Merge pull request #6141 from JonnyH/WIP/shader-framebuffer-fetch2

Implement dual-source blending in shader

Source/Core/VideoBackends/D3D/main.cpp

@@ -82,6 +82,7 @@ void VideoBackend::InitBackendInfo()
   g_Config.backend_info.bSupportsBitfield = false;
   g_Config.backend_info.bSupportsDynamicSamplerIndexing = false;
   g_Config.backend_info.bSupportsBPTCTextures = false;
+  g_Config.backend_info.bSupportsFramebufferFetch = false;
 
   IDXGIFactory2* factory;
   IDXGIAdapter* ad;

Source/Core/VideoBackends/Null/NullBackend.cpp

@@ -46,6 +46,7 @@ void VideoBackend::InitBackendInfo()
   g_Config.backend_info.bSupportsGPUTextureDecoding = false;
   g_Config.backend_info.bSupportsST3CTextures = false;
   g_Config.backend_info.bSupportsBPTCTextures = false;
+  g_Config.backend_info.bSupportsFramebufferFetch = false;
 
   // aamodes: We only support 1 sample, so no MSAA
   g_Config.backend_info.Adapters.clear();

Source/Core/VideoBackends/OGL/ProgramShaderCache.cpp

@@ -64,26 +64,26 @@ static std::string s_glsl_header = "";
 
 static std::string GetGLSLVersionString()
 {
-  GLSL_VERSION v = g_ogl_config.eSupportedGLSLVersion;
+  GlslVersion v = g_ogl_config.eSupportedGLSLVersion;
   switch (v)
   {
-  case GLSLES_300:
+  case GlslEs300:
     return "#version 300 es";
-  case GLSLES_310:
+  case GlslEs310:
     return "#version 310 es";
-  case GLSLES_320:
+  case GlslEs320:
     return "#version 320 es";
-  case GLSL_130:
+  case Glsl130:
     return "#version 130";
-  case GLSL_140:
+  case Glsl140:
     return "#version 140";
-  case GLSL_150:
+  case Glsl150:
     return "#version 150";
-  case GLSL_330:
+  case Glsl330:
     return "#version 330";
-  case GLSL_400:
+  case Glsl400:
     return "#version 400";
-  case GLSL_430:
+  case Glsl430:
     return "#version 430";
   default:
     // Shouldn't ever hit this
@@ -431,7 +431,7 @@ bool ProgramShaderCache::CompileComputeShader(SHADER& shader, const std::string&
   // but not GLSL 4.3. Mesa is one example.
   std::string header;
   if (g_ActiveConfig.backend_info.bSupportsComputeShaders &&
-      g_ogl_config.eSupportedGLSLVersion < GLSL_430)
+      g_ogl_config.eSupportedGLSLVersion < Glsl430)
   {
     header = "#extension GL_ARB_compute_shader : enable\n";
   }
@@ -839,8 +839,8 @@ void ProgramShaderCache::DestroyShaders()
 
 void ProgramShaderCache::CreateHeader()
 {
-  GLSL_VERSION v = g_ogl_config.eSupportedGLSLVersion;
-  bool is_glsles = v >= GLSLES_300;
+  GlslVersion v = g_ogl_config.eSupportedGLSLVersion;
+  bool is_glsles = v >= GlslEs300;
   std::string SupportedESPointSize;
   std::string SupportedESTextureBuffer;
   switch (g_ogl_config.SupportedESPointSize)
@@ -858,14 +858,14 @@ void ProgramShaderCache::CreateHeader()
 
   switch (g_ogl_config.SupportedESTextureBuffer)
   {
-  case ES_TEXBUF_TYPE::TEXBUF_EXT:
+  case EsTexbufType::TexbufExt:
     SupportedESTextureBuffer = "#extension GL_EXT_texture_buffer : enable";
     break;
-  case ES_TEXBUF_TYPE::TEXBUF_OES:
+  case EsTexbufType::TexbufOes:
     SupportedESTextureBuffer = "#extension GL_OES_texture_buffer : enable";
     break;
-  case ES_TEXBUF_TYPE::TEXBUF_CORE:
-  case ES_TEXBUF_TYPE::TEXBUF_NONE:
+  case EsTexbufType::TexbufCore:
+  case EsTexbufType::TexbufNone:
     SupportedESTextureBuffer = "";
     break;
   }
@@ -885,6 +885,24 @@ void ProgramShaderCache::CreateHeader()
     }
   }
 
