Merge 869d316596 into 74a743a5e4

src/core/fullscreenui_settings.cpp

@@ -2795,6 +2795,10 @@ void FullscreenUI::DrawInterfaceSettingsPage()
   DrawToggleSetting(bsi, FSUI_ICONVSTR(ICON_FA_CHART_BAR, "Show GPU Statistics"),
                     FSUI_VSTR("Shows information about the emulated GPU in the top-right corner of the display."),
                     "Display", "ShowGPUStatistics", false);
+  DrawToggleSetting(bsi, FSUI_ICONVSTR(ICON_FA_CHART_BAR, "Show Bandwidth Statistics"),
+                    FSUI_VSTR("Shows VRAM memory bandwidth (writes, readbacks, copies, display) per frame in the "
+                              "top-right corner of the display."),
+                    "Display", "ShowBandwidthStatistics", false);
   DrawToggleSetting(
     bsi, FSUI_ICONVSTR(ICON_FA_USER_CLOCK, "Show Latency Statistics"),
     FSUI_VSTR("Shows information about input and audio latency in the top-right corner of the display."), "Display",

src/core/gpu_backend.cpp

@@ -117,7 +117,8 @@ bool GPUBackend::Initialize(bool clear_vram, Error* error)
 
 bool GPUBackend::UpdateSettings(const GPUSettings& old_settings, Error* error)
 {
-  if (g_gpu_settings.display_show_gpu_stats != old_settings.display_show_gpu_stats)
+  if (g_gpu_settings.display_show_gpu_stats != old_settings.display_show_gpu_stats ||
+      g_gpu_settings.display_show_bandwidth_stats != old_settings.display_show_bandwidth_stats)
     GPUBackend::ResetStatistics();
 
   return true;
@@ -654,6 +655,31 @@ void GPUBackend::GetMemoryStatsString(SmallStringBase& str) const
              s_stats.host_num_copies, s_stats.host_num_uploads);
 }
 
+void GPUBackend::GetBandwidthStatsString(SmallStringBase& str) const
+{
+  // VRAM memory bandwidth in KB per frame, by category. "BW" is the total;
+  // W = CPU->GPU writes, R = GPU->CPU readbacks, C = GPU->GPU copies, D = display extraction.
+  const u32 total_kb = static_cast<u32>((s_stats.vram_bandwidth.total() + (1024 - 1)) / 1024);
+  const u32 write_kb = static_cast<u32>((s_stats.vram_bandwidth.write + (1024 - 1)) / 1024);
+  const u32 read_kb = static_cast<u32>((s_stats.vram_bandwidth.read + (1024 - 1)) / 1024);
+  const u32 copy_kb = static_cast<u32>((s_stats.vram_bandwidth.copy + (1024 - 1)) / 1024);
+  const u32 display_kb = static_cast<u32>((s_stats.vram_bandwidth.display + (1024 - 1)) / 1024);
+
+  str.format("{}KB " BOLD("BW") " | {} " BOLD("W") " | {} " BOLD("R") " | {} " BOLD("C") " | {} " BOLD("D"),
+             total_kb, write_kb, read_kb, copy_kb, display_kb);
+}
+
+void GPUBackend::GetRenderStatsString(SmallStringBase& str) const
+{
+  // Rendering / fill-rate in thousands of shaded pixels per frame. "FP" is the total
+  // shaded-pixel count (bounding-box estimate, includes overdraw); "TX" is the subset
+  // that samples a texture (i.e. the texture-fetch cost).
+  const u32 fill_k = static_cast<u32>((s_stats.render.fill + 1000 - 1) / 1000);
+  const u32 tex_k = static_cast<u32>((s_stats.render.tex + 1000 - 1) / 1000);
+
+  str.format("{}K " BOLD("FP") " | {}K " BOLD("TX"), fill_k, tex_k);
+}
+
 #undef BOLD
 
 void GPUBackend::ResetStatistics()
@@ -688,6 +714,16 @@ void GPUBackend::UpdateStatistics(u32 frame_count)
   UPDATE_GPU_STAT(num_downloads);
   UPDATE_GPU_STAT(num_uploads);
 
