-
Notifications
You must be signed in to change notification settings - Fork 75
/
Copy pathodla.py
229 lines (205 loc) · 12.4 KB
/
odla.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
# Copyright (C) 2019-2021 Alibaba Group Holding Limited.
#
# Licensed 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.
from ctypes import *
from enum import Enum
from time import time
import logging
import os
from pathlib import Path
from halo import halo
class Device(Enum):
CUDA = 1
IPU = 2
X86 = 3
class ValueShape(Structure):
_fields_ = [("size", c_int32), ("dims", c_int64 * 10)]
class ValueType(Structure):
_fields_ = [("element_type", c_int32), ("shape", ValueShape)]
class ODLAModel:
def __init__(self, so_file):
self.logger = logging.getLogger(__name__)
self.so_file = so_file
self.h = None
self.save_temps = False
self.intermediate_files = []
def __del__(self):
self.logger.info(str(self.intermediate_files))
for file in self.intermediate_files:
if not self.save_temps:
Path(file).unlink()
if self.h is not None:
self.h.odla_DestroyContext(self.ctx)
self.h.odla_DestroyComputation(self.comp)
self.h.odla_DestroyDevice(self.device)
def Load(self,model,input_shapes,output_names,format,batch,qps):
if self.h is None:
self.h = CDLL(self.so_file)
self.comp = c_void_p(0)
self.device = c_void_p(0)
model_info = halo.ModelInfo()
model_info.input_qps = qps
model_info.adaptive_bsz = batch
rsc_est = c_void_p(0)
if qps>0 and batch==1:
halo.AnalyzeModel(model,input_shapes,1,format,model_info)
rsc_est = (c_char_p)(model_info.output_rsc_est)
self.h.odla_AllocateDevice(c_void_p(0), 0, pointer(self.device), rsc_est)
self.files = halo.CompileModel(
model,
input_shapes,
output_names,
batch,
format,
)
self.intermediate_files = [*self.files]
cc_file = str(self.files[0]).encode("utf-8")
bin_file = str(self.files[1]).encode("utf-8")
self.comp = self.h.model_helper(cc_file, bin_file)
self.ctx = c_void_p(0)
self.h.odla_CreateContext(pointer(self.ctx))
n = c_int32(-1)
self.h.odla_GetNumOfArgsFromComputation(self.comp, pointer(n))
# self.nr_args = n.value
if("bert" in model):
self.nr_args = 3
elif("shoucai" in model):
self.nr_args = 16
else:
self.nr_args = 1
# nr_args = c_int32(-1)
self.h.odla_GetNumOfOutputsFromComputation(self.comp, pointer(n))
# self.nr_outputs = n.value
if("bert" in model):
self.nr_outputs = 2
else:
self.nr_outputs = 1
self.in_vals = []
for idx in range(0, self.nr_args):
arg_v = c_void_p(0)
self.h.odla_GetArgFromComputationByIdx(self.comp, idx, pointer(arg_v))
self.out_vals = []
for idx in range(0, self.nr_outputs):
out = c_void_p(0)
self.h.odla_GetOutputFromComputationByIdx(self.comp, idx, pointer(out))
def Execute(self, data, model, batch):
print(f"model:{model},batch:{batch}")
# bind input
if("bert" in model):
self.h.odla_BindToArgument(c_void_p(0), data[0].ctypes.data_as(c_void_p), self.ctx)
self.h.odla_BindToArgument(c_void_p(1), data[1].ctypes.data_as(c_void_p), self.ctx)
self.h.odla_BindToArgument(c_void_p(2), data[2].ctypes.data_as(c_void_p), self.ctx)
elif("shoucai" in model):
self.h.odla_BindToArgument(c_void_p(0), data[0].ctypes.data_as(c_void_p), self.ctx)
self.h.odla_BindToArgument(c_void_p(1), data[1].ctypes.data_as(c_void_p), self.ctx)
self.h.odla_BindToArgument(c_void_p(2), data[2].ctypes.data_as(c_void_p), self.ctx)
self.h.odla_BindToArgument(c_void_p(3), data[3].ctypes.data_as(c_void_p), self.ctx)
self.h.odla_BindToArgument(c_void_p(4), data[4].ctypes.data_as(c_void_p), self.ctx)
self.h.odla_BindToArgument(c_void_p(5), data[5].ctypes.data_as(c_void_p), self.ctx)
self.h.odla_BindToArgument(c_void_p(6), data[6].ctypes.data_as(c_void_p), self.ctx)
self.h.odla_BindToArgument(c_void_p(7), data[7].ctypes.data_as(c_void_p), self.ctx)
self.h.odla_BindToArgument(c_void_p(8), data[8].ctypes.data_as(c_void_p), self.ctx)
self.h.odla_BindToArgument(c_void_p(9), data[9].ctypes.data_as(c_void_p), self.ctx)
self.h.odla_BindToArgument(c_void_p(10), data[10].ctypes.data_as(c_void_p), self.ctx)
self.h.odla_BindToArgument(c_void_p(11), data[11].ctypes.data_as(c_void_p), self.ctx)
self.h.odla_BindToArgument(c_void_p(12), data[12].ctypes.data_as(c_void_p), self.ctx)
self.h.odla_BindToArgument(c_void_p(13), data[13].ctypes.data_as(c_void_p), self.ctx)
self.h.odla_BindToArgument(c_void_p(14), data[14].ctypes.data_as(c_void_p), self.ctx)
self.h.odla_BindToArgument(c_void_p(15), data[15].ctypes.data_as(c_void_p), self.ctx)
else:
self.h.odla_BindToArgument(c_void_p(0), data[0].ctypes.data_as(c_void_p), self.ctx)
# output buffer
buffers = []
if("bert" in model):
buf1 = (c_float * 256 * batch)()
buffers.append(buf1)
self.h.odla_BindToOutput(c_void_p(0), buf1, self.ctx)
buf2 = (c_float * 256 * batch)()
buffers.append(buf2)
self.h.odla_BindToOutput(c_void_p(1), buf2, self.ctx)
else:
if("resnet50" in model):
buf = (c_float * 1*1000 * batch)()
elif("dbnet" in model):
assert((batch==1) and "dbnet only support 1 batch.")
