Merge pull request RPCS3#1196 from vlj/d3d12

D3d12/RSX: Some light improvement to program management.
Bigpet · Aug 26, 2015 · 99a9fd8 · 99a9fd8
2 parents 3468a7e + e0aa74d
commit 99a9fd8
Show file tree

Hide file tree

Showing 9 changed files with 114 additions and 1 deletion.
diff --git a/rpcs3/Emu/RSX/Common/ProgramStateCache.h b/rpcs3/Emu/RSX/Common/ProgramStateCache.h
@@ -223,6 +223,8 @@ class ProgramStateCache
 		{
 			size_t hashValue = 0;
 			hashValue ^= std::hash<unsigned>()(key.vpIdx);
+			hashValue ^= std::hash<unsigned>()(key.fpIdx);
+			hashValue ^= std::hash<typename BackendTraits::PipelineProperties>()(key.properties);
 			return hashValue;
 		}
 	};

diff --git a/rpcs3/Emu/RSX/D3D12/D3D12.h b/rpcs3/Emu/RSX/D3D12/D3D12.h
@@ -240,6 +240,7 @@ inline D3D12_COMPARISON_FUNC getCompareFunc(u32 op)
 {
 	switch (op)
 	{
+	case CELL_GCM_ZERO:
 	case CELL_GCM_NEVER: return D3D12_COMPARISON_FUNC_NEVER;
 	case CELL_GCM_LESS: return D3D12_COMPARISON_FUNC_LESS;
 	case CELL_GCM_EQUAL: return D3D12_COMPARISON_FUNC_EQUAL;

diff --git a/rpcs3/Emu/RSX/D3D12/D3D12GSRender.h b/rpcs3/Emu/RSX/D3D12/D3D12GSRender.h
@@ -383,6 +383,10 @@ class D3D12GSRender : public GSRender
 	virtual void semaphorePGRAPHTextureReadRelease(u32 offset, u32 value) override;
 	virtual void semaphorePGRAPHBackendRelease(u32 offset, u32 value) override;
 	virtual void semaphorePFIFOAcquire(u32 offset, u32 value) override;
+	virtual void notifyProgramChange() override;
+	virtual void notifyBlendStateChange() override;
+	virtual void notifyDepthStencilStateChange() override;
+	virtual void notifyRasterizerStateChange() override;
 
 private:
 	void InitD2DStructures();

diff --git a/rpcs3/Emu/RSX/D3D12/D3D12PipelineState.cpp b/rpcs3/Emu/RSX/D3D12/D3D12PipelineState.cpp
@@ -28,8 +28,28 @@ void Shader::Compile(const std::string &code, SHADER_TYPE st)
 	}
 }
 
+void D3D12GSRender::notifyProgramChange()
+{
+	m_PSO = nullptr;
+}
+void D3D12GSRender::notifyBlendStateChange()
+{
+	m_PSO = nullptr;
+}
+void D3D12GSRender::notifyDepthStencilStateChange()
+{
+	m_PSO = nullptr;
+}
+void D3D12GSRender::notifyRasterizerStateChange()
+{
+	m_PSO = nullptr;
+}
+
 bool D3D12GSRender::LoadProgram()
 {
+	if (m_PSO != nullptr)
+		return true;
+
 	if (!m_cur_fragment_prog)
 	{
 		LOG_WARNING(RSX, "LoadProgram: m_cur_shader_prog == NULL");

diff --git a/rpcs3/Emu/RSX/D3D12/D3D12PipelineState.h b/rpcs3/Emu/RSX/D3D12/D3D12PipelineState.h
@@ -47,6 +47,33 @@ struct D3D12PipelineProperties
 	}
 };
 
