Merge pull request #1503 from kayru/d3d_optimization_cache

D3D: Filter redundant API calls by caching state in StateManager

Source/Core/VideoBackends/D3D/D3DState.cpp

@@ -33,46 +33,170 @@ template<typename T> AutoState<T>::~AutoState()
 	state = nullptr;
 }
 
-StateManager::StateManager() : cur_blendstate(nullptr), cur_depthstate(nullptr), cur_raststate(nullptr) {}
+StateManager::StateManager()
+	: m_currentBlendState(nullptr)
+	, m_currentDepthState(nullptr)
+	, m_currentRasterizerState(nullptr)
+	, m_dirtyFlags(~0u)
+	, m_pending()
+	, m_current()
+{
+}
 
-void StateManager::PushBlendState(const ID3D11BlendState* state) { blendstates.push(AutoBlendState(state)); }
-void StateManager::PushDepthState(const ID3D11DepthStencilState* state) { depthstates.push(AutoDepthStencilState(state)); }
-void StateManager::PushRasterizerState(const ID3D11RasterizerState* state) { raststates.push(AutoRasterizerState(state)); }
-void StateManager::PopBlendState() { blendstates.pop(); }
-void StateManager::PopDepthState() { depthstates.pop(); }
-void StateManager::PopRasterizerState() { raststates.pop(); }
+void StateManager::PushBlendState(const ID3D11BlendState* state) { m_blendStates.push(AutoBlendState(state)); }
+void StateManager::PushDepthState(const ID3D11DepthStencilState* state) { m_depthStates.push(AutoDepthStencilState(state)); }
+void StateManager::PushRasterizerState(const ID3D11RasterizerState* state) { m_rasterizerStates.push(AutoRasterizerState(state)); }
+void StateManager::PopBlendState() { m_blendStates.pop(); }
+void StateManager::PopDepthState() { m_depthStates.pop(); }
+void StateManager::PopRasterizerState() { m_rasterizerStates.pop(); }
 
 void StateManager::Apply()
 {
-	if (!blendstates.empty())
+	if (!m_blendStates.empty())
 	{
-		if (cur_blendstate != blendstates.top().GetPtr())
+		if (m_currentBlendState != m_blendStates.top().GetPtr())
 		{
-			cur_blendstate = (ID3D11BlendState*)blendstates.top().GetPtr();
-			D3D::context->OMSetBlendState(cur_blendstate, nullptr, 0xFFFFFFFF);
+			m_currentBlendState = (ID3D11BlendState*)m_blendStates.top().GetPtr();
+			D3D::context->OMSetBlendState(m_currentBlendState, nullptr, 0xFFFFFFFF);
 		}
 	}
 	else ERROR_LOG(VIDEO, "Tried to apply without blend state!");
 
-	if (!depthstates.empty())
+	if (!m_depthStates.empty())
 	{
-		if (cur_depthstate != depthstates.top().GetPtr())
+		if (m_currentDepthState != m_depthStates.top().GetPtr())
 		{
-			cur_depthstate = (ID3D11DepthStencilState*)depthstates.top().GetPtr();
-			D3D::context->OMSetDepthStencilState(cur_depthstate, 0);
+			m_currentDepthState = (ID3D11DepthStencilState*)m_depthStates.top().GetPtr();
+			D3D::context->OMSetDepthStencilState(m_currentDepthState, 0);
 		}
 	}
 	else ERROR_LOG(VIDEO, "Tried to apply without depth state!");
 
