update optimize code

mllib-dal/src/main/native/CorrelationImpl.cpp

@@ -197,9 +197,9 @@ static void doCorrelationOneAPICompute(
 
 JNIEXPORT jlong JNICALL
 Java_com_intel_oap_mllib_stat_CorrelationDALImpl_cCorrelationTrainDAL(
- JNIEnv *env, jobject obj, jint rank, jlong pNumTabData, jlong numRows,
- jlong numCols, jint executorNum, jint executorCores,
- jint computeDeviceOrdinal, jintArray gpuIdxArray, jobject resultObj) {
+ JNIEnv *env, jobject obj, jint rank, jlong pNumTabData, jlong numRows, jlong numCols,
+ jint executorNum, jint executorCores, jint computeDeviceOrdinal,
+ jintArray gpuIdxArray, jstring ip_port, jobject resultObj) {
  logger::println(logger::INFO,
  "oneDAL (native): use DPC++ kernels; device %s",
  ComputeDeviceString[computeDeviceOrdinal].c_str());
@@ -225,23 +225,17 @@ Java_com_intel_oap_mllib_stat_CorrelationDALImpl_cCorrelationTrainDAL(
  }
 #ifdef CPU_GPU_PROFILE
  case ComputeDevice::gpu: {
- int nGpu = env->GetArrayLength(gpuIdxArray);
  logger::println(
  logger::INFO,
- "oneDAL (native): use GPU kernels with %d GPU(s) rankid %d", nGpu,
- rank);
+ "oneDAL (native): use GPU kernels with rankid %d", rank);
 
- jint *gpuIndices = env->GetIntArrayElements(gpuIdxArray, 0);
+ const char *str = env->GetStringUTFChars(ip_port, nullptr);
+ ccl::string ccl_ip_port(str);
+ auto comm = createDalCommunicator(executorNum, rank, ccl_ip_port);
 
- auto queue = getAssignedGPU(device, gpuIndices);
-
- ccl::shared_ptr_class<ccl::kvs> &kvs = getKvs();
- auto comm =
- preview::spmd::make_communicator<preview::spmd::backend::ccl>(
- queue, executorNum, rank, kvs);
  doCorrelationOneAPICompute(env, pNumTabData, numRows, numCols, comm,
  resultObj);
- env->ReleaseIntArrayElements(gpuIdxArray, gpuIndices, 0);
+ env->ReleaseStringUTFChars(ip_port, str);
  break;
  }
 #endif

mllib-dal/src/main/native/DecisionForestClassifierImpl.cpp

@@ -307,35 +307,28 @@ Java_com_intel_oap_mllib_classification_RandomForestClassifierDALImpl_cRFClassif
  jint minObservationsSplitNode, jdouble minWeightFractionLeafNode,
  jdouble minImpurityDecreaseSplitNode, jint maxTreeDepth, jlong seed,
  jint maxBins, jboolean bootstrap, jintArray gpuIdxArray,
- jobject resultObj) {
+ jstring ip_port, jobject resultObj) {
  logger::println(logger::INFO, "oneDAL (native): use DPC++ kernels");
 
  ComputeDevice device = getComputeDeviceByOrdinal(computeDeviceOrdinal);
  switch (device) {
  case ComputeDevice::gpu: {
- int nGpu = env->GetArrayLength(gpuIdxArray);
  logger::println(
  logger::INFO,
- "oneDAL (native): use GPU kernels with %d GPU(s) rankid %d", nGpu,
- rank);
-
- jint *gpuIndices = env->GetIntArrayElements(gpuIdxArray, 0);
-
- ComputeDevice device = getComputeDeviceByOrdinal(computeDeviceOrdinal);
+ "oneDAL (native): use GPU kernels with rankid %d", rank);
 
- auto queue = getAssignedGPU(device, gpuIndices);
+ const char *str = env->GetStringUTFChars(ip_port, nullptr);
+ ccl::string ccl_ip_port(str);
+ auto comm = createDalCommunicator(executorNum, rank, ccl_ip_port);
 
