System: Rewrite memory save state handling

Makes it more friendly to GPU thread.

src/core/gpu.cpp

@@ -285,16 +285,8 @@ void GPU::SoftReset()
   UpdateGPUIdle();
 }
 
-bool GPU::DoState(StateWrapper& sw, GPUTexture** host_texture, bool update_display)
+bool GPU::DoState(StateWrapper& sw, bool update_display)
 {
-  FlushRender();
-
-  if (sw.IsReading())
-  {
-    // perform a reset to discard all pending draws/fb state
-    Reset(host_texture == nullptr);
-  }
-
   sw.Do(&m_GPUSTAT.bits);
 
   sw.Do(&m_draw_mode.mode_reg.bits);
@@ -390,32 +382,74 @@ bool GPU::DoState(StateWrapper& sw, GPUTexture** host_texture, bool update_displ
   sw.Do(&m_max_run_ahead);
   sw.Do(&m_fifo_size);
 
+  if (!sw.DoMarker("GPU-VRAM"))
+    return false;
+
+  sw.DoBytes(g_vram, VRAM_WIDTH * VRAM_HEIGHT * sizeof(u16));
+
   if (sw.IsReading())
   {
     m_draw_mode.texture_page_changed = true;
     m_drawing_area_changed = true;
     SetClampedDrawingArea();
     UpdateDMARequest();
+    UpdateCRTCConfig();
+    UpdateCommandTickEvent();
+
+    // If we're paused, need to update the display FB.
+    if (update_display)
+      UpdateDisplay();
   }
 
-  if (!host_texture)
-  {
-    if (!sw.DoMarker("GPU-VRAM"))
-      return false;
+  return !sw.HasError();
+}
 
-    sw.DoBytes(g_vram, VRAM_WIDTH * VRAM_HEIGHT * sizeof(u16));
-  }
+bool GPU::DoMemoryState(StateWrapper& sw, System::MemorySaveState& mss, bool update_display)
+{
+  sw.Do(&m_GPUSTAT.bits);
+
+  sw.DoBytes(&m_draw_mode, sizeof(m_draw_mode));
+  sw.DoBytes(&m_drawing_area, sizeof(m_drawing_area));
+  sw.DoBytes(&m_drawing_offset, sizeof(m_drawing_offset));
+
+  sw.Do(&m_console_is_pal);
+  sw.Do(&m_set_texture_disable_mask);
+
+  sw.DoBytes(&m_crtc_state, sizeof(m_crtc_state));
+
+  sw.Do(&m_blitter_state);
+  sw.Do(&m_pending_command_ticks);
+  sw.Do(&m_command_total_words);
+  sw.Do(&m_GPUREAD_latch);
+
+  sw.Do(&m_current_clut_reg_bits);
+  sw.Do(&m_current_clut_is_8bit);
+  sw.DoBytes(g_gpu_clut, sizeof(g_gpu_clut));
+
+  sw.DoBytes(&m_vram_transfer, sizeof(m_vram_transfer));
+
+  sw.Do(&m_fifo);
+  sw.Do(&m_blit_buffer);
+  sw.Do(&m_blit_remaining_words);
+  sw.Do(&m_render_command.bits);
+
+  sw.Do(&m_max_run_ahead);
+  sw.Do(&m_fifo_size);
 
   if (sw.IsReading())
   {
+    m_draw_mode.texture_page_changed = true;
+    m_drawing_area_changed = true;
+    SetClampedDrawingArea();
+    UpdateDMARequest();
     UpdateCRTCConfig();
+    UpdateCommandTickEvent();
+
     if (update_display)
       UpdateDisplay();
-
-    UpdateCommandTickEvent();
   }
 
-  return !sw.HasError();
+  return true;
 }
 
 void GPU::RestoreDeviceContext()

src/core/gpu.h

@@ -45,6 +45,10 @@ namespace Threading {
 class Thread;
 }
 
+namespace System {
+struct MemorySaveState;
+}
+
 class GPU
 {
 public:
@@ -95,7 +99,8 @@ class GPU
 
   virtual bool Initialize(Error* error);
   virtual void Reset(bool clear_vram);
-  virtual bool DoState(StateWrapper& sw, GPUTexture** save_to_texture, bool update_display);
+  virtual bool DoState(StateWrapper& sw, bool update_display);
+  virtual bool DoMemoryState(StateWrapper& sw, System::MemorySaveState& mss, bool update_display);
 
   // Graphics API state reset/restore - call when drawing the UI etc.
   // TODO: replace with "invalidate cached state"

src/core/gpu_hw.cpp

@@ -9,7 +9,7 @@
 #include "gpu_sw_rasterizer.h"
 #include "host.h"
 #include "settings.h"
-#include "system.h"
+#include "system_private.h"
 
 #include "util/imgui_manager.h"
 #include "util/postprocessing.h"
@@ -315,14 +315,16 @@ void GPU_HW::Reset(bool clear_vram)
     ClearFramebuffer();
 }
 