-	if (!raststates.empty())
+	if (!m_rasterizerStates.empty())
 	{
-		if (cur_raststate != raststates.top().GetPtr())
+		if (m_currentRasterizerState != m_rasterizerStates.top().GetPtr())
 		{
-			cur_raststate = (ID3D11RasterizerState*)raststates.top().GetPtr();
-			D3D::context->RSSetState(cur_raststate);
+			m_currentRasterizerState = (ID3D11RasterizerState*)m_rasterizerStates.top().GetPtr();
+			D3D::context->RSSetState(m_currentRasterizerState);
 		}
 	}
 	else ERROR_LOG(VIDEO, "Tried to apply without rasterizer state!");
+
+	if (!m_dirtyFlags)
+	{
+		return;
+	}
+
+	const u32 dirtyTextures = DirtyFlag_Texture0 | DirtyFlag_Texture1 | DirtyFlag_Texture2 | DirtyFlag_Texture3
+		| DirtyFlag_Texture4 | DirtyFlag_Texture5 | DirtyFlag_Texture6 | DirtyFlag_Texture7;
+	const u32 dirtySamplers = DirtyFlag_Sampler0 | DirtyFlag_Sampler1 | DirtyFlag_Sampler2 | DirtyFlag_Sampler3
+		| DirtyFlag_Sampler4 | DirtyFlag_Sampler5 | DirtyFlag_Sampler6 | DirtyFlag_Sampler7;
+	const u32 dirtyConstants = DirtyFlag_PixelConstants | DirtyFlag_VertexConstants | DirtyFlag_GeometryConstants;
+	const u32 dirtyShaders = DirtyFlag_PixelShader | DirtyFlag_VertexShader | DirtyFlag_GeometryShader;
+	const u32 dirtyBuffers = DirtyFlag_VertexBuffer | DirtyFlag_IndexBuffer;
+
+	if (m_dirtyFlags & dirtyConstants)
+	{
+		if (m_current.pixelConstants[0] != m_pending.pixelConstants[0] ||
+			m_current.pixelConstants[1] != m_pending.pixelConstants[1])
+		{
+			D3D::context->PSSetConstantBuffers(0, m_pending.pixelConstants[1] ? 2 : 1, m_pending.pixelConstants);
+			m_current.pixelConstants[0] = m_pending.pixelConstants[0];
+			m_current.pixelConstants[1] = m_pending.pixelConstants[1];
+		}
+
+		if (m_current.vertexConstants != m_pending.vertexConstants)
+		{
+			D3D::context->VSSetConstantBuffers(0, 1, &m_pending.vertexConstants);
+			m_current.vertexConstants = m_pending.vertexConstants;
+		}
+
+		if (m_current.geometryConstants != m_pending.geometryConstants)
+		{
+			D3D::context->GSSetConstantBuffers(0, 1, &m_pending.geometryConstants);
+			m_current.geometryConstants = m_pending.geometryConstants;
+		}
+	}
+
+	if (m_dirtyFlags & (dirtyBuffers|DirtyFlag_InputAssembler))
+	{
+		if (m_current.vertexBuffer != m_pending.vertexBuffer ||
+			m_current.vertexBufferStride != m_pending.vertexBufferStride ||
+			m_current.vertexBufferOffset != m_pending.vertexBufferOffset)
+		{
+			D3D::context->IASetVertexBuffers(0, 1, &m_pending.vertexBuffer, &m_pending.vertexBufferStride, &m_pending.vertexBufferOffset);
+			m_current.vertexBuffer = m_pending.vertexBuffer;
+			m_current.vertexBufferStride = m_pending.vertexBufferStride;
+			m_current.vertexBufferOffset = m_pending.vertexBufferOffset;
+		}
+
+		if (m_current.indexBuffer != m_pending.indexBuffer)
+		{
+			D3D::context->IASetIndexBuffer(m_pending.indexBuffer, DXGI_FORMAT_R16_UINT, 0);
+			m_current.indexBuffer = m_pending.indexBuffer;
+		}
+
+		if (m_current.topology != m_pending.topology)
+		{
+			D3D::context->IASetPrimitiveTopology(m_pending.topology);
+			m_current.topology = m_pending.topology;
+		}
+
+		if (m_current.inputLayout != m_pending.inputLayout)
+		{
+			D3D::context->IASetInputLayout(m_pending.inputLayout);
+			m_current.inputLayout = m_pending.inputLayout;
+		}
+	}
+
+	if (m_dirtyFlags & dirtyTextures)
+	{
+		for (int i = 0; i < 8; ++i)
+		{
+			// TODO: bit scan through dirty flags
+			if (m_current.textures[i] != m_pending.textures[i])
+			{
+				D3D::context->PSSetShaderResources(i, 1, &m_pending.textures[i]);
+				m_current.textures[i] = m_pending.textures[i];
+			}
+		}
+	}
+
+	if (m_dirtyFlags & dirtySamplers)
+	{
+		for (int i = 0; i < 8; ++i)
+		{
+			// TODO: bit scan through dirty flags
+			if (m_current.samplers[i] != m_pending.samplers[i])
+			{
+				D3D::context->PSSetSamplers(i, 1, &m_pending.samplers[i]);
+				m_current.samplers[i] = m_pending.samplers[i];
+			}
+		}
+	}
+
+	if (m_dirtyFlags & dirtyShaders)
+	{
+		if (m_current.pixelShader != m_pending.pixelShader)
+		{
+			D3D::context->PSSetShader(m_pending.pixelShader, nullptr, 0);
+			m_current.pixelShader = m_pending.pixelShader;
+		}
+
+		if (m_current.vertexShader != m_pending.vertexShader)
+		{
+			D3D::context->VSSetShader(m_pending.vertexShader, nullptr, 0);
+			m_current.vertexShader = m_pending.vertexShader;
+		}
+
+		if (m_current.geometryShader != m_pending.geometryShader)
+		{
+			D3D::context->GSSetShader(m_pending.geometryShader, nullptr, 0);
+			m_current.geometryShader = m_pending.geometryShader;
+		}
+	}
+
+	m_dirtyFlags = 0;
 }
 
 }  // namespace D3D

Source/Core/VideoBackends/D3D/D3DState.h

@@ -73,7 +73,6 @@ private:
 	std::unordered_map<u32, ID3D11RasterizerState*> m_raster;
 	std::unordered_map<u32, ID3D11BlendState*> m_blend;
 	std::unordered_map<u64, ID3D11SamplerState*> m_sampler;
-
 };
 
 namespace D3D
@@ -111,16 +110,199 @@ public:
 	void PopDepthState();
 	void PopRasterizerState();
 
-	// call this before any drawing operation if states could have changed meanwhile
+	void SetTexture(u32 index, ID3D11ShaderResourceView* texture)
+	{
+		if (m_current.textures[index] != texture)
+		{
+			m_dirtyFlags |= DirtyFlag_Texture0 << index;
+			m_pending.textures[index] = texture;
+		}
+	}
+
+	void SetSampler(u32 index, ID3D11SamplerState* sampler)
+	{
+		if (m_current.samplers[index] != sampler)
+		{
+			m_dirtyFlags |= DirtyFlag_Sampler0 << index;
+			m_pending.samplers[index] = sampler;
+		}
+	}
+
+	void SetPixelConstants(ID3D11Buffer* buffer0, ID3D11Buffer* buffer1 = nullptr)
+	{
+		if (m_current.pixelConstants[0] != buffer0)
+		{
+			m_dirtyFlags |= DirtyFlag_PixelConstants;
+			m_pending.pixelConstants[0] = buffer0;
+		}
+
+		if (m_current.pixelConstants[1] != buffer1)
+		{
+			m_dirtyFlags |= DirtyFlag_PixelConstants;
+			m_pending.pixelConstants[1] = buffer1;
+		}
+	}
+
+	void SetVertexConstants(ID3D11Buffer* buffer)
+	{
+		if (m_current.vertexConstants != buffer)
+		{
+			m_dirtyFlags |= DirtyFlag_VertexConstants;
+			m_pending.vertexConstants = buffer;
+		}
+	}
+
+	void SetGeometryConstants(ID3D11Buffer* buffer)
+	{
+		if (m_current.geometryConstants != buffer)
+		{
+			m_dirtyFlags |= DirtyFlag_GeometryConstants;
+			m_pending.geometryConstants = buffer;
+		}
+	}
+
+	void SetVertexBuffer(ID3D11Buffer* buffer, u32 stride, u32 offset)
+	{
+		if (m_current.vertexBuffer != buffer ||
+			m_current.vertexBufferStride != stride ||
+			m_current.vertexBufferOffset != offset)
+		{
+			m_dirtyFlags |= DirtyFlag_VertexBuffer;
+			m_pending.vertexBuffer = buffer;
+			m_pending.vertexBufferStride = stride;
+			m_pending.vertexBufferOffset = offset;
+		}
+	}
+
+	void SetIndexBuffer(ID3D11Buffer* buffer)
+	{
+		if (m_current.indexBuffer != buffer)
+		{
+			m_dirtyFlags |= DirtyFlag_IndexBuffer;
+			m_pending.indexBuffer = buffer;
+		}
+	}
+
+	void SetPrimitiveTopology(D3D11_PRIMITIVE_TOPOLOGY topology)
+	{
+		if (m_current.topology != topology)
+		{
+			m_dirtyFlags |= DirtyFlag_InputAssembler;
+			m_pending.topology = topology;
+		}
+	}
+
+	void SetInputLayout(ID3D11InputLayout* layout)
+	{
+		if (m_current.inputLayout != layout)
+		{
+			m_dirtyFlags |= DirtyFlag_InputAssembler;
+			m_pending.inputLayout = layout;
+		}
+	}
+
+	void SetPixelShader(ID3D11PixelShader* shader)
+	{
+		if (m_current.pixelShader != shader)
+		{
+			m_dirtyFlags |= DirtyFlag_PixelShader;
+			m_pending.pixelShader = shader;
+		}
+	}
+
+	void SetPixelShaderDynamic(ID3D11PixelShader* shader, ID3D11ClassInstance * const * classInstances, u32 classInstancesCount)
+	{
+		D3D::context->PSSetShader(shader, classInstances, classInstancesCount);
+		m_pending.pixelShader = shader;
+	}
+
+	void SetVertexShader(ID3D11VertexShader* shader)
+	{
+		if (m_current.vertexShader != shader)
+		{
+			m_dirtyFlags |= DirtyFlag_VertexShader;
+			m_pending.vertexShader = shader;
+		}
+	}
+
+	void SetGeometryShader(ID3D11GeometryShader* shader)
+	{
+		if (m_current.geometryShader != shader)
+		{
+			m_dirtyFlags |= DirtyFlag_GeometryShader;
+			m_pending.geometryShader = shader;
+		}
+	}
+
+	// call this immediately before any drawing operation or to explicitly apply pending resource state changes
 	void Apply();
 
