vleeuwenmenno/Ryujinx
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Move shader resource descriptor creation out of the backend (#2290)

Browse Source 
* Move shader resource descriptor creation out of the backend

* Remove now unused code, and other nits

* Shader cache version bump

* Nits

* Set format for bindless image load/store

* Fix buffer write flag
This commit is contained in:
gdkchan
2021-05-19 18:15:26 -03:00
committed by GitHub
parent b5c72b44de
commit 49745cfa37
25 changed files with 565 additions and 516 deletions
Download Patch File Download Diff File Expand all files Collapse all files

30
Ryujinx.Graphics.Shader/Translation/EmitterContext.cs
View File

@ -53,6 +53,36 @@ namespace Ryujinx.Graphics.Shader.Translation
_operations.Add(operation);
}
public TextureOperation CreateTextureOperation(
Instruction inst,
SamplerType type,
TextureFlags flags,
int handle,
int compIndex,
Operand dest,
params Operand[] sources)
{
return CreateTextureOperation(inst, type, TextureFormat.Unknown, flags, handle, compIndex, dest, sources);
}
public TextureOperation CreateTextureOperation(
Instruction inst,
SamplerType type,
TextureFormat format,
TextureFlags flags,
int handle,
int compIndex,
Operand dest,
params Operand[] sources)
{
if (!flags.HasFlag(TextureFlags.Bindless))
{
Config.SetUsedTexture(inst, type, format, flags, TextureOperation.DefaultCbufSlot, handle);
}
return new TextureOperation(inst, type, format, flags, handle, compIndex, dest, sources);
}
public void FlagAttributeRead(int attribute)
{
if (Config.Stage == ShaderStage.Vertex && attribute == AttributeConsts.InstanceId)

9
Ryujinx.Graphics.Shader/Translation/EmitterContextInsts.cs
View File

@ -518,6 +518,15 @@ namespace Ryujinx.Graphics.Shader.Translation
public static Operand LoadConstant(this EmitterContext context, Operand a, Operand b)
{
if (a.Type == OperandType.Constant)
{
context.Config.SetUsedConstantBuffer(a.Value);
}
else
{
context.Config.SetUsedFeature(FeatureFlags.CbIndexing);
}
return context.Add(Instruction.LoadConstant, Local(), a, b);
}

4
Ryujinx.Graphics.Shader/Translation/FeatureFlags.cs
View File

@ -16,7 +16,7 @@ namespace Ryujinx.Graphics.Shader.Translation
FragCoordXY = 1 << 1,
Bindless = 1 << 2,
InstanceId = 1 << 3
InstanceId = 1 << 3,
CbIndexing = 1 << 4
}
}

16
Ryujinx.Graphics.Shader/Translation/Optimizations/BindlessElimination.cs
View File

@ -33,7 +33,7 @@ namespace Ryujinx.Graphics.Shader.Translation.Optimizations
if (bindlessHandle.Type == OperandType.ConstantBuffer)
{
texOp.SetHandle(bindlessHandle.GetCbufOffset(), bindlessHandle.GetCbufSlot());
SetHandle(config, texOp, bindlessHandle.GetCbufOffset(), bindlessHandle.GetCbufSlot());
continue;
}
@ -56,7 +56,7 @@ namespace Ryujinx.Graphics.Shader.Translation.Optimizations
continue;
}
texOp.SetHandle(src0.GetCbufOffset() | (src1.GetCbufOffset() << 16), src0.GetCbufSlot());
SetHandle(config, texOp, src0.GetCbufOffset() | (src1.GetCbufOffset() << 16), src0.GetCbufSlot());
}
else if (texOp.Inst == Instruction.ImageLoad || texOp.Inst == Instruction.ImageStore)
{
@ -64,11 +64,19 @@ namespace Ryujinx.Graphics.Shader.Translation.Optimizations
if (src0.Type == OperandType.ConstantBuffer)
{
texOp.SetHandle(src0.GetCbufOffset(), src0.GetCbufSlot());
texOp.Format = config.GetTextureFormat(texOp.Handle, texOp.CbufSlot);
int cbufOffset = src0.GetCbufOffset();
int cbufSlot = src0.GetCbufSlot();
texOp.Format = config.GetTextureFormat(cbufOffset, cbufSlot);
SetHandle(config, texOp, cbufOffset, cbufSlot);
}
}
}
}
private static void SetHandle(ShaderConfig config, TextureOperation texOp, int cbufOffset, int cbufSlot)
{
texOp.SetHandle(cbufOffset, cbufSlot);
config.SetUsedTexture(texOp.Inst, texOp.Type, texOp.Format, texOp.Flags, cbufSlot, cbufOffset);
}
}
}

