Merge branch 'gpu/dma-apple' into 'master'

Gpu: exclude RGBA16Float texture format for fast DMA copy on Apple silicon.

See merge request [ryubing/ryujinx!206](https://git.ryujinx.app/ryubing/ryujinx/-/merge_requests/206)

src/Ryujinx.Graphics.GAL/Capabilities.cs

@@ -27,6 +27,7 @@ namespace Ryujinx.Graphics.GAL
         public readonly bool SupportsSparseBuffer;
         public readonly bool Supports5BitComponentFormat;
         public readonly bool SupportsBlendEquationAdvanced;
+        public readonly bool SupportsFastDmaTextureCopy;
         public readonly bool SupportsFragmentShaderInterlock;
         public readonly bool SupportsFragmentShaderOrderingIntel;
         public readonly bool SupportsGeometryShader;
@@ -95,6 +96,7 @@ namespace Ryujinx.Graphics.GAL
             bool supports5BitComponentFormat,
             bool supportsSparseBuffer,
             bool supportsBlendEquationAdvanced,
+            bool supportsFastDmaTextureCopy,
             bool supportsFragmentShaderInterlock,
             bool supportsFragmentShaderOrderingIntel,
             bool supportsGeometryShader,
@@ -157,6 +159,7 @@ namespace Ryujinx.Graphics.GAL
             Supports5BitComponentFormat = supports5BitComponentFormat;
             SupportsSparseBuffer = supportsSparseBuffer;
             SupportsBlendEquationAdvanced = supportsBlendEquationAdvanced;
+            SupportsFastDmaTextureCopy = supportsFastDmaTextureCopy;
             SupportsFragmentShaderInterlock = supportsFragmentShaderInterlock;
             SupportsFragmentShaderOrderingIntel = supportsFragmentShaderOrderingIntel;
             SupportsGeometryShader = supportsGeometryShader;

src/Ryujinx.Graphics.Gpu/Engine/Dma/DmaClass.cs

@@ -1,6 +1,7 @@
 using Ryujinx.Common;
 using Ryujinx.Common.Memory;
 using Ryujinx.Graphics.Device;
+using Ryujinx.Graphics.GAL;
 using Ryujinx.Graphics.Gpu.Engine.Threed;
 using Ryujinx.Graphics.Gpu.Memory;
 using Ryujinx.Graphics.Texture;
@@ -287,9 +288,9 @@ namespace Ryujinx.Graphics.Gpu.Engine.Dma
                 bool completeSource = IsTextureCopyComplete(src, srcLinear, srcBpp, srcStride, xCount, yCount);
                 bool completeDest = IsTextureCopyComplete(dst, dstLinear, dstBpp, dstStride, xCount, yCount);
 
-                // Check if the source texture exists on the GPU, if it does, do a GPU side copy.
+                // Check if the source texture exists on the GPU, if it does, do a GPU-side copy.
                 // Otherwise, we would need to flush the source texture which is costly.
-                // We don't expect the source to be linear in such cases, as linear source usually indicates buffer or CPU written data.
+                // We don't expect the source to be linear in such cases, as linear source usually indicates buffer or CPU-written data.
 
                 if (completeSource && completeDest && !srcLinear && isIdentityRemap)
                 {
@@ -307,27 +308,35 @@ namespace Ryujinx.Graphics.Gpu.Engine.Dma
 
                     if (source != null && source.Height == yCount)
                     {
-                        source.SynchronizeMemory();
+                        // HACK: Exclude RGBA16Float texture format for fast DMA copy on Apple Silicon.
+                        // Fixes Sonic Frontiers when VK_EXT_external_memory_host is not available.
+                        bool skipDma = !_context.Capabilities.SupportsFastDmaTextureCopy &&
+                                       source.Info.FormatInfo.Format == Format.R16G16B16A16Float;
 
-                        Image.Texture target = memoryManager.Physical.TextureCache.FindOrCreateTexture(
+                        if (!skipDma)
-                            memoryManager,
-                            source.Info.FormatInfo,
-                            dstGpuVa,
-                            xCount,
-                            yCount,
-                            dstStride,
-                            dstLinear,
-                            dst.MemoryLayout.UnpackGobBlocksInY(),
-                            dst.MemoryLayout.UnpackGobBlocksInZ());
-
-                        if (source.ScaleFactor != target.ScaleFactor)
                         {
-                            target.PropagateScale(source);
+                            source.SynchronizeMemory();
-                        }
 
-                        source.HostTexture.CopyTo(target.HostTexture, 0, 0);
+                            Image.Texture target = memoryManager.Physical.TextureCache.FindOrCreateTexture(
-                        target.SignalModified();
+                                memoryManager,
-                        return;
+                                source.Info.FormatInfo,
+                                dstGpuVa,
+                                xCount,
+                                yCount,
+                                dstStride,
+                                dstLinear,
+                                dst.MemoryLayout.UnpackGobBlocksInY(),
+                                dst.MemoryLayout.UnpackGobBlocksInZ());
+
+                            if (source.ScaleFactor != target.ScaleFactor)
+                            {
+                                target.PropagateScale(source);
+                            }
+
+                            source.HostTexture.CopyTo(target.HostTexture, 0, 0);
+                            target.SignalModified();
+                            return;
+                        }
                     }
                 }
 

src/Ryujinx.Graphics.OpenGL/OpenGLRenderer.cs

@@ -169,6 +169,7 @@ namespace Ryujinx.Graphics.OpenGL
                 supports5BitComponentFormat: true,
                 supportsSparseBuffer: false,
                 supportsBlendEquationAdvanced: HwCapabilities.SupportsBlendEquationAdvanced,
+                supportsFastDmaTextureCopy: true,
                 supportsFragmentShaderInterlock: HwCapabilities.SupportsFragmentShaderInterlock,
                 supportsFragmentShaderOrderingIntel: HwCapabilities.SupportsFragmentShaderOrdering,
                 supportsGeometryShader: true,

src/Ryujinx.Graphics.Vulkan/VulkanRenderer.cs

@@ -760,6 +760,7 @@ namespace Ryujinx.Graphics.Vulkan
                 supports5BitComponentFormat: supports5BitComponentFormat,
                 supportsSparseBuffer: features2.Features.SparseBinding && mainQueueProperties.QueueFlags.HasFlag(QueueFlags.SparseBindingBit),
                 supportsBlendEquationAdvanced: Capabilities.SupportsBlendEquationAdvanced,
+                supportsFastDmaTextureCopy: !IsMoltenVk,
                 supportsFragmentShaderInterlock: Capabilities.SupportsFragmentShaderInterlock,
                 supportsFragmentShaderOrderingIntel: false,
                 supportsGeometryShader: Capabilities.SupportsGeometryShader,