raphiaura/KytyPS5
1
0
Fork 0
mirror of https://github.com/KytyPS5/KytyPS5.git synced 2026-10-25 20:36:54 +01:00
Code Issues Packages Projects Releases Wiki Activity

shader: share nested pointer-table image and sampler selection

Browse Source 
Co-authored-by: fxpw <fxpw@outlook.com>
This commit is contained in:
nmzik
2026-10-14 16:40:15 +02:00
parent 3a925efddf
commit 487bbf95fd
7 changed files with 484 additions and 27 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/graphics/shader/recompiler/ir/passes/ResourceTracking.cpp
+90 -7
View File
@@ -45,10 +45,13 @@ uint32_t PossibleU32Bits(Value value, const auto& normalize) {
if (inst->NumArgs() != 0u) bits = merged;
break;
}
case ValueOpcode::ShiftLeftLogical32: {
case ValueOpcode::ShiftLeftLogical32:
case ValueOpcode::ShiftRightLogical32: {
const auto shift = inst->Arg(1).Resolve();
if (shift.IsImmediate() && shift.GetType() == Type::U32)
bits = arg(0) << (shift.U32() & 31u);
if (shift.IsImmediate() && shift.GetType() == Type::U32) {
bits = inst->GetOpcode() == ValueOpcode::ShiftLeftLogical32
? arg(0) << (shift.U32() & 31u) : arg(0) >> (shift.U32() & 31u);
}
break;
}
default: break;
@@ -397,6 +400,13 @@ public:
std::vector<std::pair<const IndirectDescriptorPlan*, Value>> buffer_selections;
for (const auto& plan: m_indirect_descriptors) {
auto key = plan.key;
if (m_sources[plan.source].indirect_descriptor->pointer_table) {
auto* block = plan.handle->Parent();
const auto where = std::ranges::find_if(block->Instructions(),
[&](const Inst& inst) { return &inst == plan.handle; });
key = Value(&*block->PrependNewInst(
where, ValueOpcode::ShiftLeftLogical32, {key, Value(2u)}));
}
if (plan.table_read != nullptr) {
const auto* anchor = plan.table_indexed ? plan.handle : plan.table_read;
auto* block = anchor->Parent();
@@ -964,7 +974,8 @@ private:
if (current.indirect_descriptor.has_value()) {
const auto& a = *current.indirect_descriptor;
const auto& b = *descriptor.indirect_descriptor;
if (a.selector != b.selector || a.table_source != b.table_source ||
if (a.selector != b.selector || a.pointer_table != b.pointer_table ||
a.table_source != b.table_source ||
a.table_offset != b.table_offset || a.table_immediate != b.table_immediate ||
a.table_stride != b.table_stride || a.indexed != b.indexed ||
a.workgroup_axis != b.workgroup_axis || a.sources != b.sources ||
@@ -2713,6 +2724,74 @@ private:
return true;
}
bool TryMakeNestedDescriptor(Inst& handle, const DescriptorSource& descriptor,
IndirectDescriptorPlan& plan) {
if (handle.GetOpcode() != ValueOpcode::GetImageResource &&
handle.GetOpcode() != ValueOpcode::GetSamplerResource) return false;
const Inst* pointer = nullptr;
uint32_t descriptor_offset = 0;
for (uint32_t word = 0; word < handle.NumArgs(); ++word) {
const auto* read = descriptor.dwords[word].Resolve().TryInstruction();
uint32_t index, offset;
const auto* memory = read != nullptr ? ScalarReadMemory(*read, index) : nullptr;
if (memory == nullptr || memory->kind != ResourceKind::ScalarAddress ||
memory->offset > INT32_MAX || !ImmediateU32(read->Arg(1), offset) ||
!AvailableBefore(*read, handle)) return false;
const auto* address = read->Arg(0).Resolve().TryInstruction();
const uint64_t current_offset = uint64_t {offset & ~3u} + (memory->offset & ~3u);
if (address == nullptr || address->GetOpcode() != ValueOpcode::GetAddressResource ||
address->NumArgs() != 2u || current_offset > UINT32_MAX ||
(pointer != nullptr && (!EquivalentValue(m_program, Value(pointer), Value(address)) ||
current_offset != uint64_t {descriptor_offset} + word * 4u)))
return false;
if (word == 0u) descriptor_offset = static_cast<uint32_t>(current_offset);
pointer = address;
}
// An out-of-bounds pointer pair is zero. Guest mappings exclude the first
// 16 KiB page, so these pointed reads have the existing null-image semantics.
