Uploaded image for project: 'JDK'
  1. JDK
  2. JDK-8190494

Different results with UseAVX=3 when calling AVX-512 native function via JNI

    Details

    • Subcomponent:
    • Resolved In Build:
      b36
    • CPU:
      x86_64
    • OS:
      linux
    • Verification:
      Verified

      Description

      FULL PRODUCT VERSION :
      java version "9.0.1"
      Java(TM) SE Runtime Environment (build 9.0.1+11)
      Java HotSpot(TM) 64-Bit Server VM (build 9.0.1+11, mixed mode)


      FULL OS VERSION :
      Linux pnod0336 3.10.0-514.6.1.el7.x86_64 #1 SMP Wed Jan 18 13:06:36 UTC 2017 x86_64 x86_64 x86_64 GNU/Linux


      EXTRA RELEVANT SYSTEM CONFIGURATION :
      Intel(R) Xeon Phi(TM) CPU 7210 @ 1.30GHz

      A DESCRIPTION OF THE PROBLEM :
      Results of the following simple program are different between runs on KNL with -XX:UseAVX=3 and without this flag set.
      Java does some computation (probably kicking in AVX-512 vectorization) and in between it calls a native method via JNI. The native implementation is compiled for AVX-512 using Intel Compiler:
      icc version 17.0.3 (gcc version 4.8.5 compatibility)

      STEPS TO FOLLOW TO REPRODUCE THE PROBLEM :
      See the code below.
      cat ./reproduce.sh:
      echo "Compiling native code"
      icc ./systeminfoJNI.c -I$JAVA_HOME/include -I$JAVA_HOME/include/linux -xMIC-AVX512 -fPIC -std=gnu99 -o libSystemInfo.so -shared
      echo "Compiling java"
      javac TestAVX.java
      echo "Executing test program with -XX:UseAVX=3"
      java -Djava.library.path=. -XX:UseAVX=3 TestAVX
      echo "Executing test program without UseAVX"
      java -Djava.library.path=. TestAVX


      EXPECTED VERSUS ACTUAL BEHAVIOR :
      Sample program prints 0 when UseAVX=3. It should print the same value as with no AVX-512 enabled:
      1711
      REPRODUCIBILITY :
      This bug can be reproduced always.

      ---------- BEGIN SOURCE ----------
      class TestAVX {
              public static class SystemInfo{
                      static {
                              System.loadLibrary("SystemInfo");
                      }
                      public native int getNumberOfHardwareThreads();
              }

              static class TraceCacheSizeEstimator {
              private static final float CORE_WEIGHT_KNL_HOST = .5f;
              static final float LIVE_SAMP_FRAC = 213f;
              static final float FAST_OUTSAMPS = 1.e6f;

              public final long defaultTraceCacheSize;
              public final long minimumTraceCacheSize;

              public final long traceObjectSize;
              public TraceCacheSizeEstimator(
                                             int numberOfSamples, int numberOfHorizonPoints, int numberOfBins, int numberOfFilters,
                                             SystemInfo systemInfo) {

                  float averageNumberOfOutputLocationsInAperture = 13213213.21f;
                  float averageNumberOfLiveSamplesInAperture = LIVE_SAMP_FRAC * numberOfSamples * averageNumberOfOutputLocationsInAperture;
                  averageNumberOfLiveSamplesInAperture = Math.max(averageNumberOfLiveSamplesInAperture, FAST_OUTSAMPS);
                  traceObjectSize = 25665;
                  float hostMicEquivalents = getHostMicEquivalents(systemInfo);
                  float allMicEquivalents = hostMicEquivalents ;

                  defaultTraceCacheSize = (long) (allMicEquivalents * 3423.90f);
                  minimumTraceCacheSize = (long) (allMicEquivalents * 3213213.f);
              }
              private float getHostMicEquivalents(SystemInfo systemInfo ){
                  return CORE_WEIGHT_KNL_HOST * Math.min(1.0f, 2.0f * 256 / systemInfo.getNumberOfHardwareThreads());
              }
          }

          private static TraceCacheSizeEstimator create(){
              int numberOfSamples = 15;
              int numberOfHorizonPoints = 1;
              int numberOfBins = 81;
              int numberOfFilters = 4;
              return new TraceCacheSizeEstimator(
                                                 numberOfSamples, numberOfHorizonPoints, numberOfBins, numberOfFilters,
                                                 new SystemInfo());

          }
          public static void main(String[] argv) throws Exception {
              TraceCacheSizeEstimator trc = create();

              System.out.println(String.format("%d", trc.defaultTraceCacheSize));
          }


       systeminfoJNI.c :
      #define _GNU_SOURCE
      #include <sched.h>
      #include "jni.h"
      int duthread_ncpus() {
              cpu_set_t cs;
              CPU_ZERO(&cs);
              sched_getaffinity(0, sizeof(cs), &cs);

              int count = 0;
              for (int i = 0; i < CPU_SETSIZE; i++)
              {
                      if (CPU_ISSET(i, &cs))
                              count++;
              }
              return count;
      }

      JNIEXPORT jint JNICALL Java_TestAVX_00024SystemInfo_getNumberOfHardwareThreads
        (JNIEnv * env, jobject obj){
              return duthread_ncpus();
      }

      ---------- END SOURCE ----------

      CUSTOMER SUBMITTED WORKAROUND :
      Not using AVX-512.

        Attachments

          Activity

            People

            • Assignee:
              vdeshpande Vivek Deshpande
              Reporter:
              webbuggrp Webbug Group
            • Votes:
              0 Vote for this issue
              Watchers:
              7 Start watching this issue

              Dates

              • Created:
                Updated:
                Resolved: