[SYCL][XPTI] Report memory allocation info from SYCL runtime

sycl/doc/SYCLInstrumentationUsingXPTI.md

@@ -256,3 +256,12 @@ All trace point types in bold provide semantic information about the graph, node
 | `wait_end`         | <div style="text-align: left"><li>**trace_type**: `xpti::trace_point_type_t::wait_end` that marks the beginning of the wait on an `event`</li> <li> **parent**: `nullptr`</li> <li> **event**: The event ID will reflect the ID of the command group object submission that created this event or a new event based on the combination of the string "queue.wait" and the address of the event. </li> <li> **instance**: Unique ID to allow the correlation of the `wait_begin` event with the `wait_end` event. </li> <li> **user_data**: String indicating `queue.wait` and the address of the event  as `const char *` </li></div>                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                   | **`sycl_device`**, `sym_function_name`, `sym_source_file_name`, `sym_line_no`                                                                                                                                                                            |
 | `barrier_begin`       | <div style="text-align: left"><li>**trace_type**: `xpti::trace_point_type_t::barrier_begin` that marks the beginning of a barrier while enqueuing a command group object</li> <li> **parent**: The global graph event that is created during the `graph_create` event.</li> <li> **event**: The event ID will reflect the ID of the command group object that has encountered a barrier during the enqueue operation.  </li> <li> **instance**: Unique ID to allow the correlation of the `barrier_begin` event with the `barrier_end` event. </li> <li> **user_data**: String indicating `enqueue.barrier` and the reason for the barrier as a `const char *` </li> <p></p>The reason for the barrier could be one of `Buffer locked by host accessor`, `Blocked by host task` or `Unknown reason`.</div>   | <li> Computational Kernels </li> `sycl_device`, `kernel_name`, `from_source`, `sym_function_name`, `sym_source_file_name`, `sym_line_no` <li>Memory operations</li> `memory_object`, `offset`, `access_range`, `allocation_type`, `copy_from`, `copy_to` |
 | `barrier_end`         | <div style="text-align: left"><li>**trace_type**: `xpti::trace_point_type_t::barrier_end` that marks the end of the barrier that is encountered during enqueue.</li> <li> **parent**: The global graph event that is created during the `graph_create` event.</li> <li> **event**: The event ID will reflect the ID of the command group object that has encountered a barrier during the enqueue operation. </li> <li> **instance**: Unique ID to allow the correlation of the `barrier_begin` event with the `barrier_end` event. </li> <li> **user_data**: String indicating `enqueue.barrier` and the reason for the barrier as a `const char *` </li> <p></p>The reason for the barrier could be one of `Buffer locked by host accessor`, `Blocked by host task` or `Unknown reason`.</div>   | <li> Computational Kernels </li> `sycl_device`, `kernel_name`, `from_source`, `sym_function_name`, `sym_source_file_name`, `sym_line_no` <li>Memory operations</li> `memory_object`, `offset`, `access_range`, `allocation_type`, `copy_from`, `copy_to` |
+
+## Level Zero Plugin Stream `"oneapi.level_zero.experimental.mem_alloc"` Notification Signatures
+
+| Trace Point Type           | Parameter Description | Metadata |
+| :------------------------: | :-------------------- | :------- |
+| `mem_alloc_begin` | <div style="text-align: left"><li>**trace_type**: `xpti::trace_point_type_t::mem_alloc_begin` that marks the beginning of memory allocation process</li> <li> **parent**: Event ID created for all functions in the `oneapi.level_zero.experimental.mem_alloc` layer.</li> <li> **event**: `nullptr` - since the stream of data just captures functions being called.</li> <li> **instance**: Unique ID to allow the correlation of the `mem_alloc_begin` event with the `mem_alloc_end` event. </li> <li> **user_data**: A pointer to `mem_alloc_data_t` object, that includes memory object ID (if any), allocation size, and guard zone size (if any). </li></div> | None     |
+| `mem_alloc_end`   | <div style="text-align: left"><li>**trace_type**: `xpti::trace_point_type_t::mem_alloc_end` that marks the end of memory allocation process</li> <li> **parent**: Event ID created for all functions in the `oneapi.level_zero.experimental.mem_alloc` layer.</li> <li> **event**: `nullptr` - since the stream of data just captures functions being called.</li> <li> **instance**: Unique ID to allow the correlation of the `mem_alloc_begin` event with the `mem_alloc_end` event. This value is guaranteed to be the same value  received by the trace event for the corresponding `mem_alloc_begin`.</li> <li> **user_data**: A pointer to `mem_alloc_data_t` object, that includes memory object ID (if any), allocated pointer, allocation size, and guard zone size (if any). </li></div> | None     |
+| `mem_release_begin` | <div style="text-align: left"><li>**trace_type**: `xpti::trace_point_type_t::mem_release_begin` that marks the beginning of memory allocation process</li> <li> **parent**: Event ID created for all functions in the `oneapi.level_zero.experimental.mem_alloc` layer.</li> <li> **event**: `nullptr` - since the stream of data just captures functions being called.</li> <li> **instance**: Unique ID to allow the correlation of the `mem_release_begin` event with the `mem_release_end` event. </li> <li> **user_data**: A pointer to `mem_alloc_data_t` object, that includes memory object ID (if any) and released pointer. </li></div> | None     |
+| `mem_release_end`   | <div style="text-align: left"><li>**trace_type**: `xpti::trace_point_type_t::mem_release_end` that marks the end of memory allocation process</li> <li> **parent**: Event ID created for all functions in the `oneapi.level_zero.experimental.mem_alloc` layer.</li> <li> **event**: `nullptr` - since the stream of data just captures functions being called.</li> <li> **instance**: Unique ID to allow the correlation of the `mem_release_begin` event with the `mem_release_end` event. This value is guaranteed to be the same value  received by the trace event for the corresponding `mem_release_begin`.</li> <li> **user_data**: A pointer to `mem_alloc_data_t` object, that includes memory object ID (if any) and released pointer. </li></div> | None     |