+template<typename T>
+size_t hashStructContent(const T& structure)
+{
+	char *data = (char*)(&structure);
+	size_t result = 0;
+	for (unsigned i = 0; i < sizeof(T); i++)
+		result ^= std::hash<char>()(data[i]);
+	return result;
+}
+
+namespace std
+{
+	template <>
+	struct hash<D3D12PipelineProperties> {
+		size_t operator()(const D3D12PipelineProperties &pipelineProperties) const {
+			size_t seed = hash<unsigned>()(pipelineProperties.DepthStencilFormat) ^
+				(hash<unsigned>()(pipelineProperties.RenderTargetsFormat) << 2) ^
+				(hash<unsigned>()(pipelineProperties.Topology) << 2) ^
+				(hash<unsigned>()(pipelineProperties.numMRT) << 4);
+			seed ^= hashStructContent(pipelineProperties.Blend);
+			seed ^= hashStructContent(pipelineProperties.DepthStencil);
+			seed ^= hashStructContent(pipelineProperties.Rasterization);
+			return hash<size_t>()(seed);
+		}
+	};
+}
+
 /** Storage for a shader
 *   Embeds the D3DBlob
 */

diff --git a/rpcs3/Emu/RSX/GL/GLGSRender.h b/rpcs3/Emu/RSX/GL/GLGSRender.h
@@ -199,4 +199,9 @@ class GLGSRender final : public GSRender
 	virtual void semaphorePGRAPHTextureReadRelease(u32 offset, u32 value) override;
 	virtual void semaphorePGRAPHBackendRelease(u32 offset, u32 value) override;
 	virtual void semaphorePFIFOAcquire(u32 offset, u32 value) override;
+
+	virtual void notifyProgramChange() override {}
+	virtual void notifyBlendStateChange() override {}
+	virtual void notifyDepthStencilStateChange() override {}
+	virtual void notifyRasterizerStateChange() override {}
 };
diff --git a/rpcs3/Emu/RSX/Null/NullGSRender.h b/rpcs3/Emu/RSX/Null/NullGSRender.h
@@ -61,4 +61,9 @@ class NullGSRender final : public GSRender
 	virtual void semaphorePFIFOAcquire(u32 offset, u32 value) override
 	{
 	}
+
+	virtual void notifyProgramChange() override {}
+	virtual void notifyBlendStateChange() override {}
+	virtual void notifyDepthStencilStateChange() override {}
+	virtual void notifyRasterizerStateChange() override {}
 };
diff --git a/rpcs3/Emu/RSX/RSXThread.cpp b/rpcs3/Emu/RSX/RSXThread.cpp
@@ -527,6 +527,7 @@ void RSXThread::DoCmd(const u32 fcmd, const u32 cmd, const u32 args_addr, const
 		m_color_mask_r = a0 & 0x0010000 ? true : false;
 		m_color_mask_g = a0 & 0x0000100 ? true : false;
 		m_color_mask_b = a0 & 0x0000001 ? true : false;
+		notifyRasterizerStateChange();
 		break;
 	}
 
@@ -572,26 +573,30 @@ void RSXThread::DoCmd(const u32 fcmd, const u32 cmd, const u32 args_addr, const
 	case NV4097_SET_CULL_FACE_ENABLE:
 	{
 		m_set_cull_face = ARGS(0) ? true : false;
+		notifyRasterizerStateChange();
 		break;
 	}
 
 	case NV4097_SET_CULL_FACE:
 	{
 		m_cull_face = ARGS(0);
+		notifyRasterizerStateChange();
 		break;
 	}
 
 	// Front face 
 	case NV4097_SET_FRONT_FACE:
 	{
 		m_front_face = ARGS(0);
+		notifyRasterizerStateChange();
 		break;
 	}
 
 	// Blending
 	case NV4097_SET_BLEND_ENABLE:
 	{
 		m_set_blend = ARGS(0) ? true : false;
+		notifyBlendStateChange();
 		break;
 	}
 
@@ -600,6 +605,7 @@ void RSXThread::DoCmd(const u32 fcmd, const u32 cmd, const u32 args_addr, const
 		m_set_blend_mrt1 = ARGS(0) & 0x02 ? true : false;
 		m_set_blend_mrt2 = ARGS(0) & 0x04 ? true : false;
 		m_set_blend_mrt3 = ARGS(0) & 0x08 ? true : false;
+		notifyBlendStateChange();
 		break;
 	}
 
@@ -615,6 +621,7 @@ void RSXThread::DoCmd(const u32 fcmd, const u32 cmd, const u32 args_addr, const
 			m_blend_dfactor_rgb = ARGS(1) & 0xffff;
 			m_blend_dfactor_alpha = ARGS(1) >> 16;
 		}
+		notifyBlendStateChange();
 		break;
 	}
 
