Replace constant buffer access on shader with new Load instruction (#4646)
This commit is contained in:
gdkchan
GitHub
parent
fb27042e01
commit
402f05b8ef
42 changed files with 788 additions and 625 deletions
|
|
@ -22,4 +22,35 @@
|
|||
|
||||
Array = 1 << 10
|
||||
}
|
||||
|
||||
static class AggregateTypeExtensions
|
||||
{
|
||||
public static int GetSizeInBytes(this AggregateType type)
|
||||
{
|
||||
int elementSize = (type & AggregateType.ElementTypeMask) switch
|
||||
{
|
||||
AggregateType.Bool or
|
||||
AggregateType.FP32 or
|
||||
AggregateType.S32 or
|
||||
AggregateType.U32 => 4,
|
||||
AggregateType.FP64 => 8,
|
||||
_ => 0
|
||||
};
|
||||
|
||||
switch (type & AggregateType.ElementCountMask)
|
||||
{
|
||||
case AggregateType.Vector2:
|
||||
elementSize *= 2;
|
||||
break;
|
||||
case AggregateType.Vector3:
|
||||
elementSize *= 3;
|
||||
break;
|
||||
case AggregateType.Vector4:
|
||||
elementSize *= 4;
|
||||
break;
|
||||
}
|
||||
|
||||
return elementSize;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
|
|
|||
|
|
@ -234,8 +234,8 @@ namespace Ryujinx.Graphics.Shader.Translation
|
|||
{
|
||||
Operand x = this.Load(StorageKind.Output, IoVariable.Position, null, Const(0));
|
||||
Operand y = this.Load(StorageKind.Output, IoVariable.Position, null, Const(1));
|
||||
Operand xScale = this.Load(StorageKind.Input, IoVariable.SupportBlockViewInverse, null, Const(0));
|
||||
Operand yScale = this.Load(StorageKind.Input, IoVariable.SupportBlockViewInverse, null, Const(1));
|
||||
Operand xScale = this.Load(StorageKind.ConstantBuffer, SupportBuffer.Binding, Const((int)SupportBufferField.ViewportInverse), Const(0));
|
||||
Operand yScale = this.Load(StorageKind.ConstantBuffer, SupportBuffer.Binding, Const((int)SupportBufferField.ViewportInverse), Const(1));
|
||||
Operand negativeOne = ConstF(-1.0f);
|
||||
|
||||
this.Store(StorageKind.Output, IoVariable.Position, null, Const(0), this.FPFusedMultiplyAdd(x, xScale, negativeOne));
|
||||
|
|
@ -420,7 +420,7 @@ namespace Ryujinx.Graphics.Shader.Translation
|
|||
// Perform B <-> R swap if needed, for BGRA formats (not supported on OpenGL).
|
||||
if (!supportsBgra && (component == 0 || component == 2))
|
||||
{
|
||||
Operand isBgra = this.Load(StorageKind.Input, IoVariable.FragmentOutputIsBgra, null, Const(rtIndex));
|
||||
Operand isBgra = this.Load(StorageKind.ConstantBuffer, SupportBuffer.Binding, Const((int)SupportBufferField.FragmentIsBgra), Const(rtIndex));
|
||||
|
||||
Operand lblIsBgra = Label();
|
||||
Operand lblEnd = Label();
|
||||
|
|
|
|||
|
|
@ -549,11 +549,31 @@ namespace Ryujinx.Graphics.Shader.Translation
|
|||
return context.Add(fpType | Instruction.IsNan, Local(), a);
|
||||
}
|
||||
|
||||
public static Operand Load(this EmitterContext context, StorageKind storageKind, int binding)
|
||||
{
|
||||
return context.Add(Instruction.Load, storageKind, Local(), Const(binding));
|
||||
}
|
||||
|
||||
public static Operand Load(this EmitterContext context, StorageKind storageKind, int binding, Operand e0)
|
||||
{
|
||||
return context.Add(Instruction.Load, storageKind, Local(), Const(binding), e0);
|
||||
}
|
||||
|
||||
public static Operand Load(this EmitterContext context, StorageKind storageKind, int binding, Operand e0, Operand e1)
|
||||
{
|
||||
return context.Add(Instruction.Load, storageKind, Local(), Const(binding), e0, e1);
|
||||
}
|
||||
|
||||
public static Operand Load(this EmitterContext context, StorageKind storageKind, int binding, Operand e0, Operand e1, Operand e2)
|
||||
{
|
||||
return context.Add(Instruction.Load, storageKind, Local(), Const(binding), e0, e1, e2);
|
||||
}
|
||||
|
||||
