forked from dgasmith/opt_einsum
-
Notifications
You must be signed in to change notification settings - Fork 0
/
setperm.py
344 lines (293 loc) · 11.2 KB
/
setperm.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
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
import numpy as np
import record
import math
#from numba import cuda, float32, int32
import permutate
import checkperm
#A: device array input, permutation: cpu int array, Ap: device array for output, itemsize: byte size of element in A
#index_from/to: einsum notation string
#A != Ap: no inline permutation. Ap.shape is ignored (no need to reshape to the permutation order)
def perm( A, index_from, Ap, index_to ) :
#I don't have time to think through the trivial case failure (probably assuming inner/outer have at least one difference)
if index_from == index_to :
return
# assert( A.gpu_data != Ap.gpu_data ) #can't use this while in simulator
itemsize = A.dtype.itemsize
T = 20
# rN = 5
rN = len(index_from)
# out = np.zeros(rN, dtype=int)
out = parse( index_from, index_to )
# out[:] = [3,1,2,0,4] #outvirt didn't look right
shape = A.shape
# shape = np.zeros(rN, dtype=int)
#out[:] = [4,1,2,3,0]
# shape[:] = [6,40,10,3,3]
#T = 20
#rN = 2
#out = np.zeros(rN, dtype=int)
##out[:] = [3,1,2,0,4] outvirt didn't look right
#shape = np.zeros(rN, dtype=int)
#out[:] = [1,0]
#shape[:] = [100,100]
shapeOut = np.zeros(rN, dtype=int)
for i in range(rN) :
shapeOut[i] = shape[ out[i] ]
#debug: ok stupid thing is I can't reshape Ap because it sends a result, not inline.
Apr = Ap.reshape( tuple(shapeOut), order='C' )
N = 1
for i in range(rN) :
N = N * shape[i]
if N < T :
T = N
#order not relevant for flat array, but original order is for shape to be strided properly in permutation
#note ravel() will copy data if order is not same as stored order, while reshape will simply stride data differently.
#can't access _dummy in cuda simulator
# assert( A._dummy.flags['C_CONTIGUOUS'] ) #!for now. Could easily adjust striding in permutation
fA = A.reshape( N, order='C' )
fAp = Ap.reshape( N, order='C' )
# A.shape = (N,)
# Ap.shape = (N,)
# A.strides = (itemsize,)
# Ap.strides = (itemsize,)
overlap = np.zeros(rN, dtype=int)
#that's the specs.
#determine inner cubes according to threads T
shapeVirt = np.zeros( rN+2, dtype=int )
mapv = np.zeros( rN + 1, dtype=int ) #with dummy element. mapv not passed to kernel
mapv[rN] = rN + 2
mapr = np.zeros( rN+2, dtype=int ) #reverse map from virt to real with split indicators
i = 1; m = 1
# while T // m != 0 and i <= rN :
while 1.0 * T / m > 1.0 and i <= rN :
inrc = m
m = m * shape[rN-i]
i = i + 1
inrN = i - 1 #inner read cube size
assert( inrN <= rN )
i = 1; m = 1
# while T // m != 0 and i <= rN :
while 1.0 * T / m > 1.0 and i <= rN :
inwc = m
m = m * shape[out[rN-i]]
i = i + 1
inwN = i - 1 #inner write cube size
#mode overlap: the inner always owns the overlap
#!also redundant with other block that uses it. only need virtual overlap
for i in range(1,inwN+1) :
if out[rN-i] >= rN-inrN :
overlap[ out[rN-i] ] = 1
#split overlap
if out[rN-inwN] == rN-inrN :
if inwc > inrc : #write split has lower fill/happens first
s1s = T // inwc #first split dimension is how many times T fills dim-1 inner write cube, ie less than read cube
s2s = T // (s1s*inrc) #second split dim how many times T fills dim-1 inner read cube w 1st split (at least >= 1X)
s3s = math.ceil( 1.0 * shape[rN-inrN] / (s2s*s1s) ) #closing virtual buffer dimension to cover whole index
#inner write cube: s1s * inwc. inner read cube: s2s*s1s*inrc (either may be the larger still)
#the final thread count T' is the max of the two, and threads in diff have to be masked out.
