-
Notifications
You must be signed in to change notification settings - Fork 3.4k
/
task.py
410 lines (345 loc) · 13.3 KB
/
task.py
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
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
# Licensed to the Apache Software Foundation (ASF) under one
# or more contributor license agreements. See the NOTICE file
# distributed with this work for additional information
# regarding copyright ownership. The ASF licenses this file
# to you under the Apache License, Version 2.0 (the
# "License"); you may not use this file except in compliance
# with the License. You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing,
# software distributed under the License is distributed on an
# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
# KIND, either express or implied. See the License for the
# specific language governing permissions and limitations
# under the License.
# pylint: disable=unused-variable
"""Definition of task function.
Task can be constructed from tuple of func, args, and kwargs.
func is a state-less function, or a string that
registers the standard task.
"""
import numpy as np
from ... import tensor, expr, container, target as _target
from ..util import get_const_int, get_const_tuple, get_func_name
from .dispatcher import DispatchContext, ApplyConfig, dispatcher
from .space import ConfigSpace
def _raise_error(*args, **kwargs): # pylint: disable=unused-argument
raise RuntimeError("The function of this task is not found. Possibly the function "
"of this task is registered in another python file "
"which is not imported in this run")
class Task(object):
"""A Tunable Task
Parameters
----------
name: str
The name of the task.
args: Tuple
Positional argument of func
"""
def __init__(self, name, args):
self.name = name
self.args = args
self.kwargs = {} # currently unused
# init null config space
self.config_space = None
self.func = TASK_TABLE.get(name, _raise_error)
# auxiliary info, available after `init_space` is called
self.workload = None
self.flop = None
self.target = None
self.target_host = None
def instantiate(self, config):
"""Instantiate this task function (template) with a config.
Returns corresponding schedule.
Parameters
----------
config: template.ConfigEntity
parameter config for this template
Returns
-------
sch: tvm.schedule.Schedule
The tvm schedule
arg_bufs: Array of tvm.tensor.Tensor
The input/output buffers
"""
config.flop = 0
with ApplyConfig(config):
sch, arg_bufs = self.func(*self.args, **self.kwargs)
if not self.flop:
config.flop = config.flop or compute_flop(sch)
self.flop = config.flop
return sch, arg_bufs
def __getstate__(self):
# custom pickle implementation is required for
# some unpickable local task functions.
# So we only pickle the name of the function
# and restore the function by name when unpickling it.
return {
"name": self.name,
"args": self.args,
"kwargs": self.kwargs,
"config_space": self.config_space,
"workload": self.workload,
"flop": self.flop,
"target": self.target,
"target_host": self.target_host
}
def __setstate__(self, state):
self.name = state["name"]
self.args = state["args"]
self.kwargs = state["kwargs"]
self.config_space = state["config_space"]
self.func = TASK_TABLE.get(state["name"], _raise_error)
self.workload = state["workload"]
self.flop = state["flop"]
self.target = state["target"]
self.target_host = state["target_host"]
def __repr__(self):
return "Task(func_name=%s, args=%s, kwargs=%s, workload=%s)" % (
self.name, self.args, self.kwargs, self.workload
)
TASK_TABLE = {
}
def register(name, func=None, override=False):
"""Register a task function.
Parameters
----------
name : str
The name to identify the task.
func : callable
The function to be registered.
override : bool
Whether override existing registration.
Returns
-------
func: callable
The registered function
"""
def _do_reg(myf):
if name in TASK_TABLE and not override:
raise ValueError(
"Key %s is already registered" % name)
TASK_TABLE[name] = myf
return myf
if func:
return _do_reg(func)
return _do_reg
def create(func_name, args, target, target_host=None, template_key=None):
"""Create a tuning task and initialize its search space
Parameters
----------
func_name : str or callable
The task function
args : List
Positional arguments
target : Target
The compilation target
target_host: Target, optional
The compilation target for host side
Returns
-------
tsk: Task
a task object
"""
if callable(func_name):
# register this function if it is not registered before
func = func_name
func_name = func.func_name if hasattr(func, 'func_name') else func.__name__
if func_name in TASK_TABLE:
assert func == TASK_TABLE[func_name], "Find name conflict in task registration. " \
"Consider to choose another name for this task"
else:
register(func_name, func=func)
func = TASK_TABLE[func_name]
ret = Task(func_name, args)
if isinstance(target, str):
target = _target.create(target)
# init config space
ret.config_space = ConfigSpace()
ret.config_space.template_key = template_key or ""
ctx = ApplyConfig(ret.config_space)
with ctx:
with target:
sch, _ = func(*args)
ret.config_space.code_hash = getattr(sch, 'code_hash', None)
ret.workload = ctx.workload
ret.flop = ret.config_space.flop or compute_flop(sch)
ret.target = target
ret.target_host = target_host
return ret
def args_to_workload(x, topi_compute_func=None):
"""Convert argument list to hashable workload tuple.
This function will convert list to tuple, tvm node to python value and
flatten tvm.tensor.Tensor to a tuple
Parameters
----------
x: primitive hashable types or tensor.Tensor
The original value
topi_compute_func: topi compute function
The function name will be added as first element of the workload tuple
Returns
-------
ret: hashable
The hashable value
"""
if isinstance(x, tensor.Tensor):
workload = get_const_tuple(x.shape) + (x.dtype, )
elif isinstance(x, (tuple, list, container.Array)):
workload = tuple([args_to_workload(a) for a in x])
elif isinstance(x, (str, int, float, np.int, np.float, expr.Var)):
workload = x
elif isinstance(x, (expr.StringImm, expr.UIntImm, expr.IntImm, expr.FloatImm)):
workload = x.value
elif x is None:
workload = 0
else:
raise RuntimeError('Do not support type "%s" in argument. Consider to use'
'primitive types or tvm.expr.Var only' % type(x))
return (get_func_name(topi_compute_func), ) + workload if topi_compute_func else workload
def template(func):
"""
Decorate a function as a tunable schedule template
Parameters
----------
func: callable
A callable template function.