public static Operand Load(this EmitterContext context, StorageKind storageKind, IoVariable ioVariable, Operand primVertex = null)
|
||||
{
|
||||
return primVertex != null
|
||||
? context.Add(Instruction.Load, storageKind, Local(), Const((int)ioVariable), primVertex)
|
||||
: context.Add(Instruction.Load, storageKind, Local(), Const((int)ioVariable));
|
||||
? context.Load(storageKind, (int)ioVariable, primVertex)
|
||||
: context.Load(storageKind, (int)ioVariable);
|
||||
}
|
||||
|
||||
public static Operand Load(
|
||||
|
|
@ -564,8 +584,8 @@ namespace Ryujinx.Graphics.Shader.Translation
|
|||
Operand elemIndex)
|
||||
{
|
||||
return primVertex != null
|
||||
? context.Add(Instruction.Load, storageKind, Local(), Const((int)ioVariable), primVertex, elemIndex)
|
||||
: context.Add(Instruction.Load, storageKind, Local(), Const((int)ioVariable), elemIndex);
|
||||
? context.Load(storageKind, (int)ioVariable, primVertex, elemIndex)
|
||||
: context.Load(storageKind, (int)ioVariable, elemIndex);
|
||||
}
|
||||
|
||||
public static Operand Load(
|
||||
|
|
@ -577,22 +597,8 @@ namespace Ryujinx.Graphics.Shader.Translation
|
|||
Operand elemIndex)
|
||||
{
|
||||
return primVertex != null
|
||||
? context.Add(Instruction.Load, storageKind, Local(), Const((int)ioVariable), primVertex, arrayIndex, elemIndex)
|
||||
: context.Add(Instruction.Load, storageKind, Local(), Const((int)ioVariable), arrayIndex, elemIndex);
|
||||
}
|
||||
|
||||
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);
|
||||
? context.Load(storageKind, (int)ioVariable, primVertex, arrayIndex, elemIndex)
|
||||
: context.Load(storageKind, (int)ioVariable, arrayIndex, elemIndex);
|
||||
}
|
||||
|
||||
public static Operand LoadGlobal(this EmitterContext context, Operand a, Operand b)
|
||||
|
|
|
|||
|
|
@ -19,7 +19,6 @@ namespace Ryujinx.Graphics.Shader.Translation
|
|||
InstanceId = 1 << 3,
|
||||
DrawParameters = 1 << 4,
|
||||
RtLayer = 1 << 5,
|
||||
CbIndexing = 1 << 6,
|
||||
IaIndexing = 1 << 7,
|
||||
OaIndexing = 1 << 8,
|
||||
FixedFuncAttr = 1 << 9
|
||||
|
|
|
|||
|
|
@ -32,25 +32,49 @@ namespace Ryujinx.Graphics.Shader.Translation.Optimizations
|
|||
continue;
|
||||
}
|
||||
|
||||
if (handleAsgOp.Inst != Instruction.LoadConstant)
|
||||
if (handleAsgOp.Inst != Instruction.Load ||
|
||||
handleAsgOp.StorageKind != StorageKind.ConstantBuffer ||
|
||||
handleAsgOp.SourcesCount != 4)
|
||||
{
|
||||
continue;
|
||||
}
|
||||
|
||||
Operand ldcSrc0 = handleAsgOp.GetSource(0);
|
||||
|
||||
if (ldcSrc0.Type != OperandType.Constant ||
|
||||
!config.ResourceManager.TryGetConstantBufferSlot(ldcSrc0.Value, out int src0CbufSlot) ||
|
||||
src0CbufSlot != 2)
|
||||
{
|
||||
continue;
|
||||
}
|
||||
|
||||
Operand ldcSrc1 = handleAsgOp.GetSource(1);
|
||||
|
||||
if (ldcSrc0.Type != OperandType.Constant || ldcSrc0.Value != 2)
|
||||
// We expect field index 0 to be accessed.
|
||||
if (ldcSrc1.Type != OperandType.Constant || ldcSrc1.Value != 0)
|
||||
{
|
||||
continue;
|
||||
}
|
||||
|
||||
if (!(ldcSrc1.AsgOp is Operation shrOp) || shrOp.Inst != Instruction.ShiftRightU32)
|
||||
Operand ldcSrc2 = handleAsgOp.GetSource(2);
|
||||
|
||||
// FIXME: This is missing some checks, for example, a check to ensure that the shift value is 2.
|
||||
// Might be not worth fixing since if that doesn't kick in, the result will be no texture
|
||||
// to access anyway which is also wrong.
|
||||
// Plus this whole transform is fundamentally flawed as-is since we have no way to know the array size.
|
||||
// Eventually, this should be entirely removed in favor of a implementation that supports true bindless
|
||||
// texture access.