+  // VRAM memory bandwidth (per-frame average).
+  s_stats.vram_bandwidth.write = (s_counters.vram_write_bytes + round) / frame_count;
+  s_stats.vram_bandwidth.read = (s_counters.vram_read_bytes + round) / frame_count;
+  s_stats.vram_bandwidth.copy = (s_counters.vram_copy_bytes + round) / frame_count;
+  s_stats.vram_bandwidth.display = (s_counters.vram_display_bytes + round) / frame_count;
+
+  // Rendering / fill-rate (per-frame average).
+  s_stats.render.fill = (s_counters.fill_pixels + round) / frame_count;
+  s_stats.render.tex = (s_counters.tex_pixels + round) / frame_count;
+
 #undef UPDATE_GPU_STAT
 #undef UPDATE_COUNTER
 

src/core/gpu_backend.h

@@ -92,6 +92,8 @@ public:
 
   void GetStatsString(SmallStringBase& str) const;
   void GetMemoryStatsString(SmallStringBase& str) const;
+  void GetBandwidthStatsString(SmallStringBase& str) const;
+  void GetRenderStatsString(SmallStringBase& str) const;
 
   void ResetStatistics();
   void UpdateStatistics(u32 frame_count);
@@ -130,6 +132,20 @@ protected:
     u32 num_vertices;
     u32 num_primitives;
     u32 num_depth_buffer_clears;
+
+    // VRAM memory bandwidth in bytes (hardware backend only). These measure the actual
+    // memory-bus traffic of VRAM operations, which is the bottleneck on low-bandwidth
+    // SoCs (e.g. Allwinner H700 / panfrost).
+    size_t vram_write_bytes;   // CPU->GPU uploads + framebuffer writes (+ reads when checking mask)
+    size_t vram_read_bytes;    // GPU->CPU readbacks
+    size_t vram_copy_bytes;    // GPU->GPU VRAM copies (read + write)
+    size_t vram_display_bytes; // display extraction (read + write)
+
+    // Rendering / fill-rate in shaded pixels (hardware backend only). These measure the
+    // GPU rasterization cost, which (unlike VRAM bandwidth) scales with what is actually
+    // on screen - the likely bottleneck for animated-texture scenes on weak GPUs.
+    size_t fill_pixels; // total shaded pixels (bounding-box estimate, includes overdraw)
+    size_t tex_pixels;  // shaded pixels that sample a texture (texture-fetch cost)
   };
 
   struct Stats : Counters
@@ -143,6 +159,26 @@ protected:
     u32 host_num_uploads;
 
     u8 gpu_busy_pct;
+
+    /// Per-frame average VRAM memory bandwidth in bytes, by category.
+    struct Bandwidth
+    {
+      size_t write;
+      size_t read;
+      size_t copy;
+      size_t display;
+
+      ALWAYS_INLINE size_t total() const { return (write + read + copy + display); }
+    };
+    Bandwidth vram_bandwidth;
+
+    /// Per-frame average rendering / fill-rate in shaded pixels.
+    struct Render
+    {
+      size_t fill;
+      size_t tex;
+    };
+    Render render;
   };
 
   virtual void ReadVRAM(u32 x, u32 y, u32 width, u32 height) = 0;

src/core/gpu_hw.cpp

@@ -3007,6 +3007,8 @@ void GPU_HW::DrawSprite(const GPUBackendDrawRectangleCommand* cmd)
     return;
   }
 
+  AccountRenderPixels(clamped_rect, cmd->texture_enable);
+
   // Treat non-textured sprite draws as fills, so we don't break the TC on framebuffer clears.
   bool draw_with_software_renderer = m_draw_with_software_renderer;
   if (m_use_texture_cache && !cmd->transparency_enable && !cmd->shading_enable && !cmd->texture_enable &&
@@ -3139,6 +3141,11 @@ void GPU_HW::DrawPolygon(const GPUBackendDrawPolygonCommand* cmd)
   {
     SetBatchDepthBuffer(cmd, false);
 
+    if (!clamped_draw_rect_012.rempty())
+      AccountRenderPixels(clamped_draw_rect_012, cmd->texture_enable);
+    if (!clamped_draw_rect_123.rempty())
+      AccountRenderPixels(clamped_draw_rect_123, cmd->texture_enable);
+
     FinishPolygonDraw(cmd, vertices, num_vertices, false, false, clamped_draw_rect_012, clamped_draw_rect_123);
   }
 
@@ -3186,6 +3193,11 @@ void GPU_HW::DrawPrecisePolygon(const GPUBackendDrawPrecisePolygonCommand* cmd)
       CheckForDepthClear(cmd, average_z);
     }
 
+    if (!clamped_draw_rect_012.rempty())
+      AccountRenderPixels(clamped_draw_rect_012, cmd->texture_enable);
+    if (!clamped_draw_rect_123.rempty())
+      AccountRenderPixels(clamped_draw_rect_123, cmd->texture_enable);
+
     FinishPolygonDraw(cmd, vertices, num_vertices, true, is_3d, clamped_draw_rect_012, clamped_draw_rect_123);
   }
 
@@ -3719,6 +3731,9 @@ void GPU_HW::DownloadVRAMFromGPU(u32 x, u32 y, u32 width, u32 height)
                                                  VRAM_WIDTH * sizeof(u16));
   }
 