 private:
-	std::stack<AutoBlendState> blendstates;
-	std::stack<AutoDepthStencilState> depthstates;
-	std::stack<AutoRasterizerState> raststates;
-	ID3D11BlendState* cur_blendstate;
-	ID3D11DepthStencilState* cur_depthstate;
-	ID3D11RasterizerState* cur_raststate;
+
+	std::stack<AutoBlendState> m_blendStates;
+	std::stack<AutoDepthStencilState> m_depthStates;
+	std::stack<AutoRasterizerState> m_rasterizerStates;
+
+	ID3D11BlendState* m_currentBlendState;
+	ID3D11DepthStencilState* m_currentDepthState;
+	ID3D11RasterizerState* m_currentRasterizerState;
+
+	enum DirtyFlags
+	{
+		DirtyFlag_Texture0 = 1 << 0,
+		DirtyFlag_Texture1 = 1 << 1,
+		DirtyFlag_Texture2 = 1 << 2,
+		DirtyFlag_Texture3 = 1 << 3,
+		DirtyFlag_Texture4 = 1 << 4,
+		DirtyFlag_Texture5 = 1 << 5,
+		DirtyFlag_Texture6 = 1 << 6,
+		DirtyFlag_Texture7 = 1 << 7,
+
+		DirtyFlag_Sampler0 = 1 << 8,
+		DirtyFlag_Sampler1 = 1 << 9,
+		DirtyFlag_Sampler2 = 1 << 10,
+		DirtyFlag_Sampler3 = 1 << 11,
+		DirtyFlag_Sampler4 = 1 << 12,
+		DirtyFlag_Sampler5 = 1 << 13,
+		DirtyFlag_Sampler6 = 1 << 14,
+		DirtyFlag_Sampler7 = 1 << 15,
+
+		DirtyFlag_PixelConstants = 1 << 16,
+		DirtyFlag_VertexConstants = 1 << 17,
+		DirtyFlag_GeometryConstants = 1 << 18,
+
+		DirtyFlag_VertexBuffer = 1 << 19,
+		DirtyFlag_IndexBuffer = 1 << 20,
+
+		DirtyFlag_PixelShader = 1 << 21,
+		DirtyFlag_VertexShader = 1 << 22,
+		DirtyFlag_GeometryShader = 1 << 23,
+
+		DirtyFlag_InputAssembler = 1 << 24,
+	};
+
+	u32 m_dirtyFlags;
+
+	struct Resources
+	{
+		ID3D11ShaderResourceView* textures[8];
+		ID3D11SamplerState* samplers[8];
+		ID3D11Buffer* pixelConstants[2];
+		ID3D11Buffer* vertexConstants;
+		ID3D11Buffer* geometryConstants;
+		ID3D11Buffer* vertexBuffer;
+		ID3D11Buffer* indexBuffer;
+		u32 vertexBufferStride;
+		u32 vertexBufferOffset;
+		D3D11_PRIMITIVE_TOPOLOGY topology;
+		ID3D11InputLayout* inputLayout;
+		ID3D11PixelShader* pixelShader;
+		ID3D11VertexShader* vertexShader;
+		ID3D11GeometryShader* geometryShader;
+	};
+
+	Resources m_pending;
+	Resources m_current;
 };
 
 extern StateManager* stateman;

Source/Core/VideoBackends/D3D/D3DUtil.cpp

@@ -346,14 +346,13 @@ int CD3DFont::DrawTextScaled(float x, float y, float size, float spacing, u32 dw
 	// set general pipeline state
 	D3D::stateman->PushBlendState(m_blendstate);
 	D3D::stateman->PushRasterizerState(m_raststate);
-	D3D::stateman->Apply();
 
-	D3D::context->PSSetShader(m_pshader, nullptr, 0);
-	D3D::context->VSSetShader(m_vshader, nullptr, 0);
+	D3D::stateman->SetPixelShader(m_pshader);
+	D3D::stateman->SetVertexShader(m_vshader);
 
-	D3D::context->IASetInputLayout(m_InputLayout);
-	D3D::context->IASetPrimitiveTopology(D3D11_PRIMITIVE_TOPOLOGY_TRIANGLELIST);
-	D3D::context->PSSetShaderResources(0, 1, &m_pTexture);
+	D3D::stateman->SetInputLayout(m_InputLayout);
+	D3D::stateman->SetPrimitiveTopology(D3D11_PRIMITIVE_TOPOLOGY_TRIANGLELIST);
+	D3D::stateman->SetTexture(0, m_pTexture);
 
 	float fStartX = sx;
 	for (char c : text)
@@ -392,7 +391,9 @@ int CD3DFont::DrawTextScaled(float x, float y, float size, float spacing, u32 dw
 		{
 			context->Unmap(m_pVB, 0);
 
