Upload twill/__init__.py
Browse files- twill/__init__.py +26 -0
twill/__init__.py
ADDED
|
@@ -0,0 +1,26 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""
|
| 2 |
+
Twill: Optimal Software Pipelining and Warp Specialization for Tensor Core GPUs
|
| 3 |
+
|
| 4 |
+
Implementation of the paper by Rupanshu Soi et al. (arXiv:2512.18134)
|
| 5 |
+
|
| 6 |
+
Twill formulates the joint SWP + WS optimization as:
|
| 7 |
+
Phase 1: ZLP-based Optimal Modulo Scheduling (CBC solver via PuLP)
|
| 8 |
+
Phase 2: SMT-based Joint SWP + WS (Z3 solver)
|
| 9 |
+
|
| 10 |
+
With cost normalization to make cycle counts tractable.
|
| 11 |
+
"""
|
| 12 |
+
|
| 13 |
+
from twill.graph import DependenceGraph, Instruction, DependenceEdge, MachineDescription
|
| 14 |
+
from twill.cost_normalization import normalize_costs
|
| 15 |
+
from twill.modulo_scheduler import optimal_modulo_schedule
|
| 16 |
+
from twill.smt_joint import swp_and_ws
|
| 17 |
+
from twill.twill_solver import twill_solve
|
| 18 |
+
from twill.codegen import generate_pipelined_code
|
| 19 |
+
from twill.visualization import visualize_schedule
|
| 20 |
+
|
| 21 |
+
__version__ = "0.1.0"
|
| 22 |
+
__all__ = [
|
| 23 |
+
"DependenceGraph", "Instruction", "DependenceEdge", "MachineDescription",
|
| 24 |
+
"normalize_costs", "optimal_modulo_schedule", "swp_and_ws",
|
| 25 |
+
"twill_solve", "generate_pipelined_code", "visualize_schedule",
|
| 26 |
+
]
|