Polyhedral Compiler Static Versioning for Dynamic Tensors
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing polyhedral compilation techniques struggle to effectively optimize loop-based programs with dynamic tensor sizes, as run-time tensor sizes are unknown at compilation time, limiting the mapper's ability to make informed mapping decisions.
Innovation Solution
The approach involves versioning the source code at compile-time by generating multiple versions of the internal representation (IR) with distinct contexts based on potential run-time parameter values, allowing the polyhedral compiler to explore tradeoffs between parallelization, data-locality, and data-layout transformation, and using affine approximation or artificial neural networks for linearization of non-linear constraints.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If multiple versions of IR with distinct contexts are generated to explore tradeoffs between parallelization and data locality, then optimization quality improves, but device complexity increases
Solution Approach 1:
The patent segments the compilation process by creating multiple versions of the internal representation (IR) with distinct contexts, where each version explores different optimization tradeoffs. This segmentation allows the system to evaluate parallelization versus data locality separately for each version, improving overall optimization quality while managing complexity through structured division of the compilation task.
Solution Approach 2:
The patent introduces dynamic version selection at runtime based on actual parameter values. The wrapper function dynamically chooses which versioned IR to execute by evaluating runtime parameters against context constraints, allowing the system to adapt to different execution scenarios and achieve optimal performance without committing to a single static optimization strategy.
2Adaptability or versatility
If polyhedral compilation is performed at runtime for new parameter instances, then adaptability improves, but productivity decreases
Solution Approach 1:
The patent performs preliminary polyhedral compilation at compile time for multiple potential parameter values, creating versioned IRs in advance. This preliminary action stores pre-compiled versions that can be quickly selected at runtime based on actual parameter values, avoiding the need to perform full polyhedral compilation during runtime execution.
Solution Approach 2:
The patent creates multiple copies of the IR with different context constraints, where each copy represents a pre-compiled version for specific parameter ranges. At runtime, the system copies or selects the appropriate pre-compiled version based on actual parameters, avoiding repeated compilation work while maintaining adaptability to different parameter instances.
3Ease of manufacture
If non-linear constraints are linearized using affine approximation or neural networks, then ease of manufacture improves, but measurement precision decreases
Solution Approach 1:
The patent transforms non-linear constraints into linear forms through parameter changes, using either affine approximation (changing the mathematical representation) or neural network-based linearization (changing how constraints are expressed). This transformation makes the constraints easier to process by polyhedral compilation tools while accepting a tradeoff in precision, allowing the system to handle complex runtime parameters more efficiently.
Data Source
AI summary
An approach is presented to enhancing the optimization process in a polyhedral compiler by introducing compile-time versioning, i.e., the production of several versions of optimized code under varying assumptions on its run-time parameters. We illustrate this process by enabling versioning in the polyhedral processor placement pass. We propose an efficient code generation method and validate that versioning can be useful in a polyhedral compiler by performing benchmarking on a small set of deep learning layers defined for dynamically-sized tensors.