[L0] Add initial USM alloc enqueue API

Co-authored-by: Michał Staniewski <[email protected]>

Author: kswiecicki

source/adapters/level_zero/context.cpp

@@ -204,6 +204,17 @@ ur_result_t ur_context_handle_t_::initialize() {
                                  .Configs[usm::DisjointPoolMemType::Device])
                             .second));
 
+    MemProvider = umf::memoryProviderMakeUnique<L0DeviceMemoryProvider>(
+                      reinterpret_cast<ur_context_handle_t>(this), Device)
+                      .second;
+    AsyncDeviceMemPools.emplace(
+        std::piecewise_construct, std::make_tuple(Device->ZeDevice),
+        std::make_tuple(umf::poolMakeUniqueFromOps(
+                            umfDisjointPoolOps(), std::move(MemProvider),
+                            &DisjointPoolConfigInstance
+                                 .Configs[usm::DisjointPoolMemType::Device])
+                            .second));
+
     MemProvider = umf::memoryProviderMakeUnique<L0SharedMemoryProvider>(
                       reinterpret_cast<ur_context_handle_t>(this), Device)
                       .second;
@@ -215,6 +226,17 @@ ur_result_t ur_context_handle_t_::initialize() {
                                  .Configs[usm::DisjointPoolMemType::Shared])
                             .second));
 
+    MemProvider = umf::memoryProviderMakeUnique<L0SharedMemoryProvider>(
+                      reinterpret_cast<ur_context_handle_t>(this), Device)
+                      .second;
+    AsyncSharedMemPools.emplace(
+        std::piecewise_construct, std::make_tuple(Device->ZeDevice),
+        std::make_tuple(umf::poolMakeUniqueFromOps(
+                            umfDisjointPoolOps(), std::move(MemProvider),
+                            &DisjointPoolConfigInstance
+                                 .Configs[usm::DisjointPoolMemType::Shared])
+                            .second));
+
     MemProvider = umf::memoryProviderMakeUnique<L0SharedReadOnlyMemoryProvider>(
                       reinterpret_cast<ur_context_handle_t>(this), Device)
                       .second;
@@ -227,6 +249,18 @@ ur_result_t ur_context_handle_t_::initialize() {
                      .Configs[usm::DisjointPoolMemType::SharedReadOnly])
                 .second));
 
+    MemProvider = umf::memoryProviderMakeUnique<L0SharedReadOnlyMemoryProvider>(
+                      reinterpret_cast<ur_context_handle_t>(this), Device)
+                      .second;
+    AsyncSharedReadOnlyMemPools.emplace(
+        std::piecewise_construct, std::make_tuple(Device->ZeDevice),
+        std::make_tuple(
+            umf::poolMakeUniqueFromOps(
+                umfDisjointPoolOps(), std::move(MemProvider),
+                &DisjointPoolConfigInstance
+                     .Configs[usm::DisjointPoolMemType::SharedReadOnly])
+                .second));
+
     MemProvider = umf::memoryProviderMakeUnique<L0DeviceMemoryProvider>(
                       reinterpret_cast<ur_context_handle_t>(this), Device)
                       .second;
@@ -279,6 +313,15 @@ ur_result_t ur_context_handle_t_::initialize() {
           &DisjointPoolConfigInstance.Configs[usm::DisjointPoolMemType::Host])
           .second;
 
+  MemProvider = umf::memoryProviderMakeUnique<L0HostMemoryProvider>(
+                    reinterpret_cast<ur_context_handle_t>(this), nullptr)
+                    .second;
+  AsyncHostMemPool =
+      umf::poolMakeUniqueFromOps(
+          umfDisjointPoolOps(), std::move(MemProvider),
+          &DisjointPoolConfigInstance.Configs[usm::DisjointPoolMemType::Host])
+          .second;
+
   MemProvider = umf::memoryProviderMakeUnique<L0HostMemoryProvider>(
                     reinterpret_cast<ur_context_handle_t>(this), nullptr)
                     .second;

source/adapters/level_zero/context.hpp

@@ -124,6 +124,15 @@ struct ur_context_handle_t_ : _ur_object {
       SharedReadOnlyMemProxyPools;
   umf::pool_unique_handle_t HostMemProxyPool;
 
+  // USM pools for async allocations.
+  std::unordered_map<ze_device_handle_t, umf::pool_unique_handle_t>
+      AsyncDeviceMemPools;
+  std::unordered_map<ze_device_handle_t, umf::pool_unique_handle_t>
+      AsyncSharedMemPools;
+  std::unordered_map<ze_device_handle_t, umf::pool_unique_handle_t>
+      AsyncSharedReadOnlyMemPools;
+  umf::pool_unique_handle_t AsyncHostMemPool;
+
   // Map associating pools created with urUsmPoolCreate and internal pools
   std::list<ur_usm_pool_handle_t> UsmPoolHandles{};
 

source/adapters/level_zero/queue.cpp

@@ -1333,7 +1333,7 @@ ur_queue_handle_t_::executeCommandList(ur_command_list_ptr_t CommandList,
       Device->Platform->ContextsMutex, std::defer_lock);
 
   if (IndirectAccessTrackingEnabled) {
-    // We are going to submit kernels for execution. If indirect access flag is
+    // We are going to submit kernels for execution.  If indirect access flag is
     // set for a kernel then we need to make a snapshot of existing memory
     // allocations in all contexts in the platform. We need to lock the mutex
     // guarding the list of contexts in the platform to prevent creation of new