sycl/source/detail/device_image_impl.hpp

@@ -17,6 +17,7 @@
 #include <detail/context_impl.hpp>
 #include <detail/device_impl.hpp>
 #include <detail/kernel_id_impl.hpp>
+#include <detail/mem_alloc_helper.hpp>
 #include <detail/plugin.hpp>
 #include <detail/program_manager/program_manager.hpp>
 
@@ -185,11 +186,11 @@ class device_image_impl {
     std::lock_guard<std::mutex> Lock{MSpecConstAccessMtx};
     if (nullptr == MSpecConstsBuffer && !MSpecConstsBlob.empty()) {
       const detail::plugin &Plugin = getSyclObjImpl(MContext)->getPlugin();
-      Plugin.call<PiApiKind::piMemBufferCreate>(
-          detail::getSyclObjImpl(MContext)->getHandleRef(),
-          PI_MEM_FLAGS_ACCESS_RW | PI_MEM_FLAGS_HOST_PTR_USE,
-          MSpecConstsBlob.size(), MSpecConstsBlob.data(), &MSpecConstsBuffer,
-          nullptr);
+      memBufferCreateHelper(Plugin,
+                            detail::getSyclObjImpl(MContext)->getHandleRef(),
+                            PI_MEM_FLAGS_ACCESS_RW | PI_MEM_FLAGS_HOST_PTR_USE,
+                            MSpecConstsBlob.size(), MSpecConstsBlob.data(),
+                            &MSpecConstsBuffer, nullptr);
     }
     return MSpecConstsBuffer;
   }

sycl/source/detail/mem_alloc_helper.hpp

@@ -0,0 +1,32 @@
+//==-------- mem_alloc_helper.hpp - SYCL mem alloc helper ------------------==//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+#pragma once
+
+#include <CL/sycl/detail/pi.h>
+
+__SYCL_INLINE_NAMESPACE(cl) {
+namespace sycl {
+namespace detail {
+void memBufferCreateHelper(const plugin &Plugin, pi_context Ctx,
+                           pi_mem_flags Flags, size_t Size, void *HostPtr,
+                           pi_mem *RetMem,
+                           const pi_mem_properties *Props = nullptr);
+void memReleaseHelper(const plugin &Plugin, pi_mem Mem);
+void memBufferMapHelper(const plugin &Plugin, pi_queue command_queue,
+                        pi_mem buffer, pi_bool blocking_map,
+                        pi_map_flags map_flags, size_t offset, size_t size,
+                        pi_uint32 num_events_in_wait_list,
+                        const pi_event *event_wait_list, pi_event *event,
+                        void **ret_map);
+void memUnmapHelper(const plugin &Plugin, pi_queue command_queue, pi_mem memobj,
+                    void *mapped_ptr, pi_uint32 num_events_in_wait_list,
+                    const pi_event *event_wait_list, pi_event *event);
+} // namespace detail
+} // namespace sycl
+} // __SYCL_INLINE_NAMESPACE(cl)

