diff --git a/CMakeLists.txt b/CMakeLists.txt
index 63171dfd4..7b497822d 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -57,7 +57,7 @@ if(NOT LIBMGBA_ONLY)
 	set(BUILD_SHARED ON CACHE BOOL "Build a shared library")
 	set(SKIP_LIBRARY OFF CACHE BOOL "Skip building the library (useful for only building libretro or OpenEmu cores)")
 	set(BUILD_GL ON CACHE BOOL "Build with OpenGL")
-	set(BUILD_GLES2 OFF CACHE BOOL "Build with OpenGL|ES 2")
+	set(BUILD_GLES2 ON CACHE BOOL "Build with OpenGL|ES 2")
 	set(BUILD_GLES3 OFF CACHE BOOL "Build with OpenGL|ES 3")
 	set(USE_EPOXY ON CACHE STRING "Build with libepoxy")
 	set(DISABLE_DEPS OFF CACHE BOOL "Build without dependencies")
@@ -447,7 +447,7 @@ set(FEATURE_DEFINES)
 set(FEATURE_FLAGS)
 set(FEATURES)
 set(ENABLES)
-if(CMAKE_SYSTEM_NAME MATCHES .*BSD)
+if(CMAKE_SYSTEM_NAME MATCHES ".*BSD|DragonFly")
 	set(LIBEDIT_LIBRARIES -ledit)
 	if (CMAKE_SYSTEM_NAME STREQUAL OpenBSD)
 		list(APPEND LIBEDIT_LIBRARIES -ltermcap)
@@ -462,9 +462,9 @@ if(BUILD_GL)
 	endif()
 endif()
 if(NOT BUILD_GL)
-	set(OPENGLE_LIBRARY "" CACHE PATH "" FORCE)
+	set(OPENGL_LIBRARY "" CACHE PATH "" FORCE)
 endif()
-if(BUILD_GLES2 AND NOT BUILD_RASPI)
+if(BUILD_GLES2 AND NOT BUILD_RASPI AND NOT CMAKE_SYSTEM_NAME MATCHES "^(Windows|Darwin|Linux|.*BSD|DragonFly|Haiku)$")
 	find_path(OPENGLES2_INCLUDE_DIR NAMES GLES2/gl2.h)
 	find_library(OPENGLES2_LIBRARY NAMES GLESv2 GLESv2_CM)
 	if(NOT OPENGLES2_INCLUDE_DIR OR NOT OPENGLES2_LIBRARY)
@@ -474,6 +474,16 @@ endif()
 if(NOT BUILD_GLES2)
 	set(OPENGLES2_LIBRARY "" CACHE PATH "" FORCE)
 endif()
+if(BUILD_GL)
+	list(APPEND OS_SRC ${CMAKE_CURRENT_SOURCE_DIR}/src/platform/opengl/gl.c)
+	list(APPEND OS_LIB ${OPENGL_LIBRARY})
+	include_directories(${OPENGL_INCLUDE_DIR})
+endif()
+if(BUILD_GLES2)
+	list(APPEND OS_SRC ${CMAKE_CURRENT_SOURCE_DIR}/src/platform/opengl/gles2.c)
+	list(APPEND OS_LIB ${OPENGLES2_LIBRARY})
+	include_directories(${OPENGLES2_INCLUDE_DIR})
+endif()
 if(BUILD_GLES3)
 	find_path(OPENGLES3_INCLUDE_DIR NAMES GLES3/gl3.h)
 	find_library(OPENGLES3_LIBRARY NAMES GLESv3 GLESv2)
@@ -733,6 +743,10 @@ if(USE_EPOXY)
 	link_directories(${EPOXY_LIBRARY_DIRS})
 	set(OPENGLES2_LIBRARY ${EPOXY_LIBRARIES})
 	set(CPACK_DEBIAN_PACKAGE_DEPENDS "${CPACK_DEBIAN_PACKAGE_DEPENDS},libepoxy0")
+elseif(BUILD_GL)
+	set(CPACK_DEBIAN_PACKAGE_DEPENDS "${CPACK_DEBIAN_PACKAGE_DEPENDS},libgl1|libgles2")
+elseif(BUILD_GLES2)
+	set(CPACK_DEBIAN_PACKAGE_DEPENDS "${CPACK_DEBIAN_PACKAGE_DEPENDS},libgles2")
 endif()
 
 if(USE_SQLITE3)
diff --git a/include/mgba/core/core.h b/include/mgba/core/core.h
index 124b02411..9cad7a3e8 100644
--- a/include/mgba/core/core.h
+++ b/include/mgba/core/core.h
@@ -64,6 +64,7 @@ struct mCore {
 	void (*deinit)(struct mCore*);
 
 	enum mPlatform (*platform)(const struct mCore*);
+	bool (*supportsFeature)(const struct mCore*, enum mCoreFeature);
 
 	void (*setSync)(struct mCore*, struct mCoreSync*);
 	void (*loadConfig)(struct mCore*, const struct mCoreConfig*);
diff --git a/include/mgba/core/interface.h b/include/mgba/core/interface.h
index 2e9ff63a7..51ba6447f 100644
--- a/include/mgba/core/interface.h
+++ b/include/mgba/core/interface.h
@@ -80,6 +80,10 @@ enum mColorFormat {
 	mCOLOR_ANY    = -1
 };
 
+enum mCoreFeature {
+	mCORE_FEATURE_OPENGL = 1,
+};
+
 struct mCoreCallbacks {
 	void* context;
 	void (*videoFrameStarted)(void* context);
diff --git a/include/mgba/internal/gba/renderers/gl.h b/include/mgba/internal/gba/renderers/gl.h
index 435c13c71..6ac96a50e 100644
--- a/include/mgba/internal/gba/renderers/gl.h
+++ b/include/mgba/internal/gba/renderers/gl.h
@@ -16,6 +16,8 @@ CXX_GUARD_START
 #include <mgba/internal/gba/renderers/common.h>
 #include <mgba/internal/gba/video.h>
 
