Replace constant buffer access on shader with new Load instruction (#4646)

This commit is contained in:
gdkchan 2023-05-20 16:19:26 -03:00 committed by GitHub
parent fb27042e01
commit 402f05b8ef
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
42 changed files with 788 additions and 625 deletions

View file

@ -23,9 +23,7 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Spirv
public int InputVertices { get; }
public Dictionary<int, Instruction> UniformBuffers { get; } = new Dictionary<int, Instruction>();
public Instruction SupportBuffer { get; set; }
public Instruction UniformBuffersArray { get; set; }
public Dictionary<int, Instruction> ConstantBuffers { get; } = new Dictionary<int, Instruction>();
public Instruction StorageBuffersArray { get; set; }
public Instruction LocalMemory { get; set; }
public Instruction SharedMemory { get; set; }
@ -217,7 +215,6 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Spirv
{
IrOperandType.Argument => GetArgument(type, operand),
IrOperandType.Constant => GetConstant(type, operand),
IrOperandType.ConstantBuffer => GetConstantBuffer(type, operand),
IrOperandType.LocalVariable => GetLocal(type, operand),
IrOperandType.Undefined => GetUndefined(type),
_ => throw new ArgumentException($"Invalid operand type \"{operand.Type}\".")
@ -274,31 +271,6 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Spirv
};
}
public Instruction GetConstantBuffer(AggregateType type, AstOperand operand)
{
var i1 = Constant(TypeS32(), 0);
var i2 = Constant(TypeS32(), operand.CbufOffset >> 2);
var i3 = Constant(TypeU32(), operand.CbufOffset & 3);
Instruction elemPointer;
if (UniformBuffersArray != null)
{
var ubVariable = UniformBuffersArray;
var i0 = Constant(TypeS32(), operand.CbufSlot);
elemPointer = AccessChain(TypePointer(StorageClass.Uniform, TypeFP32()), ubVariable, i0, i1, i2, i3);
}
else
{
var ubVariable = UniformBuffers[operand.CbufSlot];
elemPointer = AccessChain(TypePointer(StorageClass.Uniform, TypeFP32()), ubVariable, i1, i2, i3);
}
return BitcastIfNeeded(type, AggregateType.FP32, Load(TypeFP32(), elemPointer));
}
public Instruction GetLocalPointer(AstOperand local)
{
return _locals[local];

View file

@ -98,8 +98,7 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Spirv
DeclareLocalMemory(context, localMemorySize);
}
DeclareSupportBuffer(context);
DeclareUniformBuffers(context, context.Config.GetConstantBufferDescriptors());
DeclareConstantBuffers(context, context.Config.Properties.ConstantBuffers.Values);
DeclareStorageBuffers(context, context.Config.GetStorageBufferDescriptors());
DeclareSamplers(context, context.Config.GetTextureDescriptors());
DeclareImages(context, context.Config.GetImageDescriptors());
@ -127,84 +126,63 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Spirv
return variable;
}
private static void DeclareSupportBuffer(CodeGenContext context)
private static void DeclareConstantBuffers(CodeGenContext context, IEnumerable<BufferDefinition> buffers)
{
if (!context.Config.Stage.SupportsRenderScale() && !(context.Config.LastInVertexPipeline && context.Config.GpuAccessor.QueryViewportTransformDisable()))
HashSet<SpvInstruction> decoratedTypes = new HashSet<SpvInstruction>();
foreach (BufferDefinition buffer in buffers)
{
return;
}
int alignment = buffer.Layout == BufferLayout.Std140 ? 16 : 4;
int alignmentMask = alignment - 1;
int offset = 0;
var isBgraArrayType = context.TypeArray(context.TypeU32(), context.Constant(context.TypeU32(), SupportBuffer.FragmentIsBgraCount));
var viewportInverseVectorType = context.TypeVector(context.TypeFP32(), 4);
var renderScaleArrayType = context.TypeArray(context.TypeFP32(), context.Constant(context.TypeU32(), SupportBuffer.RenderScaleMaxCount));
SpvInstruction[] structFieldTypes = new SpvInstruction[buffer.Type.Fields.Length];
