GPU/HW: Slight re-shuffling of field offsets

Free up some bits in the middle.
stenzek · Jan 7, 2025 · 670dc46 · 670dc46
1 parent 0030bc2
commit 670dc46
Show file tree

Hide file tree

Showing 3 changed files with 19 additions and 25 deletions.
diff --git a/src/core/gpu_backend.h b/src/core/gpu_backend.h
@@ -191,10 +191,11 @@ class GPUBackend
   void DestroyDeinterlaceTextures();
   bool ApplyChromaSmoothing();
 
-  GSVector4i m_clamped_drawing_area = {};
-
   s32 m_display_width = 0;
   s32 m_display_height = 0;
+
+  GSVector4i m_clamped_drawing_area = {};
+
   s32 m_display_origin_left = 0;
   s32 m_display_origin_top = 0;
   s32 m_display_vram_width = 0;

diff --git a/src/core/gpu_hw.cpp b/src/core/gpu_hw.cpp
@@ -3050,10 +3050,9 @@ ALWAYS_INLINE_RELEASE void GPU_HW::CheckForTexPageOverlap(const GPUBackendDrawCo
     if (m_texpage_dirty & TEXPAGE_DIRTY_PAGE_RECT)
     {
       DebugAssert(!(m_texpage_dirty & (TEXPAGE_DIRTY_DRAWN_RECT | TEXPAGE_DIRTY_WRITTEN_RECT)));
-      DebugAssert(m_batch.texture_mode == BatchTextureMode::PageTexture &&
-                  m_batch.texture_cache_key.page < NUM_VRAM_PAGES);
+      DebugAssert(m_batch.texture_mode == BatchTextureMode::PageTexture && m_texture_cache_key.page < NUM_VRAM_PAGES);
 
-      if (GPUTextureCache::AreSourcePagesDrawn(m_batch.texture_cache_key, m_current_uv_rect))
+      if (GPUTextureCache::AreSourcePagesDrawn(m_texture_cache_key, m_current_uv_rect))
       {
         // UVs intersect with drawn area, can't use TC
         if (m_batch_index_count > 0)
@@ -3063,7 +3062,7 @@ ALWAYS_INLINE_RELEASE void GPU_HW::CheckForTexPageOverlap(const GPUBackendDrawCo
         }
 
         // We need to swap the dirty tracking over to drawn/written.
-        const GSVector4i page_rect = GetTextureRect(m_batch.texture_cache_key.page, m_batch.texture_cache_key.mode);
+        const GSVector4i page_rect = GetTextureRect(m_texture_cache_key.page, m_texture_cache_key.mode);
         m_texpage_dirty = (m_vram_dirty_draw_rect.rintersects(page_rect) ? TEXPAGE_DIRTY_DRAWN_RECT : 0) |
                           (m_vram_dirty_write_rect.rintersects(page_rect) ? TEXPAGE_DIRTY_WRITTEN_RECT : 0);
         m_compute_uv_range = (ShouldCheckForTexPageOverlap() || m_clamp_uvs);
@@ -3572,7 +3571,7 @@ void GPU_HW::PrepareDraw(const GPUBackendDrawCommand* cmd)
 {
   // TODO: avoid all this for vertex loading, only do when the type of draw changes
   BatchTextureMode texture_mode = cmd->texture_enable ? m_batch.texture_mode : BatchTextureMode::Disabled;
-  GPUTextureCache::SourceKey texture_cache_key = m_batch.texture_cache_key;
+  GPUTextureCache::SourceKey texture_cache_key = m_texture_cache_key;
   if (cmd->texture_enable)
   {
     // texture page changed - check that the new page doesn't intersect the drawing area
@@ -3680,9 +3679,9 @@ void GPU_HW::PrepareDraw(const GPUBackendDrawCommand* cmd)
   {
     if (texture_mode != m_batch.texture_mode || transparency_mode != m_batch.transparency_mode ||
         (transparency_mode == GPUTransparencyMode::BackgroundMinusForeground && !m_allow_shader_blend) ||
-        dithering_enable != m_batch.dithering || m_batch_ubo_data.u_texture_window_bits != cmd->window ||
+        dithering_enable != m_batch.dithering || m_texture_window_bits != cmd->window ||
         m_batch_ubo_data.u_set_mask_while_drawing != BoolToUInt32(cmd->set_mask_while_drawing) ||
-        (texture_mode == BatchTextureMode::PageTexture && m_batch.texture_cache_key != texture_cache_key))
+        (texture_mode == BatchTextureMode::PageTexture && m_texture_cache_key != texture_cache_key))
     {
       FlushRender();
     }
@@ -3729,13 +3728,13 @@ void GPU_HW::PrepareDraw(const GPUBackendDrawCommand* cmd)
     m_batch.texture_mode = texture_mode;
     m_batch.transparency_mode = transparency_mode;
     m_batch.dithering = dithering_enable;
-    m_batch.texture_cache_key = texture_cache_key;
+    m_texture_cache_key = texture_cache_key;
 
