/
run-blas-tester.sh
executable file
·246 lines (226 loc) · 8.59 KB
/
run-blas-tester.sh
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
#!/usr/bin/env bash
run_l1() {
temp=`mktemp`
echo -n "$4,$1,$3," >> ${2}.csv
#N sizes selected arbitrary to achieve enough runtime to get no 0.00 mflops values; more indepth knowhow needed
#(maybe based on L2SIZE) to determine proper values so that we test cpu fpu perf and not memory bandwidth limitations
#also for any sensible arch comparison ensure that cpus run at fixed frequency and no throttling is happening
OMP_NUM_THREADS=$4 BLAS-Tester-20160411/bin/${1}blastst -N 1000000 10000000 1000000 -R $3 > $temp
#we expect 10 PASSed runs of each test
[ `grep PASS $temp | wc -l` -ne 10 ] && echo "test failed!" && exit 1
#find which column is actually MFLOPs
mflopscol=$((`cat $temp | grep TST | tr -s ' ' | tr ' ' '\n' | nl | grep MFLOP | awk '{ print $1 }'`+1))
#wtf ... some are shifted
[ "$3" = "axpy" -a "$1" = "xcl1" ] && mflopscol=$(($mflopscol+1))
[ "$3" = "axpy" -a "$1" = "xzl1" ] && mflopscol=$(($mflopscol+1))
grep PASS $temp | tr -s ' ' | cut -d' ' -f $mflopscol | sort -n > ${temp}x
#mflops vary wildly; does discarding slowest and fastest and averaging the rest make sense??
( echo -n 'scale=2; ('; cat ${temp}x | tail -9 | head -8 | tr '\n' '+'; echo '0)/8' ) | bc >> ${2}.csv
rm $temp ${temp}x
}
run_l2() {
temp=`mktemp`
echo -n "$4,$1,$3," >> ${2}.csv
OMP_NUM_THREADS=$4 BLAS-Tester-20160411/bin/${1}blastst -N 1000 10000 1000 -R $3 > $temp
[ `grep PASS $temp | wc -l` -ne 10 ] && echo "test failed!" && exit 1
mflopscol=$((`cat $temp | grep TST | tr -s ' ' | tr ' ' '\n' | nl | grep MFLOP | awk '{ print $1 }'`+1))
[ "$3" = "gbmv" -a "$1" = "xcl2" ] && mflopscol=$(($mflopscol+2))
[ "$3" = "gbmv" -a "$1" = "xzl2" ] && mflopscol=$(($mflopscol+2))
[ "$3" = "gemv" -a "$1" = "xcl2" ] && mflopscol=$(($mflopscol+2))
[ "$3" = "gemv" -a "$1" = "xzl2" ] && mflopscol=$(($mflopscol+2))
[ "$3" = "hbmv" -a "$1" = "xcl2" ] && mflopscol=$(($mflopscol+2))
[ "$3" = "hbmv" -a "$1" = "xzl2" ] && mflopscol=$(($mflopscol+2))
[ "$3" = "hpmv" -a "$1" = "xcl2" ] && mflopscol=$(($mflopscol+2))
[ "$3" = "hpmv" -a "$1" = "xzl2" ] && mflopscol=$(($mflopscol+2))
[ "$3" = "hemv" -a "$1" = "xcl2" ] && mflopscol=$(($mflopscol+2))
[ "$3" = "hemv" -a "$1" = "xzl2" ] && mflopscol=$(($mflopscol+2))
[ "$3" = "gerc" -a "$1" = "xcl2" ] && mflopscol=$(($mflopscol+1))
[ "$3" = "geru" -a "$1" = "xcl2" ] && mflopscol=$(($mflopscol+1))
[ "$3" = "ger2c" -a "$1" = "xcl2" ] && mflopscol=$(($mflopscol+1))
[ "$3" = "ger2u" -a "$1" = "xcl2" ] && mflopscol=$(($mflopscol+1))
[ "$3" = "hpr2" -a "$1" = "xcl2" ] && mflopscol=$(($mflopscol+1))
[ "$3" = "her2" -a "$1" = "xcl2" ] && mflopscol=$(($mflopscol+1))
grep PASS $temp | tr -s ' ' | cut -d' ' -f $mflopscol | sort -n > ${temp}x
( echo -n 'scale=2; ('; cat ${temp}x | tail -9 | head -8 | tr '\n' '+'; echo '0)/8' ) | bc >> ${2}.csv