-			D3D::context->IASetVertexBuffers(0, 1, &m_pVB, &stride, &bufoffset);
+			D3D::stateman->SetVertexBuffer(m_pVB, stride, bufoffset);
+
+			D3D::stateman->Apply();
 			D3D::context->Draw(3 * dwNumTriangles, 0);
 
 			dwNumTriangles = 0;
@@ -408,7 +409,9 @@ int CD3DFont::DrawTextScaled(float x, float y, float size, float spacing, u32 dw
 	context->Unmap(m_pVB, 0);
 	if (dwNumTriangles > 0)
 	{
-		D3D::context->IASetVertexBuffers(0, 1, &m_pVB, &stride, &bufoffset);
+		D3D::stateman->SetVertexBuffer(m_pVB, stride, bufoffset);
+
+		D3D::stateman->Apply();
 		D3D::context->Draw(3 * dwNumTriangles, 0);
 	}
 	D3D::stateman->PopBlendState();
@@ -494,12 +497,12 @@ void ShutdownUtils()
 
 void SetPointCopySampler()
 {
-	D3D::context->PSSetSamplers(0, 1, &point_copy_sampler);
+	D3D::stateman->SetSampler(0, point_copy_sampler);
 }
 
 void SetLinearCopySampler()
 {
-	D3D::context->PSSetSamplers(0, 1, &linear_copy_sampler);
+	D3D::stateman->SetSampler(0, linear_copy_sampler);
 }
 
 void drawShadedTexQuad(ID3D11ShaderResourceView* texture,
@@ -543,17 +546,18 @@ void drawShadedTexQuad(ID3D11ShaderResourceView* texture,
 	UINT stride = sizeof(STQVertex);
 	UINT offset = 0;
 
-	D3D::context->IASetPrimitiveTopology(D3D11_PRIMITIVE_TOPOLOGY_TRIANGLESTRIP);
-	D3D::context->IASetInputLayout(layout);
-	D3D::context->IASetVertexBuffers(0, 1, &util_vbuf->GetBuffer(), &stride, &offset);
-	D3D::context->PSSetShader(PShader, nullptr, 0);
-	D3D::context->PSSetShaderResources(0, 1, &texture);
-	D3D::context->VSSetShader(Vshader, nullptr, 0);
+	D3D::stateman->SetPrimitiveTopology(D3D11_PRIMITIVE_TOPOLOGY_TRIANGLESTRIP);
+	D3D::stateman->SetInputLayout(layout);
+	D3D::stateman->SetVertexBuffer(util_vbuf->GetBuffer(), stride, offset);
+	D3D::stateman->SetPixelShader(PShader);
+	D3D::stateman->SetTexture(0, texture);
+	D3D::stateman->SetVertexShader(Vshader);
+
 	D3D::stateman->Apply();
 	D3D::context->Draw(4, stq_offset);
 
-	ID3D11ShaderResourceView* texres = nullptr;
-	context->PSSetShaderResources(0, 1, &texres); // immediately unbind the texture
+	D3D::stateman->SetTexture(0, nullptr); // immediately unbind the texture
+	D3D::stateman->Apply();
 }
 
 void drawShadedTexSubQuad(ID3D11ShaderResourceView* texture,
@@ -600,17 +604,18 @@ void drawShadedTexSubQuad(ID3D11ShaderResourceView* texture,
 	UINT stride = sizeof(STSQVertex);
 	UINT offset = 0;
 
-	context->IASetPrimitiveTopology(D3D11_PRIMITIVE_TOPOLOGY_TRIANGLESTRIP);
-	context->IASetVertexBuffers(0, 1, &util_vbuf->GetBuffer(), &stride, &offset);
-	context->IASetInputLayout(layout);
-	context->PSSetShaderResources(0, 1, &texture);
-	context->PSSetShader(PShader, nullptr, 0);
-	context->VSSetShader(Vshader, nullptr, 0);
+	stateman->SetPrimitiveTopology(D3D11_PRIMITIVE_TOPOLOGY_TRIANGLESTRIP);
+	stateman->SetVertexBuffer(util_vbuf->GetBuffer(), stride, offset);
+	stateman->SetInputLayout(layout);
+	stateman->SetTexture(0, texture);
+	stateman->SetPixelShader(PShader);
+	stateman->SetVertexShader(Vshader);
+
 	stateman->Apply();
 	context->Draw(4, stsq_offset);
 
-	ID3D11ShaderResourceView* texres = nullptr;
-	context->PSSetShaderResources(0, 1, &texres); // immediately unbind the texture
+	stateman->SetTexture(0, nullptr); // immediately unbind the texture
+	stateman->Apply();
 }
 
 // Fills a certain area of the current render target with the specified color
@@ -639,14 +644,14 @@ void drawColorQuad(u32 Color, float x1, float y1, float x2, float y2)
 		draw_quad_data.col = Color;
 	}
 
-	context->VSSetShader(VertexShaderCache::GetClearVertexShader(), nullptr, 0);
-	context->PSSetShader(PixelShaderCache::GetClearProgram(), nullptr, 0);
-	context->IASetInputLayout(VertexShaderCache::GetClearInputLayout());
+	stateman->SetVertexShader(VertexShaderCache::GetClearVertexShader());
+	stateman->SetPixelShader(PixelShaderCache::GetClearProgram());
+	stateman->SetInputLayout(VertexShaderCache::GetClearInputLayout());
 
 	UINT stride = sizeof(ColVertex);
 	UINT offset = 0;
-	context->IASetPrimitiveTopology(D3D11_PRIMITIVE_TOPOLOGY_TRIANGLESTRIP);
-	context->IASetVertexBuffers(0, 1, &util_vbuf->GetBuffer(), &stride, &offset);
+	stateman->SetPrimitiveTopology(D3D11_PRIMITIVE_TOPOLOGY_TRIANGLESTRIP);
+	stateman->SetVertexBuffer(util_vbuf->GetBuffer(), stride, offset);
 
 	stateman->Apply();
 	context->Draw(4, cq_offset);
@@ -669,14 +674,16 @@ void drawClearQuad(u32 Color, float z, ID3D11PixelShader* PShader, ID3D11VertexS
 		clear_quad_data.col = Color;
 		clear_quad_data.z = z;
 	}
-	context->VSSetShader(Vshader, nullptr, 0);
-	context->PSSetShader(PShader, nullptr, 0);
-	context->IASetInputLayout(layout);
+
+	stateman->SetVertexShader(Vshader);
+	stateman->SetPixelShader(PShader);
+	stateman->SetInputLayout(layout);
 
