Add support for HLE macros and accelerate MultiDrawElementsIndirectCount #2 (#2557)

* Add support for HLE macros and accelerate MultiDrawElementsIndirectCount

* Add missing barrier

* Fix index buffer count

* Add support check for each macro hle before use

* Add missing xml doc

Co-authored-by: gdkchan <gab.dark.100@gmail.com>

3 files changed, 46 insertions, 12 deletions

diff --git a/Ryujinx.Graphics.Gpu/Engine/GPFifo/GPFifoClass.cs b/Ryujinx.Graphics.Gpu/Engine/GPFifo/GPFifoClass.cs
index 28822f4e..fe49b0f2 100644
--- a/Ryujinx.Graphics.Gpu/Engine/GPFifo/GPFifoClass.cs
+++ b/Ryujinx.Graphics.Gpu/Engine/GPFifo/GPFifoClass.cs
@@ -161,6 +161,8 @@ namespace Ryujinx.Graphics.Gpu.Engine.GPFifo
         /// <param name="argument">Method call argument</param>
         public void SetReference(int argument)
         {
+            _context.Renderer.Pipeline.CommandBufferBarrier();
+
             _context.CreateHostSyncIfNeeded();
         }
 
@@ -195,10 +197,11 @@ namespace Ryujinx.Graphics.Gpu.Engine.GPFifo
         /// Pushes an argument to a macro.
         /// </summary>
         /// <param name="index">Index of the macro</param>
+        /// <param name="gpuVa">GPU virtual address where the command word is located</param>
         /// <param name="argument">Argument to be pushed to the macro</param>
-        public void MmePushArgument(int index, int argument)
+        public void MmePushArgument(int index, ulong gpuVa, int argument)
         {
-            _macros[index].PushArgument(argument);
+            _macros[index].PushArgument(gpuVa, argument);
         }
 
         /// <summary>
@@ -208,7 +211,7 @@ namespace Ryujinx.Graphics.Gpu.Engine.GPFifo
         /// <param name="argument">Initial argument passed to the macro</param>
         public void MmeStart(int index, int argument)
         {
-            _macros[index].StartExecution(argument);
+            _macros[index].StartExecution(_context, _parent, _macroCode, argument);
         }
 
         /// <summary>
diff --git a/Ryujinx.Graphics.Gpu/Engine/GPFifo/GPFifoDevice.cs b/Ryujinx.Graphics.Gpu/Engine/GPFifo/GPFifoDevice.cs
index ada3bc4b..b3de738d 100644
--- a/Ryujinx.Graphics.Gpu/Engine/GPFifo/GPFifoDevice.cs
+++ b/Ryujinx.Graphics.Gpu/Engine/GPFifo/GPFifoDevice.cs
@@ -54,11 +54,12 @@ namespace Ryujinx.Graphics.Gpu.Engine.GPFifo
             /// <summary>
             /// Fetch the command buffer.
             /// </summary>
-            public void Fetch(MemoryManager memoryManager)
+            /// <param name="flush">If true, flushes potential GPU written data before reading the command buffer</param>
+            public void Fetch(MemoryManager memoryManager, bool flush = true)
             {
                 if (Words == null)
                 {
-                    Words = MemoryMarshal.Cast<byte, int>(memoryManager.GetSpan(EntryAddress, (int)EntryCount * 4, true)).ToArray();
+                    Words = MemoryMarshal.Cast<byte, int>(memoryManager.GetSpan(EntryAddress, (int)EntryCount * 4, flush)).ToArray();
                 }
             }
         }
@@ -73,6 +74,7 @@ namespace Ryujinx.Graphics.Gpu.Engine.GPFifo
         private readonly AutoResetEvent _event;
 
         private bool _interrupt;
+        private int _flushSkips;
 
         /// <summary>
         /// Creates a new instance of the GPU General Purpose FIFO device.
@@ -188,8 +190,16 @@ namespace Ryujinx.Graphics.Gpu.Engine.GPFifo
             // Process command buffers.
             while (_ibEnable && !_interrupt && _commandBufferQueue.TryDequeue(out CommandBuffer entry))
             {
+                bool flushCommandBuffer = true;
+
+                if (_flushSkips != 0)
+                {
+                    _flushSkips--;
+                    flushCommandBuffer = false;
+                }
+
                 _currentCommandBuffer = entry;
-                _currentCommandBuffer.Fetch(entry.Processor.MemoryManager);
+                _currentCommandBuffer.Fetch(entry.Processor.MemoryManager, flushCommandBuffer);
 
