D3D12: Implement GPU-based bounding box

Source/Core/VideoBackends/D3D12/BoundingBox.cpp

@@ -2,43 +2,160 @@
 // Licensed under GPLv2+
 // Refer to the license.txt file included.
 
+#include <memory>
+
 #include "Common/CommonTypes.h"
 #include "Common/MsgHandler.h"
 #include "VideoBackends/D3D12/BoundingBox.h"
+#include "VideoBackends/D3D12/D3DBase.h"
+#include "VideoBackends/D3D12/D3DCommandListManager.h"
+#include "VideoBackends/D3D12/D3DDescriptorHeapManager.h"
+#include "VideoBackends/D3D12/D3DStreamBuffer.h"
+#include "VideoBackends/D3D12/D3DUtil.h"
+#include "VideoBackends/D3D12/FramebufferManager.h"
+#include "VideoBackends/D3D12/Render.h"
 #include "VideoCommon/VideoConfig.h"
 
-// D3D12TODO: Support bounding box behavior.
 namespace DX12
 {
 
-ID3D11UnorderedAccessView* BBox::GetUAV()
+constexpr size_t BBOX_BUFFER_SIZE = sizeof(int) * 4;
-{
+constexpr size_t BBOX_STREAM_BUFFER_SIZE = BBOX_BUFFER_SIZE * 128;
-	// D3D12TODO: Implement this;
+
-	return nullptr;
+static ID3D12Resource* s_bbox_buffer;
-}
+static ID3D12Resource* s_bbox_staging_buffer;
+static void* s_bbox_staging_buffer_map;
+static std::unique_ptr<D3DStreamBuffer> s_bbox_stream_buffer;
+static D3D12_GPU_DESCRIPTOR_HANDLE s_bbox_descriptor_handle;
 
 void BBox::Init()
 {
-	if (g_ActiveConfig.backend_info.bSupportsBBox)
+	if (!g_ActiveConfig.backend_info.bSupportsBBox)
-	{
+		return;
-		// D3D12TODO: Implement this;
+
-	}
+	CD3DX12_RESOURCE_DESC buffer_desc(CD3DX12_RESOURCE_DESC::Buffer(BBOX_BUFFER_SIZE, D3D12_RESOURCE_FLAG_ALLOW_UNORDERED_ACCESS, 0));
+	CD3DX12_RESOURCE_DESC staging_buffer_desc(CD3DX12_RESOURCE_DESC::Buffer(BBOX_BUFFER_SIZE, D3D12_RESOURCE_FLAG_NONE, 0));
+
+	CheckHR(D3D::device12->CreateCommittedResource(
+		&CD3DX12_HEAP_PROPERTIES(D3D12_HEAP_TYPE_DEFAULT),
+		D3D12_HEAP_FLAG_NONE,
+		&buffer_desc,
+		D3D12_RESOURCE_STATE_UNORDERED_ACCESS,
+		nullptr,
+		IID_PPV_ARGS(&s_bbox_buffer)));
+
+	CheckHR(D3D::device12->CreateCommittedResource(
+		&CD3DX12_HEAP_PROPERTIES(D3D12_HEAP_TYPE_READBACK),
+		D3D12_HEAP_FLAG_NONE,
+		&staging_buffer_desc,
+		D3D12_RESOURCE_STATE_COPY_DEST,
+		nullptr,
+		IID_PPV_ARGS(&s_bbox_staging_buffer)));
+
+	s_bbox_stream_buffer = std::make_unique<D3DStreamBuffer>(BBOX_STREAM_BUFFER_SIZE, BBOX_STREAM_BUFFER_SIZE, nullptr);
+
+	// D3D12 root signature UAV must be raw or structured buffers, not typed. Since we used a typed buffer,
+	// we have to use a descriptor table. Luckily, we only have to allocate this once, and it never changes.
+	D3D12_CPU_DESCRIPTOR_HANDLE cpu_descriptor_handle;
+	if (!D3D::gpu_descriptor_heap_mgr->Allocate(&cpu_descriptor_handle, &s_bbox_descriptor_handle, nullptr, false))
+		PanicAlert("Failed to create bounding box UAV descriptor");
+
+	D3D12_UNORDERED_ACCESS_VIEW_DESC view_desc = { DXGI_FORMAT_R32_SINT, D3D12_UAV_DIMENSION_BUFFER };
+	view_desc.Buffer.FirstElement = 0;
+	view_desc.Buffer.NumElements = 4;
+	view_desc.Buffer.StructureByteStride = 0;
+	view_desc.Buffer.CounterOffsetInBytes = 0;
+	view_desc.Buffer.Flags = D3D12_BUFFER_UAV_FLAG_NONE;
+	D3D::device12->CreateUnorderedAccessView(s_bbox_buffer, nullptr, &view_desc, cpu_descriptor_handle);
+
+	Bind();
+}
+
+void BBox::Bind()
+{
+	if (s_bbox_buffer)
+		D3D::current_command_list->SetGraphicsRootDescriptorTable(DESCRIPTOR_TABLE_PS_UAV, s_bbox_descriptor_handle);
+}
+
+void BBox::Invalidate()
+{
+	if (!s_bbox_staging_buffer_map)
+		return;
+
+	s_bbox_staging_buffer->Unmap(0, nullptr);
+	s_bbox_staging_buffer_map = nullptr;
 }
 