|
||||
if (!(ldcSrc2.AsgOp is Operation shrOp) || shrOp.Inst != Instruction.ShiftRightU32)
|
||||
{
|
||||
continue;
|
||||
}
|
||||
|
||||
if (!(shrOp.GetSource(0).AsgOp is Operation addOp) || addOp.Inst != Instruction.Add)
|
||||
if (!(shrOp.GetSource(0).AsgOp is Operation shrOp2) || shrOp2.Inst != Instruction.ShiftRightU32)
|
||||
{
|
||||
continue;
|
||||
}
|
||||
|
||||
if (!(shrOp2.GetSource(0).AsgOp is Operation addOp) || addOp.Inst != Instruction.Add)
|
||||
{
|
||||
continue;
|
||||
}
|
||||
|
|
|
|||
|
|
@ -8,7 +8,7 @@ namespace Ryujinx.Graphics.Shader.Translation.Optimizations
|
|||
{
|
||||
static class ConstantFolding
|
||||
{
|
||||
public static void RunPass(Operation operation)
|
||||
public static void RunPass(ShaderConfig config, Operation operation)
|
||||
{
|
||||
if (!AreAllSourcesConstant(operation))
|
||||
{
|
||||
|
|
@ -153,8 +153,21 @@ namespace Ryujinx.Graphics.Shader.Translation.Optimizations
|
|||
EvaluateFPUnary(operation, (x) => float.IsNaN(x));
|
||||
break;
|
||||
|
||||
case Instruction.LoadConstant:
|
||||
operation.TurnIntoCopy(Cbuf(operation.GetSource(0).Value, operation.GetSource(1).Value));
|
||||
case Instruction.Load:
|
||||
if (operation.StorageKind == StorageKind.ConstantBuffer && operation.SourcesCount == 4)
|
||||
{
|
||||
int binding = operation.GetSource(0).Value;
|
||||
int fieldIndex = operation.GetSource(1).Value;
|
||||
|
||||
if (config.ResourceManager.TryGetConstantBufferSlot(binding, out int cbufSlot) && fieldIndex == 0)
|
||||
{
|
||||
int vecIndex = operation.GetSource(2).Value;
|
||||
int elemIndex = operation.GetSource(3).Value;
|
||||
int cbufOffset = vecIndex * 4 + elemIndex;
|
||||
|
||||
operation.TurnIntoCopy(Cbuf(cbufSlot, cbufOffset));
|
||||
}
|
||||
}
|
||||
break;
|
||||
|
||||
case Instruction.Maximum:
|
||||
|
|
|
|||
|
|
@ -347,21 +347,23 @@ namespace Ryujinx.Graphics.Shader.Translation.Optimizations
|
|||
return wordOffset;
|
||||
}
|
||||
|
||||
Operand[] sources = new Operand[operation.SourcesCount];
|
||||
Operand cbufOffset = GetCbufOffset();
|
||||
Operand vecIndex = Local();
|
||||
Operand elemIndex = Local();
|
||||
|
||||
node.List.AddBefore(node, new Operation(Instruction.ShiftRightU32, 0, vecIndex, cbufOffset, Const(2)));
|
||||
node.List.AddBefore(node, new Operation(Instruction.BitwiseAnd, 0, elemIndex, cbufOffset, Const(3)));
|
||||
|
||||
Operand[] sources = new Operand[4];
|
||||
|
||||
int cbSlot = UbeFirstCbuf + storageIndex;
|
||||
|
||||
sources[0] = Const(cbSlot);
|
||||
sources[1] = GetCbufOffset();
|
||||
sources[0] = Const(config.ResourceManager.GetConstantBufferBinding(cbSlot));
|
||||
sources[1] = Const(0);
|
||||
sources[2] = vecIndex;
|
||||
sources[3] = elemIndex;
|
||||
|
||||
config.SetUsedConstantBuffer(cbSlot);
|
||||
|
||||
for (int index = 2; index < operation.SourcesCount; index++)
|
||||
{
|
||||
sources[index] = operation.GetSource(index);
|
||||
}
|
||||
|
||||
Operation ldcOp = new Operation(Instruction.LoadConstant, operation.Dest, sources);
|
||||
Operation ldcOp = new Operation(Instruction.Load, StorageKind.ConstantBuffer, operation.Dest, sources);
|
||||
|
||||
for (int index = 0; index < operation.SourcesCount; index++)
|
||||
{
|
||||
|
|
|
|||
|
|
@ -9,7 +9,7 @@ namespace Ryujinx.Graphics.Shader.Translation.Optimizations
|
|||
{
|
||||
public static void RunPass(BasicBlock[] blocks, ShaderConfig config)
|
||||
{
|
||||
RunOptimizationPasses(blocks);
|
||||
RunOptimizationPasses(blocks, config);
|
||||
|
||||
int sbUseMask = 0;
|
||||
int ubeUseMask = 0;
|
||||
|
|
@ -31,10 +31,10 @@ namespace Ryujinx.Graphics.Shader.Translation.Optimizations
|
|||
config.SetAccessibleBufferMasks(sbUseMask, ubeUseMask);
|
||||
|
||||
// Run optimizations one last time to remove any code that is now optimizable after above passes.