buf = (c_float * 1228800 * batch)()
elif("crnn" in model):
assert((batch==1) and "crnn only support 1 batch.")
buf = (c_float * 918146 * batch)()
elif("shoucai" in model):
#Add info for Shoucai Model
assert((batch==1) and "shoucai only support 1 batch!")
buf = (c_float * 425 * batch)()
else:
assert(False and f"unknown model.{model}")
buffers.append(buf)
self.h.odla_BindToOutput(c_void_p(0), buf, self.ctx)
# send data
if("resnet50" in model):
self.h.model_data(self.ctx, (c_int32 * 1)(*[224*224*3*4*batch]), (c_int32 * 1)(*[1000*4*batch]))
elif("dbnet" in model):
self.h.model_data(self.ctx, (c_int32 * 1)(*[1*3*960*1280*4*batch]), (c_int32 * 1)(*[1228800 * 4*batch]))
elif("crnn" in model):
self.h.model_data(self.ctx, (c_int32 * 1)(*[63840*4*batch]), (c_int32 * 1)(*[918146*4*batch]))
elif("bert" in model):
self.h.model_data(self.ctx, (c_int32 * 3)(*[512*4*batch, 256*4*batch, 256*8*batch]), (c_int32 * 2)(*[256*4*batch,256*4*batch]))
elif ("shoucai" in model):
#per Input File
#self.h.model_data(self.ctx, (c_int32 * 16)(*[8*4*batch, 1*4*batch, 592*4*batch, 1*4*batch, 512*4*batch, 512*4*batch, 73728*4*batch, 27200*4*batch, 13600*4*batch, 13600*4*batch, 122400*4*batch, 350200*4*batch, 15552*4*batch, 1161984*4*batch, 178*4*batch, 425*8*batch]), (c_int32*1)(*[425*4*batch]))
#per Input Shape
self.h.model_data(self.ctx, (c_int32 * 18)(*[425*4*batch, 8*4*batch, 1*4*batch, 592*4*batch, 1*4*batch, 512*4*batch, 512*4*batch, 73728*4*batch, 425*4*batch, 425*4*batch, 27200*4*batch,13600*4*batch, 13600*4*batch, 122400*4*batch, 350200*4*batch, 15552*4*batch, 1161984*4*batch, 178*4*batch]), (c_int32*1)(*[425*4*batch]))
'''
self.h.model_data(self._ctx, (c_int32 * 16)(*[8*4*batch, \ # embedding_ui_oage_shared_embedding.txt
1*4*batch, \ # input_from_feature_columns_concat_3.txt
592*4*batch, \ # input_from_feature_columns_concat.txt
1*4*batch, \ #input_from_feature_columns_concat_5.txt
512*4*batch, \ #all_clk_seq_1_time.txt
512*4*batch, \#all_clk_seq_1_st.txt
73728*4*batch, \ #seq_input_from_feature_columns_concat_1.txt
27200*4*batch, \#embedding_item_id_d_shard_embedding_2.txt
13600*4*batch, \#embedding_item_cate_id_d_shared_embedding_2.txt
13600*4*batch, \#embedding_item_seller_id_d_shared_embedding_2.txt
122400*4*batch, \#input_from_feature_columns_concat_4.txt
350200*4*batch, \#input_from_feature_columns_concat_1.txt
15552*4*batch, \#seq_input_from_feature_columns_concat.txt
1161984*4*batch,\ #seq_input_from_feature_columns_concat_2.txt
178*4*batch, \#input_from_feature_columns_concat_7.txt
425*8*batch]), \#Unique_preprocess_int64.txt
(c_int32*1)(*[425*4*batch])) #output
self.h.model_data(self._ctx, (c_int32 * 18)(*[425*4*batch, \# LookupPkOP
8*4*batch, \ # embedding_ui_oage_shared_embedding.txt
1*4*batch, \ # input_from_feature_columns_concat_3.txt
592*4*batch, \ # input_from_feature_columns_concat.txt
1*4*batch, \ #input_from_feature_columns_concat_5.txt
512*4*batch, \ #all_clk_seq_1_time.txt
512*4*batch, \#all_clk_seq_1_st.txt
73728*4*batch, \ #seq_input_from_feature_columns_concat_1.txt
425*4*batch, \ #batch_fill_attributes_for_gul_rank_item_feature
425*4*batch， \ #batch_fill_attributes_for_gul_rank_item_feature_1
27200*4*batch, \#embedding_item_id_d_shard_embedding_2.txt
13600*4*batch, \#embedding_item_cate_id_d_shared_embedding_2.txt
13600*4*batch, \#embedding_item_seller_id_d_shared_embedding_2.txt
122400*4*batch, \#input_from_feature_columns_concat_4.txt
350200*4*batch, \#input_from_feature_columns_concat_1.txt
15552*4*batch, \#seq_input_from_feature_columns_concat.txt
1161984*4*batch,\ #seq_input_from_feature_columns_concat_2.txt
178*4*batch]), \#input_from_feature_columns_concat_7.txt
(c_int32*1)(*[425*4*batch])) #output
'''
else:
assert(False and f"unknown model.{model}")
s = time()
self.h.odla_ExecuteComputation(self.comp, self.ctx, 0, self.device)
t = time()
self.logger.info("Execution time:" + str(t - s) + " sec(s)")
return buffers