10
Ryujinx.Graphics.Shader/Translation/Optimizations/BindlessToIndexed.cs
View File

@ -7,7 +7,7 @@ namespace Ryujinx.Graphics.Shader.Translation.Optimizations
{
static class BindlessToIndexed
{
public static void RunPass(BasicBlock block)
public static void RunPass(BasicBlock block, ShaderConfig config)
{
// We can turn a bindless texture access into a indexed access,
// as long the following conditions are true:
@ -62,7 +62,7 @@ namespace Ryujinx.Graphics.Shader.Translation.Optimizations
continue;
}
texOp.TurnIntoIndexed(addSrc1.Value / 4);
TurnIntoIndexed(config, texOp, addSrc1.Value / 4);
Operand index = Local();
@ -75,5 +75,11 @@ namespace Ryujinx.Graphics.Shader.Translation.Optimizations
texOp.SetSource(0, index);
}
}
private static void TurnIntoIndexed(ShaderConfig config, TextureOperation texOp, int handle)
{
texOp.TurnIntoIndexed(handle);
config.SetUsedTexture(texOp.Inst, texOp.Type, texOp.Format, texOp.Flags, texOp.CbufSlot, handle);
}
}
}

17
Ryujinx.Graphics.Shader/Translation/Optimizations/GlobalToStorage.cs
View File

@ -58,11 +58,16 @@ namespace Ryujinx.Graphics.Shader.Translation.Optimizations
{
Operation operation = (Operation)node.Value;
bool isAtomic = operation.Inst.IsAtomic();
bool isWrite = isAtomic || operation.Inst == Instruction.StoreGlobal;
config.SetUsedStorageBuffer(storageIndex, isWrite);
Operand GetStorageOffset()
{
Operand addrLow = operation.GetSource(0);
Operand baseAddrLow = Cbuf(0, GetStorageCbOffset(config.Stage, storageIndex));
Operand baseAddrLow = config.CreateCbuf(0, GetStorageCbOffset(config.Stage, storageIndex));
Operand baseAddrTrunc = Local();
@ -96,7 +101,7 @@ namespace Ryujinx.Graphics.Shader.Translation.Optimizations
Operation storageOp;
if (operation.Inst.IsAtomic())
if (isAtomic)
{
Instruction inst = (operation.Inst & ~Instruction.MrMask) | Instruction.MrStorage;
@ -133,7 +138,7 @@ namespace Ryujinx.Graphics.Shader.Translation.Optimizations
{
Operand addrLow = operation.GetSource(0);
Operand baseAddrLow = Cbuf(0, UbeBaseOffset + storageIndex * StorageDescSize);
Operand baseAddrLow = config.CreateCbuf(0, UbeBaseOffset + storageIndex * StorageDescSize);
Operand baseAddrTrunc = Local();
@ -157,9 +162,13 @@ namespace Ryujinx.Graphics.Shader.Translation.Optimizations
Operand[] sources = new Operand[operation.SourcesCount];
sources[0] = Const(UbeFirstCbuf + storageIndex);
int cbSlot = UbeFirstCbuf + storageIndex;
sources[0] = Const(cbSlot);
sources[1] = GetCbufOffset();
config.SetUsedConstantBuffer(cbSlot);
for (int index = 2; index < operation.SourcesCount; index++)
{
sources[index] = operation.GetSource(index);

