Merge pull request #289 from bbernhar/roll_gpgmm

Roll GPGMM

Author: fujunwei

DEPS

@@ -55,7 +55,7 @@ deps = {
   },
   # GPGMM support for fast DML allocation and residency management.
   'third_party/gpgmm': {
-    'url': '{github_git}/intel/gpgmm.git@61fcfcbd872b7643423ac8c9555f3a6a366904ee',
+    'url': '{github_git}/intel/gpgmm.git@ff03c9a0cea262e534d25257512f2ba2bdb8c2d4',
     'condition': 'checkout_win',
   },
   'third_party/oneDNN': {

src/webnn/native/dmlx/deps/src/device.cpp

@@ -21,7 +21,7 @@ SVDescriptorHeap::SVDescriptorHeap(
 
 ID3D12DescriptorHeap* SVDescriptorHeap::GetDescriptorHeap() const {
     ComPtr<ID3D12DescriptorHeap> descriptorHeap;
-    m_Heap->As(&descriptorHeap);
+    m_Heap.As(&descriptorHeap);
     return descriptorHeap.Get();
 }
 
@@ -584,8 +584,8 @@ HRESULT Device::ExecuteCommandListAndWait()
 
     ID3D12CommandList* commandLists[] = { m_commandList.Get() };
     if (m_residencyManager != nullptr){
-        gpgmm::d3d12::ResidencySet* residencySets[] = { &m_residencySet };
-        m_residencyManager->ExecuteCommandLists(m_commandQueue.Get(), commandLists, residencySets, ARRAYSIZE(commandLists));
+        gpgmm::d3d12::ResidencyList* residencyLists[] = { &m_residencyList };
+        m_residencyManager->ExecuteCommandLists(m_commandQueue.Get(), commandLists, residencyLists, ARRAYSIZE(commandLists));
     } else {
         m_commandQueue->ExecuteCommandLists(ARRAYSIZE(commandLists), commandLists);
     }
@@ -594,7 +594,7 @@ HRESULT Device::ExecuteCommandListAndWait()
 
     ReturnIfFailed(m_commandAllocator->Reset());
     ReturnIfFailed(m_commandList->Reset(m_commandAllocator.Get(), nullptr));
-    ReturnIfFailed(m_residencySet.Reset());
+    ReturnIfFailed(m_residencyList.Reset());
     return S_OK;
 }
 
@@ -643,7 +643,7 @@ HRESULT Device::EnsureCpuBufferSize(uint64_t requestedSizeInBytes, _Inout_ ComPt
         ReturnIfFailed(CreateCpuCustomBuffer(m_resourceAllocator.Get(), newSize, buffer));
     }
 
-    UpdateResidencyIfNeeded(buffer.Get(), &m_residencySet);
+    UpdateResidencyIfNeeded(buffer.Get(), &m_residencyList);
 
     return S_OK;
 }
@@ -660,7 +660,7 @@ HRESULT Device::EnsureDefaultBufferSize(uint64_t requestedSizeInBytes, _Inout_ C
         ReturnIfFailed(CreateDefaultBuffer(m_resourceAllocator.Get(), newSize, buffer));
     }
 
-    UpdateResidencyIfNeeded(buffer.Get(), &m_residencySet);
+    UpdateResidencyIfNeeded(buffer.Get(), &m_residencyList);
 
     return S_OK;
 }
@@ -719,7 +719,7 @@ HRESULT Device::EnsureReadBackHeapSize(uint64_t requestedSizeInBytes)
         ReturnIfFailed(CreateReadBackBuffer(m_resourceAllocator.Get(), newSize, m_readbackHeap));
     }
 
-    UpdateResidencyIfNeeded(m_readbackHeap.Get(), &m_residencySet);
+    UpdateResidencyIfNeeded(m_readbackHeap.Get(), &m_residencyList);
 
     return S_OK;
 }

src/webnn/native/dmlx/deps/src/device.h

@@ -98,7 +98,7 @@ namespace pydml
         Microsoft::WRL::ComPtr<gpgmm::d3d12::ResourceAllocation> m_temporaryResource;
         Microsoft::WRL::ComPtr<gpgmm::d3d12::ResourceAllocation> m_persistentResource;
 
-        gpgmm::d3d12::ResidencySet m_residencySet;
+        gpgmm::d3d12::ResidencyList m_residencyList;
 
         bool m_useCpuCustomHeapResources = false;
         bool m_useGpu = true;

src/webnn/native/dmlx/deps/src/util.h

@@ -116,9 +116,9 @@ struct DmlBufferArrayBinding
     }
 };
 
-inline void UpdateResidencyIfNeeded(gpgmm::d3d12::ResourceAllocation* resource, gpgmm::d3d12::ResidencySet* residencySet){
+inline void UpdateResidencyIfNeeded(gpgmm::d3d12::ResourceAllocation* resource, gpgmm::d3d12::ResidencyList* residencyList){
     if (resource == nullptr) return;
-    residencySet->Insert(resource->GetMemory());
+    residencyList->Add(resource->GetMemory());
 }
 
 inline HRESULT CreateResource(