[Vulkan/SPIR-V] Some pipeline layout parts + exec conditionals

This commit is contained in:
Triang3l
2020-10-25 15:09:39 +03:00
parent 972d66d835
commit fdbed73463
4 changed files with 607 additions and 3 deletions

View File

@@ -31,6 +31,9 @@ void SpirvShaderTranslator::Reset() {
main_switch_op_.reset();
main_switch_next_pc_phi_operands_.clear();
cf_exec_conditional_merge_ = nullptr;
cf_instruction_predicate_merge_ = nullptr;
}
void SpirvShaderTranslator::StartTranslation() {
@@ -50,12 +53,15 @@ void SpirvShaderTranslator::StartTranslation() {
type_bool_ = builder_->makeBoolType();
type_int_ = builder_->makeIntType(32);
type_int4_ = builder_->makeVectorType(type_int_, 4);
type_uint_ = builder_->makeUintType(32);
type_uint4_ = builder_->makeVectorType(type_uint_, 4);
type_float_ = builder_->makeFloatType(32);
type_float2_ = builder_->makeVectorType(type_float_, 2);
type_float3_ = builder_->makeVectorType(type_float_, 3);
type_float4_ = builder_->makeVectorType(type_float_, 4);
const_int_0_ = builder_->makeIntConstant(0);
const_uint_0_ = builder_->makeUintConstant(0);
id_vector_temp_.clear();
id_vector_temp_.reserve(4);
for (uint32_t i = 0; i < 4; ++i) {
@@ -71,6 +77,40 @@ void SpirvShaderTranslator::StartTranslation() {
const_float4_0_ =
builder_->makeCompositeConstant(type_float4_, id_vector_temp_);
// Common uniform buffer - bool and loop constants.
id_vector_temp_.clear();
id_vector_temp_.reserve(2);
// 256 bool constants.
id_vector_temp_.push_back(builder_->makeArrayType(
type_uint4_, builder_->makeUintConstant(2), sizeof(uint32_t) * 4));
// Currently (as of October 24, 2020) makeArrayType only uses the stride to
// check if deduplication can be done - the array stride decoration needs to
// be applied explicitly.
builder_->addDecoration(id_vector_temp_.back(), spv::DecorationArrayStride,
sizeof(uint32_t) * 4);
// 32 loop constants.
id_vector_temp_.push_back(builder_->makeArrayType(
type_uint4_, builder_->makeUintConstant(8), sizeof(uint32_t) * 4));
builder_->addDecoration(id_vector_temp_.back(), spv::DecorationArrayStride,
sizeof(uint32_t) * 4);
spv::Id type_bool_loop_constants =
builder_->makeStructType(id_vector_temp_, "XeBoolLoopConstants");
builder_->addMemberName(type_bool_loop_constants, 0, "bool_constants");
builder_->addMemberDecoration(type_bool_loop_constants, 0,
spv::DecorationOffset, 0);
builder_->addMemberName(type_bool_loop_constants, 1, "loop_constants");
builder_->addMemberDecoration(type_bool_loop_constants, 1,
spv::DecorationOffset, sizeof(uint32_t) * 8);
builder_->addDecoration(type_bool_loop_constants, spv::DecorationBlock);
uniform_bool_loop_constants_ = builder_->createVariable(
spv::NoPrecision, spv::StorageClassUniform, type_bool_loop_constants,
"xe_uniform_bool_loop_constants");
builder_->addDecoration(uniform_bool_loop_constants_,
spv::DecorationDescriptorSet,
int(kDescriptorSetBoolLoopConstants));
builder_->addDecoration(uniform_bool_loop_constants_, spv::DecorationBinding,
0);
if (IsSpirvVertexOrTessEvalShader()) {
StartVertexOrTessEvalShaderBeforeMain();
}
@@ -118,7 +158,7 @@ void SpirvShaderTranslator::StartTranslation() {
}
// Open the main loop.
spv::Block* main_loop_pre_header = builder_->getBuildPoint();
spv::Block& main_loop_pre_header = *builder_->getBuildPoint();
main_loop_header_ = &builder_->makeNewBlock();
spv::Block& main_loop_body = builder_->makeNewBlock();
// Added later because the body has nested control flow, but according to the
@@ -142,7 +182,7 @@ void SpirvShaderTranslator::StartTranslation() {
id_vector_temp_.clear();
id_vector_temp_.reserve(4);
id_vector_temp_.push_back(const_int_0_);
id_vector_temp_.push_back(main_loop_pre_header->getId());
id_vector_temp_.push_back(main_loop_pre_header.getId());
main_loop_pc_next_ = builder_->getUniqueId();
id_vector_temp_.push_back(main_loop_pc_next_);
id_vector_temp_.push_back(main_loop_continue_->getId());
@@ -191,6 +231,8 @@ void SpirvShaderTranslator::StartTranslation() {
}
std::vector<uint8_t> SpirvShaderTranslator::CompleteTranslation() {
// Close flow control within the last switch case.