-bool GPU_HW::DoState(StateWrapper& sw, GPUTexture** host_texture, bool update_display)
+bool GPU_HW::DoState(StateWrapper& sw, bool update_display)
 {
+  FlushRender();
+
   // Need to download local VRAM copy before calling the base class, because it serializes this.
   if (m_sw_renderer)
   {
     m_sw_renderer->Sync(true);
   }
-  else if (sw.IsWriting() && !host_texture)
+  else if (sw.IsWriting())
   {
     // If SW renderer readbacks aren't enabled, the CLUT won't be populated, which means it'll be invalid if the user
     // loads this state with software instead of hardware renderers. So force-update the CLUT.
@@ -331,65 +333,107 @@ bool GPU_HW::DoState(StateWrapper& sw, GPUTexture** host_texture, bool update_di
       GPU::ReadCLUT(g_gpu_clut, GPUTexturePaletteReg{Truncate16(m_current_clut_reg_bits)}, m_current_clut_is_8bit);
   }
 
-  if (!GPU::DoState(sw, host_texture, update_display))
+  if (!GPU::DoState(sw, false))
     return false;
 
-  if (host_texture)
+  if (sw.IsReading())
   {
-    GPUTexture* tex = *host_texture;
-    if (sw.IsReading())
-    {
-      if (tex->GetWidth() != m_vram_texture->GetWidth() || tex->GetHeight() != m_vram_texture->GetHeight() ||
-          tex->GetSamples() != m_vram_texture->GetSamples())
-      {
-        return false;
-      }
-
-      g_gpu_device->CopyTextureRegion(m_vram_texture.get(), 0, 0, 0, 0, tex, 0, 0, 0, 0, tex->GetWidth(),
-                                      tex->GetHeight());
-    }
-    else
-    {
-      if (!tex || tex->GetWidth() != m_vram_texture->GetWidth() || tex->GetHeight() != m_vram_texture->GetHeight() ||
-          tex->GetSamples() != m_vram_texture->GetSamples())
-      {
-        delete tex;
-
-        // We copy to/from the save state texture, but we can't have multisampled non-RTs.
-        tex = g_gpu_device
-                ->FetchTexture(
-                  m_vram_texture->GetWidth(), m_vram_texture->GetHeight(), 1, 1, m_vram_texture->GetSamples(),
-                  m_vram_texture->IsMultisampled() ? GPUTexture::Type::RenderTarget : GPUTexture::Type::Texture,
-                  GPUTexture::Format::RGBA8, GPUTexture::Flags::None)
-                .release();
-        *host_texture = tex;
-        if (!tex)
-          return false;
-      }
+    // Wipe out state.
+    m_batch = {};
+    m_current_depth = 1;
+    SetClampedDrawingArea();
 
-      g_gpu_device->CopyTextureRegion(tex, 0, 0, 0, 0, m_vram_texture.get(), 0, 0, 0, 0, tex->GetWidth(),
-                                      tex->GetHeight());
-    }
-  }
-  else if (sw.IsReading())
-  {
     // Need to update the VRAM copy on the GPU with the state data.
     // Would invalidate the TC, but base DoState() calls Reset().
     UpdateVRAMOnGPU(0, 0, VRAM_WIDTH, VRAM_HEIGHT, g_vram, VRAM_WIDTH * sizeof(u16), false, false, VRAM_SIZE_RECT);
+
+    // invalidate the whole VRAM read texture when loading state
+    DebugAssert(!m_batch_vertex_ptr && !m_batch_index_ptr);
+    ClearVRAMDirtyRectangle();
+    SetFullVRAMDirtyRectangle();
+    UpdateVRAMReadTexture(true, false);
+    ClearVRAMDirtyRectangle();
+    ResetBatchVertexDepth();
+
+    // refresh display, has to be done here because of the upload above
+    if (update_display)
+      UpdateDisplay();
   }
 
