diff --git a/WickedEngine/wiGraphicsDevice.h b/WickedEngine/wiGraphicsDevice.h
index 4021a9324..6489091ae 100644
--- a/WickedEngine/wiGraphicsDevice.h
+++ b/WickedEngine/wiGraphicsDevice.h
@@ -307,6 +307,12 @@ namespace wi::graphics
 			Barrier(&barrier, 1, cmd);
 		}
 
+		// Execute a global GPU memory barrier
+		void Barrier(CommandList cmd)
+		{
+			Barrier(GPUBarrier::Memory(), cmd);
+		}
+
 		struct GPULinearAllocator
 		{
 			GPUBuffer buffer;
diff --git a/WickedEngine/wiGraphicsDevice_DX12.cpp b/WickedEngine/wiGraphicsDevice_DX12.cpp
index c548f4e92..956d3e540 100644
--- a/WickedEngine/wiGraphicsDevice_DX12.cpp
+++ b/WickedEngine/wiGraphicsDevice_DX12.cpp
@@ -54,6 +54,7 @@ namespace dx12_internal
 #endif
 	static PFN_D3D12_CREATE_DEVICE D3D12CreateDevice = nullptr;
 	static PFN_D3D12_CREATE_VERSIONED_ROOT_SIGNATURE_DESERIALIZER D3D12CreateVersionedRootSignatureDeserializer = nullptr;
+	static PFN_D3D12_SERIALIZE_ROOT_SIGNATURE D3D12SerializeRootSignature = nullptr;
 #endif // PLATFORM_WINDOWS_DESKTOP
 
 	// Engine -> Native converters
@@ -2190,6 +2191,16 @@ std::mutex queue_locker;
 			wi::platform::Exit();
 		}
 
+		D3D12SerializeRootSignature = (PFN_D3D12_SERIALIZE_ROOT_SIGNATURE)wiGetProcAddress(dx12, "D3D12SerializeRootSignature");
+		assert(D3D12SerializeRootSignature != nullptr);
+		if (D3D12SerializeRootSignature == nullptr)
+		{
+			std::stringstream ss("");
+			ss << "Failed to load D3D12SerializeRootSignature! ERROR: " << std::hex << GetLastError();
+			wi::helper::messageBox(ss.str(), "Error!");
+			wi::platform::Exit();
+		}
+
 		if (validationMode != ValidationMode::Disabled)
 		{
 			// Enable the debug layer.
@@ -2855,6 +2866,42 @@ std::mutex queue_locker;
 			wi::platform::Exit();
 		}
 