+  std::string framebuffer_fetch_string;
+  switch (g_ogl_config.SupportedFramebufferFetch)
+  {
+  case EsFbFetchType::FbFetchExt:
+    framebuffer_fetch_string = "#extension GL_EXT_shader_framebuffer_fetch: enable\n"
+                               "#define FB_FETCH_VALUE real_ocol0\n"
+                               "#define FRAGMENT_INOUT inout";
+    break;
+  case EsFbFetchType::FbFetchArm:
+    framebuffer_fetch_string = "#extension GL_ARM_shader_framebuffer_fetch: enable\n"
+                               "#define FB_FETCH_VALUE gl_LastFragColorARM\n"
+                               "#define FRAGMENT_INOUT out";
+    break;
+  case EsFbFetchType::FbFetchNone:
+    framebuffer_fetch_string = "";
+    break;
+  }
+
   s_glsl_header = StringFromFormat(
       "%s\n"
       "%s\n"  // ubo
@@ -902,6 +920,7 @@ void ProgramShaderCache::CreateHeader()
       "%s\n"  // ES texture buffer
       "%s\n"  // ES dual source blend
       "%s\n"  // shader image load store
+      "%s\n"  // shader framebuffer fetch
 
       // Precision defines for GLSL ES
       "%s\n"
@@ -928,11 +947,11 @@ void ProgramShaderCache::CreateHeader()
 
       ,
       GetGLSLVersionString().c_str(),
-      v < GLSL_140 ? "#extension GL_ARB_uniform_buffer_object : enable" : "", earlyz_string.c_str(),
-      (g_ActiveConfig.backend_info.bSupportsBindingLayout && v < GLSLES_310) ?
+      v < Glsl140 ? "#extension GL_ARB_uniform_buffer_object : enable" : "", earlyz_string.c_str(),
+      (g_ActiveConfig.backend_info.bSupportsBindingLayout && v < GlslEs310) ?
           "#extension GL_ARB_shading_language_420pack : enable" :
           "",
-      (g_ogl_config.bSupportsMSAA && v < GLSL_150) ?
+      (g_ogl_config.bSupportsMSAA && v < Glsl150) ?
           "#extension GL_ARB_texture_multisample : enable" :
           "",
       // Attribute and fragment output bindings are still done via glBindAttribLocation and
@@ -955,15 +974,15 @@ void ProgramShaderCache::CreateHeader()
       !is_glsles && g_ActiveConfig.backend_info.bSupportsFragmentStoresAndAtomics ?
           "#extension GL_ARB_shader_storage_buffer_object : enable" :
           "",
-      v < GLSL_400 && g_ActiveConfig.backend_info.bSupportsGSInstancing ?
+      v < Glsl400 && g_ActiveConfig.backend_info.bSupportsGSInstancing ?
           "#extension GL_ARB_gpu_shader5 : enable" :
           "",
-      v < GLSL_400 && g_ActiveConfig.backend_info.bSupportsSSAA ?
+      v < Glsl400 && g_ActiveConfig.backend_info.bSupportsSSAA ?
           "#extension GL_ARB_sample_shading : enable" :
           "",
       SupportedESPointSize.c_str(),
       g_ogl_config.bSupportsAEP ? "#extension GL_ANDROID_extension_pack_es31a : enable" : "",
-      v < GLSL_140 && g_ActiveConfig.backend_info.bSupportsPaletteConversion ?
+      v < Glsl140 && g_ActiveConfig.backend_info.bSupportsPaletteConversion ?
           "#extension GL_ARB_texture_buffer_object : enable" :
           "",
       SupportedESTextureBuffer.c_str(),
@@ -973,16 +992,16 @@ void ProgramShaderCache::CreateHeader()
 