                 // If we are changing the current channel,
                 // we need to force all the host state to be updated.
@@ -199,13 +209,25 @@ namespace Ryujinx.Graphics.Gpu.Engine.GPFifo
                     entry.Processor.ForceAllDirty();
                 }
 
-                entry.Processor.Process(_currentCommandBuffer.Words);
+                entry.Processor.Process(entry.EntryAddress, _currentCommandBuffer.Words);
             }
 
             _interrupt = false;
         }
 
         /// <summary>
+        /// Sets the number of flushes that should be skipped for subsequent command buffers.
+        /// </summary>
+        /// <remarks>
+        /// This can improve performance when command buffer data only needs to be consumed by the GPU.
+        /// </remarks>
+        /// <param name="count">The amount of flushes that should be skipped</param>
+        internal void SetFlushSkips(int count)
+        {
+            _flushSkips = count;
+        }
+
+        /// <summary>
         /// Interrupts command processing. This will break out of the DispatchCalls loop.
         /// </summary>
         public void Interrupt()
diff --git a/Ryujinx.Graphics.Gpu/Engine/GPFifo/GPFifoProcessor.cs b/Ryujinx.Graphics.Gpu/Engine/GPFifo/GPFifoProcessor.cs
index ea34d6cd..096b795c 100644
--- a/Ryujinx.Graphics.Gpu/Engine/GPFifo/GPFifoProcessor.cs
+++ b/Ryujinx.Graphics.Gpu/Engine/GPFifo/GPFifoProcessor.cs
@@ -29,6 +29,11 @@ namespace Ryujinx.Graphics.Gpu.Engine.GPFifo
         public MemoryManager MemoryManager => _channel.MemoryManager;
 
         /// <summary>
+        /// 3D Engine.
+        /// </summary>
+        public ThreedClass ThreedClass => _3dClass;
+
+        /// <summary>
         /// Internal GPFIFO state.
         /// </summary>
         private struct DmaState
@@ -70,13 +75,16 @@ namespace Ryujinx.Graphics.Gpu.Engine.GPFifo
         /// <summary>
         /// Processes a command buffer.
         /// </summary>
+        /// <param name="baseGpuVa">Base GPU virtual address of the command buffer</param>
         /// <param name="commandBuffer">Command buffer</param>
-        public void Process(ReadOnlySpan<int> commandBuffer)
+        public void Process(ulong baseGpuVa, ReadOnlySpan<int> commandBuffer)
         {
             for (int index = 0; index < commandBuffer.Length; index++)
             {
                 int command = commandBuffer[index];
 
+                ulong gpuVa = baseGpuVa + (ulong)index * 4;
+
                 if (_state.MethodCount != 0)
                 {
                     if (TryFastI2mBufferUpdate(commandBuffer, ref index))
@@ -84,7 +92,7 @@ namespace Ryujinx.Graphics.Gpu.Engine.GPFifo
                         continue;
                     }
 
-                    Send(_state.Method, command, _state.SubChannel, _state.MethodCount <= 1);
+                    Send(gpuVa, _state.Method, command, _state.SubChannel, _state.MethodCount <= 1);
 
                     if (!_state.NonIncrementing)
                     {
@@ -120,7 +128,7 @@ namespace Ryujinx.Graphics.Gpu.Engine.GPFifo
                             _state.NonIncrementing = meth.SecOp == SecOp.NonIncMethod;
                             break;
                         case SecOp.ImmdDataMethod:
-                            Send(meth.MethodAddress, meth.ImmdData, meth.MethodSubchannel, true);
+                            Send(gpuVa, meth.MethodAddress, meth.ImmdData, meth.MethodSubchannel, true);
                             break;
                     }
                 }
@@ -198,8 +206,9 @@ namespace Ryujinx.Graphics.Gpu.Engine.GPFifo
         /// <summary>
         /// Sends a uncompressed method for processing by the graphics pipeline.
         /// </summary>
+        /// <param name="gpuVa">GPU virtual address where the command word is located</param>
         /// <param name="meth">Method to be processed</param>
-        private void Send(int offset, int argument, int subChannel, bool isLastCall)
+        private void Send(ulong gpuVa, int offset, int argument, int subChannel, bool isLastCall)
         {
             if (offset < 0x60)
             {
@@ -243,7 +252,7 @@ namespace Ryujinx.Graphics.Gpu.Engine.GPFifo
 
                     if ((offset & 1) != 0)
                     {
-                        _fifoClass.MmePushArgument(macroIndex, argument);
+                        _fifoClass.MmePushArgument(macroIndex, gpuVa, argument);
                     }
                     else
                     {