+		// Dummy rootsignature is created for validating multidraw command signatures (required):
+		ComPtr<ID3D12RootSignature> rootsig;
+		ComPtr<ID3DBlob> serializedRootSig = nullptr;
+		ComPtr<ID3DBlob> errorBlob = nullptr;
+		D3D12_ROOT_PARAMETER rootparam = {};
+		rootparam.ParameterType = D3D12_ROOT_PARAMETER_TYPE_32BIT_CONSTANTS;
+		rootparam.ShaderVisibility = D3D12_SHADER_VISIBILITY_ALL;
+		rootparam.Constants.Num32BitValues = PUSH_CONSTANT_COUNT;
+		rootparam.Constants.ShaderRegister = 999;
+		D3D12_ROOT_SIGNATURE_DESC rootsig_desc = {};
+		rootsig_desc.Flags = D3D12_ROOT_SIGNATURE_FLAG_ALLOW_INPUT_ASSEMBLER_INPUT_LAYOUT;
+		rootsig_desc.NumParameters = 1;
+		rootsig_desc.pParameters = &rootparam;
+		dx12_check(D3D12SerializeRootSignature(&rootsig_desc, D3D_ROOT_SIGNATURE_VERSION_1, serializedRootSig.GetAddressOf(), errorBlob.GetAddressOf()));
+		dx12_check(device->CreateRootSignature(0, serializedRootSig->GetBufferPointer(), serializedRootSig->GetBufferSize(), PPV_ARGS(rootsig)));
+
+		D3D12_INDIRECT_ARGUMENT_DESC drawInstancedCountArgs[2];
+		drawInstancedCountArgs[0].Type = D3D12_INDIRECT_ARGUMENT_TYPE_CONSTANT;
+		drawInstancedCountArgs[0].Constant.RootParameterIndex = 0;
+		drawInstancedCountArgs[0].Constant.Num32BitValuesToSet = 1;
+		drawInstancedCountArgs[1].Type = D3D12_INDIRECT_ARGUMENT_TYPE_DRAW;
+		cmd_desc.ByteStride = sizeof(uint32_t) + sizeof(D3D12_DRAW_ARGUMENTS);
+		cmd_desc.NumArgumentDescs = arraysize(drawInstancedCountArgs);
+		cmd_desc.pArgumentDescs = drawInstancedCountArgs;
+		dx12_check(device->CreateCommandSignature(&cmd_desc, rootsig.Get(), PPV_ARGS(drawInstancedIndirectCountCommandSignature)));
+
+		D3D12_INDIRECT_ARGUMENT_DESC drawIndexedInstancedCountArgs[2];
+		drawIndexedInstancedCountArgs[0].Type = D3D12_INDIRECT_ARGUMENT_TYPE_CONSTANT;
+		drawIndexedInstancedCountArgs[0].Constant.RootParameterIndex = 0;
+		drawIndexedInstancedCountArgs[0].Constant.Num32BitValuesToSet = 1;
+		drawIndexedInstancedCountArgs[1].Type = D3D12_INDIRECT_ARGUMENT_TYPE_DRAW_INDEXED;
+		cmd_desc.ByteStride = sizeof(uint32_t) + sizeof(D3D12_DRAW_INDEXED_ARGUMENTS);
+		cmd_desc.NumArgumentDescs = arraysize(drawIndexedInstancedCountArgs);
+		cmd_desc.pArgumentDescs = drawIndexedInstancedCountArgs;
+		dx12_check(device->CreateCommandSignature(&cmd_desc, rootsig.Get(), PPV_ARGS(drawIndexedInstancedIndirectCountCommandSignature)));
+
 		if (CheckCapability(GraphicsDeviceCapability::MESH_SHADER))
 		{
 			D3D12_INDIRECT_ARGUMENT_DESC dispatchMeshArgs[1];
@@ -2866,12 +2913,22 @@ std::mutex queue_locker;
 #endif // PLATFORM_XBOX
 			cmd_desc.NumArgumentDescs = 1;
 			cmd_desc.pArgumentDescs = dispatchMeshArgs;
-			hr = dx12_check(device->CreateCommandSignature(&cmd_desc, nullptr, PPV_ARGS(dispatchMeshIndirectCommandSignature)));
-			if (FAILED(hr))
-			{
-				wilog_messagebox("ID3D12Device::CreateCommandSignature[dispatchMeshIndirect] failed! ERROR: %s", wi::helper::GetPlatformErrorString(hr).c_str());
-				wi::platform::Exit();
-			}
+			dx12_check(device->CreateCommandSignature(&cmd_desc, nullptr, PPV_ARGS(dispatchMeshIndirectCommandSignature)));
+
+			D3D12_INDIRECT_ARGUMENT_DESC dispatchMeshCountArgs[2];
+			dispatchMeshCountArgs[0].Type = D3D12_INDIRECT_ARGUMENT_TYPE_CONSTANT;
+			dispatchMeshCountArgs[0].Constant.RootParameterIndex = 0;
+			dispatchMeshCountArgs[0].Constant.Num32BitValuesToSet = 1;
+#ifdef PLATFORM_XBOX
+			wi::graphics::xbox::FillDispatchMeshIndirectArgumentDesc(dispatchMeshCountArgs[1], cmd_desc);
+#else
+			dispatchMeshCountArgs[1].Type = D3D12_INDIRECT_ARGUMENT_TYPE_DISPATCH_MESH;
+			cmd_desc.ByteStride = sizeof(D3D12_DISPATCH_MESH_ARGUMENTS);
+#endif // PLATFORM_XBOX
+			cmd_desc.ByteStride += sizeof(uint32_t);
+			cmd_desc.NumArgumentDescs = arraysize(dispatchMeshCountArgs);
+			cmd_desc.pArgumentDescs = dispatchMeshCountArgs;
+			dx12_check(device->CreateCommandSignature(&cmd_desc, rootsig.Get(), PPV_ARGS(dispatchMeshIndirectCommandSignature)));
 		}
 