- ccl::shared_ptr_class<ccl::kvs> &kvs = getKvs();
- auto comm =
- preview::spmd::make_communicator<preview::spmd::backend::ccl>(
- queue, executorNum, rank, kvs);
  jobject hashmapObj = doRFClassifierOneAPICompute(
  env, pNumTabFeature, featureRows, featureCols, pNumTabLabel,
  labelCols, executorNum, computeDeviceOrdinal, classCount, treeCount,
  numFeaturesPerNode, minObservationsLeafNode,
  minObservationsSplitNode, minWeightFractionLeafNode,
  minImpurityDecreaseSplitNode, maxTreeDepth, seed, maxBins,
  bootstrap, comm, resultObj);
+ env->ReleaseStringUTFChars(ip_port, str);
  return hashmapObj;
  }
  default: {

mllib-dal/src/main/native/DecisionForestRegressorImpl.cpp

@@ -292,38 +292,32 @@ static jobject doRFRegressorOneAPICompute(
 
 JNIEXPORT jobject JNICALL
 Java_com_intel_oap_mllib_regression_RandomForestRegressorDALImpl_cRFRegressorTrainDAL(
- JNIEnv *env, jobject obj, jint rank, jlong pNumTabFeature,
- jlong featureRows, jlong featureCols, jlong pNumTabLabel, jlong labelCols,
- jint executorNum, jint computeDeviceOrdinal, jint treeCount,
- jint numFeaturesPerNode, jint minObservationsLeafNode, jint maxTreeDepth,
- jlong seed, jint maxbins, jboolean bootstrap, jintArray gpuIdxArray,
- jobject resultObj) {
+ JNIEnv *env, jobject obj, jint rank, jlong pNumTabFeature, jlong featureRows,
+ jlong featureCols, jlong pNumTabLabel, jlong labelCols, jint executorNum,
+ jint computeDeviceOrdinal, jint treeCount, jint numFeaturesPerNode,
+ jint minObservationsLeafNode, jint maxTreeDepth, jlong seed, jint maxbins,
+ jboolean bootstrap, jintArray gpuIdxArray, jstring ip_port, jobject resultObj) {
  logger::println(logger::INFO,
  "OneDAL (native): use DPC++ kernels; device %s",
  ComputeDeviceString[computeDeviceOrdinal].c_str());
 
  ComputeDevice device = getComputeDeviceByOrdinal(computeDeviceOrdinal);
  switch (device) {
  case ComputeDevice::gpu: {
- int nGpu = env->GetArrayLength(gpuIdxArray);
  logger::println(
  logger::INFO,
- "OneDAL (native): use GPU kernels with %d GPU(s) rankid %d", nGpu,
- rank);
-
- jint *gpuIndices = env->GetIntArrayElements(gpuIdxArray, 0);
+ "OneDAL (native): use GPU kernels with rankid %d", rank);
 
- auto queue = getAssignedGPU(device, gpuIndices);
+ const char *str = env->GetStringUTFChars(ip_port, nullptr);
+ ccl::string ccl_ip_port(str);
+ auto comm = createDalCommunicator(executorNum, rank, ccl_ip_port);
 
- ccl::shared_ptr_class<ccl::kvs> &kvs = getKvs();
- auto comm =
- preview::spmd::make_communicator<preview::spmd::backend::ccl>(
- queue, executorNum, rank, kvs);
  jobject hashmapObj = doRFRegressorOneAPICompute(
  env, pNumTabFeature, featureRows, featureCols, pNumTabLabel,
  labelCols, executorNum, computeDeviceOrdinal, treeCount,
  numFeaturesPerNode, minObservationsLeafNode, maxTreeDepth, seed,
  maxbins, bootstrap, comm, resultObj);
+ env->ReleaseStringUTFChars(ip_port, str);
  return hashmapObj;
  }
  default: {

mllib-dal/src/main/native/GPU.cpp

@@ -25,7 +25,6 @@ static std::vector<sycl::device> get_gpus() {
 }
 
 static int getLocalRank(ccl::communicator &comm, int size, int rank) {
- const int MPI_MAX_PROCESSOR_NAME = 128;
  /* Obtain local rank among nodes sharing the same host name */
  char zero = static_cast<char>(0);
  std::vector<char> name(MPI_MAX_PROCESSOR_NAME + 1, zero);
@@ -128,8 +127,6 @@ preview::spmd::communicator<preview::spmd::device_memory_access::usm> createDalC
 
  logger::println(logger::INFO, "OneCCL singleton init took %f secs",
  duration / 1000);
- logger::Logger::getInstance(c_breakdown_name).printLogToFile("rankID was %d, OneCCL singleton init took %f secs.", rank, duration / 1000 );
-
 
  t1 = std::chrono::high_resolution_clock::now();
 