sycl/source/detail/memory_manager.cpp

@@ -9,17 +9,106 @@
 #include <CL/sycl/detail/memory_manager.hpp>
 #include <detail/context_impl.hpp>
 #include <detail/event_impl.hpp>
+#include <detail/mem_alloc_helper.hpp>
 #include <detail/queue_impl.hpp>
 
 #include <algorithm>
 #include <cassert>
 #include <cstring>
 #include <vector>
 
+#ifdef XPTI_ENABLE_INSTRUMENTATION
+#include <xpti/xpti_data_types.h>
+#include <xpti/xpti_trace_framework.hpp>
+#endif
+
 __SYCL_INLINE_NAMESPACE(cl) {
 namespace sycl {
 namespace detail {
 
+#ifdef XPTI_ENABLE_INSTRUMENTATION
+uint8_t GMemAllocStreamID;
+xpti::trace_event_data_t *GMemAllocEvent;
+#endif
+
+uint64_t emitMemAllocBeginTrace(uintptr_t ObjHandle, size_t AllocSize,
+                                size_t GuardZone) {
+  (void)ObjHandle;
+  (void)AllocSize;
+  (void)GuardZone;
+  uint64_t CorrelationID = 0;
+#ifdef XPTI_ENABLE_INSTRUMENTATION
+  if (xptiTraceEnabled()) {
+    xpti::mem_alloc_data_t MemAlloc{ObjHandle, 0 /* alloc ptr */, AllocSize,
+                                    GuardZone};
+
+    CorrelationID = xptiGetUniqueId();
+    xptiNotifySubscribers(
+        GMemAllocStreamID,
+        static_cast<uint16_t>(xpti::trace_point_type_t::mem_alloc_begin),
+        GMemAllocEvent, nullptr, CorrelationID, &MemAlloc);
+  }
+#endif
+  return CorrelationID;
+}
+
+void emitMemAllocEndTrace(uintptr_t ObjHandle, uintptr_t AllocPtr,
+                          size_t AllocSize, size_t GuardZone,
+                          uint64_t CorrelationID) {
+  (void)ObjHandle;
+  (void)AllocPtr;
+  (void)AllocSize;
+  (void)GuardZone;
+  (void)CorrelationID;
+#ifdef XPTI_ENABLE_INSTRUMENTATION
+  if (xptiTraceEnabled()) {
+    xpti::mem_alloc_data_t MemAlloc{ObjHandle, AllocPtr, AllocSize, GuardZone};
+
+    xptiNotifySubscribers(
+        GMemAllocStreamID,
+        static_cast<uint16_t>(xpti::trace_point_type_t::mem_alloc_end),
+        GMemAllocEvent, nullptr, CorrelationID, &MemAlloc);
+  }
+#endif
+}
+
+uint64_t emitMemReleaseBeginTrace(uintptr_t ObjHandle, uintptr_t AllocPtr) {
+  (void)ObjHandle;
+  (void)AllocPtr;
+#ifdef XPTI_ENABLE_INSTRUMENTATION
+  uint64_t CorrelationID = 0;
+  if (xptiTraceEnabled()) {
+    xpti::mem_alloc_data_t MemAlloc{ObjHandle, AllocPtr, 0 /* alloc size */,
+                                    0 /* guard zone */};
+
+    CorrelationID = xptiGetUniqueId();
+    xptiNotifySubscribers(
+        GMemAllocStreamID,
+        static_cast<uint16_t>(xpti::trace_point_type_t::mem_release_begin),
+        GMemAllocEvent, nullptr, CorrelationID, &MemAlloc);
+  }
+#endif
+  return CorrelationID;
+}
+
+void emitMemReleaseEndTrace(uintptr_t ObjHandle, uintptr_t AllocPtr,
+                            uint64_t CorrelationID) {
+  (void)ObjHandle;
+  (void)AllocPtr;
+  (void)CorrelationID;
+#ifdef XPTI_ENABLE_INSTRUMENTATION
+  if (xptiTraceEnabled()) {
+    xpti::mem_alloc_data_t MemAlloc{ObjHandle, AllocPtr, 0 /* alloc size */,
+                                    0 /* guard zone */};
+
+    xptiNotifySubscribers(
+        GMemAllocStreamID,
+        static_cast<uint16_t>(xpti::trace_point_type_t::mem_release_end),
+        GMemAllocEvent, nullptr, CorrelationID, &MemAlloc);
+  }
+#endif
+}
+
 static void waitForEvents(const std::vector<EventImplPtr> &Events) {
   // Assuming all events will be on the same device or
   // devices associated with the same Backend.
@@ -34,6 +123,97 @@ static void waitForEvents(const std::vector<EventImplPtr> &Events) {
   }
 }
 
