System: Add memory-only save states and rewind

src/core/gpu.cpp

@@ -75,7 +75,7 @@ std::tuple<u32, u32> GPU::GetEffectiveDisplayResolution()
   return std::tie(m_crtc_state.display_vram_width, m_crtc_state.display_vram_height);
 }
 
-void GPU::Reset()
+void GPU::Reset(bool clear_vram)
 {
   SoftReset();
   m_set_texture_disable_mask = false;
@@ -119,12 +119,12 @@ void GPU::SoftReset()
   UpdateCommandTickEvent();
 }
 
-bool GPU::DoState(StateWrapper& sw, bool update_display)
+bool GPU::DoState(StateWrapper& sw, HostDisplayTexture** host_texture, bool update_display)
 {
   if (sw.IsReading())
   {
     // perform a reset to discard all pending draws/fb state
-    Reset();
+    Reset(host_texture == nullptr);
   }
 
   sw.Do(&m_GPUSTAT.bits);
@@ -214,16 +214,27 @@ bool GPU::DoState(StateWrapper& sw, bool update_display)
     UpdateDMARequest();
   }
 
-  if (!sw.DoMarker("GPU-VRAM"))
-    return false;
+  if (!host_texture)
+  {
+    if (!sw.DoMarker("GPU-VRAM"))
+      return false;
+
+    if (sw.IsReading())
+    {
+      // Still need a temporary here.
+      HeapArray<u16, VRAM_WIDTH* VRAM_HEIGHT> temp;
+      sw.DoBytes(temp.data(), VRAM_WIDTH * VRAM_HEIGHT * sizeof(u16));
+      UpdateVRAM(0, 0, VRAM_WIDTH, VRAM_HEIGHT, temp.data(), false, false);
+    }
+    else
+    {
+      ReadVRAM(0, 0, VRAM_WIDTH, VRAM_HEIGHT);
+      sw.DoBytes(m_vram_ptr, VRAM_WIDTH * VRAM_HEIGHT * sizeof(u16));
+    }
+  }
 
   if (sw.IsReading())
   {
-    // Still need a temporary here.
-    HeapArray<u16, VRAM_WIDTH * VRAM_HEIGHT> temp;
-    sw.DoBytes(temp.data(), VRAM_WIDTH * VRAM_HEIGHT * sizeof(u16));
-    UpdateVRAM(0, 0, VRAM_WIDTH, VRAM_HEIGHT, temp.data(), false, false);
-
     UpdateCRTCConfig();
     if (update_display)
       UpdateDisplay();
@@ -231,11 +242,6 @@ bool GPU::DoState(StateWrapper& sw, bool update_display)
     UpdateCRTCTickEvent();
     UpdateCommandTickEvent();
   }
-  else
-  {
-    ReadVRAM(0, 0, VRAM_WIDTH, VRAM_HEIGHT);
-    sw.DoBytes(m_vram_ptr, VRAM_WIDTH * VRAM_HEIGHT * sizeof(u16));
-  }
 
   return !sw.HasError();
 }

src/core/gpu.h

@@ -15,6 +15,7 @@
 class StateWrapper;
 
 class HostDisplay;
+class HostDisplayTexture;
 
 class TimingEvent;
 class Timers;
@@ -75,8 +76,8 @@ public:
   virtual bool IsHardwareRenderer() const = 0;
 
   virtual bool Initialize(HostDisplay* host_display);
-  virtual void Reset();
-  virtual bool DoState(StateWrapper& sw, bool update_display);
+  virtual void Reset(bool clear_vram);
+  virtual bool DoState(StateWrapper& sw, HostDisplayTexture** save_to_texture, bool update_display);
 
   // Graphics API state reset/restore - call when drawing the UI etc.
   virtual void ResetGraphicsAPIState();

src/core/gpu_backend.cpp

@@ -19,7 +19,7 @@ bool GPUBackend::Initialize()
   return true;
 }
 
-void GPUBackend::Reset()
+void GPUBackend::Reset(bool clear_vram)
 {
   Sync();
   m_drawing_area = {};

src/core/gpu_backend.h

@@ -23,7 +23,7 @@ public:
 
   virtual bool Initialize();
   virtual void UpdateSettings();
-  virtual void Reset();
+  virtual void Reset(bool clear_vram);
   virtual void Shutdown();
 
   GPUBackendFillVRAMCommand* NewFillVRAMCommand();

src/core/gpu_hw.cpp

@@ -91,9 +91,9 @@ bool GPU_HW::Initialize(HostDisplay* host_display)
   return true;
 }
 
-void GPU_HW::Reset()
+void GPU_HW::Reset(bool clear_vram)
 {
-  GPU::Reset();
+  GPU::Reset(clear_vram);
 
   m_batch_current_vertex_ptr = m_batch_start_vertex_ptr;
 
@@ -107,9 +107,9 @@ void GPU_HW::Reset()
   SetFullVRAMDirtyRectangle();
 }
 
-bool GPU_HW::DoState(StateWrapper& sw, bool update_display)
+bool GPU_HW::DoState(StateWrapper& sw, HostDisplayTexture** host_texture, bool update_display)
 {
-  if (!GPU::DoState(sw, update_display))
+  if (!GPU::DoState(sw, host_texture, update_display))
     return false;
 
   // invalidate the whole VRAM read texture when loading state

src/core/gpu_hw.h

@@ -32,8 +32,8 @@ public:
   virtual bool IsHardwareRenderer() const override;
 
   virtual bool Initialize(HostDisplay* host_display) override;
-  virtual void Reset() override;
-  virtual bool DoState(StateWrapper& sw, bool update_display) override;
+  virtual void Reset(bool clear_vram) override;
+  virtual bool DoState(StateWrapper& sw, HostDisplayTexture** host_texture, bool update_display) override;
 
   void UpdateResolutionScale() override final;
   std::tuple<u32, u32> GetEffectiveDisplayResolution() override final;

src/core/gpu_hw_d3d11.cpp

@@ -2,6 +2,7 @@
 #include "common/assert.h"
 #include "common/d3d11/shader_compiler.h"
 #include "common/log.h"
+#include "common/state_wrapper.h"
 #include "common/timer.h"
 #include "gpu_hw_shadergen.h"
 #include "host_display.h"
@@ -82,11 +83,49 @@ bool GPU_HW_D3D11::Initialize(HostDisplay* host_display)
   return true;
 }
 