 void BBox::Shutdown()
 {
-	// D3D12TODO: Implement this;
+	Invalidate();
+
+	if (s_bbox_buffer)
+	{
+		D3D::command_list_mgr->DestroyResourceAfterCurrentCommandListExecuted(s_bbox_buffer);
+		s_bbox_buffer = nullptr;
+	}
+
+	if (s_bbox_staging_buffer)
+	{
+		D3D::command_list_mgr->DestroyResourceAfterCurrentCommandListExecuted(s_bbox_staging_buffer);
+		s_bbox_staging_buffer = nullptr;
+	}
+
+	s_bbox_stream_buffer.reset();
 }
 
 void BBox::Set(int index, int value)
 {
-	// D3D12TODO: Implement this;
+	// If the buffer is currently mapped, compare the value, and update the staging buffer.
+	if (s_bbox_staging_buffer_map)
+	{
+		int current_value;
+		memcpy(&current_value, reinterpret_cast<u8*>(s_bbox_staging_buffer_map) + (index * sizeof(int)), sizeof(int));
+		if (current_value == value)
+		{
+			// Value hasn't changed. So skip updating completely.
+			return;
+		}
+
+		memcpy(reinterpret_cast<u8*>(s_bbox_staging_buffer_map) + (index * sizeof(int)), &value, sizeof(int));
+	}
+
+	if (s_bbox_stream_buffer->AllocateSpaceInBuffer(sizeof(int), sizeof(int)))
+	{
+		// Command list was executed, reset state
+		g_renderer->SetViewport();
+		FramebufferManager::RestoreEFBRenderTargets();
+	}
+
+	// Allocate temporary bytes in upload buffer, then copy to real buffer.
+	memcpy(s_bbox_stream_buffer->GetCPUAddressOfCurrentAllocation(), &value, sizeof(int));
+	D3D::ResourceBarrier(D3D::current_command_list, s_bbox_buffer, D3D12_RESOURCE_STATE_UNORDERED_ACCESS, D3D12_RESOURCE_STATE_COPY_DEST, 0);
+	D3D::current_command_list->CopyBufferRegion(s_bbox_buffer, index * sizeof(int), s_bbox_stream_buffer->GetBuffer(), s_bbox_stream_buffer->GetOffsetOfCurrentAllocation(), sizeof(int));
+	D3D::ResourceBarrier(D3D::current_command_list, s_bbox_buffer, D3D12_RESOURCE_STATE_COPY_DEST, D3D12_RESOURCE_STATE_UNORDERED_ACCESS, 0);
 }
 