 	UINT stride = sizeof(ClearVertex);
 	UINT offset = 0;
-	context->IASetPrimitiveTopology(D3D11_PRIMITIVE_TOPOLOGY_TRIANGLESTRIP);
-	context->IASetVertexBuffers(0, 1, &util_vbuf->GetBuffer(), &stride, &offset);
+	stateman->SetPrimitiveTopology(D3D11_PRIMITIVE_TOPOLOGY_TRIANGLESTRIP);
+	stateman->SetVertexBuffer(util_vbuf->GetBuffer(), stride, offset);
+
 	stateman->Apply();
 	context->Draw(4, clearq_offset);
 }

Source/Core/VideoBackends/D3D/LineGeometryShader.cpp

@@ -6,6 +6,7 @@
 
 #include "VideoBackends/D3D/D3DBase.h"
 #include "VideoBackends/D3D/D3DShader.h"
+#include "VideoBackends/D3D/D3DState.h"
 #include "VideoBackends/D3D/LineGeometryShader.h"
 #include "VideoCommon/VertexShaderGen.h"
 
@@ -223,8 +224,8 @@ bool LineGeometryShader::SetShader(u32 components, float lineWidth,
 			DEBUG_LOG(VIDEO, "Line params: width %f, texOffset %f, vpWidth %f, vpHeight %f",
 				lineWidth, texOffset, vpWidth, vpHeight);
 
-			D3D::context->GSSetShader(shaderIt->second, nullptr, 0);
-			D3D::context->GSSetConstantBuffers(0, 1, &m_paramsBuffer);
+			D3D::stateman->SetGeometryShader(shaderIt->second);
+			D3D::stateman->SetGeometryConstants(m_paramsBuffer);
 
 			return true;
 		}

Source/Core/VideoBackends/D3D/NativeVertexFormat.cpp

@@ -4,6 +4,7 @@
 
 #include "VideoBackends/D3D/D3DBase.h"
 #include "VideoBackends/D3D/D3DBlob.h"
+#include "VideoBackends/D3D/D3DState.h"
 #include "VideoBackends/D3D/VertexManager.h"
 #include "VideoBackends/D3D/VertexShaderCache.h"
 #include "VideoCommon/NativeVertexFormat.h"
@@ -137,7 +138,7 @@ void D3DVertexFormat::SetupVertexPointers()
 		if (FAILED(hr)) PanicAlert("Failed to create input layout, %s %d\n", __FILE__, __LINE__);
 		DX11::D3D::SetDebugObjectName((ID3D11DeviceChild*)m_layout, "input layout used to emulate the GX pipeline");
 	}
-	DX11::D3D::context->IASetInputLayout(m_layout);
+	DX11::D3D::stateman->SetInputLayout(m_layout);
 }
 
 } // namespace DX11

Source/Core/VideoBackends/D3D/PSTextureEncoder.cpp

@@ -1083,21 +1083,20 @@ size_t PSTextureEncoder::Encode(u8* dst, unsigned int dstFormat,
 	if (SetStaticShader(dstFormat, srcFormat, isIntensity, scaleByHalf))
 #endif
 	{
-		D3D::context->VSSetShader(m_vShader, nullptr, 0);
+		D3D::stateman->SetVertexShader(m_vShader);
 
 		D3D::stateman->PushBlendState(m_efbEncodeBlendState);
 		D3D::stateman->PushDepthState(m_efbEncodeDepthState);
 		D3D::stateman->PushRasterizerState(m_efbEncodeRastState);
-		D3D::stateman->Apply();
 
 		D3D11_VIEWPORT vp = CD3D11_VIEWPORT(0.f, 0.f, FLOAT(cacheLinesPerRow*2), FLOAT(numBlocksY));
 		D3D::context->RSSetViewports(1, &vp);
 
-		D3D::context->IASetInputLayout(m_quadLayout);
-		D3D::context->IASetPrimitiveTopology(D3D11_PRIMITIVE_TOPOLOGY_TRIANGLESTRIP);
+		D3D::stateman->SetInputLayout(m_quadLayout);
+		D3D::stateman->SetPrimitiveTopology(D3D11_PRIMITIVE_TOPOLOGY_TRIANGLESTRIP);
 		UINT stride = sizeof(QuadVertex);
 		UINT offset = 0;
-		D3D::context->IASetVertexBuffers(0, 1, &m_quad, &stride, &offset);
+		D3D::stateman->SetVertexBuffer(m_quad, stride, offset);
 
 		EFBRectangle fullSrcRect;
 		fullSrcRect.left = 0;
@@ -1117,8 +1116,6 @@ size_t PSTextureEncoder::Encode(u8* dst, unsigned int dstFormat,
 		params.TexBottom = float(targetRect.bottom) / g_renderer->GetTargetHeight();
 		D3D::context->UpdateSubresource(m_encodeParams, 0, nullptr, &params, 0, 0);
 
-		D3D::context->VSSetConstantBuffers(0, 1, &m_encodeParams);
-
 		D3D::context->OMSetRenderTargets(1, &m_outRTV, nullptr);
 
 		ID3D11ShaderResourceView* pEFB = (srcFormat == PEControl::Z24) ?
@@ -1128,12 +1125,14 @@ size_t PSTextureEncoder::Encode(u8* dst, unsigned int dstFormat,
 			// expecting the blurred edges around multisampled shapes.
 			FramebufferManager::GetResolvedEFBColorTexture()->GetSRV();
 
-		D3D::context->PSSetConstantBuffers(0, 1, &m_encodeParams);
-		D3D::context->PSSetShaderResources(0, 1, &pEFB);
-		D3D::context->PSSetSamplers(0, 1, &m_efbSampler);
+		D3D::stateman->SetVertexConstants(m_encodeParams);
+		D3D::stateman->SetPixelConstants(m_encodeParams);
+		D3D::stateman->SetTexture(0, pEFB);
+		D3D::stateman->SetSampler(0, m_efbSampler);
 
 		// Encode!
 