int[] structFieldOffsets = new int[buffer.Type.Fields.Length];
context.Decorate(isBgraArrayType, Decoration.ArrayStride, (LiteralInteger)SupportBuffer.FieldSize);
context.Decorate(renderScaleArrayType, Decoration.ArrayStride, (LiteralInteger)SupportBuffer.FieldSize);
for (int fieldIndex = 0; fieldIndex < buffer.Type.Fields.Length; fieldIndex++)
{
StructureField field = buffer.Type.Fields[fieldIndex];
int fieldSize = (field.Type.GetSizeInBytes() + alignmentMask) & ~alignmentMask;
var supportBufferStructType = context.TypeStruct(false, context.TypeU32(), isBgraArrayType, viewportInverseVectorType, context.TypeS32(), renderScaleArrayType);
structFieldTypes[fieldIndex] = context.GetType(field.Type, field.ArrayLength);
structFieldOffsets[fieldIndex] = offset;
context.MemberDecorate(supportBufferStructType, 0, Decoration.Offset, (LiteralInteger)SupportBuffer.FragmentAlphaTestOffset);
context.MemberDecorate(supportBufferStructType, 1, Decoration.Offset, (LiteralInteger)SupportBuffer.FragmentIsBgraOffset);
context.MemberDecorate(supportBufferStructType, 2, Decoration.Offset, (LiteralInteger)SupportBuffer.ViewportInverseOffset);
context.MemberDecorate(supportBufferStructType, 3, Decoration.Offset, (LiteralInteger)SupportBuffer.FragmentRenderScaleCountOffset);
context.MemberDecorate(supportBufferStructType, 4, Decoration.Offset, (LiteralInteger)SupportBuffer.GraphicsRenderScaleOffset);
context.Decorate(supportBufferStructType, Decoration.Block);
if (field.Type.HasFlag(AggregateType.Array))
{
// We can't decorate the type more than once.
if (decoratedTypes.Add(structFieldTypes[fieldIndex]))
{
context.Decorate(structFieldTypes[fieldIndex], Decoration.ArrayStride, (LiteralInteger)fieldSize);
}
var supportBufferPointerType = context.TypePointer(StorageClass.Uniform, supportBufferStructType);
var supportBufferVariable = context.Variable(supportBufferPointerType, StorageClass.Uniform);
offset += fieldSize * field.ArrayLength;
}
else
{
offset += fieldSize;
}
}
context.Decorate(supportBufferVariable, Decoration.DescriptorSet, (LiteralInteger)0);
context.Decorate(supportBufferVariable, Decoration.Binding, (LiteralInteger)0);
var ubStructType = context.TypeStruct(false, structFieldTypes);
context.AddGlobalVariable(supportBufferVariable);
if (decoratedTypes.Add(ubStructType))
{
context.Decorate(ubStructType, Decoration.Block);
context.SupportBuffer = supportBufferVariable;
}
for (int fieldIndex = 0; fieldIndex < structFieldOffsets.Length; fieldIndex++)
{
context.MemberDecorate(ubStructType, fieldIndex, Decoration.Offset, (LiteralInteger)structFieldOffsets[fieldIndex]);
}
}
private static void DeclareUniformBuffers(CodeGenContext context, BufferDescriptor[] descriptors)
{
if (descriptors.Length == 0)
{
return;
}
uint ubSize = Constants.ConstantBufferSize / 16;
var ubArrayType = context.TypeArray(context.TypeVector(context.TypeFP32(), 4), context.Constant(context.TypeU32(), ubSize), true);
context.Decorate(ubArrayType, Decoration.ArrayStride, (LiteralInteger)16);
var ubStructType = context.TypeStruct(true, ubArrayType);
context.Decorate(ubStructType, Decoration.Block);
context.MemberDecorate(ubStructType, 0, Decoration.Offset, (LiteralInteger)0);
if (context.Config.UsedFeatures.HasFlag(FeatureFlags.CbIndexing))
{
int count = descriptors.Max(x => x.Slot) + 1;
var ubStructArrayType = context.TypeArray(ubStructType, context.Constant(context.TypeU32(), count));
var ubPointerType = context.TypePointer(StorageClass.Uniform, ubStructArrayType);
var ubPointerType = context.TypePointer(StorageClass.Uniform, ubStructType);
var ubVariable = context.Variable(ubPointerType, StorageClass.Uniform);
context.Name(ubVariable, $"{GetStagePrefix(context.Config.Stage)}_u");
context.Decorate(ubVariable, Decoration.DescriptorSet, (LiteralInteger)0);