+void memBufferCreateHelper(const plugin &Plugin, pi_context Ctx,
+                           pi_mem_flags Flags, size_t Size, void *HostPtr,
+                           pi_mem *RetMem, const pi_mem_properties *Props) {
+  uint64_t CorrID = 0;
+  // We only want to instrument piMemBufferCreate
+  {
+    CorrID =
+        emitMemAllocBeginTrace(0 /* mem object */, Size, 0 /* guard zone */);
+    xpti::utils::finally _{[&] {
+      // C-style cast is required for MSVC
+      uintptr_t MemObjID = (uintptr_t)(*RetMem);
+      pi_native_handle Ptr = 0;
+      // Always use call_nocheck here, because call may throw an exception,
+      // and this lambda will be called from destructor, which in combination
+      // rewards us with UB.
+      Plugin.call_nocheck<PiApiKind::piextMemGetNativeHandle>(*RetMem, &Ptr);
+      emitMemAllocEndTrace(MemObjID, (uintptr_t)(Ptr), Size, 0 /* guard zone */,
+                           CorrID);
+    }};
+    Plugin.call<PiApiKind::piMemBufferCreate>(Ctx, Flags, Size, HostPtr, RetMem,
+                                              Props);
+  }
+}
+
+void memReleaseHelper(const plugin &Plugin, pi_mem Mem) {
+  // FIXME piMemRelease does not guarante memory release. It is only true if
+  // reference counter is 1. However, SYCL runtime currently only calls
+  // piMemRetain only for OpenCL interop
+  uint64_t CorrID = 0;
+  // C-style cast is required for MSVC
+  uintptr_t MemObjID = (uintptr_t)(Mem);
+  uintptr_t Ptr = 0;
+  // Do not make unnecessary PI calls without instrumentation enabled
+  if (xptiTraceEnabled()) {
+    pi_native_handle PtrHandle = 0;
+    Plugin.call<PiApiKind::piextMemGetNativeHandle>(Mem, &PtrHandle);
+    Ptr = (uintptr_t)(PtrHandle);
+  }
+  // We only want to instrument piMemRelease
+  {
+    CorrID = emitMemReleaseBeginTrace(MemObjID, Ptr);
+    xpti::utils::finally _{
+        [&] { emitMemReleaseEndTrace(MemObjID, Ptr, CorrID); }};
+    Plugin.call<PiApiKind::piMemRelease>(Mem);
+  }
+}
+
+void memBufferMapHelper(const plugin &Plugin, pi_queue Queue, pi_mem Buffer,
+                        pi_bool Blocking, pi_map_flags Flags, size_t Offset,
+                        size_t Size, pi_uint32 NumEvents,
+                        const pi_event *WaitList, pi_event *Event,
+                        void **RetMap) {
+  uint64_t CorrID = 0;
+  uintptr_t MemObjID = (uintptr_t)(Buffer);
+  // We only want to instrument piEnqueueMemBufferMap
+  {
+    CorrID = emitMemAllocBeginTrace(MemObjID, Size, 0 /* guard zone */);
+    xpti::utils::finally _{[&] {
+      emitMemAllocEndTrace(MemObjID, (uintptr_t)(*RetMap), Size,
+                           0 /* guard zone */, CorrID);
+    }};
+    Plugin.call<PiApiKind::piEnqueueMemBufferMap>(
+        Queue, Buffer, Blocking, Flags, Offset, Size, NumEvents, WaitList,
+        Event, RetMap);
+  }
+}
+
+void memUnmapHelper(const plugin &Plugin, pi_queue Queue, pi_mem Mem,
+                    void *MappedPtr, pi_uint32 NumEvents,
+                    const pi_event *WaitList, pi_event *Event) {
+  uint64_t CorrID = 0;
+  uintptr_t MemObjID = (uintptr_t)(Mem);
+  uintptr_t Ptr = (uintptr_t)(MappedPtr);
+  // We only want to instrument piEnqueueMemUnmap
+  {
+    CorrID = emitMemReleaseBeginTrace(MemObjID, Ptr);
+    xpti::utils::finally _{[&] {
+      // There's no way for SYCL to know, when the pointer is freed, so we have
+      // to explicitly wait for the end of data transfers here in order to
+      // report correct events.
+      // Always use call_nocheck here, because call may throw an exception,
+      // and this lambda will be called from destructor, which in combination
+      // rewards us with UB.
+      Plugin.call_nocheck<PiApiKind::piEventsWait>(1, Event);
+      emitMemReleaseEndTrace(MemObjID, Ptr, CorrID);
+    }};
+    Plugin.call<PiApiKind::piEnqueueMemUnmap>(Queue, Mem, MappedPtr, NumEvents,
+                                              WaitList, Event);
+  }
+}
+
 void MemoryManager::release(ContextImplPtr TargetContext, SYCLMemObjI *MemObj,
                             void *MemAllocation,
                             std::vector<EventImplPtr> DepEvents,
@@ -67,7 +247,7 @@ void MemoryManager::releaseMemObj(ContextImplPtr TargetContext,
   }
 