Tf = max( s1s*inwc, s2s*s1s*inrc )
innerReadN = inrN + 1; innerRead = np.zeros( innerReadN, dtype=int )
innerWriteN = inwN; innerWrite = np.zeros( innerWriteN, dtype=int )
innerRead[0] = rN-inrN + 1 #second piece
innerRead[1] = rN-inrN + 2 #third piece
innerWrite[0] = rN-inwN + 2 #reflects the split ind -> ind, ind+1, ind+2
readExtra, writeExtra = 1, 0
else : #inwc <= inrc:
s1s = T // inrc
s2s = T // (s1s*inwc)
s3s = math.ceil( 1.0 * shape[rN-inrN] / (s2s*s1s) ) #closing virtual buffer dimension to cover whole index
Tf = max( s1s*inrc, s2s*s1s*inwc )
innerReadN = inrN; innerRead = np.zeros( innerReadN, dtype=int )
innerWriteN = inwN + 1; innerWrite = np.zeros( innerWriteN, dtype=int )
innerWrite[0] = rN-inrN + 1
innerWrite[1] = rN-inrN + 2
innerRead[0] = rN-inrN + 2
readExtra, writeExtra = 0, 1
for i in range(rN-inrN) :
mapv[i] = i
mapr[i] = i
shapeVirt[i] = shape[i]
mapv[rN-inrN] = rN-inrN
mapr[rN-inrN] = -2
mapr[rN-inrN +1] = -1
mapr[rN-inrN +2] = rN-inrN
shapeVirt[rN-inrN] = s3s
shapeVirt[rN-inrN +1] = s2s
shapeVirt[rN-inrN +2] = s1s
for i in range(rN-inrN+1, rN) :
mapv[i] = i+2
mapr[i+2] = i
shapeVirt[i+2] = shape[i]
for i in range(1,inrN) : #remaining non-split indices
innerRead[ readExtra + i ] = (rN-inrN)+i + 2 #+2 for virtual index shift after triple split
for i in range(1,inwN) :
if out[ (rN-inwN)+i ] < rN-inrN : #before split means no shift
innerWrite[ writeExtra + i ] = out[ (rN-inwN)+i ]
else :
innerWrite[ writeExtra + i ] = out[ (rN-inwN)+i ] + 2
else : #separate splits
srs = T // inrc
scr = math.ceil( 1.0 * shape[rN-inrN] / srs )
sws = T // inwc
scw = math.ceil( 1.0 * shape[out[rN-inwN]] / sws )
Tf = max( srs*inrc, sws*inwc )
#does one fully contain the other split?
#!this block is redundant. Below can just populate the inner arrays and set N after fact
if overlap[ out[rN-inwN] ] : #inner read contains write split
innerReadN = inrN + 1;
else :
innerReadN = inrN;
innerRead = np.zeros( innerReadN, dtype=int )
if overlap[ rN-inrN ] : #inner write contains read split
innerWriteN = inwN + 1;
else :
innerWriteN = inwN;
innerWrite = np.zeros( innerWriteN, dtype=int )
if rN-inrN < out[rN-inwN] : #which split comes first
s1, s2 = rN-inrN, out[rN-inwN]
s1o, s1c, s2o, s2c = scr, srs, scw, sws
else :
s2, s1 = rN-inrN, out[rN-inwN]
s2o, s2c, s1o, s1c = scr, srs, scw, sws #split open close
for i in range( s1 ) : #to first split
mapv[i] = i
mapr[i] = i
shapeVirt[i] = shape[i]
mapv[s1] = s1
mapr[s1] = -1
mapr[s1+1] = s1
shapeVirt[s1] = s1o
shapeVirt[s1+1] = s1c
for i in range( s1+1, s2 ) : #to second split
mapv[i] = i+1
mapr[i+1] = i
shapeVirt[i+1] = shape[i]
mapv[s2] = s2+1
mapr[s2+1] = -1
mapr[s2+2] = s2
shapeVirt[s2+1] = s2o
shapeVirt[s2+2] = s2c
for i in range(s2+1, rN) :
mapv[i] = i+2
mapr[i+2] = i
shapeVirt[i+2] = shape[i]
#first index is split, need right half. Others may include other split.
innerRead[0] = mapv[ rN-inrN ] +1 #second half
j = 1
for i in range(1,inrN) : #remaining
vi = mapv[ (rN-inrN)+i ]
while vi != mapv[ (rN-inrN)+i +1 ] :
innerRead[j] = vi
j = j + 1
vi = vi + 1
innerWrite[0] = mapv[ out[rN-inwN] ] +1 #second half
j = 1
for i in range(1,inwN) :
vi = mapv[ out[ (rN-inwN)+i ] ]
while vi != mapv[ out[ (rN-inwN)+i ] +1 ] :
innerWrite[j] = vi
j = j + 1
vi = vi + 1
#!j has the right length could just set inner N here.