context.Decorate(ubVariable, Decoration.Binding, (LiteralInteger)context.Config.FirstConstantBufferBinding);
context.Name(ubVariable, buffer.Name);
context.Decorate(ubVariable, Decoration.DescriptorSet, (LiteralInteger)buffer.Set);
context.Decorate(ubVariable, Decoration.Binding, (LiteralInteger)buffer.Binding);
context.AddGlobalVariable(ubVariable);
context.UniformBuffersArray = ubVariable;
}
else
{
var ubPointerType = context.TypePointer(StorageClass.Uniform, ubStructType);
foreach (var descriptor in descriptors)
{
var ubVariable = context.Variable(ubPointerType, StorageClass.Uniform);
context.Name(ubVariable, $"{GetStagePrefix(context.Config.Stage)}_c{descriptor.Slot}");
context.Decorate(ubVariable, Decoration.DescriptorSet, (LiteralInteger)0);
context.Decorate(ubVariable, Decoration.Binding, (LiteralInteger)descriptor.Binding);
context.AddGlobalVariable(ubVariable);
context.UniformBuffers.Add(descriptor.Slot, ubVariable);
}
context.ConstantBuffers.Add(buffer.Binding, ubVariable);
}
}
@ -394,25 +372,11 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Spirv
{
foreach (var ioDefinition in info.IoDefinitions)
{
var ioVariable = ioDefinition.IoVariable;
// Those are actually from constant buffer, rather than being actual inputs or outputs,
// so we must ignore them here as they are declared as part of the support buffer.
// TODO: Delete this after we represent this properly on the IR (as a constant buffer rather than "input").
if (ioVariable == IoVariable.FragmentOutputIsBgra ||
ioVariable == IoVariable.SupportBlockRenderScale ||
ioVariable == IoVariable.SupportBlockViewInverse)
{
continue;
}
bool isOutput = ioDefinition.StorageKind.IsOutput();
bool isPerPatch = ioDefinition.StorageKind.IsPerPatch();
PixelImap iq = PixelImap.Unused;
if (context.Config.Stage == ShaderStage.Fragment)
{
var ioVariable = ioDefinition.IoVariable;
if (ioVariable == IoVariable.UserDefined)
{
iq = context.Config.ImapTypes[ioDefinition.Location].GetFirstUsedType();
@ -429,6 +393,9 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Spirv
}
}
bool isOutput = ioDefinition.StorageKind.IsOutput();
bool isPerPatch = ioDefinition.StorageKind.IsPerPatch();
DeclareInputOrOutput(context, ioDefinition, isOutput, isPerPatch, iq);
}
}

View file

@ -98,7 +98,6 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Spirv
Add(Instruction.ImageStore, GenerateImageStore);
Add(Instruction.IsNan, GenerateIsNan);
Add(Instruction.Load, GenerateLoad);
Add(Instruction.LoadConstant, GenerateLoadConstant);
Add(Instruction.LoadLocal, GenerateLoadLocal);
Add(Instruction.LoadShared, GenerateLoadShared);
Add(Instruction.LoadStorage, GenerateLoadStorage);
@ -313,10 +312,11 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Spirv
for (int i = 0; i < args.Length; i++)
{
var operand = (AstOperand)operation.GetSource(i + 1);
var operand = operation.GetSource(i + 1);
if (i >= function.InArguments.Length)
{
args[i] = context.GetLocalPointer(operand);
args[i] = context.GetLocalPointer((AstOperand)operand);
}
else
{
@ -867,68 +867,6 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Spirv
return GenerateLoadOrStore(context, operation, isStore: false);
}
private static OperationResult GenerateLoadConstant(CodeGenContext context, AstOperation operation)
{
var src1 = operation.GetSource(0);
var src2 = context.Get(AggregateType.S32, operation.GetSource(1));
var i1 = context.Constant(context.TypeS32(), 0);
var i2 = context.ShiftRightArithmetic(context.TypeS32(), src2, context.Constant(context.TypeS32(), 2));
var i3 = context.BitwiseAnd(context.TypeS32(), src2, context.Constant(context.TypeS32(), 3));
SpvInstruction value = null;
if (context.Config.GpuAccessor.QueryHostHasVectorIndexingBug())
{
// Test for each component individually.
