Added method to remove null_masks if the column has no nulls (#9061)

This PR adds a method to remove the validity vector in cases where there are columns in a Table with no nulls but still have a validity vector.

Authors:
  - Raza Jafri (https://github.com/razajafri)

Approvers:
  - Robert (Bobby) Evans (https://github.com/revans2)
  - David Wendt (https://github.com/davidwendt)

URL: #9061

java/src/main/java/ai/rapids/cudf/ColumnView.java

@@ -101,11 +101,39 @@ public ColumnView(DType type, long rows, Optional<Long> nullCount,
         || !nullCount.isPresent();
   }
 
+  /**
+   * Create a new column view based off of data already on the device. Ref count on the buffers
+   * is not incremented and none of the underlying buffers are owned by this view. The returned
+   * ColumnView is only valid as long as the underlying buffers remain valid. If the buffers are
+   * closed before this ColumnView is closed, it will result in undefined behavior.
+   *
+   * If ownership is needed, call {@link ColumnView#copyToColumnVector}
+   *
+   * @param type           the type of the vector
+   * @param rows           the number of rows in this vector.
+   * @param nullCount      the number of nulls in the dataset.
+   * @param dataBuffer     a host buffer required for nested types including strings and string
+   *                       categories. The ownership doesn't change on this buffer
+   * @param validityBuffer an optional validity buffer. Must be provided if nullCount != 0.
+   *                       The ownership doesn't change on this buffer
+   * @param offsetBuffer   The offsetbuffer for columns that need an offset buffer
+   */
+  public ColumnView(DType type, long rows, Optional<Long> nullCount,
+                    BaseDeviceMemoryBuffer dataBuffer,
+                    BaseDeviceMemoryBuffer validityBuffer, BaseDeviceMemoryBuffer offsetBuffer) {
+    this(type, (int) rows, nullCount.orElse(UNKNOWN_NULL_COUNT).intValue(),
+        dataBuffer, validityBuffer, offsetBuffer, null);
+    assert (!type.isNestedType());
+    assert (nullCount.isPresent() && nullCount.get() <= Integer.MAX_VALUE)
+        || !nullCount.isPresent();
+  }
+
   private ColumnView(DType type, long rows, int nullCount,
                      BaseDeviceMemoryBuffer dataBuffer, BaseDeviceMemoryBuffer validityBuffer,
                      BaseDeviceMemoryBuffer offsetBuffer, ColumnView[] children) {
     this(ColumnVector.initViewHandle(type, (int) rows, nullCount, dataBuffer, validityBuffer,
-        offsetBuffer, Arrays.stream(children).mapToLong(c -> c.getNativeView()).toArray()));
+        offsetBuffer, children == null ? new long[]{} :
+            Arrays.stream(children).mapToLong(c -> c.getNativeView()).toArray()));
   }
 
   /** Creates a ColumnVector from a column view handle
@@ -140,6 +168,32 @@ public final DType getType() {
     return type;
   }
 
+  /**
+   * Returns the child column views for this view
+   * Please note that it is the responsibility of the caller to close these views.
+   * @return an array of child column views
+   */
+  public final ColumnView[] getChildColumnViews() {
+    int numChildren = getNumChildren();
+    if (!getType().isNestedType()) {
+      return null;
+    }
+    ColumnView[] views = new ColumnView[numChildren];
+    try {
+      for (int i = 0; i < numChildren; i++) {
+        views[i] = getChildColumnView(i);
+      }
+      return views;
+    } catch(Throwable t) {
+      for (ColumnView v: views) {
+        if (v != null) {
+          v.close();
+        }
+      }
+      throw t;
+    }
+  }
+
   /**
    * Returns the child column view at a given index.
    * Please note that it is the responsibility of the caller to close this view.

java/src/main/java/ai/rapids/cudf/Table.java

@@ -170,10 +170,19 @@ public long getDeviceMemorySize() {
     return total;
   }
 
+  /**
+   * This method is internal and exposed purely for testing purpopses
+   */
+  static Table removeNullMasksIfNeeded(Table table) {
+    return new Table(removeNullMasksIfNeeded(table.nativeHandle));
+  }
+
   /////////////////////////////////////////////////////////////////////////////
   // NATIVE APIs
   /////////////////////////////////////////////////////////////////////////////
-
+
+  private static native long[] removeNullMasksIfNeeded(long tableView) throws CudfException;
+
   private static native ContiguousTable[] contiguousSplit(long inputTable, int[] indices);
 
   private static native long[] partition(long inputTable, long partitionView,

java/src/main/native/src/TableJni.cpp

@@ -929,13 +929,71 @@ jlongArray combine_join_results(JNIEnv *env, cudf::table &left_results,
   return combine_join_results(env, std::move(left_cols), std::move(right_cols));
 }
 