@@ -145,7 +142,6 @@ preview::spmd::communicator<preview::spmd::device_memory_access::usm> createDalC
  .count();
  logger::println(logger::INFO, "OneCCL (native): create kvs took %f secs",
  duration / 1000);
- logger::Logger::getInstance(c_breakdown_name).printLogToFile("rankID was %d, OneCCL create communicator took %f secs.", rank, duration / 1000 );
  sycl::queue queue{gpus[0]};
  t1 = std::chrono::high_resolution_clock::now();
  auto comm =
@@ -155,7 +151,5 @@ preview::spmd::communicator<preview::spmd::device_memory_access::usm> createDalC
  duration =
  (float)std::chrono::duration_cast<std::chrono::milliseconds>(t2 - t1)
  .count();
- logger::Logger::getInstance(c_breakdown_name).printLogToFile("rankID was %d, create communicator took %f secs.", rank, duration / 1000 );
  return comm;
 }
-

mllib-dal/src/main/native/KMeansImpl.cpp

@@ -305,10 +305,10 @@ static jlong doKMeansOneAPICompute(
  */
 JNIEXPORT jlong JNICALL
 Java_com_intel_oap_mllib_clustering_KMeansDALImpl_cKMeansOneapiComputeWithInitCenters(
- JNIEnv *env, jobject obj, jint rank, jlong pNumTabData, jlong numRows,
- jlong numCols, jlong pNumTabCenters, jint clusterNum, jdouble tolerance,
- jint iterationNum, jint executorNum, jint executorCores,
- jint computeDeviceOrdinal, jintArray gpuIdxArray, jobject resultObj) {
+ JNIEnv *env, jobject obj, jint rank, jlong pNumTabData, jlong numRows, jlong numCols,
+ jlong pNumTabCenters, jint clusterNum, jdouble tolerance, jint iterationNum,
+ jint executorNum, jint executorCores, jint computeDeviceOrdinal,
+ jintArray gpuIdxArray, jstring ip_port, jobject resultObj) {
  logger::println(logger::INFO,
  "OneDAL (native): use DPC++ kernels; device %s",
  ComputeDeviceString[computeDeviceOrdinal].c_str());

mllib-dal/src/main/native/LinearRegressionImpl.cpp

@@ -214,7 +214,8 @@ ridge_regression_compute(size_t rankId, ccl::communicator &comm,
 }
 
 #ifdef CPU_GPU_PROFILE
-static jlong doLROneAPICompute(JNIEnv *env, size_t rankId, sycl::queue &queue,
+static jlong doLROneAPICompute(JNIEnv *env, size_t rankId,
+ preview::spmd::communicator<preview::spmd::device_memory_access::usm> comm,
  jlong pNumTabFeature, jlong featureRows,
  jlong featureCols, jlong pNumTabLabel,
  jlong labelCols, jboolean jfitIntercept,
@@ -224,9 +225,6 @@ static jlong doLROneAPICompute(JNIEnv *env, size_t rankId, sycl::queue &queue,
  const bool isRoot = (rankId == ccl_root);
  bool fitIntercept = bool(jfitIntercept);
 
- ccl::shared_ptr_class<ccl::kvs> &kvs = getKvs();
- auto comm = preview::spmd::make_communicator<preview::spmd::backend::ccl>(
- queue, executorNum, rankId, kvs);
  homogen_table xtrain = *reinterpret_cast<homogen_table *>(
  createHomogenTableWithArrayPtr(pNumTabFeature, featureRows, featureCols,
  comm.get_queue())
@@ -264,7 +262,7 @@ Java_com_intel_oap_mllib_regression_LinearRegressionDALImpl_cLinearRegressionTra
  jlong featureCols, jlong label, jlong labelCols, jboolean fitIntercept,
  jdouble regParam, jdouble elasticNetParam, jint executorNum,
  jint executorCores, jint computeDeviceOrdinal, jintArray gpuIdxArray,
- jobject resultObj) {
+ jstring ip_port, jobject resultObj) {
 
  logger::println(logger::INFO,
  "oneDAL (native): use DPC++ kernels; device %s",
@@ -279,19 +277,18 @@ Java_com_intel_oap_mllib_regression_LinearRegressionDALImpl_cLinearRegressionTra
  jlong resultptr = 0L;
  if (useGPU) {
 #ifdef CPU_GPU_PROFILE
- int nGpu = env->GetArrayLength(gpuIdxArray);
  logger::println(
  logger::INFO,
- "oneDAL (native): use GPU kernels with %d GPU(s) rankid %d", nGpu,
- rank);
+ "oneDAL (native): use GPU kernels with rankid %d", rank);
 
