raphiaura/KytyPS5
1
0
Fork 0
mirror of https://github.com/KytyPS5/KytyPS5.git synced 2026-10-23 12:26:56 +02:00
Code Issues Packages Projects Releases Wiki Activity

shader: resolve workgroup-selected descriptor prefixes from dispatch arguments

Browse Source 
Co-authored-by: fxpw <fxpw@outlook.com>
This commit is contained in:
nmzik
2026-10-14 20:29:48 +02:00
parent 82470c7181
commit ceccd746d8
9 changed files with 358 additions and 36 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/graphics/shader/recompiler/ir/passes/ResourceTracking.cpp
+14 -6
View File
@@ -2204,6 +2204,7 @@ private:
const auto* memory = ScalarReadMemory(*read, memory_index);
Value index;
Value count;
uint32_t workgroup_axis = UINT32_MAX;
DescriptorSource count_source;
uint32_t row_stride = 0;
uint32_t offset = 0, stride = 0;
@@ -2212,19 +2213,21 @@ private:
if (MatchTableOffset(read->Arg(1), index, offset, stride) &&
((offset | stride | memory->offset) & 3u) == 0u &&
uint64_t {offset} + memory->offset <= UINT32_MAX) {
workgroup_axis = WorkgroupAxis(index);
const auto* bound = BoundedLoop(index, read->Parent());
if (bound != nullptr && bound->GetOpcode() == ValueOpcode::ULessThan32) {
count = bound->Arg(1);
}
}
// Raw pointers have no V# bounds for enumerating an unbounded selector domain.
if (memory->kind == ResourceKind::ScalarAddress && count.IsEmpty()) return false;
const bool rows = count.IsEmpty() && memory->offset == 0u &&
if (memory->kind == ResourceKind::ScalarAddress && count.IsEmpty() &&
workgroup_axis == UINT32_MAX) return false;
const bool rows = count.IsEmpty() && workgroup_axis == UINT32_MAX && memory->offset == 0u &&
MatchCountedRows(*read, count_source, row_stride);
if (rows) {
stride = 4u;
offset = 0u;
} else if (count.IsEmpty()) {
} else if (count.IsEmpty() && workgroup_axis == UINT32_MAX) {
auto address = read->Arg(1);
auto active = Value(true);
const auto* lane = address.Resolve().TryInstruction();
@@ -2255,17 +2258,18 @@ private:
}
// A bounded image heap already supports GPU selection. Only enumerate a
// narrower proven selector domain instead of scanning unrelated material data.
if (require_bound && count.IsEmpty() && !rows) return false;
if (require_bound && count.IsEmpty() && workgroup_axis == UINT32_MAX && !rows) return false;
if (!MakeRuntimeTableSource(*read, material_source)) return false;
selector.source = InternSource(material_source);
selector.stride = stride;
selector.offset = offset + (memory->offset & ~3u);
indirect.selector = std::move(selector);
indirect.key_count = count;
indirect.workgroup_axis = workgroup_axis;
if (rows) indirect.selector->rows.emplace(
DescriptorSource::IndirectDescriptor::SelectorRead::CountedRows {
.source = InternSource(count_source), .stride = row_stride});
if (!count.IsEmpty()) indirect.selector_first = Value(0u);
if (!count.IsEmpty() || workgroup_axis != UINT32_MAX) indirect.selector_first = Value(0u);
return true;
}
@@ -3025,6 +3029,11 @@ private:
indirect.table_scalar = true;
return finish();
}
DescriptorSource material_source;
if (!plan.split_offsets && MatchScalarMaterialKey(key, indirect, material_source, true)) {
plan.retain_reads = true;
return finish();
}
if (!legacy_store) {
if (table_source.dword_count == 2u && !plan.split_offsets) {
const auto maximum = SelectorMaximum(key);
@@ -3082,7 +3091,6 @@ private:
!ImmediateU32(read->Arg(2), offset) || offset != 0u ||
!ImmediateU32(read->Arg(3), scalar) || scalar != 0u ||
!EquivalentValue(m_program, selected->Arg(0), read->Arg(4))) return false;
DescriptorSource material_source;
if (!MakeRuntimeTableSource(*read, material_source) ||
!BoundedSelectedIndex(read->Arg(1), read->Arg(4), indirect)) return false;
indirect.selector.emplace(DescriptorSource::IndirectDescriptor::SelectorRead{.source = InternSource(material_source),