2
Ryujinx.Graphics.Shader/Translation/Optimizations/Optimizer.cs
View File

@ -16,7 +16,7 @@ namespace Ryujinx.Graphics.Shader.Translation.Optimizations
for (int blkIndex = 0; blkIndex < blocks.Length; blkIndex++)
{
GlobalToStorage.RunPass(blocks[blkIndex], config);
BindlessToIndexed.RunPass(blocks[blkIndex]);
BindlessToIndexed.RunPass(blocks[blkIndex], config);
BindlessElimination.RunPass(blocks[blkIndex], config);
}

17
Ryujinx.Graphics.Shader/Translation/Rewriter.cs
View File

@ -48,6 +48,9 @@ namespace Ryujinx.Graphics.Shader.Translation
{
Operation operation = (Operation)node.Value;
bool isAtomic = operation.Inst.IsAtomic();
bool isWrite = isAtomic || operation.Inst == Instruction.StoreGlobal;
Operation storageOp;
Operand PrependOperation(Instruction inst, params Operand[] sources)
@ -67,11 +70,13 @@ namespace Ryujinx.Graphics.Shader.Translation
for (int slot = 0; slot < StorageMaxCount; slot++)
{
config.SetUsedStorageBuffer(slot, isWrite);
int cbOffset = GetStorageCbOffset(config.Stage, slot);
Operand baseAddrLow = Cbuf(0, cbOffset);
Operand baseAddrHigh = Cbuf(0, cbOffset + 1);
Operand size = Cbuf(0, cbOffset + 2);
Operand baseAddrLow = config.CreateCbuf(0, cbOffset);
Operand baseAddrHigh = config.CreateCbuf(0, cbOffset + 1);
Operand size = config.CreateCbuf(0, cbOffset + 2);
Operand offset = PrependOperation(Instruction.Subtract, addrLow, baseAddrLow);
Operand borrow = PrependOperation(Instruction.CompareLessU32, addrLow, baseAddrLow);
@ -104,7 +109,7 @@ namespace Ryujinx.Graphics.Shader.Translation
sources[index] = operation.GetSource(index);
}
if (operation.Inst.IsAtomic())
if (isAtomic)
{
Instruction inst = (operation.Inst & ~Instruction.MrMask) | Instruction.MrStorage;
@ -303,6 +308,7 @@ namespace Ryujinx.Graphics.Shader.Translation
node.List.AddBefore(node, new TextureOperation(
Instruction.TextureSize,
texOp.Type,
texOp.Format,
texOp.Flags,
texOp.Handle,
index,
@ -350,6 +356,7 @@ namespace Ryujinx.Graphics.Shader.Translation
node.List.AddBefore(node, new TextureOperation(
Instruction.Lod,
texOp.Type,
texOp.Format,
texOp.Flags,
texOp.Handle,
1,
@ -374,6 +381,7 @@ namespace Ryujinx.Graphics.Shader.Translation
node.List.AddBefore(node, new TextureOperation(
Instruction.TextureSize,
texOp.Type,
texOp.Format,
texOp.Flags,
texOp.Handle,
index,
@ -409,6 +417,7 @@ namespace Ryujinx.Graphics.Shader.Translation
TextureOperation newTexOp = new TextureOperation(
Instruction.TextureSample,
texOp.Type,
texOp.Format,
texOp.Flags & ~(TextureFlags.Offset | TextureFlags.Offsets),
texOp.Handle,
componentIndex,