+		D3D::stateman->Apply();
 		D3D::context->Draw(4, 0);
 
 		// Copy to staging buffer
@@ -1143,22 +1142,21 @@ size_t PSTextureEncoder::Encode(u8* dst, unsigned int dstFormat,
 
 		// Clean up state
 
-		IUnknown* nullDummy = nullptr;
-
-		D3D::context->PSSetSamplers(0, 1, (ID3D11SamplerState**)&nullDummy);
-		D3D::context->PSSetShaderResources(0, 1, (ID3D11ShaderResourceView**)&nullDummy);
-		D3D::context->PSSetConstantBuffers(0, 1, (ID3D11Buffer**)&nullDummy);
-
 		D3D::context->OMSetRenderTargets(0, nullptr, nullptr);
 
-		D3D::context->VSSetConstantBuffers(0, 1, (ID3D11Buffer**)&nullDummy);
+		D3D::stateman->SetSampler(0, nullptr);
+		D3D::stateman->SetTexture(0, nullptr);
+		D3D::stateman->SetPixelConstants(nullptr);
+		D3D::stateman->SetVertexConstants(nullptr);
+
+		D3D::stateman->SetPixelShader(nullptr);
+		D3D::stateman->SetVertexBuffer(nullptr, 0, 0);
 
 		D3D::stateman->PopRasterizerState();
 		D3D::stateman->PopDepthState();
 		D3D::stateman->PopBlendState();
 
-		D3D::context->PSSetShader(nullptr, nullptr, 0);
-		D3D::context->VSSetShader(nullptr, nullptr, 0);
+		D3D::stateman->Apply();
 
 		// Transfer staging buffer to GameCube/Wii RAM
 
@@ -1276,7 +1274,7 @@ bool PSTextureEncoder::SetStaticShader(unsigned int dstFormat, PEControl::PixelF
 	{
 		if (it->second)
 		{
-			D3D::context->PSSetShader(it->second, nullptr, 0);
+			D3D::stateman->SetPixelShader(it->second);
 			return true;
 		}
 		else
@@ -1436,7 +1434,7 @@ bool PSTextureEncoder::SetDynamicShader(unsigned int dstFormat,
 	if (m_generatorSlot != UINT(-1))
 		m_linkageArray[m_generatorSlot] = m_generatorClass[generatorNum];
 
-	D3D::context->PSSetShader(m_dynamicShader,
+	D3D::stateman->SetPixelShaderDynamic(m_dynamicShader,
 		m_linkageArray.empty() ? nullptr : &m_linkageArray[0],
 		(UINT)m_linkageArray.size());
 

Source/Core/VideoBackends/D3D/PointGeometryShader.cpp

@@ -6,6 +6,7 @@
 
 #include "VideoBackends/D3D/D3DBase.h"
 #include "VideoBackends/D3D/D3DShader.h"
+#include "VideoBackends/D3D/D3DState.h"
 #include "VideoBackends/D3D/PointGeometryShader.h"
 #include "VideoCommon/VertexShaderGen.h"
 
@@ -217,8 +218,8 @@ bool PointGeometryShader::SetShader(u32 components, float pointSize,
 			DEBUG_LOG(VIDEO, "Point params: size %f, texOffset %f, vpWidth %f, vpHeight %f",
 				pointSize, texOffset, vpWidth, vpHeight);
 
-			D3D::context->GSSetShader(shaderIt->second, nullptr, 0);
-			D3D::context->GSSetConstantBuffers(0, 1, &m_paramsBuffer);
+			D3D::stateman->SetGeometryShader(shaderIt->second);
+			D3D::stateman->SetGeometryConstants(m_paramsBuffer);
 
 			return true;
 		}

Source/Core/VideoBackends/D3D/Render.cpp

@@ -357,7 +357,7 @@ u32 Renderer::AccessEFB(EFBAccessType type, u32 x, u32 y, u32 poke_data)
 		// depth buffers can only be completely CopySubresourceRegion'ed, so we're using drawShadedTexQuad instead
 		D3D11_VIEWPORT vp = CD3D11_VIEWPORT(0.f, 0.f, 1.f, 1.f);
 		D3D::context->RSSetViewports(1, &vp);
-		D3D::context->PSSetConstantBuffers(0, 1, &access_efb_cbuf);
+		D3D::stateman->SetPixelConstants(0, access_efb_cbuf);
 		D3D::context->OMSetRenderTargets(1, &FramebufferManager::GetEFBDepthReadTexture()->GetRTV(), nullptr);
 		D3D::SetPointCopySampler();
 		D3D::drawShadedTexQuad(FramebufferManager::GetEFBDepthTexture()->GetSRV(),
@@ -1007,15 +1007,12 @@ void Renderer::ApplyState(bool bUseDstAlpha)
 	gx_state.raster.wireframe = g_ActiveConfig.bWireFrame;
 	D3D::stateman->PushRasterizerState(gx_state_cache.Get(gx_state.raster));
 
-	ID3D11SamplerState* samplerstate[8];
 	for (unsigned int stage = 0; stage < 8; stage++)
 	{
+		// TODO: cache SamplerState directly, not d3d object
 		gx_state.sampler[stage].max_anisotropy = g_ActiveConfig.iMaxAnisotropy;
-		samplerstate[stage] = gx_state_cache.Get(gx_state.sampler[stage]);
+		D3D::stateman->SetSampler(stage, gx_state_cache.Get(gx_state.sampler[stage]));
 	}
-	D3D::context->PSSetSamplers(0, 8, samplerstate);
-
-	D3D::stateman->Apply();
 
 	if (bUseDstAlpha)
 	{
@@ -1024,12 +1021,14 @@ void Renderer::ApplyState(bool bUseDstAlpha)
 		SetLogicOpMode();
 	}
 
-	ID3D11Buffer* const_buffers[2] = {PixelShaderCache::GetConstantBuffer(), VertexShaderCache::GetConstantBuffer()};
-	D3D::context->PSSetConstantBuffers(0, 1 + g_ActiveConfig.bEnablePixelLighting, const_buffers);
-	D3D::context->VSSetConstantBuffers(0, 1, const_buffers+1);
+	ID3D11Buffer* vertexConstants = VertexShaderCache::GetConstantBuffer();
+	ID3D11Buffer* pixelConstants = PixelShaderCache::GetConstantBuffer();
 