-void GPU_HW_D3D11::Reset()
+void GPU_HW_D3D11::Reset(bool clear_vram)
 {
-  GPU_HW::Reset();
+  GPU_HW::Reset(clear_vram);
 
-  ClearFramebuffer();
+  if (clear_vram)
+    ClearFramebuffer();
+}
+
+bool GPU_HW_D3D11::DoState(StateWrapper& sw, HostDisplayTexture** host_texture, bool update_display)
+{
+  if (host_texture)
+  {
+    const CD3D11_BOX src_box(0, 0, 0, static_cast<LONG>(m_vram_texture.GetWidth()),
+                             static_cast<LONG>(m_vram_texture.GetHeight()), 1);
+    ComPtr<ID3D11Resource> resource;
+
+    if (sw.IsReading())
+    {
+      HostDisplayTexture* tex = *host_texture;
+      if (tex->GetWidth() != m_vram_texture.GetWidth() || tex->GetHeight() != m_vram_texture.GetHeight() ||
+          tex->GetSamples() != m_vram_texture.GetSamples())
+      {
+        return false;
+      }
+
+      static_cast<ID3D11ShaderResourceView*>(tex->GetHandle())->GetResource(resource.GetAddressOf());
+      m_context->CopySubresourceRegion(m_vram_texture.GetD3DTexture(), 0, 0, 0, 0, resource.Get(), 0, &src_box);
+    }
+    else
+    {
+      std::unique_ptr<HostDisplayTexture> tex =
+        m_host_display->CreateTexture(m_vram_texture.GetWidth(), m_vram_texture.GetHeight(), 1, 1,
+                                      m_vram_texture.GetSamples(), HostDisplayPixelFormat::RGBA8, nullptr, 0, false);
+      if (!tex)
+        return false;
+
+      static_cast<ID3D11ShaderResourceView*>(tex->GetHandle())->GetResource(resource.GetAddressOf());
+      m_context->CopySubresourceRegion(resource.Get(), 0, 0, 0, 0, m_vram_texture.GetD3DTexture(), 0, &src_box);
+      *host_texture = tex.release();
+    }
+  }
+
+  return GPU_HW::DoState(sw, host_texture, update_display);
 }
 
 void GPU_HW_D3D11::ResetGraphicsAPIState()

src/core/gpu_hw_d3d11.h

@@ -21,7 +21,8 @@ public:
   ~GPU_HW_D3D11() override;
 
   bool Initialize(HostDisplay* host_display) override;
-  void Reset() override;
+  void Reset(bool clear_vram) override;
+  bool DoState(StateWrapper& sw, HostDisplayTexture** host_texture, bool update_display) override;
 
   void ResetGraphicsAPIState() override;
   void RestoreGraphicsAPIState() override;

src/core/gpu_hw_opengl.cpp

@@ -1,6 +1,7 @@
 #include "gpu_hw_opengl.h"
 #include "common/assert.h"
 #include "common/log.h"
+#include "common/state_wrapper.h"
 #include "common/timer.h"
 #include "gpu_hw_shadergen.h"
 #include "host_display.h"
@@ -95,11 +96,106 @@ bool GPU_HW_OpenGL::Initialize(HostDisplay* host_display)
   return true;
 }
 
-void GPU_HW_OpenGL::Reset()
+void GPU_HW_OpenGL::Reset(bool clear_vram)
 {
-  GPU_HW::Reset();
+  GPU_HW::Reset(clear_vram);
 
-  ClearFramebuffer();
+  if (clear_vram)
+    ClearFramebuffer();
+}
+
+bool GPU_HW_OpenGL::DoState(StateWrapper& sw, HostDisplayTexture** host_texture, bool update_display)
+{
+  if (host_texture)
+  {
+    if (sw.IsReading())
+    {
+      HostDisplayTexture* tex = *host_texture;
+      if (tex->GetWidth() != m_vram_texture.GetWidth() || tex->GetHeight() != m_vram_texture.GetHeight() ||
+          tex->GetSamples() != m_vram_texture.GetSamples())
+      {
+        return false;
+      }
+
+      CopyFramebufferForState(
+        m_vram_texture.GetGLTarget(), static_cast<GLuint>(reinterpret_cast<uintptr_t>(tex->GetHandle())), 0, 0, 0,
+        m_vram_texture.GetGLId(), m_vram_fbo_id, 0, 0, m_vram_texture.GetWidth(), m_vram_texture.GetHeight());
+    }
+    else
+    {
+      std::unique_ptr<HostDisplayTexture> tex =
+        m_host_display->CreateTexture(m_vram_texture.GetWidth(), m_vram_texture.GetHeight(), 1, 1,
+                                      m_vram_texture.GetSamples(), HostDisplayPixelFormat::RGBA8, nullptr, 0, false);
+      if (!tex)
+        return false;
+
+      CopyFramebufferForState(m_vram_texture.GetGLTarget(), m_vram_texture.GetGLId(), m_vram_fbo_id, 0, 0,
+                              static_cast<GLuint>(reinterpret_cast<uintptr_t>(tex->GetHandle())), 0, 0, 0,
+                              m_vram_texture.GetWidth(), m_vram_texture.GetHeight());
+      *host_texture = tex.release();
+    }
+  }
+
+  return GPU_HW::DoState(sw, host_texture, update_display);
+}
+
+void GPU_HW_OpenGL::CopyFramebufferForState(GLenum target, GLuint src_texture, u32 src_fbo, u32 src_x, u32 src_y,
+                                            GLuint dst_texture, u32 dst_fbo, u32 dst_x, u32 dst_y, u32 width,
+                                            u32 height)
+{
+  if (target != GL_TEXTURE_2D && GLAD_GL_VERSION_4_3)
+  {
+    glCopyImageSubData(src_texture, target, 0, src_x, src_y, 0, dst_texture, target, 0, dst_x, dst_y, 0, width, height,
+                       1);
+  }
+  else if (target != GL_TEXTURE_2D && GLAD_GL_EXT_copy_image)
+  {
+    glCopyImageSubDataEXT(src_texture, target, 0, src_x, src_y, 0, dst_texture, target, 0, dst_x, dst_y, 0, width,
+                          height, 1);
+  }
+  else if (target != GL_TEXTURE_2D && GLAD_GL_OES_copy_image)
+  {
+    glCopyImageSubDataOES(src_texture, target, 0, src_x, src_y, 0, dst_texture, target, 0, dst_x, dst_y, 0, width,
+                          height, 1);
+  }
+  else
+  {
+    if (src_fbo == 0)
+    {
+      glBindFramebuffer(GL_READ_FRAMEBUFFER, m_state_copy_fbo_id);
+      glFramebufferTexture2D(GL_READ_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, target, src_texture, 0);
+    }
+    else
+    {
+      glBindFramebuffer(GL_READ_FRAMEBUFFER, src_fbo);
+    }
+
+    if (dst_fbo == 0)
+    {
+      glBindFramebuffer(GL_DRAW_FRAMEBUFFER, m_state_copy_fbo_id);
+      glFramebufferTexture2D(GL_DRAW_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, target, dst_texture, 0);
+    }
+    else
+    {
+      glBindFramebuffer(GL_DRAW_FRAMEBUFFER, dst_fbo);
+    }
+
+    glDisable(GL_SCISSOR_TEST);
+    glBlitFramebuffer(src_x, src_y, src_x + width, src_y + height, dst_x, dst_y, dst_x + width, dst_y + height,
+                      GL_COLOR_BUFFER_BIT, GL_NEAREST);
+    glEnable(GL_SCISSOR_TEST);
+
+    if (src_fbo == 0)
+    {
+      glFramebufferTexture2D(GL_READ_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_TEXTURE_2D, 0, 0);
+    }
+    else if (dst_fbo == 0)
+    {
+      glFramebufferTexture2D(GL_DRAW_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_TEXTURE_2D, 0, 0);
+    }
+
+    glBindFramebuffer(GL_DRAW_FRAMEBUFFER, m_vram_fbo_id);
+  }
 }
 