       ,
       g_ogl_config.bSupportsImageLoadStore &&
-              ((!is_glsles && v < GLSL_430) || (is_glsles && v < GLSLES_310)) ?
+              ((!is_glsles && v < Glsl430) || (is_glsles && v < GlslEs310)) ?
           "#extension GL_ARB_shader_image_load_store : enable" :
           "",
-      is_glsles ? "precision highp float;" : "", is_glsles ? "precision highp int;" : "",
-      is_glsles ? "precision highp sampler2DArray;" : "",
+      framebuffer_fetch_string.c_str(), is_glsles ? "precision highp float;" : "",
+      is_glsles ? "precision highp int;" : "", is_glsles ? "precision highp sampler2DArray;" : "",
       (is_glsles && g_ActiveConfig.backend_info.bSupportsPaletteConversion) ?
           "precision highp usamplerBuffer;" :
           "",
-      v > GLSLES_300 ? "precision highp sampler2DMS;" : "",
-      v >= GLSLES_310 ? "precision highp image2DArray;" : "");
+      v > GlslEs300 ? "precision highp sampler2DMS;" : "",
+      v >= GlslEs310 ? "precision highp image2DArray;" : "");
 }
 
 void ProgramShaderCache::PrecompileUberShaders()

Source/Core/VideoBackends/OGL/Render.cpp

@@ -508,12 +508,12 @@ Renderer::Renderer()
             1 :
             GLExtensions::Supports("GL_EXT_geometry_point_size") ? 2 : 0;
     g_ogl_config.SupportedESTextureBuffer = GLExtensions::Supports("VERSION_GLES_3_2") ?
-                                                ES_TEXBUF_TYPE::TEXBUF_CORE :
+                                                EsTexbufType::TexbufCore :
                                                 GLExtensions::Supports("GL_OES_texture_buffer") ?
-                                                ES_TEXBUF_TYPE::TEXBUF_OES :
+                                                EsTexbufType::TexbufOes :
                                                 GLExtensions::Supports("GL_EXT_texture_buffer") ?
-                                                ES_TEXBUF_TYPE::TEXBUF_EXT :
-                                                ES_TEXBUF_TYPE::TEXBUF_NONE;
+                                                EsTexbufType::TexbufExt :
+                                                EsTexbufType::TexbufNone;
 
     g_ogl_config.bSupportsGLSLCache = true;
     g_ogl_config.bSupportsGLSync = true;
@@ -522,16 +522,31 @@ Renderer::Renderer()
     // depth clamping.
     g_Config.backend_info.bSupportsDepthClamp = false;
 
