raphiaura/KytyPS5
1
0
Fork 0
mirror of https://github.com/KytyPS5/KytyPS5.git synced 2026-10-21 11:26:56 +02:00
Code Issues Packages Projects Releases Wiki Activity

shader: share bounded record descriptors for packed loads and byte stores

Browse Source 
Co-authored-by: fxpw <fxpw@outlook.com>
This commit is contained in:
nmzik
2026-10-14 05:07:24 +02:00
parent 4cf5a96b79
commit 26d326d1b0
12 changed files with 581 additions and 142 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/graphics/shader/recompiler/ir/passes/ResourceMaterialization.cpp
+34 -10
View File
@@ -227,16 +227,38 @@ bool ReadScalarTable(uint64_t base, uint64_t size, uint64_t offset,
return ReadTable(base & AddressMask & ~uint64_t {3}, size, offset & ~uint64_t {3}, runtime, words);
}
bool NormalizeIndirectStoreBuffer(DescriptorValue& value, bool indexed, bool formatted) {
bool NormalizeIndirectBuffer(DescriptorValue& value, bool indexed, const BufferResource& resource) {
ShaderBufferResource descriptor;
if (!DecodeBufferDescriptor(value, descriptor)) return false;
const bool formatted_read = resource.formatted && resource.read && !resource.written;
if (descriptor.Type() != 0u || descriptor.Format() == Prospero::BufferFormat::kInvalid ||
descriptor.Base48() == 0u || descriptor.GetSize() == 0u) {
(!formatted_read && (descriptor.Base48() == 0u || descriptor.GetSize() == 0u))) {
value.dwords.fill(0u);
} else if (descriptor.Stride() != (indexed ? 4u : 1u) || descriptor.SwizzleEnabled() ||
descriptor.AddTid() || descriptor.OutOfBounds() != (indexed ? 0u : 3u)) {
return true;
}
if (descriptor.SwizzleEnabled() || descriptor.AddTid() ||
descriptor.OutOfBounds() != (indexed ? 0u : 3u)) {
return SpecializationFail("indirect buffers require exact linear index or byte bounds");
}
if (formatted_read) {
const auto format = Format::GetFormatInfo(descriptor.Format());
if (!indexed || format.type == Format::ComponentType::Unknown)
return SpecializationFail("indirect formatted loads require a supported indexed format");
if (descriptor.Base48() == 0u || descriptor.Stride() == 0u || descriptor.NumRecords() == 0u) {
// Every index is out of bounds. Preserve the format and swizzle constants.
value.dwords[1] &= ~(0x3fffu << 16u);
value.dwords[2] = 0u;
return true;
}
const auto alignment = std::min(4u, format.byte_size);
if (descriptor.Stride() < format.byte_size ||
(descriptor.Stride() & (alignment - 1u)) != 0u ||
(descriptor.Base48() & (alignment - 1u)) != 0u)
return SpecializationFail("indirect formatted loads require aligned full-format records");
} else if ((!indexed && descriptor.Stride() != 1u) ||
(indexed && resource.max_byte_extent != 1u && descriptor.Stride() != 4u)) {
return SpecializationFail("indirect buffer stores require exact linear DWORD or byte bounds");
} else if (formatted && (descriptor.Format() != Prospero::BufferFormat::k32UInt ||
} else if (resource.formatted && (descriptor.Format() != Prospero::BufferFormat::k32UInt ||
descriptor.DstSelXYZW() != DstSel(4, 0, 0, 0) ||
(descriptor.Base48() & 3u) != 0u)) {
return SpecializationFail("indirect formatted X stores require an aligned 32-bit UInt identity mapping");
@@ -248,7 +270,8 @@ bool IsBoundedDescriptorTable(const ResourcePlan& program,
const DescriptorSource::IndirectDescriptor& indirect) {
const auto* table = Source(program, indirect.table_source);
return indirect.sources.empty() && !indirect.selector &&
indirect.workgroup_axis == UINT32_MAX && table != nullptr && table->dword_count == 4u;
indirect.key_count.IsEmpty() && indirect.workgroup_axis == UINT32_MAX &&
table != nullptr && table->dword_count == 4u;
}
template <typename Specialization, typename Normalize>
@@ -372,7 +395,7 @@ bool MaterializeIndirectDescriptor(const ResourcePlan& p
key_count = runtime.workgroup_counts[indirect.workgroup_axis];
if (key_count == 0u) return false;
} else {
if (table_value.dword_count != 2u || !evaluate_count(key_count))
if (!evaluate_count(key_count))
return false;
}
if (!indirect.selector_first.IsEmpty() && !clean.Evaluate(indirect.selector_first, first_key))
@@ -1308,8 +1331,8 @@ bool MaterializeResources(const ResourcePlan& program, const SrtRuntime& runtime
program, *source->indirect_descriptor, i, 4u, observed, clean, snapshot,
snapshot.buffers, specialization.buffers, ShaderInfo::MaxBuffers,
[&](DescriptorValue& value) {
return NormalizeIndirectStoreBuffer(
value, source->indirect_descriptor->indexed_store, base.formatted);
return NormalizeIndirectBuffer(
value, source->indirect_descriptor->indexed, base);
}))
return false;
}
@@ -1526,7 +1549,8 @@ void ApplyResourceSpecialization(Program& program, const ResourceSpecialization&
if (BufferAccessOf(inst.GetOpcode()) == BufferAccess::Read) {
const auto& memory = memory_info[inst.Flags<MemoryFlags>().index];
if (memory.kind == ResourceKind::Buffer &&
specialization.buffers[memory.resource].zero_stride_oob) {
specialization.buffers[memory.resource].zero_stride_oob &&
buffers[memory.resource].indirect_root == BufferResource::NoIndirectBuffer) {
// Bounds mode 0 checks offset >= stride, so zero stride
// makes every vector read out of bounds regardless of its address.
const auto count = BufferComponentCount(inst.GetOpcode());