rm $temp ${temp}x
}
run_l3() {
temp=`mktemp`
echo -n "$4,$1,$3," >> ${2}.csv
OMP_NUM_THREADS=$4 BLAS-Tester-20160411/bin/${1}blastst -N 100 1000 100 -R $3 > $temp
[ `grep PASS $temp | wc -l` -ne 10 ] && echo "test failed!" && exit 1
mflopscol=$((`cat $temp | grep TST | tr -s ' ' | tr ' ' '\n' | nl | grep MFLOP | awk '{ print $1 }'`+1))
[ "$3" = "her2k" -a "$1" = "xcl3" ] && mflopscol=$(($mflopscol+1))
[ "$3" = "her2k" -a "$1" = "xzl3" ] && mflopscol=$(($mflopscol+1))
[ "$3" = "trmm" -a "$1" = "xcl3" ] && mflopscol=$(($mflopscol+1))
[ "$3" = "trmm" -a "$1" = "xzl3" ] && mflopscol=$(($mflopscol+1))
[ "$3" = "trsm" -a "$1" = "xcl3" ] && mflopscol=$(($mflopscol+1))
[ "$3" = "trsm" -a "$1" = "xzl3" ] && mflopscol=$(($mflopscol+1))
[ "$3" = "gemm" -a "$1" = "xcl3" ] && mflopscol=$(($mflopscol+2))
[ "$3" = "gemm" -a "$1" = "xzl3" ] && mflopscol=$(($mflopscol+2))
[ "$3" = "hemm" -a "$1" = "xcl3" ] && mflopscol=$(($mflopscol+2))
[ "$3" = "hemm" -a "$1" = "xzl3" ] && mflopscol=$(($mflopscol+2))
[ "$3" = "symm" -a "$1" = "xcl3" ] && mflopscol=$(($mflopscol+2))
[ "$3" = "symm" -a "$1" = "xzl3" ] && mflopscol=$(($mflopscol+2))
[ "$3" = "syrk" -a "$1" = "xcl3" ] && mflopscol=$(($mflopscol+2))
[ "$3" = "syrk" -a "$1" = "xzl3" ] && mflopscol=$(($mflopscol+2))
[ "$3" = "syr2k" -a "$1" = "xcl3" ] && mflopscol=$(($mflopscol+2))
[ "$3" = "syr2k" -a "$1" = "xzl3" ] && mflopscol=$(($mflopscol+2))
grep PASS $temp | tr -s ' ' | cut -d' ' -f $mflopscol | sort -n > ${temp}x
( echo -n 'scale=2; ('; cat ${temp}x | tail -9 | head -8 | tr '\n' '+'; echo '0)/8' ) | bc >> ${2}.csv
rm $temp ${temp}x
}
run_tests() {
export OMP_PLACES=sockets
export OMP_PROC_BIND=close
module=$1
cores=$2
[ -e ${module}.csv ] || echo nthreads,test,routine,MFLOPS > ${module}.csv
tst=xcl1
echo -n "$tst: "
#c nrm2,amax,asum,axpy,copy,swap,dotc,dotu - these behave in sane way: no 0.00 mflops and 10 runs
for r in nrm2 amax asum axpy copy swap dotc dotu
do
echo -n "$r "
run_l1 $tst $module $r $cores
done
echo
tst=xzl1
echo -n "$tst: "
#z nrm2,amax,asum,axpy,copy,swap,dotc,dotu
for r in nrm2 amax asum axpy copy swap dotc dotu
do
echo -n "$r "
run_l1 $tst $module $r $cores
done
echo
tst=xdl1
echo -n "$tst: "
#d nrm2,amax,asum,axpy,copy,swap,dot,rotm
for r in nrm2 amax asum axpy copy swap dot rotm
do
echo -n "$r "
run_l1 $tst $module $r $cores
done
echo
tst=xsl1
echo -n "$tst: "
#s nrm2,amax,asum,axpy,copy,swap,dot,rotm,dsdot,sdsdot
for r in nrm2 amax asum axpy copy swap dot rotm dsdot sdsdot
do
echo -n "$r "
run_l1 $tst $module $r $cores
done
echo
tst=xcl2
echo -n "$tst: "
#c gbmv,gemv,hbmv,hpmv,hemv,tbmv,tpmv,trmv,tbsv,tpsv,trsv,gerc,geru,ger2c,ger2u,hpr,her,hpr2,her2
for r in gbmv gemv hbmv hpmv hemv tbmv tpmv trmv tbsv tpsv trsv gerc geru ger2c ger2u hpr her hpr2 her2
do
echo -n "$r "
run_l2 $tst $module $r $cores
done
echo