for (int i = 0; i < 4; i++)
{
var component = context.Constant(context.TypeS32(), i);
SpvInstruction elemPointer;
if (context.UniformBuffersArray != null)
{
var ubVariable = context.UniformBuffersArray;
var i0 = context.Get(AggregateType.S32, src1);
elemPointer = context.AccessChain(context.TypePointer(StorageClass.Uniform, context.TypeFP32()), ubVariable, i0, i1, i2, component);
}
else
{
var ubVariable = context.UniformBuffers[((AstOperand)src1).Value];
elemPointer = context.AccessChain(context.TypePointer(StorageClass.Uniform, context.TypeFP32()), ubVariable, i1, i2, component);
}
SpvInstruction newValue = context.Load(context.TypeFP32(), elemPointer);
value = value != null ? context.Select(context.TypeFP32(), context.IEqual(context.TypeBool(), i3, component), newValue, value) : newValue;
}
}
else
{
SpvInstruction elemPointer;
if (context.UniformBuffersArray != null)
{
var ubVariable = context.UniformBuffersArray;
var i0 = context.Get(AggregateType.S32, src1);
elemPointer = context.AccessChain(context.TypePointer(StorageClass.Uniform, context.TypeFP32()), ubVariable, i0, i1, i2, i3);
}
else
{
var ubVariable = context.UniformBuffers[((AstOperand)src1).Value];
elemPointer = context.AccessChain(context.TypePointer(StorageClass.Uniform, context.TypeFP32()), ubVariable, i1, i2, i3);
}
value = context.Load(context.TypeFP32(), elemPointer);
}
return new OperationResult(AggregateType.FP32, value);
}
private static OperationResult GenerateLoadLocal(CodeGenContext context, AstOperation operation)
{
return GenerateLoadLocalOrShared(context, operation, StorageClass.Private, context.LocalMemory);
@ -1990,12 +1928,32 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Spirv
{
StorageKind storageKind = operation.StorageKind;
SpvInstruction pointer;
StorageClass storageClass;
SpvInstruction baseObj;
AggregateType varType;
int srcIndex = 0;
switch (storageKind)
{
case StorageKind.ConstantBuffer:
if (!(operation.GetSource(srcIndex++) is AstOperand bindingIndex) || bindingIndex.Type != OperandType.Constant)
{
throw new InvalidOperationException($"First input of {operation.Inst} with {storageKind} storage must be a constant operand.");
}
if (!(operation.GetSource(srcIndex) is AstOperand fieldIndex) || fieldIndex.Type != OperandType.Constant)
{
throw new InvalidOperationException($"Second input of {operation.Inst} with {storageKind} storage must be a constant operand.");
}
BufferDefinition buffer = context.Config.Properties.ConstantBuffers[bindingIndex.Value];
StructureField field = buffer.Type.Fields[fieldIndex.Value];
storageClass = StorageClass.Uniform;
varType = field.Type & AggregateType.ElementTypeMask;
baseObj = context.ConstantBuffers[bindingIndex.Value];
break;
case StorageKind.Input:
case StorageKind.InputPerPatch:
case StorageKind.Output:
@ -2038,33 +1996,6 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Spirv
{
varType = context.Config.GetFragmentOutputColorType(location);
}
else if (ioVariable == IoVariable.FragmentOutputIsBgra)
{
var pointerType = context.TypePointer(StorageClass.Uniform, context.TypeU32());
var elemIndex = context.Get(AggregateType.S32, operation.GetSource(srcIndex++));
pointer = context.AccessChain(pointerType, context.SupportBuffer, context.Constant(context.TypeU32(), 1), elemIndex);
varType = AggregateType.U32;
