raphiaura/KytyPS5
1
0
Fork 0
mirror of https://github.com/KytyPS5/KytyPS5.git synced 2026-10-20 02:47:02 +02:00
Code Issues Packages Projects Releases Wiki Activity

shader: fold reads through empty scalar descriptor tables

Browse Source 
Co-authored-by: fxpw <fxpw@outlook.com>
This commit is contained in:
nmzik
2026-10-14 15:17:03 +02:00
parent 4b549b9cb3
commit f3c017b168
4 changed files with 170 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/graphics/shader/recompiler/ir/passes/ResourceMaterialization.cpp
+28 -4
View File
@@ -60,6 +60,15 @@ bool NullImageDescriptor(const DescriptorValue& descriptor) {
return descriptor.dwords[0] == 0 && (descriptor.dwords[1] & 0xffu) == 0;
}
bool NullBufferDescriptor(Value value) {
const auto* handle = value.Resolve().TryInstruction();
if (handle == nullptr || handle->GetOpcode() != ValueOpcode::GetBufferResource ||
handle->NumArgs() != 4u) return false;
for (size_t index = 0; index < 4u; ++index)
if (handle->Arg(index).Resolve() != Value(0u)) return false;
return true;
}
bool ValidImageDescriptor(const DescriptorValue& descriptor, bool r128 = false) {
const auto& words = descriptor.dwords;
// Reject texture descriptors with nonzero reserved bits.
@@ -1378,6 +1387,7 @@ bool MaterializeResources(const ResourcePlan& program, const SrtRuntime& runtime
base.formatted ? descriptor.Format() : Prospero::BufferFormat::kInvalid;
buffer.descriptor_swizzle = base.formatted ? descriptor.DstSelXYZW() : DstSel(4, 5, 6, 7);
buffer.zero_stride_oob = descriptor.OutOfBounds() == 0u && stride == 0u;
buffer.scalar_empty = base.scalar && descriptor.GetSize() == 0u;
}
snapshot.images.resize(program.info.images.size());
specialization.images.resize(program.info.images.size());
@@ -1543,6 +1553,17 @@ void ApplyResourceSpecialization(Program& program, const ResourceSpecialization&
}
auto& memory_info = program.memory_info;
// Resolve empty scalar-table words before folding the indirect descriptors that use them.
for (auto* block: program.blocks) {
for (auto& inst: *block) {
if (inst.GetOpcode() != ValueOpcode::ReadConstBuffer) continue;
const auto& memory = memory_info[inst.Flags<MemoryFlags>().index];
if (memory.kind == ResourceKind::ScalarBuffer &&
specialization.buffers[memory.resource].scalar_empty &&
buffers[memory.resource].indirect_root == BufferResource::NoIndirectBuffer)
inst.ReplaceUsesWith(Value(0u));
}
}
const ImageRemap image_remap(specialization);
for (auto* block: program.blocks) {
for (auto it = block->begin(); it != block->end(); ++it) {
@@ -1563,14 +1584,17 @@ 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 &&
const bool native_oob = memory.kind == ResourceKind::Buffer &&
specialization.buffers[memory.resource].zero_stride_oob &&
buffers[memory.resource].indirect_root == BufferResource::NoIndirectBuffer) {
buffers[memory.resource].indirect_root == BufferResource::NoIndirectBuffer;
const bool null_indirect = memory.kind == ResourceKind::IndirectBuffer &&
NullBufferDescriptor(inst.Arg(0));
if (native_oob || null_indirect) {
// Bounds mode 0 checks offset >= stride, so zero stride
// makes every vector read out of bounds regardless of its address.
// makes every vector read out of bounds. An all-zero V# is unbound.
const auto count = BufferComponentCount(inst.GetOpcode());
std::array<Value, 4> values {Value(0u), Value(0u), Value(0u), Value(0u)};
if (memory.formatted && !memory.typed) {
if (native_oob && memory.formatted && !memory.typed) {
const auto& buffer = buffers[memory.resource];
const auto format = Format::GetFormatInfo(buffer.descriptor_format);
const bool d16 = memory.data_bits == 16u;
src/graphics/shader/recompiler/ir/passes/ResourceMaterialization.h
+1
View File
@@ -13,6 +13,7 @@ struct ResourceSpecialization {
Prospero::BufferFormat descriptor_format = Prospero::BufferFormat::kInvalid;
uint32_t descriptor_swizzle = DstSel(4, 5, 6, 7);
bool zero_stride_oob = false;
bool scalar_empty = false;
uint32_t indirect_root = BufferResource::NoIndirectBuffer;
uint32_t indirect_mapping_offset = 0;
uint32_t indirect_search_iterations = 0;