Merge pull request #4291 from degasus/shader_gen

PixelShaderGen: Fix UID issues.

Source/Core/VideoBackends/D3D/PixelShaderCache.cpp

@@ -578,7 +578,7 @@ bool PixelShaderCache::SetShader(DSTALPHA_MODE dstAlphaMode)
   }
 
   // Need to compile a new shader
-  ShaderCode code = GeneratePixelShaderCode(dstAlphaMode, APIType::D3D, uid.GetUidData());
+  ShaderCode code = GeneratePixelShaderCode(APIType::D3D, uid.GetUidData());
 
   D3DBlob* pbytecode;
   if (!D3D::CompilePixelShader(code.GetBuffer(), &pbytecode))

Source/Core/VideoBackends/D3D12/ShaderCache.cpp

@@ -231,8 +231,7 @@ void ShaderCache::HandlePSUIDChange(PixelShaderUid ps_uid, DSTALPHA_MODE ps_dst_
   }
   else
   {
-    ShaderCode ps_code =
+    ShaderCode ps_code = GeneratePixelShaderCode(APIType::D3D, ps_uid.GetUidData());
-        GeneratePixelShaderCode(ps_dst_alpha_mode, APIType::D3D, ps_uid.GetUidData());
     ID3DBlob* ps_bytecode = nullptr;
 
     if (!D3D::CompilePixelShader(ps_code.GetBuffer(), &ps_bytecode))

Source/Core/VideoBackends/Null/ShaderCache.cpp

@@ -60,7 +60,7 @@ bool ShaderCache<Uid>::SetShader(DSTALPHA_MODE dst_alpha_mode, u32 primitive_typ
   }
 
   // Need to compile a new shader
-  ShaderCode code = GenerateCode(dst_alpha_mode, APIType::OpenGL, uid);
+  ShaderCode code = GenerateCode(APIType::OpenGL, uid);
   m_shaders.emplace(uid, code.GetBuffer());
 
   GFX_DEBUGGER_PAUSE_AT(NEXT_PIXEL_SHADER_CHANGE, true);

Source/Core/VideoBackends/Null/ShaderCache.h

@@ -26,7 +26,7 @@ public:
 
 protected:
   virtual Uid GetUid(DSTALPHA_MODE dst_alpha_mode, u32 primitive_type, APIType api_type) = 0;
-  virtual ShaderCode GenerateCode(DSTALPHA_MODE dst_alpha_mode, APIType api_type, Uid uid) = 0;
+  virtual ShaderCode GenerateCode(APIType api_type, Uid uid) = 0;
 
 private:
   std::map<Uid, std::string> m_shaders;
@@ -45,8 +45,7 @@ protected:
   {
     return GetVertexShaderUid();
   }
-  ShaderCode GenerateCode(DSTALPHA_MODE dst_alpha_mode, APIType api_type,
+  ShaderCode GenerateCode(APIType api_type, VertexShaderUid uid) override
-                          VertexShaderUid uid) override
   {
     return GenerateVertexShaderCode(api_type, uid.GetUidData());
   }
@@ -63,8 +62,7 @@ protected:
   {
     return GetGeometryShaderUid(primitive_type);
   }
-  ShaderCode GenerateCode(DSTALPHA_MODE dst_alpha_mode, APIType api_type,
+  ShaderCode GenerateCode(APIType api_type, GeometryShaderUid uid) override
-                          GeometryShaderUid uid) override
   {
     return GenerateGeometryShaderCode(api_type, uid.GetUidData());
   }
@@ -80,10 +78,9 @@ protected:
   {
     return GetPixelShaderUid(dst_alpha_mode);
   }
-  ShaderCode GenerateCode(DSTALPHA_MODE dst_alpha_mode, APIType api_type,
+  ShaderCode GenerateCode(APIType api_type, PixelShaderUid uid) override
-                          PixelShaderUid uid) override
   {
-    return GeneratePixelShaderCode(dst_alpha_mode, api_type, uid.GetUidData());
+    return GeneratePixelShaderCode(api_type, uid.GetUidData());
   }
 };
 

Source/Core/VideoBackends/OGL/ProgramShaderCache.cpp

@@ -208,7 +208,7 @@ SHADER* ProgramShaderCache::SetShader(DSTALPHA_MODE dstAlphaMode, u32 primitive_
   newentry.in_cache = 0;
 
