GPU: Implement frameskip and remove forced framebuffer swap hack.

src/citra/config.cpp

@@ -59,6 +59,7 @@ void Config::ReadValues() {
     // Core
     Settings::values.cpu_core = glfw_config->GetInteger("Core", "cpu_core", Core::CPU_Interpreter);
     Settings::values.gpu_refresh_rate = glfw_config->GetInteger("Core", "gpu_refresh_rate", 30);
+    Settings::values.frame_skip = glfw_config->GetInteger("Core", "frame_skip", 0);
 
     // Data Storage
     Settings::values.use_virtual_sd = glfw_config->GetBoolean("Data Storage", "use_virtual_sd", true);

src/citra/default_ini.h

@@ -29,6 +29,7 @@ pad_sright =
 [Core]
 cpu_core = ## 0: Interpreter (default), 1: FastInterpreter (experimental)
 gpu_refresh_rate = ## 30 (default)
+frame_skip = ## 0: No frameskip (default), 1 : 2x frameskip, 2 : 4x frameskip, etc.
 
 [Data Storage]
 use_virtual_sd =

src/citra_qt/config.cpp

@@ -45,6 +45,7 @@ void Config::ReadValues() {
     qt_config->beginGroup("Core");
     Settings::values.cpu_core = qt_config->value("cpu_core", Core::CPU_Interpreter).toInt();
     Settings::values.gpu_refresh_rate = qt_config->value("gpu_refresh_rate", 30).toInt();
+    Settings::values.frame_skip = qt_config->value("frame_skip", 0).toInt();
     qt_config->endGroup();
 
     qt_config->beginGroup("Data Storage");
@@ -80,6 +81,7 @@ void Config::SaveValues() {
     qt_config->beginGroup("Core");
     qt_config->setValue("cpu_core", Settings::values.cpu_core);
     qt_config->setValue("gpu_refresh_rate", Settings::values.gpu_refresh_rate);
+    qt_config->setValue("frame_skip", Settings::values.frame_skip);
     qt_config->endGroup();
 
     qt_config->beginGroup("Data Storage");

src/core/hw/gpu.cpp

@@ -21,10 +21,14 @@ namespace GPU {
 
 Regs g_regs;
 
-static u64 frame_ticks      = 0;    ///< 268MHz / 60 frames per second
-static u32 cur_line         = 0;    ///< Current vertical screen line
-static u64 last_frame_ticks = 0;    ///< CPU tick count from last frame
+bool g_skip_frame = false;              ///< True if the current frame was skipped
+
+static u64 frame_ticks      = 0;        ///< 268MHz / gpu_refresh_rate frames per second
+static u64 line_ticks       = 0;        ///< Number of ticks for a screen line
+static u32 cur_line         = 0;        ///< Current screen line
 static u64 last_update_tick = 0;        ///< CPU ticl count from last GPU update
+static u64 frame_count      = 0;        ///< Number of frames drawn
+static bool last_skip_frame = false;    ///< True if the last frame was skipped
 
 template <typename T>
 inline void Read(T &var, const u32 raw_addr) {
@@ -178,38 +182,40 @@ template void Write<u8>(u32 addr, const u8 data);
 void Update() {
     auto& framebuffer_top = g_regs.framebuffer_config[0];
 
-    // Update the frame after a certain number of CPU ticks have elapsed. This assumes that the
-    // active frame in memory is always complete to render. There also may be issues with this
-    // becoming out-of-synch with GSP synchrinization code (as follows). At this time, this seems to
-    // be the most effective solution for both homebrew and retail applications. With retail, this
-    // could be moved below (and probably would guarantee more accurate synchronization). However,
-    // primitive homebrew relies on a vertical blank interrupt to happen inevitably (regardless of a
-    // threading reschedule).
-
-    if ((Core::g_app_core->GetTicks() - last_frame_ticks) > (GPU::frame_ticks)) {
-        VideoCore::g_renderer->SwapBuffers();
-        last_frame_ticks = Core::g_app_core->GetTicks();
-    }
-
     // Synchronize GPU on a thread reschedule: Because we cannot accurately predict a vertical
     // blank, we need to simulate it. Based on testing, it seems that retail applications work more
     // accurately when this is signalled between thread switches.
 
