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 | class Program:
def __init__(self, eqs: str, *, parser: str = "legacy"):
self.src = eqs
if parser == "legacy":
self.ir: ProgramIR = parse(eqs) # type: ignore[misc]
elif parser in {"v2", "expr"}:
from .parser_expr import parse_program as _parse_program_v2
from .macro_rewrite import expand_macros as _expand_macros
from .ast_lowering import lower_to_ir as _lower_to_ir
ast_prog = _parse_program_v2(eqs)
ast_prog = _expand_macros(ast_prog)
self.ir = _lower_to_ir(ast_prog)
else:
raise ValueError("Unknown parser mode; use 'legacy' or 'v2'")
validate_program_shapes(self.ir)
self.digest = compute_program_hash(self.src)
def compile(
self,
backend: str = "auto",
device: str = "auto",
cache_dir: Optional[str] = None,
config: Optional[ExecutionConfig] = None,
execution: Optional[str] = None,
policies: Optional[RuntimePolicies] = None,
**backend_kwargs,
):
"""Compile the program to a runnable for the selected backend.
When ``backend`` is ``"auto"`` (default), choose a backend based on:
- Execution mode and projection strategy (NumPy for demand/Monte Carlo).
- Streaming usage (NumPy when streaming is present).
- Hardware availability and requested device.
- A light heuristic over the IR to detect DL-like workloads (einsum-heavy,
attention/MLP ops) where Torch/JAX are preferable.
The selection prefers Torch on CUDA/MPS for attention/MLP-style programs,
otherwise tries JAX when JIT cost can amortize. Small/streaming workloads
use NumPy to avoid dispatch/JIT overheads.
"""
cfg = config or ExecutionConfig()
if execution is not None:
cfg = replace(cfg, mode=execution)
if device != "auto":
cfg = replace(cfg, device=device)
cfg = cfg.normalized()
target_device = cfg.device
policy_obj = policies or RuntimePolicies()
cache_manager = CacheManager(cache_dir) if cache_dir else None
# Auto backend selection -------------------------------------------------
if backend == "auto":
backend = self._choose_backend_auto(cfg)
if backend == "numpy":
if cfg.device not in {"auto", "cpu"}:
raise ValueError(
f"NumPy backend only supports CPU execution; received device='{cfg.device}'"
)
if cfg.precision == "auto":
cfg = replace(cfg, precision="fp32")
elif cfg.precision != "fp32":
raise ValueError(
f"NumPy backend only supports fp32 precision; received precision='{cfg.precision}'"
)
if cfg.mode == "demand":
runner = DemandNumpyRunner(self.ir, config=cfg, policies=policy_obj)
else:
runner = NumpyRunner(self.ir, config=cfg, policies=policy_obj)
if cache_manager is not None:
cache_key = build_cache_key(
program_src=self.src,
backend="numpy",
artifact="metadata",
device=target_device,
execution_config=cfg,
policies=policy_obj,
)
cache_manager.write_metadata(
"numpy",
cache_key,
{
"device": target_device,
"execution": cfg.mode,
"digest": self.digest,
},
)
return runner
compiler = self._resolve_backend(backend)
return compiler(
self,
device=target_device,
cache_manager=cache_manager,
execution_config=cfg,
policies=policy_obj,
**backend_kwargs,
)
def explain(self, *, json: bool = False) -> Any:
sources: List[Dict[str, Any]] = []
sinks: List[Dict[str, Any]] = []
equations: List[Dict[str, Any]] = []
for idx, eq in enumerate(self.ir.equations):
if eq.is_source:
sources.append(
{
"name": eq.lhs.name,
"path": eq.src_file,
"indices": lhs_indices(eq),
"line": eq.line,
"column": eq.column,
}
)
continue
if eq.is_sink:
sinks.append(
{
"name": eq.rhs.name if isinstance(eq.rhs, TensorRef) else eq.lhs.name,
"path": eq.sink_file,
"indices": lhs_indices(eq),
"line": eq.line,
"column": eq.column,
}
)
continue
lhs = lhs_indices(eq)
rhs = rhs_indices(eq)
projected = [axis for axis in rhs if axis not in lhs]
summary = equation_index_summary(eq, projected)
equations.append(
{
"id": idx,
"name": eq.lhs.name,
"projection": eq.projection,
"equation": eq.source,
"index_summary": summary,
"index_table": format_index_summary(summary),
"line": eq.line,
"column": eq.column,
}
)
payload = {
"digest": self.digest,
"sources": sources,
"equations": equations,
"sinks": sinks,
}
if json:
return json_ready(payload)
lines: List[str] = []
for src in sources:
indices = ",".join(src["indices"]) if src["indices"] else "-"
lines.append(f"[src] {src['name']} <- {src.get('path') or '<memory>'} idx[{indices}]")
for entry in equations:
table = entry["index_table"]
proj = entry["projection"]
lines.append(f"[eq] {entry['name']} {table} proj={proj}")
for sink in sinks:
indices = ",".join(sink["indices"]) if sink["indices"] else "-"
lines.append(
f"[sink] {sink.get('path') or '<memory>'} <- {sink['name']} idx[{indices}]"
)
return "\n".join(lines)
def _resolve_backend(self, backend: str):
if backend == "torch":
from ..torch_backend.compile import compile as torch_compile
return torch_compile
if backend == "jax":
from ..jax_backend.compile import compile as jax_compile
return jax_compile
raise BackendError(f"Unknown backend '{backend}'")
# Heuristics ---------------------------------------------------------------
def _choose_backend_auto(self, cfg: ExecutionConfig) -> str:
"""Pick a backend based on IR, execution config, and hardware.
