Merge pull request #2092 from Sonicadvance1/gles_occlusion_queries

Implement full occlusion queries for the Nexus 9.

Source/Core/VideoBackends/OGL/GLExtensions/GLExtensions.cpp

@@ -776,6 +776,15 @@ PFNGLPUSHDEBUGGROUPPROC glPushDebugGroup;
 PFNGLBUFFERSTORAGEPROC glBufferStorage;
 PFNGLNAMEDBUFFERSTORAGEEXTPROC glNamedBufferStorageEXT;
 
+// GL_NV_occlusion_query_samples
+PFNGLGENOCCLUSIONQUERIESNVPROC glGenOcclusionQueriesNV;
+PFNGLDELETEOCCLUSIONQUERIESNVPROC glDeleteOcclusionQueriesNV;
+PFNGLISOCCLUSIONQUERYNVPROC glIsOcclusionQueryNV;
+PFNGLBEGINOCCLUSIONQUERYNVPROC glBeginOcclusionQueryNV;
+PFNGLENDOCCLUSIONQUERYNVPROC glEndOcclusionQueryNV;
+PFNGLGETOCCLUSIONQUERYIVNVPROC glGetOcclusionQueryivNV;
+PFNGLGETOCCLUSIONQUERYUIVNVPROC glGetOcclusionQueryuivNV;
+
 // Creates a GLFunc object that requires a feature
 #define GLFUNC_REQUIRES(x, y) { (void**)&x, #x, y }
 // Creates a GLFunc object with a different function suffix
@@ -1254,6 +1263,15 @@ const GLFunc gl_function_array[] =
 	// EXT_geometry_shader
 	GLFUNC_SUFFIX(glFramebufferTexture, EXT, "GL_EXT_geometry_shader !VERSION_3_2"),
 
+	// NV_occlusion_query_samples
+	GLFUNC_REQUIRES(glGenOcclusionQueriesNV,    "GL_NV_occlusion_query_samples"),
+	GLFUNC_REQUIRES(glDeleteOcclusionQueriesNV, "GL_NV_occlusion_query_samples"),
+	GLFUNC_REQUIRES(glIsOcclusionQueryNV,       "GL_NV_occlusion_query_samples"),
+	GLFUNC_REQUIRES(glBeginOcclusionQueryNV,    "GL_NV_occlusion_query_samples"),
+	GLFUNC_REQUIRES(glEndOcclusionQueryNV,      "GL_NV_occlusion_query_samples"),
+	GLFUNC_REQUIRES(glGetOcclusionQueryivNV,    "GL_NV_occlusion_query_samples"),
+	GLFUNC_REQUIRES(glGetOcclusionQueryuivNV,   "GL_NV_occlusion_query_samples"),
+
 	// gl_1_1
 	// OpenGL 1.1 is at the end due to a bug in Android's EGL stack.
 	// eglGetProcAddress can only return a finite amount of function pointers

Source/Core/VideoBackends/OGL/GLExtensions/GLExtensions.h

@@ -31,6 +31,7 @@
 #include "VideoBackends/OGL/GLExtensions/gl_3_1.h"
 #include "VideoBackends/OGL/GLExtensions/gl_3_2.h"
 #include "VideoBackends/OGL/GLExtensions/KHR_debug.h"
+#include "VideoBackends/OGL/GLExtensions/NV_occlusion_query_samples.h"
 #include "VideoBackends/OGL/GLExtensions/NV_primitive_restart.h"
 
 namespace GLExtensions

Source/Core/VideoBackends/OGL/GLExtensions/NV_occlusion_query_samples.h

@@ -0,0 +1,13 @@
+// Copyright 2013 Dolphin Emulator Project
+// Licensed under GPLv2
+// Refer to the license.txt file included.
+
+#include "VideoBackends/OGL/GLExtensions/gl_common.h"
+
+extern PFNGLGENOCCLUSIONQUERIESNVPROC glGenOcclusionQueriesNV;
+extern PFNGLDELETEOCCLUSIONQUERIESNVPROC glDeleteOcclusionQueriesNV;
+extern PFNGLISOCCLUSIONQUERYNVPROC glIsOcclusionQueryNV;
+extern PFNGLBEGINOCCLUSIONQUERYNVPROC glBeginOcclusionQueryNV;
+extern PFNGLENDOCCLUSIONQUERYNVPROC glEndOcclusionQueryNV;
+extern PFNGLGETOCCLUSIONQUERYIVNVPROC glGetOcclusionQueryivNV;
+extern PFNGLGETOCCLUSIONQUERYUIVNVPROC glGetOcclusionQueryuivNV;

Source/Core/VideoBackends/OGL/OGL.vcxproj

@@ -83,6 +83,7 @@
     <ClInclude Include="GLExtensions\gl_3_2.h" />
     <ClInclude Include="GLExtensions\gl_common.h" />
     <ClInclude Include="GLExtensions\KHR_debug.h" />
+    <ClInclude Include="GLExtensions\NV_occlusion_query_samples.h" />
     <ClInclude Include="GLExtensions\NV_primitive_restart.h" />
     <ClInclude Include="GLInterface\WGL.h" />
     <ClInclude Include="GLUtil.h" />

