forked from apache/tvm
-
Notifications
You must be signed in to change notification settings - Fork 0
/
unet_conv2d_NCHWc_direct.py
231 lines (195 loc) · 9.76 KB
/
unet_conv2d_NCHWc_direct.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
from __future__ import absolute_import as _abs
import numpy as np
import tvm
from tvm import autotvm
from topi.generic import schedule_conv2d_nchw, schedule_conv2d_NCHWc_
from topi.util import traverse_inline, get_const_tuple, const_matrix
from topi.nn import pad, conv2d, conv2d_NCHWc, conv2d_alter_layout
from topi.nn.util import get_const_int, get_pad_tuple
import topi.nn
def _conv_NCHWc_arg_to_workload(data, kernel, num_filter, kernel_size, stride, padding, layout, out_layout, out_dtype):
"""convert argument to workload"""
return ('conv2d_NCHWc', ) + autotvm.task.args_to_workload(
[data, kernel, stride, padding, layout, out_layout, out_dtype])
def _decl_spatial_pack_NCHWc(cfg, data, kernel, num_filter, kernel_size, stride, padding, layout, out_layout, out_dtype):
# import ipdb
# ipdb.set_trace()
# assert layout == "NCHW", "Only support NCHW"
# create workload according to raw arguments
wkl = _conv_NCHWc_arg_to_workload(
data, kernel, num_filter, kernel_size,
stride, padding, layout, out_layout, out_dtype)
out_dtype = out_dtype or data.dtype
N, CII, IH, IW, CIII = get_const_tuple(data.shape)
COO, CII, KH, KW, CIII_, VC = get_const_tuple(kernel.shape)
pad_top, pad_left, pad_bottom, pad_right = get_pad_tuple(padding, (KH, KW))
HSTR, WSTR = stride if isinstance(stride, (tuple, list)) else (stride, stride)
OH = (IH + pad_top + pad_bottom - KH) // HSTR + 1
OW = (IW + pad_left + pad_right - KW) // WSTR + 1
data_pad = pad(data, [0, 0, pad_top, pad_left, 0], [0, 0, pad_bottom, pad_right, 0], name="data_pad")
# ==================== define configuration space ====================
n, coo, oh, ow, vc = cfg.axis(N), cfg.axis(COO), cfg.axis(OH), cfg.axis(OW), cfg.axis(VC)
cii, ciii, kh, kw = cfg.reduce_axis(CII), cfg.reduce_axis(CIII), cfg.reduce_axis(KH), cfg.reduce_axis(KW)
oh, vh = cfg.define_split('tile_oh', oh, num_outputs=2, filter=lambda x: x.size[-1] <= 8)
ow, vw = cfg.define_split('tile_ow', ow, num_outputs=2, filter=lambda x: x.size[-1] <= 8)
cfg.define_reorder("reorder_0",
[n, coo, cii, oh, ow, kh, kw, vc, ciii, vh, vw],
policy='candidate', candidate=[
[n, coo, cii, oh, ow, kh, kw, vc, ciii, vh, vw],
[n, coo, cii, oh, ow, kh, kw, ciii, vh, vw, vc],
[n, coo, cii, oh, ow, kh, kw, vc, vh, ciii, vw],
[n, coo, cii, oh, ow, kh, kw, ciii, vh, vc, vw],
[n, coo, oh, cii, ow, kh, kw, ciii, vh, vw, vc],
])
cfg.define_reorder("reorder_1",
[n, coo, oh, ow, vh, vw, vc],
policy='candidate', candidate=[
[n, coo, oh, ow, vh, vw, vc],
[n, coo, oh, ow, vc, vh, vw],
[n, coo, oh, ow, vh, vc, vw]
])
cfg.define_annotate("ann_reduce", [kh, kw, ciii], policy='try_unroll')
cfg.define_annotate("ann_spatial", [vh, vw, vc], policy='try_unroll_vec')
# cfg.define_annotate("ann_spatial", [vh, vw, vc], policy='try_unroll_vec')
# fallback support
# if cfg.is_fallback:
# if num_tile == 2: # arm cpu
# ref_log = autotvm.tophub.load_reference_log('cpu', 'rk3399', 'conv2d', 'direct')
# cfg.fallback_with_reference_log(ref_log)
# elif num_tile == 3: # mali gpu
# ref_log = autotvm.tophub.load_reference_log('mali', 'rk3399', 'conv2d', 'direct')
# cfg.fallback_with_reference_log(ref_log)
# ====================================================================
VH = cfg["tile_oh"].size[-1]
VW = cfg["tile_ow"].size[-1]
# input = (N, CII, IH, IW, CIII)
# -> transpose
############################################################
# input_tile_shape = (N, CII, OH // VH, OH // VH, VH + KH, VW + KW, CIII)
# oshape = (N, COO, OH // VH, OW // VH, VH, VW, COOO)
############################################################
# -> transpose
# O_shape = (N, COO, OH, OW, COOO)
dvshape = (N, CII, OH // VH, OW // VW, VH*HSTR + KH-1, VW*WSTR + KW-1, CIII)
ovshape = (N, COO, OH // VH, OW // VW, VH, VW, VC)
oshape = (N, COO, OH, OW, VC)
data_vec = tvm.compute(dvshape, lambda n, cii, h, w, vh, vw, ciii:
data_pad[n][cii][h*VH*HSTR+vh][w*VW*WSTR+vw][ciii],
name='data_vec')
