From c3a5716a95ea93cba9488189fb36d594db5083bc Mon Sep 17 00:00:00 2001 From: gdkchan Date: Tue, 21 Feb 2023 19:21:57 -0300 Subject: [PATCH] Add copy dependency for some incompatible texture formats (#4380) * Add copy dependency for some incompatible texture formats * Simplify compatibility check --- .../Image/TextureCompatibility.cs | 103 ++++--- .../Image/TextureCopyIncompatible.cs | 252 +++++++++++++++++ Ryujinx.Graphics.OpenGL/Image/TextureView.cs | 10 + Ryujinx.Graphics.OpenGL/OpenGLRenderer.cs | 2 + Ryujinx.Graphics.Vulkan/HelperShader.cs | 158 ++++++++++- ...olorCopyShorteningComputeShaderSource.comp | 36 +++ .../ColorCopyWideningComputeShaderSource.comp | 31 +++ .../Shaders/ShaderBinaries.cs | 259 ++++++++++++++++++ Ryujinx.Graphics.Vulkan/TextureView.cs | 10 + 9 files changed, 814 insertions(+), 47 deletions(-) create mode 100644 Ryujinx.Graphics.OpenGL/Image/TextureCopyIncompatible.cs create mode 100644 Ryujinx.Graphics.Vulkan/Shaders/ColorCopyShorteningComputeShaderSource.comp create mode 100644 Ryujinx.Graphics.Vulkan/Shaders/ColorCopyWideningComputeShaderSource.comp diff --git a/Ryujinx.Graphics.Gpu/Image/TextureCompatibility.cs b/Ryujinx.Graphics.Gpu/Image/TextureCompatibility.cs index e8061951be..4b84333df8 100644 --- a/Ryujinx.Graphics.Gpu/Image/TextureCompatibility.cs +++ b/Ryujinx.Graphics.Gpu/Image/TextureCompatibility.cs @@ -214,41 +214,6 @@ namespace Ryujinx.Graphics.Gpu.Image return true; } - /// - /// Checks if two formats are compatible, according to the host API copy format compatibility rules. - /// - /// First comparand - /// Second comparand - /// Host GPU capabilities - /// True if the formats are compatible, false otherwise - public static bool FormatCompatible(TextureInfo lhs, TextureInfo rhs, Capabilities caps) - { - FormatInfo lhsFormat = lhs.FormatInfo; - FormatInfo rhsFormat = rhs.FormatInfo; - - if (lhsFormat.Format.IsDepthOrStencil() || rhsFormat.Format.IsDepthOrStencil()) - { - return lhsFormat.Format == rhsFormat.Format; - } - - if (IsFormatHostIncompatible(lhs, caps) || IsFormatHostIncompatible(rhs, caps)) - { - return lhsFormat.Format == rhsFormat.Format; - } - - if (lhsFormat.IsCompressed && rhsFormat.IsCompressed) - { - FormatClass lhsClass = GetFormatClass(lhsFormat.Format); - FormatClass rhsClass = GetFormatClass(rhsFormat.Format); - - return lhsClass == rhsClass; - } - else - { - return lhsFormat.BytesPerPixel == rhsFormat.BytesPerPixel; - } - } - /// /// Checks if the texture format matches with the specified texture information. /// @@ -391,6 +356,13 @@ namespace Ryujinx.Graphics.Gpu.Image Size lhsSize = GetSizeInBlocks(lhs, level); Size rhsSize = GetSizeInBlocks(rhs); + bool alignedWidthMatches = lhsAlignedSize.Width == rhsAlignedSize.Width; + + if (lhs.FormatInfo.BytesPerPixel != rhs.FormatInfo.BytesPerPixel && IsIncompatibleFormatAliasingAllowed(lhs.FormatInfo, rhs.FormatInfo)) + { + alignedWidthMatches = lhsSize.Width * lhs.FormatInfo.BytesPerPixel == rhsSize.Width * rhs.FormatInfo.BytesPerPixel; + } + TextureViewCompatibility result = TextureViewCompatibility.Full; // For copies, we can copy a subset of the 3D texture slices, @@ -404,7 +376,7 @@ namespace Ryujinx.Graphics.Gpu.Image // so the width may not match in this case for different uses of the same texture. // To account for this, we compare the aligned width here. // We expect height to always match exactly, if the texture is the same. - if (lhsAlignedSize.Width == rhsAlignedSize.Width && lhsSize.Height == rhsSize.Height) + if (alignedWidthMatches && lhsSize.Height == rhsSize.Height) { return (exact && lhsSize.Width != rhsSize.Width) || lhsSize.Width < rhsSize.Width ? TextureViewCompatibility.CopyOnly @@ -659,21 +631,62 @@ namespace Ryujinx.Graphics.Gpu.Image /// The view compatibility level of the texture formats public static TextureViewCompatibility ViewFormatCompatible(TextureInfo lhs, TextureInfo rhs, Capabilities caps) { - if (FormatCompatible(lhs, rhs, caps)) + FormatInfo lhsFormat = lhs.FormatInfo; + FormatInfo rhsFormat = rhs.FormatInfo; + + if (lhsFormat.Format.IsDepthOrStencil() || rhsFormat.Format.IsDepthOrStencil()) { - if (lhs.FormatInfo.IsCompressed != rhs.FormatInfo.IsCompressed) - { - return TextureViewCompatibility.CopyOnly; - } - else - { - return TextureViewCompatibility.Full; - } + return lhsFormat.Format == rhsFormat.Format ? TextureViewCompatibility.Full : TextureViewCompatibility.Incompatible; + } + + if (IsFormatHostIncompatible(lhs, caps) || IsFormatHostIncompatible(rhs, caps)) + { + return lhsFormat.Format == rhsFormat.Format ? TextureViewCompatibility.Full : TextureViewCompatibility.Incompatible; + } + + if (lhsFormat.IsCompressed && rhsFormat.IsCompressed) + { + FormatClass lhsClass = GetFormatClass(lhsFormat.Format); + FormatClass rhsClass = GetFormatClass(rhsFormat.Format); + + return lhsClass == rhsClass ? TextureViewCompatibility.Full : TextureViewCompatibility.Incompatible; + } + else if (lhsFormat.BytesPerPixel == rhsFormat.BytesPerPixel) + { + return lhs.FormatInfo.IsCompressed == rhs.FormatInfo.IsCompressed + ? TextureViewCompatibility.Full + : TextureViewCompatibility.CopyOnly; + } + else if (IsIncompatibleFormatAliasingAllowed(lhsFormat, rhsFormat)) + { + return TextureViewCompatibility.CopyOnly; } return TextureViewCompatibility.Incompatible; } + /// + /// Checks if aliasing of two formats that would normally be considered incompatible be allowed, + /// using copy dependencies. + /// + /// Format information of the first textureFormat information of the second texture + /// True if aliasing should be allowed, false otherwise + private static bool IsIncompatibleFormatAliasingAllowed(FormatInfo lhsFormat, FormatInfo rhsFormat) + { + // Some games will try to alias textures with incompatible foramts, with different BPP (bytes per pixel). + // We allow that in some cases as long Width * BPP is equal on both textures. + // This is very conservative right now as we want to avoid copies as much as possible, + // so we only consider the formats we have seen being aliased. + + if (rhsFormat.BytesPerPixel < lhsFormat.BytesPerPixel) + { + (lhsFormat, rhsFormat) = (rhsFormat, lhsFormat); + } + + return lhsFormat.Format == Format.R8Unorm && rhsFormat.Format == Format.R8G8B8A8Unorm; + } + /// /// Check if the target of the first texture view information is compatible with the target of the second texture view information. /// This follows the host API target compatibility rules. diff --git a/Ryujinx.Graphics.OpenGL/Image/TextureCopyIncompatible.cs b/Ryujinx.Graphics.OpenGL/Image/TextureCopyIncompatible.cs new file mode 