CloseExecConditionals();
bool has_main_switch = !label_addresses().empty();
// After the final exec (if it happened to be not exece, which would already
// have a break branch), break from the switch if it exists, or from the
@@ -283,6 +325,12 @@ void SpirvShaderTranslator::ProcessLabel(uint32_t cf_index) {
// 0 already added in the beginning.
return;
}
assert_false(label_addresses().empty());
// Close flow control within the previous switch case.
CloseExecConditionals();
spv::Function& function = builder_->getBuildPoint()->getParent();
// Create the next switch case and fallthrough to it.
spv::Block* new_case = new spv::Block(builder_->getUniqueId(), function);
@@ -299,6 +347,57 @@ void SpirvShaderTranslator::ProcessLabel(uint32_t cf_index) {
builder_->setBuildPoint(new_case);
}
void SpirvShaderTranslator::ProcessExecInstructionBegin(
const ParsedExecInstruction& instr) {
UpdateExecConditionals(instr.type, instr.bool_constant_index,
instr.condition);
}
void SpirvShaderTranslator::ProcessExecInstructionEnd(
const ParsedExecInstruction& instr) {
if (instr.is_end) {
// Break out of the main switch (if exists) and the main loop.
CloseInstructionPredication();
if (!builder_->getBuildPoint()->isTerminated()) {
builder_->createBranch(label_addresses().empty() ? main_loop_merge_
: main_switch_merge_);
}
}
UpdateExecConditionals(instr.type, instr.bool_constant_index,
instr.condition);
}
void SpirvShaderTranslator::ProcessJumpInstruction(
const ParsedJumpInstruction& instr) {
// Treat like exec, merge with execs if possible, since it's an if too.
ParsedExecInstruction::Type type;
if (instr.type == ParsedJumpInstruction::Type::kConditional) {
type = ParsedExecInstruction::Type::kConditional;
} else if (instr.type == ParsedJumpInstruction::Type::kPredicated) {
type = ParsedExecInstruction::Type::kPredicated;
} else {
type = ParsedExecInstruction::Type::kUnconditional;
}
UpdateExecConditionals(type, instr.bool_constant_index, instr.condition);
// UpdateExecConditionals may not necessarily close the instruction-level
// predicate check (it's not necessary if the execs are merged), but here the
// instruction itself is on the control flow level, so the predicate check is
// on the control flow level too.
CloseInstructionPredication();
JumpToLabel(instr.target_address);
}
void SpirvShaderTranslator::EnsureBuildPointAvailable() {
if (!builder_->getBuildPoint()->isTerminated()) {
return;
}
spv::Block& new_block = builder_->makeNewBlock();
new_block.setUnreachable();
builder_->setBuildPoint(&new_block);
}
void SpirvShaderTranslator::StartVertexOrTessEvalShaderBeforeMain() {
// Create the inputs.
if (IsSpirvTessEvalShader()) {
@@ -373,5 +472,135 @@ void SpirvShaderTranslator::CompleteVertexOrTessEvalShaderAfterMain(
entry_point->addIdOperand(output_per_vertex_);
}
void SpirvShaderTranslator::UpdateExecConditionals(
ParsedExecInstruction::Type type, uint32_t bool_constant_index,
bool condition) {
// Check if we can merge the new exec with the previous one, or the jump with
// the previous exec. The instruction-level predicate check is also merged in
// this case.
if (type == ParsedExecInstruction::Type::kConditional) {
// Can merge conditional with conditional, as long as the bool constant and
// the expected values are the same.
if (cf_exec_conditional_merge_ &&
cf_exec_bool_constant_or_predicate_ == bool_constant_index &&
cf_exec_condition_ == condition) {
return;
}
} else if (type == ParsedExecInstruction::Type::kPredicated) {
// Can merge predicated with predicated if the conditions are the same and
// the previous exec hasn't modified the predicate register.