 int BBox::Get(int index)
 {
-	// D3D12TODO: Implement this;
+	if (!s_bbox_staging_buffer_map)
-	return 0;
+	{
+		D3D::command_list_mgr->CPUAccessNotify();
+
+		// Copy from real buffer to staging buffer, then block until we have the results.
+		D3D::ResourceBarrier(D3D::current_command_list, s_bbox_buffer, D3D12_RESOURCE_STATE_UNORDERED_ACCESS, D3D12_RESOURCE_STATE_COPY_SOURCE, 0);
+		D3D::current_command_list->CopyBufferRegion(s_bbox_staging_buffer, 0, s_bbox_buffer, 0, BBOX_BUFFER_SIZE);
+		D3D::ResourceBarrier(D3D::current_command_list, s_bbox_buffer, D3D12_RESOURCE_STATE_COPY_SOURCE, D3D12_RESOURCE_STATE_UNORDERED_ACCESS, 0);
+
+		D3D::command_list_mgr->ExecuteQueuedWork(true);
+		g_renderer->SetViewport();
+		FramebufferManager::RestoreEFBRenderTargets();
+
+		CheckHR(s_bbox_staging_buffer->Map(0, nullptr, &s_bbox_staging_buffer_map));
+	}
+
+	int value;
+	memcpy(&value, &reinterpret_cast<int*>(s_bbox_staging_buffer_map)[index], sizeof(int));
+	return value;
 }
 
 };

Source/Core/VideoBackends/D3D12/BoundingBox.h

@@ -11,8 +11,9 @@ namespace DX12
 class BBox
 {
 public:
-	static ID3D11UnorderedAccessView* GetUAV();
 	static void Init();
+	static void Bind();
+	static void Invalidate();
 	static void Shutdown();
 
 	static void Set(int index, int value);

Source/Core/VideoBackends/D3D12/D3DBase.cpp

@@ -637,7 +637,15 @@ void CreateRootSignatures()
 		D3D12_DESCRIPTOR_RANGE_OFFSET_APPEND // UINT OffsetInDescriptorsFromTableStart;
 	};
 
-	D3D12_ROOT_PARAMETER root_parameters[6];
+	D3D12_DESCRIPTOR_RANGE desc_range_uav = {
+		D3D12_DESCRIPTOR_RANGE_TYPE_UAV,     // D3D12_DESCRIPTOR_RANGE_TYPE RangeType;
+		1,                                   // UINT NumDescriptors;
+		2,                                   // UINT BaseShaderRegister;
+		0,                                   // UINT RegisterSpace;
+		D3D12_DESCRIPTOR_RANGE_OFFSET_APPEND // UINT OffsetInDescriptorsFromTableStart;
+	};
+
+	D3D12_ROOT_PARAMETER root_parameters[NUM_GRAPHICS_ROOT_PARAMETERS];
 
 	root_parameters[DESCRIPTOR_TABLE_PS_SRV].ParameterType = D3D12_ROOT_PARAMETER_TYPE_DESCRIPTOR_TABLE;
 	root_parameters[DESCRIPTOR_TABLE_PS_SRV].DescriptorTable.NumDescriptorRanges = 1;
@@ -669,7 +677,10 @@ void CreateRootSignatures()
 	root_parameters[DESCRIPTOR_TABLE_PS_CBVTWO].Descriptor.ShaderRegister = 1;
 	root_parameters[DESCRIPTOR_TABLE_PS_CBVTWO].ShaderVisibility = D3D12_SHADER_VISIBILITY_PIXEL;
 
-	// D3D12TODO: Add bounding box UAV to root signature.
+	root_parameters[DESCRIPTOR_TABLE_PS_UAV].ParameterType = D3D12_ROOT_PARAMETER_TYPE_DESCRIPTOR_TABLE;
+	root_parameters[DESCRIPTOR_TABLE_PS_UAV].DescriptorTable.NumDescriptorRanges = 1;
+	root_parameters[DESCRIPTOR_TABLE_PS_UAV].DescriptorTable.pDescriptorRanges = &desc_range_uav;
+	root_parameters[DESCRIPTOR_TABLE_PS_UAV].ShaderVisibility = D3D12_SHADER_VISIBILITY_PIXEL;
 
 	D3D12_ROOT_SIGNATURE_DESC root_signature_desc = {};
 	root_signature_desc.pParameters = root_parameters;

Source/Core/VideoBackends/D3D12/D3DBase.h

@@ -44,17 +44,21 @@ class D3DCommandListManager;
 class D3DDescriptorHeapManager;
 class D3DTexture2D;
 
