gremling/KytyPS5
1
0 Fork 0
mirror of https://github.com/KytyPS5/KytyPS5.git synced 2026-10-27 16:02:37 +00:00
Code Issues Packages Projects Releases Wiki Activity

shader: resolve GPU-selected parent pointers and bounded child image tables

Browse Source 
Co-authored-by: fxpw <fxpw@outlook.com>
This commit is contained in:
nmzikandfxpw committed 2026-10-15 22:28:01 +02:00
1 parent 2285a25182
commit e291589177
5 files changed
+583 -146
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

src/graphics/shader/recompiler/ir/passes/ResourceMaterialization.cpp
+46 -30
View File
@@ -588,9 +588,18 @@ bool MaterializeIndirectDescriptor(const ResourcePlan& p
table.OutOfBounds() != 0u || (table_base & 3u) != 0u ||
(table.Stride() & 3u) != 0u || table.Stride() < indirect.table_record_bytes)))
return false;
const bool affine = indirect.table_stride != 0u;
const bool image_heap = dword_count == 8u && !indirect.pointer_offset;
const uint64_t step = indirect.table_maximum ? indirect.table_stride
const auto* pointer = indirect.pointer_table ? &*indirect.pointer_table : nullptr;
const bool indexed_pointer = pointer != nullptr && pointer->indexed;
const auto child_count = pointer != nullptr ? pointer->count : 1u;
if (indexed_pointer &&
(!indirect.table_maximum || table.Type() != 0u || table.SwizzleEnabled() || table.AddTid() ||
table.OutOfBounds() != 0u || (table_base & 3u) != 0u || (table.Stride() & 3u) != 0u ||
uint64_t {indirect.table_immediate} + 8u > table.Stride() ||
*indirect.table_maximum >= table.NumRecords())) return false;
const bool affine = indirect.table_stride != 0u || indexed_pointer;
const bool image_heap = dword_count == 8u && pointer == nullptr;
const uint64_t step = indexed_pointer ? table.Stride()
: indirect.table_maximum ? indirect.table_stride
: affine ? std::max(uint64_t {1} << std::countr_zero(indirect.table_stride), uint64_t {4})
: 4u;
const uint64_t first = ((indirect.table_maximum ? indirect.table_offset
@@ -601,7 +610,8 @@ bool MaterializeIndirectDescriptor(const ResourcePlan& p
const auto count = table_size >= first + 4u
? std::min((table_size - first - 4u) / step + 1u, domain)
: 0u;
if (!image_heap && count > MaxIndirectDescriptorProbes) return false;
if (child_count == 0u || (!image_heap && count * child_count > MaxIndirectDescriptorProbes))
return false;
// Affine SMEM offsets wrap in U32 before adding the aligned immediate.
// Image heaps retain one bulk snapshot and their absolute byte-offset mapping.
if (image_heap) {
@@ -612,34 +622,40 @@ bool MaterializeIndirectDescriptor(const ResourcePlan& p
descriptors[resource_index] = {.dword_count = dword_count};
snapshot.flattened_srt.resize(mapping_offset + 2u, 0u);
for (uint32_t word = 0; word < count; ++word) {
DescriptorValue candidate {.dword_count = dword_count};
const auto table_offset = affine
? first + word * step : uint64_t {word} * 4u;
if (indirect.pointer_offset) {
const auto table_offset = affine ? first + word * step : uint64_t {word} * 4u;
uint64_t pointer_base = 0u;
if (pointer != nullptr) {
std::array<uint32_t, 2> address {};
if (!ReadScalarTable(table_base, table_size, table_offset, runtime, address) ||
!ReadScalarTable((uint64_t {address[1]} << 32u) | address[0], UINT64_MAX,
*indirect.pointer_offset, runtime,
std::span(candidate.dwords).first(read_dwords))) return false;
} else if (!image_heap) {
if (!ReadScalarTable(table_base, table_size, table_offset, runtime,
std::span(candidate.dwords).first(read_dwords))) return false;
} else {
const auto table_word = table_offset / 4u;
std::copy_n(keys.begin() + table_word,
std::min<size_t>(read_dwords, keys.size() - table_word),
candidate.dwords.begin());
if (!ReadScalarTable(table_base, table_size, table_offset, runtime, address)) return false;
pointer_base = (uint64_t {address[1]} << 32u) | address[0];
}
for (uint32_t child = 0; child < child_count; ++child) {
DescriptorValue candidate {.dword_count = dword_count};
if (pointer != nullptr) {
if (!ReadScalarTable(pointer_base, UINT64_MAX,
uint64_t {child} * pointer->stride + pointer->offset, runtime,
std::span(candidate.dwords).first(read_dwords))) return false;
} else if (!image_heap) {
if (!ReadScalarTable(table_base, table_size, table_offset, runtime,
std::span(candidate.dwords).first(read_dwords))) return false;
} else {
const auto table_word = table_offset / 4u;
std::copy_n(keys.begin() + table_word,
std::min<size_t>(read_dwords, keys.size() - table_word),
candidate.dwords.begin());
}
if (!normalize(candidate))
return normalization_failed(table_offset, word, step,
indirect.table_maximum ? "key" : "probe");
uint32_t ordinal;
if (!intern_candidate(candidate, ordinal)) return false;
if (ordinal == 0u) continue;
const auto mapping_word = image_heap ? table_offset / 4u : word * child_count + child;
snapshot.flattened_srt.resize(mapping_offset + 2u + mapping_word, 0u);
snapshot.flattened_srt[mapping_offset + 1u + mapping_word] = ordinal;
}
if (!normalize(candidate))
return normalization_failed(table_offset, word, step,
indirect.table_maximum ? "key" : "probe");
uint32_t ordinal;
if (!intern_candidate(candidate, ordinal)) return false;
if (ordinal == 0u) continue;
const auto mapping_word = image_heap ? table_offset / 4u : word;
snapshot.flattened_srt.resize(mapping_offset + 2u + mapping_word, 0u);
snapshot.flattened_srt[mapping_offset + 1u + mapping_word] = ordinal;
}
if (descriptors.size() == children_begin) {
snapshot.flattened_srt.resize(mapping_offset);
} else {
@@ -871,7 +887,7 @@ static bool BuildResourceSpecialization(const ResourcePlan& program, ResourceSna
if (i >= program.info.images.size() && source != nullptr &&
source->indirect_descriptor &&
IsBoundedDescriptorTable(program, *source->indirect_descriptor) &&
!source->indirect_descriptor->pointer_offset) {
!source->indirect_descriptor->pointer_table) {
const auto& indirect = *source->indirect_descriptor;
const size_t mapping = specialization.images[base_index].indirect_mapping_offset;
const auto& flat = snapshot.flattened_srt;