From ea30e543e62b4379a8ca3f5b66061b0987e4cedf Mon Sep 17 00:00:00 2001
From: riperiperi <rhy3756547@hotmail.com>
Date: Fri, 5 Jul 2024 09:35:17 +0100
Subject: [PATCH] Fix preload cbs optimization (for real) (#34)

* Mostly fix preload cbs. There seems to be some random flickering...

* fix index buffer usage range

* fix missing preflush submit before present
---
 src/Ryujinx.Graphics.Metal/BufferHolder.cs |  4 ++--
 src/Ryujinx.Graphics.Metal/Pipeline.cs     | 26 ++++++++++++++++------
 2 files changed, 21 insertions(+), 9 deletions(-)

diff --git a/src/Ryujinx.Graphics.Metal/BufferHolder.cs b/src/Ryujinx.Graphics.Metal/BufferHolder.cs
index e0089322f..47e9cd0e3 100644
--- a/src/Ryujinx.Graphics.Metal/BufferHolder.cs
+++ b/src/Ryujinx.Graphics.Metal/BufferHolder.cs
@@ -191,14 +191,14 @@ namespace Ryujinx.Graphics.Metal
             }
 
             if (cbs != null &&
-                _pipeline.RenderPassActive &&
+                cbs.Value.Encoders.CurrentEncoderType == EncoderType.Render &&
                 !(_buffer.HasCommandBufferDependency(cbs.Value) &&
                   _waitable.IsBufferRangeInUse(cbs.Value.CommandBufferIndex, offset, dataSize)))
             {
                 // If the buffer hasn't been used on the command buffer yet, try to preload the data.
                 // This avoids ending and beginning render passes on each buffer data upload.
 
-                cbs = _pipeline.PreloadCbs;
+                cbs = _pipeline.GetPreloadCommandBuffer();
             }
 
             if (allowCbsWait)
diff --git a/src/Ryujinx.Graphics.Metal/Pipeline.cs b/src/Ryujinx.Graphics.Metal/Pipeline.cs
index 32b520e0e..5e4e55ea3 100644
--- a/src/Ryujinx.Graphics.Metal/Pipeline.cs
+++ b/src/Ryujinx.Graphics.Metal/Pipeline.cs
@@ -36,7 +36,6 @@ namespace Ryujinx.Graphics.Metal
         internal CommandBufferScoped Cbs { get; private set; }
         internal CommandBufferEncoder Encoders => Cbs.Encoders;
         internal EncoderType CurrentEncoderType => Encoders.CurrentEncoderType;
-        internal bool RenderPassActive { get; private set; }
 
         public Pipeline(MTLDevice device, MetalRenderer renderer)
         {
@@ -137,7 +136,7 @@ namespace Ryujinx.Graphics.Metal
 
             Cbs.CommandBuffer.PresentDrawable(drawable);
 
-            CommandBuffer = (Cbs = _renderer.CommandBufferPool.ReturnAndRent(Cbs)).CommandBuffer;
+            FlushCommandsImpl();
 
             // TODO: Auto flush counting
             _renderer.SyncManager.GetAndResetWaitTicks();
@@ -146,6 +145,13 @@ namespace Ryujinx.Graphics.Metal
             dst.Dispose();
         }
 
+        public CommandBufferScoped GetPreloadCommandBuffer()
+        {
+            PreloadCbs ??= _renderer.CommandBufferPool.Rent();
+
+            return PreloadCbs.Value;
+        }
+
         public void FlushCommandsIfWeightExceeding(IAuto disposedResource, ulong byteWeight)
         {
             bool usedByCurrentCb = disposedResource.HasCommandBufferDependency(Cbs);
@@ -383,12 +389,16 @@ namespace Ryujinx.Graphics.Metal
 
             var indexBuffer = _encoderStateManager.IndexBuffer;
 
+            ulong offset = _encoderStateManager.IndexBufferOffset;
+            MTLIndexType type = _encoderStateManager.IndexType;
+            int indexSize = type == MTLIndexType.UInt32 ? sizeof(int) : sizeof(short);
+
             renderCommandEncoder.DrawIndexedPrimitives(
                 primitiveType,
                 (ulong)indexCount,
-                _encoderStateManager.IndexType,
-                indexBuffer.Get(Cbs, 0, indexCount * sizeof(int)).Value,
-                _encoderStateManager.IndexBufferOffset,
+                type,
+                indexBuffer.Get(Cbs, (int)offset, indexCount * indexSize).Value,
+                offset,
                 (ulong)instanceCount,
                 firstVertex,
                 (ulong)firstInstance);
@@ -533,11 +543,13 @@ namespace Ryujinx.Graphics.Metal
 
         public void SetPrimitiveRestart(bool enable, int index)
         {
-            // TODO: Supported for LineStrip and TriangleStrip
+            // Always active for LineStrip and TriangleStrip
             // https://github.com/gpuweb/gpuweb/issues/1220#issuecomment-732483263
             // https://developer.apple.com/documentation/metal/mtlrendercommandencoder/1515520-drawindexedprimitives
             // https://stackoverflow.com/questions/70813665/how-to-render-multiple-trianglestrips-using-metal
-            Logger.Warning?.Print(LogClass.Gpu, "Not Implemented!");
+
+            // Emulating disabling this is very difficult. It's unlikely for an index buffer to use the largest possible index,
+            // so it's fine nearly all of the time.
         }
 
         public void SetPrimitiveTopology(PrimitiveTopology topology)