[SYCL] USM shared memory allocator for L0 plugin (#2366)

* [SYCL] USM shared memory allocator for L0 plugin

In L0 each allocation results in at least a memory page
no matter how small the size was requested. This adds
significant overhead when an app does many small allocations.

This patch adds a memory allocator on top of L0 USM API in
order to solve the problem.

High level description:
There is a predefine list of bucket sizes. When allocation function
is called, the best fitted bucket is found for the requested size.
If there is a free chunk in the bucket then it's returned, otherwise
a new slab of size 64k requested from the system via L0 API.
This slab is split into chunks of the size of corresponding bucket.
Now there is a free chunk in the slab which is returned.
When a chunk is returned it's marked as non-free until the user
call free with a corresponding pointer.

In addition to that add an environment variable
SYCL_PI_LEVEL0_DISABLE_USM_ALLOCATOR which disables the allocator
and returns back to the original behavior.

* [SYCL] Restrict USM allocators tests to Level zero

There are issues with running these tests on OpenCL
on windows.
The primary goal of these tests is to test USM allocator
which is currently emabled only for Level zero, so disable
them on OpenCL for now.

* [SYCL] Fix pointer alignment

Due to incorrect pointer alignment in case when
allcation alignment > size we returned to the user
and incorrect chunk(i.e. next to the one which was
marked as non-free) which is later detected as "double free"
error.

Also corrected the existing testcases to test this scenario

Author: Ilya Stepykin

sycl/doc/EnvironmentVariables.md

@@ -26,6 +26,7 @@ subject to change. Do not rely on these variables in production code.
 | SYCL_QUEUE_THREAD_POOL_SIZE | Positive integer | Number of threads in thread pool of queue. |
 | SYCL_DEVICELIB_NO_FALLBACK | Any(\*) | Disable loading and linking of device library images |
 | SYCL_PI_LEVEL0_MAX_COMMAND_LIST_CACHE | Positive integer | Maximum number of oneAPI Level Zero Command lists that can be allocated with no reuse before throwing an "out of resources" error. Default is 20000, threshold may be increased based on resource availabilty and workload demand. |
+| SYCL_PI_LEVEL0_DISABLE_USM_ALLOCATOR | Any(\*) | Disable USM allocator in Level Zero plugin (each memory request will go directly to Level Zero runtime) |
 
 `(*) Note: Any means this environment variable is effective when set to any non-null value.`
 

sycl/plugins/level_zero/CMakeLists.txt

@@ -75,6 +75,8 @@ add_library(pi_level_zero SHARED
   "${sycl_inc_dir}/CL/sycl/detail/pi.h"
   "${CMAKE_CURRENT_SOURCE_DIR}/pi_level_zero.cpp"
   "${CMAKE_CURRENT_SOURCE_DIR}/pi_level_zero.hpp"
+  "${CMAKE_CURRENT_SOURCE_DIR}/usm_allocator.cpp"
+  "${CMAKE_CURRENT_SOURCE_DIR}/usm_allocator.hpp"
 )
 
 if (MSVC)

sycl/plugins/level_zero/pi_level_zero.cpp

@@ -22,6 +22,8 @@
 
 #include <level_zero/zet_api.h>
 
+#include "usm_allocator.hpp"
+
 namespace {
 
 // Controls Level Zero calls serialization to w/a Level Zero driver being not MT
@@ -1491,10 +1493,16 @@ pi_result piContextRelease(pi_context Context) {
 
   assert(Context);
   if (--(Context->RefCount) == 0) {
+    auto ZeContext = Context->ZeContext;
     // Destroy the command list used for initializations
     ZE_CALL(zeCommandListDestroy(Context->ZeCommandListInit));
-    ZE_CALL(zeContextDestroy(Context->ZeContext));
     delete Context;
+
+    // Destruction of some members of pi_context uses L0 context
+    // and therefore it must be valid at that point.
+    // Technically it should be placed to the destructor of pi_context
+    // but this makes API error handling more complex.
+    ZE_CALL(zeContextDestroy(ZeContext));
   }
   return PI_SUCCESS;
 }
@@ -4052,7 +4060,6 @@ pi_result piextGetDeviceFunctionPointer(pi_device Device, pi_program Program,
 pi_result piextUSMHostAlloc(void **ResultPtr, pi_context Context,
                             pi_usm_mem_properties *Properties, size_t Size,
                             pi_uint32 Alignment) {
-
   assert(Context);
   // Check that incorrect bits are not set in the properties.
   assert(!Properties || (Properties && !(*Properties & ~PI_MEM_ALLOC_FLAGS)));
@@ -4066,11 +4073,17 @@ pi_result piextUSMHostAlloc(void **ResultPtr, pi_context Context,
   return PI_SUCCESS;
 }
 
