forked from josiahw/IsomapCUDA
/
KNearestNeighbours.py
196 lines (166 loc) · 8.67 KB
/
KNearestNeighbours.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
####################################################################################################################################################
#Copyright (c) 2013, Josiah Walker
#All rights reserved.
#
#Redistribution and use in source and binary forms, with or without modification, are permitted provided that the following conditions are met:
#
# Redistributions of source code must retain the above copyright notice, this list of conditions and the following disclaimer.
# Redistributions in binary form must reproduce the above copyright notice, this list of conditions and the following disclaimer in the documentation and/or #other materials provided with the distribution.
#
#THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED #WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY #DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS #OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING #NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
####################################################################################################################################################
"""
GPU based K nearest neighbours algorithm.
"""
import time
from numpy import array,zeros,amax,amin,sqrt,dot,random
import numpy
from numpy.linalg import eig
import pycuda.autoinit
import pycuda.driver as drv
from pycuda.compiler import SourceModule
import math
from DataUtils import dataConfig,loadTable,loadSplitTable,loadMatrix,loadIntMatrix,saveTable
KernelLocation = "CudaKernels/KNN/"
# KNN Algorithm --------------------------------------------
def GPUConfig(settings,memPerElement,memPerElementSq,limit=1000000000):
return settings
def KNNConfig(dataTable,srcDims, k, eps = 1000000000.,gpuMemSize = 512, settings = {}):
"""
Creates all the memory/data settings to run GPU accelerated KNN.
"""
kmax = 2**int(math.ceil(math.log(k+1)/math.log(2)))+1
settings["oldk"] = k
settings["kmax"] = kmax
k = kmax
settings = dataConfig(dataTable,settings)
settings["sourceDims"] = min(settings["sourceDims"],srcDims)
#XXX: determine memory and thread sizes from device
settings["memSize"] = gpuMemSize*1024*1024
settings["maxThreads"] = 1024
#set up chunk sizes
memoryPerElement = k*4*2+kmax*4*2 + (settings["sourceDims"]*4)*2 + 30 #+ (k+10)*4*2 #this is an estimated memory used per element
MemoryPerElementSquared = 8*2
ctr = 0
while memoryPerElement*ctr + ctr*ctr*8*2 < settings["memSize"] and ctr*settings["sourceDims"]<1000000:
ctr += 1
ctr -= 1
#settings = GPUConfig(settings,memoryPerElement,memoryPerElementsquared,1000000/settings["sourceDims"])
settings["chunkSize"] = min(ctr,settings["dataLength"])
settings["lastChunkSize"] = ((settings["dataLength"]-1) % settings["chunkSize"]) + 1
#create kernel gridsize tuples
settings["block"] = (settings["maxThreads"],1,1)
settings["grid"] = ((settings["chunkSize"]/settings["maxThreads"])+(settings["chunkSize"]%settings["maxThreads"]>0),1,1)
#precalculate all constant kernel params
settings["dimensions"] = numpy.int64(settings["sourceDims"])
settings["k"] = numpy.int64(k)
settings["eps"] = numpy.float32(eps)
settings["dataSize"] = numpy.int64(settings["dataLength"])
settings["chunkSize"] = numpy.int64(settings["chunkSize"])
settings["maxThreads"] = numpy.int64(settings["maxThreads"])
return settings
def KNN(dataTable, k, epsilon=10000000000000., srcDims = 1000000000000000, normData = False):
"""
Get a k,epsilon version k nearest neighbours
"""
#load up the configuration
knnOptions = KNNConfig(dataTable,srcDims,k,epsilon)
#load and format the table for use.