 void GPU_HW_OpenGL::ResetGraphicsAPIState()
@@ -303,7 +399,9 @@ bool GPU_HW_OpenGL::CreateFramebuffer()
     return false;
   }
 
-  glGenFramebuffers(1, &m_vram_fbo_id);
+  if (m_vram_fbo_id == 0)
+    glGenFramebuffers(1, &m_vram_fbo_id);
+
   glBindFramebuffer(GL_DRAW_FRAMEBUFFER, m_vram_fbo_id);
   glFramebufferTexture2D(GL_DRAW_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, m_vram_texture.GetGLTarget(),
                          m_vram_texture.GetGLId(), 0);
@@ -320,6 +418,9 @@ bool GPU_HW_OpenGL::CreateFramebuffer()
     }
   }
 
+  if (m_state_copy_fbo_id == 0)
+    glGenFramebuffers(1, &m_state_copy_fbo_id);
+
   SetFullVRAMDirtyRectangle();
   return true;
 }

src/core/gpu_hw_opengl.h

@@ -17,7 +17,8 @@ public:
   ~GPU_HW_OpenGL() override;
 
   bool Initialize(HostDisplay* host_display) override;
-  void Reset() override;
+  void Reset(bool clear_vram) override;
+  bool DoState(StateWrapper& sw, HostDisplayTexture** host_texture, bool update_display) override;
 
   void ResetGraphicsAPIState() override;
   void RestoreGraphicsAPIState() override;
@@ -57,6 +58,8 @@ private:
   void SetCapabilities(HostDisplay* host_display);
   bool CreateFramebuffer();
   void ClearFramebuffer();
+  void CopyFramebufferForState(GLenum target, GLuint src_texture, u32 src_fbo, u32 src_x, u32 src_y, GLuint dst_texture,
+                               u32 dst_fbo, u32 dst_x, u32 dst_y, u32 width, u32 height);
 
   bool CreateVertexBuffer();
   bool CreateUniformBuffer();
@@ -84,6 +87,7 @@ private:
   GLuint m_vram_fbo_id = 0;
   GLuint m_vao_id = 0;
   GLuint m_attributeless_vao_id = 0;
+  GLuint m_state_copy_fbo_id = 0;
 
   std::unique_ptr<GL::StreamBuffer> m_uniform_stream_buffer;
 

src/core/gpu_hw_vulkan.cpp

@@ -2,6 +2,7 @@
 #include "common/assert.h"
 #include "common/log.h"
 #include "common/scope_guard.h"
+#include "common/state_wrapper.h"
 #include "common/timer.h"
 #include "common/vulkan/builders.h"
 #include "common/vulkan/context.h"
@@ -87,12 +88,75 @@ bool GPU_HW_Vulkan::Initialize(HostDisplay* host_display)
   return true;
 }
 
-void GPU_HW_Vulkan::Reset()
+void GPU_HW_Vulkan::Reset(bool clear_vram)
 {
-  GPU_HW::Reset();
+  GPU_HW::Reset(clear_vram);
 