   ShaderCode vcode = GenerateVertexShaderCode(APIType::OpenGL, uid.vuid.GetUidData());
-  ShaderCode pcode = GeneratePixelShaderCode(dstAlphaMode, APIType::OpenGL, uid.puid.GetUidData());
+  ShaderCode pcode = GeneratePixelShaderCode(APIType::OpenGL, uid.puid.GetUidData());
   ShaderCode gcode;
   if (g_ActiveConfig.backend_info.bSupportsGeometryShaders &&
       !uid.guid.GetUidData()->IsPassthrough())

Source/Core/VideoBackends/Vulkan/ObjectCache.cpp

@@ -525,8 +525,7 @@ VkShaderModule ObjectCache::GetGeometryShaderForUid(const GeometryShaderUid& uid
   return module;
 }
 
-VkShaderModule ObjectCache::GetPixelShaderForUid(const PixelShaderUid& uid,
+VkShaderModule ObjectCache::GetPixelShaderForUid(const PixelShaderUid& uid)
-                                                 DSTALPHA_MODE dstalpha_mode)
 {
   auto it = m_ps_cache.shader_map.find(uid);
   if (it != m_ps_cache.shader_map.end())
@@ -535,8 +534,7 @@ VkShaderModule ObjectCache::GetPixelShaderForUid(const PixelShaderUid& uid,
   // Not in the cache, so compile the shader.
   ShaderCompiler::SPIRVCodeVector spv;
   VkShaderModule module = VK_NULL_HANDLE;
-  ShaderCode source_code =
+  ShaderCode source_code = GeneratePixelShaderCode(APIType::Vulkan, uid.GetUidData());
-      GeneratePixelShaderCode(dstalpha_mode, APIType::Vulkan, uid.GetUidData());
   if (ShaderCompiler::CompileFragmentShader(&spv, source_code.GetBuffer().c_str(),
                                             source_code.GetBuffer().length()))
   {

Source/Core/VideoBackends/Vulkan/ObjectCache.h

@@ -101,7 +101,7 @@ public:
   // Accesses ShaderGen shader caches
   VkShaderModule GetVertexShaderForUid(const VertexShaderUid& uid);
   VkShaderModule GetGeometryShaderForUid(const GeometryShaderUid& uid);
-  VkShaderModule GetPixelShaderForUid(const PixelShaderUid& uid, DSTALPHA_MODE dstalpha_mode);
+  VkShaderModule GetPixelShaderForUid(const PixelShaderUid& uid);
 
   // Static samplers
   VkSampler GetPointSampler() const { return m_point_sampler; }

Source/Core/VideoBackends/Vulkan/StateTracker.cpp

@@ -223,7 +223,7 @@ bool StateTracker::CheckForShaderChanges(u32 gx_primitive_type, DSTALPHA_MODE ds
 
   if (ps_uid != m_ps_uid)
   {
-    m_pipeline_state.ps = g_object_cache->GetPixelShaderForUid(ps_uid, dstalpha_mode);
+    m_pipeline_state.ps = g_object_cache->GetPixelShaderForUid(ps_uid);
     m_ps_uid = ps_uid;
     changed = true;
   }

Source/Core/VideoCommon/LightingShaderGen.cpp

@@ -72,9 +72,9 @@ static void GenerateLightShader(ShaderCode& object, const LightingUidData& uid_d
 // inColorName is color in vs and colors_ in ps
 // dest is o.colors_ in vs and colors_ in ps
 void GenerateLightingShaderCode(ShaderCode& object, const LightingUidData& uid_data, int components,
-                                const char* inColorName, const char* dest)
+                                u32 numColorChans, const char* inColorName, const char* dest)
 {
-  for (unsigned int j = 0; j < xfmem.numChan.numColorChans; j++)
+  for (unsigned int j = 0; j < numColorChans; j++)
   {
     object.Write("{\n");
 

Source/Core/VideoCommon/LightingShaderGen.h

@@ -47,5 +47,5 @@ static const char s_lighting_struct[] = "struct Light {\n"
                                         "};\n";
 
 void GenerateLightingShaderCode(ShaderCode& object, const LightingUidData& uid_data, int components,
-                                const char* inColorName, const char* dest);
+                                u32 numColorChans, const char* inColorName, const char* dest);
 void GetLightingShaderUid(LightingUidData& uid_data);

