bart | 8785c12 | 2008-05-29 08:34:27 +0000 | [diff] [blame] | 1 | #!/bin/bash |
| 2 | |
| 3 | ######################## |
| 4 | # Function definitions # |
| 5 | ######################## |
| 6 | |
bart | 3281150 | 2008-06-03 15:12:59 +0000 | [diff] [blame] | 7 | function log2 { |
| 8 | local i |
| 9 | |
| 10 | for ((i=0;i<64;i++)) |
| 11 | do |
| 12 | if [ $((2**i)) = $1 ]; then |
| 13 | echo $i |
| 14 | return 0 |
| 15 | fi |
| 16 | done |
| 17 | echo "" |
| 18 | return 1 |
| 19 | } |
| 20 | |
| 21 | function get_cache_size { |
| 22 | local s |
| 23 | s=$(</sys/devices/system/cpu/cpu0/cache/index2/size) |
| 24 | if [ "${s%M}" != "$s" ]; then |
| 25 | echo $((${s%M}*1024*1024)) |
| 26 | elif [ "${s%K}" != "$s" ]; then |
| 27 | echo $((${s%K}*1024)) |
| 28 | else |
| 29 | echo $s |
| 30 | fi |
| 31 | } |
| 32 | |
bart | 8785c12 | 2008-05-29 08:34:27 +0000 | [diff] [blame] | 33 | # Read a stream of numbers from stdin (one per line), and print the average |
| 34 | # and standard deviation. |
| 35 | function avgstddev { |
bart | f06b427 | 2008-05-30 09:52:13 +0000 | [diff] [blame] | 36 | awk '{n++;sum+=$1;sumsq+=$1*$1}END{d=sumsq/n-sum*sum/n/n;print sum/n,(d>0?sqrt(d):0)}' |
bart | 8785c12 | 2008-05-29 08:34:27 +0000 | [diff] [blame] | 37 | } |
| 38 | |
| 39 | function run_test { |
| 40 | local tmp avg1=1 stddev1=1 avg2=1 stddev2=1 |
| 41 | |
| 42 | tmp="/tmp/test-timing.$$" |
bart | 334db5e | 2008-06-05 10:14:53 +0000 | [diff] [blame] | 43 | rm -f "${tmp}" |
| 44 | |
bart | 8785c12 | 2008-05-29 08:34:27 +0000 | [diff] [blame] | 45 | echo "$@" |
| 46 | for ((i=0;i<3;i++)) |
| 47 | do |
bart | 868d73a | 2008-06-04 13:02:22 +0000 | [diff] [blame] | 48 | cat "${test_input:-/dev/null}" | \ |
| 49 | /usr/bin/time --format="%e" "$@" 2>&1 | \ |
bart | 0d4e5c2 | 2008-06-07 10:42:52 +0000 | [diff] [blame] | 50 | tee "${1}.out" | \ |
bart | 868d73a | 2008-06-04 13:02:22 +0000 | [diff] [blame] | 51 | tail -n 1 |
bart | 8785c12 | 2008-05-29 08:34:27 +0000 | [diff] [blame] | 52 | done | avgstddev > "$tmp" |
| 53 | read avg1 stddev1 < "$tmp" |
| 54 | echo "Average time: ${avg1} +/- ${stddev1} seconds" |
| 55 | |
bart | 868d73a | 2008-06-04 13:02:22 +0000 | [diff] [blame] | 56 | for p in 1 2 4 |
bart | 8785c12 | 2008-05-29 08:34:27 +0000 | [diff] [blame] | 57 | do |
bart | 0d4e5c2 | 2008-06-07 10:42:52 +0000 | [diff] [blame] | 58 | echo "$VG --tool=exp-drd --check-stack-var=yes $@ -p$p" |
bart | c4a174f | 2008-06-03 11:41:19 +0000 | [diff] [blame] | 59 | for ((i=0;i<3;i++)) |
| 60 | do |
bart | 868d73a | 2008-06-04 13:02:22 +0000 | [diff] [blame] | 61 | cat "${test_input:-/dev/null}" | \ |