   EndRenderPass();
-  ClearFramebuffer();
+
+  if (clear_vram)
+    ClearFramebuffer();
+}
+
+bool GPU_HW_Vulkan::DoState(StateWrapper& sw, HostDisplayTexture** host_texture, bool update_display)
+{
+  if (host_texture)
+  {
+    EndRenderPass();
+
+    const VkImageCopy ic{{VK_IMAGE_ASPECT_COLOR_BIT, 0u, 0u, 1u},
+                         {0, 0, 0},
+                         {VK_IMAGE_ASPECT_COLOR_BIT, 0u, 0u, 1u},
+                         {0, 0, 0},
+                         {m_vram_texture.GetWidth(), m_vram_texture.GetHeight(), 1u}};
+
+    if (sw.IsReading())
+    {
+      Vulkan::Texture* tex = static_cast<Vulkan::Texture*>((*host_texture)->GetHandle());
+      if (tex->GetWidth() != m_vram_texture.GetWidth() || tex->GetHeight() != m_vram_texture.GetHeight() ||
+          tex->GetSamples() != m_vram_texture.GetSamples())
+      {
+        return false;
+      }
+
+      VkCommandBuffer buf = g_vulkan_context->GetCurrentCommandBuffer();
+      const VkImageLayout old_tex_layout = tex->GetLayout();
+      const VkImageLayout old_vram_layout = m_vram_texture.GetLayout();
+      tex->TransitionToLayout(buf, VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL);
+      m_vram_texture.TransitionToLayout(buf, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL);
+      vkCmdCopyImage(g_vulkan_context->GetCurrentCommandBuffer(), tex->GetImage(), tex->GetLayout(),
+                     m_vram_texture.GetImage(), m_vram_texture.GetLayout(), 1, &ic);
+      m_vram_texture.TransitionToLayout(buf, old_vram_layout);
+      tex->TransitionToLayout(buf, old_tex_layout);
+    }
+    else
+    {
+      std::unique_ptr<HostDisplayTexture> htex =
+        m_host_display->CreateTexture(m_vram_texture.GetWidth(), m_vram_texture.GetHeight(), 1, 1,
+                                      m_vram_texture.GetSamples(), HostDisplayPixelFormat::RGBA8, nullptr, 0, false);
+      if (!htex)
+        return false;
+
+      Vulkan::Texture* tex = static_cast<Vulkan::Texture*>(htex->GetHandle());
+      if (tex->GetWidth() != m_vram_texture.GetWidth() || tex->GetHeight() != m_vram_texture.GetHeight() ||
+          tex->GetSamples() != m_vram_texture.GetSamples())
+      {
+        return false;
+      }
+
+      VkCommandBuffer buf = g_vulkan_context->GetCurrentCommandBuffer();
+      const VkImageLayout old_vram_layout = m_vram_texture.GetLayout();
+      tex->TransitionToLayout(buf, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL);
+      m_vram_texture.TransitionToLayout(buf, VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL);
+      vkCmdCopyImage(g_vulkan_context->GetCurrentCommandBuffer(), m_vram_texture.GetImage(), m_vram_texture.GetLayout(),
+                     tex->GetImage(), tex->GetLayout(), 1, &ic);
+      m_vram_texture.TransitionToLayout(buf, old_vram_layout);
+      tex->TransitionToLayout(buf, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL);
+      *host_texture = htex.release();
+    }
+  }
+
+  return GPU_HW::DoState(sw, host_texture, update_display);
 }
 
 void GPU_HW_Vulkan::ResetGraphicsAPIState()

src/core/gpu_hw_vulkan.h

@@ -16,7 +16,8 @@ public:
   ~GPU_HW_Vulkan() override;
 
   bool Initialize(HostDisplay* host_display) override;
-  void Reset() override;
+  void Reset(bool clear_vram) override;
+  bool DoState(StateWrapper& sw, HostDisplayTexture** host_texture, bool update_display) override;
 
   void ResetGraphicsAPIState() override;
   void RestoreGraphicsAPIState() override;

src/core/gpu_sw.cpp

@@ -75,11 +75,11 @@ bool GPU_SW::Initialize(HostDisplay* host_display)
   return true;
 }
 
-void GPU_SW::Reset()
+void GPU_SW::Reset(bool clear_vram)
 {
-  GPU::Reset();
+  GPU::Reset(clear_vram);
 
-  m_backend.Reset();
+  m_backend.Reset(clear_vram);
 }
 
 void GPU_SW::UpdateSettings()

src/core/gpu_sw.h

@@ -18,7 +18,7 @@ public:
   bool IsHardwareRenderer() const override;
 
   bool Initialize(HostDisplay* host_display) override;
-  void Reset() override;
+  void Reset(bool clear_vram) override;
   void UpdateSettings() override;
 
 protected:

src/core/gpu_sw_backend.cpp

@@ -20,11 +20,12 @@ bool GPU_SW_Backend::Initialize()
   return GPUBackend::Initialize();
 }
 
-void GPU_SW_Backend::Reset()
+void GPU_SW_Backend::Reset(bool clear_vram)
 {
-  GPUBackend::Reset();
+  GPUBackend::Reset(clear_vram);
 
-  m_vram.fill(0);
+  if (clear_vram)
+    m_vram.fill(0);
 }
 
 void GPU_SW_Backend::DrawPolygon(const GPUBackendDrawPolygonCommand* cmd)

src/core/gpu_sw_backend.h

@@ -11,7 +11,7 @@ public:
   ~GPU_SW_Backend() override;
 
   bool Initialize() override;
-  void Reset() override;
+  void Reset(bool clear_vram) override;
 
   ALWAYS_INLINE_RELEASE u16 GetPixel(const u32 x, const u32 y) const { return m_vram[VRAM_WIDTH * y + x]; }
   ALWAYS_INLINE_RELEASE const u16* GetPixelPtr(const u32 x, const u32 y) const { return &m_vram[VRAM_WIDTH * y + x]; }

src/core/host_interface.cpp

@@ -488,6 +488,9 @@ void HostInterface::SetDefaultSettings(SettingsInterface& si)
   si.SetBoolValue("Main", "LoadDevicesFromSaveStates", false);
   si.SetBoolValue("Main", "ApplyGameSettings", true);
   si.SetBoolValue("Main", "DisableAllEnhancements", false);
+  si.SetBoolValue("Main", "RewindEnable", false);
+  si.SetFloatValue("Main", "RewindFrequency", 10.0f);
+  si.SetIntValue("Main", "RewindSaveSlots", 10);
 
   si.SetStringValue("CPU", "ExecutionMode", Settings::GetCPUExecutionModeName(Settings::DEFAULT_CPU_EXECUTION_MODE));
   si.SetBoolValue("CPU", "RecompilerMemoryExceptions", false);
@@ -655,6 +658,13 @@ void HostInterface::FixIncompatibleSettings(bool display_osd_messages)
     g_settings.cpu_fastmem_mode = CPUFastmemMode::LUT;
   }
 #endif
+
+  // rewinding causes issues with mmap fastmem, so just use LUT
+  if (g_settings.rewind_enable && g_settings.IsUsingFastmem() && g_settings.cpu_fastmem_mode == CPUFastmemMode::MMap)
+  {
+    Log_WarningPrintf("Disabling mmap fastmem due to rewind being enabled");
+    g_settings.cpu_fastmem_mode = CPUFastmemMode::LUT;
+  }
 }
 