+cudf::column_view remove_validity_from_col(cudf::column_view column_view) {
+  if (!cudf::is_compound(column_view.type())) {
+    if (column_view.nullable() && column_view.null_count() == 0) {
+      // null_mask is allocated but no nulls present therefore we create a new column_view without
+      // the null_mask to avoid things blowing up in reading the parquet file
+      return cudf::column_view(column_view.type(), column_view.size(), column_view.head(), nullptr,
+                               0, column_view.offset());
+    } else {
+      return cudf::column_view(column_view);
+    }
+  } else {
+    std::unique_ptr<cudf::column_view> ret;
+    std::vector<cudf::column_view> children;
+    children.reserve(column_view.num_children());
+    for (auto it = column_view.child_begin(); it != column_view.child_end(); it++) {
+      children.push_back(remove_validity_from_col(*it));
+    }
+    if (!column_view.nullable() || column_view.null_count() != 0) {
+      ret.reset(new cudf::column_view(column_view.type(), column_view.size(), nullptr,
+                                      column_view.null_mask(), column_view.null_count(),
+                                      column_view.offset(), children));
+    } else {
+      ret.reset(new cudf::column_view(column_view.type(), column_view.size(), nullptr, nullptr, 0,
+                                      column_view.offset(), children));
+    }
+    return *ret.release();
+  }
+}
+
+cudf::table_view remove_validity_if_needed(cudf::table_view *input_table_view) {
+  std::vector<cudf::column_view> views;
+  views.reserve(input_table_view->num_columns());
+  for (auto it = input_table_view->begin(); it != input_table_view->end(); it++) {
+    views.push_back(remove_validity_from_col(*it));
+  }
+
+  return cudf::table_view(views);
+}
+
 } // namespace
 
 } // namespace jni
 } // namespace cudf
 
 extern "C" {
 
+// This is a method purely added for testing remove_validity_if_needed method
+JNIEXPORT jlongArray JNICALL Java_ai_rapids_cudf_Table_removeNullMasksIfNeeded(JNIEnv *env, jclass,
+                                                                               jlong j_table_view) {
+  JNI_NULL_CHECK(env, j_table_view, "table view handle is null", 0);
+  try {
+    cudf::table_view *tview = reinterpret_cast<cudf::table_view *>(j_table_view);
+    cudf::table_view result = cudf::jni::remove_validity_if_needed(tview);
+    cudf::table m_tbl(result);
+    std::vector<std::unique_ptr<cudf::column>> cols = m_tbl.release();
+    auto results = cudf::jni::native_jlongArray(env, cols.size());
+    int i = 0;
+    for (auto it = cols.begin(); it != cols.end(); it++) {
+      results[i++] = reinterpret_cast<jlong>(it->release());
+    }
+    return results.get_jArray();
+  }
+  CATCH_STD(env, 0);
+}
+
 JNIEXPORT jlong JNICALL Java_ai_rapids_cudf_Table_createCudfTableView(JNIEnv *env, jclass,
                                                                       jlongArray j_cudf_columns) {
   JNI_NULL_CHECK(env, j_cudf_columns, "columns are null", 0);
@@ -1357,7 +1415,8 @@ JNIEXPORT void JNICALL Java_ai_rapids_cudf_Table_writeParquetChunk(JNIEnv *env,
   JNI_NULL_CHECK(env, j_state, "null state", );
 
   using namespace cudf::io;
-  cudf::table_view *tview = reinterpret_cast<cudf::table_view *>(j_table);
+  cudf::table_view *tview_with_empty_nullmask = reinterpret_cast<cudf::table_view *>(j_table);
+  cudf::table_view tview = cudf::jni::remove_validity_if_needed(tview_with_empty_nullmask);
   cudf::jni::native_parquet_writer_handle *state =
       reinterpret_cast<cudf::jni::native_parquet_writer_handle *>(j_state);
 
@@ -1367,7 +1426,7 @@ JNIEXPORT void JNICALL Java_ai_rapids_cudf_Table_writeParquetChunk(JNIEnv *env,
   }
   try {
     cudf::jni::auto_set_device(env);
-    state->writer->write(*tview);
+    state->writer->write(tview);
   }
   CATCH_STD(env, )
 }