// Copyright (c) 2021 Google LLC. // // 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. #ifndef SOURCE_LINT_DIVERGENCE_ANALYSIS_H_ #define SOURCE_LINT_DIVERGENCE_ANALYSIS_H_ #include #include #include #include "source/opt/basic_block.h" #include "source/opt/control_dependence.h" #include "source/opt/dataflow.h" #include "source/opt/function.h" #include "source/opt/instruction.h" namespace spvtools { namespace lint { // Computes the static divergence level for blocks (control flow) and values. // // A value is uniform if all threads that execute it are guaranteed to have the // same value. Similarly, a value is partially uniform if this is true only // within each derivative group. If neither apply, it is divergent. // // Control flow through a block is uniform if for any possible execution and // point in time, all threads are executing it, or no threads are executing it. // In particular, it is never possible for some threads to be inside the block // and some threads not executing. // TODO(kuhar): Clarify the difference between uniform, divergent, and // partially-uniform execution in this analysis. // // Caveat: // As we use control dependence to determine how divergence is propagated, this // analysis can be overly permissive when the merge block for a conditional // branch or switch is later than (strictly postdominates) the expected merge // block, which is the immediate postdominator. However, this is not expected to // be a problem in practice, given that SPIR-V is generally output by compilers // and other automated tools, which would assign the earliest possible merge // block, rather than written by hand. // TODO(kuhar): Handle late merges. class DivergenceAnalysis : public opt::ForwardDataFlowAnalysis { public: // The tightest (most uniform) level of divergence that can be determined // statically for a value or control flow for a block. // // The values are ordered such that A > B means that A is potentially more // divergent than B. // TODO(kuhar): Rename |PartiallyUniform' to something less confusing. For // example, the enum could be based on scopes. enum class DivergenceLevel { // The value or control flow is uniform across the entire invocation group. kUniform = 0, // The value or control flow is uniform across the derivative group, but not // the invocation group. kPartiallyUniform = 1, // The value or control flow is not statically uniform. kDivergent = 2, }; DivergenceAnalysis(opt::IRContext& context) : ForwardDataFlowAnalysis(context, LabelPosition::kLabelsAtEnd) {} // Returns the divergence level for the given value (non-label instructions), // or control flow for the given block. DivergenceLevel GetDivergenceLevel(uint32_t id) { auto it = divergence_.find(id); if (it == divergence_.end()) { return DivergenceLevel::kUniform; } return it->second; } // Returns the divergence source for the given id. The following types of // divergence flows from A to B are possible: // // data -> data: A is used as an operand in the definition of B. // data -> control: B is control-dependent on a branch with condition A. // control -> data: B is a OpPhi instruction in which A is a block operand. // control -> control: B is control-dependent on A. uint32_t GetDivergenceSource(uint32_t id) { auto it = divergence_source_.find(id); if (it == divergence_source_.end()) { return 0; } return it->second; } // Returns the dependence source for the control dependence for the given id. // This only exists for data -> control edges. // // In other words, if block 2 is dependent on block 1 due to value 3 (e.g. // block 1 terminates with OpBranchConditional %3 %2 %4): // * GetDivergenceSource(2) = 3 // * GetDivergenceDependenceSource(2) = 1 // // Returns 0 if not applicable. uint32_t GetDivergenceDependenceSource(uint32_t id) { auto it = divergence_dependence_source_.find(id); if (it == divergence_dependence_source_.end()) { return 0; } return it->second; } void InitializeWorklist(opt::Function* function, bool is_first_iteration) override { // Since |EnqueueSuccessors| is complete, we only need one pass. if (is_first_iteration) { Setup(function); opt::ForwardDataFlowAnalysis::InitializeWorklist(function, true); } } void EnqueueSuccessors(opt::Instruction* inst) override; VisitResult Visit(opt::Instruction* inst) override; private: VisitResult VisitBlock(uint32_t id); VisitResult VisitInstruction(opt::Instruction* inst); // Computes the divergence level for the result of the given instruction // based on the current state of the analysis. This is always an // underapproximation, which will be improved as the analysis proceeds. DivergenceLevel ComputeInstructionDivergence(opt::Instruction* inst); // Computes the divergence level for a variable, which is used for loads. DivergenceLevel ComputeVariableDivergence(opt::Instruction* var); // Initializes data structures for performing dataflow on the given function. void Setup(opt::Function* function); std::unordered_map divergence_; std::unordered_map divergence_source_; std::unordered_map divergence_dependence_source_; // Stores the result of following unconditional branches starting from the // given block. This is used to detect when reconvergence needs to be // accounted for. std::unordered_map follow_unconditional_branches_; opt::ControlDependenceAnalysis cd_; }; std::ostream& operator<<(std::ostream& os, DivergenceAnalysis::DivergenceLevel level); } // namespace lint } // namespace spvtools #endif // SOURCE_LINT_DIVERGENCE_ANALYSIS_H_