ceph/src/erasure-code/jerasure/gf-complete/tools/test_simd.sh

   1 #!/bin/bash -e
   2
   3 # this scripts has a number of tests for SIMD. It can be invoked
   4 # on the host or on a QEMU machine.
   5
   6 script_dir="$( cd "$( dirname "${BASH_SOURCE[0]}" )" && pwd )"
   7 host_cpu=`uname -p`
   8 results=${script_dir}/test_simd.results
   9 nprocs=$(grep -c ^processor /proc/cpuinfo)
  10
  11 # runs unit tests and save the results
  12 test_unit(){
  13     { ./configure && make clean && make; } || { echo "Compile FAILED" >> ${results}; return 1; }
  14     make -j$nprocs check || { echo "gf_methods $i FAILED" >> ${results}; ((++failed)); }
  15     cat tools/test-suite.log >> ${results} || true
  16 }
  17
  18 # build with DEBUG_FUNCTIONS and save all methods selected
  19 # to a results file
  20 test_functions() {
  21     failed=0
  22
  23     { ./configure --enable-debug-func && make clean && make; } || { echo "Compile FAILED" >> ${results}; return 1; }
  24     for i in 128 64 32 16 8 4; do
  25         { ${script_dir}/gf_methods $i -ACD -X >> ${results}; } || { echo "gf_methods $i FAILED" >> ${results}; ((++failed)); }
  26     done
  27
  28     return ${failed}
  29 }
  30
  31 # build with DEBUG_CPU_FUNCTIONS and print out CPU detection
  32 test_detection() {
  33     failed=0
  34
  35     { ./configure --enable-debug-cpu && make clean && make; } || { echo "Compile FAILED" >> ${results}; return 1; }
  36     { ${script_dir}/gf_methods 32 -ACD -L | grep '#' >> ${results}; } || { echo "gf_methods $i FAILED" >> ${results}; ((++failed)); }
  37
  38     return ${failed}
  39 }
  40
  41 compile_arm() {
  42     failed=0
  43
  44     echo -n "Compiling with NO SIMD support..." >> ${results}
  45     { ./configure --disable-neon && make clean && make && echo "SUCCESS" >> ${results}; } || { echo "FAIL" >> ${results}; ((++failed)); }
  46
  47     echo -n "Compiling with FULL SIMD support..." >> ${results}
  48     { ./configure && make clean && make && echo "SUCCESS" >> ${results}; } || { echo "FAIL" >> ${results}; ((++failed)); }
  49
  50     return ${failed}
  51 }
  52
  53 compile_intel() {
  54     failed=0
  55
  56     echo -n "Compiling with NO SIMD support..." >> ${results}
  57     { ./configure && make clean && make && echo "SUCCESS" >> ${results}; } || { echo "FAIL" >> ${results}; ((++failed)); }
  58
  59     echo -n "Compiling with SSE2 only..." >> ${results}
  60     export ax_cv_have_sse_ext=no
  61     export ax_cv_have_sse2_ext=yes
  62     export ax_cv_have_sse3_ext=no
  63     export ax_cv_have_ssse3_ext=no
  64     export ax_cv_have_sse41_ext=no
  65     export ax_cv_have_sse42_ext=no
  66     export ax_cv_have_pclmuldq_ext=no
  67     { ./configure && make clean && make && echo "SUCCESS" >> ${results}; } || { echo "FAIL" >> ${results}; ((++failed)); }
  68
  69     echo -n "Compiling with SSE2,SSE3 only..." >> ${results}
  70     export ax_cv_have_sse_ext=no
  71     export ax_cv_have_sse2_ext=yes
  72     export ax_cv_have_sse3_ext=yes
  73     export ax_cv_have_ssse3_ext=no
  74     export ax_cv_have_sse41_ext=no
