-
Notifications
You must be signed in to change notification settings - Fork 3.4k
/
vta_conv2d.py
505 lines (448 loc) · 18.8 KB
/
vta_conv2d.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
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
# 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.
"""Namespace for supporting packed_conv2d + ewise variant of nnvm."""
from __future__ import absolute_import as _abs
from collections import namedtuple
import logging
import tvm
import topi
from nnvm.top import registry as reg, OpPattern
from nnvm.top import nn as _nn
from ..environment import get_env
Workload = namedtuple("Conv2DWorkload",
['batch', 'height', 'width', 'in_filter', 'out_filter',
'hkernel', 'wkernel', 'hpad', 'wpad', 'hstride', 'wstride'])
def find_schedules(layer, vt_only=False, best_only=False):
""" Returns a schedule for a given a layer.
Parameters
----------
layer : Workload
Convolutional layer description.
vt_only : Boolean
Produce a schedule plan with virtual threading.
best_only : Boolean
Return the "best" schedule plan.
Returns
-------
fil_sched : list
List of valid schedules.
"""
# pylint: disable=too-many-nested-blocks
env = get_env()
# Helper function to get factors
def _find_factors(n):
factors = []
for f in range(1, n + 1):
if n % f == 0:
factors.append(f)
return factors
def _get_data_movement_byte(schedule, layer):
""" Estimate data movement in bytes for the schedule plan
"""
env = get_env()
b_f = schedule.b_factor
h_f = schedule.h_factor
w_f = schedule.w_factor
ci_f = schedule.ic_factor
co_f = schedule.oc_factor
# Derive data movement
inp_elem_sizeb = env.BATCH * env.BLOCK_IN * env.INP_WIDTH
wgt_elem_sizeb = env.BLOCK_IN * env.BLOCK_OUT * env.WGT_WIDTH
out_elem_sizeb = env.BATCH * env.BLOCK_OUT * env.OUT_WIDTH
input_tile_elems = b_f * \
((h_f - 1) * layer.hstride + layer.hkernel) * \
((w_f - 1) * layer.wstride + layer.wkernel) * ci_f
weight_tile_elems = layer.hkernel * layer.wkernel * ci_f
output_tile_elems = b_f * h_f * w_f * co_f
# Derive tiling factors
b_factor = layer.batch // (b_f * env.BATCH)
h_factor = (layer.height // layer.hstride) // h_f
w_factor = (layer.width // layer.wstride) // w_f
ci_factor = layer.in_filter // (ci_f * env.BLOCK_IN)
co_factor = layer.out_filter // (co_f * env.BLOCK_OUT)
# Compute input transaction count
input_xfers = b_factor * h_factor * w_factor * co_factor * ci_factor
weight_xfers = b_factor * h_factor * w_factor * co_factor * ci_factor
output_xfers = b_factor * h_factor * w_factor * co_factor
# Compute total transfer sizes
input_xfer_byte = input_tile_elems * input_xfers * inp_elem_sizeb // 8
weight_xfer_byte = weight_tile_elems * weight_xfers * wgt_elem_sizeb // 8
output_xfer_byte = output_tile_elems * output_xfers * out_elem_sizeb // 8
total_xfer_byte = input_xfer_byte + weight_xfer_byte + output_xfer_byte
return total_xfer_byte
