1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
|
using Ryujinx.Graphics.Shader.IntermediateRepresentation;
using System.Collections.Generic;
using static Ryujinx.Graphics.Shader.IntermediateRepresentation.OperandHelper;
namespace Ryujinx.Graphics.Shader.Translation.Optimizations
{
static class BindlessToIndexed
{
private const int NvnTextureBufferIndex = 2;
public static void RunPass(BasicBlock block, ResourceManager resourceManager)
{
// We can turn a bindless texture access into a indexed access,
// as long the following conditions are true:
// - The handle is loaded using a LDC instruction.
// - The handle is loaded from the constant buffer with the handles (CB2 for NVN).
// - The load has a constant offset.
// The base offset of the array of handles on the constant buffer is the constant offset.
for (LinkedListNode<INode> node = block.Operations.First; node != null; node = node.Next)
{
if (node.Value is not TextureOperation texOp)
{
continue;
}
if ((texOp.Flags & TextureFlags.Bindless) == 0)
{
continue;
}
if (texOp.GetSource(0).AsgOp is not Operation handleAsgOp)
{
continue;
}
if (handleAsgOp.Inst != Instruction.Load ||
handleAsgOp.StorageKind != StorageKind.ConstantBuffer ||
handleAsgOp.SourcesCount != 4)
{
continue;
}
Operand ldcSrc0 = handleAsgOp.GetSource(0);
if (ldcSrc0.Type != OperandType.Constant ||
!resourceManager.TryGetConstantBufferSlot(ldcSrc0.Value, out int src0CbufSlot) ||
src0CbufSlot != NvnTextureBufferIndex)
{
continue;
}
Operand ldcSrc1 = handleAsgOp.GetSource(1);
// We expect field index 0 to be accessed.
if (ldcSrc1.Type != OperandType.Constant || ldcSrc1.Value != 0)
{
continue;
}
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 not Operation shrOp || shrOp.Inst != Instruction.ShiftRightU32)
{
continue;
}
if (shrOp.GetSource(0).AsgOp is not Operation shrOp2 || shrOp2.Inst != Instruction.ShiftRightU32)
{
continue;
}
if (shrOp2.GetSource(0).AsgOp is not Operation addOp || addOp.Inst != Instruction.Add)
{
continue;
}
Operand addSrc1 = addOp.GetSource(1);
if (addSrc1.Type != OperandType.Constant)
{
continue;
}
TurnIntoIndexed(resourceManager, texOp, addSrc1.Value / 4);
Operand index = Local();
Operand source = addOp.GetSource(0);
Operation shrBy3 = new(Instruction.ShiftRightU32, index, source, Const(3));
block.Operations.AddBefore(node, shrBy3);
texOp.SetSource(0, index);
}
}
private static void TurnIntoIndexed(ResourceManager resourceManager, TextureOperation texOp, int handle)
{
int binding = resourceManager.GetTextureOrImageBinding(
texOp.Inst,
texOp.Type | SamplerType.Indexed,
texOp.Format,
texOp.Flags & ~TextureFlags.Bindless,
NvnTextureBufferIndex,
handle);
texOp.TurnIntoIndexed(binding);
}
}
}
|