     if (HLE::g_reschedule) {
         u64 current_ticks = Core::g_app_core->GetTicks();
-        u64 line_ticks = (GPU::frame_ticks / framebuffer_top.height) * 16;
+        u32 num_lines = static_cast<u32>((current_ticks - last_update_tick) / line_ticks);
 
-        //// Synchronize line...
-        if ((current_ticks - last_update_tick) >= line_ticks) {
+        // Synchronize line...
+        if (num_lines > 0) {
             GSP_GPU::SignalInterrupt(GSP_GPU::InterruptId::PDC0);
-            cur_line++;
-            last_update_tick += line_ticks;
+            cur_line += num_lines;
+            last_update_tick += (num_lines * line_ticks);
         }
 
         // Synchronize frame...
         if (cur_line >= framebuffer_top.height) {
             cur_line = 0;
+            frame_count++;
+            last_skip_frame = g_skip_frame;
+            g_skip_frame = (frame_count & Settings::values.frame_skip) != 0;
+
+            // Swap buffers based on the frameskip mode, which is a little bit tricky. When
+            // a frame is being skipped, nothing is being rendered to the internal framebuffer(s).
+            // So, we should only swap frames if the last frame was rendered. The rules are:
+            //  - If frameskip == 0 (disabled), always swap buffers
+            //  - If frameskip == 1, swap buffers every other frame (starting from the first frame)
+            //  - If frameskip > 1, swap buffers every frameskip^n frames (starting from the second frame)
+
+            if ((((Settings::values.frame_skip != 1) ^ last_skip_frame) && last_skip_frame != g_skip_frame) || 
+                   Settings::values.frame_skip == 0) {
                 VideoCore::g_renderer->SwapBuffers();
+            }
+
             GSP_GPU::SignalInterrupt(GSP_GPU::InterruptId::PDC1);
         }
     }
@@ -217,10 +223,6 @@ void Update() {
 
 /// Initialize hardware
 void Init() {
-    frame_ticks = 268123480 / Settings::values.gpu_refresh_rate;
-    cur_line = 0;
-    last_update_tick = last_frame_ticks = Core::g_app_core->GetTicks();
-
     auto& framebuffer_top = g_regs.framebuffer_config[0];
     auto& framebuffer_sub = g_regs.framebuffer_config[1];
 
@@ -249,6 +251,13 @@ void Init() {
     framebuffer_sub.color_format = Regs::PixelFormat::RGB8;
     framebuffer_sub.active_fb = 0;
 
+    frame_ticks = 268123480 / Settings::values.gpu_refresh_rate;
+    line_ticks = (GPU::frame_ticks / framebuffer_top.height);
+    cur_line = 0;
+    last_update_tick = Core::g_app_core->GetTicks();
+    last_skip_frame = false;
+    g_skip_frame = false;
+
     LOG_DEBUG(HW_GPU, "initialized OK");
 }
 

src/core/hw/gpu.h

@@ -241,6 +241,7 @@ ASSERT_REG_POSITION(command_processor_config, 0x00638);
 static_assert(sizeof(Regs) == 0x1000 * sizeof(u32), "Invalid total size of register set");
 
 extern Regs g_regs;
+extern bool g_skip_frame;
 
 template <typename T>
 void Read(T &var, const u32 addr);

src/core/settings.h

@@ -31,6 +31,7 @@ struct Values {
     // Core
     int cpu_core;
     int gpu_refresh_rate;
+    int frame_skip;
 
     // Data Storage
     bool use_virtual_sd;

src/video_core/command_processor.cpp

@@ -9,6 +9,7 @@
 #include "primitive_assembly.h"
 #include "vertex_shader.h"
 #include "core/hle/service/gsp_gpu.h"
+#include "core/hw/gpu.h"
 
 #include "debug_utils/debug_utils.h"
 
@@ -31,6 +32,10 @@ static inline void WritePicaReg(u32 id, u32 value, u32 mask) {
     if (id >= registers.NumIds())
         return;
 
+    // If we're skipping this frame, only allow trigger IRQ
+    if (GPU::g_skip_frame && id != PICA_REG_INDEX(trigger_irq))
+        return;
+
     // TODO: Figure out how register masking acts on e.g. vs_uniform_setup.set_value
     u32 old_value = registers[id];
     registers[id] = (old_value & ~mask) | (value & mask);