-    if (m_batch_ubo_data.u_texture_window_bits != cmd->window)
+    if (m_texture_window_bits != cmd->window)
     {
-      m_batch_ubo_data.u_texture_window_bits = cmd->window;
+      m_texture_window_bits = cmd->window;
       m_texture_window_active = (cmd->window != GPUTextureWindow{{0xFF, 0xFF, 0x00, 0x00}});
-      GSVector4i::store<true>(&m_batch_ubo_data.u_texture_window[0], GSVector4i::load32(&cmd->window).u8to32());
+      GSVector4i::store<false>(&m_batch_ubo_data.u_texture_window[0], GSVector4i::load32(&cmd->window).u8to32());
       m_batch_ubo_dirty = true;
     }
 
@@ -3779,7 +3778,7 @@ void GPU_HW::FlushRender()
   const GPUTextureCache::Source* texture = nullptr;
   if (m_batch.texture_mode == BatchTextureMode::PageTexture)
   {
-    texture = LookupSource(m_batch.texture_cache_key, m_current_uv_rect,
+    texture = LookupSource(m_texture_cache_key, m_current_uv_rect,
                            m_batch.transparency_mode != GPUTransparencyMode::Disabled ?
                              GPUTextureCache::PaletteRecordFlags::HasSemiTransparentDraws :
                              GPUTextureCache::PaletteRecordFlags::None);

diff --git a/src/core/gpu_hw.h b/src/core/gpu_hw.h
@@ -15,12 +15,6 @@
 #include <tuple>
 #include <utility>
 
-class Error;
-
-class GPU_SW_Backend;
-struct GPUBackendCommand;
-struct GPUBackendDrawCommand;
-
 // TODO: Move to cpp
 // TODO: Rename to GPUHWBackend, preserved to avoid conflicts.
 class GPU_HW final : public GPUBackend
@@ -139,7 +133,7 @@ class GPU_HW final : public GPUBackend
     void SetUVLimits(u32 min_u, u32 max_u, u32 min_v, u32 max_v);
   };
 
-  struct alignas(4) BatchConfig
+  struct BatchConfig
   {
     BatchTextureMode texture_mode = BatchTextureMode::Disabled;
     GPUTransparencyMode transparency_mode = GPUTransparencyMode::Disabled;
@@ -150,13 +144,11 @@ class GPU_HW final : public GPUBackend
     bool use_depth_buffer = false;
     bool sprite_mode = false;
 
-    GPUTextureCache::SourceKey texture_cache_key = {};
-
     // Returns the render mode for this batch.
     BatchRenderMode GetRenderMode() const;
   };
 
-  struct alignas(VECTOR_ALIGNMENT) BatchUBOData
+  struct BatchUBOData
   {
     u32 u_texture_window[4]; // and_x, and_y, or_x, or_y
     float u_src_alpha_factor;
@@ -166,7 +158,6 @@ class GPU_HW final : public GPUBackend
     float u_resolution_scale;
     float u_rcp_resolution_scale;
     float u_resolution_scale_minus_one;
-    GPUTextureWindow u_texture_window_bits; // not actually used on GPU
   };
 
   struct RendererStats
@@ -327,6 +318,7 @@ class GPU_HW final : public GPUBackend
   bool m_batch_ubo_dirty = true;
   bool m_drawing_area_changed = true;
   BatchConfig m_batch;
+  GPUTextureCache::SourceKey m_texture_cache_key = {};
 
   // Changed state
   BatchUBOData m_batch_ubo_data = {};
@@ -350,6 +342,8 @@ class GPU_HW final : public GPUBackend
     u32 bits = INVALID_DRAW_MODE_BITS;
   } m_draw_mode = {};
 
+  GPUTextureWindow m_texture_window_bits;
+
   std::unique_ptr<GPUPipeline> m_wireframe_pipeline;
 
   // [wrapped][interlaced]