  75     export ax_cv_have_sse42_ext=no
  76     export ax_cv_have_pclmuldq_ext=no
  77     { ./configure && make clean && make && echo "SUCCESS" >> ${results}; } || { echo "FAIL" >> ${results}; ((++failed)); }
  78
  79     echo -n "Compiling with SSE2,SSE3,SSSE3 only..." >> ${results}
  80     export ax_cv_have_sse_ext=no
  81     export ax_cv_have_sse2_ext=yes
  82     export ax_cv_have_sse3_ext=yes
  83     export ax_cv_have_ssse3_ext=yes
  84     export ax_cv_have_sse41_ext=no
  85     export ax_cv_have_sse42_ext=no
  86     export ax_cv_have_pclmuldq_ext=no
  87     { ./configure && make clean && make && echo "SUCCESS" >> ${results}; } || { echo "FAIL" >> ${results}; ((++failed)); }
  88
  89     echo -n "Compiling with SSE2,SSE3,SSSE3,SSE4_1 only..." >> ${results}
  90     export ax_cv_have_sse_ext=no
  91     export ax_cv_have_sse2_ext=yes
  92     export ax_cv_have_sse3_ext=yes
  93     export ax_cv_have_ssse3_ext=yes
  94     export ax_cv_have_sse41_ext=yes
  95     export ax_cv_have_sse42_ext=no
  96     export ax_cv_have_pclmuldq_ext=no
  97     { ./configure && make clean && make && echo "SUCCESS" >> ${results}; } || { echo "FAIL" >> ${results}; ((++failed)); }
  98
  99     echo -n "Compiling with SSE2,SSE3,SSSE3,SSE4_2 only..." >> ${results}
 100     export ax_cv_have_sse_ext=no
 101     export ax_cv_have_sse2_ext=yes
 102     export ax_cv_have_sse3_ext=yes
 103     export ax_cv_have_ssse3_ext=yes
 104     export ax_cv_have_sse41_ext=no
 105     export ax_cv_have_sse42_ext=yes
 106     export ax_cv_have_pclmuldq_ext=no
 107     { ./configure && make clean && make && echo "SUCCESS" >> ${results}; } || { echo "FAIL" >> ${results}; ((++failed)); }
 108
 109     echo -n "Compiling with FULL SIMD support..." >> ${results}
 110     export ax_cv_have_sse_ext=no
 111     export ax_cv_have_sse2_ext=yes
 112     export ax_cv_have_sse3_ext=yes
 113     export ax_cv_have_ssse3_ext=yes
 114     export ax_cv_have_sse41_ext=yes
 115     export ax_cv_have_sse42_ext=yes
 116     export ax_cv_have_pclmuldq_ext=yes
 117     { ./configure && make clean && make && echo "SUCCESS" >> ${results}; } || { echo "FAIL" >> ${results}; ((++failed)); }
 118
 119     return ${failed}
 120 }
 121
 122 # test that we can compile the source code with different
 123 # SIMD options. We assume that we are running on processor
 124 # full SIMD support
 125 test_compile() {
 126     case $host_cpu in
 127         aarch64*|arm*) compile_arm ;;
 128         i[[3456]]86*|x86_64*|amd64*) compile_intel ;;
 129     esac
 130 }
 131
 132 # disable through build flags
 133 runtime_arm_flags() {
 134     failed=0
 135
 136     echo "====NO SIMD support..." >> ${1}
 137     { ./configure --disable-neon --enable-debug-func && make clean && make; } || { echo "Compile FAILED" >> ${1}; return 1; }
 138     for i in 128 64 32 16 8 4; do
 139         { ${script_dir}/gf_methods $i -ACD -X >> ${1}; } || { echo "gf_methods $i FAILED" >> ${1}; ((++failed)); }
 140     done
 141
 142     echo "====FULL SIMD support..." >> ${1}
 143     { ./configure --enable-debug-func && make clean && make; } || { echo "Compile FAILED" >> ${1}; return 1; }