break;
}
else if (ioVariable == IoVariable.SupportBlockRenderScale)
{
var pointerType = context.TypePointer(StorageClass.Uniform, context.TypeFP32());
var elemIndex = context.Get(AggregateType.S32, operation.GetSource(srcIndex++));
pointer = context.AccessChain(pointerType, context.SupportBuffer, context.Constant(context.TypeU32(), 4), elemIndex);
varType = AggregateType.FP32;
break;
}
else if (ioVariable == IoVariable.SupportBlockViewInverse)
{
var pointerType = context.TypePointer(StorageClass.Uniform, context.TypeFP32());
var elemIndex = context.Get(AggregateType.S32, operation.GetSource(srcIndex++));
pointer = context.AccessChain(pointerType, context.SupportBuffer, context.Constant(context.TypeU32(), 2), elemIndex);
varType = AggregateType.FP32;
break;
}
else
{
(_, varType) = IoMap.GetSpirvBuiltIn(ioVariable);
@ -2072,55 +2003,57 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Spirv
varType &= AggregateType.ElementTypeMask;
int inputsCount = (isStore ? operation.SourcesCount - 1 : operation.SourcesCount) - srcIndex;
var storageClass = isOutput ? StorageClass.Output : StorageClass.Input;
storageClass = isOutput ? StorageClass.Output : StorageClass.Input;
var ioDefinition = new IoDefinition(storageKind, ioVariable, location, component);
var dict = isPerPatch
? (isOutput ? context.OutputsPerPatch : context.InputsPerPatch)
: (isOutput ? context.Outputs : context.Inputs);
SpvInstruction baseObj = dict[ioDefinition];
SpvInstruction e0, e1, e2;
switch (inputsCount)
{
case 0:
pointer = baseObj;
break;
case 1:
e0 = context.Get(AggregateType.S32, operation.GetSource(srcIndex++));
pointer = context.AccessChain(context.TypePointer(storageClass, context.GetType(varType)), baseObj, e0);
break;
case 2:
e0 = context.Get(AggregateType.S32, operation.GetSource(srcIndex++));
e1 = context.Get(AggregateType.S32, operation.GetSource(srcIndex++));
pointer = context.AccessChain(context.TypePointer(storageClass, context.GetType(varType)), baseObj, e0, e1);
break;
case 3:
e0 = context.Get(AggregateType.S32, operation.GetSource(srcIndex++));
e1 = context.Get(AggregateType.S32, operation.GetSource(srcIndex++));
e2 = context.Get(AggregateType.S32, operation.GetSource(srcIndex++));
pointer = context.AccessChain(context.TypePointer(storageClass, context.GetType(varType)), baseObj, e0, e1, e2);
break;
default:
var indexes = new SpvInstruction[inputsCount];
int index = 0;
for (; index < inputsCount; srcIndex++, index++)
{
indexes[index] = context.Get(AggregateType.S32, operation.GetSource(srcIndex));
}
pointer = context.AccessChain(context.TypePointer(storageClass, context.GetType(varType)), baseObj, indexes);
break;
}
baseObj = dict[ioDefinition];
break;
default:
throw new InvalidOperationException($"Invalid storage kind {storageKind}.");
}
int inputsCount = (isStore ? operation.SourcesCount - 1 : operation.SourcesCount) - srcIndex;
SpvInstruction e0, e1, e2;
SpvInstruction pointer;
switch (inputsCount)
{
case 0:
pointer = baseObj;
break;
case 1:
e0 = context.Get(AggregateType.S32, operation.GetSource(srcIndex++));
pointer = context.AccessChain(context.TypePointer(storageClass, context.GetType(varType)), baseObj, e0);
break;
case 2:
e0 = context.Get(AggregateType.S32, operation.GetSource(srcIndex++));