| 62 | /usr/bin/time --format="%e" $VG --tool=exp-drd "$@" -p$p 2>&1 | \ |
bart | 0d4e5c2 | 2008-06-07 10:42:52 +0000 | [diff] [blame] | 63 | tee "${1}-drd-with-stack-var-${p}.out" | \ |
| 64 | tail -n 1 |
| 65 | done | avgstddev > "$tmp" |
| 66 | read avg2 stddev2 < "$tmp" |
| 67 | echo "Average time: ${avg2} +/- ${stddev2} seconds" |
| 68 | awk "END{print "'"'"Ratio ="'"'", ${avg2}/${avg1}, "'"'"+/-"'"'", ${avg2}/${avg1}*(${stddev1}/${avg1}+${stddev2}/${avg2})}" </dev/null |
| 69 | |
| 70 | echo "$VG --tool=exp-drd --check-stack-var=no $@ -p$p" |
| 71 | for ((i=0;i<3;i++)) |
| 72 | do |
| 73 | cat "${test_input:-/dev/null}" | \ |
| 74 | /usr/bin/time --format="%e" $VG --tool=exp-drd "$@" -p$p 2>&1 | \ |
| 75 | tee "${1}-drd-without-stack-var-${p}.out" | \ |
bart | 868d73a | 2008-06-04 13:02:22 +0000 | [diff] [blame] | 76 | tail -n 1 |
bart | c4a174f | 2008-06-03 11:41:19 +0000 | [diff] [blame] | 77 | done | avgstddev > "$tmp" |
| 78 | read avg2 stddev2 < "$tmp" |
| 79 | echo "Average time: ${avg2} +/- ${stddev2} seconds" |
| 80 | awk "END{print "'"'"Ratio ="'"'", ${avg2}/${avg1}, "'"'"+/-"'"'", ${avg2}/${avg1}*(${stddev1}/${avg1}+${stddev2}/${avg2})}" </dev/null |
| 81 | done |
bart | 8785c12 | 2008-05-29 08:34:27 +0000 | [diff] [blame] | 82 | |
| 83 | echo '' |
| 84 | |
| 85 | rm -f "$tmp" |
| 86 | } |
| 87 | |
| 88 | |
| 89 | # Script body |
| 90 | |
bart | 3281150 | 2008-06-03 15:12:59 +0000 | [diff] [blame] | 91 | DRD_SCRIPTS_DIR="$(dirname $0)" |
bart | 7acf380 | 2008-06-06 10:17:26 +0000 | [diff] [blame] | 92 | if [ "${DRD_SCRIPTS_DIR:0:1}" != "/" ]; then |
| 93 | DRD_SCRIPTS_DIR="$PWD/$DRD_SCRIPTS_DIR" |
bart | 3281150 | 2008-06-03 15:12:59 +0000 | [diff] [blame] | 94 | fi |
| 95 | |
| 96 | SPLASH2="${DRD_SCRIPTS_DIR}/../splash2" |
bart | c4a174f | 2008-06-03 11:41:19 +0000 | [diff] [blame] | 97 | if [ ! -e "${SPLASH2}" ]; then |
| 98 | echo "Error: splash2 directory not found (${SPLASH2})." |
bart | 8785c12 | 2008-05-29 08:34:27 +0000 | [diff] [blame] | 99 | exit 1 |
| 100 | fi |
| 101 | |
| 102 | if [ "$VG" = "" ]; then |
bart | 3281150 | 2008-06-03 15:12:59 +0000 | [diff] [blame] | 103 | VG="${DRD_SCRIPTS_DIR}/../../vg-in-place" |
bart | 8785c12 | 2008-05-29 08:34:27 +0000 | [diff] [blame] | 104 | fi |
| 105 | |
| 106 | if [ ! -e "$VG" ]; then |
| 107 | echo "Could not find $VG." |
| 108 | exit 1 |
| 109 | fi |
| 110 | |
bart | 0d4e5c2 | 2008-06-07 10:42:52 +0000 | [diff] [blame] | 111 | ############################################################################## |
| 112 | # Results: (-p1) (-p2) (-p4) ITC (-p4) ITC (-p4) |
| 113 | # original w/ filter |