data = loadTable(dataTable,knnOptions)
#check if we should normalise the data (this is really quick and dirty, replace it with something better)
if normData:
dmax = max([amax(d) for d in data])
dmin = max([amin(d) for d in data])
data = [(d-dmin)/(dmax-dmin+0.00000001) for d in data]
#create the CUDA kernels
header = "#define MAXKMAX ("+str(knnOptions['kmax'])+")\n#define CHUNKSIZE ("+str(knnOptions['chunkSize'])+")\n"
program = SourceModule(header+open(KernelLocation+"KNN.nvcc").read())
prg = program.get_function("KNN")
sortprogram = SourceModule(header+open(KernelLocation+"KNNSORT.nvcc").read())
sort = sortprogram.get_function("KNNSORT")
t0 = time.time()
#make a default distance list
distances0 = (zeros((knnOptions['chunkSize']*knnOptions['k'])) + knnOptions['eps']).astype(numpy.float32)
indices0 = zeros((knnOptions['chunkSize']*knnOptions['k'])).astype(numpy.uint32)
dists = [distances0.copy() for i in xrange(len(data))]
indices = [indices0.copy() for i in xrange(len(data))]
#calculate KNN
offset = 0
source_gpu = drv.mem_alloc(data[0].nbytes)
target_gpu = drv.mem_alloc(data[0].nbytes)
indices_gpu = drv.mem_alloc(indices[0].nbytes)
dists_gpu = drv.mem_alloc(dists[0].nbytes)
data = [s.T.flatten() for s in data]
for source in data:
drv.memcpy_htod(source_gpu, source)
drv.memcpy_htod(indices_gpu, indices[offset])
drv.memcpy_htod(dists_gpu, dists[offset])
for t in xrange(len(data)):
drv.memcpy_htod(target_gpu, data[t])
prg(source_gpu,
target_gpu,
indices_gpu,
dists_gpu,
knnOptions["dimensions"],
knnOptions['k'],
knnOptions['eps'],
knnOptions['dataSize'],
knnOptions['chunkSize'],
numpy.int64(offset*knnOptions['chunkSize']),
numpy.int64(t*knnOptions['chunkSize']),
block=knnOptions['block'],
grid=knnOptions['grid'])
sort(dists_gpu,
indices_gpu,
numpy.int64(knnOptions['oldk']),
knnOptions['dataSize'],
knnOptions['chunkSize'],
numpy.int64(offset*knnOptions['chunkSize']),
block=knnOptions['block'],
grid=knnOptions['grid'])
drv.memcpy_dtoh(indices[offset], indices_gpu)
drv.memcpy_dtoh(dists[offset], dists_gpu)
offset += 1
del source_gpu
del indices_gpu
del dists_gpu
del target_gpu
#organise data and add neighbours
alldists = numpy.concatenate(dists).reshape((-1,knnOptions['k']))[:(knnOptions['dataSize']),:knnOptions['oldk']].tolist()
allindices = numpy.concatenate(indices).reshape((-1,knnOptions['k']))[:(knnOptions['dataSize']),:knnOptions['oldk']].tolist()
for i in xrange(len(alldists)): #remove excess entries
if knnOptions['eps'] in alldists[i]:
ind = alldists[i].index(knnOptions['eps'])
alldists[i] = alldists[i][:ind]
allindices[i] = allindices[i][:ind]
ind = allindices[i].index(i)
alldists[i] = alldists[i][:ind]
allindices[i] = allindices[i][:ind]
for i in xrange(len(alldists)):
j = 0
for ind in allindices[i]: #add mirrored entries
if not (i in allindices[ind]):
allindices[ind].append(i)
alldists[ind].append(alldists[i][j])
j += 1
#have a list for start and end indices for knn
klist = [0]
for i in xrange(len(alldists)): #pad all entries to the same length for the next algorithm
klist.append(len(alldists[i])+klist[-1])
alldists = array([j for i in alldists for j in i],dtype=numpy.float32)
allindices = array([j for i in allindices for j in i],dtype=numpy.uint32)
klist = array(klist,dtype=numpy.uint32)
print time.time()-t0, " seconds to process KNN"
"""
#save the link matrix to compare to the matlab results
f = open("linkmatrix.csv",'w')
mat = [[0]*len(allindices) for i in xrange(len(allindices))]
for i in xrange(len(allindices)):
for j in xrange(len(allindices[i])):
if alldists[i][j] < knnOptions['eps']:
mat[i][allindices[i][j]] = mat[allindices[i][j]][i] = 1
for m in mat:
f.write(str(m).strip('[]')+'\n')
f.close()
"""
return allindices, alldists, klist #[allindices[i]+alldists[i] for i in xrange(len(alldists))]