Source/Core/VideoBackends/OGL/OGL.vcxproj.filters

@@ -187,6 +187,9 @@
     <ClInclude Include="GLExtensions\GLExtensions.h">
       <Filter>GLExtensions</Filter>
     </ClInclude>
+    <ClInclude Include="GLExtensions\NV_occlusion_query_samples.h">
+      <Filter>GLExtensions</Filter>
+    </ClInclude>
     <ClInclude Include="GLExtensions\NV_primitive_restart.h">
       <Filter>GLExtensions</Filter>
     </ClInclude>

Source/Core/VideoBackends/OGL/PerfQuery.cpp

@@ -9,24 +9,90 @@
 
 namespace OGL
 {
+PerfQueryBase* GetPerfQuery()
+{
+	if (GLInterface->GetMode() == GLInterfaceMode::MODE_OPENGLES3 &&
+	    GLExtensions::Supports("GL_NV_occlusion_query_samples"))
+		return new PerfQueryGLESNV();
+	else if (GLInterface->GetMode() == GLInterfaceMode::MODE_OPENGLES3)
+		return new PerfQueryGL(GL_ANY_SAMPLES_PASSED);
+	else
+		return new PerfQueryGL(GL_SAMPLES_PASSED);
+}
 
 PerfQuery::PerfQuery()
 	: m_query_read_pos()
 	, m_query_count()
+{
+	ResetQuery();
+}
+
+void PerfQuery::EnableQuery(PerfQueryGroup type)
+{
+	m_query->EnableQuery(type);
+}
+
+void PerfQuery::DisableQuery(PerfQueryGroup type)
+{
+	m_query->DisableQuery(type);
+}
+
+bool PerfQuery::IsFlushed() const
+{
+	return 0 == m_query_count;
+}
+
+// TODO: could selectively flush things, but I don't think that will do much
+void PerfQuery::FlushResults()
+{
+	m_query->FlushResults();
+}
+
+void PerfQuery::ResetQuery()
+{
+	m_query_count = 0;
+	std::fill_n(m_results, ArraySize(m_results), 0);
+}
+
+u32 PerfQuery::GetQueryResult(PerfQueryType type)
+{
+	u32 result = 0;
+
+	if (type == PQ_ZCOMP_INPUT_ZCOMPLOC || type == PQ_ZCOMP_OUTPUT_ZCOMPLOC)
+	{
+		result = m_results[PQG_ZCOMP_ZCOMPLOC];
+	}
+	else if (type == PQ_ZCOMP_INPUT || type == PQ_ZCOMP_OUTPUT)
+	{
+		result = m_results[PQG_ZCOMP];
+	}
+	else if (type == PQ_BLEND_INPUT)
+	{
+		result = m_results[PQG_ZCOMP] + m_results[PQG_ZCOMP_ZCOMPLOC];
+	}
+	else if (type == PQ_EFB_COPY_CLOCKS)
+	{
+		result = m_results[PQG_EFB_COPY_CLOCKS];
+	}
+
+	return result / 4;
+}
+
+// Implementations
+PerfQueryGL::PerfQueryGL(GLenum query_type)
+	: m_query_type(query_type)
 {
 	for (ActiveQuery& query : m_query_buffer)
 		glGenQueries(1, &query.query_id);
-
-	ResetQuery();
 }
 
-PerfQuery::~PerfQuery()
+PerfQueryGL::~PerfQueryGL()
 {
 	for (ActiveQuery& query : m_query_buffer)
 		glDeleteQueries(1, &query.query_id);
 }
 
-void PerfQuery::EnableQuery(PerfQueryGroup type)
+void PerfQueryGL::EnableQuery(PerfQueryGroup type)
 {
 	// Is this sane?
 	if (m_query_count > m_query_buffer.size() / 2)
@@ -43,28 +109,42 @@ void PerfQuery::EnableQuery(PerfQueryGroup type)
 	{
 		auto& entry = m_query_buffer[(m_query_read_pos + m_query_count) % m_query_buffer.size()];
 
-		glBeginQuery(GLInterface->GetMode() == GLInterfaceMode::MODE_OPENGL ? GL_SAMPLES_PASSED : GL_ANY_SAMPLES_PASSED, entry.query_id);
+		glBeginQuery(m_query_type, entry.query_id);
 		entry.query_type = type;
 
 		++m_query_count;
 	}
 }
-
-void PerfQuery::DisableQuery(PerfQueryGroup type)
+void PerfQueryGL::DisableQuery(PerfQueryGroup type)
 {
 	// stop query
 	if (type == PQG_ZCOMP_ZCOMPLOC || type == PQG_ZCOMP)
 	{
-		glEndQuery(GLInterface->GetMode() == GLInterfaceMode::MODE_OPENGL ? GL_SAMPLES_PASSED : GL_ANY_SAMPLES_PASSED);
+		glEndQuery(m_query_type);
 	}
 }
 