-	D3D::context->PSSetShader(PixelShaderCache::GetActiveShader(), nullptr, 0);
-	D3D::context->VSSetShader(VertexShaderCache::GetActiveShader(), nullptr, 0);
+	D3D::stateman->SetPixelConstants(pixelConstants, g_ActiveConfig.bEnablePixelLighting ? vertexConstants : nullptr);
+	D3D::stateman->SetVertexConstants(vertexConstants);
+
+	D3D::stateman->SetPixelShader(PixelShaderCache::GetActiveShader());
+	D3D::stateman->SetVertexShader(VertexShaderCache::GetActiveShader());
 }
 
 void Renderer::RestoreState()
@@ -1046,8 +1045,6 @@ void Renderer::ApplyCullDisable()
 
 	ID3D11RasterizerState* raststate = gx_state_cache.Get(rast);
 	D3D::stateman->PushRasterizerState(raststate);
-
-	D3D::stateman->Apply();
 }
 
 void Renderer::RestoreCull()

Source/Core/VideoBackends/D3D/Television.cpp

@@ -7,6 +7,7 @@
 #include "Core/HW/Memmap.h"
 #include "VideoBackends/D3D/D3DBase.h"
 #include "VideoBackends/D3D/D3DShader.h"
+#include "VideoBackends/D3D/D3DState.h"
 #include "VideoBackends/D3D/D3DUtil.h"
 #include "VideoBackends/D3D/Television.h"
 #include "VideoBackends/D3D/VertexShaderCache.h"
@@ -150,7 +151,7 @@ void Television::Render()
 		MathUtil::Rectangle<int> sourceRc(0, 0, int(m_curWidth), int(m_curHeight));
 		MathUtil::Rectangle<float> destRc(-1.f, 1.f, 1.f, -1.f);
 
-		D3D::context->PSSetSamplers(0, 1, &m_samplerState);
+		D3D::stateman->SetSampler(0, m_samplerState);
 
 		D3D::drawShadedTexSubQuad(
 			m_yuyvTextureSRV, &sourceRc,

Source/Core/VideoBackends/D3D/TextureCache.cpp

@@ -4,6 +4,7 @@
 
 #include "Core/HW/Memmap.h"
 #include "VideoBackends/D3D/D3DBase.h"
+#include "VideoBackends/D3D/D3DState.h"
 #include "VideoBackends/D3D/D3DUtil.h"
 #include "VideoBackends/D3D/FramebufferManager.h"
 #include "VideoBackends/D3D/PixelShaderCache.h"
@@ -29,7 +30,7 @@ TextureCache::TCacheEntry::~TCacheEntry()
 
 void TextureCache::TCacheEntry::Bind(unsigned int stage)
 {
-	D3D::context->PSSetShaderResources(stage, 1, &texture->GetSRV());
+	D3D::stateman->SetTexture(stage, texture->GetSRV());
 }
 
 bool TextureCache::TCacheEntry::Save(const std::string& filename, unsigned int level)
@@ -143,7 +144,7 @@ void TextureCache::TCacheEntry::FromRenderTarget(u32 dstAddr, unsigned int dstFo
 			CHECK(SUCCEEDED(hr), "Create efb copy constant buffer %d", cbufid);
 			D3D::SetDebugObjectName((ID3D11DeviceChild*)efbcopycbuf[cbufid], "a constant buffer used in TextureCache::CopyRenderTargetToTexture");
 		}
-		D3D::context->PSSetConstantBuffers(0, 1, &efbcopycbuf[cbufid]);
+		D3D::stateman->SetPixelConstants(efbcopycbuf[cbufid]);
 
 		const TargetRectangle targetSource = g_renderer->ConvertEFBRectangle(srcRect);
 		// TODO: try targetSource.asRECT();

Source/Core/VideoBackends/D3D/VertexManager.cpp

@@ -4,6 +4,7 @@
 
 #include "VideoBackends/D3D/BoundingBox.h"
 #include "VideoBackends/D3D/D3DBase.h"
+#include "VideoBackends/D3D/D3DState.h"
 #include "VideoBackends/D3D/PixelShaderCache.h"
 #include "VideoBackends/D3D/Render.h"
 #include "VideoBackends/D3D/VertexManager.h"
@@ -129,17 +130,19 @@ void VertexManager::Draw(u32 stride)
 	u32 components = VertexLoaderManager::GetCurrentVertexFormat()->m_components;
 	u32 indices = IndexGenerator::GetIndexLen();
 
-	u32 zero = 0;
-	D3D::context->IASetVertexBuffers(0, 1, &m_buffers[m_currentBuffer], &stride, &zero);
-	D3D::context->IASetIndexBuffer(m_buffers[m_currentBuffer], DXGI_FORMAT_R16_UINT, 0);
+	D3D::stateman->SetVertexBuffer(m_buffers[m_currentBuffer], stride, 0);
+	D3D::stateman->SetIndexBuffer(m_buffers[m_currentBuffer]);
 
 	u32 baseVertex = m_vertexDrawOffset / stride;
 	u32 startIndex = m_indexDrawOffset / sizeof(u16);
 
 	if (current_primitive_type == PRIMITIVE_TRIANGLES)
 	{
-		D3D::context->IASetPrimitiveTopology(D3D11_PRIMITIVE_TOPOLOGY_TRIANGLESTRIP);
+		D3D::stateman->SetPrimitiveTopology(D3D11_PRIMITIVE_TOPOLOGY_TRIANGLESTRIP);
+
+		D3D::stateman->Apply();
 		D3D::context->DrawIndexed(indices, startIndex, baseVertex);
+
 		INCSTAT(stats.thisFrame.numDrawCalls);
 	}
 	else if (current_primitive_type == PRIMITIVE_LINES)
@@ -157,13 +160,14 @@ void VertexManager::Draw(u32 stride)
 		if (m_lineShader.SetShader(components, lineWidth,
 			texOffset, vpWidth, vpHeight, texOffsetEnable))
 		{
-			((DX11::Renderer*)g_renderer)->ApplyCullDisable(); // Disable culling for lines and points
-			D3D::context->IASetPrimitiveTopology(D3D11_PRIMITIVE_TOPOLOGY_LINELIST);
+			D3D::stateman->SetPrimitiveTopology(D3D11_PRIMITIVE_TOPOLOGY_LINELIST);
+
+			D3D::stateman->Apply();
 			D3D::context->DrawIndexed(indices, startIndex, baseVertex);
+
 			INCSTAT(stats.thisFrame.numDrawCalls);
 