- jint *gpuIndices = env->GetIntArrayElements(gpuIdxArray, 0);
- auto queue = getAssignedGPU(device, gpuIndices);
+ const char *str = env->GetStringUTFChars(ip_port, nullptr);
+ ccl::string ccl_ip_port(str);
+ auto comm = createDalCommunicator(executorNum, rank, ccl_ip_port);
 
- resultptr = doLROneAPICompute(env, rank, queue, feature, featureRows,
-  featureCols, label, labelCols,
-   fitIntercept, executorNum, resultObj);
- env->ReleaseIntArrayElements(gpuIdxArray, gpuIndices, 0);
+ resultptr = doLROneAPICompute(
+ env, rank, comm, feature, featureRows, featureCols,
+ label, labelCols, fitIntercept, executorNum, resultObj);
+ env->ReleaseStringUTFChars(ip_port, str);
 #endif
  } else {
  ccl::communicator &cclComm = getComm();

mllib-dal/src/main/native/OneCCL.cpp

@@ -55,34 +55,32 @@ JNIEXPORT jint JNICALL Java_com_intel_oap_mllib_OneCCL_00024_c_1init(
  auto t1 = std::chrono::high_resolution_clock::now();
 
  ccl::init();
+ auto t2 = std::chrono::high_resolution_clock::now();
+
 
  const char *str = env->GetStringUTFChars(ip_port, 0);
  ccl::string ccl_ip_port(str);
- const char *device = env->GetStringUTFChars(use_device, 0);
- ccl::string ccl_ip_port(str);
 
  auto &singletonCCLInit = CCLInitSingleton::get(size, rank, ccl_ip_port);
 
  g_kvs.push_back(singletonCCLInit.kvs);
-
-#ifdef CPU_ONLY_PROFILE
  g_comms.push_back(
  ccl::create_communicator(size, rank, singletonCCLInit.kvs));
-
- auto t2 = std::chrono::high_resolution_clock::now();
  auto duration =
  (float)std::chrono::duration_cast<std::chrono::milliseconds>(t2 - t1)
  .count();
  logger::println(logger::INFO, "OneCCL (native): init took %f secs",
  duration / 1000);
-#endif
+
+ rank_id = getComm().rank();
+ comm_size = getComm().size();
 
  jclass cls = env->GetObjectClass(param);
  jfieldID fid_comm_size = env->GetFieldID(cls, "commSize", "J");
  jfieldID fid_rank_id = env->GetFieldID(cls, "rankId", "J");
 
- env->SetLongField(param, size, comm_size);
- env->SetLongField(param, rank, rank_id);
+ env->SetLongField(param, fid_comm_size, comm_size);
+ env->SetLongField(param, fid_rank_id, rank_id);
  env->ReleaseStringUTFChars(ip_port, str);
 
  return 1;

mllib-dal/src/main/native/PCAImpl.cpp

@@ -250,9 +250,9 @@ static void doPCAOneAPICompute(
 
 JNIEXPORT jlong JNICALL
 Java_com_intel_oap_mllib_feature_PCADALImpl_cPCATrainDAL(
- JNIEnv *env, jobject obj, jint rank, jlong pNumTabData, jlong numRows,
- jlong numCols, jint executorNum, jint executorCores,
- jint computeDeviceOrdinal, jintArray gpuIdxArray, jobject resultObj) {
+ JNIEnv *env, jobject obj, jint rank, jlong pNumTabData, jlong numRows, jlong numCols,
+ jint executorNum, jint executorCores, jint computeDeviceOrdinal,
+ jintArray gpuIdxArray, jstring ip_port, jobject resultObj) {
  logger::println(logger::INFO,
  "oneDAL (native): use DPC++ kernels; device %s",
  ComputeDeviceString[computeDeviceOrdinal].c_str());
@@ -277,22 +277,16 @@ Java_com_intel_oap_mllib_feature_PCADALImpl_cPCATrainDAL(
  }
 #ifdef CPU_GPU_PROFILE
  case ComputeDevice::gpu: {
- int nGpu = env->GetArrayLength(gpuIdxArray);
  logger::println(
  logger::INFO,
- "oneDAL (native): use GPU kernels with %d GPU(s) rankid %d", nGpu,
- rank);
+ "oneDAL (native): use GPU kernels with rankid %d", rank);
 