 		allocationhandler->descriptors_res.init(this, D3D12_DESCRIPTOR_HEAP_TYPE_CBV_SRV_UAV, 4096);
@@ -6754,7 +6811,7 @@ std::mutex queue_locker;
 		auto args_internal = to_internal(args);
 		auto count_internal = to_internal(count);
 		CommandList_DX12& commandlist = GetCommandList(cmd);
-		commandlist.GetGraphicsCommandList()->ExecuteIndirect(drawInstancedIndirectCommandSignature.Get(), max_count, args_internal->resource.Get(), args_offset, count_internal->resource.Get(), count_offset);
+		commandlist.GetGraphicsCommandList()->ExecuteIndirect(drawInstancedIndirectCountCommandSignature.Get(), max_count, args_internal->resource.Get(), args_offset, count_internal->resource.Get(), count_offset);
 	}
 	void GraphicsDevice_DX12::DrawIndexedInstancedIndirectCount(const GPUBuffer* args, uint64_t args_offset, const GPUBuffer* count, uint64_t count_offset, uint32_t max_count, CommandList cmd)
 	{
@@ -6762,7 +6819,7 @@ std::mutex queue_locker;
 		auto args_internal = to_internal(args);
 		auto count_internal = to_internal(count);
 		CommandList_DX12& commandlist = GetCommandList(cmd);
-		commandlist.GetGraphicsCommandList()->ExecuteIndirect(drawIndexedInstancedIndirectCommandSignature.Get(), max_count, args_internal->resource.Get(), args_offset, count_internal->resource.Get(), count_offset);
+		commandlist.GetGraphicsCommandList()->ExecuteIndirect(drawIndexedInstancedIndirectCountCommandSignature.Get(), max_count, args_internal->resource.Get(), args_offset, count_internal->resource.Get(), count_offset);
 	}
 	void GraphicsDevice_DX12::Dispatch(uint32_t threadGroupCountX, uint32_t threadGroupCountY, uint32_t threadGroupCountZ, CommandList cmd)
 	{
@@ -6796,7 +6853,7 @@ std::mutex queue_locker;
 		auto args_internal = to_internal(args);
 		auto count_internal = to_internal(count);
 		CommandList_DX12& commandlist = GetCommandList(cmd);
-		commandlist.GetGraphicsCommandList()->ExecuteIndirect(dispatchMeshIndirectCommandSignature.Get(), max_count, args_internal->resource.Get(), args_offset, count_internal->resource.Get(), count_offset);
+		commandlist.GetGraphicsCommandList()->ExecuteIndirect(dispatchMeshIndirectCountCommandSignature.Get(), max_count, args_internal->resource.Get(), args_offset, count_internal->resource.Get(), count_offset);
 	}
 	void GraphicsDevice_DX12::CopyResource(const GPUResource* pDst, const GPUResource* pSrc, CommandList cmd)
 	{
diff --git a/WickedEngine/wiGraphicsDevice_DX12.h b/WickedEngine/wiGraphicsDevice_DX12.h
index 79279aa76..500e99518 100644
--- a/WickedEngine/wiGraphicsDevice_DX12.h
+++ b/WickedEngine/wiGraphicsDevice_DX12.h
@@ -60,6 +60,10 @@ namespace wi::graphics
 		Microsoft::WRL::ComPtr<ID3D12CommandSignature> drawIndexedInstancedIndirectCommandSignature;
 		Microsoft::WRL::ComPtr<ID3D12CommandSignature> dispatchMeshIndirectCommandSignature;
 
+		Microsoft::WRL::ComPtr<ID3D12CommandSignature> drawInstancedIndirectCountCommandSignature;
+		Microsoft::WRL::ComPtr<ID3D12CommandSignature> drawIndexedInstancedIndirectCountCommandSignature;
+		Microsoft::WRL::ComPtr<ID3D12CommandSignature> dispatchMeshIndirectCountCommandSignature;
+
 		wi::vector<GUID> video_decode_profile_list;
 
 		bool deviceRemoved = false;
diff --git a/WickedEngine/wiPlatform.h b/WickedEngine/wiPlatform.h
index 87498ae62..647d92ac6 100644
--- a/WickedEngine/wiPlatform.h
+++ b/WickedEngine/wiPlatform.h
@@ -18,6 +18,7 @@
 #define PLATFORM_WINDOWS_DESKTOP
 #endif // WINAPI_FAMILY_GAMES
 #define wiLoadLibrary(name) LoadLibraryA(name)
+#define wiFreeLibrary(handle) FreeLibrary(handle)
 #define wiGetProcAddress(handle,name) GetProcAddress(handle, name)
 #elif defined(__SCE__)
 #define PLATFORM_PS5
@@ -43,6 +44,7 @@
 #if defined(PLATFORM_LINUX) || defined(PLATFORM_APPLE)
 #include <dlfcn.h>
 #define wiLoadLibrary(name) dlopen(name, RTLD_LAZY)
+#define wiFreeLibrary(handle) dlclose(handle)
 #define wiGetProcAddress(handle,name) dlsym(handle, name)
 typedef void* HMODULE;
 #endif // defined(PLATFORM_LINUX) || defined(PLATFORM_APPLE)
diff --git a/WickedEngine/wiRenderer.cpp b/WickedEngine/wiRenderer.cpp
index 294bfd22f..a269157ac 100644
--- a/WickedEngine/wiRenderer.cpp
+++ b/WickedEngine/wiRenderer.cpp
@@ -11305,7 +11305,7 @@ void ComputeLuminance(
 			cmd
 		);
 
-		device->Barrier(GPUBarrier::Memory(), cmd);
+		device->Barrier(GPUBarrier::Memory(&res.luminance), cmd);
 	}
 