+  // VRAM bandwidth: framebuffer read (4 bytes/pixel) + CPU readback copy (2 bytes/pixel).
+  GPUBackend::s_counters.vram_read_bytes += static_cast<size_t>(copy_rect.width()) * copy_rect.height() * 6;
+
   RestoreDeviceContext();
 }
 
@@ -3824,6 +3839,11 @@ void GPU_HW::UpdateVRAMOnGPU(u32 x, u32 y, u32 width, u32 height, const void* da
 
   DrawScreenQuad(scaled_bounds, m_vram_texture->GetSizeVec(), GSVector4::zero(), &uniforms, sizeof(uniforms));
 
+  // VRAM bandwidth: upload (2 bytes/pixel) + framebuffer write (4 bytes/pixel) +
+  // framebuffer read for the mask check (4 bytes/pixel).
+  GPUBackend::s_counters.vram_write_bytes +=
+    static_cast<size_t>(width) * height * (2 + 4 + (check_mask ? 4 : 0));
+
   RestoreDeviceContext();
 }
 
@@ -3880,6 +3900,10 @@ void GPU_HW::CopyVRAM(u32 src_x, u32 src_y, u32 dst_x, u32 dst_y, u32 width, u32
     return;
   }
 
+  // VRAM bandwidth: a copy reads and writes the framebuffer (4 bytes/pixel each).
+  // (The recursive chunk path above counts itself; the local-memory TC path is CPU-side.)
+  GPUBackend::s_counters.vram_copy_bytes += static_cast<size_t>(width) * height * 8;
+
   if (use_shader || IsUsingMultisampling())
   {
     if (intersect_with_draw || intersect_with_write)
@@ -4375,6 +4399,11 @@ void GPU_HW::UpdateDisplay(const GPUBackendUpdateDisplayCommand* cmd)
                                       static_cast<float>(line_skip ? 2 : 1)};
     g_gpu_device->DrawWithPushConstants(3, 0, &uniforms, sizeof(uniforms));
 
+    // VRAM bandwidth: the display extraction reads the VRAM framebuffer and writes the
+    // extract texture (4 bytes/pixel each, at the scaled resolution).
+    GPUBackend::s_counters.vram_display_bytes +=
+      static_cast<size_t>(scaled_display_width) * scaled_display_height * 8;
+
     m_vram_extract_texture->MakeReadyForSampling();
     if (depth_source)
     {

src/core/gpu_hw.h

@@ -99,11 +99,20 @@ public:
   void UpdateDisplay(const GPUBackendUpdateDisplayCommand* cmd) override;
 
 private:
+  /// Accounts for the shaded-pixel (fill-rate) cost of a drawn rectangle.
+  ALWAYS_INLINE void AccountRenderPixels(const GSVector4i& rect, bool textured)
+  {
+    const size_t pixels = static_cast<size_t>(rect.width()) * rect.height();
+    s_counters.fill_pixels += pixels;
+    if (textured)
+      s_counters.tex_pixels += pixels;
+  }
+
   enum : u32
   {
     MAX_BATCH_VERTEX_COUNTER_IDS = 65536 - 2,
     MAX_VERTICES_FOR_RECTANGLE = 6 * (((MAX_PRIMITIVE_WIDTH + (TEXTURE_PAGE_WIDTH - 1)) / TEXTURE_PAGE_WIDTH) + 1u) *
-                                 (((MAX_PRIMITIVE_HEIGHT + (TEXTURE_PAGE_HEIGHT - 1)) / TEXTURE_PAGE_HEIGHT) + 1u),
+                                      (((MAX_PRIMITIVE_HEIGHT + (TEXTURE_PAGE_HEIGHT - 1)) / TEXTURE_PAGE_HEIGHT) + 1u),
     NUM_TEXTURE_MODES = static_cast<u32>(BatchTextureMode::MaxCount),
   };
   enum : u8

src/core/imgui_overlays.cpp

@@ -389,9 +389,9 @@ void ImGuiManager::DrawPerformanceOverlay(const GPUBackend* gpu, float& position
 #define COLOR(text) "\x04" text "\x03"
 
   if (!(g_gpu_settings.display_show_fps || g_gpu_settings.display_show_speed || g_gpu_settings.display_show_gpu_stats ||
-        g_gpu_settings.display_show_resolution || g_gpu_settings.display_show_latency_stats ||
-        g_gpu_settings.display_show_cpu_usage || g_gpu_settings.display_show_gpu_usage ||
-        g_gpu_settings.display_show_frame_times ||
+        g_gpu_settings.display_show_bandwidth_stats || g_gpu_settings.display_show_resolution ||
+        g_gpu_settings.display_show_latency_stats || g_gpu_settings.display_show_cpu_usage ||
+        g_gpu_settings.display_show_gpu_usage || g_gpu_settings.display_show_frame_times ||
         (g_gpu_settings.display_show_status_indicators &&
          (VideoThread::IsSystemPaused() || System::IsFastForwardEnabled() || System::IsTurboEnabled()))))
   {
@@ -456,6 +456,17 @@ void ImGuiManager::DrawPerformanceOverlay(const GPUBackend* gpu, float& position
       position_y += spacing;
     }
 