|
||||
RunOptimizationPasses(blocks);
|
||||
RunOptimizationPasses(blocks, config);
|
||||
}
|
||||
|
||||
private static void RunOptimizationPasses(BasicBlock[] blocks)
|
||||
private static void RunOptimizationPasses(BasicBlock[] blocks, ShaderConfig config)
|
||||
{
|
||||
bool modified;
|
||||
|
||||
|
|
@ -73,7 +73,7 @@ namespace Ryujinx.Graphics.Shader.Translation.Optimizations
|
|||
continue;
|
||||
}
|
||||
|
||||
ConstantFolding.RunPass(operation);
|
||||
ConstantFolding.RunPass(config, operation);
|
||||
Simplification.RunPass(operation);
|
||||
|
||||
if (DestIsLocalVar(operation))
|
||||
|
|
|
|||
126
src/Ryujinx.Graphics.Shader/Translation/ResourceManager.cs
Normal file
126
src/Ryujinx.Graphics.Shader/Translation/ResourceManager.cs
Normal file
|
|
@ -0,0 +1,126 @@
|
|||
using Ryujinx.Graphics.Shader.StructuredIr;
|
||||
using System;
|
||||
using System.Collections.Generic;
|
||||
using System.Globalization;
|
||||
|
||||
namespace Ryujinx.Graphics.Shader.Translation
|
||||
{
|
||||
class ResourceManager
|
||||
{
|
||||
private static readonly string[] _stagePrefixes = new string[] { "cp", "vp", "tcp", "tep", "gp", "fp" };
|
||||
|
||||
private readonly IGpuAccessor _gpuAccessor;
|
||||
private readonly ShaderProperties _properties;
|
||||
private readonly string _stagePrefix;
|
||||
|
||||
private readonly int[] _cbSlotToBindingMap;
|
||||
|
||||
private readonly HashSet<int> _usedConstantBufferBindings;
|
||||
|
||||
public ResourceManager(ShaderStage stage, IGpuAccessor gpuAccessor, ShaderProperties properties)
|
||||
{
|
||||
_gpuAccessor = gpuAccessor;
|
||||
_properties = properties;
|
||||
_stagePrefix = GetShaderStagePrefix(stage);
|
||||
|
||||
_cbSlotToBindingMap = new int[18];
|
||||
_cbSlotToBindingMap.AsSpan().Fill(-1);
|
||||
|
||||
_usedConstantBufferBindings = new HashSet<int>();
|
||||
|
||||
properties.AddConstantBuffer(0, new BufferDefinition(BufferLayout.Std140, 0, 0, "support_buffer", SupportBuffer.GetStructureType()));
|
||||
}
|
||||
|
||||
public int GetConstantBufferBinding(int slot)
|
||||
{
|
||||
int binding = _cbSlotToBindingMap[slot];
|
||||
if (binding < 0)
|
||||
{
|
||||
binding = _gpuAccessor.QueryBindingConstantBuffer(slot);
|
||||
_cbSlotToBindingMap[slot] = binding;
|
||||
string slotNumber = slot.ToString(CultureInfo.InvariantCulture);
|
||||
AddNewConstantBuffer(binding, $"{_stagePrefix}_c{slotNumber}");
|
||||
}
|
||||
|
||||
return binding;
|
||||
}
|
||||
|
||||
public bool TryGetConstantBufferSlot(int binding, out int slot)
|
||||
{
|
||||
for (slot = 0; slot < _cbSlotToBindingMap.Length; slot++)
|
||||
{
|
||||
if (_cbSlotToBindingMap[slot] == binding)
|
||||
{
|
||||
return true;
|
||||
}
|
||||
}
|
||||
|
||||
slot = 0;
|
||||
return false;
|
||||
}
|
||||
|
||||
public void SetUsedConstantBufferBinding(int binding)
|
||||
{
|
||||
_usedConstantBufferBindings.Add(binding);
|
||||
}
|
||||
|
||||
public BufferDescriptor[] GetConstantBufferDescriptors()
|
||||
{
|
||||
var descriptors = new BufferDescriptor[_usedConstantBufferBindings.Count];
|
||||
|
||||
int descriptorIndex = 0;
|
||||
|
||||
for (int slot = 0; slot < _cbSlotToBindingMap.Length; slot++)
|
||||
{
|
||||
int binding = _cbSlotToBindingMap[slot];
|
||||
|
||||
if (binding >= 0 && _usedConstantBufferBindings.Contains(binding))
|
||||
{
|
||||
descriptors[descriptorIndex++] = new BufferDescriptor(binding, slot);
|
||||
}
|
||||
}
|
||||
|
||||
if (descriptors.Length != descriptorIndex)
|
||||
{
|
||||
Array.Resize(ref descriptors, descriptorIndex);
|
||||
}
|
||||
|
||||
return descriptors;
|