295
Ryujinx.Graphics.Shader/Translation/ShaderConfig.cs
View File

@ -1,9 +1,16 @@
using Ryujinx.Graphics.Shader.IntermediateRepresentation;
using System;
using System.Collections.Generic;
using System.Linq;
using System.Numerics;
namespace Ryujinx.Graphics.Shader.Translation
{
class ShaderConfig
{
// TODO: Non-hardcoded array size.
public const int SamplerArraySize = 4;
public ShaderStage Stage { get; }
public bool GpPassthrough { get; }
@ -24,8 +31,6 @@ namespace Ryujinx.Graphics.Shader.Translation
public TranslationFlags Flags { get; }
public TranslationCounts Counts { get; }
public int Size { get; private set; }
public byte ClipDistancesWritten { get; private set; }
@ -34,42 +39,80 @@ namespace Ryujinx.Graphics.Shader.Translation
public HashSet<int> TextureHandlesForCache { get; }
private readonly TranslationCounts _counts;
private int _usedConstantBuffers;
private int _usedStorageBuffers;
private int _usedStorageBuffersWrite;
private struct TextureInfo : IEquatable<TextureInfo>
{
public int CbufSlot { get; }
public int Handle { get; }
public bool Indexed { get; }
public TextureFormat Format { get; }
public TextureInfo(int cbufSlot, int handle, bool indexed, TextureFormat format)
{
CbufSlot = cbufSlot;
Handle = handle;
Indexed = indexed;
Format = format;
}
public override bool Equals(object obj)
{
return obj is TextureInfo other && Equals(other);
}
public bool Equals(TextureInfo other)
{
return CbufSlot == other.CbufSlot && Handle == other.Handle && Indexed == other.Indexed && Format == other.Format;
}
public override int GetHashCode()
{
return HashCode.Combine(CbufSlot, Handle, Indexed, Format);
}
}
private struct TextureMeta
{
public bool AccurateType;
public SamplerType Type;
public TextureUsageFlags UsageFlags;
}
private readonly Dictionary<TextureInfo, TextureMeta> _usedTextures;
private readonly Dictionary<TextureInfo, TextureMeta> _usedImages;
private BufferDescriptor[] _cachedConstantBufferDescriptors;
private BufferDescriptor[] _cachedStorageBufferDescriptors;
private TextureDescriptor[] _cachedTextureDescriptors;
private TextureDescriptor[] _cachedImageDescriptors;
public ShaderConfig(IGpuAccessor gpuAccessor, TranslationFlags flags, TranslationCounts counts)
{
Stage = ShaderStage.Compute;
GpPassthrough = false;
OutputTopology = OutputTopology.PointList;
MaxOutputVertices = 0;
LocalMemorySize = 0;
ImapTypes = null;
OmapTargets = null;
OmapSampleMask = false;
OmapDepth = false;
GpuAccessor = gpuAccessor;
Flags = flags;
Size = 0;
UsedFeatures = FeatureFlags.None;
Counts = counts;
_counts = counts;
TextureHandlesForCache = new HashSet<int>();
_usedTextures = new Dictionary<TextureInfo, TextureMeta>();
_usedImages = new Dictionary<TextureInfo, TextureMeta>();
}
public ShaderConfig(ShaderHeader header, IGpuAccessor gpuAccessor, TranslationFlags flags, TranslationCounts counts)
public ShaderConfig(ShaderHeader header, IGpuAccessor gpuAccessor, TranslationFlags flags, TranslationCounts counts) : this(gpuAccessor, flags, counts)
{
Stage = header.Stage;
GpPassthrough = header.Stage == ShaderStage.Geometry && header.GpPassthrough;
OutputTopology = header.OutputTopology;
MaxOutputVertices = header.MaxOutputVertexCount;
LocalMemorySize = header.ShaderLocalMemoryLowSize + header.ShaderLocalMemoryHighSize;
ImapTypes = header.ImapTypes;