 144     for i in 128 64 32 16 8 4; do
 145         { ${script_dir}/gf_methods $i -ACD -X >> ${1}; } || { echo "gf_methods $i FAILED" >> ${1}; ((++failed)); }
 146     done
 147
 148     return ${failed}
 149 }
 150
 151 # build once with FULL SIMD and disable at runtime through environment
 152 runtime_arm_env() {
 153     failed=0
 154
 155     { ./configure --enable-debug-func && make clean && make; } || { echo "Compile FAILED" >> ${1}; return 1; }
 156
 157     echo "====NO SIMD support..." >> ${1}
 158     export GF_COMPLETE_DISABLE_NEON=1
 159     for i in 128 64 32 16 8 4; do
 160         { ${script_dir}/gf_methods $i -ACD -X >> ${1}; } || { echo "gf_methods $i FAILED" >> ${1}; ((++failed)); }
 161     done
 162
 163     echo "====FULL SIMD support..." >> ${1}
 164     unset GF_COMPLETE_DISABLE_NEON
 165     for i in 128 64 32 16 8 4; do
 166         { ${script_dir}/gf_methods $i -ACD -X >> ${1}; } || { echo "gf_methods $i FAILED" >> ${1}; ((++failed)); }
 167     done
 168
 169     return ${failed}
 170 }
 171
 172 runtime_intel_flags() {
 173     failed=0
 174
 175     echo "====NO SIMD support..." >> ${1}
 176     { ./configure --disable-sse --enable-debug-func && make clean && make; } || { echo "FAIL" >> ${1}; ((++failed)); }
 177     for i in 128 64 32 16 8 4; do
 178         { ${script_dir}/gf_methods $i -ACD -X >> ${1}; } || { echo "gf_methods $i FAILED" >> ${1}; ((++failed)); }
 179     done
 180
 181    echo "====SSE2 support..." >> ${1}
 182     export ax_cv_have_sse_ext=no
 183     export ax_cv_have_sse2_ext=yes
 184     export ax_cv_have_sse3_ext=no
 185     export ax_cv_have_ssse3_ext=no
 186     export ax_cv_have_sse41_ext=no
 187     export ax_cv_have_sse42_ext=no
 188     export ax_cv_have_pclmuldq_ext=no
 189     { ./configure --enable-debug-func && make clean && make; } || { echo "FAIL" >> ${1}; ((++failed)); }
 190     for i in 128 64 32 16 8 4; do
 191         { ${script_dir}/gf_methods $i -ACD -X >> ${1}; } || { echo "gf_methods $i FAILED" >> ${1}; ((++failed)); }
 192     done
 193
 194     echo "====SSE2,SSE3 support..." >> ${1}
 195     export ax_cv_have_sse_ext=no
 196     export ax_cv_have_sse2_ext=yes
 197     export ax_cv_have_sse3_ext=yes
 198     export ax_cv_have_ssse3_ext=no
 199     export ax_cv_have_sse41_ext=no
 200     export ax_cv_have_sse42_ext=no
 201     export ax_cv_have_pclmuldq_ext=no
 202     { ./configure --enable-debug-func && make clean && make; } || { echo "FAIL" >> ${1}; ((++failed)); }
 203     for i in 128 64 32 16 8 4; do
 204         { ${script_dir}/gf_methods $i -ACD -X >> ${1}; } || { echo "gf_methods $i FAILED" >> ${1}; ((++failed)); }
 205     done
 206
 207     echo "====SSE2,SSE3,SSSE3 support..." >> ${1}
 208     export ax_cv_have_sse_ext=no
 209     export ax_cv_have_sse2_ext=yes
 210     export ax_cv_have_sse3_ext=yes
 211     export ax_cv_have_ssse3_ext=yes
 212     export ax_cv_have_sse41_ext=no
 213     export ax_cv_have_sse42_ext=no
 214     export ax_cv_have_pclmuldq_ext=no
 215     { ./configure --enable-debug-func && make clean && make; } || { echo "FAIL" >> ${1}; ((++failed)); }