- jint *gpuIndices = env->GetIntArrayElements(gpuIdxArray, 0);
+ const char *str = env->GetStringUTFChars(ip_port, nullptr);
+ ccl::string ccl_ip_port(str);
+ auto comm = createDalCommunicator(executorNum, rank, ccl_ip_port);
 
- auto queue = getAssignedGPU(device, gpuIndices);
-
- ccl::shared_ptr_class<ccl::kvs> &kvs = getKvs();
- auto comm =
- preview::spmd::make_communicator<preview::spmd::backend::ccl>(
- queue, executorNum, rank, kvs);
  doPCAOneAPICompute(env, pNumTabData, numRows, numCols, comm, resultObj);
- env->ReleaseIntArrayElements(gpuIdxArray, gpuIndices, 0);
+ env->ReleaseStringUTFChars(ip_port, str);
  break;
  }
 #endif

mllib-dal/src/main/native/SummarizerImpl.cpp

@@ -268,15 +268,9 @@ static void doSummarizerOneAPICompute(
 
 JNIEXPORT jlong JNICALL
 Java_com_intel_oap_mllib_stat_SummarizerDALImpl_cSummarizerTrainDAL(
-<<<<<<< HEAD
- JNIEnv *env, jobject obj, jint rank, jlong pNumTabData, jlong numRows,
- jlong numCols, jint executorNum, jint executorCores,
- jint computeDeviceOrdinal, jintArray gpuIdxArray, jobject resultObj) {
-=======
  JNIEnv *env, jobject obj, jint rank, jlong pNumTabData, jlong numRows, jlong numCols,
  jint executorNum, jint executorCores, jint computeDeviceOrdinal,
- jintArray gpuIdxArray, jobject resultObj) {
->>>>>>> remove oneccl communicator
+ jintArray gpuIdxArray, jstring ip_port, jobject resultObj) {
  logger::println(logger::INFO,
  "oneDAL (native): use DPC++ kernels; device %s",
  ComputeDeviceString[computeDeviceOrdinal].c_str());
@@ -301,28 +295,21 @@ Java_com_intel_oap_mllib_stat_SummarizerDALImpl_cSummarizerTrainDAL(
  }
 #ifdef CPU_GPU_PROFILE
  case ComputeDevice::gpu: {
- int nGpu = env->GetArrayLength(gpuIdxArray);
  logger::println(
  logger::INFO,
- "oneDAL (native): use GPU kernels with %d GPU(s) rankid %d", nGpu,
- rank);
+ "oneDAL (native): use GPU kernels with rankid %d", rank);
+ const char *str = env->GetStringUTFChars(ip_port, nullptr);
+ ccl::string ccl_ip_port(str);
+ auto comm = createDalCommunicator(executorNum, rank, ccl_ip_port);
 
- jint *gpuIndices = env->GetIntArrayElements(gpuIdxArray, 0);
- auto queue = getAssignedGPU(device, gpuIndices);
-
- ccl::shared_ptr_class<ccl::kvs> &kvs = getKvs();
- auto comm =
- preview::spmd::make_communicator<preview::spmd::backend::ccl>(
- queue, executorNum, rank, kvs);
  doSummarizerOneAPICompute(env, pNumTabData, numRows, numCols, comm,
  resultObj);
- env->ReleaseIntArrayElements(gpuIdxArray, gpuIndices, 0);
+ env->ReleaseStringUTFChars(ip_port, str);
  break;
  }
 #endif
  default: {
- deviceError("Summarizer",
- ComputeDeviceString[computeDeviceOrdinal].c_str());
+ deviceError("Summarizer", ComputeDeviceString[computeDeviceOrdinal].c_str());
  }
  }
  return 0;