| 114 | # ............................................................................ |
| 115 | # Cholesky 40 47 82 239 82 |
| 116 | # FFT 16 17 47 90 41 |
| 117 | # LU, contiguous blocks 39 41 45 428 128 |
| 118 | # LU, non-contiguous blocks 39 41 49 428 128 |
| 119 | # Ocean, contiguous partitions 17 19 25 90 28 |
| 120 | # Ocean, non-continguous partns 18 21 30 90 28 |
| 121 | # Radiosity 78 78 78 485 163 |
| 122 | # Radix 10 12 15 222 56 |
| 123 | # Raytrace 56 56 56 172 53 |
| 124 | # Water-n2 34 34 34 189 39 |
| 125 | # Water-sp 33 33 33 183 34 |
| 126 | # ............................................................................ |
| 127 | # Hardware: dual-core Intel Xeon 5130, 2.0 GHz, 4 MB L2 cache, 4 GB RAM. |
| 128 | # Software: Ubuntu 7.10 server, 64-bit, gcc 4.1.3. |
| 129 | ############################################################################## |
| 130 | # Results: native DRD (-p4) DRD (-p4) ITC (-p4) ITC (-p4) |
| 131 | # time original w/ filter original w/ filter |
| 132 | # ............................................................................ |
| 133 | # Cholesky 0.21 74 76 239 82 |
| 134 | # FFT 0.16 12 11 90 41 |
| 135 | # LU, contiguous blocks 0.57 49 49 428 128 |
| 136 | # LU, non-contiguous blocks 0.60 52 52 428 128 |
| 137 | # Ocean, contiguous partitions 14.42 24 24 90 28 |
| 138 | # Ocean, non-continguous partns 0.21 30 30 90 28 |
| 139 | # Radiosity 2.35 75 76 485 163 |
| 140 | # Radix 2.80 15 16 222 56 |
| 141 | # Raytrace 90.79 55 54 172 53 |
| 142 | # Water-n2 0.15 27 27 189 39 |
| 143 | # Water-sp 0.15 26 26 183 34 |
| 144 | # ............................................................................ |
| 145 | # Hardware: dual-core Intel Core2 Duo E6750, 2.66 GHz, 4 MB L2 cache, 2 GB RAM. |
| 146 | # Software: openSUSE 10.3, 64-bit, gcc 4.2.1. |
| 147 | ############################################################################## |
bart | 8785c12 | 2008-05-29 08:34:27 +0000 | [diff] [blame] | 148 | |
bart | 0d4e5c2 | 2008-06-07 10:42:52 +0000 | [diff] [blame] | 149 | cache_size=$(($(get_cache_size)/2)) |
bart | 3281150 | 2008-06-03 15:12:59 +0000 | [diff] [blame] | 150 | log2_cache_size=$(log2 ${cache_size}) |
| 151 | |
| 152 | # Cholesky |
bart | 3281150 | 2008-06-03 15:12:59 +0000 | [diff] [blame] | 153 | ( |
bart | 868d73a | 2008-06-04 13:02:22 +0000 | [diff] [blame] | 154 | cd ${SPLASH2}/codes/kernels/cholesky/inputs |
| 155 | for f in *Z |
| 156 | do |
| 157 | gzip -cd <$f >${f%.Z} |
| 158 | done |
bart | 0d4e5c2 | 2008-06-07 10:42:52 +0000 | [diff] [blame] | 159 | run_test ../CHOLESKY -C${cache_size} tk29.O |
bart | 3281150 | 2008-06-03 15:12:59 +0000 | [diff] [blame] | 160 | ) |