Source/Core/VideoCommon/PixelShaderGen.cpp

@@ -207,7 +207,7 @@ PixelShaderUid GetPixelShaderUid(DSTALPHA_MODE dstAlphaMode)
     // The lighting shader only needs the two color bits of the 23bit component bit array.
     uid_data->components =
         (VertexLoaderManager::g_current_components & (VB_HAS_COL0 | VB_HAS_COL1)) >> VB_COL_SHIFT;
-    ;
+    uid_data->numColorChans = xfmem.numChan.numColorChans;
     GetLightingShaderUid(uid_data->lighting);
   }
 
@@ -342,11 +342,10 @@ static void WriteTevRegular(ShaderCode& out, const char* components, int bias, i
 static void SampleTexture(ShaderCode& out, const char* texcoords, const char* texswap, int texmap,
                           bool stereo, APIType ApiType);
 static void WriteAlphaTest(ShaderCode& out, const pixel_shader_uid_data* uid_data, APIType ApiType,
-                           DSTALPHA_MODE dstAlphaMode, bool per_pixel_depth);
+                           bool per_pixel_depth);
 static void WriteFog(ShaderCode& out, const pixel_shader_uid_data* uid_data);
 
-ShaderCode GeneratePixelShaderCode(DSTALPHA_MODE dstAlphaMode, APIType ApiType,
+ShaderCode GeneratePixelShaderCode(APIType ApiType, const pixel_shader_uid_data* uid_data)
-                                   const pixel_shader_uid_data* uid_data)
 {
   ShaderCode out;
 
@@ -501,21 +500,10 @@ ShaderCode GeneratePixelShaderCode(DSTALPHA_MODE dstAlphaMode, APIType ApiType,
       out.Write("[earlydepthstencil]\n");
     }
   }
-  else if (bpmem.UseEarlyDepthTest() &&
-           (uid_data->fast_depth_calc || bpmem.alpha_test.TestResult() == AlphaTest::UNDETERMINED))
-  {
-    static bool warn_once = true;
-    if (warn_once)
-      WARN_LOG(VIDEO, "Early z test enabled but not possible to emulate with current "
-                      "configuration. Make sure to enable fast depth calculations. If this message "
-                      "still shows up your hardware isn't able to emulate the feature properly (a "
-                      "GPU with D3D 11.0 / OGL 4.2 support is required).");
-    warn_once = false;
-  }
 
   if (ApiType == APIType::OpenGL || ApiType == APIType::Vulkan)
   {
-    if (dstAlphaMode == DSTALPHA_DUAL_SOURCE_BLEND)
+    if (uid_data->dstAlphaMode == DSTALPHA_DUAL_SOURCE_BLEND)
     {
       if (DriverDetails::HasBug(DriverDetails::BUG_BROKEN_FRAGMENT_SHADER_INDEX_DECORATION))
       {
@@ -587,7 +575,8 @@ ShaderCode GeneratePixelShaderCode(DSTALPHA_MODE dstAlphaMode, APIType ApiType,
   {
     out.Write("void main(\n");
     out.Write("  out float4 ocol0 : SV_Target0,%s%s\n  in float4 rawpos : SV_Position,\n",
-              dstAlphaMode == DSTALPHA_DUAL_SOURCE_BLEND ? "\n  out float4 ocol1 : SV_Target1," :
+              uid_data->dstAlphaMode == DSTALPHA_DUAL_SOURCE_BLEND ?
+                  "\n  out float4 ocol1 : SV_Target1," :
                   "",
               uid_data->per_pixel_depth ? "\n  out float depth : SV_Depth," : "");
 
@@ -649,7 +638,7 @@ ShaderCode GeneratePixelShaderCode(DSTALPHA_MODE dstAlphaMode, APIType ApiType,
     // out.SetConstantsUsed(C_PLIGHTS, C_PLIGHTS+31); // TODO: Can be optimized further
     // out.SetConstantsUsed(C_PMATERIALS, C_PMATERIALS+3);
     GenerateLightingShaderCode(out, uid_data->lighting, uid_data->components << VB_COL_SHIFT,
-                               "colors_", "col");
+                               uid_data->numColorChans, "colors_", "col");
   }
 