if (!cf_exec_predicate_written_ && cf_exec_conditional_merge_ &&
cf_exec_bool_constant_or_predicate_ == kCfExecBoolConstantPredicate &&
cf_exec_condition_ == condition) {
return;
}
} else {
// Can merge unconditional with unconditional.
assert_true(type == ParsedExecInstruction::Type::kUnconditional);
if (!cf_exec_conditional_merge_) {
return;
}
}
CloseExecConditionals();
if (type == ParsedExecInstruction::Type::kUnconditional) {
return;
}
EnsureBuildPointAvailable();
spv::Id condition_id;
if (type == ParsedExecInstruction::Type::kConditional) {
id_vector_temp_.clear();
id_vector_temp_.reserve(3);
// Bool constants (member 0).
id_vector_temp_.push_back(const_int_0_);
// 128-bit vector.
id_vector_temp_.push_back(
builder_->makeIntConstant(int(bool_constant_index >> 7)));
// 32-bit scalar of a 128-bit vector.
id_vector_temp_.push_back(
builder_->makeIntConstant(int((bool_constant_index >> 5) & 2)));
spv::Id bool_constant_scalar =
builder_->createLoad(builder_->createAccessChain(
spv::StorageClassUniform,
uniform_bool_loop_constants_, id_vector_temp_),
spv::NoPrecision);
condition_id = builder_->createBinOp(
spv::OpINotEqual, type_bool_,
builder_->createBinOp(
spv::OpBitwiseAnd, type_uint_, bool_constant_scalar,
builder_->makeUintConstant(uint32_t(1)
<< (bool_constant_index & 31))),
const_uint_0_);
cf_exec_bool_constant_or_predicate_ = bool_constant_index;
} else if (type == ParsedExecInstruction::Type::kPredicated) {
condition_id = builder_->createLoad(var_main_predicate_, spv::NoPrecision);
cf_exec_bool_constant_or_predicate_ = kCfExecBoolConstantPredicate;
} else {
assert_unhandled_case(type);
return;
}
cf_exec_condition_ = condition;
spv::Function& function = builder_->getBuildPoint()->getParent();
cf_exec_conditional_merge_ =
new spv::Block(builder_->getUniqueId(), function);
{
std::unique_ptr<spv::Instruction> selection_merge_op =
std::make_unique<spv::Instruction>(spv::OpSelectionMerge);
selection_merge_op->addIdOperand(cf_exec_conditional_merge_->getId());
selection_merge_op->addImmediateOperand(spv::SelectionControlMaskNone);
builder_->getBuildPoint()->addInstruction(std::move(selection_merge_op));
}
spv::Block& inner_block = builder_->makeNewBlock();
builder_->createConditionalBranch(
condition_id, condition ? &inner_block : cf_exec_conditional_merge_,
condition ? cf_exec_conditional_merge_ : &inner_block);
builder_->setBuildPoint(&inner_block);
}
void SpirvShaderTranslator::CloseInstructionPredication() {
if (!cf_instruction_predicate_merge_) {
return;
}
spv::Block& inner_block = *builder_->getBuildPoint();
if (!inner_block.isTerminated()) {
builder_->createBranch(cf_instruction_predicate_merge_);
}
inner_block.getParent().addBlock(cf_instruction_predicate_merge_);
builder_->setBuildPoint(cf_instruction_predicate_merge_);
cf_instruction_predicate_merge_ = nullptr;
}
void SpirvShaderTranslator::CloseExecConditionals() {
// Within the exec - instruction-level predicate check.
CloseInstructionPredication();
// Exec level.
if (cf_exec_conditional_merge_) {
spv::Block& inner_block = *builder_->getBuildPoint();
if (!inner_block.isTerminated()) {
builder_->createBranch(cf_exec_conditional_merge_);
}
inner_block.getParent().addBlock(cf_exec_conditional_merge_);
builder_->setBuildPoint(cf_exec_conditional_merge_);
cf_exec_conditional_merge_ = nullptr;
}
// Nothing relies on the predicate value being unchanged now.
cf_exec_predicate_written_ = false;
}
void SpirvShaderTranslator::JumpToLabel(uint32_t address) {
assert_false(label_addresses().empty());
spv::Block& origin_block = *builder_->getBuildPoint();
if (origin_block.isTerminated()) {
// Unreachable jump for some reason.
return;
}
main_switch_next_pc_phi_operands_.push_back(
builder_->makeIntConstant(int(address)));
main_switch_next_pc_phi_operands_.push_back(origin_block.getId());
builder_->createBranch(main_loop_continue_);
}
} // namespace gpu
} // namespace xe