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
|
import time
import sys
import subprocess
import numba as nb
import numpy as np
import awkward as ak
import awkward.forth
import fastparquet
import fastparquet.core
import fastparquet.thrift_structures
compress = sys.argv[1]
N = int(sys.argv[2])
is_split = sys.argv[3] == "split"
s = "-split" if is_split else ""
filename = f"/home/jpivarski/storage/data/chep-2021-jagged-jagged-jagged/{compress}{s}-jagged{N}.parquet"
subprocess.call(f"vmtouch -t {filename} > /dev/null", shell=True)
subprocess.call(f"vmtouch {filename} | fgrep Pages", shell=True)
file = open(filename, "rb")
parquetfile = fastparquet.ParquetFile(file)
if is_split:
def unsplit(data):
values = np.empty(len(data), np.uint8)
quarter = len(data) // 4
values[::4] = data[:quarter]
values[1::4] = data[quarter:2*quarter]
values[2::4] = data[2*quarter:3*quarter]
values[3::4] = data[3*quarter:]
return values.view(np.float32)
else:
def unsplit(data):
return data.view(np.float32)
def bit_width_of(value):
for i in range(0, 64):
if value == 0:
return i
value >>= 1
bit_width = bit_width_of(N)
if N > 0:
read_repetition_levels = awkward.forth.ForthMachine32("""
input stream
output reps uint8
stream I-> stack
begin
stream varint-> stack
dup 1 and 0= if
( run-length encoding )
stream {rle_format}-> reps
1 rshift 1-
reps dup
else
( bit-packed )
1 rshift 8 *
stream #{bit_width}bit-> reps
then
dup stream pos 4 - <=
until
""".format(bit_width=bit_width, rle_byte_width=1, rle_format="B"))
jagged1 = awkward.forth.ForthMachine32("""
input reps
output offsets0 int32
variable count0
begin
reps b-> stack
dup 1 = if
1 count0 +!
then
0 = if
count0 @ offsets0 +<- stack
1 count0 !
then
reps end
until
count0 @ offsets0 +<- stack
""")
jagged2 = awkward.forth.ForthMachine32("""
input reps
output offsets0 int32
output offsets1 int32
variable count0
variable count1
begin
reps b-> stack
dup 2 = if
1 count1 +!
then
dup 1 = if
count1 @ offsets1 +<- stack
1 count1 !
1 count0 +!
then
0 = if
count1 @ offsets1 +<- stack
1 count1 !
count0 @ offsets0 +<- stack
1 count0 !
then
reps end
until
count1 @ offsets1 +<- stack
count0 @ offsets0 +<- stack
""")
jagged3 = awkward.forth.ForthMachine32("""
input reps
output offsets0 int32 output offsets1 int32 output offsets2 int32
variable count0 variable count1 variable count2
begin
reps b-> stack
dup 3 = if
1 count2 +!
then
dup 2 = if
count2 @ offsets2 +<- stack
1 count2 !
1 count1 +!
then
dup 1 = if
count2 @ offsets2 +<- stack
1 count2 !
count1 @ offsets1 +<- stack
1 count1 !
1 count0 +!
then
0 = if
count2 @ offsets2 +<- stack
1 count2 !
count1 @ offsets1 +<- stack
1 count1 !
count0 @ offsets0 +<- stack
1 count0 !
then
reps end
until
count2 @ offsets2 +<- stack
count1 @ offsets1 +<- stack
count0 @ offsets0 +<- stack
""")
begintime = time.time()
for rg in parquetfile.filter_row_groups([]):
col = rg.columns[0]
file.seek(col.meta_data.data_page_offset)
ph = fastparquet.thrift_structures.read_thrift(file, fastparquet.thrift_structures.parquet_thrift.PageHeader)
raw_bytes = np.frombuffer(fastparquet.core._read_page(file, ph, col.meta_data), np.uint8)
data = raw_bytes[-(ph.data_page_header.num_values * 4):]
values = unsplit(data)
max_repetition_level = col.meta_data.path_in_schema.count("list")
assert max_repetition_level == N
bit_width = bit_width_of(max_repetition_level)
rep_length, = raw_bytes[:4].view(np.uint32)
if N != 0:
read_repetition_levels.run({"stream": raw_bytes})
repetition_levels = np.asarray(read_repetition_levels["reps"])[:len(values)]
if N == 0:
array = ak.layout.NumpyArray(values)
elif N == 1:
jagged1.run({"reps": repetition_levels})
array = ak.layout.ListOffsetArray32(
ak.layout.Index32(jagged1["offsets0"]),
ak.layout.NumpyArray(values)
)
elif N == 2:
jagged2.run({"reps": repetition_levels})
array = ak.layout.ListOffsetArray32(
ak.layout.Index32(jagged2["offsets0"]),
ak.layout.ListOffsetArray32(
ak.layout.Index32(jagged2["offsets1"]),
ak.layout.NumpyArray(values)
)
)
elif N == 3:
jagged3.run({"reps": repetition_levels})
array = ak.layout.ListOffsetArray32(
ak.layout.Index32(jagged3["offsets0"]),
ak.layout.ListOffsetArray32(
ak.layout.Index32(jagged3["offsets1"]),
ak.layout.ListOffsetArray32(
ak.layout.Index32(jagged3["offsets2"]),
ak.layout.NumpyArray(values)
)
)
)
tmp = ak.Array(array)
endtime = time.time()
print(f"AwkwardForth {compress}{s}-jagged{N}", endtime - begintime, "seconds")
|