# Scheduling exploration
batch_factors = _find_factors(layer.batch // env.BATCH)
height_factors = _find_factors(layer.height // layer.hstride)
width_factors = _find_factors(layer.width // layer.wstride)
cin_factors = _find_factors(layer.in_filter // env.BLOCK_IN)
cout_factors = _find_factors(layer.out_filter // env.BLOCK_OUT)
ht_factors = [1, 2]
cot_factors = [1, 2]
# Explore schedules
schedules = []
for b_f in batch_factors:
for h_f in height_factors:
for w_f in width_factors:
for ci_f in cin_factors:
for co_f in cout_factors:
# FIXME: 2D load pattern matching imposes restrictions on schedule
valid = (w_f == layer.width // layer.wstride) or \
(w_f != layer.width // layer.wstride and co_f == 1) and \
ci_f == 1
if valid:
schedules.append([b_f, h_f, w_f, ci_f, co_f])
# Filter the schedules that wouldn't work in the available BRAM sizes
inp_elem_sizeb = env.BATCH * env.BLOCK_IN * env.INP_WIDTH
wgt_elem_sizeb = env.BLOCK_IN * env.BLOCK_OUT * env.WGT_WIDTH
out_elem_sizeb = env.BATCH * env.BLOCK_OUT * env.OUT_WIDTH
inp_brams_sizeb = env.INP_BUFF_SIZE * 8
wgt_brams_sizeb = env.WGT_BUFF_SIZE * 8
out_brams_sizeb = env.OUT_BUFF_SIZE * 8
fil_sched = []
xfer_size = []
for sched in schedules:
b_f, h_f, w_f, ci_f, co_f = sched
for h_t in ht_factors:
for co_t in cot_factors:
# Make sure to filter cases where we apply threading on two axes
# or cases where the threading factors for h and co are not
# factors of h and co
if (h_t == 2 and co_t == 2) or (h_f % h_t != 0) or (co_f % co_t != 0):
continue
# Adjust tile sizes if threading is applied
h_f //= h_t
co_f //= co_t
# Derive tile sizes
input_tile_elems = b_f * \
((h_f - 1) * layer.hstride + layer.hkernel) * \
((w_f - 1) * layer.wstride + layer.wkernel) * ci_f
weight_tile_elems = layer.hkernel * layer.wkernel * ci_f * co_f
output_tile_elems = b_f * h_f * w_f * co_f
# Derive valid schedule filter
valid = True
# If in vitrual-threaded mode, only allow for threaded plans
valid &= (vt_only and (h_t == 2 or co_t == 2)) or not vt_only
# Check that we don't exceed input/weight/output capacity
valid &= input_tile_elems * inp_elem_sizeb <= inp_brams_sizeb // (co_t * h_t)
valid &= weight_tile_elems * wgt_elem_sizeb <= wgt_brams_sizeb
valid &= output_tile_elems * out_elem_sizeb <= out_brams_sizeb // (co_t * h_t)
# Make sure that we don't write to the same acc location within 2 consecutive cycles
valid &= h_f > 2 and w_f > 2
# TODO: check that we don't exceed instruction or micro-op count
if valid:
schedule = Schedule(b_factor=b_f, oc_factor=co_f, ic_factor=ci_f, h_factor=h_f,
w_factor=w_f, oc_nthread=co_t, h_nthread=h_t)
fil_sched.append(schedule)
xfer_size.append(_get_data_movement_byte(schedule, layer))
if best_only:
return [fil_sched[xfer_size.index(min(xfer_size))]]
return fil_sched
def packed_conv2d(data,
kernel,
padding,
strides,
out_dtype="int32"):
""" Packed conv2d function.
