mirror of
https://github.com/KytyPS5/KytyPS5.git
synced 2026-10-20 02:47:02 +02:00
shader: share proven selector domains for paired images and samplers
This commit is contained in:
@@ -164,6 +164,16 @@ bool EquivalentValue(const ResourcePlan& program, Value left, Value right) {
|
||||
return EquivalentValue(program, left, right, visited);
|
||||
}
|
||||
|
||||
bool EquivalentSelectorDomain(const ResourcePlan& program,
|
||||
const DescriptorSource::IndirectDescriptor& left,
|
||||
const DescriptorSource::IndirectDescriptor& right) {
|
||||
return left.selector == right.selector && left.signed_count == right.signed_count &&
|
||||
left.workgroup_axis == right.workgroup_axis &&
|
||||
EquivalentValue(program, left.key_count, right.key_count) &&
|
||||
EquivalentValue(program, left.selector_first, right.selector_first) &&
|
||||
EquivalentValue(program, left.selector_mask, right.selector_mask);
|
||||
}
|
||||
|
||||
Value ResolveInvariantPhi(const ResourcePlan& program, Value value) {
|
||||
value = value.Resolve();
|
||||
const auto* root = value.TryInstruction();
|
||||
|
||||
@@ -707,6 +707,9 @@ bool HasShaderMemoryWrites(const Program& program);
|
||||
void ValidateProgram(const Program& program, bool require_ssa);
|
||||
void ResolveControlFlowIdentities(Program& program);
|
||||
bool EquivalentValue(const ResourcePlan& program, Value left, Value right);
|
||||
bool EquivalentSelectorDomain(const ResourcePlan& program,
|
||||
const DescriptorSource::IndirectDescriptor& left,
|
||||
const DescriptorSource::IndirectDescriptor& right);
|
||||
Value ResolveInvariantPhi(const ResourcePlan& program, Value value);
|
||||
Value ResolveActiveU32(Value value, Value active);
|
||||
|
||||
|
||||
@@ -1300,11 +1300,7 @@ ResourcePlan ExtractResourcePlan(const Program& program) {
|
||||
uint32_t group = 0;
|
||||
for (; group < plan.selector_keys.size(); ++group) {
|
||||
const auto& b = *program.descriptor_sources[plan.selector_keys[group].source].indirect_descriptor;
|
||||
if (a.selector == b.selector && a.signed_count == b.signed_count &&
|
||||
a.workgroup_axis == b.workgroup_axis &&
|
||||
EquivalentValue(program, a.key_count, b.key_count) &&
|
||||
EquivalentValue(program, a.selector_first, b.selector_first) &&
|
||||
EquivalentValue(program, a.selector_mask, b.selector_mask)) break;
|
||||
if (EquivalentSelectorDomain(program, a, b)) break;
|
||||
}
|
||||
if (group == plan.selector_keys.size()) plan.selector_keys.push_back({.source = source});
|
||||
plan.selector_groups[source] = group;
|
||||
|
||||
@@ -2851,7 +2851,8 @@ private:
|
||||
}
|
||||
|
||||
bool TryMakeTableDescriptor(Inst& handle, const DescriptorSource& descriptor,
|
||||
IndirectDescriptorPlan& plan) {
|
||||
IndirectDescriptorPlan& plan,
|
||||
const IndirectDescriptorPlan* paired_sampler) {
|
||||
const bool sampler = handle.GetOpcode() == ValueOpcode::GetSamplerResource;
|
||||
if ((!sampler && handle.GetOpcode() != ValueOpcode::GetImageResource) ||
|
||||
handle.NumArgs() != (sampler ? 4u : 8u)) return false;
|
||||
@@ -2934,15 +2935,39 @@ private:
|
||||
// Keep the GPU selector and descriptor reads, and preserve the proven scalar
|
||||
// offset domain within the bounded image table.
