GPU/HW: Fix mask bit when rendering with transparency and no DSB

Fixes some sprites in Bloody Roar on Mali GPUs.

src/core/gpu_hw.cpp

@@ -1168,7 +1168,7 @@ void GPU_HW::FlushRender()
     m_batch_ubo_dirty = false;
   }
 
-  if (m_batch.NeedsTwoPassRendering())
+  if (NeedsTwoPassRendering())
   {
     m_renderer_stats.num_batches += 2;
     DrawBatchVertices(BatchRenderMode::OnlyOpaque, m_batch_base_vertex, vertex_count);

src/core/gpu_hw.h

@@ -102,14 +102,6 @@ protected:
     bool check_mask_before_draw;
     bool use_depth_buffer;
 
-    // We need two-pass rendering when using BG-FG blending and texturing, as the transparency can be enabled
-    // on a per-pixel basis, and the opaque pixels shouldn't be blended at all.
-    bool NeedsTwoPassRendering() const
-    {
-      return transparency_mode == GPUTransparencyMode::BackgroundMinusForeground &&
-             texture_mode != GPUTextureMode::Disabled;
-    }
-
     // Returns the render mode for this batch.
     BatchRenderMode GetRenderMode() const
     {
@@ -254,6 +246,15 @@ protected:
     return true;
   }
 
+  /// We need two-pass rendering when using BG-FG blending and texturing, as the transparency can be enabled
+  /// on a per-pixel basis, and the opaque pixels shouldn't be blended at all.
+  bool NeedsTwoPassRendering() const
+  {
+    return (m_batch.transparency_mode == GPUTransparencyMode::BackgroundMinusForeground &&
+            m_batch.texture_mode != GPUTextureMode::Disabled) ||
+           (m_batch.transparency_mode != GPUTransparencyMode::Disabled && !m_supports_dual_source_blend);
+  }
+
   void FillVRAM(u32 x, u32 y, u32 width, u32 height, u32 color) override;
   void UpdateVRAM(u32 x, u32 y, u32 width, u32 height, const void* data, bool set_mask, bool check_mask) override;
   void CopyVRAM(u32 src_x, u32 src_y, u32 dst_x, u32 dst_y, u32 width, u32 height) override;

src/core/gpu_hw_opengl.cpp

@@ -361,8 +361,6 @@ void GPU_HW_OpenGL::SetCapabilities(HostDisplay* host_display)
   m_supports_dual_source_blend =
     (max_dual_source_draw_buffers > 0) &&
     (GLAD_GL_VERSION_3_3 || GLAD_GL_ARB_blend_func_extended || GLAD_GL_EXT_blend_func_extended);
-  if (!m_supports_dual_source_blend)
-    Log_WarningPrintf("Dual-source blending is not supported, this may break some mask effects.");
 
   m_supports_geometry_shaders =
     GLAD_GL_VERSION_3_2 || GLAD_GL_ARB_geometry_shader4 || GLAD_GL_OES_geometry_shader || GLAD_GL_ES_VERSION_3_2;
@@ -754,9 +752,19 @@ void GPU_HW_OpenGL::SetBlendMode()
                               GL_FUNC_REVERSE_SUBTRACT :
                               GL_FUNC_ADD,
                             GL_FUNC_ADD);
+    if (m_supports_dual_source_blend)
+    {
       glBlendFuncSeparate(GL_ONE, m_supports_dual_source_blend ? GL_SRC1_ALPHA : GL_SRC_ALPHA, GL_ONE, GL_ZERO);
     }
     else
+    {
+      const float factor =
+        (m_current_transparency_mode == GPUTransparencyMode::HalfBackgroundPlusHalfForeground) ? 0.5f : 1.0f;
+      glBlendFuncSeparate(GL_ONE, GL_CONSTANT_ALPHA, GL_ONE, GL_ZERO);
+      glBlendColor(0.0f, 0.0f, 0.0f, factor);
+    }
+  }
+  else
   {
     glDisable(GL_BLEND);
   }

src/core/gpu_hw_shadergen.cpp

@@ -939,7 +939,7 @@ float4 SampleFromVRAM(uint4 texpage, float2 coords)
         o_col0 = float4(color, oalpha);
         o_col1 = float4(0.0, 0.0, 0.0, u_dst_alpha_factor / ialpha);
       #else
-        o_col0 = float4(color, u_dst_alpha_factor / ialpha);
+        o_col0 = float4(color, oalpha);
       #endif
 
       #if !PGXP_DEPTH
@@ -952,19 +952,11 @@ float4 SampleFromVRAM(uint4 texpage, float2 coords)
         discard;
       #endif
 
-      #if TRANSPARENCY_ONLY_OPAQUE
-        // We don't output the second color here because it's not used (except for filtering).
-        o_col0 = float4(color, oalpha);
-        #if USE_DUAL_SOURCE
-          o_col1 = float4(0.0, 0.0, 0.0, 1.0 - ialpha);
-        #endif
-      #else
       #if USE_DUAL_SOURCE
         o_col0 = float4(color, oalpha);
         o_col1 = float4(0.0, 0.0, 0.0, 1.0 - ialpha);
       #else
-          o_col0 = float4(color, 1.0 - ialpha);
+        o_col0 = float4(color, oalpha);
-        #endif
       #endif
 
