| """ |
| Twill: Optimal Software Pipelining and Warp Specialization for Tensor Core GPUs |
| |
| Implementation of the paper by Rupanshu Soi et al. (arXiv:2512.18134) |
| with GauS differentiable solver by Yaohui Cai et al. (arXiv:2602.20427) |
| |
| Twill formulates the joint SWP + WS optimization as: |
| Phase 1: ZLP-based Optimal Modulo Scheduling (CBC solver via PuLP) |
| Phase 2: SMT-based Joint SWP + WS (Z3 solver) |
| |
| GauS provides a scalable differentiable alternative using: |
| Gaussian reparameterization + Augmented Lagrangian Method (ALM) |
| |
| With cost normalization to make cycle counts tractable. |
| """ |
|
|
| from twill.graph import DependenceGraph, Instruction, DependenceEdge, MachineDescription |
| from twill.cost_normalization import normalize_costs |
| from twill.modulo_scheduler import optimal_modulo_schedule |
| from twill.smt_joint import swp_and_ws |
| from twill.twill_solver import twill_solve |
| from twill.codegen import generate_pipelined_code |
| from twill.visualization import visualize_schedule |
| from twill.gaus_solver import GauSSolver, GausGraph, gaus_solve_twill_graph |
|
|
| __version__ = "0.2.0" |
| __all__ = [ |
| "DependenceGraph", "Instruction", "DependenceEdge", "MachineDescription", |
| "normalize_costs", "optimal_modulo_schedule", "swp_and_ws", |
| "twill_solve", "generate_pipelined_code", "visualize_schedule", |
| "GauSSolver", "GausGraph", "gaus_solve_twill_graph", |
| ] |
|
|