@@ -623,6 +630,7 @@ void RSXThread::DoCmd(const u32 fcmd, const u32 cmd, const u32 args_addr, const
 		m_set_blend_dfactor = true;
 		m_blend_dfactor_rgb = ARGS(0) & 0xffff;
 		m_blend_dfactor_alpha = ARGS(0) >> 16;
+		notifyBlendStateChange();
 		break;
 	}
 
@@ -633,6 +641,7 @@ void RSXThread::DoCmd(const u32 fcmd, const u32 cmd, const u32 args_addr, const
 		m_blend_color_g = (ARGS(0) >> 8) & 0xff;
 		m_blend_color_b = (ARGS(0) >> 16) & 0xff;
 		m_blend_color_a = (ARGS(0) >> 24) & 0xff;
+		notifyBlendStateChange();
 		break;
 	}
 
@@ -650,6 +659,7 @@ void RSXThread::DoCmd(const u32 fcmd, const u32 cmd, const u32 args_addr, const
 		m_set_blend_equation = true;
 		m_blend_equation_rgb = ARGS(0) & 0xffff;
 		m_blend_equation_alpha = ARGS(0) >> 16;
+		notifyBlendStateChange();
 		break;
 	}
 
@@ -762,13 +772,15 @@ void RSXThread::DoCmd(const u32 fcmd, const u32 cmd, const u32 args_addr, const
 	{
 		m_set_depth_func = true;
 		m_depth_func = ARGS(0);
+		notifyDepthStencilStateChange();
 		break;
 	}
 
 	case NV4097_SET_DEPTH_MASK:
 	{
 		m_set_depth_mask = true;
 		m_depth_mask = ARGS(0);
+		notifyDepthStencilStateChange();
 		break;
 	}
 
@@ -1052,6 +1064,7 @@ void RSXThread::DoCmd(const u32 fcmd, const u32 cmd, const u32 args_addr, const
 		m_cur_fragment_prog->offset = a0 & ~0x3;
 		m_cur_fragment_prog->addr = GetAddress(m_cur_fragment_prog->offset, (a0 & 0x3) - 1);
 		m_cur_fragment_prog->ctrl = 0x40;
+		notifyProgramChange();
 		break;
 	}
 
@@ -1101,6 +1114,7 @@ void RSXThread::DoCmd(const u32 fcmd, const u32 cmd, const u32 args_addr, const
 				LOG_WARNING(RSX, "NV4097_SET_TRANSFORM_PROGRAM_LOAD: start = %d", start);
 			}
 		}
+		notifyProgramChange();
 		break;
 	}
 
@@ -1128,6 +1142,7 @@ void RSXThread::DoCmd(const u32 fcmd, const u32 cmd, const u32 args_addr, const
 		{
 			m_cur_vertex_prog->data.push_back(ARGS(i));
 		}
+		notifyProgramChange();
 		break;
 	}
 
@@ -1220,12 +1235,14 @@ void RSXThread::DoCmd(const u32 fcmd, const u32 cmd, const u32 args_addr, const
 	case NV4097_SET_LOGIC_OP_ENABLE:
 	{
 		m_set_logic_op = ARGS(0) ? true : false;
+		notifyBlendStateChange();
 		break;
 	}
 
 	case NV4097_SET_LOGIC_OP:
 	{
 		m_logic_op = ARGS(0);
+		notifyBlendStateChange();
 		break;
 	}
 
@@ -1240,6 +1257,7 @@ void RSXThread::DoCmd(const u32 fcmd, const u32 cmd, const u32 args_addr, const
 	case NV4097_SET_STENCIL_TEST_ENABLE:
 	{
 		m_set_stencil_test = ARGS(0) ? true : false;
+		notifyDepthStencilStateChange();
 		break;
 	}
 
@@ -1259,6 +1277,7 @@ void RSXThread::DoCmd(const u32 fcmd, const u32 cmd, const u32 args_addr, const
 	{
 		m_set_stencil_mask = true;
 		m_stencil_mask = ARGS(0);
+		notifyDepthStencilStateChange();
 		break;
 	}
 