       #if !PGXP_DEPTH

src/core/gpu_hw_vulkan.cpp

@@ -174,7 +174,8 @@ void GPU_HW_Vulkan::ResetGraphicsAPIState()
   EndRenderPass();
 
   if (m_host_display->GetDisplayTextureHandle() == &m_vram_texture)
-    m_vram_texture.TransitionToLayout(g_vulkan_context->GetCurrentCommandBuffer(), VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL);
+    m_vram_texture.TransitionToLayout(g_vulkan_context->GetCurrentCommandBuffer(),
+                                      VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL);
 }
 
 void GPU_HW_Vulkan::RestoreGraphicsAPIState()
@@ -522,9 +523,9 @@ bool GPU_HW_Vulkan::CreateFramebuffer()
       !m_vram_read_texture.Create(texture_width, texture_height, 1, 1, texture_format, VK_SAMPLE_COUNT_1_BIT,
                                   VK_IMAGE_VIEW_TYPE_2D, VK_IMAGE_TILING_OPTIMAL,
                                   VK_IMAGE_USAGE_SAMPLED_BIT | VK_IMAGE_USAGE_TRANSFER_DST_BIT) ||
-      !m_display_texture.Create(GPU_MAX_DISPLAY_WIDTH * m_resolution_scale,
+      !m_display_texture.Create(GPU_MAX_DISPLAY_WIDTH * m_resolution_scale, GPU_MAX_DISPLAY_HEIGHT * m_resolution_scale,
-                                GPU_MAX_DISPLAY_HEIGHT * m_resolution_scale, 1, 1, texture_format,
+                                1, 1, texture_format, VK_SAMPLE_COUNT_1_BIT, VK_IMAGE_VIEW_TYPE_2D,
-                                VK_SAMPLE_COUNT_1_BIT, VK_IMAGE_VIEW_TYPE_2D, VK_IMAGE_TILING_OPTIMAL,
+                                VK_IMAGE_TILING_OPTIMAL,
                                 VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT | VK_IMAGE_USAGE_SAMPLED_BIT |
                                   VK_IMAGE_USAGE_TRANSFER_SRC_BIT | VK_IMAGE_USAGE_TRANSFER_DST_BIT) ||
       !m_vram_readback_texture.Create(VRAM_WIDTH, VRAM_HEIGHT, 1, 1, texture_format, VK_SAMPLE_COUNT_1_BIT,
@@ -923,10 +924,11 @@ bool GPU_HW_Vulkan::CompilePipelines()
                    (static_cast<BatchRenderMode>(render_mode) != BatchRenderMode::TransparencyDisabled &&
                     static_cast<BatchRenderMode>(render_mode) != BatchRenderMode::OnlyOpaque)) ||
                   m_texture_filtering != GPUTextureFilter::Nearest)
+              {
+                if (m_supports_dual_source_blend)
                 {
                   gpbuilder.SetBlendAttachment(
-                  0, true, VK_BLEND_FACTOR_ONE,
+                    0, true, VK_BLEND_FACTOR_ONE, VK_BLEND_FACTOR_SRC1_ALPHA,
-                  m_supports_dual_source_blend ? VK_BLEND_FACTOR_SRC1_ALPHA : VK_BLEND_FACTOR_SRC_ALPHA,
                     (static_cast<GPUTransparencyMode>(transparency_mode) ==
                        GPUTransparencyMode::BackgroundMinusForeground &&
                      static_cast<BatchRenderMode>(render_mode) != BatchRenderMode::TransparencyDisabled &&
@@ -935,6 +937,24 @@ bool GPU_HW_Vulkan::CompilePipelines()
                       VK_BLEND_OP_ADD,
                     VK_BLEND_FACTOR_ONE, VK_BLEND_FACTOR_ZERO, VK_BLEND_OP_ADD);
                 }
+                else
+                {
+                  const float factor = (static_cast<GPUTransparencyMode>(transparency_mode) ==
+                                        GPUTransparencyMode::HalfBackgroundPlusHalfForeground) ?
+                                         0.5f :
+                                         1.0f;
+                  gpbuilder.SetBlendAttachment(
+                    0, true, VK_BLEND_FACTOR_ONE, VK_BLEND_FACTOR_CONSTANT_ALPHA,
+                    (static_cast<GPUTransparencyMode>(transparency_mode) ==
+                       GPUTransparencyMode::BackgroundMinusForeground &&
+                     static_cast<BatchRenderMode>(render_mode) != BatchRenderMode::TransparencyDisabled &&
+                     static_cast<BatchRenderMode>(render_mode) != BatchRenderMode::OnlyOpaque) ?
+                      VK_BLEND_OP_REVERSE_SUBTRACT :
+                      VK_BLEND_OP_ADD,
+                    VK_BLEND_FACTOR_ONE, VK_BLEND_FACTOR_ZERO, VK_BLEND_OP_ADD);
+                  gpbuilder.SetBlendConstants(0.0f, 0.0f, 0.0f, factor);
+                }
+              }
 
               gpbuilder.SetDynamicViewportAndScissorState();
 

src/core/shader_cache_version.h

@@ -1,4 +1,4 @@
 #pragma once
 #include "types.h"
 
-static constexpr u32 SHADER_CACHE_VERSION = 1;
+static constexpr u32 SHADER_CACHE_VERSION = 2;