.hpp>
 #include <CL/sycl/property_list.hpp>
 #include <CL/sycl/stl.hpp>
+#include <detail/config.hpp>
 #include <detail/context_impl.hpp>
 #include <detail/device_impl.hpp>
 #include <detail/event_impl.hpp>
@@ -50,6 +51,22 @@ enum QueueOrder { Ordered, OOO };
 
 class queue_impl {
 public:
+  // \return a default context for the platform if it includes the device
+  // passed and default contexts are enabled, a new context otherwise.
+  static ContextImplPtr getDefaultOrNew(const DeviceImplPtr &Device) {
+    if (!SYCLConfig<SYCL_ENABLE_DEFAULT_CONTEXTS>::get())
+      return detail::getSyclObjImpl(
+          context{createSyclObjFromImpl<device>(Device), {}, {}});
+
+    ContextImplPtr DefaultContext = detail::getSyclObjImpl(
+        Device->get_platform().ext_oneapi_get_default_context());
+
+    if (DefaultContext->hasDevice(Device))
+      return DefaultContext;
+
+    return detail::getSyclObjImpl(
+        context{createSyclObjFromImpl<device>(Device), {}, {}});
+  }
   /// Constructs a SYCL queue from a device using an async_handler and
   /// property_list provided.
   ///
@@ -59,14 +76,7 @@ class queue_impl {
   /// \param PropList is a list of properties to use for queue construction.
   queue_impl(const DeviceImplPtr &Device, const async_handler &AsyncHandler,
              const property_list &PropList)
-      : queue_impl(Device,
-                   detail::getSyclObjImpl(
-                       context(createSyclObjFromImpl<device>(Device), {},
-                               (DefaultContextType == CUDAContextT::primary)
-                                   ? property_list{property::context::cuda::
-                                                       use_primary_context()}
-                                   : property_list{})),
-                   AsyncHandler, PropList){};
+      : queue_impl(Device, getDefaultOrNew(Device), AsyncHandler, PropList){};
 
   /// Constructs a SYCL queue with an async_handler and property_list provided
   /// form a device and a context.

sycl/source/detail/scheduler/scheduler.cpp

@@ -420,8 +420,10 @@ void Scheduler::deallocateStreamBuffers(stream_impl *Impl) {
 
 Scheduler::Scheduler() {
   sycl::device HostDevice;
+  sycl::context HostContext{HostDevice};
   DefaultHostQueue = QueueImplPtr(
-      new queue_impl(detail::getSyclObjImpl(HostDevice), /*AsyncHandler=*/{},
+      new queue_impl(detail::getSyclObjImpl(HostDevice),
+                     detail::getSyclObjImpl(HostContext), /*AsyncHandler=*/{},
                      /*PropList=*/{}));
 }
 

sycl/source/platform.cpp

@@ -11,7 +11,9 @@
 #include <CL/sycl/info/info_desc.hpp>
 #include <CL/sycl/platform.hpp>
 #include <detail/backend_impl.hpp>
+#include <detail/config.hpp>
 #include <detail/force_device.hpp>
+#include <detail/global_handler.hpp>
 #include <detail/platform_impl.hpp>
 
 __SYCL_INLINE_NAMESPACE(cl) {
@@ -65,5 +67,26 @@ bool platform::has(aspect Aspect) const { return impl->has(Aspect); }
 
 #undef __SYCL_PARAM_TRAITS_SPEC
 
+context platform::ext_oneapi_get_default_context() const {
+  if (!detail::SYCLConfig<detail::SYCL_ENABLE_DEFAULT_CONTEXTS>::get())
+    throw std::runtime_error("SYCL default contexts are not enabled");
+
+  // Keeping the default context for platforms in the global cache to avoid
+  // shared_ptr based circular dependency between platform and context classes
+  std::unordered_map<detail::PlatformImplPtr, detail::ContextImplPtr>
+      &PlatformToDefaultContextCache =
+          detail::GlobalHandler::instance().getPlatformToDefaultContextCache();
+
+  std::lock_guard Lock{detail::GlobalHandler::instance()
+                           .getPlatformToDefaultContextCacheMutex()};
+
+  auto It = PlatformToDefaultContextCache.find(impl);
+  if (PlatformToDefaultContextCache.end() == It)
+    std::tie(It, std::ignore) = PlatformToDefaultContextCache.insert(
+        {impl, detail::getSyclObjImpl(context{get_devices()})});
+
+  return detail::createSyclObjFromImpl<context>(It->second);
+}
+
 } // namespace sycl
 } // __SYCL_INLINE_NAMESPACE(cl)

sycl/test/abi/sycl_symbols_linux.dump

@@ -4336,6 +4336,7 @@ _ZNK2cl4sycl7samplerneERKS1_
 _ZNK2cl4sycl8platform11get_backendEv
 _ZNK2cl4sycl8platform11get_devicesENS0_4info11device_typeE
 _ZNK2cl4sycl8platform13has_extensionERKNSt7__cxx1112basic_stringIcSt11char_traitsIcESaIcEEE
+_ZNK2cl4sycl8platform30ext_oneapi_get_default_contextEv
 _ZNK2cl4sycl8platform3getEv
 _ZNK2cl4sycl8platform3hasENS0_6aspectE
 _ZNK2cl4sycl8platform7is_hostEv

sycl/test/abi/sycl_symbols_windows.dump

@@ -1747,6 +1747,7 @@
 ?extractArgsAndReqs@handler@sycl@cl@@AEAAXXZ
 ?extractArgsAndReqsFromLambda@handler@sycl@cl@@AEAAXPEAD_KPEBUkernel_param_desc_t@detail@23@@Z
 ?extractArgsAndReqsFromLambda@handler@sycl@cl@@AEAAXPEAD_KPEBUkernel_param_desc_t@detail@23@_N@Z
+?ext_oneapi_get_default_context@platform@sycl@cl@@QEBA?AVcontext@23@XZ
 ?fabs@__host_std@cl@@YA?AV?$vec@M$00@sycl@2@V342@@Z
 ?fabs@__host_std@cl@@YA?AV?$vec@M$01@sycl@2@V342@@Z
 ?fabs@__host_std@cl@@YA?AV?$vec@M$02@sycl@2@V342@@Z

sycl/unittests/CMakeLists.txt

@@ -22,3 +22,4 @@ add_subdirectory(SYCL2020)
 add_subdirectory(thread_safety)
 add_subdirectory(program_manager)
 add_subdirectory(assert)
+add_subdirectory(Extensions)

sycl/unittests/Extensions/CMakeLists.txt

@@ -0,0 +1,8 @@
+set(CMAKE_CXX_EXTENSIONS OFF)
+
+# Enable exception handling for these unit tests
+set(LLVM_REQUIRES_EH 1)
+add_sycl_unittest(ExtensionsTests OBJECT
+  DefaultContext.cpp
+)
+