OmapTargets = header.OmapTargets;
OmapSampleMask = header.OmapSampleMask;
OmapDepth = header.OmapDepth;
GpuAccessor = gpuAccessor;
Flags = flags;
Size = 0;
UsedFeatures = FeatureFlags.None;
Counts = counts;
TextureHandlesForCache = new HashSet<int>();
Stage = header.Stage;
GpPassthrough = header.Stage == ShaderStage.Geometry && header.GpPassthrough;
OutputTopology = header.OutputTopology;
MaxOutputVertices = header.MaxOutputVertexCount;
LocalMemorySize = header.ShaderLocalMemoryLowSize + header.ShaderLocalMemoryHighSize;
ImapTypes = header.ImapTypes;
OmapTargets = header.OmapTargets;
OmapSampleMask = header.OmapSampleMask;
OmapDepth = header.OmapDepth;
}
public int GetDepthRegister()
@ -126,5 +169,199 @@ namespace Ryujinx.Graphics.Shader.Translation
{
UsedFeatures |= flags;
}
public Operand CreateCbuf(int slot, int offset)
{
SetUsedConstantBuffer(slot);
return OperandHelper.Cbuf(slot, offset);
}
public void SetUsedConstantBuffer(int slot)
{
_usedConstantBuffers |= 1 << slot;
}
public void SetUsedStorageBuffer(int slot, bool write)
{
int mask = 1 << slot;
_usedStorageBuffers |= mask;
if (write)
{
_usedStorageBuffersWrite |= mask;
}
}
public void SetUsedTexture(
Instruction inst,
SamplerType type,
TextureFormat format,
TextureFlags flags,
int cbufSlot,
int handle)
{
inst &= Instruction.Mask;
bool isImage = inst == Instruction.ImageLoad || inst == Instruction.ImageStore;
bool isWrite = inst == Instruction.ImageStore;
bool accurateType = inst != Instruction.TextureSize && inst != Instruction.Lod;
if (isImage)
{
SetUsedTextureOrImage(_usedImages, cbufSlot, handle, type, format, true, isWrite, false);
}
else
{
SetUsedTextureOrImage(_usedTextures, cbufSlot, handle, type, TextureFormat.Unknown, flags.HasFlag(TextureFlags.IntCoords), false, accurateType);
}
}
private static void SetUsedTextureOrImage(
Dictionary<TextureInfo, TextureMeta> dict,
int cbufSlot,
int handle,
SamplerType type,
TextureFormat format,
bool intCoords,
bool write,
bool accurateType)
{
var dimensions = type.GetDimensions();
var isArray = type.HasFlag(SamplerType.Array);
var isIndexed = type.HasFlag(SamplerType.Indexed);
var usageFlags = TextureUsageFlags.None;
if (intCoords)
{
usageFlags |= TextureUsageFlags.NeedsScaleValue;
var canScale = (dimensions == 2 && !isArray) || (dimensions == 3 && isArray);
if (!canScale)
{
// Resolution scaling cannot be applied to this texture right now.
// Flag so that we know to blacklist scaling on related textures when binding them.
usageFlags |= TextureUsageFlags.ResScaleUnsupported;
}
}
if (write)
{
usageFlags |= TextureUsageFlags.ImageStore;
}
int arraySize = isIndexed ? SamplerArraySize : 1;
for (int layer = 0; layer < arraySize; layer++)
{
var info = new TextureInfo(cbufSlot, handle + layer * 2, isIndexed, format);
var meta = new TextureMeta()
{
AccurateType = accurateType,
Type = type,
UsageFlags = usageFlags
};
if (dict.TryGetValue(info, out var existingMeta))
{
meta.UsageFlags |= existingMeta.UsageFlags;
// If the texture we have has inaccurate type information, then
// we prefer the most accurate one.
if (existingMeta.AccurateType)
{
meta.AccurateType = true;
meta.Type = existingMeta.Type;
}
dict[info] = meta;
}
else
{
dict.Add(info, meta);
}
}
}
public BufferDescriptor[] GetConstantBufferDescriptors()
{
