Hardware-optimized symmetric high-performance automated parallelization system with loop-level parallelization and method thereof
The compiler system addresses inefficiencies in multiprocessor code parallelization by partitioning code into compute block nodes and using matrices/tensors to optimize latency and parallelism, achieving improved throughput and hardware adaptation.
Patent Information
- Application Number
- JP2025562086
- Authority / Receiving Office
- JP · JP
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2023-04-24
- Publication Date
- 2026-05-26
AI Technical Summary
Conventional compilers struggle to efficiently parallelize code for multiprocessor systems, particularly in managing latency and data dependencies, leading to suboptimal performance in modern computing architectures with increasing core counts and heterogeneous hardware.
A compiler system that automatically optimizes code for multiprocessor systems by partitioning it into compute block nodes, using matrices and tensors to minimize latency and maximize parallelism across multiple processing units, incorporating a parser, matrix builder, and optimizer to generate optimized machine code.
Enhances throughput by efficiently managing processor and data dependencies, optimizing low-level parallelism, and adapting to hardware-specific requirements, overcoming limitations of prior art systems.