+#if defined(BUILD_GLES2) || defined(BUILD_GLES3)
+
 #ifdef USE_EPOXY
 #include <epoxy/gl.h>
 #elif defined(BUILD_GL)
@@ -27,7 +29,7 @@ CXX_GUARD_START
 #include <GL/glext.h>
 #endif
 #else
-#include <GLES2/gl2.h>
+#include <GLES3/gl3.h>
 #endif
 
 struct GBAVideoGLAffine {
@@ -137,6 +139,9 @@ struct GBAVideoGLRenderer {
 
 	GLuint outputTex;
 
+#ifdef BUILD_GLES3
+	uint16_t shadowPalette[512];
+#endif
 	GLuint paletteTex;
 	bool paletteDirty;
 
@@ -182,6 +187,8 @@ struct GBAVideoGLRenderer {
 
 void GBAVideoGLRendererCreate(struct GBAVideoGLRenderer* renderer);
 
+#endif
+
 CXX_GUARD_END
 
 #endif
\ No newline at end of file
diff --git a/src/gb/core.c b/src/gb/core.c
index 20e383393..264f29bf1 100644
--- a/src/gb/core.c
+++ b/src/gb/core.c
@@ -138,6 +138,14 @@ static enum mPlatform _GBCorePlatform(const struct mCore* core) {
 	return PLATFORM_GB;
 }
 
+static bool _GBCoreSupportsFeature(const struct mCore* core, enum mCoreFeature feature) {
+	UNUSED(core);
+	switch (feature) {
+	default:
+		return false;
+	}
+}
+
 static void _GBCoreSetSync(struct mCore* core, struct mCoreSync* sync) {
 	struct GB* gb = core->board;
 	gb->sync = sync;
@@ -888,6 +896,7 @@ struct mCore* GBCoreCreate(void) {
 	core->init = _GBCoreInit;
 	core->deinit = _GBCoreDeinit;
 	core->platform = _GBCorePlatform;
+	core->supportsFeature = _GBCoreSupportsFeature;
 	core->setSync = _GBCoreSetSync;
 	core->loadConfig = _GBCoreLoadConfig;
 	core->desiredVideoDimensions = _GBCoreDesiredVideoDimensions;
diff --git a/src/gba/core.c b/src/gba/core.c
index 5284a04b9..839404d03 100644
--- a/src/gba/core.c
+++ b/src/gba/core.c
@@ -17,7 +17,9 @@
 #ifndef DISABLE_THREADING
 #include <mgba/feature/thread-proxy.h>
 #endif
+#ifdef BUILD_GLES2
 #include <mgba/internal/gba/renderers/gl.h>
+#endif
 #include <mgba/internal/gba/renderers/proxy.h>
 #include <mgba/internal/gba/renderers/video-software.h>
 #include <mgba/internal/gba/savedata.h>
@@ -124,7 +126,9 @@ struct mVideoLogContext;
 struct GBACore {
 	struct mCore d;
 	struct GBAVideoSoftwareRenderer renderer;
+#ifdef BUILD_GLES2
 	struct GBAVideoGLRenderer glRenderer;
+#endif
 	struct GBAVideoProxyRenderer proxyRenderer;
 	struct mVideoLogContext* logContext;
 	struct mCoreCallbacks logCallbacks;
@@ -170,8 +174,10 @@ static bool _GBACoreInit(struct mCore* core) {
 	GBAVideoSoftwareRendererCreate(&gbacore->renderer);
 	gbacore->renderer.outputBuffer = NULL;
 
+#ifdef BUILD_GLES2
 	GBAVideoGLRendererCreate(&gbacore->glRenderer);
 	gbacore->glRenderer.outputTex = -1;
+#endif
 
 #ifndef DISABLE_THREADING
 	mVideoThreadProxyCreate(&gbacore->threadProxy);
@@ -212,6 +218,20 @@ static enum mPlatform _GBACorePlatform(const struct mCore* core) {
 	return PLATFORM_GBA;
 }
 
+static bool _GBACoreSupportsFeature(const struct mCore* core, enum mCoreFeature feature) {
+	UNUSED(core);
+	switch (feature) {
+	case mCORE_FEATURE_OPENGL:
+#ifdef BUILD_GLES2
+		return true;
+#else
+		return false;
+#endif
+	default:
+		return false;
+	}
+}
+
 static void _GBACoreSetSync(struct mCore* core, struct mCoreSync* sync) {
 	struct GBA* gba = core->board;
 	gba->sync = sync;
@@ -261,9 +281,12 @@ static void _GBACoreLoadConfig(struct mCore* core, const struct mCoreConfig* con
 }
 
 static void _GBACoreDesiredVideoDimensions(struct mCore* core, unsigned* width, unsigned* height) {
+#ifdef BUILD_GLES2
 	struct GBACore* gbacore = (struct GBACore*) core;
-	int fakeBool;
 	int scale = gbacore->glRenderer.scale;
+#else
+	int scale = 1;
+#endif
 
