TextureConversionShader: Don't sample from adjacent rows when not needed

Source/Core/VideoBackends/D3D/TextureCache.cpp

@@ -240,7 +240,8 @@ void TextureCache::CopyEFBToCacheEntry(TCacheEntry* entry, bool is_depth_copy,
   }
 
   auto uid = TextureConversionShaderGen::GetShaderUid(dst_format, is_depth_copy, is_intensity,
-                                                      scale_by_half);
+                                                      scale_by_half,
+                                                      NeedsCopyFilterInShader(filter_coefficients));
   ID3D11PixelShader* pixel_shader = GetEFBToTexPixelShader(uid);
   if (!pixel_shader)
     return;

Source/Core/VideoBackends/OGL/TextureCache.cpp

@@ -515,7 +515,8 @@ void TextureCache::CopyEFBToCacheEntry(TCacheEntry* entry, bool is_depth_copy,
   glViewport(0, 0, destination_texture->GetConfig().width, destination_texture->GetConfig().height);
 
   auto uid = TextureConversionShaderGen::GetShaderUid(dst_format, is_depth_copy, is_intensity,
-                                                      scale_by_half);
+                                                      scale_by_half,
+                                                      NeedsCopyFilterInShader(filter_coefficients));
 
   auto it = m_efb_copy_programs.emplace(uid, EFBCopyShader());
   EFBCopyShader& shader = it.first->second;

Source/Core/VideoBackends/Vulkan/TextureCache.cpp

@@ -274,7 +274,8 @@ void TextureCache::CopyEFBToCacheEntry(TCacheEntry* entry, bool is_depth_copy,
                                                      VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL);
 
   auto uid = TextureConversionShaderGen::GetShaderUid(dst_format, is_depth_copy, is_intensity,
-                                                      scale_by_half);
+                                                      scale_by_half,
+                                                      NeedsCopyFilterInShader(filter_coefficients));
 
   auto it = m_efb_copy_to_tex_shaders.emplace(uid, VkShaderModule(VK_NULL_HANDLE));
   VkShaderModule& shader = it.first->second;

Source/Core/VideoCommon/TextureCacheBase.cpp

@@ -1499,8 +1499,8 @@ void TextureCacheBase::LoadTextureLevelZeroFromMemory(TCacheEntry* entry_to_upda
   }
 }
 
-TextureCacheBase::CopyFilterCoefficientArray
+TextureCacheBase::CopyFilterCoefficientArray TextureCacheBase::GetRAMCopyFilterCoefficients(
-TextureCacheBase::GetRAMCopyFilterCoefficients(const CopyFilterCoefficients::Values& coefficients)
+    const CopyFilterCoefficients::Values& coefficients) const
 {
   // To simplify the backend, we precalculate the three coefficients in common. Coefficients 0, 1
   // are for the row above, 2, 3, 4 are for the current pixel, and 5, 6 are for the row below.
@@ -1510,8 +1510,8 @@ TextureCacheBase::GetRAMCopyFilterCoefficients(const CopyFilterCoefficients::Val
           static_cast<u32>(coefficients[5]) + static_cast<u32>(coefficients[6])};
 }
 
-TextureCacheBase::CopyFilterCoefficientArray
+TextureCacheBase::CopyFilterCoefficientArray TextureCacheBase::GetVRAMCopyFilterCoefficients(
-TextureCacheBase::GetVRAMCopyFilterCoefficients(const CopyFilterCoefficients::Values& coefficients)
+    const CopyFilterCoefficients::Values& coefficients) const
 {
   // If the user disables the copy filter, only apply it to the VRAM copy.
   // This way games which are sensitive to changes to the RAM copy of the XFB will be unaffected.
@@ -1528,6 +1528,12 @@ TextureCacheBase::GetVRAMCopyFilterCoefficients(const CopyFilterCoefficients::Va
   return res;
 }
 
+bool TextureCacheBase::NeedsCopyFilterInShader(const CopyFilterCoefficientArray& coefficients) const
+{
+  // If the top/bottom coefficients are zero, no point sampling/blending from these rows.
+  return coefficients[0] != 0 || coefficients[2] != 0;
+}
+
 void TextureCacheBase::CopyRenderTargetToTexture(
     u32 dstAddr, EFBCopyFormat dstFormat, u32 width, u32 height, u32 dstStride, bool is_depth_copy,
     const EFBRectangle& srcRect, bool isIntensity, bool scaleByHalf, float y_scale, float gamma,
@@ -1650,11 +1656,12 @@ void TextureCacheBase::CopyRenderTargetToTexture(
 
   if (copy_to_ram)
   {
+    CopyFilterCoefficientArray coefficients = GetRAMCopyFilterCoefficients(filter_coefficients);
     PEControl::PixelFormat srcFormat = bpmem.zcontrol.pixel_format;
-    EFBCopyParams format(srcFormat, dstFormat, is_depth_copy, isIntensity);
+    EFBCopyParams format(srcFormat, dstFormat, is_depth_copy, isIntensity,
+                         NeedsCopyFilterInShader(coefficients));
     CopyEFB(dst, format, tex_w, bytes_per_row, num_blocks_y, dstStride, srcRect, scaleByHalf,
-            y_scale, gamma, clamp_top, clamp_bottom,
+            y_scale, gamma, clamp_top, clamp_bottom, coefficients);
-            GetRAMCopyFilterCoefficients(filter_coefficients));
   }
   else
   {

Source/Core/VideoCommon/TextureCacheBase.h

@@ -47,21 +47,23 @@ struct TextureAndTLUTFormat
 struct EFBCopyParams
 {
   EFBCopyParams(PEControl::PixelFormat efb_format_, EFBCopyFormat copy_format_, bool depth_,
-                bool yuv_)
+                bool yuv_, bool copy_filter_)
-      : efb_format(efb_format_), copy_format(copy_format_), depth(depth_), yuv(yuv_)
+      : efb_format(efb_format_), copy_format(copy_format_), depth(depth_), yuv(yuv_),
+        copy_filter(copy_filter_)
   {
   }
 
   bool operator<(const EFBCopyParams& rhs) const
   {
-    return std::tie(efb_format, copy_format, depth, yuv) <
+    return std::tie(efb_format, copy_format, depth, yuv, copy_filter) <
-           std::tie(rhs.efb_format, rhs.copy_format, rhs.depth, rhs.yuv);
+           std::tie(rhs.efb_format, rhs.copy_format, rhs.depth, rhs.yuv, rhs.copy_filter);
   }
 
   PEControl::PixelFormat efb_format;
   EFBCopyFormat copy_format;
   bool depth;
   bool yuv;
+  bool copy_filter;
 };
 
 struct TextureLookupInformation
@@ -106,6 +108,7 @@ private:
   static const int FRAMECOUNT_INVALID = 0;
 
 public:
+  // Reduced version of the full coefficient array, reduced to a single value for each row.
   using CopyFilterCoefficientArray = std::array<u32, 3>;
 
   struct TCacheEntry
@@ -278,6 +281,9 @@ public:
 protected:
   TextureCacheBase();
 
+  // Returns false if the top/bottom row coefficients are zero.
+  bool NeedsCopyFilterInShader(const CopyFilterCoefficientArray& coefficients) const;
+
   alignas(16) u8* temp = nullptr;
   size_t temp_size = 0;
 
@@ -329,9 +335,9 @@ private:
 
   // Precomputing the coefficients for the previous, current, and next lines for the copy filter.
   CopyFilterCoefficientArray
-  GetRAMCopyFilterCoefficients(const CopyFilterCoefficients::Values& coefficients);
+  GetRAMCopyFilterCoefficients(const CopyFilterCoefficients::Values& coefficients) const;
   CopyFilterCoefficientArray
-  GetVRAMCopyFilterCoefficients(const CopyFilterCoefficients::Values& coefficients);
+  GetVRAMCopyFilterCoefficients(const CopyFilterCoefficients::Values& coefficients) const;
 
   TexAddrCache textures_by_address;
   TexHashCache textures_by_hash;

Source/Core/VideoCommon/TextureConversionShader.cpp

@@ -180,21 +180,33 @@ static void WriteSampleFunction(char*& p, const EFBCopyParams& params, APIType A
   // The filter is only applied to the RGB channels, the alpha channel is left intact.
   WRITE(p, "float4 SampleEFB(float2 uv, float2 pixel_size, int xoffset)\n");
   WRITE(p, "{\n");
-  WRITE(p, "  float4 prev_row = ");
+  if (params.copy_filter)
-  WriteSampleOp(-1);
+  {
-  WRITE(p, ";\n");
+    WRITE(p, "  float4 prev_row = ");
-  WRITE(p, "  float4 current_row = ");
+    WriteSampleOp(-1);
-  WriteSampleOp(0);
+    WRITE(p, ";\n");
-  WRITE(p, ";\n");
+    WRITE(p, "  float4 current_row = ");
-  WRITE(p, "  float4 next_row = ");
+    WriteSampleOp(0);
-  WriteSampleOp(1);
+    WRITE(p, ";\n");
-  WRITE(p, ";\n");
+    WRITE(p, "  float4 next_row = ");
-  WRITE(p,
+    WriteSampleOp(1);
+    WRITE(p, ";\n");
+    WRITE(
+        p,
         "  float3 col = float3(clamp((int3(prev_row.rgb * 255.0) * filter_coefficients[0] +\n"
         "                             int3(current_row.rgb * 255.0) * filter_coefficients[1] +\n"
         "                             int3(next_row.rgb * 255.0) * filter_coefficients[2]) >> 6,\n"
         "                            int3(0, 0, 0), int3(255, 255, 255))) / 255.0;\n");
-  WRITE(p, "  return float4(col, current_row.a);\n");
+    WRITE(p, "  return float4(col, current_row.a);\n");
+  }
+  else
+  {
+    WRITE(p, "  float4 current_row = ");
+    WriteSampleOp(0);
+    WRITE(p, ";\n");
+    WRITE(p, "  return float4(clamp(int3(current_row.rgb * 255.0) * filter_coefficients[1], "
+             "int3(0, 0, 0), int3(255, 255, 255)), current_row.a);\n");
+  }
   WRITE(p, "}\n");
 }
 

Source/Core/VideoCommon/TextureConverterShaderGen.cpp

@@ -15,7 +15,7 @@
 namespace TextureConversionShaderGen
 {
 TCShaderUid GetShaderUid(EFBCopyFormat dst_format, bool is_depth_copy, bool is_intensity,
-                         bool scale_by_half)
+                         bool scale_by_half, bool copy_filter)
 {
   TCShaderUid out;
   UidData* uid_data = out.GetUidData<UidData>();
@@ -26,6 +26,7 @@ TCShaderUid GetShaderUid(EFBCopyFormat dst_format, bool is_depth_copy, bool is_i
   uid_data->is_depth_copy = is_depth_copy;
   uid_data->is_intensity = is_intensity;
   uid_data->scale_by_half = scale_by_half;
+  uid_data->copy_filter = copy_filter;
 
   return out;
 }
@@ -91,12 +92,21 @@ ShaderCode GenerateShader(APIType api_type, const UidData* uid_data)
 
   // The copy filter applies to both color and depth copies. This has been verified on hardware.
   // The filter is only applied to the RGB channels, the alpha channel is left intact.
-  out.Write("  float4 prev_row = SampleEFB(uv0, -1.0f);\n"
+  if (uid_data->copy_filter)
-            "  float4 current_row = SampleEFB(uv0, 0.0f);\n"
+  {
-            "  float4 next_row = SampleEFB(uv0, 1.0f);\n"
+    out.Write("  float4 prev_row = SampleEFB(uv0, -1.0f);\n"
-            "  float4 texcol = float4(prev_row.rgb * filter_coefficients[0] +\n"
+              "  float4 current_row = SampleEFB(uv0, 0.0f);\n"
-            "                         current_row.rgb * filter_coefficients[1] +\n"
+              "  float4 next_row = SampleEFB(uv0, 1.0f);\n"
-            "                         next_row.rgb * filter_coefficients[2], current_row.a);\n");
+              "  float4 texcol = float4(prev_row.rgb * filter_coefficients[0] +\n"
+              "                         current_row.rgb * filter_coefficients[1] +\n"
+              "                         next_row.rgb * filter_coefficients[2], current_row.a);\n");
+  }
+  else
+  {
+    out.Write(
+        "  float4 current_row = SampleEFB(uv0, 0.0f);\n"
+        "  float4 texcol = float4(current_row.rgb * filter_coefficients[1], current_row.a);\n");
+  }
 
   if (uid_data->is_depth_copy)
   {

Source/Core/VideoCommon/TextureConverterShaderGen.h

@@ -22,6 +22,7 @@ struct UidData
   u32 is_depth_copy : 1;
   u32 is_intensity : 1;
   u32 scale_by_half : 1;
+  u32 copy_filter : 1;
 };
 #pragma pack()
 
@@ -30,6 +31,6 @@ using TCShaderUid = ShaderUid<UidData>;
 ShaderCode GenerateShader(APIType api_type, const UidData* uid_data);
 
 TCShaderUid GetShaderUid(EFBCopyFormat dst_format, bool is_depth_copy, bool is_intensity,
-                         bool scale_by_half);
+                         bool scale_by_half, bool copy_filter);
 
 }  // namespace TextureConversionShaderGen