 void HostInterface::SaveSettings(SettingsInterface& si)
@@ -676,6 +686,8 @@ void HostInterface::CheckForSettingsChanges(const Settings& old_settings)
 
   if (System::IsValid())
   {
+    System::ClearMemorySaveStates();
+
     if (g_settings.cpu_overclock_active != old_settings.cpu_overclock_active ||
         (g_settings.cpu_overclock_active &&
          (g_settings.cpu_overclock_numerator != old_settings.cpu_overclock_numerator ||
@@ -755,7 +767,8 @@ void HostInterface::CheckForSettingsChanges(const Settings& old_settings)
         g_settings.display_active_start_offset != old_settings.display_active_start_offset ||
         g_settings.display_active_end_offset != old_settings.display_active_end_offset ||
         g_settings.display_line_start_offset != old_settings.display_line_start_offset ||
-        g_settings.display_line_end_offset != old_settings.display_line_end_offset)
+        g_settings.display_line_end_offset != old_settings.display_line_end_offset ||
+        g_settings.rewind_enable != old_settings.rewind_enable)
     {
       if (g_settings.IsUsingCodeCache())
         CPU::CodeCache::Reinitialize();
@@ -793,6 +806,13 @@ void HostInterface::CheckForSettingsChanges(const Settings& old_settings)
       System::UpdateMemoryCards();
     }
 
+    if (g_settings.rewind_enable != old_settings.rewind_enable ||
+        g_settings.rewind_save_frequency != old_settings.rewind_save_frequency ||
+        g_settings.rewind_save_slots != old_settings.rewind_save_slots)
+    {
+      System::UpdateMemorySaveStateSettings();
+    }
+
     if (g_settings.texture_replacements.enable_vram_write_replacements !=
           old_settings.texture_replacements.enable_vram_write_replacements ||
         g_settings.texture_replacements.preload_textures != old_settings.texture_replacements.preload_textures)
@@ -999,6 +1019,7 @@ void HostInterface::ModifyResolutionScale(s32 increment)
     g_gpu->RestoreGraphicsAPIState();
     g_gpu->UpdateSettings();
     g_gpu->ResetGraphicsAPIState();
+    System::ClearMemorySaveStates();
   }
 }
 

src/core/settings.cpp

@@ -123,6 +123,9 @@ void Settings::Load(SettingsInterface& si)
   apply_game_settings = si.GetBoolValue("Main", "ApplyGameSettings", true);
   auto_load_cheats = si.GetBoolValue("Main", "AutoLoadCheats", false);
   disable_all_enhancements = si.GetBoolValue("Main", "DisableAllEnhancements", false);
+  rewind_enable = si.GetBoolValue("Main", "RewindEnable", false);
+  rewind_save_frequency = si.GetFloatValue("Main", "RewindFrequency", 10.0f);
+  rewind_save_slots = static_cast<u32>(si.GetIntValue("Main", "RewindSaveSlots", 10));
 
   cpu_execution_mode =
     ParseCPUExecutionMode(
@@ -295,6 +298,9 @@ void Settings::Save(SettingsInterface& si) const
   si.SetBoolValue("Main", "ApplyGameSettings", apply_game_settings);
   si.SetBoolValue("Main", "AutoLoadCheats", auto_load_cheats);
   si.SetBoolValue("Main", "DisableAllEnhancements", disable_all_enhancements);
+  si.SetBoolValue("Main", "RewindEnable", rewind_enable);
+  si.SetFloatValue("Main", "RewindFrequency", rewind_save_frequency);
+  si.SetIntValue("Main", "RewindSaveSlots", rewind_save_slots);
 
   si.SetStringValue("CPU", "ExecutionMode", GetCPUExecutionModeName(cpu_execution_mode));
   si.SetBoolValue("CPU", "OverclockEnable", cpu_overclock_enable);

src/core/settings.h

@@ -94,6 +94,10 @@ struct Settings
   bool auto_load_cheats = false;
   bool disable_all_enhancements = false;
 
+  bool rewind_enable = false;
+  float rewind_save_frequency = 10.0f;
+  u32 rewind_save_slots = 10;
+
   GPURenderer gpu_renderer = GPURenderer::Software;
   std::string gpu_adapter;
   std::string display_post_process_chain;

src/core/system.cpp

@@ -9,6 +9,7 @@
 #include "common/log.h"
 #include "common/state_wrapper.h"
 #include "common/string_util.h"
+#include "common/timestamp.h"
 #include "controller.h"
 #include "cpu_code_cache.h"
 #include "cpu_core.h"
@@ -30,7 +31,10 @@
 #include "texture_replacements.h"
 #include "timers.h"
 #include <cctype>
+#include <cinttypes>
+#include <cmath>
 #include <cstdio>
+#include <deque>
 #include <fstream>
 #include <limits>
 Log_SetChannel(System);
@@ -51,6 +55,12 @@ SystemBootParameters::~SystemBootParameters() = default;
 
 namespace System {
 
+struct MemorySaveState
+{
+  std::unique_ptr<HostDisplayTexture> vram_texture;
+  std::unique_ptr<GrowableMemoryByteStream> state_stream;
+};
+
 static bool LoadEXE(const char* filename);
 static bool SetExpansionROM(const char* filename);
 
@@ -58,7 +68,10 @@ static bool SetExpansionROM(const char* filename);
 static std::unique_ptr<CDImage> OpenCDImage(const char* path, bool force_preload);
 
 static bool DoLoadState(ByteStream* stream, bool force_software_renderer, bool update_display);
-static bool DoState(StateWrapper& sw, bool update_display);
+static bool DoState(StateWrapper& sw, HostDisplayTexture** host_texture, bool update_display);
+static void DoRunFrame();
+static void DoRewind();
+static void DoMemorySaveStates();
 static bool CreateGPU(GPURenderer renderer);
 
 static bool Initialize(bool force_software_renderer);
@@ -104,6 +117,15 @@ static std::string s_media_playlist_filename;
 
 static std::unique_ptr<CheatList> s_cheat_list;
 