 	*width = GBA_VIDEO_HORIZONTAL_PIXELS * scale;
 	*height = GBA_VIDEO_VERTICAL_PIXELS * scale;
@@ -277,8 +300,13 @@ static void _GBACoreSetVideoBuffer(struct mCore* core, color_t* buffer, size_t s
 }
 
 static void _GBACoreSetVideoGLTex(struct mCore* core, unsigned texid) {
+#ifdef BUILD_GLES2
 	struct GBACore* gbacore = (struct GBACore*) core;
 	gbacore->glRenderer.outputTex = texid;
+#else
+	UNUSED(core);
+	UNUSED(texid);
+#endif
 }
 
 static void _GBACoreGetPixels(struct mCore* core, const void** buffer, size_t* stride) {
@@ -403,16 +431,22 @@ static void _GBACoreChecksum(const struct mCore* core, void* data, enum mCoreChe
 static void _GBACoreReset(struct mCore* core) {
 	struct GBACore* gbacore = (struct GBACore*) core;
 	struct GBA* gba = (struct GBA*) core->board;
-	if (gbacore->renderer.outputBuffer || gbacore->glRenderer.outputTex != (unsigned) -1) {
+	if (gbacore->renderer.outputBuffer
+#ifdef BUILD_GLES2
+	    || gbacore->glRenderer.outputTex != (unsigned) -1
+#endif
+	) {
 		struct GBAVideoRenderer* renderer;
 		if (gbacore->renderer.outputBuffer) {
 			renderer = &gbacore->renderer.d;
 		}
 		int fakeBool;
+#ifdef BUILD_GLES2
 		if (gbacore->glRenderer.outputTex != (unsigned) -1 && mCoreConfigGetIntValue(&core->config, "hwaccelVideo", &fakeBool) && fakeBool) {
 			renderer = &gbacore->glRenderer.d;
 			mCoreConfigGetIntValue(&core->config, "videoScale", &gbacore->glRenderer.scale);
 		}
+#endif
 #ifndef DISABLE_THREADING
 		if (mCoreConfigGetIntValue(&core->config, "threadedVideo", &fakeBool) && fakeBool) {
 			if (!core->videoLogger) {
@@ -952,6 +986,7 @@ struct mCore* GBACoreCreate(void) {
 	core->init = _GBACoreInit;
 	core->deinit = _GBACoreDeinit;
 	core->platform = _GBACorePlatform;
+	core->supportsFeature = _GBACoreSupportsFeature;
 	core->setSync = _GBACoreSetSync;
 	core->loadConfig = _GBACoreLoadConfig;
 	core->desiredVideoDimensions = _GBACoreDesiredVideoDimensions;
diff --git a/src/gba/renderers/gl.c b/src/gba/renderers/gl.c
index ca133d81b..bdd065cb8 100644
--- a/src/gba/renderers/gl.c
+++ b/src/gba/renderers/gl.c
@@ -5,6 +5,8 @@
  * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
 #include <mgba/internal/gba/renderers/gl.h>
 
+#if defined(BUILD_GLES2) || defined(BUILD_GLES3)
+
 #include <mgba/core/cache-set.h>
 #include <mgba/internal/arm/macros.h>
 #include <mgba/internal/gba/io.h>
@@ -90,7 +92,7 @@ static const char* const _renderTile256 =
 	"	return color;\n"
 	"}";
 
-const static struct GBAVideoGLUniform _uniformsMode0[] = {
+static const struct GBAVideoGLUniform _uniformsMode0[] = {
 	{ "loc", GBA_GL_VS_LOC, },
 	{ "maxPos", GBA_GL_VS_MAXPOS, },
 	{ "vram", GBA_GL_BG_VRAM, },
@@ -155,7 +157,7 @@ static const char* const _fetchTileNoOverflow =
 	"	return renderTile(coord);\n"
 	"}";
 
-const static struct GBAVideoGLUniform _uniformsMode2[] = {
+static const struct GBAVideoGLUniform _uniformsMode2[] = {
 	{ "loc", GBA_GL_VS_LOC, },
 	{ "maxPos", GBA_GL_VS_MAXPOS, },
 	{ "vram", GBA_GL_BG_VRAM, },
@@ -221,7 +223,7 @@ static const char* const _renderMode2 =
 	"	flags = inflags / flagCoeff;\n"
 	"}";
 
-const static struct GBAVideoGLUniform _uniformsObj[] = {
+static const struct GBAVideoGLUniform _uniformsObj[] = {
 	{ "loc", GBA_GL_VS_LOC, },
 	{ "maxPos", GBA_GL_VS_MAXPOS, },
 	{ "vram", GBA_GL_OBJ_VRAM, },
@@ -268,7 +270,7 @@ static const char* const _renderObj =
 	"	window = objwin.yz;\n"
 	"}";
 
-const static struct GBAVideoGLUniform _uniformsComposite[] = {
+static const struct GBAVideoGLUniform _uniformsComposite[] = {
 	{ "loc", GBA_GL_VS_LOC, },
 	{ "maxPos", GBA_GL_VS_MAXPOS, },
 	{ "scale", GBA_GL_COMPOSITE_SCALE, },
@@ -329,7 +331,7 @@ static const char* const _composite =
 	"	}\n"
 	"}";
 
-const static struct GBAVideoGLUniform _uniformsFinalize[] = {
+static const struct GBAVideoGLUniform _uniformsFinalize[] = {
 	{ "loc", GBA_GL_VS_LOC, },
 	{ "maxPos", GBA_GL_VS_MAXPOS, },
 	{ "scale", GBA_GL_FINALIZE_SCALE, },
@@ -413,8 +415,10 @@ void _compileShader(struct GBAVideoGLRenderer* glRenderer, GLuint program, const
 		mLOG(GBA_VIDEO, ERROR, "Program link failure: %s", log);
 	}
 	glDeleteShader(fs);
+#ifndef BUILD_GLES3
 	glBindFragDataLocation(program, 0, "color");
 	glBindFragDataLocation(program, 1, "flags");
+#endif
 }
 
 static void _initFramebufferTexture(GLuint tex, GLenum format, GLenum attachment, int scale) {
@@ -550,18 +554,24 @@ void GBAVideoGLRendererInit(struct GBAVideoRenderer* renderer) {
 	shaderBuffer[2] = _renderTile16;
 	_compileShader(glRenderer, glRenderer->objProgram[0], shaderBuffer, 3, vs, log);
 	_lookupUniforms(glRenderer->objProgram[0], glRenderer->objUniforms[0], _uniformsObj);
+#ifndef BUILD_GLES3
 	glBindFragDataLocation(glRenderer->objProgram[0], 2, "window");
+#endif
 