if (_cachedConstantBufferDescriptors != null)
{
return _cachedConstantBufferDescriptors;
}
int usedMask = _usedConstantBuffers;
if (UsedFeatures.HasFlag(FeatureFlags.CbIndexing))
{
usedMask = FillMask(usedMask);
}
return _cachedConstantBufferDescriptors = GetBufferDescriptors(usedMask, 0, _counts.IncrementUniformBuffersCount);
}
public BufferDescriptor[] GetStorageBufferDescriptors()
{
return _cachedStorageBufferDescriptors ??= GetBufferDescriptors(FillMask(_usedStorageBuffers), _usedStorageBuffersWrite, _counts.IncrementStorageBuffersCount);
}
private static int FillMask(int mask)
{
// When the storage or uniform buffers are used as array, we must allocate a binding
// even for the "gaps" that are not used on the shader.
// For this reason, fill up the gaps so that all slots up to the highest one are
// marked as "used".
return mask != 0 ? (int)(uint.MaxValue >> BitOperations.LeadingZeroCount((uint)mask)) : 0;
}
private static BufferDescriptor[] GetBufferDescriptors(int usedMask, int writtenMask, Func<int> getBindingCallback)
{
var descriptors = new BufferDescriptor[BitOperations.PopCount((uint)usedMask)];
for (int i = 0; i < descriptors.Length; i++)
{
int slot = BitOperations.TrailingZeroCount(usedMask);
descriptors[i] = new BufferDescriptor(getBindingCallback(), slot);
if ((writtenMask & (1 << slot)) != 0)
{
descriptors[i].SetFlag(BufferUsageFlags.Write);
}
usedMask &= ~(1 << slot);
}
return descriptors;
}
public TextureDescriptor[] GetTextureDescriptors()
{
return _cachedTextureDescriptors ??= GetTextureOrImageDescriptors(_usedTextures, _counts.IncrementTexturesCount);
}
public TextureDescriptor[] GetImageDescriptors()
{
return _cachedImageDescriptors ??= GetTextureOrImageDescriptors(_usedImages, _counts.IncrementImagesCount);
}
private static TextureDescriptor[] GetTextureOrImageDescriptors(Dictionary<TextureInfo, TextureMeta> dict, Func<int> getBindingCallback)
{
var descriptors = new TextureDescriptor[dict.Count];
int i = 0;
foreach (var kv in dict.OrderBy(x => x.Key.Indexed).OrderBy(x => x.Key.Handle))
{
var info = kv.Key;
var meta = kv.Value;
int binding = getBindingCallback();
descriptors[i] = new TextureDescriptor(binding, meta.Type, info.Format, info.CbufSlot, info.Handle);
descriptors[i].SetFlag(meta.UsageFlags);
i++;
}
return descriptors;
}
}
}

14
Ryujinx.Graphics.Shader/Translation/Translator.cs
View File

@ -87,18 +87,16 @@ namespace Ryujinx.Graphics.Shader.Translation
StructuredProgramInfo sInfo = StructuredProgram.MakeStructuredProgram(funcs, config);
GlslProgram program = GlslGenerator.Generate(sInfo, config);
string glslCode = GlslGenerator.Generate(sInfo, config);
shaderProgramInfo = new ShaderProgramInfo(
program.CBufferDescriptors,
program.SBufferDescriptors,
program.TextureDescriptors,
program.ImageDescriptors,
config.GetConstantBufferDescriptors(),
config.GetStorageBufferDescriptors(),
config.GetTextureDescriptors(),
config.GetImageDescriptors(),
config.UsedFeatures.HasFlag(FeatureFlags.InstanceId),
config.ClipDistancesWritten);
string glslCode = program.Code;
return new ShaderProgram(config.Stage, glslCode);
}
@ -112,7 +110,7 @@ namespace Ryujinx.Graphics.Shader.Translation
Block[][] cfg;
ulong maxEndAddress = 0;
bool hasBindless = false;
bool hasBindless;
if ((flags & TranslationFlags.Compute) != 0)
{