|
||||
if (sampler || plan.split_offsets) return false;
|
||||
indirect.workgroup_axis = UINT32_MAX;
|
||||
plan.table_read = plan.reads[0];
|
||||
if (plan.table_indexed) {
|
||||
indirect.table_stride = 0u;
|
||||
indirect.table_offset = 0u;
|
||||
indirect.table_immediate = 0u;
|
||||
indirect.table_record_bytes = table_immediate + plan.read_dwords * 4u;
|
||||
} else {
|
||||
indirect.table_scalar = true;
|
||||
bool shared_selector = false;
|
||||
if (!plan.table_indexed && paired_sampler != nullptr &&
|
||||
m_sources[paired_sampler->source].indirect_descriptor->selector &&
|
||||
EquivalentValue(m_program, key, paired_sampler->key) &&
|
||||
std::ranges::all_of(handle.Uses(), [&](const Use& use) {
|
||||
return use.user->GetOpcode() == ValueOpcode::ImageSampleRaw &&
|
||||
use.user->Arg(1).Resolve().TryInstruction() == paired_sampler->handle &&
|
||||
(m_program.memory_info[use.user->Flags<MemoryFlags>().index].image_sample_flags &
|
||||
Decoder::ImageSampleFlagAdjust) == 0u;
|
||||
})) {
|
||||
// This sampler already requires these exact keys. Prove the pairing before
|
||||
// interning, so image-only handles keep their independent heap domain.
|
||||
auto selected = indirect;
|
||||
DescriptorSource selected_source;
|
||||
if (MatchScalarMaterialKey(key, selected, selected_source, false, true) &&
|
||||
EquivalentSelectorDomain(m_program, selected,
|
||||
*m_sources[paired_sampler->source].indirect_descriptor)) {
|
||||
indirect = std::move(selected);
|
||||
material_source = std::move(selected_source);
|
||||
shared_selector = true;
|
||||
}
|
||||
}
|
||||
if (!shared_selector) {
|
||||
indirect.workgroup_axis = UINT32_MAX;
|
||||
plan.table_read = plan.reads[0];
|
||||
if (plan.table_indexed) {
|
||||
indirect.table_stride = 0u;
|
||||
indirect.table_offset = 0u;
|
||||
indirect.table_immediate = 0u;
|
||||
indirect.table_record_bytes = table_immediate + plan.read_dwords * 4u;
|
||||
} else {
|
||||
indirect.table_scalar = true;
|
||||
}
|
||||
}
|
||||
plan.retain_reads = true;
|
||||
}
|
||||
@@ -3415,6 +3440,7 @@ private:
|
||||
}
|
||||
const auto flags = inst.Flags<MemoryFlags>();
|
||||
const auto& memory = m_program.memory_info[flags.index];
|
||||
const IndirectDescriptorPlan* paired_sampler = nullptr;
|
||||
const auto plan_handle = [&](Value value, uint32_t width, uint32_t base, bool sampler) {
|
||||
auto* handle = value.Resolve().TryInstruction();
|
||||
if (handle == nullptr || FindIndirectDescriptor(*handle) != nullptr) return;
|
||||
@@ -3437,7 +3463,7 @@ private:
|
||||
}
|
||||
if (TryMakeNestedDescriptor(*handle, descriptor, plan) ||
|
||||
(sampler && TryMakeFiniteDescriptor(*handle, plan)) ||
|
||||
TryMakeTableDescriptor(*handle, descriptor, plan) ||
|
||||
TryMakeTableDescriptor(*handle, descriptor, plan, paired_sampler) ||
|
||||
(!sampler && (TryMakeFiniteDescriptor(*handle, plan) ||
|
||||
TryMakeIndirectBuffer(*handle, descriptor, plan)))) {
|
||||
for (const auto& previous: m_indirect_descriptors) {
|
||||
@@ -3448,11 +3474,14 @@ private:
|
||||
m_indirect_descriptors.push_back(std::move(plan));
|
||||
}
|
||||
};
|
||||
if (inst.GetOpcode() == ValueOpcode::ImageSampleRaw &&
|
||||
(memory.image_sample_flags & Decoder::ImageSampleFlagAdjust) == 0u) {
|
||||
plan_handle(inst.Arg(1), 4u, memory.sampler * 4u, true);
|
||||
if (const auto* handle = inst.Arg(1).Resolve().TryInstruction())
|
||||
paired_sampler = FindIndirectDescriptor(*handle);
|
||||
}
|
||||
plan_handle(inst.Arg(0), IndirectBufferAccess(inst) ? 4u : 8u,
|
||||
memory.resource * 4u, false);
|
||||
if (inst.GetOpcode() == ValueOpcode::ImageSampleRaw &&
|
||||
(memory.image_sample_flags & Decoder::ImageSampleFlagAdjust) == 0u)
|
||||
plan_handle(inst.Arg(1), 4u, memory.sampler * 4u, true);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
Reference in New Issue
Block a user