File size: 11,077 Bytes
67e6b48 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 | #!/usr/bin/env python3
"""
Test suite for Twill: Optimal Software Pipelining and Warp Specialization.
Reproduces the key results from the paper:
1. Simplified FA schedule (Figure 1) - GEMM and EXP have equal cost
2. Full FMHA forward on Hopper (Section 6.2.1) - rediscovers FA3 pipeline
3. Full FMHA forward on Blackwell (Section 6.2.2) - rediscovers FA4 strategy
4. Simple GEMM pipeline - load-compute overlap
"""
import sys
import time
import numpy as np
sys.path.insert(0, '/app')
from twill.graph import (
DependenceGraph, Instruction, DependenceEdge,
hopper_machine, blackwell_machine, make_rrt,
)
from twill.cost_normalization import normalize_costs
from twill.modulo_scheduler import (
optimal_modulo_schedule, compute_modular_rrt, validate_schedule,
)
from twill.smt_joint import swp_and_ws
from twill.twill_solver import twill_solve
from twill.codegen import generate_pipelined_code, generate_pseudocode, generate_cuda_skeleton
from twill.visualization import visualize_schedule, print_modular_rrt
from twill.kernels import (
flash_attention_forward_simplified,
flash_attention_forward_hopper,
flash_attention_forward_blackwell,
simple_gemm_pipeline,
)
def test_cost_normalization():
"""Test cost normalization (Section 5.2)."""
print("\n" + "=" * 70)
print("TEST: Cost Normalization (Section 5.2)")
print("=" * 70)
# Example from the paper: GEMM ~1000 cycles on Hopper
costs = {"WGMMA": 1000, "EXP": 1000, "TMA": 500}
normalized, F = normalize_costs(costs, U=10)
print(f"Original: {costs}")
print(f"Normalized: {normalized} (F={F})")
# Check ratio preservation
for k1 in costs:
for k2 in costs:
if costs[k2] > 0 and normalized[k2] > 0:
orig_ratio = costs[k1] / costs[k2]
norm_ratio = normalized[k1] / normalized[k2]
print(f" Ratio {k1}/{k2}: orig={orig_ratio:.3f}, norm={norm_ratio:.3f}")
# Larger example
costs2 = {"WGMMA": 1024, "EXP": 960, "TMA_K": 512, "TMA_V": 512, "TMEM_R": 128, "TMEM_W": 128}
normalized2, F2 = normalize_costs(costs2, U=20)
print(f"\nOriginal: {costs2}")
print(f"Normalized: {normalized2} (F={F2})")
print("\nβ Cost normalization test passed")
return True
def test_simplified_flash_attention():
"""Test the simplified FA example from Figure 1.
Expected result from the paper:
- I = 2 (one GEMM and one EXP per 2 cycles)
- L = 4 (schedule length)
- ceil(L/I) = 2 copies overlapped
- S is extracted into prologue
"""
print("\n" + "=" * 70)
print("TEST: Simplified Flash Attention (Figure 1)")
print("=" * 70)
graph = flash_attention_forward_simplified()
print(f"Graph: {graph}")
print(f"Min I (resource bound): {graph.compute_min_initiation_interval()}")
# Phase 1: Modulo scheduling
print("\n--- Phase 1: Modulo Scheduling ---")
schedule = optimal_modulo_schedule(graph, target_I=2, verbose=False)
if schedule is None:
print("Phase 1 FAILED for I=2, trying I=3...")
schedule = optimal_modulo_schedule(graph, target_I=3, verbose=False)
assert schedule is not None, "No modulo schedule found!"
print(f"Schedule: {schedule}")
valid, violations = validate_schedule(graph, schedule)
print(f"Valid: {valid}")
if not valid:
for v in violations:
print(f" VIOLATION: {v}")
print(f"\n{print_modular_rrt(graph, schedule)}")
# Full Twill solve
print("\n--- Full Twill Solve ---")
result = twill_solve(
graph,
max_I=5,
enable_cost_normalization=False, # Costs already normalized (1 cycle each)
enable_memory_constraints=False, # Simplified example
enable_warp_constraints=True,
verbose=True,
)
if result:
print(f"\n{result}")
# Visualization
viz = visualize_schedule(graph, result.joint_result,
output_path="/app/simplified_fa_schedule.png",
title="Simplified Flash Attention (Figure 1)")
print(viz)
# Code generation
pseudo = generate_pseudocode(graph, result.joint_result)
print(f"\n--- Generated Pseudocode ---")
print(pseudo)
# CUDA skeleton
cuda = generate_cuda_skeleton(graph, result.joint_result)
print(f"\n--- Generated CUDA Skeleton ---")
print(cuda)
return True
else:
print("FAILED: No solution found!")
return False
def test_hopper_fmha_forward():
"""Test full FMHA forward on Hopper (Section 6.2.1).