Rules of thumb:
- Demand mode and Monte Carlo projections use NumPy.
- Streaming programs use NumPy.
- Prefer Torch on CUDA/MPS for attention/MLP-like workloads.
- Otherwise try JAX if available for heavier, batched workloads.
- Fall back to NumPy for small programs and when no accel is present.
"""
# Fast exits where Torch/JAX intentionally fall back to NumPy
if cfg.mode == "demand" or cfg.projection_strategy == "monte_carlo":
return "numpy"
if self.ir.has_streaming():
return "numpy"
# Inspect IR for workload hints
features = self._program_features()
equations = features.get("equations", 0)
einsum_terms = features.get("einsum_terms", 0)
has_attention = bool(features.get("has_attention", False))
op_count = features.get("op_count", 0)
# A simple workload score: einsums and attention weigh more
score = einsum_terms + op_count + (3 if has_attention else 0)
small_workload = equations <= 3 and score <= 2
target = (cfg.device or "auto").lower()
# Probe availability lazily
torch_available = False
torch_gpu_or_mps = False
try: # pragma: no cover - import/availability depends on environment
import torch as _torch # type: ignore
torch_available = _torch is not None
torch_gpu_or_mps = (
_torch.cuda.is_available() if hasattr(_torch, "cuda") else False
) or (
hasattr(_torch, "backends")
and hasattr(_torch.backends, "mps")
and bool(_torch.backends.mps.is_available())
)
except Exception: # pragma: no cover - torch optional
torch_available = False
torch_gpu_or_mps = False
jax_available = False
jax_has_gpu = False
try: # pragma: no cover - import/availability depends on environment
import jax as _jax # type: ignore
jax_available = _jax is not None
# If jax is present, ask for gpu devices
try:
jax_has_gpu = bool(_jax.devices("gpu"))
except Exception:
jax_has_gpu = False
except Exception: # pragma: no cover - jax optional
jax_available = False
jax_has_gpu = False
# GPU/MPS targets ------------------------------------------------------
wants_gpu_like = (
target.startswith("cuda")
or target == "mps"
or (target == "auto" and (torch_gpu_or_mps or jax_has_gpu))
)
if not small_workload and wants_gpu_like:
# Prefer Torch for attention/DL-like programs when available
if torch_available and (
torch_gpu_or_mps or target.startswith("cuda") or target == "mps"
):
return "torch"
if jax_available and (jax_has_gpu or target == "auto"):
return "jax"
# No accel but heavy: fall through to CPU choices below
# CPU targets or small workloads --------------------------------------
if small_workload:
return "numpy"
# Heavier CPU workloads: prefer Torch if available, else JAX.
if torch_available:
return "torch"
if jax_available:
return "jax"
return "numpy"
def _program_features(self) -> Dict[str, int]:
"""Collect lightweight structural features from the IR for heuristics."""
equations = 0
einsum_terms = 0
op_count = 0
has_attention = 0
from .ir import FuncCall, Term # local import to avoid cycles
for eq in self.ir.equations:
if eq.is_source or eq.is_sink:
continue
equations += 1
rhs = eq.rhs
if isinstance(rhs, Term):
if len(getattr(rhs, "factors", []) or []) >= 2:
einsum_terms += 1
if isinstance(rhs, FuncCall):
op_count += 1
name = (rhs.name or "").lower()
if name in {"attention"}:
has_attention = 1
return {
"equations": equations,
"einsum_terms": einsum_terms,
"op_count": op_count,
"has_attention": has_attention,
}
|