||||
}
|
||||
|
||||
private void AddNewConstantBuffer(int binding, string name)
|
||||
{
|
||||
StructureType type = new StructureType(new[]
|
||||
{
|
||||
new StructureField(AggregateType.Array | AggregateType.Vector4 | AggregateType.FP32, "data", Constants.ConstantBufferSize / 16)
|
||||
});
|
||||
|
||||
_properties.AddConstantBuffer(binding, new BufferDefinition(BufferLayout.Std140, 0, binding, name, type));
|
||||
}
|
||||
|
||||
public void InheritFrom(ResourceManager other)
|
||||
{
|
||||
for (int i = 0; i < other._cbSlotToBindingMap.Length; i++)
|
||||
{
|
||||
int binding = other._cbSlotToBindingMap[i];
|
||||
|
||||
if (binding >= 0)
|
||||
{
|
||||
_cbSlotToBindingMap[i] = binding;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
public static string GetShaderStagePrefix(ShaderStage stage)
|
||||
{
|
||||
uint index = (uint)stage;
|
||||
|
||||
if (index >= _stagePrefixes.Length)
|
||||
{
|
||||
return "invalid";
|
||||
}
|
||||
|
||||
return _stagePrefixes[index];
|
||||
}
|
||||
}
|
||||
}
|
||||
|
|
@ -1,7 +1,6 @@
|
|||
using Ryujinx.Graphics.Shader.Decoders;
|
||||
using Ryujinx.Graphics.Shader.IntermediateRepresentation;
|
||||
using Ryujinx.Graphics.Shader.StructuredIr;
|
||||
using System.Collections.Generic;
|
||||
using System.Diagnostics;
|
||||
using System.Linq;
|
||||
using System.Numerics;
|
||||
|
||||
|
|
@ -16,6 +15,7 @@ namespace Ryujinx.Graphics.Shader.Translation
|
|||
{
|
||||
bool isVertexShader = config.Stage == ShaderStage.Vertex;
|
||||
bool hasConstantBufferDrawParameters = config.GpuAccessor.QueryHasConstantBufferDrawParameters();
|
||||
bool hasVectorIndexingBug = config.GpuAccessor.QueryHostHasVectorIndexingBug();
|
||||
bool supportsSnormBufferTextureFormat = config.GpuAccessor.QueryHostSupportsSnormBufferTextureFormat();
|
||||
|
||||
for (int blkIndex = 0; blkIndex < blocks.Length; blkIndex++)
|
||||
|
|
@ -45,6 +45,11 @@ namespace Ryujinx.Graphics.Shader.Translation
|
|||
}
|
||||
}
|
||||
|
||||
if (hasVectorIndexingBug)
|
||||
{
|
||||
InsertVectorComponentSelect(node, config);
|
||||
}
|
||||
|
||||
LinkedListNode<INode> nextNode = node.Next;
|
||||
|
||||
if (operation is TextureOperation texOp)
|
||||
|
|
@ -71,6 +76,84 @@ namespace Ryujinx.Graphics.Shader.Translation
|
|||
}
|
||||
}
|
||||
|
||||
private static void InsertVectorComponentSelect(LinkedListNode<INode> node, ShaderConfig config)
|
||||
{
|
||||
Operation operation = (Operation)node.Value;
|
||||
|
||||
if (operation.Inst != Instruction.Load ||
|
||||
operation.StorageKind != StorageKind.ConstantBuffer ||
|
||||
operation.SourcesCount < 3)
|
||||
{
|
||||
return;
|
||||
}
|
||||
|
||||
Operand bindingIndex = operation.GetSource(0);
|
||||
Operand fieldIndex = operation.GetSource(1);
|
||||
Operand elemIndex = operation.GetSource(operation.SourcesCount - 1);
|
||||
|
||||
if (bindingIndex.Type != OperandType.Constant ||
|
||||
fieldIndex.Type != OperandType.Constant ||
|
||||
elemIndex.Type == OperandType.Constant)
|
||||
{
|
||||
return;
|
||||
}
|
||||
|
||||
BufferDefinition buffer = config.Properties.ConstantBuffers[bindingIndex.Value];
|
||||
StructureField field = buffer.Type.Fields[fieldIndex.Value];
|
||||
|
||||
int elemCount = (field.Type & AggregateType.ElementCountMask) switch
|
||||
{
|
||||
AggregateType.Vector2 => 2,
|
||||
AggregateType.Vector3 => 3,
|
||||
AggregateType.Vector4 => 4,
|
||||
_ => 1
|
||||
};
|
||||
|
||||
if (elemCount == 1)
|
||||
{
|
||||
return;
|
||||
}
|
||||
|
||||
Operand result = null;
|
||||
|
||||
for (int i = 0; i < elemCount; i++)
|
||||
{
|
||||