100644 index 0000000000..c8fbfbc6a6 --- /dev/null +++ b/Ryujinx.Graphics.OpenGL/Image/TextureCopyIncompatible.cs @@ -0,0 +1,252 @@ +using OpenTK.Graphics.OpenGL; +using Ryujinx.Graphics.GAL; +using System; +using System.Collections.Generic; +using System.Globalization; +using System.Numerics; + +namespace Ryujinx.Graphics.OpenGL.Image +{ + class TextureCopyIncompatible + { + private const string ComputeShaderShortening = @"#version 450 core + +layout (binding = 0, $SRC_FORMAT$) uniform uimage2D src; +layout (binding = 1, $DST_FORMAT$) uniform uimage2D dst; + +layout (local_size_x = 32, local_size_y = 32, local_size_z = 1) in; + +void main() +{ + uvec2 coords = gl_GlobalInvocationID.xy; + ivec2 imageSz = imageSize(src); + + if (int(coords.x) >= imageSz.x || int(coords.y) >= imageSz.y) + { + return; + } + + uint coordsShifted = coords.x << $RATIO_LOG2$; + + uvec2 dstCoords0 = uvec2(coordsShifted, coords.y); + uvec2 dstCoords1 = uvec2(coordsShifted + 1, coords.y); + uvec2 dstCoords2 = uvec2(coordsShifted + 2, coords.y); + uvec2 dstCoords3 = uvec2(coordsShifted + 3, coords.y); + + uvec4 rgba = imageLoad(src, ivec2(coords)); + + imageStore(dst, ivec2(dstCoords0), rgba.rrrr); + imageStore(dst, ivec2(dstCoords1), rgba.gggg); + imageStore(dst, ivec2(dstCoords2), rgba.bbbb); + imageStore(dst, ivec2(dstCoords3), rgba.aaaa); +}"; + + private const string ComputeShaderWidening = @"#version 450 core + +layout (binding = 0, $SRC_FORMAT$) uniform uimage2D src; +layout (binding = 1, $DST_FORMAT$) uniform uimage2D dst; + +layout (local_size_x = 32, local_size_y = 32, local_size_z = 1) in; + +void main() +{ + uvec2 coords = gl_GlobalInvocationID.xy; + ivec2 imageSz = imageSize(dst); + + if (int(coords.x) >= imageSz.x || int(coords.y) >= imageSz.y) + { + return; + } + + uvec2 srcCoords = uvec2(coords.x << $RATIO_LOG2$, coords.y); + + uint r = imageLoad(src, ivec2(srcCoords) + ivec2(0, 0)).r; + uint g = imageLoad(src, ivec2(srcCoords) + ivec2(1, 0)).r; + uint b = imageLoad(src, ivec2(srcCoords) + ivec2(2, 0)).r; + uint a = imageLoad(src, ivec2(srcCoords) + ivec2(3, 0)).r; + + imageStore(dst, ivec2(coords), uvec4(r, g, b, a)); +}"; + + private readonly OpenGLRenderer _renderer; + private readonly Dictionary _shorteningProgramHandles; + private readonly Dictionary _wideningProgramHandles; + + public TextureCopyIncompatible(OpenGLRenderer renderer) + { + _renderer = renderer; + _shorteningProgramHandles = new Dictionary(); + _wideningProgramHandles = new Dictionary(); + } + + public void CopyIncompatibleFormats(ITextureInfo src, ITextureInfo dst, int srcLayer, int dstLayer, int srcLevel, int dstLevel, int depth, int levels) + { + TextureCreateInfo srcInfo = src.Info; + TextureCreateInfo dstInfo = dst.Info; + + int srcBpp = src.Info.BytesPerPixel; + int dstBpp = dst.Info.BytesPerPixel; + + // Calculate ideal component size, given our constraints: + // - Component size must not exceed bytes per pixel of source and destination image formats. + // - Maximum component size is 4 (R32). + int componentSize = Math.Min(Math.Min(srcBpp, dstBpp), 4); + + int srcComponentsCount = srcBpp / componentSize; + int dstComponentsCount = dstBpp / componentSize; + + var srcFormat = GetFormat(componentSize, srcComponentsCount); + var dstFormat = GetFormat(componentSize, dstComponentsCount); + + GL.UseProgram(srcBpp < dstBpp + ? GetWideningShader(componentSize, srcComponentsCount, dstComponentsCount) + : GetShorteningShader(componentSize, srcComponentsCount, dstComponentsCount)); + + for (int l = 0; l < levels; l++) + { + int srcWidth = Math.Max(1, src.Info.Width >> l); + int srcHeight = Math.Max(1, src.Info.Height >> l); + + int dstWidth = Math.Max(1, dst.Info.Width >> l); + int dstHeight = Math.Max(1, dst.Info.Height >> l); + + int width = Math.Min(srcWidth, dstWidth); + int height = Math.Min(srcHeight, dstHeight); + + for (int z = 0; z < depth; z++) + { + GL.BindImageTexture(0, src.Handle, srcLevel + l, false, srcLayer + z, TextureAccess.ReadOnly, srcFormat); + GL.BindImageTexture(1, dst.Handle, dstLevel + l, false, dstLayer + z, TextureAccess.WriteOnly, dstFormat); + + GL.DispatchCompute((width + 31) / 32, (height + 31) / 32, 1); + } + } + + Pipeline pipeline = (Pipeline)_renderer.Pipeline; + + pipeline.RestoreProgram(); + pipeline.RestoreImages1And2(); + } + + private static SizedInternalFormat GetFormat(int componentSize, int componentsCount) + { + if (componentSize == 1) + { + return componentsCount switch + { + 1 => SizedInternalFormat.R8ui, + 2 => SizedInternalFormat.Rg8ui, + 4 => SizedInternalFormat.Rgba8ui, + _ => throw new ArgumentException($"Invalid components count {componentsCount}.") + }; + } + else if (componentSize == 2) + { + return componentsCount switch + { + 1 => SizedInternalFormat.R16ui, + 2 => SizedInternalFormat.Rg16ui, + 4 => SizedInternalFormat.Rgba16ui, + _ => throw new ArgumentException($"Invalid components count {componentsCount}.") + }; + } + else if (componentSize == 4) + { + return componentsCount switch + { + 1 => SizedInternalFormat.R32ui, + 2 => SizedInternalFormat.Rg32ui, + 4 => SizedInternalFormat.Rgba32ui, + _ => throw new ArgumentException($"Invalid components count {componentsCount}.") + }; + } + else + { + throw new ArgumentException($"Invalid component size {componentSize}."); + } + } + + private int GetShorteningShader(int componentSize, int srcComponentsCount, int dstComponentsCount) + { + return GetShader(ComputeShaderShortening, _shorteningProgramHandles, componentSize, srcComponentsCount, dstComponentsCount); + } + + private int GetWideningShader(int componentSize, int srcComponentsCount, int dstComponentsCount) + { + return GetShader(ComputeShaderWidening, _wideningProgramHandles, componentSize, srcComponentsCount, dstComponentsCount); + } + + private int GetShader( + string code, + Dictionary programHandles, + int componentSize, + int srcComponentsCount, + int dstComponentsCount) + { + int componentSizeLog2 = BitOperations.Log2((uint)componentSize); + + int srcIndex = componentSizeLog2 + BitOperations.Log2((uint)srcComponentsCount) * 3; + int dstIndex = componentSizeLog2 + BitOperations.Log2((uint)dstComponentsCount) * 3; + + int key = srcIndex | (dstIndex << 8); + + if (!programHandles.TryGetValue(key, out int programHandle)) + { + int csHandle = GL.CreateShader(ShaderType.ComputeShader); + + string[] formatTable = new[] { "r8ui", "r16ui", "r32ui", "rg8ui", "rg16ui", "rg32ui", "rgba8ui", "rgba16ui", "rgba32ui" }; + + string srcFormat = formatTable[srcIndex]; + string dstFormat = formatTable[dstIndex]; + + int srcBpp = srcComponentsCount * componentSize; + int dstBpp = dstComponentsCount * componentSize; + + int ratio = srcBpp < dstBpp ? dstBpp / srcBpp : srcBpp / dstBpp; + int ratioLog2 = BitOperations.Log2((uint)ratio); + + GL.ShaderSource(csHandle, code + .Replace("$SRC_FORMAT$", srcFormat) + .Replace("$DST_FORMAT$", dstFormat) + .Replace("$RATIO_LOG2$", ratioLog2.ToString(CultureInfo.InvariantCulture))); + + GL.CompileShader(csHandle); + + programHandle = GL.CreateProgram(); + + GL.AttachShader(programHandle, csHandle); + GL.LinkProgram(programHandle); + GL.DetachShader(programHandle, csHandle); + GL.DeleteShader(csHandle); + + GL.GetProgram(programHandle, GetProgramParameterName.LinkStatus, out int status); + + if (status == 0) + { + throw new Exception(GL.GetProgramInfoLog(programHandle)); + } + + programHandles.Add(key, programHandle); + } + + return programHandle; + } + + public void Dispose() + { + foreach (int handle in _shorteningProgramHandles.Values) + { + GL.DeleteProgram(handle); + } + + _shorteningProgramHandles.Clear(); + + foreach (int handle in _wideningProgramHandles.Values) + { + GL.DeleteProgram(handle); + } + + _wideningProgramHandles.Clear(); + } + } +} diff --git a/Ryujinx.Graphics.OpenGL/Image/TextureView.cs b/Ryujinx.Graphics.OpenGL/Image/TextureView.cs index 68cd2d30f1..44df441f71 100644 --- a/Ryujinx.Graphics.OpenGL/Image/TextureView.cs +++ b/Ryujinx.Graphics.OpenGL/Image/TextureView.cs @@ -127,6 +127,12 @@ namespace Ryujinx.Graphics.OpenGL.Image int layers = Math.Min(Info.GetLayers(), destinationView.Info.GetLayers() - firstLayer); _renderer.TextureCopyMS.CopyNonMSToMS(this, destinationView, 0, firstLayer, layers); } + else if (destinationView.Info.BytesPerPixel != Info.BytesPerPixel) + { + int layers = Math.Min(Info.GetLayers(), destinationView.Info.GetLayers() - firstLayer); + int levels = Math.Min(Info.Levels, destinationView.Info.Levels - firstLevel); + _renderer.TextureCopyIncompatible.CopyIncompatibleFormats(this, destinationView, 0, firstLayer, 0, firstLevel, layers, levels); + } else { _renderer.TextureCopy.CopyUnscaled(this, destinationView, 0, firstLayer, 0, firstLevel); @@ -145,6 +151,10 @@ namespace Ryujinx.Graphics.OpenGL.Image { _renderer.TextureCopyMS.CopyNonMSToMS(this, destinationView, srcLayer, dstLayer, 1); } + else if (destinationView.Info.BytesPerPixel != Info.BytesPerPixel) + { + _renderer.TextureCopyIncompatible.CopyIncompatibleFormats(this, destinationView, srcLayer, dstLayer, srcLevel, dstLevel, 1, 1); + } else { _renderer.TextureCopy.CopyUnscaled(this, destinationView, srcLayer, dstLayer, srcLevel, dstLevel, 1, 1); diff --git a/Ryujinx.Graphics.OpenGL/OpenGLRenderer.cs b/Ryujinx.Graphics.OpenGL/OpenGLRenderer.cs index 722c4b4da6..efbd17c1b8 100644 --- a/Ryujinx.Graphics.OpenGL/OpenGLRenderer.cs +++ b/Ryujinx.Graphics.OpenGL/OpenGLRenderer.cs @@ -24,6 +24,7 @@ namespace Ryujinx.Graphics.OpenGL private TextureCopy _textureCopy; private TextureCopy _backgroundTextureCopy; internal TextureCopy TextureCopy => BackgroundContextWorker.InBackground ? _backgroundTextureCopy : _textureCopy; + internal TextureCopyIncompatible TextureCopyIncompatible { get; } internal TextureCopyMS TextureCopyMS { get; } private Sync _sync; @@ -49,6 +50,7 @@ namespace Ryujinx.Graphics.OpenGL _window = new Window(this); _textureCopy = new TextureCopy(this); _backgroundTextureCopy = new TextureCopy(this); + TextureCopyIncompatible = new TextureCopyIncompatible(this); TextureCopyMS = new TextureCopyMS(this); _sync = new Sync(); PersistentBuffers = new PersistentBuffers(); diff --git a/Ryujinx.Graphics.Vulkan/HelperShader.cs b/Ryujinx.Graphics.Vulkan/HelperShader.cs index b8c21fe8e0..c67389aa40 100644 --- a/Ryujinx.Graphics.Vulkan/HelperShader.cs +++ b/Ryujinx.Graphics.Vulkan/HelperShader.cs @@ -5,6 +5,7 @@ using Ryujinx.Graphics.Vulkan.Shaders; using Silk.NET.Vulkan; using System; using System.Collections.Generic; +using System.Numerics; using VkFormat = Silk.NET.Vulkan.Format; namespace Ryujinx.Graphics.Vulkan @@ -32,7 +33,9 @@ namespace Ryujinx.Graphics.Vulkan private readonly IProgram _programStrideChange; private readonly IProgram _programConvertIndexBuffer; private readonly IProgram _programConvertIndirectData; + private readonly IProgram _programColorCopyShortening; private readonly IProgram _programColorCopyToNonMs; + private readonly IProgram _programColorCopyWidening; private readonly IProgram _programColorDrawToMs; private readonly IProgram _programDepthBlit; private readonly IProgram _programDepthBlitMs; @@ -112,15 +115,25 @@ namespace Ryujinx.Graphics.Vulkan new ShaderSource(ShaderBinaries.ChangeBufferStrideShaderSource, strideChangeBindings, ShaderStage.Compute, TargetLanguage.Spirv), }); - var colorCopyToNonMsBindings = new ShaderBindings( + var colorCopyBindings = new ShaderBindings( new[] { 0 }, Array.Empty(), new[] { 0 }, new[] { 0 }); + _programColorCopyShortening = gd.CreateProgramWithMinimalLayout(new[] + { + new ShaderSource(ShaderBinaries.ColorCopyShorteningComputeShaderSource, colorCopyBindings, ShaderStage.Compute, TargetLanguage.Spirv), + }); + _programColorCopyToNonMs = gd.CreateProgramWithMinimalLayout(new[] { - new ShaderSource(ShaderBinaries.ColorCopyToNonMsComputeShaderSource, colorCopyToNonMsBindings, ShaderStage.Compute, TargetLanguage.Spirv), + new ShaderSource(ShaderBinaries.ColorCopyToNonMsComputeShaderSource, colorCopyBindings, ShaderStage.Compute, TargetLanguage.Spirv), + }); + + _programColorCopyWidening = gd.CreateProgramWithMinimalLayout(new[] + { + new ShaderSource(ShaderBinaries.ColorCopyWideningComputeShaderSource, colorCopyBindings, ShaderStage.Compute, TargetLanguage.Spirv), }); var colorDrawToMsVertexBindings = new ShaderBindings( @@ -922,6 +935,107 @@ namespace Ryujinx.Graphics.Vulkan convertedCount * outputIndexSize); } + public void CopyIncompatibleFormats( + VulkanRenderer gd, + CommandBufferScoped cbs, + TextureView src, + TextureView dst, + int srcLayer, + int dstLayer, + int srcLevel, + int dstLevel, + int depth, + int levels) + { + const int ParamsBufferSize = 4; + + Span shaderParams = stackalloc int[sizeof(int)]; + + int srcBpp = src.Info.BytesPerPixel; + int dstBpp = dst.Info.BytesPerPixel; + + int ratio = srcBpp < dstBpp ? dstBpp / srcBpp : srcBpp / dstBpp; + + shaderParams[0] = BitOperations.Log2((uint)ratio); + + var bufferHandle = gd.BufferManager.CreateWithHandle(gd, ParamsBufferSize, false); + + gd.BufferManager.SetData(bufferHandle, 0, shaderParams); + + TextureView.InsertImageBarrier( + gd.Api, + cbs.CommandBuffer, + src.GetImage().Get(cbs).Value, + TextureStorage.DefaultAccessMask, + AccessFlags.ShaderReadBit, + PipelineStageFlags.AllCommandsBit, + PipelineStageFlags.ComputeShaderBit, + ImageAspectFlags.ColorBit, + src.FirstLayer + srcLayer, + src.FirstLevel + srcLevel, + depth, + levels); + + _pipeline.SetCommandBuffer(cbs); + + _pipeline.SetProgram(srcBpp < dstBpp ? _programColorCopyWidening : _programColorCopyShortening); + + // Calculate ideal component size, given our constraints: + // - Component size must not exceed bytes per pixel of source and destination image formats. + // - Maximum component size is 4 (R32). + int componentSize = Math.Min(Math.Min(srcBpp, dstBpp), 4); + + var srcFormat = GetFormat(componentSize, srcBpp / componentSize); + var dstFormat = GetFormat(componentSize, dstBpp / componentSize); + + _pipeline.SetUniformBuffers(stackalloc[] { new BufferAssignment(0, new BufferRange(bufferHandle, 0, ParamsBufferSize)) }); + + for (int l = 0; l < levels; l++) + { + for (int z = 0; z < depth; z++) + { + var srcView = Create2DLayerView(src, srcLayer + z, srcLevel + l, srcFormat); + var dstView = Create2DLayerView(dst, dstLayer + z, dstLevel + l); + + _pipeline.SetTextureAndSampler(ShaderStage.Compute, 0, srcView, null); + _pipeline.SetImage(0, dstView, dstFormat); + + int dispatchX = (Math.Min(srcView.Info.Width, dstView.Info.Width) + 31) / 32; + int dispatchY = (Math.Min(srcView.Info.Height, dstView.Info.Height) + 31) / 32; + + _pipeline.DispatchCompute(dispatchX, dispatchY, 1); + + if (srcView != src) + { + srcView.Release(); + } + + if (dstView != dst) + { + dstView.Release(); + } + } + } + + gd.BufferManager.Delete(bufferHandle); + + _pipeline.Finish(gd, cbs); + + TextureView.InsertImageBarrier( + gd.Api, + cbs.CommandBuffer, + dst.GetImage().Get(cbs).Value, + AccessFlags.ShaderWriteBit, + TextureStorage.DefaultAccessMask, + PipelineStageFlags.ComputeShaderBit, + PipelineStageFlags.AllCommandsBit, + ImageAspectFlags.ColorBit, + dst.FirstLayer + dstLayer, + dst.FirstLevel + dstLevel, + depth, + levels); + } + public void CopyMSToNonMS(VulkanRenderer gd, CommandBufferScoped cbs, TextureView src, TextureView dst, int srcLayer, int dstLayer, int depth) { const int ParamsBufferSize = 16; @@ -1196,6 +1310,44 @@ namespace Ryujinx.Graphics.Vulkan }; } + private static GAL.Format GetFormat(int componentSize, int componentsCount) + { + if (componentSize == 1) + { + return componentsCount switch + { + 1 => GAL.Format.R8Uint, + 2 => GAL.Format.R8G8Uint, + 4 => GAL.Format.R8G8B8A8Uint, + _ => throw new ArgumentException($"Invalid components count {componentsCount}.") + }; + } + else if (componentSize == 2) + { + return componentsCount switch + { + 1 => GAL.Format.R16Uint, + 2 => GAL.Format.R16G16Uint, + 4 => GAL.Format.R16G16B16A16Uint, + _ => throw new ArgumentException($"Invalid components count {componentsCount}.") + }; + } + else if (componentSize == 4) + { + return componentsCount switch + { + 1 => GAL.Format.R32Uint, + 2 => GAL.Format.R32G32Uint, + 4 => GAL.Format.R32G32B32A32Uint, + _ => throw new ArgumentException($"Invalid components count {componentsCount}.") + }; + } + else + { + throw new ArgumentException($"Invalid component size {componentSize}."); + } + } + public void ConvertIndexBufferIndirect( VulkanRenderer gd, CommandBufferScoped cbs, @@ -1336,7 +1488,9 @@ namespace Ryujinx.Graphics.Vulkan _programStrideChange.Dispose(); _programConvertIndexBuffer.Dispose(); _programConvertIndirectData.Dispose(); + _programColorCopyShortening.Dispose(); _programColorCopyToNonMs.Dispose(); + _programColorCopyWidening.Dispose(); _programColorDrawToMs.Dispose(); _programDepthBlit.Dispose(); _programDepthBlitMs.Dispose(); diff --git a/Ryujinx.Graphics.Vulkan/Shaders/ColorCopyShorteningComputeShaderSource.comp b/Ryujinx.Graphics.Vulkan/Shaders/ColorCopyShorteningComputeShaderSource.comp new file mode 100644 index 0000000000..78cc1cc6fe --- /dev/null +++ b/Ryujinx.Graphics.Vulkan/Shaders/ColorCopyShorteningComputeShaderSource.comp @@ -0,0 +1,36 @@ +#version 450 core + +layout (std140, binding = 0) uniform ratio_in +{ + int ratio; +}; + +layout (set = 2, binding = 0) uniform usampler2D src; +layout (set = 3, binding = 0) writeonly uniform uimage2D dst; + +layout (local_size_x = 32, local_size_y = 32, local_size_z = 1) in; + +void main() +{ + uvec2 coords = gl_GlobalInvocationID.xy; + ivec2 textureSz = textureSize(src, 0); + + if (int(coords.x) >= textureSz.x || int(coords.y) >= textureSz.y) + { + return; + } + + uint coordsShifted = coords.x << ratio; + + uvec2 dstCoords0 = uvec2(coordsShifted, coords.y); + uvec2 dstCoords1 = uvec2(coordsShifted + 1, coords.y); + uvec2 dstCoords2 = uvec2(coordsShifted + 2, coords.y); + uvec2 dstCoords3 = uvec2(coordsShifted + 3, coords.y); + + uvec4 rgba = texelFetch(src, ivec2(coords), 0); + + imageStore(dst, ivec2(dstCoords0), rgba.rrrr); + imageStore(dst, ivec2(dstCoords1), rgba.gggg); + imageStore(dst, ivec2(dstCoords2), rgba.bbbb); + imageStore(dst, ivec2(dstCoords3), rgba.aaaa); +} \ No newline at end of file diff --git a/Ryujinx.Graphics.Vulkan/Shaders/ColorCopyWideningComputeShaderSource.comp b/Ryujinx.Graphics.Vulkan/Shaders/ColorCopyWideningComputeShaderSource.comp new file mode 100644 index 0000000000..a9be454fab --- /dev/null +++ b/Ryujinx.Graphics.Vulkan/Shaders/ColorCopyWideningComputeShaderSource.comp @@ -0,0 +1,31 @@ +#version 450 core + +layout (std140, binding = 0) uniform ratio_in +{ + int ratio; +}; + +layout (set = 2, binding = 0) uniform usampler2D src; +layout (set = 3, binding = 0) writeonly uniform uimage2D dst; + +layout (local_size_x = 32, local_size_y = 32, local_size_z = 1) in; + +void main() +{ + uvec2 coords = gl_GlobalInvocationID.xy; + ivec2 imageSz = imageSize(dst); + + if (int(coords.x) >= imageSz.x || int(coords.y) >= imageSz.y) + { + return; + } + + uvec2 srcCoords = uvec2(coords.x << ratio, coords.y); + + uint r = texelFetchOffset(src, ivec2(srcCoords), 0, ivec2(0, 0)).r; + uint g = texelFetchOffset(src, ivec2(srcCoords), 0, ivec2(1, 0)).r; + uint b = texelFetchOffset(src, ivec2(srcCoords), 0, ivec2(2, 0)).r; + uint a = texelFetchOffset(src, ivec2(srcCoords), 0, ivec2(3, 0)).r; + + imageStore(dst, ivec2(coords), uvec4(r, g, b, a)); +} \ No newline at end of file diff --git a/Ryujinx.Graphics.Vulkan/Shaders/ShaderBinaries.cs b/Ryujinx.Graphics.Vulkan/Shaders/ShaderBinaries.cs index c9df894bc6..7fd047a23e 100644 --- a/Ryujinx.Graphics.Vulkan/Shaders/ShaderBinaries.cs +++ b/Ryujinx.Graphics.Vulkan/Shaders/ShaderBinaries.cs @@ -669,6 +669,138 @@ namespace Ryujinx.Graphics.Vulkan.Shaders 0x35, 0x00, 0x00, 0x00, 0x33, 0x00, 0x00, 0x00, 0xFD, 0x00, 0x01, 0x00, 0x38, 0x00, 0x01, 0x00, }; + public static readonly byte[] ColorCopyShorteningComputeShaderSource = new byte[] + { + 0x03, 0x02, 0x23, 0x07, 0x00, 0x05, 0x01, 0x00, 0x0B, 0x00, 0x08, 0x00, 0x79, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x11, 0x00, 0x02, 0x00, 0x01, 0x00, 0x00, 0x00, 0x11, 0x00, 0x02, 0x00, + 0x32, 0x00, 0x00, 0x00, 0x11, 0x00, 0x02, 0x00, 0x38, 0x00, 0x00, 0x00, 0x0B, 0x00, 0x06, 0x00, + 0x01, 0x00, 0x00, 0x00, 0x47, 0x4C, 0x53, 0x4C, 0x2E, 0x73, 0x74, 0x64, 0x2E, 0x34, 0x35, 0x30, + 0x00, 0x00, 0x00, 0x00, 0x0E, 0x00, 0x03, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0x0F, 0x00, 0x09, 0x00, 0x05, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x6D, 0x61, 0x69, 0x6E, + 0x00, 0x00, 0x00, 0x00, 0x0C, 0x00, 0x00, 0x00, 0x16, 0x00, 0x00, 0x00, 0x38, 0x00, 0x00, 0x00, + 0x60, 0x00, 0x00, 0x00, 0x10, 0x00, 0x06, 0x00, 0x04, 0x00, 0x00, 0x00, 0x11, 0x00, 0x00, 0x00, + 0x20, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x03, 0x00, 0x03, 0x00, + 0x02, 0x00, 0x00, 0x00, 0xC2, 0x01, 0x00, 0x00, 0x05, 0x00, 0x04, 0x00, 0x04, 0x00, 0x00, 0x00, + 0x6D, 0x61, 0x69, 0x6E, 0x00, 0x00, 0x00, 0x00, 0x05, 0x00, 0x08, 0x00, 0x0C, 0x00, 0x00, 0x00, + 0x67, 0x6C, 0x5F, 0x47, 0x6C, 0x6F, 0x62, 0x61, 0x6C, 0x49, 0x6E, 0x76, 0x6F, 0x63, 0x61, 0x74, + 0x69, 0x6F, 0x6E, 0x49, 0x44, 0x00, 0x00, 0x00, 0x05, 0x00, 0x03, 0x00, 0x16, 0x00, 0x00, 0x00, + 0x73, 0x72, 0x63, 0x00, 0x05, 0x00, 0x05, 0x00, 0x36, 0x00, 0x00, 0x00, 0x72, 0x61, 0x74, 0x69, + 0x6F, 0x5F, 0x69, 0x6E, 0x00, 0x00, 0x00, 0x00, 0x06, 0x00, 0x05, 0x00, 0x36, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x72, 0x61, 0x74, 0x69, 0x6F, 0x00, 0x00, 0x00, 0x05, 0x00, 0x03, 0x00, + 0x38, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x05, 0x00, 0x03, 0x00, 0x60, 0x00, 0x00, 0x00, + 0x64, 0x73, 0x74, 0x00, 0x47, 0x00, 0x04, 0x00, 0x0C, 0x00, 0x00, 0x00, 0x0B, 0x00, 0x00, 0x00, + 0x1C, 0x00, 0x00, 0x00, 0x47, 0x00, 0x04, 0x00, 0x16, 0x00, 0x00, 0x00, 0x22, 0x00, 0x00, 0x00, + 0x02, 0x00, 0x00, 0x00, 0x47, 0x00, 0x04, 0x00, 0x16, 0x00, 0x00, 0x00, 0x21, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x48, 0x00, 0x05, 0x00, 0x36, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x23, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x47, 0x00, 0x03, 0x00, 0x36, 0x00, 0x00, 0x00, + 0x02, 0x00, 0x00, 0x00, 0x47, 0x00, 0x04, 0x00, 0x38, 0x00, 0x00, 0x00, 0x22, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x47, 0x00, 0x04, 0x00, 0x38, 0x00, 0x00, 0x00, 0x21, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x47, 0x00, 0x04, 0x00, 0x60, 0x00, 0x00, 0x00, 0x22, 0x00, 0x00, 0x00, + 0x03, 0x00, 0x00, 0x00, 0x47, 0x00, 0x04, 0x00, 0x60, 0x00, 0x00, 0x00, 0x21, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x47, 0x00, 0x03, 0x00, 0x60, 0x00, 0x00, 0x00, 0x19, 0x00, 0x00, 0x00, + 0x47, 0x00, 0x04, 0x00, 0x76, 0x00, 0x00, 0x00, 0x0B, 0x00, 0x00, 0x00, 0x19, 0x00, 0x00, 0x00, + 0x13, 0x00, 0x02, 0x00, 0x02, 0x00, 0x00, 0x00, 0x21, 0x00, 0x03, 0x00, 0x03, 0x00, 0x00, 0x00, + 0x02, 0x00, 0x00, 0x00, 0x15, 0x00, 0x04, 0x00, 0x06, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x17, 0x00, 0x04, 0x00, 0x07, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, + 0x02, 0x00, 0x00, 0x00, 0x17, 0x00, 0x04, 0x00, 0x0A, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, + 0x03, 0x00, 0x00, 0x00, 0x20, 0x00, 0x04, 0x00, 0x0B, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0x0A, 0x00, 0x00, 0x00, 0x3B, 0x00, 0x04, 0x00, 0x0B, 0x00, 0x00, 0x00, 0x0C, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0x15, 0x00, 0x04, 0x00, 0x0F, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0x17, 0x00, 0x04, 0x00, 0x10, 0x00, 0x00, 0x00, 0x0F, 0x00, 0x00, 0x00, + 0x02, 0x00, 0x00, 0x00, 0x19, 0x00, 0x09, 0x00, 0x13, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x1B, 0x00, 0x03, 0x00, 0x14, 0x00, 0x00, 0x00, + 0x13, 0x00, 0x00, 0x00, 0x20, 0x00, 0x04, 0x00, 0x15, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x14, 0x00, 0x00, 0x00, 0x3B, 0x00, 0x04, 0x00, 0x15, 0x00, 0x00, 0x00, 0x16, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x2B, 0x00, 0x04, 0x00, 0x0F, 0x00, 0x00, 0x00, 0x18, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x14, 0x00, 0x02, 0x00, 0x1B, 0x00, 0x00, 0x00, 0x2B, 0x00, 0x04, 0x00, + 0x06, 0x00, 0x00, 0x00, 0x1C, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x2B, 0x00, 0x04, 0x00, + 0x06, 0x00, 0x00, 0x00, 0x28, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x1E, 0x00, 0x03, 0x00, + 0x36, 0x00, 0x00, 0x00, 0x0F, 0x00, 0x00, 0x00, 0x20, 0x00, 0x04, 0x00, 0x37, 0x00, 0x00, 0x00, + 0x02, 0x00, 0x00, 0x00, 0x36, 0x00, 0x00, 0x00, 0x3B, 0x00, 0x04, 0x00, 0x37, 0x00, 0x00, 0x00, + 0x38, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x20, 0x00, 0x04, 0x00, 0x39, 0x00, 0x00, 0x00, + 0x02, 0x00, 0x00, 0x00, 0x0F, 0x00, 0x00, 0x00, 0x2B, 0x00, 0x04, 0x00, 0x06, 0x00, 0x00, 0x00, + 0x4A, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x2B, 0x00, 0x04, 0x00, 0x06, 0x00, 0x00, 0x00, + 0x51, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, 0x17, 0x00, 0x04, 0x00, 0x56, 0x00, 0x00, 0x00, + 0x06, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x19, 0x00, 0x09, 0x00, 0x5E, 0x00, 0x00, 0x00, + 0x06, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x20, 0x00, 0x04, 0x00, + 0x5F, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x5E, 0x00, 0x00, 0x00, 0x3B, 0x00, 0x04, 0x00, + 0x5F, 0x00, 0x00, 0x00, 0x60, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x2B, 0x00, 0x04, 0x00, + 0x06, 0x00, 0x00, 0x00, 0x75, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00, 0x2C, 0x00, 0x06, 0x00, + 0x0A, 0x00, 0x00, 0x00, 0x76, 0x00, 0x00, 0x00, 0x75, 0x00, 0x00, 0x00, 0x75, 0x00, 0x00, 0x00, + 0x28, 0x00, 0x00, 0x00, 0x36, 0x00, 0x05, 0x00, 0x02, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, 0xF8, 0x00, 0x02, 0x00, 0x05, 0x00, 0x00, 0x00, + 0xF7, 0x00, 0x03, 0x00, 0x77, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0xFB, 0x00, 0x03, 0x00, + 0x1C, 0x00, 0x00, 0x00, 0x78, 0x00, 0x00, 0x00, 0xF8, 0x00, 0x02, 0x00, 0x78, 0x00, 0x00, 0x00, + 0x3D, 0x00, 0x04, 0x00, 0x0A, 0x00, 0x00, 0x00, 0x0D, 0x00, 0x00, 0x00, 0x0C, 0x00, 0x00, 0x00, + 0x4F, 0x00, 0x07, 0x00, 0x07, 0x00, 0x00, 0x00, 0x0E, 0x00, 0x00, 0x00, 0x0D, 0x00, 0x00, 0x00, + 0x0D, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x3D, 0x00, 0x04, 0x00, + 0x14, 0x00, 0x00, 0x00, 0x17, 0x00, 0x00, 0x00, 0x16, 0x00, 0x00, 0x00, 0x64, 0x00, 0x04, 0x00, + 0x13, 0x00, 0x00, 0x00, 0x19, 0x00, 0x00, 0x00, 0x17, 0x00, 0x00, 0x00, 0x67, 0x00, 0x05, 0x00, + 0x10, 0x00, 0x00, 0x00, 0x1A, 0x00, 0x00, 0x00, 