Bind the Gap: Compiling Real Software to Hardware FFT Accelerators
summary
Paper claims a compiler pipeline that automatically maps real C FFT implementations to API‑driven FFT accelerators by resolving four concrete mismatches: code, data, domain, and behavior. It uses neural embeddings to find FFT regions, program synthesis (sketching) to generate data bindings and behavioral conversions, plus range checks to gate offloads. The tool compiles ~25 GitHub FFTs to FFTW, NXP PowerQuad, and FFTa and reports significant speedups, though tiny transforms can slow down due to copy overhead.
AI-written summary · about
chapters (12)
- 0:00The Accelerator Era
- 1:33API-Programmable Accelerators
- 2:44FFT Accelerators
- 4:16Inflexible Chips: The Downside
- 6:14Why so few clones?: Code Mismatch
- 6:49Why so few clones?: Data Mismatch
- 7:54Why so few clones?: Domain Mismatch
- 8:52Why so few clones?: Behaviour Mismatch
- 12:01Binding Synthesis
- 12:50Range-Check Generation
- 13:34Behavioural Synthesis
- 14:02Benchmarks
related talks
24mMapping Sparse Applications to Accelerated Computing Systems
17mTowards Semantics Lifting for Scientific Computing: A Case Study on FFT
20mA Tensor Compiler with Automatic Data Packing for Simple and Efficient Fully Homomorphic(…)
43mSoftware Development Trends for FPGAs
19mSSFFT: Energy-Efficient Selective Scaling for Fast Fourier Transform in Embedded GPUs
01h 00mBeating Hand Tuned Assembly in Compiled Languages
53mLocknote: How Badly Do We Want Correct Compilers?
01h 00mBringing Clang and C++ to GPUs: An Open-Source, CUDA-Compatible GPU C++ Compiler