if (pointer == nullptr || uint64_t {descriptor_offset} + handle.NumArgs() * 4u > 0x4000u)
return false;
const auto* low = pointer->Arg(0).Resolve().TryInstruction();
const auto* high = pointer->Arg(1).Resolve().TryInstruction();
uint32_t low_index, high_index;
const auto* low_memory = low != nullptr ? ScalarReadMemory(*low, low_index) : nullptr;
const auto* high_memory = high != nullptr ? ScalarReadMemory(*high, high_index) : nullptr;
if (low_memory == nullptr || high_memory == nullptr ||
low_memory->kind != ResourceKind::ScalarBuffer ||
high_memory->kind != ResourceKind::ScalarBuffer ||
low_memory->offset > INT32_MAX || high_memory->offset > INT32_MAX ||
(low_memory->offset & ~3u) + 4u != (high_memory->offset & ~3u) ||
!EquivalentValue(m_program, low->Arg(0), high->Arg(0)) ||
!EquivalentValue(m_program, low->Arg(1), high->Arg(1)) ||
!AvailableBefore(*low, *pointer) || !AvailableBefore(*high, *pointer)) return false;
Value key;
DescriptorSource::IndirectDescriptor indirect;
if (!MatchTableOffset(low->Arg(1), key, indirect.table_offset, indirect.table_stride) ||
indirect.table_stride < 8u || (indirect.table_stride & 3u) != 0u) return false;
const auto maximum = SelectorMaximum(key);
if (maximum > UINT32_MAX / 4u ||
uint64_t {maximum} * indirect.table_stride + indirect.table_offset > UINT32_MAX ||
(key.TryInstruction() != nullptr && !AvailableBefore(*key.Instruction(), handle)))
return false;
DescriptorSource table_source;
if (!MakeRuntimeTableSource(*low, table_source) || table_source.dword_count != 4u)
return false;
indirect.table_source = InternSource(table_source);
indirect.table_immediate = low_memory->offset & ~3u;
indirect.table_scalar = true;
indirect.pointer_table = {{maximum, descriptor_offset}};
DescriptorSource source;
source.dword_count = handle.NumArgs();
source.dwords.fill(Value(0u));
std::copy_n(table_source.dwords.begin(), 4u, source.dwords.begin());
source.indirect_descriptor = indirect;
plan.handle = &handle;
plan.source = InternSource(source);
plan.key = key;
plan.roots = source.dwords;
return true;
}
bool TryMakeIndirectImage(Inst& handle, const DescriptorSource& descriptor,
IndirectDescriptorPlan& plan) {
if (handle.GetOpcode() != ValueOpcode::GetImageResource || handle.NumArgs() != 8u) return false;
@@ -3019,6 +3098,9 @@ private:
return SelectorMaximum(inst->Arg(arg), bounds, depth + 1u, guard);
};
switch (inst->GetOpcode()) {
case ValueOpcode::ReadLane:
// ReadLane ignores EXEC; only context-free bounds on every source lane apply.
return PossibleU32Bits(inst->Arg(0));
case ValueOpcode::Phi: {
const auto initial = InitialCandidateMask(value);
if (!initial.IsEmpty()) return SelectorMaximum(initial, bounds, depth + 1u, guard);
@@ -3240,10 +3322,11 @@ private:
uint32_t bad_dword = 0;
if (ValidateSource(descriptor, bad_dword)) return;
IndirectDescriptorPlan plan;
if (sampler ? TryMakeFiniteDescriptor(*handle, plan)
if (TryMakeNestedDescriptor(*handle, descriptor, plan) ||
(sampler ? TryMakeFiniteDescriptor(*handle, plan)
: (TryMakeIndirectImage(*handle, descriptor, plan) ||
TryMakeFiniteDescriptor(*handle, plan) ||
TryMakeIndirectBuffer(*handle, descriptor, plan))) {
TryMakeIndirectBuffer(*handle, descriptor, plan)))) {
for (const auto& previous: m_indirect_descriptors) {
if (plan.reads[0] != nullptr && previous.reads[0] == plan.reads[0] &&
!EquivalentValue(m_program, previous.key, plan.key))
@@ -3389,7 +3472,7 @@ private:
const auto* descriptor = Source(source);
if (descriptor != nullptr && descriptor->indirect_descriptor) {
const auto& sources = descriptor->indirect_descriptor->sources;
if (sources.empty()) Fail(pc, "finite sampler has no candidates");
if (sources.empty()) return root;
std::vector<uint32_t> resources {root};
for (size_t i = 1; i < sources.size(); ++i) {
const auto candidate = AddSampler(sources[i], pc);