   const detail::plugin &Plugin = TargetContext->getPlugin();
-  Plugin.call<PiApiKind::piMemRelease>(pi::cast<RT::PiMem>(MemAllocation));
+  memReleaseHelper(Plugin, pi::cast<RT::PiMem>(MemAllocation));
 }
 
 void *MemoryManager::allocate(ContextImplPtr TargetContext, SYCLMemObjI *MemObj,
@@ -165,9 +345,8 @@ MemoryManager::allocateBufferObject(ContextImplPtr TargetContext, void *UserPtr,
 
   RT::PiMem NewMem = nullptr;
   const detail::plugin &Plugin = TargetContext->getPlugin();
-  Plugin.call<PiApiKind::piMemBufferCreate>(TargetContext->getHandleRef(),
-                                            CreationFlags, Size, UserPtr,
-                                            &NewMem, nullptr);
+  memBufferCreateHelper(Plugin, TargetContext->getHandleRef(), CreationFlags,
+                        Size, UserPtr, &NewMem, nullptr);
   return NewMem;
 }
 
@@ -623,10 +802,9 @@ void *MemoryManager::map(SYCLMemObjI *, void *Mem, QueueImplPtr Queue,
   void *MappedPtr = nullptr;
   const size_t BytesToMap = AccessRange[0] * AccessRange[1] * AccessRange[2];
   const detail::plugin &Plugin = Queue->getPlugin();
-  Plugin.call<PiApiKind::piEnqueueMemBufferMap>(
-      Queue->getHandleRef(), pi::cast<RT::PiMem>(Mem), CL_FALSE, Flags,
-      AccessOffset[0], BytesToMap, DepEvents.size(), DepEvents.data(),
-      &OutEvent, &MappedPtr);
+  memBufferMapHelper(Plugin, Queue->getHandleRef(), pi::cast<RT::PiMem>(Mem),
+                     CL_FALSE, Flags, AccessOffset[0], BytesToMap,
+                     DepEvents.size(), DepEvents.data(), &OutEvent, &MappedPtr);
   return MappedPtr;
 }
 
@@ -639,9 +817,8 @@ void MemoryManager::unmap(SYCLMemObjI *, void *Mem, QueueImplPtr Queue,
   // Using the plugin of the Queue.
 
   const detail::plugin &Plugin = Queue->getPlugin();
-  Plugin.call<PiApiKind::piEnqueueMemUnmap>(
-      Queue->getHandleRef(), pi::cast<RT::PiMem>(Mem), MappedPtr,
-      DepEvents.size(), DepEvents.data(), &OutEvent);
+  memUnmapHelper(Plugin, Queue->getHandleRef(), pi::cast<RT::PiMem>(Mem),
+                 MappedPtr, DepEvents.size(), DepEvents.data(), &OutEvent);
 }
 
 void MemoryManager::copy_usm(const void *SrcMem, QueueImplPtr SrcQueue,