+static bool s_memory_saves_enabled = false;
+
+static std::deque<MemorySaveState> s_rewind_states;
+static s32 s_rewind_load_frequency = -1;
+static s32 s_rewind_load_counter = -1;
+static s32 s_rewind_save_frequency = -1;
+static s32 s_rewind_save_counter = -1;
+static bool s_rewinding_first_save = false;
+
 State GetState()
 {
   return s_state;
@@ -521,12 +543,13 @@ std::optional<DiscRegion> GetRegionForPath(const char* image_path)
 
 bool RecreateGPU(GPURenderer renderer, bool update_display /* = true*/)
 {
+  ClearMemorySaveStates();
   g_gpu->RestoreGraphicsAPIState();
 
   // save current state
   std::unique_ptr<ByteStream> state_stream = ByteStream_CreateGrowableMemoryStream();
   StateWrapper sw(state_stream.get(), StateWrapper::Mode::Write, SAVE_STATE_VERSION);
-  const bool state_valid = g_gpu->DoState(sw, false) && TimingEvents::DoState(sw);
+  const bool state_valid = g_gpu->DoState(sw, nullptr, false) && TimingEvents::DoState(sw);
   if (!state_valid)
     Log_ErrorPrintf("Failed to save old GPU state when switching renderers");
 
@@ -548,7 +571,7 @@ bool RecreateGPU(GPURenderer renderer, bool update_display /* = true*/)
     state_stream->SeekAbsolute(0);
     sw.SetMode(StateWrapper::Mode::Read);
     g_gpu->RestoreGraphicsAPIState();
-    g_gpu->DoState(sw, update_display);
+    g_gpu->DoState(sw, nullptr, update_display);
     TimingEvents::DoState(sw);
     g_gpu->ResetGraphicsAPIState();
   }
@@ -802,6 +825,7 @@ bool Initialize(bool force_software_renderer)
   }
 
   UpdateThrottlePeriod();
+  UpdateMemorySaveStateSettings();
   return true;
 }
 
@@ -810,6 +834,8 @@ void Shutdown()
   if (s_state == State::Shutdown)
     return;
 
+  ClearMemorySaveStates();
+
   g_texture_replacements.Shutdown();
 
   g_sio.Shutdown();
@@ -868,11 +894,11 @@ bool CreateGPU(GPURenderer renderer)
   }
 
   // we put this here rather than in Initialize() because of the virtual calls
-  g_gpu->Reset();
+  g_gpu->Reset(true);
   return true;
 }
 
-bool DoState(StateWrapper& sw, bool update_display)
+bool DoState(StateWrapper& sw, HostDisplayTexture** host_texture, bool update_display)
 {
   if (!sw.DoMarker("System"))
     return false;
@@ -897,7 +923,7 @@ bool DoState(StateWrapper& sw, bool update_display)
     return false;
 
   g_gpu->RestoreGraphicsAPIState();
-  const bool gpu_result = sw.DoMarker("GPU") && g_gpu->DoState(sw, update_display);
+  const bool gpu_result = sw.DoMarker("GPU") && g_gpu->DoState(sw, host_texture, update_display);
   g_gpu->ResetGraphicsAPIState();
   if (!gpu_result)
     return false;
@@ -963,7 +989,7 @@ void Reset()
   Bus::Reset();
   g_dma.Reset();
   g_interrupt_controller.Reset();
-  g_gpu->Reset();
+  g_gpu->Reset(true);
   g_cdrom.Reset();
   g_pad.Reset();
   g_timers.Reset();
@@ -1076,6 +1102,7 @@ bool DoLoadState(ByteStream* state, bool force_software_renderer, bool update_di
   }
 
   UpdateRunningGame(media_filename.c_str(), media.get());
+  ClearMemorySaveStates();
 
   if (s_state == State::Starting)
   {
@@ -1117,7 +1144,7 @@ bool DoLoadState(ByteStream* state, bool force_software_renderer, bool update_di
     return false;
 
   StateWrapper sw(state, StateWrapper::Mode::Read, header.version);
-  if (!DoState(sw, update_display))
+  if (!DoState(sw, nullptr, update_display))
     return false;
 
   if (s_state == State::Starting)
@@ -1184,7 +1211,7 @@ bool SaveState(ByteStream* state, u32 screenshot_size /* = 128 */)
     g_gpu->RestoreGraphicsAPIState();
 
     StateWrapper sw(state, StateWrapper::Mode::Write, SAVE_STATE_VERSION);
-    const bool result = DoState(sw, false);
+    const bool result = DoState(sw, nullptr, false);
 
     g_gpu->ResetGraphicsAPIState();
 
@@ -1224,10 +1251,8 @@ void SingleStepCPU()
   g_gpu->ResetGraphicsAPIState();
 }
 
-void RunFrame()
+void DoRunFrame()
 {
-  s_frame_timer.Reset();
-
   g_gpu->RestoreGraphicsAPIState();
 
   if (CPU::g_state.use_debug_dispatcher)
@@ -1266,6 +1291,22 @@ void RunFrame()
   g_gpu->ResetGraphicsAPIState();
 }
 
+void RunFrame()
+{
+  s_frame_timer.Reset();
+
+  if (s_rewind_load_counter >= 0)
+  {
+    DoRewind();
+    return;
+  }
+
+  DoRunFrame();
+
+  if (s_memory_saves_enabled)
+    DoMemorySaveStates();
+}
+
 float GetTargetSpeed()
 {
   return s_target_speed;
@@ -1739,12 +1780,14 @@ bool InsertMedia(const char* path)
   if (g_settings.IsUsingCodeCache())
     CPU::CodeCache::Reinitialize();
 
+  ClearMemorySaveStates();
   return true;
 }
 
 void RemoveMedia()
 {
   g_cdrom.RemoveMedia();
+  ClearMemorySaveStates();
 }
 
 void UpdateRunningGame(const char* path, CDImage* image)
@@ -1907,4 +1950,152 @@ void SetCheatList(std::unique_ptr<CheatList> cheats)
   s_cheat_list = std::move(cheats);
 }
 
