/
bitonic_sort.py
161 lines (126 loc) · 5.24 KB
/
bitonic_sort.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
import pyopencl as cl
import numpy as np
import os
class BitonicSort(object):
def __init__(self, ctx, queue, local_size_limit=None):
self.ctx = ctx
if local_size_limit is None:
local_size_limit = min(device.max_work_group_size for device in self.ctx.devices)//2
self.queue = queue
cur_dir = os.path.dirname(os.path.abspath(__file__))
with open('%s/BitonicSort_b.cl' % cur_dir) as f:
self.prg = cl.Program(ctx, f.read() % { 'local_size_limit': local_size_limit }).build()
self.local_size_limit = local_size_limit
def sort_in_place(self, d_key, d_val, array_length, dir):
self.sort(d_key, d_val, d_key, d_val, array_length, dir)
def sort(self, d_dst_key, d_dst_val, d_src_key, d_src_val, array_length, dir):
if array_length < self.local_size_limit:
batch = self.local_size_limit // array_length
else:
batch = 1
self._sort(d_dst_key, d_dst_val, d_src_key, d_src_val, batch, array_length, dir)
def _sort(self, d_dst_key, d_dst_val, d_src_key, d_src_val, batch, array_length, dir):
"""
dir: 0 for descending sort, 1 for ascending.
"""
assert array_length >= 2
queue = self.queue
prg = self.prg
local_size_limit = self.local_size_limit
def int_log2(L):
if not L:
return 0
log2 = 0
while L & 1 == 0:
L >>= 1
log2 += 1
return log2
# only power-of-two array lengths are supported
log2L = int_log2(array_length)
assert 2**log2L == array_length
if array_length <= local_size_limit:
assert (batch * array_length) % local_size_limit == 0
local_work_size = (local_size_limit // 2, )
global_work_size = (batch * array_length // 2, )
kernel_args = (
d_dst_key,
d_dst_val,
d_src_key,
d_src_val,
np.uint32(array_length),
np.uint32(dir),
)
prg.bitonicSortLocal(queue, global_work_size, local_work_size, *kernel_args)
queue.finish()
else:
# launch bitonicSortLocal1
local_work_size = (local_size_limit // 2, )
global_work_size = (batch * array_length // 2, )
kernel_args = (
d_dst_key,
d_dst_val,
d_src_key,
d_src_val,
)
prg.bitonicSortLocal1(queue, global_work_size, local_work_size, *kernel_args)
queue.finish()
size = 2*local_size_limit
while size <= array_length:
stride = size // 2
while stride > 0:
if stride >= local_size_limit:
# launch bitonicMergeGlobal
local_work_size = (local_size_limit // 4, )
global_work_size = (batch * array_length // 2, )
kernel_args = (
d_dst_key,
d_dst_val,
d_src_key,
d_src_val,
np.uint32(array_length),
np.uint32(size),
np.uint32(stride),
np.uint32(dir),
)
prg.bitonicMergeGlobal(queue, global_work_size, local_work_size, *kernel_args)
queue.finish()
else:
# launch bitonicMergeLocal
local_work_size = (local_size_limit // 2, )
global_work_size = (batch * array_length // 2, )
kernel_args = (
d_dst_key,
d_dst_val,
d_src_key,
d_src_val,
np.uint32(array_length),
np.uint32(stride),
np.uint32(size),
np.uint32(dir),
)
prg.bitonicMergeLocal(queue, global_work_size, local_work_size, *kernel_args)
queue.finish()
stride >>= 1
size <<= 1
if __name__ == '__main__':
N = 2<<20
keys = np.random.randint(1, 300, size=N).astype(np.uint32)
vals = np.random.randint(1, 300, size=N).astype(np.uint32)
indices = keys.argsort()
sorted_keys = keys[indices]
sorted_vals = keys[indices]
ctx = cl.create_some_context()
queue = cl.CommandQueue(ctx)
mf = cl.mem_flags
d_keys = cl.Buffer(ctx, mf.READ_ONLY | mf.COPY_HOST_PTR, hostbuf=keys)
d_vals = cl.Buffer(ctx, mf.READ_ONLY | mf.COPY_HOST_PTR, hostbuf=vals)
s = BitonicSort(ctx, queue)
from time import time
t = time()
s.sort_in_place(d_keys, d_vals, N, 1)
print("%ims" % ((time()-t)*1000))
cl.enqueue_copy(queue, keys, d_keys)
cl.enqueue_copy(queue, vals, d_vals)
queue.finish()
print(np.linalg.norm(keys-sorted_keys))
print(np.linalg.norm(vals-sorted_vals))