raphiaura/KytyPS5
1
0
Fork 0
mirror of https://github.com/KytyPS5/KytyPS5.git synced 2026-10-20 10:56:58 +02:00
Code Issues Packages Projects Releases Wiki Activity

shader: bound GPU-selected record descriptors by scalar table geometry

Browse Source
This commit is contained in:
nmzik
2026-10-14 14:45:56 +02:00
parent 150bcb06a9
commit 9cc0bba80b
4 changed files with 320 additions and 48 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/graphics/shader/recompiler/ir/passes/ResourceMaterialization.cpp
+22 -7
View File
@@ -357,17 +357,32 @@ bool MaterializeIndirectDescriptor(const ResourcePlan& p
table.OutOfBounds() != 0u || (table_base & 3u) != 0u ||
(table.Stride() & 3u) != 0u || table.Stride() < indirect.table_record_bytes)))
return false;
// Descriptor bytes are refreshed together. The GPU selects the DWORD address;
// no material selector or unused table field is interpreted on the CPU.
keys.resize(table_size / 4u);
if (!ReadScalarTable(table_base, table_size, 0u, runtime, keys)) return false;
const bool affine = indirect.table_stride != 0u;
const uint64_t step = affine ? uint64_t {1} << std::countr_zero(indirect.table_stride) : 4u;
const uint64_t first = ((indirect.table_offset % step) & ~uint64_t {3}) +
(indirect.table_immediate & ~3u);
const auto count = table_size >= first + 4u
? std::min((table_size - first - 4u) / step + 1u, (uint64_t {1} << 32u) / step)
: 0u;
if (affine && count > MaxIndirectDescriptorProbes) return false;
// Affine SMEM offsets wrap in U32 before adding the aligned immediate.
// Read only their congruence class; unrestricted image tables retain a bulk read.
if (!affine) {
keys.resize(table_size / 4u);
if (!ReadScalarTable(table_base, table_size, 0u, runtime, keys)) return false;
}
const auto mapping_offset = snapshot.flattened_srt.size();
descriptors[resource_index] = {.dword_count = dword_count};
snapshot.flattened_srt.resize(mapping_offset + 2u, 0u);
for (uint32_t word = 0; word < keys.size(); ++word) {
for (uint32_t word = 0; word < count; ++word) {
DescriptorValue candidate {.dword_count = dword_count};
std::copy_n(keys.begin() + word, std::min<size_t>(dword_count, keys.size() - word),
candidate.dwords.begin());
if (affine) {
if (!ReadScalarTable(table_base, table_size, first + word * step, runtime,
std::span(candidate.dwords).first(dword_count))) return false;
} else {
std::copy_n(keys.begin() + word, std::min<size_t>(dword_count, keys.size() - word),
candidate.dwords.begin());
}
if (!normalize(candidate)) return false;
uint32_t ordinal;
if (!intern_candidate(candidate, ordinal)) return false;
src/graphics/shader/recompiler/ir/passes/ResourceTracking.cpp
+18 -2
View File
@@ -470,7 +470,16 @@ public:
auto* block = anchor->Parent();
const auto where = std::ranges::find_if(block->Instructions(),
[&](const Inst& inst) { return &inst == anchor; });
auto& selection = *block->PrependNewInst(where, ValueOpcode::GetBufferResourceIndex, {plan.key});
const auto& indirect = *m_sources[plan.source].indirect_descriptor;
if (indirect.table_scalar) {
// Wrapped affine offsets span one power-of-two congruence class.
// Keep the existing mapping's byte key while compressing its empty gaps.
key = plan.reads[0]->Arg(1);
const auto shift = std::countr_zero(indirect.table_stride) - 2u;
if (shift != 0u) key = Value(&*block->PrependNewInst(
where, ValueOpcode::ShiftRightLogical32, {key, Value(shift)}));
}
auto& selection = *block->PrependNewInst(where, ValueOpcode::GetBufferResourceIndex, {key});
selection.SetFlags(plan.handle->Flags<uint32_t>());
selected = Value(&selection);
buffer_selections.emplace_back(&plan, selected);
@@ -2875,7 +2884,14 @@ private:
indirect.signed_count = bound->GetOpcode() == ValueOpcode::SLessThan32;
return finish();
}
if (!legacy_store || table_stride != 16u) return false;
if (!legacy_store) {
// A bounded V# supplies every reachable descriptor without reading the GPU key.
// Separate U32 offsets could wrap each DWORD to a different table position.
if (table_source.dword_count != 4u || plan.split_offsets) return false;
indirect.table_scalar = true;
return finish();
}
if (table_stride != 16u) return false;
if (table_source.dword_count == 2u) {
if (!MatchPackedScalarKey(key, indirect)) {
// EXEC implications must describe the same iteration as the selected value.