   // HACK to handle cases where the tex gen is not enabled
@@ -716,7 +705,7 @@ ShaderCode GeneratePixelShaderCode(DSTALPHA_MODE dstAlphaMode, APIType ApiType,
   // testing result)
   if (uid_data->Pretest == AlphaTest::UNDETERMINED ||
       (uid_data->Pretest == AlphaTest::FAIL && uid_data->late_ztest))
-    WriteAlphaTest(out, uid_data, ApiType, dstAlphaMode, uid_data->per_pixel_depth);
+    WriteAlphaTest(out, uid_data, ApiType, uid_data->per_pixel_depth);
 
   if (uid_data->zfreeze)
   {
@@ -787,7 +776,7 @@ ShaderCode GeneratePixelShaderCode(DSTALPHA_MODE dstAlphaMode, APIType ApiType,
       out.Write("\tdepth = float(zCoord) / 16777216.0;\n");
   }
 
-  if (dstAlphaMode == DSTALPHA_ALPHA_PASS)
+  if (uid_data->dstAlphaMode == DSTALPHA_ALPHA_PASS)
   {
     out.SetConstantsUsed(C_ALPHA, C_ALPHA);
     out.Write("\tocol0 = float4(float3(prev.rgb), float(" I_ALPHA ".a)) / 255.0;\n");
@@ -799,7 +788,7 @@ ShaderCode GeneratePixelShaderCode(DSTALPHA_MODE dstAlphaMode, APIType ApiType,
   }
 
   // Use dual-source color blending to perform dst alpha in a single pass
-  if (dstAlphaMode == DSTALPHA_DUAL_SOURCE_BLEND)
+  if (uid_data->dstAlphaMode == DSTALPHA_DUAL_SOURCE_BLEND)
   {
     out.SetConstantsUsed(C_ALPHA, C_ALPHA);
 
@@ -1195,7 +1184,7 @@ static const char* tevAlphaFunclogicTable[] = {
 };
 
 static void WriteAlphaTest(ShaderCode& out, const pixel_shader_uid_data* uid_data, APIType ApiType,
-                           DSTALPHA_MODE dstAlphaMode, bool per_pixel_depth)
+                           bool per_pixel_depth)
 {
   static const char* alphaRef[2] = {I_ALPHA ".r", I_ALPHA ".g"};
 
@@ -1222,7 +1211,7 @@ static void WriteAlphaTest(ShaderCode& out, const pixel_shader_uid_data* uid_dat
     out.Write(")) {\n");
 
   out.Write("\t\tocol0 = float4(0.0, 0.0, 0.0, 0.0);\n");
-  if (dstAlphaMode == DSTALPHA_DUAL_SOURCE_BLEND)
+  if (uid_data->dstAlphaMode == DSTALPHA_DUAL_SOURCE_BLEND)
     out.Write("\t\tocol1 = float4(0.0, 0.0, 0.0, 0.0);\n");
   if (per_pixel_depth)
   {

Source/Core/VideoCommon/PixelShaderGen.h

@@ -52,7 +52,8 @@ struct pixel_shader_uid_data
   u32 zfreeze : 1;
   u32 msaa : 1;
   u32 ssaa : 1;
-  u32 pad : 16;
+  u32 numColorChans : 2;
+  u32 pad : 14;
 
   u32 texMtxInfo_n_projection : 8;  // 8x1 bit
   u32 tevindref_bi0 : 3;
@@ -166,6 +167,5 @@ struct pixel_shader_uid_data
 
 typedef ShaderUid<pixel_shader_uid_data> PixelShaderUid;
 
-ShaderCode GeneratePixelShaderCode(DSTALPHA_MODE dstAlphaMode, APIType ApiType,
+ShaderCode GeneratePixelShaderCode(APIType ApiType, const pixel_shader_uid_data* uid_data);
-                                   const pixel_shader_uid_data* uid_data);
 PixelShaderUid GetPixelShaderUid(DSTALPHA_MODE dstAlphaMode);

Source/Core/VideoCommon/VertexShaderGen.cpp

@@ -250,7 +250,8 @@ ShaderCode GenerateVertexShaderCode(APIType api_type, const vertex_shader_uid_da
       out.Write("o.colors_0 = float4(1.0, 1.0, 1.0, 1.0);\n");
   }
 
-  GenerateLightingShaderCode(out, uid_data->lighting, uid_data->components, "color", "o.colors_");
+  GenerateLightingShaderCode(out, uid_data->lighting, uid_data->components, uid_data->numColorChans,
+                             "color", "o.colors_");
 
   if (uid_data->numColorChans < 2)
   {