 	// Pass 2 : Reduce into 1x1 texture
@@ -11633,7 +11633,7 @@ void Visibility_Prepare(
 		{
 			device->ClearUAV(res.primitiveID_resolved, 0, cmd);
 		}
-		device->Barrier(GPUBarrier::Memory(), cmd);
+		device->Barrier(cmd);
 
 		device->BindComputeShader(&shaders[msaa ? CSTYPE_VISIBILITY_RESOLVE_MSAA : CSTYPE_VISIBILITY_RESOLVE], cmd);
 
@@ -11692,7 +11692,7 @@ void Visibility_Surface(
 	device->ClearUAV(&res.texture_roughness, 0, cmd);
 	device->ClearUAV(&res.texture_payload_0, 0, cmd);
 	device->ClearUAV(&res.texture_payload_1, 0, cmd);
-	device->Barrier(GPUBarrier::Memory(), cmd);
+	device->Barrier(cmd);
 
 	device->BindResource(&res.binned_tiles, 0, cmd);
 	device->BindUAV(&output, 0, cmd);
@@ -11747,7 +11747,7 @@ void Visibility_Surface_Reduced(
 
 	device->ClearUAV(&res.texture_normals, 0, cmd);
 	device->ClearUAV(&res.texture_roughness, 0, cmd);
-	device->Barrier(GPUBarrier::Memory(), cmd);
+	device->Barrier(cmd);
 
 	device->BindResource(&res.binned_tiles, 0, cmd);
 	device->BindUAV(&res.texture_normals, 1, cmd);
@@ -11886,7 +11886,7 @@ void SurfelGI_Coverage(
 	}
 	device->ClearUAV(&res.result_halfres, 0, cmd);
 	device->ClearUAV(&res.result, 0, cmd);
-	device->Barrier(GPUBarrier::Memory(), cmd);
+	device->Barrier(cmd);
 
 
 	// Coverage:
@@ -11978,18 +11978,11 @@ void SurfelGI(
 	if (!scene.surfelgi.cleared)
 	{
 		scene.surfelgi.cleared = true;
-		GPUBarrier barriers[] = {
-			GPUBarrier::Image(&scene.surfelgi.momentsTexture, scene.surfelgi.momentsTexture.desc.layout, ResourceState::UNORDERED_ACCESS),
-		};
-		device->Barrier(barriers, arraysize(barriers), cmd);
+		device->Barrier(GPUBarrier::Image(&scene.surfelgi.momentsTexture, scene.surfelgi.momentsTexture.desc.layout, ResourceState::UNORDERED_ACCESS), cmd);
 		device->ClearUAV(&scene.surfelgi.momentsTexture, 0, cmd);
 		device->ClearUAV(&scene.surfelgi.varianceBuffer, 0, cmd);
-		for (auto& x : barriers)
-		{
-			std::swap(x.image.layout_before, x.image.layout_after);
-		}
-		device->Barrier(barriers, arraysize(barriers), cmd);
-		device->Barrier(GPUBarrier::Memory(), cmd);
+		device->Barrier(GPUBarrier::Image(&scene.surfelgi.momentsTexture, ResourceState::UNORDERED_ACCESS, scene.surfelgi.momentsTexture.desc.layout), cmd);
+		device->Barrier(GPUBarrier::Memory(&scene.surfelgi.varianceBuffer), cmd);
 	}
 
 	// Grid reset:
@@ -16822,7 +16815,7 @@ void VolumetricClouds_Capture(
 	{
 		device->ClearUAV(&scene.cloudmap, 0, cmd);
 		device->ClearUAV(&scene.cloudmap_variance, 0, cmd);
-		device->Barrier(GPUBarrier::Memory(), cmd);
+		device->Barrier(cmd);
 	}
 
 	device->Dispatch(
@@ -17219,7 +17212,7 @@ void Postprocess_FSR(
 
 		device->ClearUAV(&temp, 0, cmd);
 		device->ClearUAV(&output, 0, cmd);
-		device->Barrier(GPUBarrier::Memory(), cmd);
+		device->Barrier(cmd);
 
 		device->Dispatch((desc.width + 15) / 16, (desc.height + 15) / 16, 1, cmd);
 
diff --git a/WickedEngine/wiVersion.cpp b/WickedEngine/wiVersion.cpp
index 880a6b912..a901ecd47 100644
--- a/WickedEngine/wiVersion.cpp
+++ b/WickedEngine/wiVersion.cpp
@@ -9,7 +9,7 @@ namespace wi::version
 	// minor features, major updates, breaking compatibility changes
 	const int minor = 72;
 	// minor bug fixes, alterations, refactors, updates
-	const int revision = 46;
+	const int revision = 47;
 
 	const std::string version_string = std::to_string(major) + "." + std::to_string(minor) + "." + std::to_string(revision);