Merge pull request #193 from Themaister/gl-sync

GL sync

Makefile

@@ -218,6 +218,7 @@ ifeq ($(HAVE_OPENGL), 1)
          LIBS += -framework OpenGL
       else
          LIBS += -lGL
+         DEFINES += -DHAVE_GL_SYNC
       endif
    endif
 

Makefile.win

@@ -114,7 +114,7 @@ endif
 
 ifeq ($(HAVE_OPENGL), 1)
    OBJ += gfx/gl.o gfx/math/matrix.o gfx/fonts/gl_font.o gfx/fonts/gl_raster_font.o gfx/gfx_context.o gfx/context/wgl_ctx.o gfx/shader_glsl.o 
-   DEFINES += -DHAVE_OPENGL -DHAVE_OVERLAY -DHAVE_GLSL
+   DEFINES += -DHAVE_OPENGL -DHAVE_OVERLAY -DHAVE_GLSL -DHAVE_GL_SYNC
    LIBS += -lopengl32 -lgdi32
 endif
 

config.def.h

@@ -239,6 +239,9 @@ static const bool disable_composition = false;
 // Video VSYNC (recommended)
 static const bool vsync = true;
 
+// Attempts to hard-synchronize CPU and GPU. Can reduce latency at cost of performance.
+static const bool hard_sync = false;
+
 // Threaded video. Will possibly increase performance significantly at cost of worse synchronization and latency.
 static const bool video_threaded = false;
 

general.h

@@ -158,6 +158,7 @@ struct settings
       unsigned fullscreen_x;
       unsigned fullscreen_y;
       bool vsync;
+      bool hard_sync;
       bool smooth;
       bool force_aspect;
       bool crop_overscan;

gfx/gl.c

@@ -124,6 +124,24 @@ static bool load_eglimage_proc(gl_t *gl)
 }
 #endif
 
+#ifdef HAVE_GL_SYNC
+static PFNGLFENCESYNCPROC pglFenceSync;
+static PFNGLDELETESYNCPROC pglDeleteSync;
+static PFNGLCLIENTWAITSYNCPROC pglClientWaitSync;
+
+static bool load_sync_proc(gl_t *gl)
+{
+   if (!gl_query_extension("ARB_sync"))
+      return false;
+
+   LOAD_GL_SYM(FenceSync);
+   LOAD_GL_SYM(DeleteSync);
+   LOAD_GL_SYM(ClientWaitSync);
+
+   return pglFenceSync && pglDeleteSync && pglClientWaitSync;
+}
+#endif
+
 #ifdef HAVE_FBO
 #if defined(_WIN32) && !defined(RARCH_CONSOLE)
 static PFNGLGENFRAMEBUFFERSPROC pglGenFramebuffers;
@@ -1448,6 +1466,19 @@ static bool gl_frame(void *data, const void *frame, unsigned width, unsigned hei
    context_swap_buffers_func();
    g_extern.frame_count++;
 
+#ifdef HAVE_GL_SYNC
+   if (gl->use_sync)
+   {
+      RARCH_PERFORMANCE_INIT(gl_fence);
+      RARCH_PERFORMANCE_START(gl_fence);
+      glClear(GL_COLOR_BUFFER_BIT);
+      GLsync sync = pglFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, 0);
+      pglClientWaitSync(sync, 0, 1000000000);
+      pglDeleteSync(sync);
+      RARCH_PERFORMANCE_STOP(gl_fence);
+   }
+#endif
+
 #if !defined(HAVE_OPENGLES) && defined(HAVE_FFMPEG)
    if (gl->pbo_readback_enable)
       gl_pbo_async_readback(gl);
@@ -1538,6 +1569,12 @@ static bool resolve_extensions(gl_t *gl)
       return false;
 #endif
 
+#ifdef HAVE_GL_SYNC
+   gl->use_sync = g_settings.video.hard_sync && load_sync_proc(gl);
+   if (gl->use_sync)
+      RARCH_LOG("[GL]: Using ARB_sync to reduce latency.\n");
+#endif
+
 #ifdef NO_GL_CLAMP_TO_BORDER
    // NOTE: This will be a serious problem for some shaders.
    gl->border_type = GL_CLAMP_TO_EDGE;

gfx/gl_common.h

@@ -282,6 +282,9 @@ typedef struct gl
    GLfloat rgui_texture_alpha;
 #endif
 
+#ifdef HAVE_GL_SYNC
+   bool use_sync;
+#endif
 } gl_t;
 
 // Windows ... <_<

retroarch.cfg

@@ -69,6 +69,9 @@
 # Video vsync.
 # video_vsync = true
 
+# Attempts to hard-synchronize CPU and GPU. Can reduce latency at cost of performance.
+# video_hard_sync = false
+
 # Use threaded video driver. Using this might improve performance at possible cost of latency and more video stuttering.
 # video_threaded = false
 

settings.c

@@ -164,6 +164,7 @@ void config_set_defaults(void)
    g_settings.video.fullscreen_y = fullscreen_y;
    g_settings.video.disable_composition = disable_composition;
    g_settings.video.vsync = vsync;
+   g_settings.video.hard_sync = hard_sync;
    g_settings.video.threaded = video_threaded;
    g_settings.video.smooth = video_smooth;
    g_settings.video.force_aspect = force_aspect;
@@ -443,6 +444,7 @@ bool config_load_file(const char *path)
    CONFIG_GET_INT(video.monitor_index, "video_monitor_index");
    CONFIG_GET_BOOL(video.disable_composition, "video_disable_composition");
    CONFIG_GET_BOOL(video.vsync, "video_vsync");
+   CONFIG_GET_BOOL(video.hard_sync, "video_hard_sync");
    CONFIG_GET_BOOL(video.threaded, "video_threaded");
    CONFIG_GET_BOOL(video.smooth, "video_smooth");
    CONFIG_GET_BOOL(video.force_aspect, "video_force_aspect");