-pi_result piextUSMDeviceAlloc(void **ResultPtr, pi_context Context,
-                              pi_device Device,
-                              pi_usm_mem_properties *Properties, size_t Size,
-                              pi_uint32 Alignment) {
+static bool ShouldUseUSMAllocator() {
+  // Enable allocator by default if it's not explicitly disabled
+  return std::getenv("SYCL_PI_LEVEL0_DISABLE_USM_ALLOCATOR") == nullptr;
+}
+
+static const bool UseUSMAllocator = ShouldUseUSMAllocator();
 
+pi_result USMDeviceAllocImpl(void **ResultPtr, pi_context Context,
+                             pi_device Device,
+                             pi_usm_mem_properties *Properties, size_t Size,
+                             pi_uint32 Alignment) {
   assert(Context);
   assert(Device);
   // Check that incorrect bits are not set in the properties.
@@ -4086,11 +4099,10 @@ pi_result piextUSMDeviceAlloc(void **ResultPtr, pi_context Context,
   return PI_SUCCESS;
 }
 
-pi_result piextUSMSharedAlloc(void **ResultPtr, pi_context Context,
-                              pi_device Device,
-                              pi_usm_mem_properties *Properties, size_t Size,
-                              pi_uint32 Alignment) {
-
+pi_result USMSharedAllocImpl(void **ResultPtr, pi_context Context,
+                             pi_device Device,
+                             pi_usm_mem_properties *Properties, size_t Size,
+                             pi_uint32 Alignment) {
   assert(Context);
   assert(Device);
   // Check that incorrect bits are not set in the properties.
@@ -4108,11 +4120,170 @@ pi_result piextUSMSharedAlloc(void **ResultPtr, pi_context Context,
   return PI_SUCCESS;
 }
 
-pi_result piextUSMFree(pi_context Context, void *Ptr) {
+pi_result USMFreeImpl(pi_context Context, void *Ptr) {
   ZE_CALL(zeMemFree(Context->ZeContext, Ptr));
   return PI_SUCCESS;
 }
 
