SPIRV-Tools/source/opt/invocation_interlock_placement_pass.cpp
Cassandra Beckley 1bc0e6f59a
Add a new legalization pass to dedupe invocation interlock instructions (#5409)
Add a new legalization pass to dedupe invocation interlock instructions

DXC will be adding support for HLSL's rasterizer ordered views by using
the SPV_EXT_fragment_shader_interlock_extension. That extension
stipulates that if an entry point has an interlock ordering execution
mode, it must dynamically execute OpBeginInvocationInterlockEXT and
OpEndInvocationInterlockEXT, in that order, exactly once. This would be
difficult to determine in DXC's SPIR-V backend, so instead we will emit
these instructions potentially multiple times, and use this legalization
pass to ensure that the final SPIR-V follows the specification.

This PR uses data-flow analysis to determine where to place begin and
end instructions; in essence, determining whether a block contains or is
preceded by a begin instruction is similar to a specialized case of a
reaching definitions analysis, where we have only a single definition,
such as `bool has_begun = false`. For this simpler case, we can compute
the set of blocks using BFS to determine the reachability of the begin
instruction.

We need to do this for both begin and end instructions, so I have
generalized portions of the code to run both forward and backward over
the CFG for each respective case.
2023-09-27 19:54:10 -04:00

494 lines
16 KiB
C++

// Copyright (c) 2023 Google Inc.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include "source/opt/invocation_interlock_placement_pass.h"
#include <algorithm>
#include <array>
#include <cassert>
#include <functional>
#include <optional>
#include <queue>
#include <stack>
#include <unordered_map>
#include <unordered_set>
#include <vector>
#include "source/enum_set.h"
#include "source/enum_string_mapping.h"
#include "source/opt/ir_context.h"
#include "source/opt/reflect.h"
#include "source/spirv_target_env.h"
#include "source/util/string_utils.h"
namespace spvtools {
namespace opt {
namespace {
constexpr uint32_t kEntryPointExecutionModelInIdx = 0;
constexpr uint32_t kEntryPointFunctionIdInIdx = 1;
constexpr uint32_t kFunctionCallFunctionIdInIdx = 0;
} // namespace
bool InvocationInterlockPlacementPass::hasSingleNextBlock(uint32_t block_id,
bool reverse_cfg) {
if (reverse_cfg) {
// We are traversing forward, so check whether there is a single successor.
BasicBlock* block = cfg()->block(block_id);
switch (block->tail()->opcode()) {
case spv::Op::OpBranchConditional:
return false;
case spv::Op::OpSwitch:
return block->tail()->NumInOperandWords() == 1;
default:
return !block->tail()->IsReturnOrAbort();
}
} else {
// We are traversing backward, so check whether there is a single
// predecessor.
return cfg()->preds(block_id).size() == 1;
}
}
void InvocationInterlockPlacementPass::forEachNext(
uint32_t block_id, bool reverse_cfg, std::function<void(uint32_t)> f) {
if (reverse_cfg) {
BasicBlock* block = cfg()->block(block_id);
block->ForEachSuccessorLabel([f](uint32_t succ_id) { f(succ_id); });
} else {
for (uint32_t pred_id : cfg()->preds(block_id)) {
f(pred_id);
}
}
}
void InvocationInterlockPlacementPass::addInstructionAtBlockBoundary(
BasicBlock* block, spv::Op opcode, bool at_end) {
if (at_end) {
assert(block->begin()->opcode() != spv::Op::OpPhi &&
"addInstructionAtBlockBoundary expects to be called with at_end == "
"true only if there is a single successor to block");
// Insert a begin instruction at the end of the block.
Instruction* begin_inst = new Instruction(context(), opcode);
begin_inst->InsertAfter(&*--block->tail());
} else {
assert(block->begin()->opcode() != spv::Op::OpPhi &&
"addInstructionAtBlockBoundary expects to be called with at_end == "
"false only if there is a single predecessor to block");
// Insert an end instruction at the beginning of the block.
