-
Notifications
You must be signed in to change notification settings - Fork 109
/
Copy pathtranspose-numpy-shmem-rma-get.py
executable file
·236 lines (200 loc) · 9.06 KB
/
transpose-numpy-shmem-rma-get.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
#!/usr/bin/env python3
#
# Copyright (c) 2020, Intel Corporation
# Copyright (c) 2023, NVIDIA
#
# 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.
# * Neither the name of Intel Corporation nor the names of its
# contributors may be used to endorse or promote products
# derived from this software without specific prior written
# permission.
#
# 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 OWNER 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.
#*******************************************************************
#
# NAME: transpose
#
# PURPOSE: This program measures the time for the transpose of a
# column-major stored matrix into a row-major stored matrix.
#
# USAGE: Program input is the matrix order and the number of times to
# repeat the operation:
#
# transpose <# iterations> <matrix_size>
#
# The output consists of diagnostics to make sure the
# transpose worked and timing statistics.
#
# HISTORY: Written by Rob Van der Wijngaart, February 2009.
# Converted to Python by Jeff Hammond, February 2016.
#
# *******************************************************************
# Layout nomenclature
# -------------------
#
# - Each rank owns one block of columns (Colblock) of the overall
# matrix to be transposed, as well as of the transposed matrix.
# - Colblock is stored contiguously in the memory of the rank.
# The stored format is column major, which means that matrix
# elements (i,j) and (i+1,j) are adjacent, and (i,j) and (i,j+1)
# are "order" words apart
# - Colblock is logically composed of #ranks Blocks, but a Block is
# not stored contiguously in memory. Conceptually, the Block is
# the unit of data that gets communicated between ranks. Block i of
# rank j is locally transposed and gathered into a buffer called Work,
# which is sent to rank i, where it is scattered into Block j of the
# transposed matrix.
# - When tiling is applied to reduce TLB misses, each block gets
# accessed by tiles.
# - The original and transposed matrices are called A and B
#
# +-----------------------------------------------------------------+
# | | | | |
# | Colblock | | | |
# | | | | |
# | | | | |
# | | | | |
# | ------------------------------- |
# | | | | |
# | | Block | | |
# | | | | |
# | | | | |
# | | | | |
# | ------------------------------- |
# | | | | |
# | | | | Overall Matrix |
# | | | | |
# | | | | |
# | | | | |
# | ------------------------------- |
# | | | | |
# | | | | |
# | | | | |
# | | | | |
# | | | | |
# +-----------------------------------------------------------------+
import sys
if sys.version_info >= (3, 3):
from time import process_time as timer
else:
from timeit import default_timer as timer
from shmem4py import shmem
import numpy
import time
def main():
me = shmem.my_pe()
np = shmem.n_pes()
# ********************************************************************
# read and test input parameters
# ********************************************************************
if (me==0):
print('Parallel Research Kernels version ') #, PRKVERSION
print('Python SHMEM/Numpy Matrix transpose: B = A^T')
if len(sys.argv) != 3:
if (me==0):
print('argument count = ', len(sys.argv))
print("Usage: ./transpose <# iterations> <matrix order>")
sys.exit()
iterations = int(sys.argv[1])
if iterations < 1:
if (me==0):
print("ERROR: iterations must be >= 1")
sys.exit()
order = int(sys.argv[2])
if order < 1:
if (me==0):
print("ERROR: order must be >= 1")
sys.exit()
if order % np != 0:
if (me==0):
print(f"ERROR: matrix order ({order}) should be divisible by # procs ({np})")
sys.exit()
block_order = int(order / np)
if (me==0):
print('Number of ranks = ', np)
print('Number of iterations = ', iterations)
print('Matrix order = ', order)
shmem.barrier_all()
# ********************************************************************
# ** Allocate space for the input and transpose matrix
# ********************************************************************
#LA = numpy.fromfunction(lambda i,j: me * block_order + i*order + j, (order,block_order), dtype='d')
#A = shmem.full((order,block_order),LA)
A = shmem.zeros((order,block_order))
B = shmem.zeros((order,block_order))
T = shmem.zeros((np,block_order,block_order))
send_flag = shmem.ones(np, dtype='i')
recv_flag = shmem.zeros(np, dtype='i')
TA = numpy.fromfunction(lambda i,j: me * block_order + i*order + j, (order,block_order), dtype=numpy.double)
A[:,:] = TA[:,:]
for k in range(0,iterations+1):
if k<1:
shmem.barrier_all()
t0 = timer()
for phase in range(0,np):
recv_from = (me + phase) % np
send_to = (me - phase + np) % np
lo = block_order * send_to
hi = block_order * (send_to+1)
shmem.wait_until(send_flag[send_to:send_to+1], shmem.CMP.EQ, 1)
send_flag[send_to] = 0
shmem.put(T[phase], A[lo : hi,:], send_to)
shmem.fence()
shmem.atomic_inc(recv_flag[phase:phase+1], send_to)
shmem.wait_until(recv_flag[phase:phase+1], shmem.CMP.EQ, k+1)
lo = block_order * recv_from
hi = block_order * (recv_from+1)
B[lo:hi,:] += T[phase].T
shmem.put(send_flag[me:me+1], numpy.array([1],dtype='i'), recv_from)
A += 1.0
t1 = timer()
shmem.barrier_all()
trans_time = t1 - t0
shmem.free(A)
shmem.free(T)
# ********************************************************************
# ** Analyze and output results.
# ********************************************************************
# allgather is non-scalable but was easier to debug
F = shmem.zeros((np,order,block_order))
shmem.fcollect(F,B)
G = numpy.concatenate(F,axis=1)
#if (me==0):
# print(G)
H = numpy.fromfunction(lambda i,j: ((iterations/2.0)+(order*j+i))*(iterations+1.0), (order,order), dtype='d')
abserr = numpy.linalg.norm(numpy.reshape(G-H,order*order),ord=1)
shmem.free(B)
shmem.free(F)
epsilon=1.e-8
nbytes = 2 * order**2 * 8 # 8 is not sizeof(double) in bytes, but allows for comparison to C etc.
if abserr < epsilon:
if (me==0):
print('Solution validates')
avgtime = trans_time/iterations
print('Rate (MB/s): ',1.e-6*nbytes/avgtime, ' Avg time (s): ', avgtime)
else:
if (me==0):
print('error ',abserr, ' exceeds threshold ',epsilon)
print("ERROR: solution did not validate")
if __name__ == '__main__':
main()