raphiaura/KytyPS5
1
0
Fork 0
mirror of https://github.com/KytyPS5/KytyPS5.git synced 2026-10-21 11:26:56 +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/ResourceMaterialization.cpp
+42 -9
View File
@@ -307,9 +307,10 @@ bool MaterializeIndirectDescriptor(const ResourcePlan& p
const auto root_resource = specializations[resource_index];
const auto evaluate_count = [&](uint32_t& count) {
if (indirect.workgroup_axis != UINT32_MAX) {
if (indirect.workgroup_axis >= runtime.workgroup_counts.size()) return false;
count = runtime.workgroup_counts[indirect.workgroup_axis];
return count != 0u;
if (runtime.workgroup_counts.size() != 3u || indirect.workgroup_axis >= 3u) return false;
count = std::ranges::find(runtime.workgroup_counts, 0u) != runtime.workgroup_counts.end()
? 0u : runtime.workgroup_counts[indirect.workgroup_axis];
return true;
}
if (!clean.Evaluate(indirect.key_count, count)) return false;
if (indirect.signed_count && std::bit_cast<int32_t>(count) < 0) count = 0;
@@ -337,14 +338,28 @@ bool MaterializeIndirectDescriptor(const ResourcePlan& p
const auto read_keys = [&](uint64_t base, uint64_t size, uint64_t first,
uint64_t step, uint64_t count) {
keys.resize(count);
if (count == 0u) return true;
if (step == 4u) {
return read_selector(base, size, first, keys);
}
for (auto& key: keys) {
if (!read_selector(base, size, first, {&key, 1}))
if (step < 4u) return false;
for (uint64_t copied = 0; copied < count;) {
const auto offset = first + copied * step;
// Scalar reads align each strided offset independently.
const auto bias = selector->global ? 0u : offset & 3u;
const auto length = std::min(count - copied,
1u + (uint64_t {MaxIndirectDescriptorProbes - 1u} * 4u - bias) / step);
keys.resize(count + (bias + (length - 1u) * step) / 4u + 1u);
const auto batch = std::span(keys).subspan(count);
if (!read_selector(base, size, offset, batch)) {
keys.resize(count);
return false;
first += step;
}
for (uint64_t index = 0; index < length; ++index)
keys[copied + index] = batch[(bias + index * step) / 4u];
copied += length;
}
keys.resize(count);
return true;
};
uint64_t table_base = 0;
@@ -564,7 +579,8 @@ bool MaterializeIndirectDescriptor(const ResourcePlan& p
uint64_t {first} + count > material.NumRecords())) return false;
const auto end = count != 0u
? (uint64_t {first} + count - 1u) * step + selector->offset + 4u : 0u;
if (end > uint64_t {UINT32_MAX} + 1u || end > size)
if (end > uint64_t {UINT32_MAX} + 1u ||
(indirect.workgroup_axis == UINT32_MAX && end > size))
return false;
if (!read_keys(base, size, uint64_t {first} * step + selector->offset, step, count))
return false;
@@ -1331,11 +1347,16 @@ bool MaterializeResources(const ResourcePlan& program, const SrtRuntime& runtime
runtime.read_specialization_memory == nullptr)) {
return false;
}
const bool capture_reads = program.capture_specialization_reads || !program.source_reads.empty();
const auto uses_workgroup_count = [](const DescriptorSource& source) {
return source.indirect_descriptor && source.indirect_descriptor->workgroup_axis != UINT32_MAX;
};
const bool capture_reads = program.capture_specialization_reads || !program.source_reads.empty() ||
(runtime.indirect_args_address != 0u &&
std::ranges::any_of(program.descriptor_sources, uses_workgroup_count));
auto& reads = snapshot.specialization_reads;
reads = program.source_reads;
SrtReadCapture capture {runtime, reads};
const auto observed = capture_reads ? capture.ObservedRuntime() : runtime;
auto observed = capture_reads ? capture.ObservedRuntime() : runtime;
SrtWalker clean(program, CleanRuntime(observed));
SrtWalker walker(program, observed,
capture_reads || program.requires_specialization_memory ? &clean : nullptr);
@@ -1343,6 +1364,18 @@ bool MaterializeResources(const ResourcePlan& program, const SrtRuntime& runtime
return false;
}
const auto active = std::span<const uint8_t>(program.active_sources);
std::array<uint32_t, 3> workgroup_counts {};
if (runtime.indirect_args_address != 0u) {
for (uint32_t source = 0; source < program.descriptor_sources.size(); ++source) {
if ((!active.empty() && !active[source]) ||
!uses_workgroup_count(program.descriptor_sources[source])) continue;
if ((runtime.indirect_args_address & 3u) != 0u ||
!ReadTable(runtime.indirect_args_address, sizeof(workgroup_counts), 0u,
observed, workgroup_counts)) return false;
observed.workgroup_counts = workgroup_counts;
break;
}
}
snapshot.uniform_fill = {};
const auto& fill = program.uniform_fill;
const auto words = fill.fill.words;