"""
if padding[0]:
pad_data = topi.nn.pad(data, [0, 0, padding[0], padding[1], 0, 0], name="pad_data")
else:
pad_data = data
assert len(data.shape) == 6
assert len(kernel.shape) == 6
oheight = topi.util.simplify((pad_data.shape[2] - kernel.shape[2]) // strides[0] + 1)
owidth = topi.util.simplify((pad_data.shape[3] - kernel.shape[3]) // strides[1] + 1)
oshape = (data.shape[0], kernel.shape[0], oheight, owidth, data.shape[4], kernel.shape[4])
ishape = topi.util.get_const_tuple(data.shape)
kshape = topi.util.get_const_tuple(kernel.shape)
assert data.dtype == "int8", data.dtype
assert kernel.dtype == "int8", kernel.dtype
d_i = tvm.reduce_axis((0, kshape[2]), name='d_i')
d_j = tvm.reduce_axis((0, kshape[3]), name='d_j')
k_o = tvm.reduce_axis((0, ishape[1]), name='k_o')
k_i = tvm.reduce_axis((0, ishape[-1]), name='k_i')
hstride, wstride = strides
res = tvm.compute(
oshape,
lambda b_o, c_o, i, j, b_i, c_i: tvm.sum(
pad_data[b_o, k_o, i*hstride+d_i, j*wstride+d_j, b_i, k_i].astype(out_dtype) *
kernel[c_o, k_o, d_i, d_j, c_i, k_i].astype(out_dtype),
axis=[k_o, d_i, d_j, k_i]),
name="res", tag="packed_conv2d")
return res
@tvm.register_func("nnvm.compiler.build_target", override=True)
def _build(funcs, target, target_host):
tvm_t = tvm.target.create(target)
if tvm_t.device_name == "vta":
return tvm.build(funcs, target="ext_dev", target_host=target_host)
if tvm_t.device_name == "rasp" or tvm_t.device_name == "vtacpu":
return tvm.build(funcs, target=target_host)
return tvm.build(funcs, target=target)
@tvm.register_func("nnvm.compiler.lower", override=True)
def _lower(sch, inputs, func_name, graph):
import traceback
# pylint: disable=broad-except
try:
f = tvm.lower(sch, inputs, name=func_name)
if "quantized_conv2d" in func_name:
logging.info(graph.ir(join_entry_attrs=["shape"]))
except Exception:
msg = traceback.format_exc()
msg += "Error during compile graph\n"
msg += "--------------------------\n"
msg += graph.ir(join_entry_attrs=["shape"])
raise RuntimeError(msg)
return f if isinstance(
f, (tvm.container.Array, tuple, list)) else [f]
@reg.register_compute("clip", level=15)
def compute_clip(attrs, inputs, _):
""" Clip operator.
"""
x = inputs[0]
a_min = attrs.get_float("a_min")
a_max = attrs.get_float("a_max")
const_min = tvm.const(a_min, x.dtype)
const_max = tvm.const(a_max, x.dtype)
with tvm.tag_scope(topi.tag.ELEMWISE):
x = tvm.compute(
x.shape, lambda *i: tvm.min(x(*i), const_max), name="clipA")
x = tvm.compute(
x.shape, lambda *i: tvm.max(x(*i), const_min), name="clipB")
return x
# override to force partition at copy
reg.register_pattern("copy", OpPattern.INJECTIVE, level=15)
def is_packed_layout(layout):
"""Check if layout is packed layout"""
if layout == "NCHW":
return False
if "n" in layout and "c" in layout:
return True
return False
@reg.register_alter_op_layout("conv2d", level=15)
def alter_conv2d_layout(attrs, inputs, out):
layout = attrs['layout']
if is_packed_layout(layout):
return None
return _nn.alter_conv2d_layout(attrs, inputs, out)
@reg.register_compute("conv2d", level=15)
def compute_conv2d(attrs, inputs, out):
""" 2D convolution algorithm.
"""
padding = attrs.get_int_tuple("padding")
strides = attrs.get_int_tuple("strides")
dilation = attrs.get_int_tuple("dilation")
groups = attrs.get_int("groups")
layout = attrs["layout"]
out_dtype = attrs['out_dtype']
assert dilation == (1, 1), "not support dilate now"
if is_packed_layout(layout):
assert groups == 1
return packed_conv2d(inputs[0], inputs[1],
padding, strides, out_dtype=out_dtype)
return _nn.compute_conv2d(attrs, inputs, out)
@reg.register_schedule("conv2d", level=15)
def schedule_conv2d(attrs, outs, target):
""" 2D convolution schedule.
"""
layout = attrs["layout"]
if is_packed_layout(layout):
target = tvm.target.create(target)
if target.device_name == "vta":
return schedule_packed_conv2d(outs)
if str(target).startswith("llvm"):
return tvm.create_schedule([x.op for x in outs])
raise RuntimeError("not support target %s" % target)
return _nn.schedule_conv2d(attrs, outs, target)
def _get_workload(data, pad_data, kernel, output):
""" Get the workload structure.