-  // invalidate the whole VRAM read texture when loading state
+  return GPUTextureCache::DoState(sw, !m_use_texture_cache);
+}
+
+bool GPU_HW::DoMemoryState(StateWrapper& sw, System::MemorySaveState& mss, bool update_display)
+{
+  // sw-for-readbacks just makes a mess here
+  if (m_sw_renderer)
+    m_sw_renderer->Sync(true);
+  if (m_sw_renderer || m_use_texture_cache)
+    sw.DoBytes(g_vram, VRAM_WIDTH * VRAM_HEIGHT * sizeof(u16));
+
+  // This could be faster too.
+  if (m_use_texture_cache)
+    GPUTextureCache::DoState(sw, m_use_texture_cache);
+
+  // Base class never fails.
+  GPU::DoMemoryState(sw, mss, false);
+
   if (sw.IsReading())
   {
+    if (m_batch_vertex_ptr)
+      UnmapGPUBuffer(0, 0);
+
+    DebugAssert(mss.vram_texture->GetWidth() == m_vram_texture->GetWidth() &&
+                mss.vram_texture->GetHeight() == m_vram_texture->GetHeight() &&
+                mss.vram_texture->GetSamples() == m_vram_texture->GetSamples());
+    g_gpu_device->CopyTextureRegion(m_vram_texture.get(), 0, 0, 0, 0, mss.vram_texture.get(), 0, 0, 0, 0,
+                                    m_vram_texture->GetWidth(), m_vram_texture->GetHeight());
+
+    // Wipe out state.
     DebugAssert(!m_batch_vertex_ptr && !m_batch_index_ptr);
+    m_batch = {};
+    SetClampedDrawingArea();
     ClearVRAMDirtyRectangle();
     SetFullVRAMDirtyRectangle();
     UpdateVRAMReadTexture(true, false);
     ClearVRAMDirtyRectangle();
     ResetBatchVertexDepth();
+
+    if (update_display)
+      UpdateDisplay();
+  }
+  else
+  {
+    FlushRender();
+
+    // saving state
+    if (!mss.vram_texture || mss.vram_texture->GetWidth() != m_vram_texture->GetWidth() ||
+        mss.vram_texture->GetHeight() != m_vram_texture->GetHeight() ||
+        mss.vram_texture->GetSamples() != m_vram_texture->GetSamples()) [[unlikely]]
+    {
+      g_gpu_device->RecycleTexture(std::move(mss.vram_texture));
+      mss.vram_texture.reset();
+    }
+    if (!mss.vram_texture)
+    {
+      // We copy to/from the save state texture, but we can't have multisampled non-RTs.
+      Error error;
+      mss.vram_texture = g_gpu_device->FetchTexture(
+        m_vram_texture->GetWidth(), m_vram_texture->GetHeight(), 1, 1, m_vram_texture->GetSamples(),
+        m_vram_texture->IsMultisampled() ? GPUTexture::Type::RenderTarget : GPUTexture::Type::Texture,
+        GPUTexture::Format::RGBA8, GPUTexture::Flags::None);
+      if (!mss.vram_texture) [[unlikely]]
+      {
+        ERROR_LOG("Failed to allocate VRAM texture for memory save state: {}", error.GetDescription());
+        return false;
+      }
+    }
+
+    g_gpu_device->CopyTextureRegion(mss.vram_texture.get(), 0, 0, 0, 0, m_vram_texture.get(), 0, 0, 0, 0,
+                                    m_vram_texture->GetWidth(), m_vram_texture->GetHeight());
   }
 
-  return GPUTextureCache::DoState(sw, !m_use_texture_cache);
+  return true;
 }
 
 void GPU_HW::RestoreDeviceContext()

src/core/gpu_hw.h

@@ -68,7 +68,8 @@ class GPU_HW final : public GPU
 
   bool Initialize(Error* error) override;
   void Reset(bool clear_vram) override;
-  bool DoState(StateWrapper& sw, GPUTexture** host_texture, bool update_display) override;
+  bool DoState(StateWrapper& sw, bool update_display) override;
+  bool DoMemoryState(StateWrapper& sw, System::MemorySaveState& mss, bool update_display) override;
 
   void RestoreDeviceContext() override;
 

src/core/gpu_sw.cpp

@@ -7,6 +7,7 @@
 #include "system.h"
 
 #include "util/gpu_device.h"
+#include "util/state_wrapper.h"
 
 #include "common/align.h"
 #include "common/assert.h"
@@ -58,19 +59,26 @@ bool GPU_SW::Initialize(Error* error)
   return true;
 }
 
-bool GPU_SW::DoState(StateWrapper& sw, GPUTexture** host_texture, bool update_display)
+bool GPU_SW::DoState(StateWrapper& sw, bool update_display)
 {
   // need to ensure the worker thread is done
   m_backend.Sync(true);
 
   // ignore the host texture for software mode, since we want to save vram here
-  if (!GPU::DoState(sw, nullptr, update_display))
+  if (!GPU::DoState(sw, update_display))
     return false;
 
   // need to still call the TC, to toss any data in the state
   return GPUTextureCache::DoState(sw, true);
 }
 
+bool GPU_SW::DoMemoryState(StateWrapper& sw, System::MemorySaveState& mss, bool update_display)
+{
+  m_backend.Sync(true);
+  sw.DoBytes(g_vram, VRAM_WIDTH * VRAM_HEIGHT * sizeof(u16));
+  return GPU::DoMemoryState(sw, mss, update_display);
+}
+
 void GPU_SW::Reset(bool clear_vram)
 {
   GPU::Reset(clear_vram);

src/core/gpu_sw.h

@@ -30,7 +30,8 @@ class GPU_SW final : public GPU
   bool IsHardwareRenderer() const override;
 
   bool Initialize(Error* error) override;
-  bool DoState(StateWrapper& sw, GPUTexture** host_texture, bool update_display) override;
+  bool DoState(StateWrapper& sw, bool update_display) override;
+  bool DoMemoryState(StateWrapper& sw, System::MemorySaveState& mss, bool update_display) override;
   void Reset(bool clear_vram) override;
   void UpdateSettings(const Settings& old_settings) override;