[ML-214] [GPU] use distributed covariance as the first step for PCA

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

@@ -24,6 +24,7 @@
 #endif
 #include "Communicator.hpp"
 #include "OutputHelpers.hpp"
+#include "oneapi/dal/algo/covariance.hpp"
 #include "oneapi/dal/algo/pca.hpp"
 #include "oneapi/dal/table/homogen.hpp"
 #endif
@@ -192,30 +193,36 @@ static void doPCAOneAPICompute(JNIEnv *env, jint rankId, jlong pNumTabData,
     const bool isRoot = (rankId == ccl_root);
     homogen_table htable =
         *reinterpret_cast<const homogen_table *>(pNumTabData);
-    const auto pca_desc = pca::descriptor{};
+
+    const auto cov_desc = covariance::descriptor{}.set_result_options(
+        covariance::result_options::cov_matrix);
     auto queue = getQueue(device);
     auto comm = preview::spmd::make_communicator<preview::spmd::backend::ccl>(
         queue, executorNum, rankId, ipPort);
-    pca::train_input local_input{htable};
+
     auto t1 = std::chrono::high_resolution_clock::now();
-    const auto result_train = preview::train(comm, pca_desc, local_input);
+    const auto result = preview::compute(comm, cov_desc, htable);
     auto t2 = std::chrono::high_resolution_clock::now();
     auto duration =
         std::chrono::duration_cast<std::chrono::milliseconds>(t2 - t1).count();
-    std::cout << "PCA (native): training step took " << duration / 1000
+    std::cout << "PCA (native): Covariance step took " << duration / 1000
               << " secs" << std::endl;
     if (isRoot) {
-        std::cout << "Eigenvectors:\n"
-                  << result_train.get_eigenvectors() << std::endl;
-        std::cout << "Eigenvalues:\n"
-                  << result_train.get_eigenvalues() << std::endl;
+        using float_t = double;
+        using method_t = pca::method::precomputed;
+        using task_t = pca::task::dim_reduction;
+        using descriptor_t = pca::descriptor<float_t, method_t, task_t>;
+        const auto pca_desc = descriptor_t().set_deterministic(true);
+
+        auto t1 = std::chrono::high_resolution_clock::now();
+        const auto result_train =
+            train(queue, pca_desc, result.get_cov_matrix());
         t2 = std::chrono::high_resolution_clock::now();
         duration =
             std::chrono::duration_cast<std::chrono::milliseconds>(t2 - t1)
                 .count();
-        std::cout << "PCA (native): rankid " << rankId
-                  << "; training step took " << duration / 1000
-                  << " secs in end. " << std::endl;
+        std::cout << "PCA (native): rankid " << rankId << "; Eigen step took "
+                  << duration / 1000 << " secs in end. " << std::endl;
         // Return all eigenvalues & eigenvectors
         // Get the class of the input object
         jclass clazz = env->GetObjectClass(resultObj);
@@ -224,6 +231,10 @@ static void doPCAOneAPICompute(JNIEnv *env, jint rankId, jlong pNumTabData,
             env->GetFieldID(clazz, "pcNumericTable", "J");
         jfieldID explainedVarianceNumericTableField =
             env->GetFieldID(clazz, "explainedVarianceNumericTable", "J");
+        std::cout << "Eigenvectors:\n"
+                  << result_train.get_eigenvectors() << std::endl;
+        std::cout << "Eigenvalues:\n"
+                  << result_train.get_eigenvalues() << std::endl;
 
         HomogenTablePtr eigenvectors =
             std::make_shared<homogen_table>(result_train.get_eigenvectors());

mllib-dal/src/main/scala/com/intel/oap/mllib/feature/PCADALImpl.scala

@@ -124,8 +124,8 @@ class PCADALImpl(val k: Int,
   }
 
   private[mllib] def getPrincipleComponentsFromOneAPI(table: HomogenTable,
-                                            k: Int,
-                                            device: Common.ComputeDevice): DenseMatrix = {
+                                                      k: Int,
+                                                      device: Common.ComputeDevice): DenseMatrix = {
     val numRows = table.getRowCount.toInt
     val numCols = table.getColumnCount.toInt
     require(k <= numRows, "k should be less or equal to row number")