+// Exception type to pass allocation errors
+class UsmAllocationException {
+  const pi_result Error;
+
+public:
+  UsmAllocationException(pi_result Err) : Error{Err} {}
+  pi_result getError() const { return Error; }
+};
+
+pi_result USMSharedMemoryAlloc::allocateImpl(void **ResultPtr, size_t Size,
+                                             pi_uint32 Alignment) {
+  return USMSharedAllocImpl(ResultPtr, Context, Device, nullptr, Size,
+                            Alignment);
+}
+
+pi_result USMDeviceMemoryAlloc::allocateImpl(void **ResultPtr, size_t Size,
+                                             pi_uint32 Alignment) {
+  return USMDeviceAllocImpl(ResultPtr, Context, Device, nullptr, Size,
+                            Alignment);
+}
+
+void *USMMemoryAllocBase::allocate(size_t Size) {
+  void *Ptr = nullptr;
+
+  auto Res = allocateImpl(&Ptr, Size, sizeof(void *));
+  if (Res != PI_SUCCESS) {
+    throw UsmAllocationException(Res);
+  }
+
+  return Ptr;
+}
+
+void *USMMemoryAllocBase::allocate(size_t Size, size_t Alignment) {
+  void *Ptr = nullptr;
+
+  auto Res = allocateImpl(&Ptr, Size, Alignment);
+  if (Res != PI_SUCCESS) {
+    throw UsmAllocationException(Res);
+  }
+  return Ptr;
+}
+
+void USMMemoryAllocBase::deallocate(void *Ptr) {
+  auto Res = USMFreeImpl(Context, Ptr);
+  if (Res != PI_SUCCESS) {
+    throw UsmAllocationException(Res);
+  }
+}
+
+pi_result piextUSMDeviceAlloc(void **ResultPtr, pi_context Context,
+                              pi_device Device,
+                              pi_usm_mem_properties *Properties, size_t Size,
+                              pi_uint32 Alignment) {
+  if (!UseUSMAllocator ||
+      // L0 spec says that allocation fails if Alignment != 2^n, in order to
+      // keep the same behavior for the allocator, just call L0 API directly and
+      // return the error code.
+      ((Alignment & (Alignment - 1)) != 0)) {
+    return USMDeviceAllocImpl(ResultPtr, Context, Device, Properties, Size,
+                              Alignment);
+  }
+
+  try {
+    auto It = Context->DeviceMemAllocContexts.find(Device);
+    if (It == Context->DeviceMemAllocContexts.end())
+      return PI_INVALID_VALUE;
+
+    *ResultPtr = It->second.allocate(Size, Alignment);
+  } catch (const UsmAllocationException &Ex) {
+    *ResultPtr = nullptr;
+    return Ex.getError();
+  }
+
+  return PI_SUCCESS;
+}
+
+pi_result piextUSMSharedAlloc(void **ResultPtr, pi_context Context,
+                              pi_device Device,
+                              pi_usm_mem_properties *Properties, size_t Size,
+                              pi_uint32 Alignment) {
+  if (!UseUSMAllocator ||
+      // L0 spec says that allocation fails if Alignment != 2^n, in order to
+      // keep the same behavior for the allocator, just call L0 API directly and
+      // return the error code.
+      ((Alignment & (Alignment - 1)) != 0)) {
+    return USMSharedAllocImpl(ResultPtr, Context, Device, Properties, Size,
+                              Alignment);
+  }
+
+  try {
+    auto It = Context->SharedMemAllocContexts.find(Device);
+    if (It == Context->SharedMemAllocContexts.end())
+      return PI_INVALID_VALUE;
+
+    *ResultPtr = It->second.allocate(Size, Alignment);
+  } catch (const UsmAllocationException &Ex) {
+    *ResultPtr = nullptr;
+    return Ex.getError();
+  }
+
+  return PI_SUCCESS;
+}
+
+pi_result piextUSMFree(pi_context Context, void *Ptr) {
+  if (!UseUSMAllocator) {
+    return USMFreeImpl(Context, Ptr);
+  }
+
+  // Query the device of the allocation to determine the right allocator context
+  ze_device_handle_t ZeDeviceHandle;
+  ze_memory_allocation_properties_t ZeMemoryAllocationProperties = {};
+
+  // Query memory type of the pointer we're freeing to determine the correct
+  // way to do it(directly or via the allocator)
+  ZE_CALL(zeMemGetAllocProperties(
+      Context->ZeContext, Ptr, &ZeMemoryAllocationProperties, &ZeDeviceHandle));
+
+  // TODO: when support for multiple devices is implemented, here
+  // we should do the following:
+  // - Find pi_device instance corresponding to ZeDeviceHandle we've just got if
+  // exist
+  // - Use that pi_device to find the right allocator context and free the
+  // pointer.
+
+  // The allocation doesn't belong to any device for which USM allocator is
+  // enabled.
+  if (Context->Device->ZeDevice != ZeDeviceHandle) {
+    return USMFreeImpl(Context, Ptr);
+  }
+
+  auto DeallocationHelper =
+      [Context,
+       Ptr](std::unordered_map<pi_device, USMAllocContext> &AllocContextMap) {
+        try {
+          auto It = AllocContextMap.find(Context->Device);
+          if (It == AllocContextMap.end())
+            return PI_INVALID_VALUE;
+
+          // The right context is found, deallocate the pointer
+          It->second.deallocate(Ptr);
+        } catch (const UsmAllocationException &Ex) {
+          return Ex.getError();
+        }
+
+        return PI_SUCCESS;
+      };
+
+  switch (ZeMemoryAllocationProperties.type) {
+  case ZE_MEMORY_TYPE_SHARED:
+    return DeallocationHelper(Context->SharedMemAllocContexts);
+  case ZE_MEMORY_TYPE_DEVICE:
+    return DeallocationHelper(Context->DeviceMemAllocContexts);
+  default:
+    // Handled below
+    break;
+  }
+  return USMFreeImpl(Context, Ptr);
+}
+
 pi_result piextKernelSetArgPointer(pi_kernel Kernel, pi_uint32 ArgIndex,
                                    size_t ArgSize, const void *ArgValue) {
 