-			D3D::context->GSSetShader(nullptr, nullptr, 0);
-			((DX11::Renderer*)g_renderer)->RestoreCull();
+			D3D::stateman->SetGeometryShader(nullptr);
 		}
 	}
 	else //if (current_primitive_type == PRIMITIVE_POINTS)
@@ -181,13 +185,14 @@ void VertexManager::Draw(u32 stride)
 		if (m_pointShader.SetShader(components, pointSize,
 			texOffset, vpWidth, vpHeight, texOffsetEnable))
 		{
-			((DX11::Renderer*)g_renderer)->ApplyCullDisable(); // Disable culling for lines and points
-			D3D::context->IASetPrimitiveTopology(D3D11_PRIMITIVE_TOPOLOGY_POINTLIST);
+			D3D::stateman->SetPrimitiveTopology(D3D11_PRIMITIVE_TOPOLOGY_POINTLIST);
+
+			D3D::stateman->Apply();
 			D3D::context->DrawIndexed(indices, startIndex, baseVertex);
+
 			INCSTAT(stats.thisFrame.numDrawCalls);
 
-			D3D::context->GSSetShader(nullptr, nullptr, 0);
-			((DX11::Renderer*)g_renderer)->RestoreCull();
+			D3D::stateman->SetGeometryShader(nullptr);
 		}
 	}
 }

Source/Core/VideoBackends/D3D/XFBEncoder.cpp

@@ -280,22 +280,21 @@ void XFBEncoder::Encode(u8* dst, u32 width, u32 height, const EFBRectangle& srcR
 
 	// Set up all the state for XFB encoding
 
-	D3D::context->PSSetShader(m_pShader, nullptr, 0);
-	D3D::context->VSSetShader(m_vShader, nullptr, 0);
+	D3D::stateman->SetPixelShader(m_pShader);
+	D3D::stateman->SetVertexShader(m_vShader);
 
 	D3D::stateman->PushBlendState(m_xfbEncodeBlendState);
 	D3D::stateman->PushDepthState(m_xfbEncodeDepthState);
 	D3D::stateman->PushRasterizerState(m_xfbEncodeRastState);
-	D3D::stateman->Apply();
 
 	D3D11_VIEWPORT vp = CD3D11_VIEWPORT(0.f, 0.f, FLOAT(width/2), FLOAT(height));
 	D3D::context->RSSetViewports(1, &vp);
 
-	D3D::context->IASetInputLayout(m_quadLayout);
-	D3D::context->IASetPrimitiveTopology(D3D11_PRIMITIVE_TOPOLOGY_TRIANGLESTRIP);
+	D3D::stateman->SetInputLayout(m_quadLayout);
+	D3D::stateman->SetPrimitiveTopology(D3D11_PRIMITIVE_TOPOLOGY_TRIANGLESTRIP);
 	UINT stride = sizeof(QuadVertex);
 	UINT offset = 0;
-	D3D::context->IASetVertexBuffers(0, 1, &m_quad, &stride, &offset);
+	D3D::stateman->SetVertexBuffer(m_quad, stride, offset);
 
 	TargetRectangle targetRect = g_renderer->ConvertEFBRectangle(srcRect);
 
@@ -309,18 +308,18 @@ void XFBEncoder::Encode(u8* dst, u32 width, u32 height, const EFBRectangle& srcR
 	params.Gamma = gamma;
 	D3D::context->UpdateSubresource(m_encodeParams, 0, nullptr, &params, 0, 0);
 
-	D3D::context->VSSetConstantBuffers(0, 1, &m_encodeParams);
-
 	D3D::context->OMSetRenderTargets(1, &m_outRTV, nullptr);
 
 	ID3D11ShaderResourceView* pEFB = FramebufferManager::GetEFBColorTexture()->GetSRV();
 
-	D3D::context->PSSetConstantBuffers(0, 1, &m_encodeParams);
-	D3D::context->PSSetShaderResources(0, 1, &pEFB);
-	D3D::context->PSSetSamplers(0, 1, &m_efbSampler);
+	D3D::stateman->SetVertexConstants(m_encodeParams);
+	D3D::stateman->SetPixelConstants(m_encodeParams);
+	D3D::stateman->SetTexture(0, pEFB);
+	D3D::stateman->SetSampler(0, m_efbSampler);
 
 	// Encode!
 
+	D3D::stateman->Apply();
 	D3D::context->Draw(4, 0);
 
 	// Copy to staging buffer
@@ -330,23 +329,20 @@ void XFBEncoder::Encode(u8* dst, u32 width, u32 height, const EFBRectangle& srcR
 
 	// Clean up state
 
-	IUnknown* nullDummy = nullptr;
-
-	D3D::context->PSSetSamplers(0, 1, (ID3D11SamplerState**)&nullDummy);
-	D3D::context->PSSetShaderResources(0, 1, (ID3D11ShaderResourceView**)&nullDummy);
-	D3D::context->PSSetConstantBuffers(0, 1, (ID3D11Buffer**)&nullDummy);
-
 	D3D::context->OMSetRenderTargets(0, nullptr, nullptr);
 
-	D3D::context->VSSetConstantBuffers(0, 1, (ID3D11Buffer**)&nullDummy);
+	D3D::stateman->SetSampler(0, nullptr);
+	D3D::stateman->SetTexture(0, nullptr);
+	D3D::stateman->SetPixelConstants(nullptr);
+	D3D::stateman->SetVertexConstants(nullptr);
+
+	D3D::stateman->SetPixelShader(nullptr);
+	D3D::stateman->SetVertexShader(nullptr);
 
 	D3D::stateman->PopRasterizerState();
 	D3D::stateman->PopDepthState();
 	D3D::stateman->PopBlendState();
 
-	D3D::context->PSSetShader(nullptr, nullptr, 0);
-	D3D::context->VSSetShader(nullptr, nullptr, 0);
-
 	// Transfer staging buffer to GameCube/Wii RAM
 
 	D3D11_MAPPED_SUBRESOURCE map = { 0 };