 	shaderBuffer[2] = _renderTile256;
 	_compileShader(glRenderer, glRenderer->objProgram[1], shaderBuffer, 3, vs, log);
 	_lookupUniforms(glRenderer->objProgram[1], glRenderer->objUniforms[1], _uniformsObj);
+#ifndef BUILD_GLES3
 	glBindFragDataLocation(glRenderer->objProgram[1], 2, "window");
+#endif
 
 	shaderBuffer[1] = _composite;
 	_compileShader(glRenderer, glRenderer->compositeProgram, shaderBuffer, 2, vs, log);
 	_lookupUniforms(glRenderer->compositeProgram, glRenderer->compositeUniforms, _uniformsComposite);
+#ifndef BUILD_GLES3
 	glBindFragDataLocation(glRenderer->compositeProgram, 2, "oldColor");
 	glBindFragDataLocation(glRenderer->compositeProgram, 3, "oldFlags");
+#endif
 
 	shaderBuffer[1] = _finalize;
 	_compileShader(glRenderer, glRenderer->finalizeProgram, shaderBuffer, 2, vs, log);
@@ -585,7 +595,6 @@ void GBAVideoGLRendererDeinit(struct GBAVideoRenderer* renderer) {
 	glDeleteProgram(glRenderer->bgProgram[3]);
 	glDeleteProgram(glRenderer->bgProgram[4]);
 	glDeleteProgram(glRenderer->bgProgram[5]);
-	glDeleteProgram(glRenderer->bgProgram[6]);
 	glDeleteProgram(glRenderer->objProgram[0]);
 	glDeleteProgram(glRenderer->objProgram[1]);
 	glDeleteProgram(glRenderer->compositeProgram);
@@ -611,9 +620,13 @@ void GBAVideoGLRendererWriteOAM(struct GBAVideoRenderer* renderer, uint32_t oam)
 }
 
 void GBAVideoGLRendererWritePalette(struct GBAVideoRenderer* renderer, uint32_t address, uint16_t value) {
+	struct GBAVideoGLRenderer* glRenderer = (struct GBAVideoGLRenderer*) renderer;
+#ifdef BUILD_GLES3
+	glRenderer->shadowPalette[address >> 1] = (value & 0x3F) | ((value & 0x7FE0) << 1);
+#else
 	UNUSED(address);
 	UNUSED(value);
-	struct GBAVideoGLRenderer* glRenderer = (struct GBAVideoGLRenderer*) renderer;
+#endif
 	glRenderer->paletteDirty = true;
 }
 
@@ -822,7 +835,11 @@ void GBAVideoGLRendererDrawScanline(struct GBAVideoRenderer* renderer, int y) {
 	struct GBAVideoGLRenderer* glRenderer = (struct GBAVideoGLRenderer*) renderer;
 	if (glRenderer->paletteDirty) {
 		glBindTexture(GL_TEXTURE_2D, glRenderer->paletteTex);
+#ifdef BUILD_GLES3
+		glTexImage2D(GL_TEXTURE_2D, 0, GL_RGB565, 16, 32, 0, GL_RGBA, GL_UNSIGNED_SHORT_5_6_5, glRenderer->shadowPalette);
+#else
 		glTexImage2D(GL_TEXTURE_2D, 0, GL_RGB5_A1, 16, 32, 0, GL_RGBA, GL_UNSIGNED_SHORT_1_5_5_5_REV, glRenderer->d.palette);
+#endif
 		glRenderer->paletteDirty = false;
 	}
 	int i;
@@ -841,33 +858,33 @@ void GBAVideoGLRendererDrawScanline(struct GBAVideoRenderer* renderer, int y) {
 	glBindFramebuffer(GL_FRAMEBUFFER, glRenderer->fbo[GBA_GL_FBO_COMPOSITE]);
 	glEnable(GL_SCISSOR_TEST);
 	glScissor(0, y * glRenderer->scale, GBA_VIDEO_HORIZONTAL_PIXELS * glRenderer->scale, glRenderer->scale);
-	glDrawBuffer(GL_COLOR_ATTACHMENT0);
+	glDrawBuffers(1, (GLenum[]) { GL_COLOR_ATTACHMENT0 });
 	glClear(GL_COLOR_BUFFER_BIT);
 	if (y == 0) {
 		glDisable(GL_SCISSOR_TEST);
-		glDrawBuffer(GL_COLOR_ATTACHMENT1);
+		glDrawBuffers(1, (GLenum[]) { GL_COLOR_ATTACHMENT1 });
 		glClearColor(1, (glRenderer->target1Bd | (glRenderer->target2Bd * 2) | (glRenderer->blendEffect * 4)) / 32.f,
 		             (glRenderer->blendEffect == BLEND_ALPHA ? glRenderer->blda : glRenderer->bldy) / 16.f, glRenderer->bldb / 16.f);
 		glClear(GL_COLOR_BUFFER_BIT);
 		
 		glClearColor(0, 0, 0, 0);
-		glDrawBuffer(GL_COLOR_ATTACHMENT3);
+		glDrawBuffers(1, (GLenum[]) { GL_COLOR_ATTACHMENT3 });
 		glClear(GL_COLOR_BUFFER_BIT);
 
 		glBindFramebuffer(GL_FRAMEBUFFER, glRenderer->fbo[GBA_GL_FBO_OBJ]);
-		glDrawBuffer(GL_COLOR_ATTACHMENT0);
+		glDrawBuffers(1, (GLenum[]) { GL_COLOR_ATTACHMENT0 });
 		glClear(GL_COLOR_BUFFER_BIT);
-		glDrawBuffer(GL_COLOR_ATTACHMENT1);
+		glDrawBuffers(1, (GLenum[]) { GL_COLOR_ATTACHMENT1 });
 		glClear(GL_COLOR_BUFFER_BIT);
 