@@ -1278,20 +1297,23 @@ void RSXThread::DoCmd(const u32 fcmd, const u32 cmd, const u32 args_addr, const
 				m_stencil_func_mask = ARGS(2);
 			}
 		}
+		notifyDepthStencilStateChange();
 		break;
 	}
 
 	case NV4097_SET_STENCIL_FUNC_REF:
 	{
 		m_set_stencil_func_ref = true;
 		m_stencil_func_ref = ARGS(0);
+		notifyDepthStencilStateChange();
 		break;
 	}
 
 	case NV4097_SET_STENCIL_FUNC_MASK:
 	{
 		m_set_stencil_func_mask = true;
 		m_stencil_func_mask = ARGS(0);
+		notifyDepthStencilStateChange();
 		break;
 	}
 
@@ -1311,13 +1333,15 @@ void RSXThread::DoCmd(const u32 fcmd, const u32 cmd, const u32 args_addr, const
 				m_stencil_zpass = ARGS(2);
 			}
 		}
+		notifyDepthStencilStateChange();
 		break;
 	}
 
 	case NV4097_SET_BACK_STENCIL_MASK:
 	{
 		m_set_back_stencil_mask = true;
 		m_back_stencil_mask = ARGS(0);
+		notifyDepthStencilStateChange();
 		break;
 	}
 
@@ -1337,20 +1361,23 @@ void RSXThread::DoCmd(const u32 fcmd, const u32 cmd, const u32 args_addr, const
 				m_back_stencil_func_mask = ARGS(2);
 			}
 		}
+		notifyDepthStencilStateChange();
 		break;
 	}
 
 	case NV4097_SET_BACK_STENCIL_FUNC_REF:
 	{
 		m_set_back_stencil_func_ref = true;
 		m_back_stencil_func_ref = ARGS(0);
+		notifyDepthStencilStateChange();
 		break;
 	}
 
 	case NV4097_SET_BACK_STENCIL_FUNC_MASK:
 	{
 		m_set_back_stencil_func_mask = true;
 		m_back_stencil_func_mask = ARGS(0);
+		notifyDepthStencilStateChange();
 		break;
 	}
 
@@ -1370,6 +1397,7 @@ void RSXThread::DoCmd(const u32 fcmd, const u32 cmd, const u32 args_addr, const
 				m_back_stencil_zpass = ARGS(2);
 			}
 		}
+		notifyDepthStencilStateChange();
 		break;
 	}
 
@@ -1484,7 +1512,6 @@ void RSXThread::DoCmd(const u32 fcmd, const u32 cmd, const u32 args_addr, const
 		m_height_scale = (float)res.height / m_height * 2.0f;
 		m_width = (u32)res.width;
 		m_height = (u32)res.height;
-
 		break;
 	}
 
@@ -1746,6 +1773,7 @@ void RSXThread::DoCmd(const u32 fcmd, const u32 cmd, const u32 args_addr, const
 
 		m_set_depth_test = a0 & 0x1 ? true : false;
 		m_set_stencil_test = a0 & 0x2 ? true : false;
+		notifyDepthStencilStateChange();
 		break;
 	}
 

diff --git a/rpcs3/Emu/RSX/RSXThread.h b/rpcs3/Emu/RSX/RSXThread.h
@@ -685,6 +685,27 @@ class RSXThread : protected named_thread_t
 	* waiting.
 	*/
 	virtual void semaphorePFIFOAcquire(u32 offset, u32 value) = 0;
+	/**
+	 * Called when vertex or fragment shader changes.
+	 * Backend can reuse same program if no change has been notified.
+	 */
+	virtual void notifyProgramChange() = 0;
+	/**
+	* Called when blend state changes.
+	* Backend can reuse same program if no change has been notified.
+	*/
+	virtual void notifyBlendStateChange() = 0;
+	/**
+	* Called when depth stencil state changes.
+	* Backend can reuse same program if no change has been notified.
+	*/
+	virtual void notifyDepthStencilStateChange() = 0;
+	/**
+	* Called when rasterizer state changes.
+	* Rasterizer state includes culling, color masking
+	* Backend can reuse same program if no change has been notified.
+	*/
+	virtual void notifyRasterizerStateChange() = 0;
 
 	void LoadVertexData(u32 first, u32 count)
 	{