Attempt to compile shaders asynchronously, #2701

src/Combiner.cpp

@@ -2,6 +2,7 @@
 #include <functional>
 #include <cstring>
 #include <stdio.h>
+#include <thread>
 #include <osal_files.h>
 
 #include "Combiner.h"
@@ -108,6 +109,13 @@ void CombinerInfo::init()
 		setCombine(EncodeCombineMode(0, 0, 0, TEXEL0, 0, 0, 0, TEXEL0, 0, 0, 0, TEXEL0, 0, 0, 0, TEXEL0));
 		gDP.otherMode.cycleType = G_CYC_FILL;
 		setCombine(EncodeCombineMode(0, 0, 0, SHADE, 0, 0, 0, SHADE, 0, 0, 0, SHADE, 0, 0, 0, SHADE));
+
+		gDP.otherMode.cycleType = G_CYC_1CYCLE;
+		setCombine(EncodeCombineMode(0, 0, 0, TEXEL0, 0, 0, 0, TEXEL0, 0, 0, 0, TEXEL0, 0, 0, 0, TEXEL0));
+		m_defaultKeys[0] = m_pCurrent->getKey();
+		gDP.otherMode.cycleType = G_CYC_2CYCLE;
+		setCombine(EncodeCombineMode(0, 0, 0, TEXEL0, 0, 0, 0, TEXEL1, 0, 0, 0, TEXEL0, 0, 0, 0, TEXEL1));
+		m_defaultKeys[1] = m_pCurrent->getKey();
 	}
 
 	m_shadowmapProgram.reset(gfxContext.createDepthFogShader());
@@ -299,13 +307,58 @@ void CombinerInfo::setCombine(u64 _mux )
 	if (iter != m_combiners.end()) {
 		m_pCurrent = iter->second;
 	} else {
+		if (gDP.otherMode.cycleType <= G_CYC_2CYCLE &&
+			config.generalEmulation.enableAsyncShadersCompilation != 0 &&
+			config.video.threadedVideo != 0)
+		{
+			iter = m_combiners.find(m_defaultKeys[gDP.otherMode.cycleType]);
+			if (iter != m_combiners.end()) {
+				m_pCurrent = iter->second;
+				if (m_unknownCombiners.insert(key).second)
+					m_combinersToCompile.push_back(key);
+				m_bChanged = true;
+				return;
+			}
+		}
+
 		m_pCurrent = Combiner_Compile(key);
 		m_pCurrent->update(true);
 		m_combiners[m_pCurrent->getKey()] = m_pCurrent;
 	}
 	m_bChanged = true;
 }
 
+void CombinerInfo::compileUnknownShaders()
+{
+#if 1
+	if (m_combinersToCompile.empty())
+		return;
+
+	std::thread task([this] {
+		std::unique_lock<std::mutex> lock(m_compileCombinersMutex);
+
+		for (auto& key : m_combinersToCompile)
+		{
+			m_pCurrent = Combiner_Compile(key);
+			m_pCurrent->update(true);
+			m_combiners[m_pCurrent->getKey()] = m_pCurrent;
+		}
+		m_combinersToCompile.clear();
+	});
+	task.detach();
+#else
+	if (m_combinersToCompile.empty())
+		return;
+	for (auto& key : m_combinersToCompile)
+	{
+		m_pCurrent = Combiner_Compile(key);
+		m_pCurrent->update(true);
+		m_combiners[m_pCurrent->getKey()] = m_pCurrent;
+	}
+	m_combinersToCompile.clear();
+#endif
+}
+
 void CombinerInfo::updateParameters()
 {
 	m_pCurrent->update(false);

src/Combiner.h

@@ -3,6 +3,8 @@
 
 #include <map>
 #include <memory>
+#include <mutex>
+#include <set>
 
 #include "GLideN64.h"
 #include "GraphicsDrawer.h"
@@ -128,6 +130,8 @@ class CombinerInfo
 	void setCombine(u64 _mux);
 	void updateParameters();
 
+	void compileUnknownShaders();
+
 	void setDepthFogCombiner();
 	graphics::ShaderProgram * getTexrectUpscaleCopyProgram();
 	graphics::ShaderProgram * getTexrectColorAndDepthUpscaleCopyProgram();
@@ -163,6 +167,11 @@ class CombinerInfo
 	graphics::CombinerProgram * m_pCurrent;
 	graphics::Combiners m_combiners;
 
+	std::mutex m_compileCombinersMutex;
+	std::set<CombinerKey> m_unknownCombiners;
+	std::vector<CombinerKey> m_combinersToCompile;
+	std::array<CombinerKey, 2> m_defaultKeys;
+
 	std::unique_ptr<graphics::ShaderProgram> m_shadowmapProgram;
 	std::unique_ptr<graphics::ShaderProgram> m_texrectUpscaleCopyProgram;
 	std::unique_ptr<graphics::ShaderProgram> m_texrectColorAndDepthUpscaleCopyProgram;

src/Config.cpp

@@ -51,6 +51,7 @@ void Config::resetToDefaults()
 	generalEmulation.enableClipping = 1;
 	generalEmulation.enableCustomSettings = 1;
 	generalEmulation.enableShadersStorage = 1;
+	generalEmulation.enableAsyncShadersCompilation = 1;
 	generalEmulation.enableLegacyBlending = 0;
 	generalEmulation.enableHybridFilter = 1;
 	generalEmulation.enableInaccurateTextureCoordinates = 0;

src/Config.h

@@ -65,6 +65,7 @@ struct Config
 		u32 enableClipping;
 		u32 enableCustomSettings;
 		u32 enableShadersStorage;
+		u32 enableAsyncShadersCompilation;
 		u32 enableLegacyBlending;
 		u32 enableHybridFilter;
 		u32 enableInaccurateTextureCoordinates;

src/gDP.cpp

@@ -983,6 +983,7 @@ void gDPFullSync()
 
 	CheckInterrupts();
 
+	CombinerInfo::get().compileUnknownShaders();
 	DebugMsg( DEBUG_NORMAL, "gDPFullSync();\n" );
 }