Expected: Twill rediscovers FA3's SWP strategy:
- S=gemm(Q,K[0]) extracted into prologue
- Ping-pong scheduling between EXP and TC
- TMA loads on variable-latency (producer) warp
"""
print("\n" + "=" * 70)
print("TEST: FMHA Forward on Hopper (Section 6.2.1)")
print("=" * 70)
graph = flash_attention_forward_hopper()
print(f"Graph: {graph}")
print(f"Instructions: {[(v.name, 'var_lat' if v.variable_latency else 'fixed') for v in graph.V]}")
print(f"Min I: {graph.compute_min_initiation_interval()}")
result = twill_solve(
graph,
max_I=10,
enable_cost_normalization=False,
enable_memory_constraints=False,
enable_warp_constraints=True,
verbose=True,
)
if result:
print(f"\n{result}")
viz = visualize_schedule(graph, result.joint_result,
output_path="/app/hopper_fmha_schedule.png",
title="FMHA Forward on Hopper")
print(viz)
pseudo = generate_pseudocode(graph, result.joint_result)
print(f"\n--- Generated Pseudocode ---")
print(pseudo)
# Verify key properties:
wa = result.warp_assignment
print(f"\n--- Verification ---")
# TMA loads should be on variable-latency warp (W_vl = 0)
k_warp = wa.warp_of("K_load")
v_warp = wa.warp_of("V_load")
print(f"K_load on warp {k_warp} (expected: 0/variable-latency)")
print(f"V_load on warp {v_warp} (expected: 0/variable-latency)")
# TC operations should be on compute warps (not warp 0)
s_warp = wa.warp_of("S")
o_warp = wa.warp_of("O")
print(f"S (GEMM) on warp {s_warp} (expected: compute warp)")
print(f"O (GEMM) on warp {o_warp} (expected: compute warp)")
# EXP should be on a compute warp
p_warp = wa.warp_of("P")
print(f"P (EXP) on warp {p_warp} (expected: compute warp)")
return True
else:
print("FAILED: No solution found!")
return False
def test_blackwell_fmha_forward():
"""Test FMHA forward on Blackwell (Section 6.2.2).
Expected: Twill rediscovers FA4's strategy:
- Variable-latency ops (TMA) on separate warp
- TC GEMMs on compute warps
- TMEM read/write for accumulator management
- Different pipeline structure than Hopper
"""
print("\n" + "=" * 70)
print("TEST: FMHA Forward on Blackwell (Section 6.2.2)")
print("=" * 70)
graph = flash_attention_forward_blackwell()
print(f"Graph: {graph}")
print(f"Instructions: {[(v.name, 'var_lat' if v.variable_latency else 'fixed') for v in graph.V]}")
print(f"Min I: {graph.compute_min_initiation_interval()}")
result = twill_solve(
graph,
max_I=15,
enable_cost_normalization=False,
enable_memory_constraints=False,
enable_warp_constraints=True,
verbose=True,
)
if result:
print(f"\n{result}")
viz = visualize_schedule(graph, result.joint_result,
output_path="/app/blackwell_fmha_schedule.png",
title="FMHA Forward on Blackwell")
print(viz)
pseudo = generate_pseudocode(graph, result.joint_result)
print(f"\n--- Generated Pseudocode ---")
print(pseudo)
cuda = generate_cuda_skeleton(graph, result.joint_result)
print(f"\n--- Generated CUDA Skeleton ---")
print(cuda)
return True
else:
print("FAILED: No solution found!")
return False
def test_simple_gemm():
"""Test simple GEMM with load-compute overlap."""
print("\n" + "=" * 70)
print("TEST: Simple GEMM Pipeline")
print("=" * 70)
graph = simple_gemm_pipeline()
print(f"Graph: {graph}")
result = twill_solve(
graph,
max_I=6,
enable_cost_normalization=False,
enable_memory_constraints=False,
enable_warp_constraints=True,
verbose=True,
)
if result:
print(f"\n{result}")
viz = visualize_schedule(graph, result.joint_result,
title="Simple GEMM Pipeline")
print(viz)
return True
else:
print("FAILED: No solution found!")
return False
def test_modulo_scheduling_only():
"""Test Phase 1 (modulo scheduling) in isolation."""
print("\n" + "=" * 70)
print("TEST: Modulo Scheduling Only (Phase 1)")
print("=" * 70)
graph = flash_attention_forward_simplified()
# Try increasing I values
for target_I in range(1, 6):
print(f"\n Trying I = {target_I}...")
sched = optimal_modulo_schedule(graph, target_I=target_I, verbose=False)
if sched:
valid, violations = validate_schedule(graph, sched)
print(f" Found: M={sched.schedule}, L={sched.length}, copies={sched.num_copies}")
print(f" Valid: {valid}")
if not valid:
for v in violations:
print(f" {v}")
else:
print(f" Infeasible")
return True
if __name__ == "__main__":
print("β" + "β" * 68 + "β")
print("β" + " Twill Test Suite ".center(68) + "β")
print("β" + " Optimal SWP + WS for Tensor Core GPUs ".center(68) + "β")
print("β" + " (arXiv:2512.18134) ".center(68) + "β")
print("β" + "β" * 68 + "β")
results = {}
start = time.time()
# Run tests
results["Cost Normalization"] = test_cost_normalization()
results["Modulo Scheduling Only"] = test_modulo_scheduling_only()
results["Simplified FA (Figure 1)"] = test_simplified_flash_attention()
results["Simple GEMM"] = test_simple_gemm()
results["Hopper FMHA Forward"] = test_hopper_fmha_forward()
results["Blackwell FMHA Forward"] = test_blackwell_fmha_forward()
elapsed = time.time() - start
# Summary
print("\n" + "=" * 70)
print("TEST SUMMARY")
print("=" * 70)
for name, passed in results.items():
status = "β PASS" if passed else "β FAIL"
print(f" {status} {name}")
print(f"\nTotal time: {elapsed:.2f}s")
print(f"Passed: {sum(results.values())}/{len(results)}")
sys.exit(0 if all(results.values()) else 1)
|