 		for (i = 0; i < 4; ++i) {
 			glBindFramebuffer(GL_FRAMEBUFFER, glRenderer->bg[i].fbo);
-			glDrawBuffer(GL_COLOR_ATTACHMENT0);
+			glDrawBuffers(1, (GLenum[]) { GL_COLOR_ATTACHMENT0 });
 			glClear(GL_COLOR_BUFFER_BIT);
-			glDrawBuffer(GL_COLOR_ATTACHMENT1);
+			glDrawBuffers(1, (GLenum[]) { GL_COLOR_ATTACHMENT1 });
 			glClear(GL_COLOR_BUFFER_BIT);
 		}
-		glDrawBuffer(GL_COLOR_ATTACHMENT0);
+		glDrawBuffers(1, (GLenum[]) { GL_COLOR_ATTACHMENT0 });
 		glEnable(GL_SCISSOR_TEST);
 	}
 	glBindFramebuffer(GL_FRAMEBUFFER, 0);
@@ -1310,3 +1327,5 @@ void GBAVideoGLRendererDrawWindow(struct GBAVideoGLRenderer* renderer, int y) {
 	}
 	glBindFramebuffer(GL_FRAMEBUFFER, 0);
 }
+
+#endif
diff --git a/src/platform/qt/CMakeLists.txt b/src/platform/qt/CMakeLists.txt
index 9b3eb63f0..98a0a9e5a 100644
--- a/src/platform/qt/CMakeLists.txt
+++ b/src/platform/qt/CMakeLists.txt
@@ -47,17 +47,6 @@ if(APPLE)
 	endif()
 endif()
 
-if(BUILD_GL)
-	list(APPEND PLATFORM_SRC ${CMAKE_SOURCE_DIR}/src/platform/opengl/gl.c)
-	if(NOT WIN32 OR USE_EPOXY)
-		list(APPEND PLATFORM_SRC ${CMAKE_SOURCE_DIR}/src/platform/opengl/gles2.c)
-	endif()
-endif()
-
-if(BUILD_GLES2)
-	list(APPEND PLATFORM_SRC ${CMAKE_SOURCE_DIR}/src/platform/opengl/gles2.c)
-endif()
-
 get_target_property(QT_TYPE Qt5::Core TYPE)
 if(QT_TYPE STREQUAL STATIC_LIBRARY)
 	set(QT_STATIC ON)
@@ -245,7 +234,7 @@ if(NOT DEFINED DATADIR)
 		set(DATADIR ${CMAKE_INSTALL_DATADIR}/${BINARY_NAME})
 	endif()
 endif()
-if(BUILD_GL OR BUILD_GLES2)
+if(BUILD_GL OR BUILD_GLES2 OR BUILD_EPOXY)
 	install(DIRECTORY ${CMAKE_SOURCE_DIR}/res/shaders DESTINATION ${DATADIR} COMPONENT ${BINARY_NAME}-qt)
 endif()
 install(FILES ${CMAKE_SOURCE_DIR}/res/nointro.dat DESTINATION ${DATADIR} COMPONENT ${BINARY_NAME}-qt)
@@ -291,7 +280,7 @@ add_executable(${BINARY_NAME}-qt WIN32 MACOSX_BUNDLE main.cpp ${CMAKE_SOURCE_DIR
 set_target_properties(${BINARY_NAME}-qt PROPERTIES MACOSX_BUNDLE_INFO_PLIST ${CMAKE_SOURCE_DIR}/res/info.plist.in COMPILE_DEFINITIONS "${FEATURE_DEFINES};${FUNCTION_DEFINES};${OS_DEFINES};${QT_DEFINES}" COMPILE_OPTIONS "${FEATURE_FLAGS}")
 
 list(APPEND QT_LIBRARIES Qt5::Widgets)
-if(BUILD_GL OR BUILD_GLES2)
+if(BUILD_GL OR BUILD_GLES2 OR BUILD_EPOXY)
 	list(APPEND QT_LIBRARIES Qt5::OpenGL ${OPENGL_LIBRARY} ${OPENGLES2_LIBRARY})
 endif()
 if(QT_STATIC)
diff --git a/src/platform/qt/CoreController.cpp b/src/platform/qt/CoreController.cpp
index 37ea45cac..2c6ec39f7 100644
--- a/src/platform/qt/CoreController.cpp
+++ b/src/platform/qt/CoreController.cpp
@@ -40,16 +40,6 @@ CoreController::CoreController(mCore* core, QObject* parent)
 	m_threadContext.core = core;
 	m_threadContext.userData = this;
 
-	QSize size(1024, 2048);
-	m_buffers[0].resize(size.width() * size.height() * sizeof(color_t));
-	m_buffers[1].resize(size.width() * size.height() * sizeof(color_t));
-	m_buffers[0].fill(0xFF);
-	m_buffers[1].fill(0xFF);
-	m_activeBuffer = &m_buffers[0];
-	m_completeBuffer = m_buffers[0];
-
-	m_threadContext.core->setVideoBuffer(m_threadContext.core, reinterpret_cast<color_t*>(m_activeBuffer->data()), size.width());
-
 	m_resetActions.append([this]() {
 		if (m_autoload) {
 			mCoreLoadState(m_threadContext.core, 0, m_loadStateFlags);
@@ -91,8 +81,10 @@ CoreController::CoreController(mCore* core, QObject* parent)
 
 		controller->m_resetActions.clear();
 