bart | 3281150 | 2008-06-03 15:12:59 +0000 | [diff] [blame] | 161 | |
| 162 | # FFT |
bart | 0d4e5c2 | 2008-06-07 10:42:52 +0000 | [diff] [blame] | 163 | run_test ${SPLASH2}/codes/kernels/fft/FFT -t -l${log2_cache_size} -m18 |
bart | 3281150 | 2008-06-03 15:12:59 +0000 | [diff] [blame] | 164 | |
| 165 | # LU, contiguous blocks. |
bart | c4a174f | 2008-06-03 11:41:19 +0000 | [diff] [blame] | 166 | run_test ${SPLASH2}/codes/kernels/lu/contiguous_blocks/LU -n1024 |
bart | 8785c12 | 2008-05-29 08:34:27 +0000 | [diff] [blame] | 167 | |
bart | 3281150 | 2008-06-03 15:12:59 +0000 | [diff] [blame] | 168 | # LU, non-contiguous blocks. |
bart | c4a174f | 2008-06-03 11:41:19 +0000 | [diff] [blame] | 169 | run_test ${SPLASH2}/codes/kernels/lu/non_contiguous_blocks/LU -n1024 |
bart | 8785c12 | 2008-05-29 08:34:27 +0000 | [diff] [blame] | 170 | |
bart | 868d73a | 2008-06-04 13:02:22 +0000 | [diff] [blame] | 171 | # Ocean |
| 172 | run_test ${SPLASH2}/codes/apps/ocean/contiguous_partitions/OCEAN -n2050 |
| 173 | run_test ${SPLASH2}/codes/apps/ocean/non_contiguous_partitions/OCEAN -n258 |
| 174 | |
bart | 3281150 | 2008-06-03 15:12:59 +0000 | [diff] [blame] | 175 | # Radiosity. |
bart | c4a174f | 2008-06-03 11:41:19 +0000 | [diff] [blame] | 176 | run_test ${SPLASH2}/codes/apps/radiosity/RADIOSITY -batch -room |
| 177 | |
bart | 868d73a | 2008-06-04 13:02:22 +0000 | [diff] [blame] | 178 | # Radix |
| 179 | run_test ${SPLASH2}/codes/kernels/radix/RADIX -n$((2**24)) |
| 180 | |
| 181 | # Raytrace |
| 182 | ( |
| 183 | cd ${SPLASH2}/codes/apps/raytrace/inputs |
| 184 | rm -f *.env *.geo *.rl |
| 185 | for f in *Z |
| 186 | do |
| 187 | gzip -cd <$f >${f%.Z} |
| 188 | done |
| 189 | run_test ../RAYTRACE balls4.env |
| 190 | ) |
bart | 334db5e | 2008-06-05 10:14:53 +0000 | [diff] [blame] | 191 | |
bart | 868d73a | 2008-06-04 13:02:22 +0000 | [diff] [blame] | 192 | # Water-n2 |
bart | 334db5e | 2008-06-05 10:14:53 +0000 | [diff] [blame] | 193 | ( |
| 194 | cd ${SPLASH2}/codes/apps/water-nsquared |
| 195 | test_input=input run_test ./WATER-NSQUARED |
| 196 | ) |
bart | 868d73a | 2008-06-04 13:02:22 +0000 | [diff] [blame] | 197 | |
| 198 | # Water-sp |
bart | 334db5e | 2008-06-05 10:14:53 +0000 | [diff] [blame] | 199 | ( |
| 200 | cd ${SPLASH2}/codes/apps/water-spatial |
| 201 | test_input=input run_test ./WATER-SPATIAL |
| 202 | ) |
bart | 868d73a | 2008-06-04 13:02:22 +0000 | [diff] [blame] | 203 | |
| 204 | |
bart | c4a174f | 2008-06-03 11:41:19 +0000 | [diff] [blame] | 205 | |
| 206 | # Local variables: |
| 207 | # compile-command: "./run-splash2" |
| 208 | # End: |