+    if (GLExtensions::Supports("GL_EXT_shader_framebuffer_fetch"))
+    {
+      g_ogl_config.SupportedFramebufferFetch = EsFbFetchType::FbFetchExt;
+    }
+    else if (GLExtensions::Supports("GL_ARM_shader_framebuffer_fetch"))
+    {
+      g_ogl_config.SupportedFramebufferFetch = EsFbFetchType::FbFetchArm;
+    }
+    else
+    {
+      g_ogl_config.SupportedFramebufferFetch = EsFbFetchType::FbFetchNone;
+    }
+    g_Config.backend_info.bSupportsFramebufferFetch =
+        g_ogl_config.SupportedFramebufferFetch != EsFbFetchType::FbFetchNone;
+
     if (GLExtensions::Version() == 300)
     {
-      g_ogl_config.eSupportedGLSLVersion = GLSLES_300;
+      g_ogl_config.eSupportedGLSLVersion = GlslEs300;
       g_ogl_config.bSupportsAEP = false;
       g_ogl_config.bSupportsTextureStorage = true;
       g_Config.backend_info.bSupportsGeometryShaders = false;
     }
     else if (GLExtensions::Version() == 310)
     {
-      g_ogl_config.eSupportedGLSLVersion = GLSLES_310;
+      g_ogl_config.eSupportedGLSLVersion = GlslEs310;
       g_ogl_config.bSupportsAEP = GLExtensions::Supports("GL_ANDROID_extension_pack_es31a");
       g_Config.backend_info.bSupportsBindingLayout = true;
       g_ogl_config.bSupportsImageLoadStore = true;
@@ -556,7 +571,7 @@ Renderer::Renderer()
     }
     else
     {
-      g_ogl_config.eSupportedGLSLVersion = GLSLES_320;
+      g_ogl_config.eSupportedGLSLVersion = GlslEs320;
       g_ogl_config.bSupportsAEP = GLExtensions::Supports("GL_ANDROID_extension_pack_es31a");
       g_Config.backend_info.bSupportsBindingLayout = true;
       g_ogl_config.bSupportsImageLoadStore = true;
@@ -589,7 +604,7 @@ Renderer::Renderer()
     }
     else if (GLExtensions::Version() == 300)
     {
-      g_ogl_config.eSupportedGLSLVersion = GLSL_130;
+      g_ogl_config.eSupportedGLSLVersion = Glsl130;
       g_ogl_config.bSupportsImageLoadStore = false;  // layout keyword is only supported on glsl150+
       g_ogl_config.bSupportsConservativeDepth =
           false;  // layout keyword is only supported on glsl150+
@@ -598,7 +613,7 @@ Renderer::Renderer()
     }
     else if (GLExtensions::Version() == 310)
     {
-      g_ogl_config.eSupportedGLSLVersion = GLSL_140;
+      g_ogl_config.eSupportedGLSLVersion = Glsl140;
       g_ogl_config.bSupportsImageLoadStore = false;  // layout keyword is only supported on glsl150+
       g_ogl_config.bSupportsConservativeDepth =
           false;  // layout keyword is only supported on glsl150+
@@ -607,16 +622,16 @@ Renderer::Renderer()
     }
     else if (GLExtensions::Version() == 320)
     {
-      g_ogl_config.eSupportedGLSLVersion = GLSL_150;
+      g_ogl_config.eSupportedGLSLVersion = Glsl150;
     }
     else if (GLExtensions::Version() == 330)
     {
-      g_ogl_config.eSupportedGLSLVersion = GLSL_330;
+      g_ogl_config.eSupportedGLSLVersion = Glsl330;
     }
     else if (GLExtensions::Version() >= 430)
     {
       // TODO: We should really parse the GL_SHADING_LANGUAGE_VERSION token.
-      g_ogl_config.eSupportedGLSLVersion = GLSL_430;
+      g_ogl_config.eSupportedGLSLVersion = Glsl430;
       g_ogl_config.bSupportsTextureStorage = true;
       g_ogl_config.bSupportsImageLoadStore = true;
       g_Config.backend_info.bSupportsSSAA = true;
@@ -628,7 +643,7 @@ Renderer::Renderer()
     }
     else
     {
-      g_ogl_config.eSupportedGLSLVersion = GLSL_400;
+      g_ogl_config.eSupportedGLSLVersion = Glsl400;
       g_Config.backend_info.bSupportsSSAA = true;
 
       if (GLExtensions::Version() == 420)
@@ -1275,44 +1290,54 @@ void Renderer::SetBlendingState(const BlendingState& state)
   bool useDualSource =
       state.usedualsrc && g_ActiveConfig.backend_info.bSupportsDualSourceBlend &&
       (!DriverDetails::HasBug(DriverDetails::BUG_BROKEN_DUAL_SOURCE_BLENDING) || state.dstalpha);
+  // Only use shader blend if we need to and we don't support dual-source blending directly
+  bool useShaderBlend = !useDualSource && state.usedualsrc && state.dstalpha &&
+                        g_ActiveConfig.backend_info.bSupportsFramebufferFetch;
 