sycl/plugins/level_zero/pi_level_zero.hpp

@@ -32,6 +32,8 @@
 
 #include <level_zero/ze_api.h>
 
+#include "usm_allocator.hpp"
+
 template <class To, class From> To pi_cast(From Value) {
   // TODO: see if more sanity checks are possible.
   assert(sizeof(From) == sizeof(To));
@@ -89,6 +91,46 @@ struct _pi_platform {
   std::atomic<int> ZeGlobalCommandListCount{0};
 };
 
+// Implements memory allocation via L0 RT for USM allocator interface.
+class USMMemoryAllocBase : public SystemMemory {
+protected:
+  pi_context Context;
+  pi_device Device;
+  // Internal allocation routine which must be implemented for each allocation
+  // type
+  virtual pi_result allocateImpl(void **ResultPtr, size_t Size,
+                                 pi_uint32 Alignment) = 0;
+
+public:
+  USMMemoryAllocBase(pi_context Ctx, pi_device Dev)
+      : Context{Ctx}, Device{Dev} {}
+  void *allocate(size_t Size) override final;
+  void *allocate(size_t Size, size_t Alignment) override final;
+  void deallocate(void *Ptr) override final;
+};
+
+// Allocation routines for shared memory type
+class USMSharedMemoryAlloc : public USMMemoryAllocBase {
+protected:
+  pi_result allocateImpl(void **ResultPtr, size_t Size,
+                         pi_uint32 Alignment) override;
+
+public:
+  USMSharedMemoryAlloc(pi_context Ctx, pi_device Dev)
+      : USMMemoryAllocBase(Ctx, Dev) {}
+};
+
+// Allocation routines for device memory type
+class USMDeviceMemoryAlloc : public USMMemoryAllocBase {
+protected:
+  pi_result allocateImpl(void **ResultPtr, size_t Size,
+                         pi_uint32 Alignment) override;
+
+public:
+  USMDeviceMemoryAlloc(pi_context Ctx, pi_device Dev)
+      : USMMemoryAllocBase(Ctx, Dev) {}
+};
+
 struct _pi_device : _pi_object {
   _pi_device(ze_device_handle_t Device, pi_platform Plt,
              bool isSubDevice = false)
@@ -145,7 +187,19 @@ struct _pi_device : _pi_object {
 struct _pi_context : _pi_object {
   _pi_context(pi_device Device)
       : Device{Device}, ZeCommandListInit{nullptr}, ZeEventPool{nullptr},
-        NumEventsAvailableInEventPool{}, NumEventsLiveInEventPool{} {}
+        NumEventsAvailableInEventPool{}, NumEventsLiveInEventPool{} {
+    // TODO: when support for multiple devices is added, here we should
+    // loop over all the devices and initialize allocator context for each
+    // pair (context, device)
+    SharedMemAllocContexts.emplace(
+        std::piecewise_construct, std::make_tuple(Device),
+        std::make_tuple(std::unique_ptr<SystemMemory>(
+            new USMSharedMemoryAlloc(this, Device))));
+    DeviceMemAllocContexts.emplace(
+        std::piecewise_construct, std::make_tuple(Device),
+        std::make_tuple(std::unique_ptr<SystemMemory>(
+            new USMDeviceMemoryAlloc(this, Device))));
+  }
 
   // A L0 context handle is primarily used during creation and management of
   // resources that may be used by multiple devices.
@@ -174,6 +228,12 @@ struct _pi_context : _pi_object {
   // and destroy the pool if there are no alive events.
   ze_result_t decrementAliveEventsInPool(ze_event_pool_handle_t pool);
 
+  // Store USM allocator context(internal allocator structures)
+  // for USM shared/host and device allocations. There is 1 allocator context
+  // per each pair of (context, device) per each memory type.
+  std::unordered_map<pi_device, USMAllocContext> SharedMemAllocContexts;
+  std::unordered_map<pi_device, USMAllocContext> DeviceMemAllocContexts;
+
 private:
   // Following member variables are used to manage assignment of events
   // to event pools.