+enum GRAPHICS_ROOT_PARAMETER : u32
+{
+	DESCRIPTOR_TABLE_PS_SRV,
+	DESCRIPTOR_TABLE_PS_SAMPLER,
+	DESCRIPTOR_TABLE_GS_CBV,
+	DESCRIPTOR_TABLE_VS_CBV,
+	DESCRIPTOR_TABLE_PS_CBVONE,
+	DESCRIPTOR_TABLE_PS_CBVTWO,
+	DESCRIPTOR_TABLE_PS_UAV,
+	NUM_GRAPHICS_ROOT_PARAMETERS
+};
+
 namespace D3D
 {
 
-#define DESCRIPTOR_TABLE_PS_SRV     0
-#define DESCRIPTOR_TABLE_PS_SAMPLER 1
-#define DESCRIPTOR_TABLE_GS_CBV     2
-#define DESCRIPTOR_TABLE_VS_CBV     3
-// #define DESCRIPTOR_TABLE_PS_UAV     4
-#define DESCRIPTOR_TABLE_PS_CBVONE  4
-#define DESCRIPTOR_TABLE_PS_CBVTWO  5
-
 HRESULT LoadDXGI();
 HRESULT LoadD3D();
 HRESULT LoadD3DCompiler();

Source/Core/VideoBackends/D3D12/FramebufferManager.cpp

@@ -3,6 +3,7 @@
 // Refer to the license.txt file included.
 
 #include "Core/HW/Memmap.h"
+#include "VideoBackends/D3D12/BoundingBox.h"
 #include "VideoBackends/D3D12/D3DBase.h"
 #include "VideoBackends/D3D12/D3DCommandListManager.h"
 #include "VideoBackends/D3D12/D3DUtil.h"
@@ -208,6 +209,8 @@ void FramebufferManager::RestoreEFBRenderTargets()
 	D3D::current_command_list->OMSetRenderTargets(1,
 		&FramebufferManager::GetEFBColorTexture()->GetRTV12(), FALSE,
 		&FramebufferManager::GetEFBDepthTexture()->GetDSV12());
+
+	BBox::Bind();
 }
 
 u32 FramebufferManager::ReadEFBColorAccessCopy(u32 x, u32 y)

Source/Core/VideoBackends/D3D12/Render.cpp

@@ -735,6 +735,7 @@ void Renderer::SwapImpl(u32 xfb_addr, u32 fb_width, u32 fb_stride, u32 fb_height
 
 	// Invalidate EFB access copies. Not strictly necessary, but this avoids having the buffers mapped when calling Present().
 	FramebufferManager::InvalidateEFBAccessCopies();
+	BBox::Invalidate();
 
 	// Prepare to copy the XFBs to our backbuffer
 	UpdateDrawRectangle(s_backbuffer_width, s_backbuffer_height);

Source/Core/VideoBackends/D3D12/VertexManager.cpp

@@ -137,10 +137,7 @@ void VertexManager::vFlush(bool use_dst_alpha)
 	ShaderCache::LoadAndSetActiveShaders(use_dst_alpha ? DSTALPHA_DUAL_SOURCE_BLEND : DSTALPHA_NONE, current_primitive_type);
 
 	if (g_ActiveConfig.backend_info.bSupportsBBox && BoundingBox::active)
-	{
+		BBox::Invalidate();
-		// D3D12TODO: Support GPU-side bounding box.
-		// D3D::context->OMSetRenderTargetsAndUnorderedAccessViews(D3D11_KEEP_RENDER_TARGETS_AND_DEPTH_STENCIL, nullptr, nullptr, 2, 1, &BBox::GetUAV(), nullptr);
-	}
 
 	u32 stride = VertexLoaderManager::GetCurrentVertexFormat()->GetVertexStride();
 

Source/Core/VideoBackends/D3D12/main.cpp

@@ -116,7 +116,7 @@ void InitBackendInfo()
 			g_Config.backend_info.bSupportsEarlyZ = shader_model_5_supported;
 
 			// Requires full UAV functionality (only available in shader model 5)
-			g_Config.backend_info.bSupportsBBox = false;
+			g_Config.backend_info.bSupportsBBox = shader_model_5_supported;
 
 			// Requires the instance attribute (only available in shader model 5)
 			g_Config.backend_info.bSupportsGSInstancing = shader_model_5_supported;