-  const GLenum src_factors[8] = {
-      GL_ZERO,
-      GL_ONE,
-      GL_DST_COLOR,
-      GL_ONE_MINUS_DST_COLOR,
-      useDualSource ? GL_SRC1_ALPHA : (GLenum)GL_SRC_ALPHA,
-      useDualSource ? GL_ONE_MINUS_SRC1_ALPHA : (GLenum)GL_ONE_MINUS_SRC_ALPHA,
-      GL_DST_ALPHA,
-      GL_ONE_MINUS_DST_ALPHA};
-  const GLenum dst_factors[8] = {
-      GL_ZERO,
-      GL_ONE,
-      GL_SRC_COLOR,
-      GL_ONE_MINUS_SRC_COLOR,
-      useDualSource ? GL_SRC1_ALPHA : (GLenum)GL_SRC_ALPHA,
-      useDualSource ? GL_ONE_MINUS_SRC1_ALPHA : (GLenum)GL_ONE_MINUS_SRC_ALPHA,
-      GL_DST_ALPHA,
-      GL_ONE_MINUS_DST_ALPHA};
-
-  if (state.blendenable)
+  if (useShaderBlend)
   {
-    glEnable(GL_BLEND);
+    glDisable(GL_BLEND);
   }
   else
   {
-    glDisable(GL_BLEND);
-  }
+    const GLenum src_factors[8] = {
+        GL_ZERO,
+        GL_ONE,
+        GL_DST_COLOR,
+        GL_ONE_MINUS_DST_COLOR,
+        useDualSource ? GL_SRC1_ALPHA : (GLenum)GL_SRC_ALPHA,
+        useDualSource ? GL_ONE_MINUS_SRC1_ALPHA : (GLenum)GL_ONE_MINUS_SRC_ALPHA,
+        GL_DST_ALPHA,
+        GL_ONE_MINUS_DST_ALPHA};
+    const GLenum dst_factors[8] = {
+        GL_ZERO,
+        GL_ONE,
+        GL_SRC_COLOR,
+        GL_ONE_MINUS_SRC_COLOR,
+        useDualSource ? GL_SRC1_ALPHA : (GLenum)GL_SRC_ALPHA,
+        useDualSource ? GL_ONE_MINUS_SRC1_ALPHA : (GLenum)GL_ONE_MINUS_SRC_ALPHA,
+        GL_DST_ALPHA,
+        GL_ONE_MINUS_DST_ALPHA};
+
+    if (state.blendenable)
+    {
+      glEnable(GL_BLEND);
+    }
+    else
+    {
+      glDisable(GL_BLEND);
+    }
 
-  // Always call glBlendEquationSeparate and glBlendFuncSeparate, even when
-  // GL_BLEND is disabled, as a workaround for some bugs (possibly graphics
-  // driver issues?). See https://bugs.dolphin-emu.org/issues/10120 : "Sonic
-  // Adventure 2 Battle: graphics crash when loading first Dark level"
-  GLenum equation = state.subtract ? GL_FUNC_REVERSE_SUBTRACT : GL_FUNC_ADD;
-  GLenum equationAlpha = state.subtractAlpha ? GL_FUNC_REVERSE_SUBTRACT : GL_FUNC_ADD;
-  glBlendEquationSeparate(equation, equationAlpha);
-  glBlendFuncSeparate(src_factors[state.srcfactor], dst_factors[state.dstfactor],
-                      src_factors[state.srcfactoralpha], dst_factors[state.dstfactoralpha]);
+    // Always call glBlendEquationSeparate and glBlendFuncSeparate, even when
+    // GL_BLEND is disabled, as a workaround for some bugs (possibly graphics
+    // driver issues?). See https://bugs.dolphin-emu.org/issues/10120 : "Sonic
+    // Adventure 2 Battle: graphics crash when loading first Dark level"
+    GLenum equation = state.subtract ? GL_FUNC_REVERSE_SUBTRACT : GL_FUNC_ADD;
+    GLenum equationAlpha = state.subtractAlpha ? GL_FUNC_REVERSE_SUBTRACT : GL_FUNC_ADD;
+    glBlendEquationSeparate(equation, equationAlpha);
+    glBlendFuncSeparate(src_factors[state.srcfactor], dst_factors[state.dstfactor],
+                        src_factors[state.srcfactoralpha], dst_factors[state.dstfactoralpha]);
+  }
 
   const GLenum logic_op_codes[16] = {
       GL_CLEAR,         GL_AND,         GL_AND_REVERSE, GL_COPY,  GL_AND_INVERTED, GL_NOOP,