Its argument should be hashable values.
Its return value should be a Tuple(Schedule, Array of Tensor)
Returns
-------
func: callable
The decorated function
Examples
--------
The following code is a tunable template for a blocked matrix multiplication
.. code-block:: python
@autotvm.template
def matmul(N, L, M, dtype):
A = tvm.placeholder((N, L), name='A', dtype=dtype)
B = tvm.placeholder((L, M), name='B', dtype=dtype)
k = tvm.reduce_axis((0, L), name='k')
C = tvm.compute((N, M), lambda i, j: tvm.sum(A[i, k] * B[k, j], axis=k), name='C')
s = tvm.create_schedule(C.op)
# schedule
y, x = s[C].op.axis
k = s[C].op.reduce_axis[0]
##### define space begin #####
cfg = autotvm.get_config()
cfg.define_split("tile_y", y, num_outputs=2)
cfg.define_split("tile_x", x, num_outputs=2)
##### define space end #####
# schedule according to config
yo, yi = cfg["tile_y"].apply(s, C, y)
xo, xi = cfg["tile_x"].apply(s, C, x)
s[C].reorder(yo, xo, k, yi, xi)
return s, [A, B, C]
"""
# pylint: disable=unused-variable
fname = get_func_name(func)
@register(fname)
@dispatcher
def config_dispatcher(*args, **kwargs):
assert not kwargs, "Do not support kwargs in template function call"
return (fname, ) + args_to_workload(args)
@config_dispatcher.register("")
def template_call(cfg, *args, **kwargs):
assert not kwargs, "Do not support kwargs in template function call"
with ApplyConfig(cfg):
return func(*args, **kwargs)
config_dispatcher.func_name = fname
return config_dispatcher
def get_config():
"""Get current config object
Returns
-------
cfg: ConfigSpace or ConfigEntity
The current config
"""
return DispatchContext.current.query(None, None)
class FlopCalculationError(RuntimeError):
"""Error happens when estimating FLOP for a compute op"""
def compute_flop(sch):
"""Calculate number of FLOP (floating number operations) of the compute ops in a schedule
Parameters
----------
sch: tvm.schedule.Schedule
schedule
Returns
-------
flop: int
number of FLOP in this schedule
"""
def _prod_length(axes):
"""compute product of the lengths of a list of axes"""
try:
num_iter = int(np.prod([get_const_int(axis.dom.extent) for axis in axes]))
except ValueError:
raise FlopCalculationError("The length of axis is not constant. ")
return num_iter
def _count_flop(exp):
"""compute flop for a single expression"""
if isinstance(exp, expr.Reduce):
num_iter = _prod_length(exp.axis)
combiner = exp.combiner.result
source = exp.source
if len(combiner) != 1:
raise FlopCalculationError("Found multiple output in the combiner of reduce op")
if len(source) != 1:
raise FlopCalculationError("Found multiple output in the source of reduce op")
return num_iter * (_count_flop(combiner[0]) + _count_flop(source[0]))
if isinstance(exp, (expr.FloatImm, expr.IntImm, expr.UIntImm)):
return 0
if isinstance(exp, expr.Cast):
return _count_flop(exp.value)
if isinstance(exp, expr.Var):
return 0
if isinstance(exp, (expr.Add, expr.Sub, expr.Mul,
expr.Div, expr.Mod,
expr.FloorDiv, expr.FloorMod,
expr.Max, expr.Min,
expr.EQ, expr.NE, expr.LT, expr.LE, expr.GT, expr.GE,
expr.And, expr.Or, expr.Not)):
base = 1
if isinstance(exp, expr.Not): # unary
return base + _count_flop(exp.a)
return base + _count_flop(exp.a) + _count_flop(exp.b)
if isinstance(exp, expr.Select):
return _count_flop(exp.condition) + max(_count_flop(exp.true_value),
_count_flop(exp.false_value))
if isinstance(exp, expr.Call):
if exp.call_type == expr.Call.Halide:
# Ignore flops from indexing expressions.
return 0
return sum([_count_flop(x) for x in exp.args])
raise FlopCalculationError("Found unsupported operator in the compute expr")
def traverse(ops):
"""accumulate flops"""
ret = 0
for op in ops:
if isinstance(op, tensor.ComputeOp):
num_element = _prod_length(op.axis)
body = op.body
if len(body) != 1:
raise FlopCalculationError("Found multiple output in the compute")
exp = body[0]
ret += num_element * _count_flop(exp)
ret += traverse([t.op for t in op.input_tensors])
elif isinstance(op, tensor.PlaceholderOp):
pass
else:
raise FlopCalculationError("Only support tvm.compute currently. "
"Other ops like tvm.scan/tvm.extern is not supported")
return ret
try:
ret = traverse(sch.outputs)
except FlopCalculationError as exc:
raise RuntimeError("FLOP estimator fails for this operator. Error msg: "
+ str(exc) + ". Please use `cfg.add_flop` to manually set "
"FLOP for this operator")
if ret == 0:
raise RuntimeError("Cannot find float number operation in this operator. "
"Please use `cfg.add_flop` to manually set "
"FLOP for this operator")
return ret