-		controller->m_activeBuffer = &controller->m_buffers[0];
-		context->core->setVideoBuffer(context->core, reinterpret_cast<color_t*>(controller->m_activeBuffer->data()), controller->screenDimensions().width());
+		if (!controller->m_hwaccel) {
+			controller->m_activeBuffer = &controller->m_buffers[0];
+			context->core->setVideoBuffer(context->core, reinterpret_cast<color_t*>(controller->m_activeBuffer->data()), controller->screenDimensions().width());
+		}
 
 		controller->finishFrame();
 	};
@@ -211,6 +203,9 @@ CoreController::~CoreController() {
 
 const color_t* CoreController::drawContext() {
 	QMutexLocker locker(&m_mutex);
+	if (m_hwaccel) {
+		return nullptr;
+	}
 	return reinterpret_cast<const color_t*>(m_completeBuffer.constData());
 }
 
@@ -340,6 +335,18 @@ void CoreController::setLogger(LogController* logger) {
 }
 
 void CoreController::start() {
+	if (!m_hwaccel) {
+		QSize size(1024, 2048);
+		m_buffers[0].resize(size.width() * size.height() * sizeof(color_t));
+		m_buffers[1].resize(size.width() * size.height() * sizeof(color_t));
+		m_buffers[0].fill(0xFF);
+		m_buffers[1].fill(0xFF);
+		m_activeBuffer = &m_buffers[0];
+		m_completeBuffer = m_buffers[0];
+
+		m_threadContext.core->setVideoBuffer(m_threadContext.core, reinterpret_cast<color_t*>(m_activeBuffer->data()), size.width());
+	}
+
 	if (!m_patched) {
 		mCoreAutoloadPatch(m_threadContext.core);
 	}
@@ -800,6 +807,16 @@ void CoreController::endVideoLog() {
 	m_vl = nullptr;
 }
 
+void CoreController::setFramebufferHandle(int fb) {
+	Interrupter interrupter(this);
+	if (fb < 0) {
+		m_hwaccel = false;
+	} else {
+		m_threadContext.core->setVideoGLTex(m_threadContext.core, fb);
+		m_hwaccel = true;
+	}
+}
+
 void CoreController::updateKeys() {
 	int activeKeys = m_activeKeys | updateAutofire() | m_inputController->pollEvents();
 	m_threadContext.core->setKeys(m_threadContext.core, activeKeys);
@@ -823,17 +840,18 @@ int CoreController::updateAutofire() {
 
 void CoreController::finishFrame() {
 	QMutexLocker locker(&m_mutex);
-	memcpy(m_completeBuffer.data(), m_activeBuffer->constData(), m_activeBuffer->size());
+	if (!m_hwaccel) {
+			memcpy(m_completeBuffer.data(), m_activeBuffer->constData(), m_activeBuffer->size());
 
-	// TODO: Generalize this to triple buffering?
-	m_activeBuffer = &m_buffers[0];
-	if (m_activeBuffer == m_completeBuffer) {
-		m_activeBuffer = &m_buffers[1];
+		// TODO: Generalize this to triple buffering?
+		m_activeBuffer = &m_buffers[0];
+		if (m_activeBuffer == m_completeBuffer) {
+			m_activeBuffer = &m_buffers[1];
+		}
+		// Copy contents to avoid issues when doing frameskip
+		memcpy(m_activeBuffer->data(), m_completeBuffer.constData(), m_activeBuffer->size());
+		m_threadContext.core->setVideoBuffer(m_threadContext.core, reinterpret_cast<color_t*>(m_activeBuffer->data()), screenDimensions().width());
 	}
-	// Copy contents to avoid issues when doing frameskip
-	memcpy(m_activeBuffer->data(), m_completeBuffer.constData(), m_activeBuffer->size());
-	m_threadContext.core->setVideoBuffer(m_threadContext.core, reinterpret_cast<color_t*>(m_activeBuffer->data()), screenDimensions().width());
-
 	for (auto& action : m_frameActions) {
 		action();
 	}
diff --git a/src/platform/qt/CoreController.h b/src/platform/qt/CoreController.h
index dbcec9cf3..e31602117 100644
--- a/src/platform/qt/CoreController.h
+++ b/src/platform/qt/CoreController.h
@@ -46,6 +46,10 @@ public:
 	static const bool VIDEO_SYNC = false;
 	static const bool AUDIO_SYNC = true;
 
+	enum class Feature {
+		OPENGL = mCORE_FEATURE_OPENGL,
+	};
+
 	class Interrupter {
 	public:
 		Interrupter(CoreController*, bool fromThread = false);
@@ -69,6 +73,8 @@ public:
 
 	mPlatform platform() const;
 	QSize screenDimensions() const;
+	bool supportsFeature(Feature feature) const { return m_threadContext.core->supportsFeature(m_threadContext.core, static_cast<mCoreFeature>(feature)); }
+	bool hardwareAccelerated() const { return m_hwaccel; }
 
 	void loadConfig(ConfigController*);
 
@@ -154,6 +160,8 @@ public slots:
 	void startVideoLog(const QString& path);
 	void endVideoLog();
 
+	void setFramebufferHandle(int fb);
+
 signals:
 	void started();
 	void paused();
@@ -188,6 +196,7 @@ private:
 	QByteArray m_buffers[2];
 	QByteArray* m_activeBuffer;
 	QByteArray m_completeBuffer;
+	bool m_hwaccel = false;
 
 	std::unique_ptr<mCacheSet> m_cacheSet;
 	std::unique_ptr<Override> m_override;
diff --git a/src/platform/qt/Display.h b/src/platform/qt/Display.h
index de8c03648..f00bdb2b1 100644
--- a/src/platform/qt/Display.h
+++ b/src/platform/qt/Display.h
@@ -49,6 +49,7 @@ public:
 	virtual bool supportsShaders() const = 0;
 	virtual VideoShader* shaders() = 0;
 	virtual VideoProxy* videoProxy() { return nullptr; }
+	virtual int framebufferHandle() { return -1; }
 
 signals:
 	void showCursor();
diff --git a/src/platform/qt/DisplayGL.cpp b/src/platform/qt/DisplayGL.cpp
index 8ff657016..6a3afecd6 100644
--- a/src/platform/qt/DisplayGL.cpp
+++ b/src/platform/qt/DisplayGL.cpp
@@ -18,7 +18,7 @@
 #ifdef BUILD_GL
 #include "platform/opengl/gl.h"
 #endif
-#if !defined(_WIN32) || defined(USE_EPOXY)
+#ifdef BUILD_GLES2
 #include "platform/opengl/gles2.h"
 #ifdef _WIN32
 #include <epoxy/wgl.h>
@@ -195,6 +195,10 @@ VideoProxy* DisplayGL::videoProxy() {
 	return nullptr;
 }
 