Operand value = Local();
|
||||
Operand[] inputs = new Operand[operation.SourcesCount];
|
||||
|
||||
for (int srcIndex = 0; srcIndex < inputs.Length - 1; srcIndex++)
|
||||
{
|
||||
inputs[srcIndex] = operation.GetSource(srcIndex);
|
||||
}
|
||||
|
||||
inputs[inputs.Length - 1] = Const(i);
|
||||
|
||||
Operation loadOp = new Operation(Instruction.Load, StorageKind.ConstantBuffer, value, inputs);
|
||||
|
||||
node.List.AddBefore(node, loadOp);
|
||||
|
||||
if (i == 0)
|
||||
{
|
||||
result = value;
|
||||
}
|
||||
else
|
||||
{
|
||||
Operand isCurrentIndex = Local();
|
||||
Operand selection = Local();
|
||||
|
||||
Operation compareOp = new Operation(Instruction.CompareEqual, isCurrentIndex, new Operand[] { elemIndex, Const(i) });
|
||||
Operation selectOp = new Operation(Instruction.ConditionalSelect, selection, new Operand[] { isCurrentIndex, value, result });
|
||||
|
||||
node.List.AddBefore(node, compareOp);
|
||||
node.List.AddBefore(node, selectOp);
|
||||
|
||||
result = selection;
|
||||
}
|
||||
}
|
||||
|
||||
operation.TurnIntoCopy(result);
|
||||
}
|
||||
|
||||
private static LinkedListNode<INode> RewriteGlobalAccess(LinkedListNode<INode> node, ShaderConfig config)
|
||||
{
|
||||
Operation operation = (Operation)node.Value;
|
||||
|
|
@ -90,6 +173,15 @@ namespace Ryujinx.Graphics.Shader.Translation
|
|||
return local;
|
||||
}
|
||||
|
||||
Operand PrependStorageOperation(Instruction inst, StorageKind storageKind, params Operand[] sources)
|
||||
{
|
||||
Operand local = Local();
|
||||
|
||||
node.List.AddBefore(node, new Operation(inst, storageKind, local, sources));
|
||||
|
||||
return local;
|
||||
}
|
||||
|
||||
Operand PrependExistingOperation(Operation operation)
|
||||
{
|
||||
Operand local = Local();
|
||||
|
|
@ -204,8 +296,6 @@ namespace Ryujinx.Graphics.Shader.Translation
|
|||
|
||||
cbeUseMask &= ~(1 << slot);
|
||||
|
||||
config.SetUsedConstantBuffer(cbSlot);
|
||||
|
||||
Operand previousResult = PrependExistingOperation(storageOp);
|
||||
|
||||
int cbOffset = GetConstantUbeOffset(slot);
|
||||
|
|
@ -216,18 +306,17 @@ namespace Ryujinx.Graphics.Shader.Translation
|
|||
Operand byteOffsetConst = PrependOperation(Instruction.Subtract, addrLow, baseAddrTruncConst);
|
||||
|
||||
Operand cbIndex = PrependOperation(Instruction.ShiftRightU32, byteOffsetConst, Const(2));
|
||||
Operand vecIndex = PrependOperation(Instruction.ShiftRightU32, cbIndex, Const(2));
|
||||
Operand elemIndex = PrependOperation(Instruction.BitwiseAnd, cbIndex, Const(3));
|
||||
|
||||
Operand[] sourcesCb = new Operand[operation.SourcesCount];
|
||||
Operand[] sourcesCb = new Operand[4];
|
||||
|
||||
sourcesCb[0] = Const(cbSlot);
|
||||
sourcesCb[1] = cbIndex;
|
||||
sourcesCb[0] = Const(config.ResourceManager.GetConstantBufferBinding(cbSlot));
|
||||
sourcesCb[1] = Const(0);
|
||||
sourcesCb[2] = vecIndex;
|
||||
sourcesCb[3] = elemIndex;
|
||||
|
||||
for (int index = 2; index < operation.SourcesCount; index++)
|
||||
{
|
||||
sourcesCb[index] = operation.GetSource(index);
|
||||
}
|
||||
|
||||
Operand ldcResult = PrependOperation(Instruction.LoadConstant, sourcesCb);
|
||||
Operand ldcResult = PrependStorageOperation(Instruction.Load, StorageKind.ConstantBuffer, sourcesCb);
|
||||
|
||||
storageOp = new Operation(Instruction.ConditionalSelect, operation.Dest, inRange, ldcResult, previousResult);
|
||||
}
|
||||
|
|
|
|||
|
|
@ -39,6 +39,10 @@ namespace Ryujinx.Graphics.Shader.Translation
|
|||
|
||||
public TranslationOptions Options { get; }
|
||||
|
||||
public ShaderProperties Properties { get; }