+    if (g_gpu_settings.display_show_bandwidth_stats)
+    {
+      gpu->GetBandwidthStatsString(text);
+      DrawPerformanceStat(dl, position_y, fixed_font, fixed_font_size, FIXED_BOLD_WEIGHT, 0, rbound, text);
+      position_y += spacing;
+
+      gpu->GetRenderStatsString(text);
+      DrawPerformanceStat(dl, position_y, fixed_font, fixed_font_size, FIXED_BOLD_WEIGHT, 0, rbound, text);
+      position_y += spacing;
+    }
+
     if (g_gpu_settings.display_show_resolution)
     {
       const u32 resolution_scale = gpu->GetResolutionScale();

src/core/settings.cpp

@@ -429,6 +429,7 @@ void Settings::Load(const SettingsInterface& si, const SettingsInterface& contro
   display_show_fps = si.GetBoolValue("Display", "ShowFPS", false);
   display_show_speed = si.GetBoolValue("Display", "ShowSpeed", false);
   display_show_gpu_stats = si.GetBoolValue("Display", "ShowGPUStatistics", false);
+  display_show_bandwidth_stats = si.GetBoolValue("Display", "ShowBandwidthStatistics", false);
   display_show_resolution = si.GetBoolValue("Display", "ShowResolution", false);
   display_show_latency_stats = si.GetBoolValue("Display", "ShowLatencyStatistics", false);
   display_show_cpu_usage = si.GetBoolValue("Display", "ShowCPU", false);
@@ -806,6 +807,7 @@ void Settings::Save(SettingsInterface& si, bool for_copy) const
     si.SetBoolValue("Display", "ShowResolution", display_show_resolution);
     si.SetBoolValue("Display", "ShowLatencyStatistics", display_show_latency_stats);
     si.SetBoolValue("Display", "ShowGPUStatistics", display_show_gpu_stats);
+    si.SetBoolValue("Display", "ShowBandwidthStatistics", display_show_bandwidth_stats);
     si.SetBoolValue("Display", "ShowCPU", display_show_cpu_usage);
     si.SetBoolValue("Display", "ShowGPU", display_show_gpu_usage);
     si.SetBoolValue("Display", "ShowFrameTimes", display_show_frame_times);

src/core/settings.h

@@ -111,6 +111,7 @@ struct GPUSettings
   bool display_show_fps : 1 = false;
   bool display_show_speed : 1 = false;
   bool display_show_gpu_stats : 1 = false;
+  bool display_show_bandwidth_stats : 1 = false;
   bool display_show_resolution : 1 = false;
   bool display_show_latency_stats : 1 = false;
   bool display_show_cpu_usage : 1 = false;

src/duckstation-qt/osdsettingswidget.cpp

@@ -72,6 +72,8 @@ OSDSettingsWidget::OSDSettingsWidget(SettingsWindow* dialog, QWidget* parent) :
   SettingWidgetBinder::BindWidgetToBoolSetting(sif, m_ui.showGPU, "Display", "ShowGPU", false);
   SettingWidgetBinder::BindWidgetToBoolSetting(sif, m_ui.showInput, "Display", "ShowInputs", false);
   SettingWidgetBinder::BindWidgetToBoolSetting(sif, m_ui.showGPUStatistics, "Display", "ShowGPUStatistics", false);
+  SettingWidgetBinder::BindWidgetToBoolSetting(sif, m_ui.showBandwidthStatistics, "Display", "ShowBandwidthStatistics",
+                                              false);
   SettingWidgetBinder::BindWidgetToBoolSetting(sif, m_ui.showLatencyStatistics, "Display", "ShowLatencyStatistics",
                                                false);
   SettingWidgetBinder::BindWidgetToBoolSetting(sif, m_ui.showFrameTimes, "Display", "ShowFrameTimes", false);

src/duckstation-qt/osdsettingswidget.ui

@@ -331,6 +331,13 @@
           </property>
          </widget>
         </item>
+        <item row="5" column="0">
+         <widget class="QCheckBox" name="showBandwidthStatistics">
+          <property name="text">
+           <string>Show Bandwidth Statistics</string>
+          </property>
+         </widget>
+        </item>
        </layout>
       </item>
      </layout>