+int DisplayGL::framebufferHandle() {
+	return m_painter->glTex();
+}
+
 PainterGL::PainterGL(int majorVersion, VideoProxy* proxy, QGLWidget* parent)
 	: m_gl(parent)
 	, m_videoProxy(proxy)
@@ -202,7 +206,7 @@ PainterGL::PainterGL(int majorVersion, VideoProxy* proxy, QGLWidget* parent)
 #ifdef BUILD_GL
 	mGLContext* glBackend;
 #endif
-#if !defined(_WIN32) || defined(USE_EPOXY)
+#ifdef BUILD_GLES2
 	mGLES2Context* gl2Backend;
 #endif
 
@@ -213,7 +217,7 @@ PainterGL::PainterGL(int majorVersion, VideoProxy* proxy, QGLWidget* parent)
 
 	QStringList extensions = QString(reinterpret_cast<const char*>(glGetString(GL_EXTENSIONS))).split(' ');
 
-#if !defined(_WIN32) || defined(USE_EPOXY)
+#ifdef BUILD_GLES2
 	if (extensions.contains("GL_ARB_framebuffer_object") && majorVersion >= 2) {
 		gl2Backend = static_cast<mGLES2Context*>(malloc(sizeof(mGLES2Context)));
 		mGLES2ContextCreate(gl2Backend);
@@ -239,7 +243,7 @@ PainterGL::PainterGL(int majorVersion, VideoProxy* proxy, QGLWidget* parent)
 	};
 
 	m_backend->init(m_backend, reinterpret_cast<WHandle>(m_gl->winId()));
-#if !defined(_WIN32) || defined(USE_EPOXY)
+#ifdef BUILD_GLES2
 	if (m_supportsShaders) {
 		m_shader.preprocessShader = static_cast<void*>(&reinterpret_cast<mGLES2Context*>(m_backend)->initialShader);
 	}
@@ -266,7 +270,7 @@ PainterGL::~PainterGL() {
 #if defined(_WIN32) && defined(USE_EPOXY)
 	epoxy_handle_external_wglMakeCurrent();
 #endif
-#if !defined(_WIN32) || defined(USE_EPOXY)
+#ifdef BUILD_GLES2
 	if (m_shader.passes) {
 		mGLES2ShaderFree(&m_shader);
 	}
@@ -339,7 +343,7 @@ void PainterGL::start() {
 	epoxy_handle_external_wglMakeCurrent();
 #endif
 
-#if !defined(_WIN32) || defined(USE_EPOXY)
+#ifdef BUILD_GLES2
 	if (m_supportsShaders && m_shader.passes) {
 		mGLES2ShaderAttach(reinterpret_cast<mGLES2Context*>(m_backend), static_cast<mGLES2Shader*>(m_shader.passes), m_shader.nPasses);
 	}
@@ -422,14 +426,16 @@ void PainterGL::performDraw() {
 
 void PainterGL::enqueue(const uint32_t* backing) {
 	m_mutex.lock();
-	uint32_t* buffer;
-	if (m_free.isEmpty()) {
-		buffer = m_queue.dequeue();
-	} else {
-		buffer = m_free.takeLast();
+	uint32_t* buffer = nullptr;
+	if (backing) {
+		if (m_free.isEmpty()) {
+			buffer = m_queue.dequeue();
+		} else {
+			buffer = m_free.takeLast();
+		}
+		QSize size = m_context->screenDimensions();
+		memcpy(buffer, backing, size.width() * size.height() * BYTES_PER_PIXEL);
 	}
-	QSize size = m_context->screenDimensions();
-	memcpy(buffer, backing, size.width() * size.height() * BYTES_PER_PIXEL);
 	m_queue.enqueue(buffer);
 	m_mutex.unlock();
 }
@@ -441,8 +447,10 @@ void PainterGL::dequeue() {
 		return;
 	}
 	uint32_t* buffer = m_queue.dequeue();
-	m_backend->postFrame(m_backend, buffer);
-	m_free.append(buffer);
+	if (buffer) {
+		m_backend->postFrame(m_backend, buffer);
+		m_free.append(buffer);
+	}
 	m_mutex.unlock();
 }
 
@@ -451,7 +459,9 @@ void PainterGL::dequeueAll() {
 	m_mutex.lock();
 	while (!m_queue.isEmpty()) {
 		buffer = m_queue.dequeue();
-		m_free.append(buffer);
+		if (buffer) {
+			m_free.append(buffer);
+		}
 	}
 	if (buffer) {
 		m_backend->postFrame(m_backend, buffer);
@@ -464,7 +474,7 @@ void PainterGL::setShaders(struct VDir* dir) {
 		return;
 	}
 	if (!m_active) {
-#if !defined(_WIN32) || defined(USE_EPOXY)
+#ifdef BUILD_GLES2
 		m_gl->makeCurrent();
 #if defined(_WIN32) && defined(USE_EPOXY)
 		epoxy_handle_external_wglMakeCurrent();
@@ -489,7 +499,7 @@ void PainterGL::clearShaders() {
 		return;
 	}
 	if (!m_active) {
-#if !defined(_WIN32) || defined(USE_EPOXY)
+#ifdef BUILD_GLES2
 		m_gl->makeCurrent();
 #if defined(_WIN32) && defined(USE_EPOXY)
 		epoxy_handle_external_wglMakeCurrent();
@@ -509,4 +519,19 @@ VideoShader* PainterGL::shaders() {
 	return &m_shader;
 }
 