0x19, 0x00, 0x00, 0x00, 0x18, 0x00, 0x00, 0x00, + 0x51, 0x00, 0x05, 0x00, 0x06, 0x00, 0x00, 0x00, 0x1F, 0x00, 0x00, 0x00, 0x0D, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x7C, 0x00, 0x04, 0x00, 0x0F, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00, + 0x1F, 0x00, 0x00, 0x00, 0x51, 0x00, 0x05, 0x00, 0x0F, 0x00, 0x00, 0x00, 0x23, 0x00, 0x00, 0x00, + 0x1A, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0xAF, 0x00, 0x05, 0x00, 0x1B, 0x00, 0x00, 0x00, + 0x24, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00, 0x23, 0x00, 0x00, 0x00, 0xA8, 0x00, 0x04, 0x00, + 0x1B, 0x00, 0x00, 0x00, 0x25, 0x00, 0x00, 0x00, 0x24, 0x00, 0x00, 0x00, 0xF7, 0x00, 0x03, 0x00, + 0x27, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0xFA, 0x00, 0x04, 0x00, 0x25, 0x00, 0x00, 0x00, + 0x26, 0x00, 0x00, 0x00, 0x27, 0x00, 0x00, 0x00, 0xF8, 0x00, 0x02, 0x00, 0x26, 0x00, 0x00, 0x00, + 0x51, 0x00, 0x05, 0x00, 0x06, 0x00, 0x00, 0x00, 0x2A, 0x00, 0x00, 0x00, 0x0D, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0x7C, 0x00, 0x04, 0x00, 0x0F, 0x00, 0x00, 0x00, 0x2B, 0x00, 0x00, 0x00, + 0x2A, 0x00, 0x00, 0x00, 0x51, 0x00, 0x05, 0x00, 0x0F, 0x00, 0x00, 0x00, 0x2D, 0x00, 0x00, 0x00, + 0x1A, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0xAF, 0x00, 0x05, 0x00, 0x1B, 0x00, 0x00, 0x00, + 0x2E, 0x00, 0x00, 0x00, 0x2B, 0x00, 0x00, 0x00, 0x2D, 0x00, 0x00, 0x00, 0xF9, 0x00, 0x02, 0x00, + 0x27, 0x00, 0x00, 0x00, 0xF8, 0x00, 0x02, 0x00, 0x27, 0x00, 0x00, 0x00, 0xF5, 0x00, 0x07, 0x00, + 0x1B, 0x00, 0x00, 0x00, 0x2F, 0x00, 0x00, 0x00, 0x24, 0x00, 0x00, 0x00, 0x78, 0x00, 0x00, 0x00, + 0x2E, 0x00, 0x00, 0x00, 0x26, 0x00, 0x00, 0x00, 0xF7, 0x00, 0x03, 0x00, 0x31, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0xFA, 0x00, 0x04, 0x00, 0x2F, 0x00, 0x00, 0x00, 0x30, 0x00, 0x00, 0x00, + 0x31, 0x00, 0x00, 0x00, 0xF8, 0x00, 0x02, 0x00, 0x30, 0x00, 0x00, 0x00, 0xF9, 0x00, 0x02, 0x00, + 0x77, 0x00, 0x00, 0x00, 0xF8, 0x00, 0x02, 0x00, 0x31, 0x00, 0x00, 0x00, 0x41, 0x00, 0x05, 0x00, + 0x39, 0x00, 0x00, 0x00, 0x3A, 0x00, 0x00, 0x00, 0x38, 0x00, 0x00, 0x00, 0x18, 0x00, 0x00, 0x00, + 0x3D, 0x00, 0x04, 0x00, 0x0F, 0x00, 0x00, 0x00, 0x3B, 0x00, 0x00, 0x00, 0x3A, 0x00, 0x00, 0x00, + 0xC4, 0x00, 0x05, 0x00, 0x06, 0x00, 0x00, 0x00, 0x3C, 0x00, 0x00, 0x00, 0x1F, 0x00, 0x00, 0x00, + 0x3B, 0x00, 0x00, 0x00, 0x51, 0x00, 0x05, 0x00, 0x06, 0x00, 0x00, 0x00, 0x40, 0x00, 0x00, 0x00, + 0x0D, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x50, 0x00, 0x05, 0x00, 0x07, 0x00, 0x00, 0x00, + 0x41, 0x00, 0x00, 0x00, 0x3C, 0x00, 0x00, 0x00, 0x40, 0x00, 0x00, 0x00, 0x80, 0x00, 0x05, 0x00, + 0x06, 0x00, 0x00, 0x00, 0x44, 0x00, 0x00, 0x00, 0x3C, 0x00, 0x00, 0x00, 0x28, 0x00, 0x00, 0x00, + 0x50, 0x00, 0x05, 0x00, 0x07, 0x00, 0x00, 0x00, 0x47, 0x00, 0x00, 0x00, 0x44, 0x00, 0x00, 0x00, + 0x40, 0x00, 0x00, 0x00, 0x80, 0x00, 0x05, 0x00, 0x06, 0x00, 0x00, 0x00, 0x4B, 0x00, 0x00, 0x00, + 0x3C, 0x00, 0x00, 0x00, 0x4A, 0x00, 0x00, 0x00, 0x50, 0x00, 0x05, 0x00, 0x07, 0x00, 0x00, 0x00, + 0x4E, 0x00, 0x00, 0x00, 0x4B, 0x00, 0x00, 0x00, 0x40, 0x00, 0x00, 0x00, 0x80, 0x00, 0x05, 0x00, + 0x06, 0x00, 0x00, 0x00, 0x52, 0x00, 0x00, 0x00, 0x3C, 0x00, 0x00, 0x00, 0x51, 0x00, 0x00, 0x00, + 0x50, 0x00, 0x05, 0x00, 0x07, 0x00, 0x00, 0x00, 0x55, 0x00, 0x00, 0x00, 0x52, 0x00, 0x00, 0x00, + 0x40, 0x00, 0x00, 0x00, 0x7C, 0x00, 0x04, 0x00, 0x10, 0x00, 0x00, 0x00, 0x5B, 0x00, 0x00, 0x00, + 0x0E, 0x00, 0x00, 0x00, 0x64, 0x00, 0x04, 0x00, 0x13, 0x00, 0x00, 0x00, 0x5C, 0x00, 0x00, 0x00, + 0x17, 0x00, 0x00, 0x00, 0x5F, 0x00, 0x07, 0x00, 0x56, 0x00, 0x00, 0x00, 0x5D, 0x00, 0x00, 0x00, + 0x5C, 0x00, 0x00, 0x00, 0x5B, 0x00, 0x00, 0x00, 0x02, 0x20, 0x00, 0x00, 0x18, 0x00, 0x00, 0x00, + 0x3D, 0x00, 0x04, 0x00, 0x5E, 0x00, 0x00, 0x00, 0x61, 0x00, 0x00, 0x00, 0x60, 0x00, 0x00, 0x00, + 0x7C, 0x00, 0x04, 0x00, 0x10, 0x00, 0x00, 0x00, 0x63, 0x00, 0x00, 0x00, 0x41, 0x00, 0x00, 0x00, + 0x4F, 0x00, 0x09, 0x00, 0x56, 0x00, 0x00, 0x00, 0x65, 0x00, 0x00, 0x00, 0x5D, 0x00, 0x00, 0x00, + 0x5D, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x63, 0x00, 0x05, 0x00, 0x61, 0x00, 0x00, 0x00, 0x63, 0x00, 0x00, 0x00, + 0x65, 0x00, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x3D, 0x00, 0x04, 0x00, 0x5E, 0x00, 0x00, 0x00, + 0x66, 0x00, 0x00, 0x00, 0x60, 0x00, 0x00, 0x00, 0x7C, 0x00, 0x04, 0x00, 0x10, 0x00, 0x00, 0x00, + 0x68, 0x00, 0x00, 0x00, 0x47, 0x00, 0x00, 0x00, 0x4F, 0x00, 0x09, 0x00, 0x56, 0x00, 0x00, 0x00, + 0x6A, 0x00, 0x00, 0x00, 0x5D, 0x00, 0x00, 0x00, 0x5D, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x63, 0x00, 0x05, 0x00, + 0x66, 0x00, 0x00, 0x00, 0x68, 0x00, 0x00, 0x00, 0x6A, 0x00, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, + 0x3D, 0x00, 0x04, 0x00, 0x5E, 0x00, 0x00, 0x00, 0x6B, 0x00, 0x00, 0x00, 0x60, 0x00, 0x00, 0x00, + 0x7C, 0x00, 0x04, 0x00, 0x10, 0x00, 0x00, 0x00, 0x6D, 0x00, 0x00, 0x00, 0x4E, 0x00, 0x00, 0x00, + 0x4F, 0x00, 0x09, 0x00, 0x56, 0x00, 0x00, 0x00, 0x6F, 0x00, 0x00, 0x00, 0x5D, 0x00, 0x00, 0x00, + 0x5D, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, + 0x02, 0x00, 0x00, 0x00, 0x63, 0x00, 0x05, 0x00, 0x6B, 0x00, 0x00, 0x00, 0x6D, 0x00, 0x00, 0x00, + 0x6F, 0x00, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x3D, 0x00, 0x04, 0x00, 0x5E, 0x00, 0x00, 0x00, + 0x70, 0x00, 0x00, 0x00, 0x60, 0x00, 0x00, 0x00, 0x7C, 0x00, 0x04, 0x00, 0x10, 0x00, 0x00, 0x00, + 0x72, 0x00, 0x00, 0x00, 0x55, 0x00, 0x00, 0x00, 0x4F, 0x00, 0x09, 0x00, 0x56, 0x00, 0x00, 0x00, + 0x74, 0x00, 0x00, 0x00, 0x5D, 0x00, 0x00, 0x00, 0x5D, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, + 0x03, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, 0x63, 0x00, 0x05, 0x00, + 0x70, 0x00, 0x00, 0x00, 0x72, 0x00, 0x00, 0x00, 0x74, 0x00, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, + 0xF9, 0x00, 0x02, 0x00, 0x77, 0x00, 0x00, 0x00, 0xF8, 0x00, 0x02, 0x00, 0x77, 0x00, 0x00, 0x00, + 0xFD, 0x00, 0x01, 0x00, 0x38, 0x00, 0x01, 0x00, + }; + public static readonly byte[] ColorCopyToNonMsComputeShaderSource = new byte[] { 0x03, 0x02, 0x23, 0x07, 0x00, 0x00, 0x01, 0x00, 0x0A, 0x00, 0x08, 0x00, 0x86, 0x00, 0x00, 0x00, @@ -801,6 +933,133 @@ namespace Ryujinx.Graphics.Vulkan.Shaders 0x84, 0x00, 0x00, 0x00, 0xFD, 0x00, 0x01, 0x00, 0x38, 0x00, 0x01, 0x00, }; + public static readonly byte[] ColorCopyWideningComputeShaderSource = new byte[] + { + 0x03, 0x02, 0x23, 0x07, 0x00, 0x05, 0x01, 0x00, 0x0B, 0x00, 0x08, 0x00, 0x72, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x11, 0x00, 0x02, 0x00, 0x01, 0x00, 0x00, 0x00, 0x11, 0x00, 0x02, 0x00, + 0x32, 0x00, 0x00, 0x00, 0x11, 0x00, 0x02, 0x00, 0x38, 0x00, 0x00, 0x00, 0x0B, 0x00, 0x06, 0x00, + 