|
||||
|
||||
public ResourceManager ResourceManager { get; }
|
||||
|
||||
public bool TransformFeedbackEnabled { get; }
|
||||
|
||||
private TransformFeedbackOutput[] _transformFeedbackOutputs;
|
||||
|
|
@ -109,7 +113,6 @@ namespace Ryujinx.Graphics.Shader.Translation
|
|||
public int AccessibleStorageBuffersMask { get; private set; }
|
||||
public int AccessibleConstantBuffersMask { get; private set; }
|
||||
|
||||
private int _usedConstantBuffers;
|
||||
private int _usedStorageBuffers;
|
||||
private int _usedStorageBuffersWrite;
|
||||
|
||||
|
|
@ -128,20 +131,17 @@ namespace Ryujinx.Graphics.Shader.Translation
|
|||
private readonly Dictionary<int, int> _sbSlots;
|
||||
private readonly Dictionary<int, int> _sbSlotsReverse;
|
||||
|
||||
private BufferDescriptor[] _cachedConstantBufferDescriptors;
|
||||
private BufferDescriptor[] _cachedStorageBufferDescriptors;
|
||||
private TextureDescriptor[] _cachedTextureDescriptors;
|
||||
private TextureDescriptor[] _cachedImageDescriptors;
|
||||
|
||||
private int _firstConstantBufferBinding;
|
||||
private int _firstStorageBufferBinding;
|
||||
|
||||
public int FirstConstantBufferBinding => _firstConstantBufferBinding;
|
||||
public int FirstStorageBufferBinding => _firstStorageBufferBinding;
|
||||
|
||||
public ShaderConfig(IGpuAccessor gpuAccessor, TranslationOptions options)
|
||||
public ShaderConfig(ShaderStage stage, IGpuAccessor gpuAccessor, TranslationOptions options)
|
||||
{
|
||||
Stage = ShaderStage.Compute;
|
||||
Stage = stage;
|
||||
GpuAccessor = gpuAccessor;
|
||||
Options = options;
|
||||
|
||||
|
|
@ -158,6 +158,9 @@ namespace Ryujinx.Graphics.Shader.Translation
|
|||
|
||||
_sbSlots = new Dictionary<int, int>();
|
||||
_sbSlotsReverse = new Dictionary<int, int>();
|
||||
|
||||
Properties = new ShaderProperties();
|
||||
ResourceManager = new ResourceManager(stage, gpuAccessor, Properties);
|
||||
}
|
||||
|
||||
public ShaderConfig(
|
||||
|
|
@ -165,9 +168,8 @@ namespace Ryujinx.Graphics.Shader.Translation
|
|||
OutputTopology outputTopology,
|
||||
int maxOutputVertices,
|
||||
IGpuAccessor gpuAccessor,
|
||||
TranslationOptions options) : this(gpuAccessor, options)
|
||||
TranslationOptions options) : this(stage, gpuAccessor, options)
|
||||
{
|
||||
Stage = stage;
|
||||
ThreadsPerInputPrimitive = 1;
|
||||
OutputTopology = outputTopology;
|
||||
MaxOutputVertices = maxOutputVertices;
|
||||
|
|
@ -179,9 +181,8 @@ namespace Ryujinx.Graphics.Shader.Translation
|
|||
}
|
||||
}
|
||||
|
||||
public ShaderConfig(ShaderHeader header, IGpuAccessor gpuAccessor, TranslationOptions options) : this(gpuAccessor, options)
|
||||
public ShaderConfig(ShaderHeader header, IGpuAccessor gpuAccessor, TranslationOptions options) : this(header.Stage, gpuAccessor, options)
|
||||
{
|
||||
Stage = header.Stage;
|
||||
GpPassthrough = header.Stage == ShaderStage.Geometry && header.GpPassthrough;
|
||||
ThreadsPerInputPrimitive = header.ThreadsPerInputPrimitive;
|
||||
OutputTopology = header.OutputTopology;
|
||||
|
|
@ -428,12 +429,13 @@ namespace Ryujinx.Graphics.Shader.Translation
|
|||
|
||||
public void InheritFrom(ShaderConfig other)
|
||||
{
|
||||
ResourceManager.InheritFrom(other.ResourceManager);
|
||||
|
||||
ClipDistancesWritten |= other.ClipDistancesWritten;
|
||||
UsedFeatures |= other.UsedFeatures;
|
||||
|
||||
UsedInputAttributes |= other.UsedInputAttributes;
|
||||
UsedOutputAttributes |= other.UsedOutputAttributes;
|
||||
_usedConstantBuffers |= other._usedConstantBuffers;
|
||||
_usedStorageBuffers |= other._usedStorageBuffers;
|