+void CalculateRewindMemoryUsage(u32 num_saves, u64* ram_usage, u64* vram_usage)
+{
+  *ram_usage = MAX_SAVE_STATE_SIZE * static_cast<u64>(num_saves);
+  *vram_usage = (VRAM_WIDTH * VRAM_HEIGHT * 4) * static_cast<u64>(std::max(g_settings.gpu_resolution_scale, 1u)) *
+                static_cast<u64>(num_saves);
+}
+
+void ClearMemorySaveStates()
+{
+  s_rewind_states.clear();
+}
+
+void UpdateMemorySaveStateSettings()
+{
+  ClearMemorySaveStates();
+
+  s_memory_saves_enabled = g_settings.rewind_enable;
+
+  if (g_settings.rewind_enable)
+  {
+    s_rewind_save_frequency = static_cast<s32>(std::ceil(g_settings.rewind_save_frequency * s_throttle_frequency));
+    s_rewind_save_counter = 0;
+
+    u64 ram_usage, vram_usage;
+    CalculateRewindMemoryUsage(g_settings.rewind_save_slots, &ram_usage, &vram_usage);
+    Log_InfoPrintf(
+      "Rewind is enabled, saving every %d frames, with %u slots and %" PRIu64 "MB RAM and %" PRIu64 "MB VRAM usage",
+      std::max(s_rewind_save_frequency, 1), g_settings.rewind_save_slots, ram_usage / 1048576, vram_usage / 1048576);
+  }
+  else
+  {
+    s_rewind_save_frequency = -1;
+    s_rewind_save_counter = -1;
+  }
+}
+
+bool SaveRewindState()
+{
+  Common::Timer save_timer;
+
+  const u32 save_slots = g_settings.rewind_save_slots;
+  while (s_rewind_states.size() >= save_slots)
+    s_rewind_states.pop_front();
+
+  MemorySaveState mss;
+  mss.state_stream = std::make_unique<GrowableMemoryByteStream>(nullptr, MAX_SAVE_STATE_SIZE);
+
+  HostDisplayTexture* host_texture = nullptr;
+  StateWrapper sw(mss.state_stream.get(), StateWrapper::Mode::Write, SAVE_STATE_VERSION);
+  if (!DoState(sw, &host_texture, false))
+  {
+    Log_ErrorPrint("Failed to create rewind state.");
+    return false;
+  }
+
+  mss.vram_texture.reset(host_texture);
+  s_rewind_states.push_back(std::move(mss));
+
+  Log_DevPrintf("Saved rewind state (%u bytes, took %.4f ms)", s_rewind_states.back().state_stream->GetSize(),
+                save_timer.GetTimeMilliseconds());
+
+  return true;
+}
+
+bool LoadRewindState(u32 skip_saves /*= 0*/, bool consume_state /*=true */)
+{
+  while (skip_saves > 0 && !s_rewind_states.empty())
+  {
+    s_rewind_states.pop_back();
+    skip_saves--;
+  }
+
+  if (s_rewind_states.empty())
+    return false;
+
+  Common::Timer load_timer;
+
+  const MemorySaveState& mss = s_rewind_states.back();
+  mss.state_stream->SeekAbsolute(0);
+
+  StateWrapper sw(mss.state_stream.get(), StateWrapper::Mode::Read, SAVE_STATE_VERSION);
+  HostDisplayTexture* host_texture = mss.vram_texture.get();
+  if (!DoState(sw, &host_texture, true))
+  {
+    g_host_interface->ReportError("Failed to load rewind state from memory, resetting.");
+    Reset();
+    return false;
+  }
+
+  if (consume_state)
+    s_rewind_states.pop_back();
+
+  Log_DevPrintf("Rewind load took %.4f ms", load_timer.GetTimeMilliseconds());
+  return true;
+}
+
+void SetRewinding(bool enabled)
+{
+  if (enabled)
+  {
+    // Try to rewind at the replay speed, or one per second maximum.
+    const float load_frequency = std::min(g_settings.rewind_save_frequency, 1.0f);
+    s_rewind_load_frequency = static_cast<s32>(std::ceil(load_frequency * s_throttle_frequency));
+    s_rewind_load_counter = 0;
+  }
+  else
+  {
+    s_rewind_load_frequency = -1;
+    s_rewind_load_counter = -1;
+  }
+
+  s_rewinding_first_save = true;
+}
+
+void DoRewind()
+{
+  s_frame_timer.Reset();
+
+  if (s_rewind_load_counter == 0)
+  {
+    const u32 skip_saves = BoolToUInt32(!s_rewinding_first_save);
+    s_rewinding_first_save = false;
+    LoadRewindState(skip_saves, false);
+    ResetPerformanceCounters();
+    s_rewind_load_counter = s_rewind_load_frequency;
+  }
+  else
+  {
+    s_rewind_load_counter--;
+  }
+}
+
+void DoMemorySaveStates()
+{
+  if (s_rewind_save_counter >= 0)
+  {
+    if (s_rewind_save_counter == 0)
+    {
+      SaveRewindState();
+      s_rewind_save_counter = s_rewind_save_frequency;
+    }
+    else
+    {
+      s_rewind_save_counter--;
+    }
+  }
+}
+
 } // namespace System

src/core/system.h

@@ -228,4 +228,14 @@ void ApplyCheatCode(const CheatCode& code);
 /// Sets or clears the provided cheat list, applying every frame.
 void SetCheatList(std::unique_ptr<CheatList> cheats);
 
+//////////////////////////////////////////////////////////////////////////
+// Memory Save States (Rewind and Runahead)
+//////////////////////////////////////////////////////////////////////////
+void CalculateRewindMemoryUsage(u32 num_saves, u64* ram_usage, u64* vram_usage);
+void ClearMemorySaveStates();
+void UpdateMemorySaveStateSettings();
+bool SaveRewindState();
+bool LoadRewindState(u32 skip_saves = 0, bool consume_state = true);
+void SetRewinding(bool enabled);
+
 } // namespace System

src/frontend-common/common_host_interface.cpp

@@ -1549,7 +1549,7 @@ void CommonHostInterface::RegisterGeneralHotkeys()
 
   RegisterHotkey(StaticString(TRANSLATABLE("Hotkeys", "General")), StaticString("ToggleCheats"),
                  StaticString(TRANSLATABLE("Hotkeys", "Toggle Cheats")), [this](bool pressed) {
-                   if (pressed)
+                   if (pressed && System::IsValid())
                      DoToggleCheats();
                  });
 