+int PainterGL::glTex() {
+#ifdef BUILD_GLES2
+	if (supportsShaders()) {
+		mGLES2Context* gl2Backend = reinterpret_cast<mGLES2Context*>(m_backend);
+		return gl2Backend->tex;
+	}
+#endif
+#ifdef BUILD_GL
+	mGLContext* glBackend = reinterpret_cast<mGLContext*>(m_backend);
+	return glBackend->tex;
+#else
+	return -1;
+#endif
+}
+
 #endif
diff --git a/src/platform/qt/DisplayGL.h b/src/platform/qt/DisplayGL.h
index 8ef6f9e1c..990269fd0 100644
--- a/src/platform/qt/DisplayGL.h
+++ b/src/platform/qt/DisplayGL.h
@@ -52,6 +52,7 @@ public:
 	bool supportsShaders() const override;
 	VideoShader* shaders() override;
 	VideoProxy* videoProxy() override;
+	int framebufferHandle() override;
 
 public slots:
 	void stopDrawing() override;
@@ -111,6 +112,8 @@ public slots:
 	void clearShaders();
 	VideoShader* shaders();
 
+	int glTex();
+
 private:
 	void performDraw();
 	void dequeue();
diff --git a/src/platform/qt/Window.cpp b/src/platform/qt/Window.cpp
index 379f5cecb..b0971ddc3 100644
--- a/src/platform/qt/Window.cpp
+++ b/src/platform/qt/Window.cpp
@@ -851,6 +851,10 @@ void Window::unimplementedBiosCall(int call) {
 
 void Window::reloadDisplayDriver() {
 	if (m_controller) {
+		if (m_controller->hardwareAccelerated()) {
+			mustRestart();
+			return;
+		}
 		m_display->stopDrawing();
 		detachWidget(m_display.get());
 	}
@@ -1715,8 +1719,15 @@ void Window::setController(CoreController* controller, const QString& fname) {
 		reloadDisplayDriver();
 	}
 
-	if (m_display->videoProxy()) {
-		m_display->videoProxy()->attach(controller);
+	if (m_config->getOption("hwaccelVideo").toInt() && m_display->supportsShaders() && controller->supportsFeature(CoreController::Feature::OPENGL)) {
+		if (m_display->videoProxy()) {
+			m_display->videoProxy()->attach(controller);
+		}
+
+		int fb = m_display->framebufferHandle();
+		if (fb >= 0) {
+			controller->setFramebufferHandle(fb);
+		}
 	}
 
 	m_controller = std::shared_ptr<CoreController>(controller);
diff --git a/src/platform/sdl/CMakeLists.txt b/src/platform/sdl/CMakeLists.txt
index aa4c6cca5..36e73b686 100644
--- a/src/platform/sdl/CMakeLists.txt
+++ b/src/platform/sdl/CMakeLists.txt
@@ -77,12 +77,11 @@ if(BUILD_PANDORA)
 else()
 	if(BUILD_GL)
 		list(APPEND MAIN_SRC ${CMAKE_SOURCE_DIR}/src/platform/sdl/gl-sdl.c)
-		list(APPEND PLATFORM_SRC ${CMAKE_SOURCE_DIR}/src/platform/opengl/gl.c ${CMAKE_SOURCE_DIR}/src/platform/sdl/gl-common.c)
-		include_directories(${OPENGL_INCLUDE_DIR})
+		list(APPEND PLATFORM_SRC ${CMAKE_SOURCE_DIR}/src/platform/sdl/gl-common.c)
 	endif()
 	if(BUILD_GLES2)
 		list(APPEND MAIN_SRC ${CMAKE_SOURCE_DIR}/src/platform/sdl/gles2-sdl.c)
-		list(APPEND PLATFORM_SRC ${CMAKE_SOURCE_DIR}/src/platform/opengl/gles2.c ${CMAKE_SOURCE_DIR}/src/platform/sdl/gl-common.c)
+		list(APPEND PLATFORM_SRC ${CMAKE_SOURCE_DIR}/src/platform/sdl/gl-common.c)
 		include_directories(${OPENGLES2_INCLUDE_DIR})
 	endif()
 	if(NOT BUILD_GL AND NOT BUILD_GLES2)
diff --git a/src/platform/sdl/gles2-sdl.c b/src/platform/sdl/gles2-sdl.c
index 6cc046769..675a1139c 100644
--- a/src/platform/sdl/gles2-sdl.c
+++ b/src/platform/sdl/gles2-sdl.c
@@ -35,7 +35,9 @@ bool mSDLGLES2Init(struct mSDLRenderer* renderer) {
 #endif
 
 	size_t size = renderer->width * renderer->height * BYTES_PER_PIXEL;
-#ifndef __APPLE__
+#ifdef _WIN32
+	renderer->outputBuffer = _aligned_malloc(size, 16);
+#elif !defined(__APPLE__)
 	renderer->outputBuffer = memalign(16, size);
 #else
 	posix_memalign((void**) &renderer->outputBuffer, 16, size);
diff --git a/tools/sanitize-deb.sh b/tools/sanitize-deb.sh
index d400f0489..df00f11dd 100755
--- a/tools/sanitize-deb.sh
+++ b/tools/sanitize-deb.sh
@@ -31,6 +31,7 @@ while [ $# -gt 0 ]; do
         rmdep libav
         rmdep libedit
         rmdep libelf
+        rmdep libgl
         rmdep libpng
         rmdep libzip
         rmdep libmagickwand
@@ -45,6 +46,7 @@ while [ $# -gt 0 ]; do
         rmdep libav
         rmdep libedit
         rmdep libelf
+        rmdep libgl
         rmdep libpng
         rmdep qt
         rmdep libzip