Add jni method about omp and others (intel-analytics#10)

* fix: add jni method about omp environments and convert back to parallel mode

* fix: set threads to 1 through mkl_set_num_threads, opm_set_num_threads can't work as OMP_NUM_THREADS

* fix: icc warning fix

* fix: typo of waitPolicyPasssive

* fix: set mkl environment variables after load the library.

* fix: typo

mkl2017-xeon-blas/jni/src/main/java/com/intel/analytics/bigdl/mkl/MKL.java

@@ -4,6 +4,7 @@
 import java.net.URL;
 import java.nio.channels.FileChannel;
 import java.nio.channels.ReadableByteChannel;
+import java.util.Locale;
 
 import static java.io.File.createTempFile;
 import static java.nio.channels.Channels.newChannel;
@@ -29,8 +30,10 @@ public class MKL {
             tmpFile = extract(jmklFileName);
             System.load(tmpFile.getAbsolutePath());
             tmpFile.delete(); // delete so temp file after loaded
-            isLoaded = true;
 
+            setMklEnv();
+
+            isLoaded = true;
         } catch (Exception e) {
             isLoaded = false;
             e.printStackTrace();
@@ -39,6 +42,49 @@ public class MKL {
         }
     }
 
+    /**
+     * This method will call 4 native methods to set mkl environments. They are
+     *
+     * 1. mkl num threads:
+     *      + function: If mkl is linked with parallel mode, one can modify the number of threads omp will use.
+     *      + default value: 1
+     * 2. mkl block time:
+     *      + function: Turns off the Intel MKL Memory Allocator for Intel MKL functions
+     *                  to directly use the system malloc/free functions.
+     *      + default value: 0
+     * 3. wait policy:
+     *      + function: Sets omp wait policy to passive.
+     *      + default value: passive
+     * 4. disable fast mm:
+     *      + function: Sets the time (milliseconds) that a thread should wait before sleeping,
+     *                  after completing the execution of a parallel region.
+     *      + default value: true
+     */
+    private static void setMklEnv() {
+        String mklBlockTimeStr = System.getProperty("bigdl.mklBlockTime", "0");
+        int mklBlockTime = Integer.parseInt(mklBlockTimeStr);
+
+        String mklNumThreadsStr = System.getProperty("bigdl.mklNumThreads", "1");
+        int mklNumThreads = Integer.parseInt(mklNumThreadsStr);
+
+        String mklDisableFastMMStr = System.getProperty("bigdl.mklDisableFastMM","true").toLowerCase();
+        boolean mklDisableFastMM = true;
+        if (mklDisableFastMMStr.equals("false")) {
+            mklDisableFastMM = false;
+        }
+
+        String mklWaitPolicy = System.getProperty("bigdl.mklWaitPolicy", "passive").toLowerCase();
+
+        // set mkl environment variables
+        setNumThreads(mklNumThreads);
+        setBlockTime(mklBlockTime);
+        waitPolicyPassive();
+
+        if (mklDisableFastMM) {
+            disableFastMM();
+        }
+    }
+
     /**
      * Check if MKL is loaded
      * @return
@@ -58,8 +104,38 @@ public static String getTmpSoFilePath() {
             return tmpFile.getAbsolutePath();
     }
 
+    // {{ mkl environments set up
+
+    /**
+     * If mkl is linked with parallel mode, one can modify the number of threads omp will use.
+     * It's an subsitute of environment variable: OMP_NUM_THREADS
+     *
+     * @param numThreads how many threads omp will use.
+     */
+    public native static void setNumThreads(int numThreads);
+
+    /**
+     * Turns off the Intel MKL Memory Allocator for Intel MKL functions
+     * to directly use the system malloc/free functions.
+     * It's an substitute of environment variable: MKL_DISABLE_FAST_MM
+     */
     public native static void disableFastMM();
 
+    /**
+     * Sets the time (milliseconds) that a thread should wait before sleeping,
+     * after completing the execution of a parallel region.
+     * It's an substitute of environment variable: KMP_BLOCKTIME
+     *
+     * @param msec the time should wait
+     */
+    public native static void setBlockTime(int msec);
+
+    /**
+     * Sets omp wait policy to passive.
+     */
+    public native static void waitPolicyPassive();
+    // }} mkl environments set up
+
     public native static void vsAdd(int n, float[] a, int aOffset, float[] b, int bOffset,
                                     float[] y, int yOffset);
 

mkl2017-xeon-blas/mkl/pom.xml

@@ -48,8 +48,10 @@
                         <linkerStartOption>-I ${JAVA_HOME}/include/</linkerStartOption>
                         <linkerStartOption>${linkerStartOptionDirectory}</linkerStartOption>
                         <linkerStartOption>-ldl</linkerStartOption>
+                        <linkerStartOption>-liomp5</linkerStartOption>
                         <linkerStartOption>-shared</linkerStartOption>
-                        <linkerStartOption>-mkl=sequential</linkerStartOption>
+                        <linkerStartOption>-mkl=parallel</linkerStartOption>
+                        <linkerStartOption>-qopenmp</linkerStartOption>
                         <linkerStartOption>-static-intel</linkerStartOption>
                         <linkerStartOption>-no-intel-extensions</linkerStartOption>
                     </linkerStartOptions>

mkl2017-xeon-blas/mkl/src/main/c/jni/mkl.c

@@ -12,11 +12,12 @@ extern "C" {
  */
 JNIEXPORT void JNICALL Java_com_intel_analytics_bigdl_mkl_MKL_setNumThreads
   (JNIEnv * env, jclass cls, jint num_threads) {
-  return ;
+  mkl_set_dynamic(0);
+  mkl_set_num_threads(num_threads);
 }
 
 /*
- * Class:     com_intel_analytics_bigdl_mkl_MKL
+ * Class:     com_intel_analytics_bigdl_mkl_disableFastMM
  * Method:    setNumThreads
  * Signature: ()V
  */
@@ -25,6 +26,26 @@ JNIEXPORT void JNICALL Java_com_intel_analytics_bigdl_mkl_MKL_disableFastMM
   mkl_disable_fast_mm();
 }
 
+/*
+ * Class:     com_intel_analytics_bigdl_mkl_setBlockTime
+ * Method:    setNumThreads
+ * Signature: ()V
+ */
+JNIEXPORT void JNICALL Java_com_intel_analytics_bigdl_mkl_MKL_setBlockTime
+  (JNIEnv * env, jclass cls, jint time) {
+  kmp_set_blocktime(0);
+}
+
+/*
+ * Class:     com_intel_analytics_bigdl_mkl_watiPolicyPassive
+ * Method:    setNumThreads
+ * Signature: ()V
+ */
+JNIEXPORT void JNICALL Java_com_intel_analytics_bigdl_mkl_MKL_waitPolicyPassive
+  (JNIEnv * env, jclass cls) {
+  kmp_set_library_throughput();
+}
+
 /*
  * Class:     com_intel_analytics_bigdl_mkl_MKL
  * Method:    getNumThreads

mkl2017-xeon-dnn/mkl/pom.xml

@@ -73,6 +73,7 @@
                         <linkerStartOption>-mkl=parallel</linkerStartOption>
                         <linkerStartOption>-static-intel</linkerStartOption>
                         <linkerStartOption>-qopenmp</linkerStartOption>
+                        <linkerStartOption>-no-intel-extensions</linkerStartOption>
                     </linkerStartOptions>
                     <linkerMiddleOptions>
                         <linkerMiddleOption>-lm</linkerMiddleOption>