Auto-vectorization of Control-Flow Divergent Loops
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing vectorization techniques struggle with loops containing control-flow divergent branch conditions, leading to inefficient code generation and underutilization of SIMD units due to over-pessimistic assumptions and high overhead from software predication.
Innovation Solution
A computer-implemented method that performs compile-time analysis under the assumption of dynamic uniformity, determines a vectorizable access pattern, and includes run-time checks to verify dynamically uniform conditions, enabling efficient vectorization of loops with control-dependent cross-iteration dependencies.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If existing vectorization techniques are used on loops with control-flow divergent branch conditions, then compilation is simpler, but SIMD unit utilization deteriorates due to over-pessimistic assumptions and software predication overhead
Solution Approach 1:
The patent performs compile-time analysis to identify loops with control-dependent cross-iteration dependencies and determines their vectorizability in advance. By analyzing the loop structure and dependencies before code generation, the compiler can make informed decisions about vectorization without requiring runtime software predication, thus improving SIMD unit utilization while maintaining compilation simplicity.
Solution Approach 2:
The patent introduces an intermediate analysis phase that acts as a mediator between the source code and the vectorized code. This intermediate analysis determines the vectorizable access patterns and generates appropriate vectorized code directly, eliminating the need for runtime software predication mechanisms while preserving the benefits of vectorization.
2Reliability
If software predication is used to handle control flow in vectorized loops, then correctness is improved, but overhead increases reducing execution efficiency
Solution Approach 1:
The patent extracts the control flow handling logic from runtime software predication and moves it to compile-time analysis. By determining the vectorizable access patterns and handling control dependencies during compilation, the need for runtime predication overhead is eliminated while maintaining correctness through static analysis of the control flow structure.
3Productivity
If loops with control-dependent cross-iteration dependencies are vectorized, then productivity is improved, but reliability deteriorates due to incorrect dependency handling
Solution Approach 1:
The patent performs preliminary analysis of control-dependent cross-iteration dependencies at compile time to determine whether the loop can be safely vectorized. By analyzing the dependency structure before code generation, the compiler can identify vectorizable patterns and generate correct vectorized code without compromising reliability, while still achieving the productivity benefits of vectorization.
Data Source
AI summary
Vectorization of program code by way of a method that includes obtaining program code to be compiled. The program code has a loop that includes a control-flow having divergent branch conditions. The method includes compiling the program code to produce compiled code. The compiling includes performing compile-time analysis of the loop based on an assumption of dynamic uniformity, and determining from the analysis a vectorizable access pattern of the loop. The compiling also includes vectorizing the loop as part of the compiling the program code, including compiling the loop. The vectorizing includes providing run-time checks in the compiled code. The run-time checks are configured for checking for dynamically uniform conditions for vector processing at run-time to control program execution flow based on a result of the checking. The method also includes outputting the compiled code for execution.