 216     for i in 128 64 32 16 8 4; do
 217         { ${script_dir}/gf_methods $i -ACD -X >> ${1}; } || { echo "gf_methods $i FAILED" >> ${1}; ((++failed)); }
 218     done
 219
 220     echo "====SSE2,SSE3,SSSE3,SSE4_1 support..." >> ${1}
 221     export ax_cv_have_sse_ext=no
 222     export ax_cv_have_sse2_ext=yes
 223     export ax_cv_have_sse3_ext=yes
 224     export ax_cv_have_ssse3_ext=yes
 225     export ax_cv_have_sse41_ext=yes
 226     export ax_cv_have_sse42_ext=no
 227     export ax_cv_have_pclmuldq_ext=no
 228     { ./configure --enable-debug-func && make clean && make; } || { echo "FAIL" >> ${1}; ((++failed)); }
 229     for i in 128 64 32 16 8 4; do
 230         { ${script_dir}/gf_methods $i -ACD -X >> ${1}; } || { echo "gf_methods $i FAILED" >> ${1}; ((++failed)); }
 231     done
 232
 233     echo "====SSE2,SSE3,SSSE3,SSE4_2 support..." >> ${1}
 234     export ax_cv_have_sse_ext=no
 235     export ax_cv_have_sse2_ext=yes
 236     export ax_cv_have_sse3_ext=yes
 237     export ax_cv_have_ssse3_ext=yes
 238     export ax_cv_have_sse41_ext=no
 239     export ax_cv_have_sse42_ext=yes
 240     export ax_cv_have_pclmuldq_ext=no
 241     { ./configure --enable-debug-func && make clean && make; } || { echo "FAIL" >> ${1}; ((++failed)); }
 242     for i in 128 64 32 16 8 4; do
 243         { ${script_dir}/gf_methods $i -ACD -X >> ${1}; } || { echo "gf_methods $i FAILED" >> ${1}; ((++failed)); }
 244     done
 245
 246     echo "====FULL SIMD support..." >> ${1}
 247     { ./configure --enable-debug-func && make clean && make; } || { echo "FAIL" >> ${1}; ((++failed)); }
 248     for i in 128 64 32 16 8 4; do
 249         { ${script_dir}/gf_methods $i -ACD -X >> ${1}; } || { echo "gf_methods $i FAILED" >> ${1}; ((++failed)); }
 250     done
 251
 252     return ${failed}
 253 }
 254
 255 runtime_intel_env() {
 256     failed=0
 257
 258     # compile a build with full SIMD support
 259     { ./configure --enable-debug-func && make clean && make; } || { echo "Compile FAILED" >> ${1}; return 1; }
 260
 261     echo "====NO SIMD support..." >> ${1}
 262     export GF_COMPLETE_DISABLE_SSE2=1
 263     export GF_COMPLETE_DISABLE_SSE3=1
 264     export GF_COMPLETE_DISABLE_SSSE3=1
 265     export GF_COMPLETE_DISABLE_SSE4=1
 266     export GF_COMPLETE_DISABLE_SSE4_PCLMUL=1
 267     for i in 128 64 32 16 8 4; do
 268         { ${script_dir}/gf_methods $i -ACD -X >> ${1}; } || { echo "gf_methods $i FAILED" >> ${1}; ((++failed)); }
 269     done
 270
 271     echo "====SSE2 support..." >> ${1}
 272     unset GF_COMPLETE_DISABLE_SSE2
 273     export GF_COMPLETE_DISABLE_SSE3=1
 274     export GF_COMPLETE_DISABLE_SSSE3=1
 275     export GF_COMPLETE_DISABLE_SSE4=1
 276     export GF_COMPLETE_DISABLE_SSE4_PCLMUL=1
 277     for i in 128 64 32 16 8 4; do
 278         { ${script_dir}/gf_methods $i -ACD -X >> ${1}; } || { echo "gf_methods $i FAILED" >> ${1}; ((++failed)); }
 279     done
 280
 281     echo "====SSE2,SSE3 support..." >> ${1}
 282     unset GF_COMPLETE_DISABLE_SSE2
 283     unset GF_COMPLETE_DISABLE_SSE3