0x01, 0x00, 0x00, 0x00, 0x47, 0x4C, 0x53, 0x4C, 0x2E, 0x73, 0x74, 0x64, 0x2E, 0x34, 0x35, 0x30, + 0x00, 0x00, 0x00, 0x00, 0x0E, 0x00, 0x03, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0x0F, 0x00, 0x09, 0x00, 0x05, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, 0x6D, 0x61, 0x69, 0x6E, + 0x00, 0x00, 0x00, 0x00, 0x0C, 0x00, 0x00, 0x00, 0x15, 0x00, 0x00, 0x00, 0x35, 0x00, 0x00, 0x00, + 0x42, 0x00, 0x00, 0x00, 0x10, 0x00, 0x06, 0x00, 0x04, 0x00, 0x00, 0x00, 0x11, 0x00, 0x00, 0x00, + 0x20, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x03, 0x00, 0x03, 0x00, + 0x02, 0x00, 0x00, 0x00, 0xC2, 0x01, 0x00, 0x00, 0x05, 0x00, 0x04, 0x00, 0x04, 0x00, 0x00, 0x00, + 0x6D, 0x61, 0x69, 0x6E, 0x00, 0x00, 0x00, 0x00, 0x05, 0x00, 0x08, 0x00, 0x0C, 0x00, 0x00, 0x00, + 0x67, 0x6C, 0x5F, 0x47, 0x6C, 0x6F, 0x62, 0x61, 0x6C, 0x49, 0x6E, 0x76, 0x6F, 0x63, 0x61, 0x74, + 0x69, 0x6F, 0x6E, 0x49, 0x44, 0x00, 0x00, 0x00, 0x05, 0x00, 0x03, 0x00, 0x15, 0x00, 0x00, 0x00, + 0x64, 0x73, 0x74, 0x00, 0x05, 0x00, 0x05, 0x00, 0x33, 0x00, 0x00, 0x00, 0x72, 0x61, 0x74, 0x69, + 0x6F, 0x5F, 0x69, 0x6E, 0x00, 0x00, 0x00, 0x00, 0x06, 0x00, 0x05, 0x00, 0x33, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x72, 0x61, 0x74, 0x69, 0x6F, 0x00, 0x00, 0x00, 0x05, 0x00, 0x03, 0x00, + 0x35, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x05, 0x00, 0x03, 0x00, 0x42, 0x00, 0x00, 0x00, + 0x73, 0x72, 0x63, 0x00, 0x47, 0x00, 0x04, 0x00, 0x0C, 0x00, 0x00, 0x00, 0x0B, 0x00, 0x00, 0x00, + 0x1C, 0x00, 0x00, 0x00, 0x47, 0x00, 0x04, 0x00, 0x15, 0x00, 0x00, 0x00, 0x22, 0x00, 0x00, 0x00, + 0x03, 0x00, 0x00, 0x00, 0x47, 0x00, 0x04, 0x00, 0x15, 0x00, 0x00, 0x00, 0x21, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x47, 0x00, 0x03, 0x00, 0x15, 0x00, 0x00, 0x00, 0x19, 0x00, 0x00, 0x00, + 0x48, 0x00, 0x05, 0x00, 0x33, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x23, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x47, 0x00, 0x03, 0x00, 0x33, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, + 0x47, 0x00, 0x04, 0x00, 0x35, 0x00, 0x00, 0x00, 0x22, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x47, 0x00, 0x04, 0x00, 0x35, 0x00, 0x00, 0x00, 0x21, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x47, 0x00, 0x04, 0x00, 0x42, 0x00, 0x00, 0x00, 0x22, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, + 0x47, 0x00, 0x04, 0x00, 0x42, 0x00, 0x00, 0x00, 0x21, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x47, 0x00, 0x04, 0x00, 0x6F, 0x00, 0x00, 0x00, 0x0B, 0x00, 0x00, 0x00, 0x19, 0x00, 0x00, 0x00, + 0x13, 0x00, 0x02, 0x00, 0x02, 0x00, 0x00, 0x00, 0x21, 0x00, 0x03, 0x00, 0x03, 0x00, 0x00, 0x00, + 0x02, 0x00, 0x00, 0x00, 0x15, 0x00, 0x04, 0x00, 0x06, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x17, 0x00, 0x04, 0x00, 0x07, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, + 0x02, 0x00, 0x00, 0x00, 0x17, 0x00, 0x04, 0x00, 0x0A, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, + 0x03, 0x00, 0x00, 0x00, 0x20, 0x00, 0x04, 0x00, 0x0B, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0x0A, 0x00, 0x00, 0x00, 0x3B, 0x00, 0x04, 0x00, 0x0B, 0x00, 0x00, 0x00, 0x0C, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0x15, 0x00, 0x04, 0x00, 0x0F, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0x17, 0x00, 0x04, 0x00, 0x10, 0x00, 0x00, 0x00, 0x0F, 0x00, 0x00, 0x00, + 0x02, 0x00, 0x00, 0x00, 0x19, 0x00, 0x09, 0x00, 0x13, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x02, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x20, 0x00, 0x04, 0x00, 0x14, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x13, 0x00, 0x00, 0x00, 0x3B, 0x00, 0x04, 0x00, 0x14, 0x00, 0x00, 0x00, + 0x15, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x14, 0x00, 0x02, 0x00, 0x18, 0x00, 0x00, 0x00, + 0x2B, 0x00, 0x04, 0x00, 0x06, 0x00, 0x00, 0x00, 0x19, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x2B, 0x00, 0x04, 0x00, 0x06, 0x00, 0x00, 0x00, 0x25, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0x1E, 0x00, 0x03, 0x00, 0x33, 0x00, 0x00, 0x00, 0x0F, 0x00, 0x00, 0x00, 0x20, 0x00, 0x04, 0x00, + 0x34, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x33, 0x00, 0x00, 0x00, 0x3B, 0x00, 0x04, 0x00, + 0x34, 0x00, 0x00, 0x00, 0x35, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x2B, 0x00, 0x04, 0x00, + 0x0F, 0x00, 0x00, 0x00, 0x36, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x20, 0x00, 0x04, 0x00, + 0x37, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x0F, 0x00, 0x00, 0x00, 0x19, 0x00, 0x09, 0x00, + 0x3F, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x1B, 0x00, 0x03, 0x00, 0x40, 0x00, 0x00, 0x00, 0x3F, 0x00, 0x00, 0x00, 0x20, 0x00, 0x04, 0x00, + 0x41, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x40, 0x00, 0x00, 0x00, 0x3B, 0x00, 0x04, 0x00, + 0x41, 0x00, 0x00, 0x00, 0x42, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x2C, 0x00, 0x05, 0x00, + 0x10, 0x00, 0x00, 0x00, 0x46, 0x00, 0x00, 0x00, 0x36, 0x00, 0x00, 0x00, 0x36, 0x00, 0x00, 0x00, + 0x17, 0x00, 0x04, 0x00, 0x48, 0x00, 0x00, 0x00, 0x06, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, + 0x2B, 0x00, 0x04, 0x00, 0x0F, 0x00, 0x00, 0x00, 0x4F, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0x2C, 0x00, 0x05, 0x00, 0x10, 0x00, 0x00, 0x00, 0x50, 0x00, 0x00, 0x00, 0x4F, 0x00, 0x00, 0x00, + 0x36, 0x00, 0x00, 0x00, 0x2B, 0x00, 0x04, 0x00, 0x0F, 0x00, 0x00, 0x00, 0x58, 0x00, 0x00, 0x00, + 0x02, 0x00, 0x00, 0x00, 0x2C, 0x00, 0x05, 0x00, 0x10, 0x00, 0x00, 0x00, 0x59, 0x00, 0x00, 0x00, + 0x58, 0x00, 0x00, 0x00, 0x36, 0x00, 0x00, 0x00, 0x2B, 0x00, 0x04, 0x00, 0x0F, 0x00, 0x00, 0x00, + 0x61, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, 0x2C, 0x00, 0x05, 0x00, 0x10, 0x00, 0x00, 0x00, + 0x62, 0x00, 0x00, 0x00, 0x61, 0x00, 0x00, 0x00, 0x36, 0x00, 0x00, 0x00, 0x2B, 0x00, 0x04, 0x00, + 0x06, 0x00, 0x00, 0x00, 0x6E, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00, 0x2C, 0x00, 0x06, 0x00, + 0x0A, 0x00, 0x00, 0x00, 0x6F, 0x00, 0x00, 0x00, 0x6E, 0x00, 0x00, 0x00, 0x6E, 0x00, 0x00, 0x00, + 0x25, 0x00, 0x00, 0x00, 0x36, 0x00, 0x05, 0x00, 0x02, 0x00, 0x00, 0x00, 0x04, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x03, 0x00, 0x00, 0x00, 0xF8, 0x00, 0x02, 0x00, 0x05, 0x00, 0x00, 0x00, + 0xF7, 0x00, 0x03, 0x00, 0x70, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0xFB, 0x00, 0x03, 0x00, + 0x19, 0x00, 0x00, 0x00, 0x71, 0x00, 0x00, 0x00, 0xF8, 0x00, 0x02, 0x00, 0x71, 0x00, 0x00, 0x00, + 0x3D, 0x00, 0x04, 0x00, 0x0A, 0x00, 0x00, 0x00, 0x0D, 0x00, 0x00, 0x00, 0x0C, 0x00, 0x00, 0x00, + 0x4F, 0x00, 0x07, 0x00, 0x07, 0x00, 0x00, 0x00, 0x0E, 0x00, 0x00, 0x00, 0x0D, 0x00, 0x00, 0x00, + 