-bool PerfQuery::IsFlushed() const
+void PerfQueryGL::WeakFlush()
 {
-	return 0 == m_query_count;
+	while (!IsFlushed())
+	{
+		auto& entry = m_query_buffer[m_query_read_pos];
+
+		GLuint result = GL_FALSE;
+		glGetQueryObjectuiv(entry.query_id, GL_QUERY_RESULT_AVAILABLE, &result);
+
+		if (GL_TRUE == result)
+		{
+			FlushOne();
+		}
+		else
+		{
+			break;
+		}
+	}
 }
 
-void PerfQuery::FlushOne()
+void PerfQueryGL::FlushOne()
 {
 	auto& entry = m_query_buffer[m_query_read_pos];
 
@@ -79,20 +159,64 @@ void PerfQuery::FlushOne()
 }
 
 // TODO: could selectively flush things, but I don't think that will do much
-void PerfQuery::FlushResults()
+void PerfQueryGL::FlushResults()
 {
 	while (!IsFlushed())
 		FlushOne();
 }
 
-void PerfQuery::WeakFlush()
+PerfQueryGLESNV::PerfQueryGLESNV()
+{
+	for (ActiveQuery& query : m_query_buffer)
+		glGenOcclusionQueriesNV(1, &query.query_id);
+}
+
+PerfQueryGLESNV::~PerfQueryGLESNV()
+{
+	for (ActiveQuery& query : m_query_buffer)
+		glDeleteOcclusionQueriesNV(1, &query.query_id);
+}
+
+void PerfQueryGLESNV::EnableQuery(PerfQueryGroup type)
+{
+	// Is this sane?
+	if (m_query_count > m_query_buffer.size() / 2)
+		WeakFlush();
+
+	if (m_query_buffer.size() == m_query_count)
+	{
+		FlushOne();
+		//ERROR_LOG(VIDEO, "Flushed query buffer early!");
+	}
+
+	// start query
+	if (type == PQG_ZCOMP_ZCOMPLOC || type == PQG_ZCOMP)
+	{
+		auto& entry = m_query_buffer[(m_query_read_pos + m_query_count) % m_query_buffer.size()];
+
+		glBeginOcclusionQueryNV(entry.query_id);
+		entry.query_type = type;
+
+		++m_query_count;
+	}
+}
+void PerfQueryGLESNV::DisableQuery(PerfQueryGroup type)
+{
+	// stop query
+	if (type == PQG_ZCOMP_ZCOMPLOC || type == PQG_ZCOMP)
+	{
+		glEndOcclusionQueryNV();
+	}
+}
+
+void PerfQueryGLESNV::WeakFlush()
 {
 	while (!IsFlushed())
 	{
 		auto& entry = m_query_buffer[m_query_read_pos];
 
 		GLuint result = GL_FALSE;
-		glGetQueryObjectuiv(entry.query_id, GL_QUERY_RESULT_AVAILABLE, &result);
+		glGetOcclusionQueryuivNV(entry.query_id, GL_PIXEL_COUNT_AVAILABLE_NV, &result);
 
 		if (GL_TRUE == result)
 		{
@@ -105,34 +229,25 @@ void PerfQuery::WeakFlush()
 	}
 }
 
-void PerfQuery::ResetQuery()
+void PerfQueryGLESNV::FlushOne()
 {
-	m_query_count = 0;
-	std::fill_n(m_results, ArraySize(m_results), 0);
-}
+	auto& entry = m_query_buffer[m_query_read_pos];
 
-u32 PerfQuery::GetQueryResult(PerfQueryType type)
-{
-	u32 result = 0;
+	GLuint result = 0;
+	glGetOcclusionQueryuivNV(entry.query_id, GL_OCCLUSION_TEST_RESULT_HP, &result);
 
-	if (type == PQ_ZCOMP_INPUT_ZCOMPLOC || type == PQ_ZCOMP_OUTPUT_ZCOMPLOC)
-	{
-		result = m_results[PQG_ZCOMP_ZCOMPLOC];
-	}
-	else if (type == PQ_ZCOMP_INPUT || type == PQ_ZCOMP_OUTPUT)
-	{
-		result = m_results[PQG_ZCOMP];
-	}
-	else if (type == PQ_BLEND_INPUT)
-	{
-		result = m_results[PQG_ZCOMP] + m_results[PQG_ZCOMP_ZCOMPLOC];
-	}
-	else if (type == PQ_EFB_COPY_CLOCKS)
-	{
-		result = m_results[PQG_EFB_COPY_CLOCKS];
-	}
+	// NOTE: Reported pixel metrics should be referenced to native resolution
+	m_results[entry.query_type] += (u64)result * EFB_WIDTH / g_renderer->GetTargetWidth() * EFB_HEIGHT / g_renderer->GetTargetHeight();
 
-	return result / 4;
+	m_query_read_pos = (m_query_read_pos + 1) % m_query_buffer.size();
+	--m_query_count;
+}
+
+// TODO: could selectively flush things, but I don't think that will do much
+void PerfQueryGLESNV::FlushResults()
+{
+	while (!IsFlushed())
+		FlushOne();
 }
 
 } // namespace