"""
o_shape = topi.util.get_const_tuple(output.shape)
d_shape = topi.util.get_const_tuple(data.shape)
k_shape = topi.util.get_const_tuple(kernel.shape)
o_b, o_c, o_h, o_w, ob_blk, o_blk = o_shape
i_b, i_c, i_h, i_w, ib_blk, i_blk = d_shape
k_o, k_i, k_h, k_w, ko_blk, ki_blk = k_shape
# For now we need to assume that input channel blocking is the same
# as the output channel blocking
assert o_blk == i_blk
assert ob_blk == ib_blk
# Make sure that dimensions match
assert o_b == i_b
assert o_blk == ko_blk
assert i_blk == ki_blk
assert k_o == o_c
assert k_i == i_c
# Scale the channel size
i_c *= i_blk
o_c *= o_blk
if pad_data is not None:
p_shape = topi.util.get_const_tuple(pad_data.shape)
h_pad = (p_shape[2] - d_shape[2]) // 2
w_pad = (p_shape[3] - d_shape[3]) // 2
else:
h_pad, w_pad = 0, 0
h_str = (i_h + h_pad*2 - k_h) // (o_h - 1)
w_str = (i_w + w_pad*2 - k_w) // (o_w - 1)
return Workload(i_b, i_h, i_w, i_c, o_c, k_h, k_w, h_pad, w_pad, h_str, w_str)
_WL2PLAN = {}
def schedule_packed_conv2d(outs):
""" Schedule the packed conv2d.
"""
assert len(outs) == 1
output = outs[0]
ewise_inputs = []
ewise_ops = []
conv2d_res = []
assert output.dtype == "int8"
assert output.op.input_tensors[0].dtype == "int32"
def _traverse(op):
if topi.tag.is_broadcast(op.tag):
if not op.same_as(output.op):
ewise_ops.append(op)
for tensor in op.input_tensors:
if isinstance(tensor.op, tvm.tensor.PlaceholderOp):
ewise_inputs.append((op, tensor))
else:
_traverse(tensor.op)
else:
assert op.tag == "packed_conv2d"
conv2d_res.append(op)
_traverse(output.op)
assert len(conv2d_res) == 1
conv2d_stage = conv2d_res[0].output(0)
data, kernel = conv2d_stage.op.input_tensors
if isinstance(data.op, tvm.tensor.ComputeOp) and "pad" in data.op.tag:
temp = data.op.input_tensors[0]
pad_data = data
data = temp
else:
pad_data = None
wrkld = _get_workload(data, pad_data, kernel, output)
if wrkld in _WL2PLAN:
plan = _WL2PLAN[wrkld]
else:
plan = find_schedules(wrkld, vt_only=True, best_only=True)[0]
logging.info("Trying to find plan for %s", wrkld)
env = get_env()
load_inp = load_wgt = load_out = store_out = env.dma_copy
alu = env.alu
gemm = env.gemm
# schedule1
oshape = topi.util.get_const_tuple(output.shape)
s = tvm.create_schedule(output.op)
# setup pad
if pad_data is not None:
cdata = pad_data
s[pad_data].set_scope(env.inp_scope)
else:
cdata = s.cache_read(data, env.inp_scope, [conv2d_stage])
ckernel = s.cache_read(kernel, env.wgt_scope, [conv2d_stage])
s[conv2d_stage].set_scope(env.acc_scope)
# cache read input
cache_read_ewise = []
for consumer, tensor in ewise_inputs:
cache_read_ewise.append(
s.cache_read(tensor, env.acc_scope, [consumer]))
# set ewise scope
for op in ewise_ops:
s[op].set_scope(env.acc_scope)
s[op].pragma(s[op].op.axis[0], alu)
# tile
oc_factor = (plan.oc_factor if plan.oc_factor
else plan.out_filter // env.BLOCK_OUT)
h_factor = (plan.h_factor if plan.h_factor else oshape[2])