Instruction* end_inst = new Instruction(context(), opcode);
end_inst->InsertBefore(&*block->begin());
}
}
bool InvocationInterlockPlacementPass::killDuplicateBegin(BasicBlock* block) {
bool found = false;
return context()->KillInstructionIf(
block->begin(), block->end(), [&found](Instruction* inst) {
if (inst->opcode() == spv::Op::OpBeginInvocationInterlockEXT) {
if (found) {
return true;
}
found = true;
}
return false;
});
}
bool InvocationInterlockPlacementPass::killDuplicateEnd(BasicBlock* block) {
std::vector<Instruction*> to_kill;
block->ForEachInst([&to_kill](Instruction* inst) {
if (inst->opcode() == spv::Op::OpEndInvocationInterlockEXT) {
to_kill.push_back(inst);
}
});
if (to_kill.size() <= 1) {
return false;
}
to_kill.pop_back();
for (Instruction* inst : to_kill) {
context()->KillInst(inst);
}
return true;
}
void InvocationInterlockPlacementPass::recordBeginOrEndInFunction(
Function* func) {
if (extracted_functions_.count(func)) {
return;
}
bool had_begin = false;
bool had_end = false;
func->ForEachInst([this, &had_begin, &had_end](Instruction* inst) {
switch (inst->opcode()) {
case spv::Op::OpBeginInvocationInterlockEXT:
had_begin = true;
break;
case spv::Op::OpEndInvocationInterlockEXT:
had_end = true;
break;
case spv::Op::OpFunctionCall: {
uint32_t function_id =
inst->GetSingleWordInOperand(kFunctionCallFunctionIdInIdx);
Function* inner_func = context()->GetFunction(function_id);
recordBeginOrEndInFunction(inner_func);
ExtractionResult result = extracted_functions_[inner_func];
had_begin = had_begin || result.had_begin;
had_end = had_end || result.had_end;
break;
}
default:
break;
}
});
ExtractionResult result = {had_begin, had_end};
extracted_functions_[func] = result;
}
bool InvocationInterlockPlacementPass::
removeBeginAndEndInstructionsFromFunction(Function* func) {
bool modified = false;
func->ForEachInst([this, &modified](Instruction* inst) {
switch (inst->opcode()) {
case spv::Op::OpBeginInvocationInterlockEXT:
context()->KillInst(inst);
modified = true;
break;
case spv::Op::OpEndInvocationInterlockEXT:
context()->KillInst(inst);
modified = true;
break;
default:
break;
}
});
return modified;
}
bool InvocationInterlockPlacementPass::extractInstructionsFromCalls(
std::vector<BasicBlock*> blocks) {
bool modified = false;
for (BasicBlock* block : blocks) {
block->ForEachInst([this, &modified](Instruction* inst) {
if (inst->opcode() == spv::Op::OpFunctionCall) {
uint32_t function_id =
inst->GetSingleWordInOperand(kFunctionCallFunctionIdInIdx);
Function* func = context()->GetFunction(function_id);
ExtractionResult result = extracted_functions_[func];
if (result.had_begin) {
Instruction* new_inst = new Instruction(
context(), spv::Op::OpBeginInvocationInterlockEXT);
new_inst->InsertBefore(inst);
modified = true;
}
if (result.had_end) {
Instruction* new_inst =
new Instruction(context(), spv::Op::OpEndInvocationInterlockEXT);
new_inst->InsertAfter(inst);
modified = true;
}
}
});
}
return modified;
}
void InvocationInterlockPlacementPass::recordExistingBeginAndEndBlock(
std::vector<BasicBlock*> blocks) {
for (BasicBlock* block : blocks) {
block->ForEachInst([this, block](Instruction* inst) {
switch (inst->opcode()) {
case spv::Op::OpBeginInvocationInterlockEXT:
begin_.insert(block->id());
break;
case spv::Op::OpEndInvocationInterlockEXT:
end_.insert(block->id());
break;
default:
break;
}
});
}
}
InvocationInterlockPlacementPass::BlockSet
InvocationInterlockPlacementPass::computeReachableBlocks(
BlockSet& previous_inside, const BlockSet& starting_nodes,
bool reverse_cfg) {
BlockSet inside = starting_nodes;
std::deque<uint32_t> worklist;
worklist.insert(worklist.begin(), starting_nodes.begin(),
starting_nodes.end());
while (!worklist.empty()) {
uint32_t block_id = worklist.front();
worklist.pop_front();
forEachNext(block_id, reverse_cfg,
[&inside, &previous_inside, &worklist](uint32_t next_id) {
previous_inside.insert(next_id);
if (inside.insert(next_id).second) {
worklist.push_back(next_id);
}
});
}
return inside;
}
bool InvocationInterlockPlacementPass::removeUnneededInstructions(
BasicBlock* block) {
bool modified = false;
if (!predecessors_after_begin_.count(block->id()) &&
after_begin_.count(block->id())) {
// None of the previous blocks are in the critical section, but this block
// is. This can only happen if this block already has at least one begin
// instruction. Leave the first begin instruction, and remove any others.
