raphiaura/KytyPS5
1
0
Fork 0
mirror of https://github.com/KytyPS5/KytyPS5.git synced 2026-10-27 13:26:52 +01:00
Code Issues Packages Projects Releases Wiki Activity

shader: enumerate set ranks of eight-word light bitmaps

Browse Source
This commit is contained in:
Claxten
2026-10-27 00:11:27 +01:00
committed by nmzik
parent ffb6ad6cab
commit 914a3bb4b8
3 changed files with 122 additions and 57 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/ResourceTrackingTests.cpp
+75 -28
View File
@@ -2505,11 +2505,15 @@ void TestDescendingImageRange() {
enum class Variant {
Valid, WrongStep, IncrementBypass, MissingNonzero, NegativeInitial,
Bitmap, BitmapAddsBit, BitmapWrongWord, BitmapUnknownInitial,
BitmapByteOffset, BitmapScalarOffset, BitmapOuterClobber, BitmapWrongLoadMask
BitmapByteOffset, BitmapScalarOffset, BitmapOuterClobber, BitmapWrongLoadMask,
Bitmap8, Bitmap8MisplacedLoad, Bitmap8OtherBuffer, Bitmap8OuterClobber, Bitmap8WrongUpperWord,
Bitmap8NarrowMask, Bitmap8OtherGuard
};
const auto make_plan = [](Variant variant) {
Fixture fixture;
const bool bitmap = variant >= Variant::Bitmap;
// Eight-word records are two four-word loads of one buffer record.
const uint32_t bitmap_words = variant >= Variant::Bitmap8 ? 8u : 4u;
auto *entry = fixture.block;
auto *header = fixture.AddBlock();
auto *guard = fixture.AddBlock();
@@ -2532,11 +2536,11 @@ void TestDescendingImageRange() {
branch(latch, header);
auto &induction = header->AppendNewInst(ValueOpcode::Phi);
induction.SetFlags(Type::U32);
std::array<Inst *, 4> outer_words {}, remaining_words {};
std::array<Value, 4> loaded_words {};
std::array<Inst *, 8> outer_words {}, remaining_words {};
std::array<Value, 8> loaded_words {};
if (bitmap)
for (auto& word : outer_words)
word = &header->AppendNewInst(ValueOpcode::Phi, {}, uint64_t(Type::U32));
for (uint32_t word = 0; word < bitmap_words; ++word)
outer_words[word] = &header->AppendNewInst(ValueOpcode::Phi, {}, uint64_t(Type::U32));
const auto bound = fixture.Emit(ValueOpcode::SGreaterThanEqual32,
{Value(&induction), Value(0u)}, 0, header);
choose(header, guard, variant == Variant::IncrementBypass ? latch : exit, bound);
@@ -2553,14 +2557,27 @@ void TestDescendingImageRange() {
read.idxen = true;
read.offen = variant == Variant::BitmapByteOffset;
read.data_dwords = 4u;
const auto packed = fixture.Emit(ValueOpcode::LoadBufferU32x4,
{buffer, lane,
Value(variant == Variant::BitmapByteOffset ? 4u : 0u),
Value(variant == Variant::BitmapScalarOffset ? 4u : 0u), load_active},
fixture.AddMemory(read, 0x10));
for (uint32_t word = 0; word < 4u; ++word) {
std::array<Value, 2> packed;
for (uint32_t part = 0; part < bitmap_words / 4u; ++part) {
read.offset = part * (variant == Variant::Bitmap8MisplacedLoad ? 32u : 16u);
const auto source = variant == Variant::Bitmap8OtherBuffer && part != 0u
? fixture.Buffer({fixture.UserData(8), fixture.UserData(9), fixture.UserData(10),
fixture.UserData(11)}) : buffer;
const auto guard = variant == Variant::Bitmap8OtherGuard && part != 0u
? fixture.Emit(ValueOpcode::LogicalAnd,
{load_active, fixture.Emit(ValueOpcode::ULessThan32, {lane, Value(16u)})})
: load_active;
packed[part] = fixture.Emit(ValueOpcode::LoadBufferU32x4,
{source, lane,
Value(variant == Variant::BitmapByteOffset ? 4u : 0u),
Value(variant == Variant::BitmapScalarOffset ? 4u : 0u), guard},
fixture.AddMemory(read, 0x10 + part * 8u));
}
for (uint32_t word = 0; word < bitmap_words; ++word) {
const bool wrong_word = variant == Variant::BitmapWrongWord ||
(variant == Variant::Bitmap8WrongUpperWord && word >= 4u);
auto component = fixture.Emit(ValueOpcode::CompositeExtractU32x4,
{packed, Value(variant == Variant::BitmapWrongWord ? (word + 1u) % 4u : word)}, 0, entry);
{packed[word / 4u], Value((wrong_word ? word + 1u : word) % 4u)}, 0, entry);
component = fixture.Emit(ValueOpcode::SelectU32,
{load_active, component, fixture.Emit(ValueOpcode::UndefU32)});
loaded_words[word] = component;
@@ -2573,7 +2590,7 @@ void TestDescendingImageRange() {
remaining_words[word]->AddPhiOperand(header, Value(outer_words[word]));
}
auto selected = fixture.Emit(ValueOpcode::UndefU32, {}, 0, guard);
for (uint32_t word = 0; word < 4u; ++word) {
for (uint32_t word = 0; word < bitmap_words; ++word) {
selected = fixture.Emit(ValueOpcode::SelectU32,
{fixture.Emit(ValueOpcode::IEqual32, {Value(&induction), Value(word)}, 0, guard),
Value(remaining_words[word]), selected}, 0, guard);
@@ -2636,32 +2653,39 @@ void TestDescendingImageRange() {
{fixture.Emit(ValueOpcode::ShiftLeftLogical32, {Value(1u), bit})})});
if (variant == Variant::BitmapAddsBit)
next = fixture.Emit(ValueOpcode::BitwiseOr32, {next, Value(1u << 21u)});
for (uint32_t word = 0; word < 4u; ++word) {
for (uint32_t word = 0; word < bitmap_words; ++word) {
// Relative register writes index with M0 & 0xff; a narrower mask aliases upper words.
const auto index = bitmap_words == 4u ? Value(&induction)
: fixture.Emit(ValueOpcode::BitwiseAnd32,
{Value(&induction), Value(variant == Variant::Bitmap8NarrowMask ? 3u : 255u)});
const auto selected_word = fixture.Emit(ValueOpcode::LogicalAnd,
{active, fixture.Emit(ValueOpcode::IEqual32, {Value(&induction), Value(word)})});
{active, fixture.Emit(ValueOpcode::IEqual32, {index, Value(word)})});
remaining_words[word]->AddPhiOperand(body, fixture.Emit(ValueOpcode::SelectU32,
{selected_word, next, Value(remaining_words[word])}));
auto carried = Value(remaining_words[word]);
if (variant == Variant::BitmapOuterClobber)
if (variant == Variant::BitmapOuterClobber || variant == Variant::Bitmap8OuterClobber)
carried = fixture.Emit(ValueOpcode::SelectU32,
{fixture.Emit(ValueOpcode::IEqual32,
{Value(&induction), Value(word + 1u)}, 0, latch),
loaded_words[(word + 1u) % 4u], carried}, 0, latch);
loaded_words[(word + 1u) % bitmap_words], carried}, 0, latch);
outer_words[word]->AddPhiOperand(latch, carried);
}
}
const auto step = fixture.Emit(ValueOpcode::IAdd32,
{Value(&induction), Value(variant == Variant::WrongStep ? UINT32_MAX - 1u : UINT32_MAX)}, 0, latch);
induction.AddPhiOperand(entry, Value(variant == Variant::NegativeInitial ? UINT32_MAX : 3u));
induction.AddPhiOperand(entry,
Value(variant == Variant::NegativeInitial ? UINT32_MAX : bitmap_words - 1u));
induction.AddPhiOperand(latch, step);
fixture.PlanAndTrack();
Check(ordinal.Instruction()->Arg(1).Resolve() == bit.Resolve(),
"descending range retained impossible relative-register stores");
const auto &indirect = fixture.program.descriptor_sources[
fixture.program.info.images[0].source].indirect_descriptor;
Check(indirect && indirect->selector.has_value() == (variant == Variant::Bitmap) &&
(!indirect->selector || indirect->selector->bitset) &&
indirect->key_count == Value(128u) &&
const bool selects = variant == Variant::Bitmap || variant == Variant::Bitmap8;
Check(indirect && indirect->selector.has_value() == selects &&
(!indirect->selector || (indirect->selector->bitset &&
indirect->selector->stride == bitmap_words * 4u)) &&
indirect->key_count == Value(bitmap_words * 32u) &&
!indirect->selector_first.IsEmpty(),
"descending bitscan range lost its bound or accepted an unproved bitmap subset");
return ExtractResourcePlan(fixture.program);
@@ -2701,10 +2725,14 @@ void TestDescendingImageRange() {
}
auto bitmap_plan = make_plan(Variant::Bitmap);
auto wide_plan = make_plan(Variant::Bitmap8);
for (const auto variant : {Variant::BitmapAddsBit, Variant::BitmapWrongWord,
Variant::BitmapUnknownInitial, Variant::BitmapByteOffset,
Variant::BitmapScalarOffset, Variant::BitmapOuterClobber,
Variant::BitmapWrongLoadMask})
Variant::BitmapWrongLoadMask, Variant::Bitmap8MisplacedLoad,
Variant::Bitmap8OtherBuffer, Variant::Bitmap8OuterClobber,
Variant::Bitmap8WrongUpperWord, Variant::Bitmap8NarrowMask,
Variant::Bitmap8OtherGuard})
make_plan(variant);
// Every descriptor is distinct, so retained keys are visible in the snapshot as well as
// its lookup table. An unrelated table slot must never be read, even when it is malformed.
@@ -2719,23 +2747,25 @@ void TestDescendingImageRange() {
const SrtRuntime bitmap_runtime {.user_data = bitmap_data, .userdata = &memory,
.read_specialization_memory = ReadLinearTestMemory};
memory.watched_address = bitmap_address;
auto *checked_plan = &bitmap_plan;
const auto check_bitmap = [&](std::initializer_list<uint32_t> ranks, uint32_t count) {
bitmap_data[2] = count;
memory.watched_reads = memory.descriptor_reads = 0u;
std::vector<uint32_t> keys;
for (const auto rank : ranks) keys.push_back(count - rank);
std::ranges::sort(keys);
Check(MaterializeResources(bitmap_plan, bitmap_runtime, snapshot, specialization) &&
const auto dwords = bitmap_data[6] * (bitmap_data[5] >> 18u);
Check(MaterializeResources(*checked_plan, bitmap_runtime, snapshot, specialization) &&
memory.watched_reads == uint32_t(bitmap_data[6] != 0u) &&
(bitmap_data[6] == 0u || memory.watched_dwords ==
std::min(bitmap_data[6] * 4u, MaxIndirectDescriptorProbes)) &&
std::min(dwords, MaxIndirectDescriptorProbes)) &&
memory.descriptor_reads == keys.size() &&
snapshot.images.size() == std::max<size_t>(keys.size(), 1u),
"bitmap selector lost bounded reads or retained unreachable descriptors");
for (uint32_t first = 0; first < bitmap_data[6] * 4u; first += MaxIndirectDescriptorProbes)
for (uint32_t first = 0; first < dwords; first += MaxIndirectDescriptorProbes)
Check(std::ranges::find(snapshot.specialization_reads,
std::pair{uint64_t{bitmap_address} + first * 4u,
uint64_t{std::min(bitmap_data[6] * 4u - first, MaxIndirectDescriptorProbes)} * 4u}) !=
uint64_t{std::min(dwords - first, MaxIndirectDescriptorProbes)} * 4u}) !=
snapshot.specialization_reads.end(),
"bitmap chunk was omitted from same-dispatch write-alias protection");
if (keys.empty()) {
@@ -2787,8 +2817,8 @@ void TestDescendingImageRange() {
memory.fail_address = UINT64_MAX;
const auto valid_data = bitmap_data;
for (const auto [word, value] : std::array<std::pair<uint32_t, uint32_t>, 6>{{
{4u, bitmap_address + 1u}, {5u, 8u << 16u}, {5u, (16u << 16u) | (1u << 31u)},
for (const auto [word, value] : std::array<std::pair<uint32_t, uint32_t>, 7>{{
{4u, bitmap_address + 1u}, {5u, 8u << 16u}, {5u, 32u << 16u}, {5u, (16u << 16u) | (1u << 31u)},
{7u, 0x5204u | (1u << 23u)}, {7u, 0x5204u | (1u << 28u)},
{7u, 0x5204u | (1u << 30u)}}}) {
bitmap_data[word] = value;
@@ -2828,6 +2858,23 @@ void TestDescendingImageRange() {
return read.first >= memory.fail_address;
}),
"a complete bitmap domain read redundant chunks or lost candidate keys");
// Eight-word records keep their upper words' ranks, and a key below the table is never read.
checked_plan = &wide_plan;
bitmap_data = valid_data;
bitmap_data[5] = 32u << 16u;
bitmap_data[6] = 2u;
std::fill_n(memory.words.begin() + bitmap_word, 16u, 0u);
memory.words[bitmap_word] = 1u | (1u << 2u);
memory.words[bitmap_word + 12u] = 1u << 1u;
memory.fail_address = 0x2000u - 32u;
check_bitmap({0u, 2u, 129u}, 2u);
memory.fail_address = UINT64_MAX;
for (const auto stride : {16u, 64u}) {
bitmap_data[5] = stride << 16u;
Check(!MaterializeResources(wide_plan, bitmap_runtime, snapshot, specialization),
"an eight-word bitmap selector accepted a different record stride");
}
}
void TestUniformizedMaterialImageKeys() {