w_factor = (plan.w_factor if plan.w_factor else oshape[3])
x_bo, x_co, x_i, x_j, x_bi, x_ci = s[output].op.axis
x_co0, x_co1 = s[output].split(x_co, factor=oc_factor)
x_i0, x_i1 = s[output].split(x_i, factor=h_factor)
x_j0, x_j1 = s[output].split(x_j, factor=w_factor)
s[output].reorder(x_bo, x_i0, x_co0, x_j0, x_co1, x_i1, x_j1, x_bi, x_ci)
store_pt = x_j0
# set all compute scopes
s[conv2d_stage].compute_at(s[output], store_pt)
for op in ewise_ops:
s[op].compute_at(s[output], store_pt)
for tensor in cache_read_ewise:
s[tensor].compute_at(s[output], store_pt)
s[tensor].pragma(s[tensor].op.axis[0], load_out)
# virtual threading along output channel axes
if plan.oc_nthread > 1:
_, v_t = s[output].split(x_co0, factor=plan.oc_nthread)
s[output].reorder(v_t, x_bo)
s[output].bind(v_t, tvm.thread_axis("cthread"))
# virtual threading along spatial rows
if plan.h_nthread > 1:
_, v_t = s[output].split(x_i0, factor=plan.h_nthread)
s[output].reorder(v_t, x_bo)
s[output].bind(v_t, tvm.thread_axis("cthread"))
x_bo, x_co, x_i, x_j, x_bi, x_ci = s[conv2d_stage].op.axis
k_o, d_i, d_j, k_i = s[conv2d_stage].op.reduce_axis
s[conv2d_stage].reorder(x_bo, k_o, x_j, d_j, d_i, x_co, x_i, x_bi, x_ci, k_i)
if plan.ic_factor:
k_o, _ = s[conv2d_stage].split(k_o, factor=plan.ic_factor)
s[cdata].compute_at(s[conv2d_stage], k_o)
s[ckernel].compute_at(s[conv2d_stage], k_o)
# Use VTA instructions
s[cdata].pragma(s[cdata].op.axis[0], load_inp)
s[ckernel].pragma(s[ckernel].op.axis[0], load_wgt)
s[conv2d_stage].tensorize(x_bi, gemm)
s[output].pragma(x_co1, store_out)
return s
class Conv2DSchedule(object):
""" 2D convolution schedule object.
"""
def __init__(self,
b_factor=1,
oc_factor=1,
ic_factor=1,
h_factor=1,
w_factor=0,
oc_nthread=0,
h_nthread=0,
debug_sync=False):
self.b_factor = b_factor
self.oc_factor = oc_factor
self.ic_factor = ic_factor
self.h_factor = h_factor
self.w_factor = w_factor
self.oc_nthread = oc_nthread
self.h_nthread = h_nthread
self.debug_sync = debug_sync
def __str__(self):
return "{}.{}.{}.{}.{}.{}.{}".format(
self.b_factor, self.oc_factor, self.ic_factor,
self.h_factor, self.w_factor,
self.oc_nthread, self.h_nthread)
Schedule = Conv2DSchedule
# Layer description of the ResNet18
RESNET = {
0: Workload(1, 224, 224, 16, 64, 7, 7, 3, 3, 2, 2),
1: Workload(1, 56, 56, 64, 64, 3, 3, 1, 1, 1, 1),
2: Workload(1, 56, 56, 64, 64, 1, 1, 0, 0, 1, 1),
3: Workload(1, 56, 56, 64, 128, 3, 3, 1, 1, 2, 2),
4: Workload(1, 56, 56, 64, 128, 1, 1, 0, 0, 2, 2),
5: Workload(1, 28, 28, 128, 128, 3, 3, 1, 1, 1, 1),
6: Workload(1, 28, 28, 128, 256, 3, 3, 1, 1, 2, 2),
7: Workload(1, 28, 28, 128, 256, 1, 1, 0, 0, 2, 2),
8: Workload(1, 14, 14, 256, 256, 3, 3, 1, 1, 1, 1),
9: Workload(1, 14, 14, 256, 512, 3, 3, 1, 1, 2, 2),
10: Workload(1, 14, 14, 256, 512, 1, 1, 0, 0, 2, 2),
11: Workload(1, 7, 7, 512, 512, 3, 3, 1, 1, 1, 1),
}
for idx in RESNET:
scheds = find_schedules(RESNET[idx], vt_only=True, best_only=True)[0]
_WL2PLAN[RESNET[idx]] = scheds