0x0D, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x3D, 0x00, 0x04, 0x00, + 0x13, 0x00, 0x00, 0x00, 0x16, 0x00, 0x00, 0x00, 0x15, 0x00, 0x00, 0x00, 0x68, 0x00, 0x04, 0x00, + 0x10, 0x00, 0x00, 0x00, 0x17, 0x00, 0x00, 0x00, 0x16, 0x00, 0x00, 0x00, 0x51, 0x00, 0x05, 0x00, + 0x06, 0x00, 0x00, 0x00, 0x1C, 0x00, 0x00, 0x00, 0x0D, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x7C, 0x00, 0x04, 0x00, 0x0F, 0x00, 0x00, 0x00, 0x1D, 0x00, 0x00, 0x00, 0x1C, 0x00, 0x00, 0x00, + 0x51, 0x00, 0x05, 0x00, 0x0F, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00, 0x17, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0xAF, 0x00, 0x05, 0x00, 0x18, 0x00, 0x00, 0x00, 0x21, 0x00, 0x00, 0x00, + 0x1D, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0x00, 0xA8, 0x00, 0x04, 0x00, 0x18, 0x00, 0x00, 0x00, + 0x22, 0x00, 0x00, 0x00, 0x21, 0x00, 0x00, 0x00, 0xF7, 0x00, 0x03, 0x00, 0x24, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0xFA, 0x00, 0x04, 0x00, 0x22, 0x00, 0x00, 0x00, 0x23, 0x00, 0x00, 0x00, + 0x24, 0x00, 0x00, 0x00, 0xF8, 0x00, 0x02, 0x00, 0x23, 0x00, 0x00, 0x00, 0x51, 0x00, 0x05, 0x00, + 0x06, 0x00, 0x00, 0x00, 0x27, 0x00, 0x00, 0x00, 0x0D, 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, + 0x7C, 0x00, 0x04, 0x00, 0x0F, 0x00, 0x00, 0x00, 0x28, 0x00, 0x00, 0x00, 0x27, 0x00, 0x00, 0x00, + 0x51, 0x00, 0x05, 0x00, 0x0F, 0x00, 0x00, 0x00, 0x2A, 0x00, 0x00, 0x00, 0x17, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0xAF, 0x00, 0x05, 0x00, 0x18, 0x00, 0x00, 0x00, 0x2B, 0x00, 0x00, 0x00, + 0x28, 0x00, 0x00, 0x00, 0x2A, 0x00, 0x00, 0x00, 0xF9, 0x00, 0x02, 0x00, 0x24, 0x00, 0x00, 0x00, + 0xF8, 0x00, 0x02, 0x00, 0x24, 0x00, 0x00, 0x00, 0xF5, 0x00, 0x07, 0x00, 0x18, 0x00, 0x00, 0x00, + 0x2C, 0x00, 0x00, 0x00, 0x21, 0x00, 0x00, 0x00, 0x71, 0x00, 0x00, 0x00, 0x2B, 0x00, 0x00, 0x00, + 0x23, 0x00, 0x00, 0x00, 0xF7, 0x00, 0x03, 0x00, 0x2E, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0xFA, 0x00, 0x04, 0x00, 0x2C, 0x00, 0x00, 0x00, 0x2D, 0x00, 0x00, 0x00, 0x2E, 0x00, 0x00, 0x00, + 0xF8, 0x00, 0x02, 0x00, 0x2D, 0x00, 0x00, 0x00, 0xF9, 0x00, 0x02, 0x00, 0x70, 0x00, 0x00, 0x00, + 0xF8, 0x00, 0x02, 0x00, 0x2E, 0x00, 0x00, 0x00, 0x41, 0x00, 0x05, 0x00, 0x37, 0x00, 0x00, 0x00, + 0x38, 0x00, 0x00, 0x00, 0x35, 0x00, 0x00, 0x00, 0x36, 0x00, 0x00, 0x00, 0x3D, 0x00, 0x04, 0x00, + 0x0F, 0x00, 0x00, 0x00, 0x39, 0x00, 0x00, 0x00, 0x38, 0x00, 0x00, 0x00, 0xC4, 0x00, 0x05, 0x00, + 0x06, 0x00, 0x00, 0x00, 0x3A, 0x00, 0x00, 0x00, 0x1C, 0x00, 0x00, 0x00, 0x39, 0x00, 0x00, 0x00, + 0x51, 0x00, 0x05, 0x00, 0x06, 0x00, 0x00, 0x00, 0x3C, 0x00, 0x00, 0x00, 0x0D, 0x00, 0x00, 0x00, + 0x01, 0x00, 0x00, 0x00, 0x50, 0x00, 0x05, 0x00, 0x07, 0x00, 0x00, 0x00, 0x3D, 0x00, 0x00, 0x00, + 0x3A, 0x00, 0x00, 0x00, 0x3C, 0x00, 0x00, 0x00, 0x3D, 0x00, 0x04, 0x00, 0x40, 0x00, 0x00, 0x00, + 0x43, 0x00, 0x00, 0x00, 0x42, 0x00, 0x00, 0x00, 0x7C, 0x00, 0x04, 0x00, 0x10, 0x00, 0x00, 0x00, + 0x45, 0x00, 0x00, 0x00, 0x3D, 0x00, 0x00, 0x00, 0x64, 0x00, 0x04, 0x00, 0x3F, 0x00, 0x00, 0x00, + 0x47, 0x00, 0x00, 0x00, 0x43, 0x00, 0x00, 0x00, 0x5F, 0x00, 0x08, 0x00, 0x48, 0x00, 0x00, 0x00, + 0x49, 0x00, 0x00, 0x00, 0x47, 0x00, 0x00, 0x00, 0x45, 0x00, 0x00, 0x00, 0x0A, 0x20, 0x00, 0x00, + 0x36, 0x00, 0x00, 0x00, 0x46, 0x00, 0x00, 0x00, 0x51, 0x00, 0x05, 0x00, 0x06, 0x00, 0x00, 0x00, + 0x4A, 0x00, 0x00, 0x00, 0x49, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x64, 0x00, 0x04, 0x00, + 0x3F, 0x00, 0x00, 0x00, 0x51, 0x00, 0x00, 0x00, 0x43, 0x00, 0x00, 0x00, 0x5F, 0x00, 0x08, 0x00, + 0x48, 0x00, 0x00, 0x00, 0x52, 0x00, 0x00, 0x00, 0x51, 0x00, 0x00, 0x00, 0x45, 0x00, 0x00, 0x00, + 0x0A, 0x20, 0x00, 0x00, 0x36, 0x00, 0x00, 0x00, 0x50, 0x00, 0x00, 0x00, 0x51, 0x00, 0x05, 0x00, + 0x06, 0x00, 0x00, 0x00, 0x53, 0x00, 0x00, 0x00, 0x52, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, + 0x64, 0x00, 0x04, 0x00, 0x3F, 0x00, 0x00, 0x00, 0x5A, 0x00, 0x00, 0x00, 0x43, 0x00, 0x00, 0x00, + 0x5F, 0x00, 0x08, 0x00, 0x48, 0x00, 0x00, 0x00, 0x5B, 0x00, 0x00, 0x00, 0x5A, 0x00, 0x00, 0x00, + 0x45, 0x00, 0x00, 0x00, 0x0A, 0x20, 0x00, 0x00, 0x36, 0x00, 0x00, 0x00, 0x59, 0x00, 0x00, 0x00, + 0x51, 0x00, 0x05, 0x00, 0x06, 0x00, 0x00, 0x00, 0x5C, 0x00, 0x00, 0x00, 0x5B, 0x00, 0x00, 0x00, + 0x00, 0x00, 0x00, 0x00, 0x64, 0x00, 0x04, 0x00, 0x3F, 0x00, 0x00, 0x00, 0x63, 0x00, 0x00, 0x00, + 0x43, 0x00, 0x00, 0x00, 0x5F, 0x00, 0x08, 0x00, 0x48, 0x00, 0x00, 0x00, 0x64, 0x00, 0x00, 0x00, + 0x63, 0x00, 0x00, 0x00, 0x45, 0x00, 0x00, 0x00, 0x0A, 0x20, 0x00, 0x00, 0x36, 0x00, 0x00, 0x00, + 0x62, 0x00, 0x00, 0x00, 0x51, 0x00, 0x05, 0x00, 0x06, 0x00, 0x00, 0x00, 0x65, 0x00, 0x00, 0x00, + 0x64, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x3D, 0x00, 0x04, 0x00, 0x13, 0x00, 0x00, 0x00, + 0x66, 0x00, 0x00, 0x00, 0x15, 0x00, 0x00, 0x00, 0x7C, 0x00, 0x04, 0x00, 0x10, 0x00, 0x00, 0x00, + 0x68, 0x00, 0x00, 0x00, 0x0E, 0x00, 0x00, 0x00, 0x50, 0x00, 0x07, 0x00, 0x48, 0x00, 0x00, 0x00, + 0x6D, 0x00, 0x00, 0x00, 0x4A, 0x00, 0x00, 0x00, 0x53, 0x00, 0x00, 0x00, 0x5C, 0x00, 0x00, 0x00, + 0x65, 0x00, 0x00, 0x00, 0x63, 0x00, 0x05, 0x00, 0x66, 0x00, 0x00, 0x00, 0x68, 0x00, 0x00, 0x00, + 0x6D, 0x00, 0x00, 0x00, 0x00, 0x20, 0x00, 0x00, 0xF9, 0x00, 0x02, 0x00, 0x70, 0x00, 0x00, 0x00, + 0xF8, 0x00, 0x02, 0x00, 0x70, 0x00, 0x00, 0x00, 0xFD, 0x00, 0x01, 0x00, 0x38, 0x00, 0x01, 0x00, + }; + public static readonly byte[] ColorDrawToMsVertexShaderSource = new byte[] { 0x03, 0x02, 0x23, 0x07, 0x00, 0x00, 0x01, 0x00, 0x0A, 0x00, 0x08, 0x00, 0x2E, 0x00, 0x00, 0x00, diff --git a/Ryujinx.Graphics.Vulkan/TextureView.cs b/Ryujinx.Graphics.Vulkan/TextureView.cs index aa050c015d..264ecf5db9 100644 --- a/Ryujinx.Graphics.Vulkan/TextureView.cs +++ b/Ryujinx.Graphics.Vulkan/TextureView.cs @@ -199,6 +199,12 @@ namespace Ryujinx.Graphics.Vulkan int layers = Math.Min(Info.GetLayers(), dst.Info.GetLayers() - firstLayer); _gd.HelperShader.CopyNonMSToMS(_gd, cbs, src, dst, 0, firstLayer, layers); } + else if (dst.Info.BytesPerPixel != Info.BytesPerPixel) + { + int layers = Math.Min(Info.GetLayers(), dst.Info.GetLayers() - firstLayer); + int levels = Math.Min(Info.Levels, dst.Info.Levels - firstLevel); + _gd.HelperShader.CopyIncompatibleFormats(_gd, cbs, src, dst, 0, firstLayer, 0, firstLevel, layers, levels); + } else { TextureCopy.Copy( @@ -244,6 +250,10 @@ namespace Ryujinx.Graphics.Vulkan { _gd.HelperShader.CopyNonMSToMS(_gd, cbs, src, dst, srcLayer, dstLayer, 1); } + else if (dst.Info.BytesPerPixel != Info.BytesPerPixel) + { + _gd.HelperShader.CopyIncompatibleFormats(_gd, cbs, src, dst, srcLayer, dstLayer, srcLevel, dstLevel, 1, 1); + } else { TextureCopy.Copy(