||||
_usedStorageBuffersWrite |= other._usedStorageBuffersWrite;
|
||||
|
||||
|
|
@ -641,11 +643,6 @@ namespace Ryujinx.Graphics.Shader.Translation
|
|||
AccessibleConstantBuffersMask = ubeMask;
|
||||
}
|
||||
|
||||
public void SetUsedConstantBuffer(int slot)
|
||||
{
|
||||
_usedConstantBuffers |= 1 << slot;
|
||||
}
|
||||
|
||||
public void SetUsedStorageBuffer(int slot, bool write)
|
||||
{
|
||||
int mask = 1 << slot;
|
||||
|
|
@ -762,27 +759,6 @@ namespace Ryujinx.Graphics.Shader.Translation
|
|||
return meta;
|
||||
}
|
||||
|
||||
public BufferDescriptor[] GetConstantBufferDescriptors()
|
||||
{
|
||||
if (_cachedConstantBufferDescriptors != null)
|
||||
{
|
||||
return _cachedConstantBufferDescriptors;
|
||||
}
|
||||
|
||||
int usedMask = _usedConstantBuffers;
|
||||
|
||||
if (UsedFeatures.HasFlag(FeatureFlags.CbIndexing))
|
||||
{
|
||||
usedMask |= (int)GpuAccessor.QueryConstantBufferUse();
|
||||
}
|
||||
|
||||
return _cachedConstantBufferDescriptors = GetUniformBufferDescriptors(
|
||||
usedMask,
|
||||
UsedFeatures.HasFlag(FeatureFlags.CbIndexing),
|
||||
out _firstConstantBufferBinding,
|
||||
GpuAccessor.QueryBindingConstantBuffer);
|
||||
}
|
||||
|
||||
public BufferDescriptor[] GetStorageBufferDescriptors()
|
||||
{
|
||||
if (_cachedStorageBufferDescriptors != null)
|
||||
|
|
@ -798,47 +774,6 @@ namespace Ryujinx.Graphics.Shader.Translation
|
|||
GpuAccessor.QueryBindingStorageBuffer);
|
||||
}
|
||||
|
||||
private static BufferDescriptor[] GetUniformBufferDescriptors(int usedMask, bool isArray, out int firstBinding, Func<int, int> getBindingCallback)
|
||||
{
|
||||
firstBinding = 0;
|
||||
int lastSlot = -1;
|
||||
bool hasFirstBinding = false;
|
||||
var descriptors = new BufferDescriptor[BitOperations.PopCount((uint)usedMask)];
|
||||
|
||||
for (int i = 0; i < descriptors.Length; i++)
|
||||
{
|
||||
int slot = BitOperations.TrailingZeroCount(usedMask);
|
||||
|
||||
if (isArray)
|
||||
{
|
||||
// The next array entries also consumes bindings, even if they are unused.
|
||||
for (int j = lastSlot + 1; j < slot; j++)
|
||||
{
|
||||
int binding = getBindingCallback(j);
|
||||
|
||||
if (!hasFirstBinding)
|
||||
{
|
||||
firstBinding = binding;
|
||||
hasFirstBinding = true;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
lastSlot = slot;
|
||||
descriptors[i] = new BufferDescriptor(getBindingCallback(slot), slot);
|
||||
|
||||
if (!hasFirstBinding)
|
||||
{
|
||||
firstBinding = descriptors[i].Binding;
|
||||
hasFirstBinding = true;
|
||||
}
|
||||
|
||||
usedMask &= ~(1 << slot);
|
||||
}
|
||||
|
||||
return descriptors;
|
||||
}
|
||||
|
||||
private BufferDescriptor[] GetStorageBufferDescriptors(
|
||||
int usedMask,
|
||||
int writtenMask,
|
||||
|
|
@ -1009,7 +944,7 @@ namespace Ryujinx.Graphics.Shader.Translation
|
|||
public ShaderProgramInfo CreateProgramInfo(ShaderIdentification identification = ShaderIdentification.None)
|
||||
{
|
||||
return new ShaderProgramInfo(
|
||||
GetConstantBufferDescriptors(),
|
||||
ResourceManager.GetConstantBufferDescriptors(),
|
||||
GetStorageBufferDescriptors(),
|
||||
GetTextureDescriptors(),
|
||||
GetImageDescriptors(),
|
||||
|
|
|
|||
|
|
@ -99,7 +99,7 @@ namespace Ryujinx.Graphics.Shader.Translation
|
|||
|
||||
if (options.Flags.HasFlag(TranslationFlags.Compute))
|
||||
{
|
||||
config = new ShaderConfig(gpuAccessor, options);
|
||||
config = new ShaderConfig(ShaderStage.Compute, gpuAccessor, options);
|
||||
|
||||
program = Decoder.Decode(config, address);
|
||||
}
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue