mirror of
https://github.com/KytyPS5/KytyPS5.git
synced 2026-10-28 16:23:05 +00:00
shader: link multiple compute helpers without synthetic return loops
This commit is contained in:
nmzik committed
1 parent
b2d3ce6dfc
commit
ab8566591f
8 files changed
+363
-109
No files matched your search
@@ -481,18 +481,14 @@ Decoder::Program DecodeFusedProgram(std::span<const uint32_t> front, std::span<c
|
||||
return result;
|
||||
}
|
||||
|
||||
void PrepareCallTarget(ShaderSource& source, const CompileOptions& options) {
|
||||
void PrepareCallTargets(ShaderSource& source, const CompileOptions& options) {
|
||||
using namespace IR;
|
||||
EXIT_NOT_IMPLEMENTED(options.stage != ShaderType::Compute);
|
||||
for (uint32_t i = 0; i < source.decoded.instructions.size(); ++i) {
|
||||
if (source.decoded.instructions[i].opcode != Decoder::Opcode::S_SWAPPC_B64) continue;
|
||||
EXIT_NOT_IMPLEMENTED(source.call.has_value() || options.stage != ShaderType::Compute);
|
||||
source.call = ShaderSource::Call {.instruction = i};
|
||||
source.calls.push_back({.instruction = i});
|
||||
}
|
||||
EXIT_IF(options.input_info.compute == nullptr);
|
||||
const auto& call = source.decoded.instructions[source.call->instruction];
|
||||
EXIT_NOT_IMPLEMENTED(call.src0.kind != Decoder::OperandKind::Sgpr ||
|
||||
call.dst.kind != Decoder::OperandKind::Sgpr ||
|
||||
call.src0.reg != call.dst.reg || call.dst.reg >= 105u);
|
||||
const auto graph = CFG::BuildGraph(source.decoded);
|
||||
Program query;
|
||||
query.stage = options.stage;
|
||||
@@ -524,7 +520,11 @@ void PrepareCallTarget(ShaderSource& source, const CompileOptions& options) {
|
||||
return value;
|
||||
};
|
||||
Register = [&](uint32_t before, uint32_t code) -> Value {
|
||||
const auto index = CFG::FindScalarDefinition(source.decoded, graph, before, code);
|
||||
std::vector<uint32_t> crossed_calls;
|
||||
const auto index = CFG::FindScalarDefinition(source.decoded, graph, before, code, crossed_calls);
|
||||
for (const auto crossed: crossed_calls) {
|
||||
std::ranges::find(source.calls, crossed, &ShaderSource::Call::instruction)->preserved_registers.set(code);
|
||||
}
|
||||
if (index == UINT32_MAX) {
|
||||
EXIT_IF(code < query.user_data_base || code - query.user_data_base >= query.user_data_count);
|
||||
return IREmitter(&block).GetUserData(static_cast<ScalarReg>(code));
|
||||
@@ -557,18 +557,25 @@ void PrepareCallTarget(ShaderSource& source, const CompileOptions& options) {
|
||||
EXIT_IF(!producer.ready || !producer.values.contains(code));
|
||||
return producer.values.at(code).Resolve();
|
||||
};
|
||||
for (uint32_t i = 0; i < 2u; ++i) {
|
||||
const auto value = Register(source.call->instruction, call.src0.reg + i);
|
||||
EXIT_IF(!ValidateRuntimeValue(query, value, RuntimeValueType::Integer));
|
||||
auto& root = query.value_storage.emplace_back(ValueOpcode::ReferenceU32);
|
||||
root.SetArg(0, value);
|
||||
source.call->target[i] = Value(&root);
|
||||
for (auto& site: source.calls) {
|
||||
const auto& call = source.decoded.instructions[site.instruction];
|
||||
EXIT_NOT_IMPLEMENTED(call.src0.kind != Decoder::OperandKind::Sgpr ||
|
||||
call.dst.kind != Decoder::OperandKind::Sgpr ||
|
||||
call.src0.reg != call.dst.reg || call.dst.reg >= 105u);
|
||||
for (uint32_t i = 0; i < 2u; ++i) {
|
||||
const auto value = Register(site.instruction, call.src0.reg + i);
|
||||
EXIT_IF(!ValidateRuntimeValue(query, value, RuntimeValueType::Integer));
|
||||
auto& root = query.value_storage.emplace_back(ValueOpcode::ReferenceU32);
|
||||
root.SetArg(0, value);
|
||||
site.target[i] = Value(&root);
|
||||
}
|
||||
}
|
||||
RewriteToSsa(query.blocks);
|
||||
ConstantPropagationPass(query.blocks, query.wave_size);
|
||||
RemoveIdentities(query);
|
||||
EliminateDeadCode(query.blocks);
|
||||
for (auto& value: source.call->target) value = value.Instruction()->Arg(0).Resolve();
|
||||
for (auto& site: source.calls)
|
||||
for (auto& value: site.target) value = value.Instruction()->Arg(0).Resolve();
|
||||
query.value_storage.clear();
|
||||
for (auto& inst: block) {
|
||||
inst.SetParent(nullptr);
|
||||
@@ -595,7 +602,7 @@ ShaderSource PrepareShaderSource(std::span<const uint32_t> code, const CompileOp
|
||||
Decoder::DecodeProgram(code, source.decoded);
|
||||
}
|
||||
if (source.decoded.has_swap_pc) {
|
||||
PrepareCallTarget(source, options);
|
||||
PrepareCallTargets(source, options);
|
||||
source.code.assign(code.begin(), code.end());
|
||||
source.decoded.code = source.code;
|
||||
}
|
||||
@@ -608,83 +615,116 @@ const Decoder::Program& RefreshShaderSource(ShaderSource& source, const IR::SrtR
|
||||
IR::SrtReadCapture capture(runtime, reads);
|
||||
const auto clean = IR::CleanRuntime(capture.ObservedRuntime());
|
||||
IR::SrtWalker walker(source.call_targets, clean);
|
||||
uint32_t low = 0, high = 0;
|
||||
const auto& call = source.decoded.instructions[source.call->instruction];
|
||||
if (!walker.Evaluate(source.call->target[0], low) || !walker.Evaluate(source.call->target[1], high))
|
||||
EXIT("shader call at pc 0x%08x has an unavailable scalar target", call.pc);
|
||||
// SWAPPC reads the old aliased pair and ignores the target's low two bits.
|
||||
const auto address = ((uint64_t {high} << 32u) | low) & ~uint64_t {3};
|
||||
EXIT_IF(address == 0u);
|
||||
for (auto& site: source.calls) {
|
||||
uint32_t low = 0, high = 0;
|
||||
if (!walker.Evaluate(site.target[0], low) || !walker.Evaluate(site.target[1], high))
|
||||
EXIT("shader call at pc 0x%08x has an unavailable scalar target",
|
||||
source.decoded.instructions[site.instruction].pc);
|
||||
// SWAPPC reads the old aliased pair and ignores the target's low two bits.
|
||||
const auto address = ((uint64_t {high} << 32u) | low) & ~uint64_t {3};
|
||||
EXIT_IF(address == 0u);
|
||||
site.address = address;
|
||||
}
|
||||
const auto query_reads = reads.size();
|
||||
if (source.linked) {
|
||||
auto& linked = *source.linked;
|
||||
const auto function = std::span(linked.code).subspan(source.code.size());
|
||||
if (clean.read_memory(clean.userdata, address, linked.observed_function) &&
|
||||
std::ranges::equal(function, linked.observed_function)) {
|
||||
return linked.decoded;
|
||||
bool unchanged = true;
|
||||
for (size_t i = 0; i < source.calls.size(); ++i) {
|
||||
const auto& function = linked.functions[i];
|
||||
const auto expected = std::span(linked.code).subspan(function.offset, function.size);
|
||||
const auto preceding = std::span(source.calls).first(i);
|
||||
const auto previous = std::ranges::find(preceding, source.calls[i].address, &ShaderSource::Call::address);
|
||||
if (previous != preceding.end()) {
|
||||
const auto& observed = linked.functions[previous - preceding.begin()];
|
||||
if (std::ranges::equal(expected, std::span(linked.code).subspan(observed.offset, observed.size))) continue;
|
||||
unchanged = false;
|
||||
break;
|
||||
}
|
||||
linked.observed_function.resize(function.size);
|
||||
if (!clean.read_memory(clean.userdata, source.calls[i].address, linked.observed_function) ||
|
||||
!std::ranges::equal(expected, linked.observed_function)) {
|
||||
unchanged = false;
|
||||
break;
|
||||
}
|
||||
}
|
||||
if (unchanged) return linked.decoded;
|
||||
reads.resize(query_reads);
|
||||
}
|
||||
ShaderSource::Linked linked {.code = source.code,
|
||||
.decoded = source.decoded};
|
||||
ShaderSource::Linked linked {.code = source.code, .decoded = source.decoded};
|
||||
const auto& last = source.decoded.instructions.back();
|
||||
EXIT_NOT_IMPLEMENTED(last.opcode != Decoder::Opcode::S_ENDPGM);
|
||||
// Keep the caller's data/footer bytes at their native offsets. The leaf's virtual CFG
|
||||
// PCs follow decoded instructions; its physical code is independently read and owned.
|
||||
const auto function_pc = last.pc + last.word_count * sizeof(uint32_t);
|
||||
for (uint32_t index = 0;;) {
|
||||
// The observed queue helper is a scalar leaf. Scalar encodings need at most one literal.
|
||||
EXIT_IF(index >= 64u || address > UINT64_MAX - (uint64_t {index} + 2u) * sizeof(uint32_t));
|
||||
std::array<uint32_t, 2> words {};
|
||||
if (!clean.read_memory(clean.userdata, address + uint64_t {index} * sizeof(uint32_t),
|
||||
std::span(words).first(1)))
|
||||
EXIT("shader call at pc 0x%08x cannot read leaf code", call.pc);
|
||||
const auto family = Decoder::GetInstructionFamily(words[0]);
|
||||
EXIT_NOT_IMPLEMENTED(family != Decoder::Family::SOP1 && family != Decoder::Family::SOP2 &&
|
||||
family != Decoder::Family::SOPC && family != Decoder::Family::SOPK &&
|
||||
family != Decoder::Family::SOPP);
|
||||
Decoder::Instruction inst;
|
||||
Decoder::DecodeInstruction(words, 0u, inst);
|
||||
if (inst.word_count == 2u) {
|
||||
if (!clean.read_memory(clean.userdata, address + uint64_t {index + 1u} * sizeof(uint32_t),
|
||||
std::span(words).last(1)))
|
||||
EXIT("shader call at pc 0x%08x cannot read leaf literal", call.pc);
|
||||
inst = {};
|
||||
Decoder::DecodeInstruction(words, 0u, inst);
|
||||
// Keep the caller's data/footer at native offsets. Expand leaf CFGs per call site
|
||||
// before SSA, so return edges neither require a dispatcher nor merge resource roots.
|
||||
uint32_t function_pc = last.pc + last.word_count * sizeof(uint32_t);
|
||||
std::map<uint64_t, std::vector<Decoder::Instruction>> functions;
|
||||
for (size_t i = 0; i < source.calls.size(); ++i) {
|
||||
const auto& site = source.calls[i];
|
||||
const auto& call = source.decoded.instructions[site.instruction];
|
||||
const auto address = source.calls[i].address;
|
||||
auto [entry, inserted] = functions.try_emplace(address);
|
||||
auto& function = entry->second;
|
||||
if (inserted) {
|
||||
for (uint32_t index = 0;;) {
|
||||
std::array<uint32_t, Decoder::MaxInstructionRawWords> words {};
|
||||
EXIT_IF(index > UINT32_MAX / sizeof(uint32_t) - words.size() ||
|
||||
address > UINT64_MAX - (uint64_t {index} + words.size()) * sizeof(uint32_t));
|
||||
uint32_t count = 1;
|
||||
Decoder::Instruction inst;
|
||||
for (;;) {
|
||||
if (!clean.read_memory(clean.userdata, address + uint64_t {index} * sizeof(uint32_t),
|
||||
std::span(words).first(count)))
|
||||
EXIT("shader call at pc 0x%08x cannot read leaf code", call.pc);
|
||||
inst = {};
|
||||
Decoder::DecodeInstruction(words, 0u, inst);
|
||||
if (inst.word_count <= count) break;
|
||||
count = inst.word_count;
|
||||
}
|
||||
inst.pc = index * sizeof(uint32_t);
|
||||
if (Decoder::IsDirectBranch(inst.opcode)) inst.branch_target += inst.pc;
|
||||
index += inst.word_count;
|
||||
EXIT_NOT_IMPLEMENTED(inst.opcode == Decoder::Opcode::S_SWAPPC_B64 ||
|
||||
inst.opcode == Decoder::Opcode::S_GETPC_B64 ||
|
||||
inst.opcode == Decoder::Opcode::S_ENDPGM);
|
||||
function.push_back(inst);
|
||||
if (inst.opcode == Decoder::Opcode::S_SETPC_B64) break;
|
||||
}
|
||||
}
|
||||
inst.pc = index * sizeof(uint32_t);
|
||||
if (Decoder::IsDirectBranch(inst.opcode)) inst.branch_target += inst.pc;
|
||||
linked.code.insert(linked.code.end(), inst.raw, inst.raw + inst.word_count);
|
||||
index += inst.word_count;
|
||||
EXIT_NOT_IMPLEMENTED(inst.opcode == Decoder::Opcode::S_SWAPPC_B64 ||
|
||||
inst.opcode == Decoder::Opcode::S_GETPC_B64 ||
|
||||
inst.opcode == Decoder::Opcode::S_ENDPGM);
|
||||
EXIT_IF(CFG::MayWriteScalarRegister(inst, call.dst.reg) || CFG::MayWriteScalarRegister(inst, call.dst.reg + 1u));