cii = tvm.reduce_axis((0, CII), name='cii')
ciii = tvm.reduce_axis((0, CIII), name='ciii')
kh = tvm.reduce_axis((0, KH), name='kh')
kw = tvm.reduce_axis((0, KW), name='kw')
conv = tvm.compute(ovshape, lambda n, coo, h, w, vh, vw, vc: \
tvm.sum(data_vec[n, cii, h, w, vh*HSTR+kh, vw*WSTR+kw, ciii].astype(out_dtype) *
kernel[coo, cii, kh, kw, ciii, vc].astype(out_dtype),
axis=[cii, ciii, kh, kw]), name='conv')
output = tvm.compute(oshape, lambda n, coo, h, w, vc:
conv[n][coo][h//VH][w//VW][h%VH][w%VW][vc],
name='output_unpack', tag='spatial_conv2d_output',
attrs={'workload': wkl})
return output
def _schedule_spatial_pack_NCHWc(cfg, s, output, last):
"""schedule implementation"""
"""schedule implementation"""
# import ipdb
# ipdb.set_trace()
conv = output.op.input_tensors[0]
data_vec = conv.op.input_tensors[0]
data_pad = data_vec.op.input_tensors[0]
# s[data_pad].compute_inline()
kernel_vec = conv.op.input_tensors[1]
n, coo, oh, ow, vh, vw, vc = s[conv].op.axis
_, dvcii, dvoh, dvow, dvvh, dvvw, dvciii = s[data_vec].op.axis
cii, ciii, kh, kw = s[conv].op.reduce_axis
data_pad = data_vec.op.input_tensors[0]
if data_pad.op.name == "data_pad":
assert type(data_pad.op) == tvm.tensor.ComputeOp
has_padding = True
else:
pass
assert type(data_pad.op) == tvm.tensor.PlaceholderOp
has_padding = False
cfg.define_knob('data_pad_inline', [0, 1, 2, 3, 4])
if cfg['data_pad_inline'].val == 1 and has_padding:
s[data_pad].compute_inline()
if cfg['data_pad_inline'].val == 2 and has_padding:
s[data_pad].vectorize(list(s[data_pad].op.axis)[-1])
if cfg['data_pad_inline'].val == 3 and has_padding:
s[data_pad].vectorize(list(s[data_pad].op.axis)[-1])
s[data_pad].compute_at(s[data_vec], dvoh)
if cfg['data_pad_inline'].val == 4 and has_padding:
s[data_pad].vectorize(list(s[data_pad].op.axis)[-1])
s[data_pad].compute_at(s[data_vec], dvow)
cfg.define_knob('data_vec_inline', [0, 1, 2, 3])
if cfg['data_vec_inline'].val == 1:
s[data_vec].compute_at(s[conv], oh)
if cfg['data_vec_inline'].val == 2:
s[data_vec].compute_at(s[conv], ow)
if cfg['data_vec_inline'].val == 3:
s[data_vec].compute_at(s[conv], coo)
# schedule conv
cfg["reorder_0"].apply(s, conv, [n, coo, cii, oh, ow, kh, kw, vc, ciii, vh, vw])
cfg["ann_reduce"].apply(s, conv, [kh, kw, ciii],
axis_lens=[get_const_int(kh.dom.extent),
get_const_int(kw.dom.extent),
get_const_int(ciii.dom.extent)],
max_unroll=16,
cfg=cfg)
cfg["ann_spatial"].apply(s, conv, [vh, vw, vc],
axis_lens=[cfg['tile_oh'].size[-1],
cfg['tile_ow'].size[-1],
get_const_int(vc.dom.extent)],
max_unroll=16,
cfg=cfg)
s[conv].vectorize(vc)
# schedule fusion
n, coo, h, w, vc = s[last].op.axis
s[last].vectorize(vc)
oh, vh = cfg['tile_oh'].apply(s, last, h)
ow, vw = cfg['tile_ow'].apply(s, last, w)
cfg["reorder_1"].apply(s, last, [n, coo, oh, ow, vh, vw, vc])
if last != output:
s[output].compute_inline()
cfg["ann_spatial"].apply(s, last, [vh, vw, vc],
axis_lens=[cfg['tile_oh'].size[-1],
cfg['tile_ow'].size[-1],
get_const_int(vc.dom.extent)],
max_unroll=16,
cfg=cfg)
else:
# s[last].vectorize(vc)
pass
cfg.define_knob('conv_inline', [0, 1, 2, 3])
if cfg['conv_inline'].val == 1:
s[conv].compute_at(s[last], ow)
if cfg['conv_inline'].val == 2:
s[conv].compute_at(s[last], oh)
if cfg['conv_inline'].val == 3:
s[conv].compute_at(s[last], coo)
# s[conv].compute_at(s[last], ow)
_, _, _, _, vh, vw, vc = s[data_vec].op.axis
cfg["ann_spatial"].apply(s, data_vec, [vh, vw, vc],
axis_lens=[cfg['tile_oh'].size[-1],
cfg['tile_ow'].size[-1],
get_const_int(vc.dom.extent)],
max_unroll=16,
cfg=cfg)
s[data_vec].vectorize(vc)
s[data_vec].unroll(vw)
# s[data_pad].compute_inline()
# mark parallel
# s[last].parallel(co)
# # s[data_vec].parallel(h)
if kernel_vec.op.name == 'kernel_vec':
co, _, _, _, _ = s[kernel_vec].op.axis
s[kernel_vec].pragma(co, 'debug_skip_region')
if autotvm.GLOBAL_SCOPE.in_tuning:
# kernel packing will be pre-computed during compilation, so we skip
# this part to make tuning records correct
s[kernel_vec].pragma(co, 'debug_skip_region')
else:
pass
# import ipdb; ipdb.set_trace()
# print(tvm.lower(s, [data_pad, kernel_vec, last], simple_mode=True))
return s