@@ -1584,7 +1584,7 @@ void CommonHostInterface::RegisterGeneralHotkeys()
 #else
   RegisterHotkey(StaticString(TRANSLATABLE("Hotkeys", "General")), StaticString("TogglePatchCodes"),
                  StaticString(TRANSLATABLE("Hotkeys", "Toggle Patch Codes")), [this](bool pressed) {
-                   if (pressed)
+                   if (pressed && System::IsValid())
                      DoToggleCheats();
                  });
 #endif
@@ -1603,25 +1603,39 @@ void CommonHostInterface::RegisterGeneralHotkeys()
 
   RegisterHotkey(StaticString(TRANSLATABLE("Hotkeys", "General")), StaticString("FrameStep"),
                  StaticString(TRANSLATABLE("Hotkeys", "Frame Step")), [this](bool pressed) {
-                   if (pressed)
+                   if (pressed && System::IsValid())
                      DoFrameStep();
                  });
+
+  RegisterHotkey(StaticString(TRANSLATABLE("Hotkeys", "General")), StaticString("Rewind"),
+                 StaticString(TRANSLATABLE("Hotkeys", "Rewind")), [this](bool pressed) {
+                   if (System::IsValid())
+                   {
+                     AddOSDMessage(pressed ? TranslateStdString("OSDMessage", "Rewinding...") :
+                                             TranslateStdString("OSDMessage", "Stopped rewinding."),
+                                   5.0f);
+                     System::SetRewinding(pressed);
+                   }
+                 });
 }
 
 void CommonHostInterface::RegisterGraphicsHotkeys()
 {
   RegisterHotkey(StaticString(TRANSLATABLE("Hotkeys", "Graphics")), StaticString("ToggleSoftwareRendering"),
                  StaticString(TRANSLATABLE("Hotkeys", "Toggle Software Rendering")), [this](bool pressed) {
-                   if (pressed)
+                   if (pressed && System::IsValid())
                      ToggleSoftwareRendering();
                  });
 
   RegisterHotkey(StaticString(TRANSLATABLE("Hotkeys", "Graphics")), StaticString("TogglePGXP"),
                  StaticString(TRANSLATABLE("Hotkeys", "Toggle PGXP")), [this](bool pressed) {
-                   if (pressed)
+                   if (pressed && System::IsValid())
                    {
                      g_settings.gpu_pgxp_enable = !g_settings.gpu_pgxp_enable;
+                     g_gpu->RestoreGraphicsAPIState();
                      g_gpu->UpdateSettings();
+                     g_gpu->ResetGraphicsAPIState();
+                     System::ClearMemorySaveStates();
                      AddOSDMessage(g_settings.gpu_pgxp_enable ?
                                      TranslateStdString("OSDMessage", "PGXP is now enabled.") :
                                      TranslateStdString("OSDMessage", "PGXP is now disabled"),
@@ -1640,13 +1654,16 @@ void CommonHostInterface::RegisterGraphicsHotkeys()
 
   RegisterHotkey(StaticString(TRANSLATABLE("Hotkeys", "Graphics")), StaticString("TogglePGXPDepth"),
                  StaticString(TRANSLATABLE("Hotkeys", "Toggle PGXP Depth Buffer")), [this](bool pressed) {
-                   if (pressed)
+                   if (pressed && System::IsValid())
                    {
                      g_settings.gpu_pgxp_depth_buffer = !g_settings.gpu_pgxp_depth_buffer;
                      if (!g_settings.gpu_pgxp_enable)
                        return;
 
+                     g_gpu->RestoreGraphicsAPIState();
                      g_gpu->UpdateSettings();
+                     g_gpu->ResetGraphicsAPIState();
+                     System::ClearMemorySaveStates();
                      AddOSDMessage(g_settings.gpu_pgxp_depth_buffer ?
                                      TranslateStdString("OSDMessage", "PGXP Depth Buffer is now enabled.") :
                                      TranslateStdString("OSDMessage", "PGXP Depth Buffer is now disabled."),
@@ -1656,31 +1673,31 @@ void CommonHostInterface::RegisterGraphicsHotkeys()
 
   RegisterHotkey(StaticString(TRANSLATABLE("Hotkeys", "Graphics")), StaticString("IncreaseResolutionScale"),
                  StaticString(TRANSLATABLE("Hotkeys", "Increase Resolution Scale")), [this](bool pressed) {
-                   if (pressed)
+                   if (pressed && System::IsValid())
                      ModifyResolutionScale(1);
                  });
 
   RegisterHotkey(StaticString(TRANSLATABLE("Hotkeys", "Graphics")), StaticString("DecreaseResolutionScale"),
                  StaticString(TRANSLATABLE("Hotkeys", "Decrease Resolution Scale")), [this](bool pressed) {
-                   if (pressed)
+                   if (pressed && System::IsValid())
                      ModifyResolutionScale(-1);
                  });
 
   RegisterHotkey(StaticString(TRANSLATABLE("Hotkeys", "Graphics")), StaticString("TogglePostProcessing"),
                  StaticString(TRANSLATABLE("Hotkeys", "Toggle Post-Processing")), [this](bool pressed) {
-                   if (pressed)
+                   if (pressed && System::IsValid())
                      TogglePostProcessing();
                  });
 
   RegisterHotkey(StaticString(TRANSLATABLE("Hotkeys", "Graphics")), StaticString("ReloadPostProcessingShaders"),
                  StaticString(TRANSLATABLE("Hotkeys", "Reload Post Processing Shaders")), [this](bool pressed) {
-                   if (pressed)
+                   if (pressed && System::IsValid())
                      ReloadPostProcessingShaders();
                  });
 
   RegisterHotkey(StaticString(TRANSLATABLE("Hotkeys", "Graphics")), StaticString("ReloadTextureReplacements"),
                  StaticString(TRANSLATABLE("Hotkeys", "Reload Texture Replacements")), [this](bool pressed) {
-                   if (pressed)
+                   if (pressed && System::IsValid())
                    {
                      AddOSDMessage(TranslateStdString("OSDMessage", "Texture replacements reloaded."), 10.0f);
                      g_texture_replacements.Reload();