|
||||
const bool returns = inst.opcode == Decoder::Opcode::S_SETPC_B64;
|
||||
if (returns) {
|
||||
EXIT_NOT_IMPLEMENTED(inst.src0.kind != Decoder::OperandKind::Sgpr || inst.src0.reg != call.dst.reg);
|
||||
inst.opcode = Decoder::Opcode::S_BRANCH;
|
||||
inst.branch_target = call.pc + call.word_count * sizeof(uint32_t);
|
||||
} else if (Decoder::IsDirectBranch(inst.opcode)) {
|
||||
EXIT_IF(inst.branch_target > UINT32_MAX - function_pc);
|
||||
inst.branch_target += function_pc;
|
||||
const uint32_t function_size = function.back().pc + function.back().word_count * sizeof(uint32_t);
|
||||
EXIT_IF(function_pc > UINT32_MAX - function_size);
|
||||
const auto function_offset = static_cast<uint32_t>(linked.code.size());
|
||||
for (auto inst: function) {
|
||||
EXIT_IF(CFG::MayWriteScalarRegister(inst, call.dst.reg) ||
|
||||
CFG::MayWriteScalarRegister(inst, call.dst.reg + 1u));
|
||||
for (uint32_t code = 0; code < site.preserved_registers.size(); ++code) {
|
||||
if (site.preserved_registers[code] && CFG::MayWriteScalarRegister(inst, code))
|
||||
EXIT("shader call at pc 0x%08x clobbers scalar target dependency s%u", call.pc, code);
|
||||
}
|
||||
linked.code.insert(linked.code.end(), inst.raw, inst.raw + inst.word_count);
|
||||
if (inst.opcode == Decoder::Opcode::S_SETPC_B64) {
|
||||
EXIT_NOT_IMPLEMENTED(inst.src0.kind != Decoder::OperandKind::Sgpr || inst.src0.reg != call.dst.reg);
|
||||
inst.opcode = Decoder::Opcode::S_BRANCH;
|
||||
inst.branch_target = call.pc + call.word_count * sizeof(uint32_t);
|
||||
} else if (Decoder::IsDirectBranch(inst.opcode)) {
|
||||
EXIT_IF(inst.branch_target >= function_size);
|
||||
inst.branch_target += function_pc;
|
||||
}
|
||||
inst.pc += function_pc;
|
||||
linked.decoded.instructions.push_back(std::move(inst));
|
||||
}
|
||||
inst.pc += function_pc;
|
||||
linked.decoded.instructions.push_back(std::move(inst));
|
||||
if (returns) break;
|
||||
}
|
||||
const auto function_end = linked.decoded.instructions.back().pc +
|
||||
linked.decoded.instructions.back().word_count * sizeof(uint32_t);
|
||||
for (size_t i = source.decoded.instructions.size(); i + 1u < linked.decoded.instructions.size(); ++i) {
|
||||
const auto& inst = linked.decoded.instructions[i];
|
||||
if (Decoder::IsDirectBranch(inst.opcode))
|
||||
EXIT_IF(inst.branch_target < function_pc || inst.branch_target >= function_end);
|
||||
linked.functions.push_back({function_offset, function_size / 4u});
|
||||
linked.decoded.instructions[site.instruction].branch_target = function_pc;
|
||||
function_pc += function_size;
|
||||
}
|
||||
reads.resize(query_reads);
|
||||
reads.emplace_back(address, (linked.code.size() - source.code.size()) * sizeof(uint32_t));
|
||||
linked.decoded.instructions[source.call->instruction].branch_target = function_pc;
|
||||
for (size_t i = 0; i < source.calls.size(); ++i) {
|
||||
const auto preceding = std::span(source.calls).first(i);
|
||||
if (std::ranges::find(preceding, source.calls[i].address, &ShaderSource::Call::address) == preceding.end())
|
||||
reads.emplace_back(source.calls[i].address, linked.functions[i].size * sizeof(uint32_t));
|
||||
}
|
||||
linked.decoded.code = linked.code;
|
||||
linked.observed_function.resize(linked.code.size() - source.code.size());
|
||||
linked.observed_function.resize(std::ranges::max_element(linked.functions, {},
|
||||
&ShaderSource::Linked::Function::size)->size);
|
||||
|
||||
source.linked = std::move(linked);
|
||||
++source.revision;
|
||||
return source.linked->decoded;
|
||||
|
||||
Reference in new issue
Block a user