#virtual overlap of inner, and union
overlapVirt = np.zeros( rN+2, dtype=int )
unionVirt = np.zeros( rN+2, dtype=int )
for i in range( innerWriteN ) :
if innerWrite[i] >= rN+2 - innerReadN :
overlapVirt[ innerWrite[i] ] = 1
unionVirt[ innerWrite[i] ] = 1
for i in range( innerReadN ) :
unionVirt[ innerRead[i] ] = 1
#block: virtual modes not contained in inners
block = np.zeros( rN+2, dtype=int )
j = 0
for i in range( rN+2 ) :
if not unionVirt[i] :
block[j] = i
j = j + 1
blockN = j
#outers are the complements: outer read is inner write - inner read, outer write is inner read - inner write
#also outer order of modes doesn't matter. But just in case there is dependency on first mode being a split
outerRead = np.zeros( rN+2, dtype=int )
j = 0
for i in range( innerWriteN ) :
if not overlapVirt[ innerWrite[i] ] :
outerRead[j] = innerWrite[i]
j = j + 1
outerReadN = j
outerWrite = np.zeros( rN+2, dtype=int )
j = 0
for i in range( innerReadN ) :
if not overlapVirt[ innerRead[i] ] :
outerWrite[j] = innerRead[i]
j = j + 1
outerWriteN = j
#output virtual: scan out along with mapv
outVirt = np.zeros( rN+2, dtype=int )
j = 0
for i in range( rN ) :
k = mapv[ out[i] ]
while k != mapv[ out[i] + 1] :
outVirt[j] = k
k = k + 1
j = j + 1
assert( j == rN+2 )
#--------------------------------
#ok send a sample matrix
# A = np.arange(N, dtype='float32') #.reshape(10,10,10,10)
# C = np.zeros(N, dtype='float32') #.reshape(10,10,10,10)
# Ad = cuda.to_device(A)
# Cd = cuda.to_device(C)
p = record.parg( innerReadN=innerReadN, outerReadN=outerReadN, innerWriteN=innerWriteN, outerWriteN=outerWriteN
, blockN=blockN, virtualN=rN+2, realN=rN )
p['shape'] = shape
p['shapeVirt'] = shapeVirt
p['mapVirt'] = mapr #right virt is mapped to orig, left to inner=-1/outer=-2 virt
ps = p['shape']
#si = A.dtype.itemsize
si = 1
#p['strides'] = [40000,4000,400,40,4]
p['strides'][rN-1] = si
for i in range( rN-2,-1,-1 ) :
p['strides'][i] = p['shape'][i+1] * p['strides'][i+1]
p['T'] = Tf #!!mask out the smaller threads in gpu !!also: -1/-2 meaning change, only -2 when left double split
p['out'] = out
p['outVirt'] = outVirt
#inner/outer virt are the real indices that get split
outerYet = False
for i in range( rN+2 ) :
if mapr[i] < 0 :
if not outerYet :
outerVirt, outerYet = mapr[ i-mapr[i] ], True
else :
assert( mapr[i] == -1 )
innerVirt = mapr[i+1]
p['innerVirt'] = innerVirt #wrt real
p['outerVirt'] = outerVirt
p['innerRead'] = innerRead[:innerReadN] #first one always virtual.
p['outerRead'] = outerRead[:outerReadN] #first one always virtual (possibly on top of virtual, or last one is virtual.
p['innerWrite'] = innerWrite[:innerWriteN]
p['outerWrite'] = outerWrite[:outerWriteN]
p['block'] = block[:blockN] #order arbitrary
blocks = 1
for i in range( blockN ) :
blocks = blocks * shapeVirt[ block[i] ]
#griddim (number of blocks), blockdim (threads per block), stream, sharedmem
from pdb import set_trace; set_trace()
#permutate.perm[ blocks, Tf, 0, Tf*Tf * A.dtype.itemsize ]( Ad, Cd, p) #Tf*Tf is upper bound when no overlap
permutate.perm[ blocks, Tf, 0, Tf*Tf * itemsize ]( fA, fAp, p) #Tf*Tf is upper bound when no overlap
#debug
Ap_ = Ap.copy_to_host()
Apr_ = Apr.copy_to_host()
A_ = A.copy_to_host()
checkperm.checkperm( A_, shape, Ap_, tuple(shapeOut), out )
#assumes no repeated indices and proper permutation
def parse( index_from, index_to ) :
out = np.zeros( len(index_from), dtype=int )
map_index = {}
for i in range(len(index_from)) :
map_index[ index_from[i] ] = i
for i in range(len(index_to)) :
out[i] = map_index[ index_to[i] ]
return out
#