tst=xzl2
echo -n "$tst: "
#z gbmv,gemv,hbmv,hpmv,hemv,tbmv,tpmv,trmv,tbsv,tpsv,trsv, rest fails with nomem: gerc,geru,ger2c,ger2u,hpr,her,hpr2,her2
for r in gbmv gemv hbmv hpmv hemv tbmv tpmv trmv tbsv tpsv trsv
do
echo -n "$r "
run_l2 $tst $module $r $cores
done
echo
tst=xdl2
echo -n "$tst: "
#d gbmv,gemv,sbmv,spmv,symv,tbmv,tpmv,trmv,tbsv,tpsv,trsv,ger,ger2,spr,syr,spr2,syr2
for r in gbmv gemv sbmv spmv symv tbmv tpmv trmv tbsv tpsv trsv ger ger2 spr syr spr2 syr2
do
echo -n "$r "
run_l2 $tst $module $r $cores
done
echo
tst=xsl2
echo -n "$tst: "
#s gbmv,gemv,sbmv,spmv,symv,tbmv,tpmv,trmv,tbsv,tpsv,trsv,ger,ger2, spr,syr,spr2,syr2
for r in gbmv gemv sbmv spmv symv tbmv tpmv trmv tbsv tpsv trsv ger ger2 spr syr spr2 syr2
do
echo -n "$r "
run_l2 $tst $module $r $cores
done
echo
tst=xcl3
echo -n "$tst: "
#c gemm,hemm,herk,her2k,symm,syr2k,syrk,trmm,trsm
for r in gemm hemm herk her2k symm syr2k syrk trmm trsm
do
echo -n "$r "
run_l3 $tst $module $r $cores
done
echo
tst=xzl3
echo -n "$tst: "
#z gemm,hemm,herk,her2k,symm,syr2k,syrk,trmm,trsm
for r in gemm hemm herk her2k symm syr2k syrk trmm trsm
do
echo -n "$r "
run_l3 $tst $module $r $cores
done
echo
tst=xdl3
echo -n "$tst: "
#d gemm,symm,syrk,syr2k,trmm,trsm
for r in gemm symm syrk syr2k trmm trsm
do
echo -n "$r "
run_l3 $tst $module $r $cores
done
echo
tst=xsl3
echo -n "$tst: "
#s gemm,symm,syrk,syr2k,trmm,trsm
for r in gemm symm syrk syr2k trmm trsm
do
echo -n "$r "
run_l3 $tst $module $r $cores
done
echo
}
#fetch if not already here
[ -f 20160411.tar.gz ] || wget -q https://github.com/xianyi/BLAS-Tester/archive/8e1f624.tar.gz -O 20160411.tar.gz
#test for these toolchains
for chain in foss/2020b gobff/2020b gobff/2020.06-amd
do
echo loading $chain
module purge
module load $chain
maxcores=$((`lscpu | grep On-line | cut -f3 -d-`+1))
cores=1
until [ $cores -gt $maxcores ]
do
echo -n "num_threads=$cores ... "
#prepare clean copy
rm -rf BLAS-Tester-20160411
mkdir BLAS-Tester-20160411 && tar -xzf 20160411.tar.gz --strip-components 1 -C BLAS-Tester-20160411 && cd BLAS-Tester-20160411
sed -i -e 's/-openmp/-qopenmp/g' Makefile.system
echo building ...
#link to proper blas lib
if [[ "$chain" == "gobff"* ]]
then
#reports indicate that amd blis comes with separate .so for multithreaded
if [ -e $EBROOTBLIS/lib/libblis-mt.so ]; then BLASLIB=$EBROOTBLIS/lib/libblis-mt.so; else BLASLIB=$EBROOTBLIS/lib/libblis.so; fi
elif [[ "$chain" == "foss"* ]]
then
BLASLIB=$EBROOTOPENBLAS/lib/libopenblas.so
else
echo "Unknown toolchain"
exit 1
fi
#does NUMTREADS here influence the way tests are built?
make -j 4 CC="mpicc" CXX="mpicxx" FC="mpif90" NUMTHREADS=$cores USE_OPENMP=1 L2SIZE=$(getconf LEVEL2_CACHE_SIZE) TEST_BLAS=$BLASLIB NO_EXTENSION=1 >/dev/null 2>&1
cd ..
f=`echo $chain | sed 's/\//-/'`
run_tests ${f} $cores
echo
cores=$(($cores+1)) #or cores+cores for more quadratic range
done
done