modified |= killDuplicateBegin(block);
} else if (predecessors_after_begin_.count(block->id())) {
// At least one previous block is in the critical section; remove all
// begin instructions in this block.
modified |= context()->KillInstructionIf(
block->begin(), block->end(), [](Instruction* inst) {
return inst->opcode() == spv::Op::OpBeginInvocationInterlockEXT;
});
}
if (!successors_before_end_.count(block->id()) &&
before_end_.count(block->id())) {
// Same as above
modified |= killDuplicateEnd(block);
} else if (successors_before_end_.count(block->id())) {
modified |= context()->KillInstructionIf(
block->begin(), block->end(), [](Instruction* inst) {
return inst->opcode() == spv::Op::OpEndInvocationInterlockEXT;
});
}
return modified;
}
BasicBlock* InvocationInterlockPlacementPass::splitEdge(BasicBlock* block,
uint32_t succ_id) {
// Create a new block to replace the critical edge.
auto new_succ_temp = MakeUnique<BasicBlock>(
MakeUnique<Instruction>(context(), spv::Op::OpLabel, 0, TakeNextId(),
std::initializer_list<Operand>{}));
auto* new_succ = new_succ_temp.get();
// Insert the new block into the function.
block->GetParent()->InsertBasicBlockAfter(std::move(new_succ_temp), block);
new_succ->AddInstruction(MakeUnique<Instruction>(
context(), spv::Op::OpBranch, 0, 0,
std::initializer_list<Operand>{
Operand(spv_operand_type_t::SPV_OPERAND_TYPE_ID, {succ_id})}));
assert(block->tail()->opcode() == spv::Op::OpBranchConditional ||
block->tail()->opcode() == spv::Op::OpSwitch);
// Update the first branch to successor to instead branch to
// the new successor. If there are multiple edges, we arbitrarily choose the
// first time it appears in the list. The other edges to `succ_id` will have
// to be split by another call to `splitEdge`.
block->tail()->WhileEachInId([new_succ, succ_id](uint32_t* branch_id) {
if (*branch_id == succ_id) {
*branch_id = new_succ->id();
return false;
}
return true;
});
return new_succ;
}
bool InvocationInterlockPlacementPass::placeInstructionsForEdge(
BasicBlock* block, uint32_t next_id, BlockSet& inside,
BlockSet& previous_inside, spv::Op opcode, bool reverse_cfg) {
bool modified = false;
if (previous_inside.count(next_id) && !inside.count(block->id())) {
// This block is not in the critical section but the next has at least one
// other previous block that is, so this block should be enter it as well.
// We need to add begin or end instructions to the edge.
modified = true;
if (hasSingleNextBlock(block->id(), reverse_cfg)) {
// This is the only next block.
// Additionally, because `next_id` is in `previous_inside`, we know that
// `next_id` has at least one previous block in `inside`. And because
// 'block` is not in `inside`, that means the `next_id` has to have at
// least one other previous block in `inside`.
// This is solely for a debug assertion. It is essentially recomputing the
// value of `previous_inside` to verify that it was computed correctly
// such that the above statement is true.
bool next_has_previous_inside = false;
// By passing !reverse_cfg to forEachNext, we are actually iterating over
// the previous blocks.
forEachNext(next_id, !reverse_cfg,
[&next_has_previous_inside, inside](uint32_t previous_id) {
if (inside.count(previous_id)) {
next_has_previous_inside = true;
}
});
assert(next_has_previous_inside &&
"`previous_inside` must be the set of blocks with at least one "
"previous block in `inside`");
addInstructionAtBlockBoundary(block, opcode, reverse_cfg);
} else {
// This block has multiple next blocks. Split the edge and insert the
// instruction in the new next block.