e1 = context.Get(AggregateType.S32, operation.GetSource(srcIndex++));
pointer = context.AccessChain(context.TypePointer(storageClass, context.GetType(varType)), baseObj, e0, e1);
break;
case 3:
e0 = context.Get(AggregateType.S32, operation.GetSource(srcIndex++));
e1 = context.Get(AggregateType.S32, operation.GetSource(srcIndex++));
e2 = context.Get(AggregateType.S32, operation.GetSource(srcIndex++));
pointer = context.AccessChain(context.TypePointer(storageClass, context.GetType(varType)), baseObj, e0, e1, e2);
break;
default:
var indexes = new SpvInstruction[inputsCount];
int index = 0;
for (; index < inputsCount; srcIndex++, index++)
{
indexes[index] = context.Get(AggregateType.S32, operation.GetSource(srcIndex));
}
pointer = context.AccessChain(context.TypePointer(storageClass, context.GetType(varType)), baseObj, indexes);
break;
}
if (isStore)
{
context.Store(pointer, context.Get(varType, operation.GetSource(srcIndex)));

View file

@ -63,7 +63,7 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Spirv
if (context.Config.Stage == ShaderStage.Vertex)
{
var scaleCountPointerType = context.TypePointer(StorageClass.Uniform, context.TypeS32());
var scaleCountElemPointer = context.AccessChain(scaleCountPointerType, context.SupportBuffer, context.Constant(context.TypeU32(), 3));
var scaleCountElemPointer = context.AccessChain(scaleCountPointerType, context.ConstantBuffers[0], context.Constant(context.TypeU32(), 3));
var scaleCount = context.Load(context.TypeS32(), scaleCountElemPointer);
scaleIndex = context.IAdd(context.TypeU32(), scaleIndex, scaleCount);
@ -71,7 +71,7 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Spirv
scaleIndex = context.IAdd(context.TypeU32(), scaleIndex, context.Constant(context.TypeU32(), 1));
var scaleElemPointer = context.AccessChain(pointerType, context.SupportBuffer, fieldIndex, scaleIndex);
var scaleElemPointer = context.AccessChain(pointerType, context.ConstantBuffers[0], fieldIndex, scaleIndex);
var scale = context.Load(context.TypeFP32(), scaleElemPointer);
var ivector2Type = context.TypeVector(context.TypeS32(), 2);
@ -201,7 +201,7 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Spirv
if (context.Config.Stage == ShaderStage.Vertex)
{
var scaleCountPointerType = context.TypePointer(StorageClass.Uniform, context.TypeS32());
var scaleCountElemPointer = context.AccessChain(scaleCountPointerType, context.SupportBuffer, context.Constant(context.TypeU32(), 3));
var scaleCountElemPointer = context.AccessChain(scaleCountPointerType, context.ConstantBuffers[0], context.Constant(context.TypeU32(), 3));
var scaleCount = context.Load(context.TypeS32(), scaleCountElemPointer);
scaleIndex = context.IAdd(context.TypeU32(), scaleIndex, scaleCount);
@ -209,7 +209,7 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Spirv
scaleIndex = context.IAdd(context.TypeU32(), scaleIndex, context.Constant(context.TypeU32(), 1));
var scaleElemPointer = context.AccessChain(pointerType, context.SupportBuffer, fieldIndex, scaleIndex);
var scaleElemPointer = context.AccessChain(pointerType, context.ConstantBuffers[0], fieldIndex, scaleIndex);
var scale = context.GlslFAbs(context.TypeFP32(), context.Load(context.TypeFP32(), scaleElemPointer));
var passthrough = context.FOrdEqual(context.TypeBool(), scale, context.Constant(context.TypeFP32(), 1f));