Merge branch 'feature/vulkan-index-buff-compute' into 'master'

Vulkan: Use compute shader for non-indirect unsupported topology index buffer conversions

See merge request ryubing/ryujinx!5

.gitignore

@@ -178,3 +178,7 @@ PublishProfiles/
 
 # Ignore MacOS Attribute Files
 ._*
+
+# Ignore distribution build files
+distribution/macos/temp/
+distribution/macos/output/

assets/locales.json

@@ -19914,7 +19914,7 @@
         "pl_PL": "Skróty Klawiszowe Klawiatury",
         "pt_BR": "Atalhos do Teclado",
         "ru_RU": "Горячие клавиши",
-        "sv_SE": "Snabbtangenter för tangentbord",
+        "sv_SE": "Snabbtangenter",
         "th_TH": "ปุ่มลัดของคีย์บอร์ด",
         "tr_TR": "Klavye Kısayolları",
         "uk_UA": "Гарячі клавіші",

distribution/macos/bundle_fix_up.py

@@ -562,26 +562,29 @@ search_path = [
 
 
 for path in content_directory.rglob("**/*.dylib"):
-    current_search_path = [path.parent]
+    if not path.name.startswith("._"):
-    current_search_path.extend(search_path)
+        current_search_path = [path.parent]
+        current_search_path.extend(search_path)
 
-    fixup_dylib(
+        print(f"Fixing path '{path}' using search path of '{current_search_path}' for context of '{content_directory}'.")
-        path,
+        fixup_dylib(
-        get_path_related_to_target_exec(content_directory, path),
+            path,
-        current_search_path,
+            get_path_related_to_target_exec(content_directory, path),
-        content_directory,
+            current_search_path,
-    )
+            content_directory,
+        )
 
 for path in content_directory.rglob("**/*.so"):
-    current_search_path = [path.parent]
+    if not path.name.startswith("._"):
-    current_search_path.extend(search_path)
+        current_search_path = [path.parent]
+        current_search_path.extend(search_path)
 
-    fixup_dylib(
+        fixup_dylib(
-        path,
+            path,
-        get_path_related_to_target_exec(content_directory, path),
+            get_path_related_to_target_exec(content_directory, path),
-        current_search_path,
+            current_search_path,
-        content_directory,
+            content_directory,
-    )
+        )
 
 
 with open(executable_path, "rb") as input:

distribution/macos/create_app_bundle.sh

@@ -30,9 +30,14 @@ cp -r "$PUBLISH_DIRECTORY/THIRDPARTY.md" "$APP_BUNDLE_DIRECTORY/Contents/Resourc
 echo -n "APPL????" > "$APP_BUNDLE_DIRECTORY/Contents/PkgInfo"
 
 # Fixup libraries and executable
+echo "Running bundle fix up python script"
 python3 bundle_fix_up.py "$APP_BUNDLE_DIRECTORY" MacOS/Ryujinx
 
+# Resign all dyplib files as ad-hoc after changing them
+find "$APP_BUNDLE_DIRECTORY/Contents/Frameworks" -type f -name "*.dylib" -exec codesign --force --sign - {} \;
+
 # Now sign it
+echo "Starting signing process"
 if ! [ -x "$(command -v codesign)" ];
 then
     if ! [ -x "$(command -v rcodesign)" ];
@@ -42,9 +47,9 @@ then
     fi
 
     # cargo install apple-codesign
-    echo "Usign rcodesign for ad-hoc signing"
+    echo "Using rcodesign for ad-hoc signing"
     rcodesign sign --entitlements-xml-path "$ENTITLEMENTS_FILE_PATH" "$APP_BUNDLE_DIRECTORY"
 else
-    echo "Usign codesign for ad-hoc signing"
+    echo "Using codesign for ad-hoc signing"
     codesign --entitlements "$ENTITLEMENTS_FILE_PATH" -f -s - "$APP_BUNDLE_DIRECTORY"
 fi

src/Ryujinx.Graphics.Vulkan/HelperShader.cs

@@ -874,57 +874,42 @@ namespace Ryujinx.Graphics.Vulkan
 
         public unsafe void ConvertIndexBuffer(VulkanRenderer gd,
             CommandBufferScoped cbs,
-            BufferHolder src,
+            BufferHolder srcIndexBuffer,
-            BufferHolder dst,
+            BufferHolder dstIndexBuffer,
             IndexBufferPattern pattern,
             int indexSize,
             int srcOffset,
             int indexCount)
         {
             // TODO: Support conversion with primitive restart enabled.
-            // TODO: Convert with a compute shader?
 
+            int primitiveCount = pattern.GetPrimitiveCount(indexCount);
             int convertedCount = pattern.GetConvertedCount(indexCount);
             int outputIndexSize = 4;
             
-            Buffer srcBuffer = src.GetBuffer().Get(cbs, srcOffset, indexCount * indexSize).Value;
+            Buffer dstBuffer = dstIndexBuffer.GetBuffer().Get(cbs, 0, convertedCount * outputIndexSize).Value;
-            Buffer dstBuffer = dst.GetBuffer().Get(cbs, 0, convertedCount * outputIndexSize).Value;
 
-            gd.Api.CmdFillBuffer(cbs.CommandBuffer, dstBuffer, 0, Vk.WholeSize, 0);
+            const int ParamsBufferSize = 16 * sizeof(int);
 
-            List<BufferCopy> bufferCopy = [];
+            Span<int> shaderParams = stackalloc int[ParamsBufferSize / sizeof(int)];
-            int outputOffset = 0;
 
-            // Try to merge copies of adjacent indices to reduce copy count.
+            shaderParams[8] = pattern.PrimitiveVertices;
-            int sequenceStart = 0;
+            shaderParams[9] = pattern.PrimitiveVerticesOut;
-            int sequenceLength = 0;
+            shaderParams[10] = indexSize;
+            shaderParams[11] = outputIndexSize;
+            shaderParams[12] = pattern.BaseIndex;
+            shaderParams[13] = pattern.IndexStride;
+            shaderParams[14] = srcOffset;
+            shaderParams[15] = primitiveCount;
 
-            foreach (int index in pattern.GetIndexMapping(indexCount))
+            pattern.OffsetIndex.CopyTo(shaderParams[..pattern.OffsetIndex.Length]);
-            {
-                if (sequenceLength > 0)
-                {
-                    if (index == sequenceStart + sequenceLength && indexSize == outputIndexSize)
-                    {
-                        sequenceLength++;
-                        continue;
-                    }
 
-                    // Commit the copy so far.
+            using var patternScoped = gd.BufferManager.ReserveOrCreate(gd, cbs, ParamsBufferSize);
-                    bufferCopy.Add(new BufferCopy((ulong)(srcOffset + sequenceStart * indexSize), (ulong)outputOffset, (ulong)(indexSize * sequenceLength)));
+            var patternBuffer = patternScoped.Holder;
-                    outputOffset += outputIndexSize * sequenceLength;
-                }
 
-                sequenceStart = index;
+            patternBuffer.SetDataUnchecked<int>(patternScoped.Offset, shaderParams);
-                sequenceLength = 1;
-            }
 
-            if (sequenceLength > 0)
+            _pipeline.SetCommandBuffer(cbs);
-            {
-                // Commit final pending copy.
-                bufferCopy.Add(new BufferCopy((ulong)(srcOffset + sequenceStart * indexSize), (ulong)outputOffset, (ulong)(indexSize * sequenceLength)));
-            }
-
-            BufferCopy[] bufferCopyArray = bufferCopy.ToArray();
 
             BufferHolder.InsertBufferBarrier(
                 gd,
@@ -937,10 +922,11 @@ namespace Ryujinx.Graphics.Vulkan
                 0,
                 convertedCount * outputIndexSize);
 
-            fixed (BufferCopy* pBufferCopy = bufferCopyArray)
+            _pipeline.SetUniformBuffers([new BufferAssignment(0, new BufferRange(patternScoped.Handle, patternScoped.Offset, ParamsBufferSize))]);
-            {
+            _pipeline.SetStorageBuffers(1, new[] { srcIndexBuffer.GetBuffer(), dstIndexBuffer.GetBuffer() });
-                gd.Api.CmdCopyBuffer(cbs.CommandBuffer, srcBuffer, dstBuffer, (uint)bufferCopyArray.Length, pBufferCopy);
+
-            }
+            _pipeline.SetProgram(_programConvertIndexBuffer);
+            _pipeline.DispatchCompute(BitUtils.DivRoundUp(primitiveCount, 16), 1, 1);
 
             BufferHolder.InsertBufferBarrier(
                 gd,
@@ -952,6 +938,8 @@ namespace Ryujinx.Graphics.Vulkan
                 PipelineStageFlags.AllCommandsBit,
                 0,
                 convertedCount * outputIndexSize);
+
+            _pipeline.Finish(gd, cbs);
         }
 
         public void CopyIncompatibleFormats(

src/Ryujinx.Graphics.Vulkan/IndexBufferPattern.cs

@@ -47,28 +47,6 @@ namespace Ryujinx.Graphics.Vulkan
             return primitiveCount * OffsetIndex.Length;
         }
 
-        public IEnumerable<int> GetIndexMapping(int indexCount)
-        {
-            int primitiveCount = GetPrimitiveCount(indexCount);
-            int index = BaseIndex;
-
-            for (int i = 0; i < primitiveCount; i++)
-            {
-                if (RepeatStart)
-                {
-                    // Used for triangle fan
-                    yield return 0;
-                }
-
-                for (int j = RepeatStart ? 1 : 0; j < OffsetIndex.Length; j++)
-                {
-                    yield return index + OffsetIndex[j];
-                }
-
-                index += IndexStride;
-            }
-        }
-
         public BufferHandle GetRepeatingBuffer(int vertexCount, out int indexCount)
         {
             int primitiveCount = GetPrimitiveCount(vertexCount);