 284     export GF_COMPLETE_DISABLE_SSSE3=1
 285     export GF_COMPLETE_DISABLE_SSE4=1
 286     export GF_COMPLETE_DISABLE_SSE4_PCLMUL=1
 287     for i in 128 64 32 16 8 4; do
 288         { ${script_dir}/gf_methods $i -ACD -X >> ${1}; } || { echo "gf_methods $i FAILED" >> ${1}; ((++failed)); }
 289     done
 290
 291     echo "====SSE2,SSE3,SSSE3 support..." >> ${1}
 292     unset GF_COMPLETE_DISABLE_SSE2
 293     unset GF_COMPLETE_DISABLE_SSE3
 294     unset GF_COMPLETE_DISABLE_SSSE3
 295     export GF_COMPLETE_DISABLE_SSE4=1
 296     export GF_COMPLETE_DISABLE_SSE4_PCLMUL=1
 297     for i in 128 64 32 16 8 4; do
 298         { ${script_dir}/gf_methods $i -ACD -X >> ${1}; } || { echo "gf_methods $i FAILED" >> ${1}; ((++failed)); }
 299     done
 300
 301     echo "====SSE2,SSE3,SSSE3,SSE4_1 support..." >> ${1}
 302     unset GF_COMPLETE_DISABLE_SSE2
 303     unset GF_COMPLETE_DISABLE_SSE3
 304     unset GF_COMPLETE_DISABLE_SSSE3
 305     unset GF_COMPLETE_DISABLE_SSE4
 306     export GF_COMPLETE_DISABLE_SSE4_PCLMUL=1
 307     for i in 128 64 32 16 8 4; do
 308         { ${script_dir}/gf_methods $i -ACD -X >> ${1}; } || { echo "gf_methods $i FAILED" >> ${1}; ((++failed)); }
 309     done
 310
 311     echo "====SSE2,SSE3,SSSE3,SSE4_2 support..." >> ${1}
 312     unset GF_COMPLETE_DISABLE_SSE2
 313     unset GF_COMPLETE_DISABLE_SSE3
 314     unset GF_COMPLETE_DISABLE_SSSE3
 315     unset GF_COMPLETE_DISABLE_SSE4
 316     export GF_COMPLETE_DISABLE_SSE4_PCLMUL=1
 317     for i in 128 64 32 16 8 4; do
 318         { ${script_dir}/gf_methods $i -ACD -X >> ${1}; } || { echo "gf_methods $i FAILED" >> ${1}; ((++failed)); }
 319     done
 320
 321     echo "====FULL SIMD support..." >> ${1}
 322     unset GF_COMPLETE_DISABLE_SSE2
 323     unset GF_COMPLETE_DISABLE_SSE3
 324     unset GF_COMPLETE_DISABLE_SSSE3
 325     unset GF_COMPLETE_DISABLE_SSE4
 326     unset GF_COMPLETE_DISABLE_SSE4_PCLMUL
 327     for i in 128 64 32 16 8 4; do
 328         { ${script_dir}/gf_methods $i -ACD -X >> ${1}; } || { echo "gf_methods $i FAILED" >> ${1}; ((++failed)); }
 329     done
 330
 331     return ${failed}
 332 }
 333
 334 test_runtime() {
 335     rm -f ${results}.left
 336     rm -f ${results}.right
 337
 338     case $host_cpu in
 339         aarch64*|arm*)
 340             runtime_arm_flags ${results}.left
 341             runtime_arm_env ${results}.right
 342             ;;
 343         i[[3456]]86*|x86_64*|amd64*)
 344             runtime_intel_flags ${results}.left
 345             runtime_intel_env ${results}.right
 346             ;;
 347     esac
 348
 349     echo "======LEFT======" > ${results}
 350     cat ${results}.left >> ${results}
 351     echo "======RIGHT======" >> ${results}
 352     cat ${results}.right >> ${results}
 353     echo "======RESULT======" >> ${results}
 354     if diff "${results}.left" "${results}.right"; then
 355         echo SUCCESS >> ${results}
 356         return 0
 357     else
 358         echo SUCCESS >> ${results}
 359         return 1
 360     fi
 361 }
 362
 363 cd ${script_dir}/..
 364 rm -f ${results}
 365
 366 test_$1
 367 exit $?