BasicBlock* new_branch;
if (reverse_cfg) {
new_branch = splitEdge(block, next_id);
} else {
new_branch = splitEdge(cfg()->block(next_id), block->id());
}
auto inst = new Instruction(context(), opcode);
inst->InsertBefore(&*new_branch->tail());
}
}
return modified;
}
bool InvocationInterlockPlacementPass::placeInstructions(BasicBlock* block) {
bool modified = false;
block->ForEachSuccessorLabel([this, block, &modified](uint32_t succ_id) {
modified |= placeInstructionsForEdge(
block, succ_id, after_begin_, predecessors_after_begin_,
spv::Op::OpBeginInvocationInterlockEXT, /* reverse_cfg= */ true);
modified |= placeInstructionsForEdge(cfg()->block(succ_id), block->id(),
before_end_, successors_before_end_,
spv::Op::OpEndInvocationInterlockEXT,
/* reverse_cfg= */ false);
});
return modified;
}
bool InvocationInterlockPlacementPass::processFragmentShaderEntry(
Function* entry_func) {
bool modified = false;
// Save the original order of blocks in the function, so we don't iterate over
// newly-added blocks.
std::vector<BasicBlock*> original_blocks;
for (auto bi = entry_func->begin(); bi != entry_func->end(); ++bi) {
original_blocks.push_back(&*bi);
}
modified |= extractInstructionsFromCalls(original_blocks);
recordExistingBeginAndEndBlock(original_blocks);
after_begin_ = computeReachableBlocks(predecessors_after_begin_, begin_,
/* reverse_cfg= */ true);
before_end_ = computeReachableBlocks(successors_before_end_, end_,
/* reverse_cfg= */ false);
for (BasicBlock* block : original_blocks) {
modified |= removeUnneededInstructions(block);
modified |= placeInstructions(block);
}
return modified;
}
bool InvocationInterlockPlacementPass::isFragmentShaderInterlockEnabled() {
if (!context()->get_feature_mgr()->HasExtension(
kSPV_EXT_fragment_shader_interlock)) {
return false;
}
if (context()->get_feature_mgr()->HasCapability(
spv::Capability::FragmentShaderSampleInterlockEXT)) {
return true;
}
if (context()->get_feature_mgr()->HasCapability(
spv::Capability::FragmentShaderPixelInterlockEXT)) {
return true;
}
if (context()->get_feature_mgr()->HasCapability(
spv::Capability::FragmentShaderShadingRateInterlockEXT)) {
return true;
}
return false;
}
Pass::Status InvocationInterlockPlacementPass::Process() {
// Skip this pass if the necessary extension or capability is missing
if (!isFragmentShaderInterlockEnabled()) {
return Status::SuccessWithoutChange;
}
bool modified = false;
std::unordered_set<Function*> entry_points;
for (Instruction& entry_inst : context()->module()->entry_points()) {
uint32_t entry_id =
entry_inst.GetSingleWordInOperand(kEntryPointFunctionIdInIdx);
entry_points.insert(context()->GetFunction(entry_id));
}
for (auto fi = context()->module()->begin(); fi != context()->module()->end();
++fi) {
Function* func = &*fi;
recordBeginOrEndInFunction(func);
if (!entry_points.count(func) && extracted_functions_.count(func)) {
modified |= removeBeginAndEndInstructionsFromFunction(func);
}
}
for (Instruction& entry_inst : context()->module()->entry_points()) {
uint32_t entry_id =
entry_inst.GetSingleWordInOperand(kEntryPointFunctionIdInIdx);
Function* entry_func = context()->GetFunction(entry_id);
auto execution_model = spv::ExecutionModel(
entry_inst.GetSingleWordInOperand(kEntryPointExecutionModelInIdx));
if (execution_model != spv::ExecutionModel::Fragment) {
continue;
}
modified |= processFragmentShaderEntry(entry_func);
}
return modified ? Pass::Status::SuccessWithChange
: Pass::Status::SuccessWithoutChange;
}
} // namespace opt
} // namespace spvtools