JNI: Rewrite growBuffersAndRows to accelerate the HostColumnBuilder (#…

…10025)

According to NVIDIA/spark-rapids#4393, current PR takes several measures to speed up the buffer growing during the build of `HostColumnVector`:
1. Introduce `rowCapacity` to cache the maximum number of rows/bytes 
2. Introduce pura Java method `byteSizeOfNullMask` to get the size of the validity buffer
3. Reorganize the code structure to reduce the number of method calls

I have tested this PR with the spark-rapids tests locally.
BTW, shall we clean up the `HostColumnVector.Builder` and replace all the usages of `Builder` with `ColumnBuilder`?

Authors:
  - Alfred Xu (https://github.com/sperlingxx)

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

URL: #10025

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

@@ -3248,7 +3248,7 @@ public final ColumnVector listIndexOf(Scalar key, FindOptions findOption) {
    * The index is set to null if one of the following is true: 
    * 1. The search key row is null.
    * 2. The list row is null.
-   * @param key ColumnView of search keys.
+   * @param keys ColumnView of search keys.
    * @param findOption Whether to find the first index of the key, or the last.
    * @return The resultant column of int32 indices
    */
@@ -3284,6 +3284,17 @@ public final Scalar getScalarElement(int index) {
     return new Scalar(getType(), getElement(getNativeView(), index));
   }
 
+  /**
+   * Get the number of bytes needed to allocate a validity buffer for the given number of rows.
+   * According to cudf::bitmask_allocation_size_bytes, the padding boundary for null mask is 64 bytes.
+   */
+  static long getValidityBufferSize(int numRows) {
+    // number of bytes required = Math.ceil(number of bits / 8)
+    long actualBytes = ((long) numRows + 7) >> 3;
+    // padding to the multiplies of the padding boundary(64 bytes)
+    return ((actualBytes + 63) >> 6) << 6;
+  }
+
   /////////////////////////////////////////////////////////////////////////////
   // INTERNAL/NATIVE ACCESS
   /////////////////////////////////////////////////////////////////////////////
@@ -3701,7 +3712,7 @@ private static native long stringReplaceWithBackrefs(long columnView, String pat
    * Native method to find the first (or last) index of each search key in the specified column,
    * in each row of a list column.
    * @param nativeView the column view handle of the list
-   * @param scalarColumnHandle handle to the search key column
+   * @param keyColumnHandle handle to the search key column
    * @param isFindFirst Whether to find the first index of the key, or the last.
    * @return column handle of the resultant column of int32 indices
    */
@@ -3881,11 +3892,6 @@ private static native long bitwiseMergeAndSetValidity(long baseHandle, long[] vi
   private static native long copyWithBooleanColumnAsValidity(long exemplarViewHandle, 
                                                              long boolColumnViewHandle) throws CudfException;
 
-  /**
-   * Get the number of bytes needed to allocate a validity buffer for the given number of rows.
-   */
-  static native long getNativeValidPointerSize(int size);
-
   ////////
   // Native cudf::column_view life cycle and metadata access methods. Life cycle methods
   // should typically only be called from the OffHeap inner class.
@@ -3975,7 +3981,7 @@ static ColumnVector createColumnVector(DType type, int rows, HostMemoryBuffer da
       DeviceMemoryBuffer mainValidDevBuff = null;
       DeviceMemoryBuffer mainOffsetsDevBuff = null;
       if (mainColValid != null) {
-        long validLen = getNativeValidPointerSize(mainColRows);
+        long validLen = getValidityBufferSize(mainColRows);
         mainValidDevBuff = DeviceMemoryBuffer.allocate(validLen);
         mainValidDevBuff.copyFromHostBuffer(mainColValid, 0, validLen);
       }
@@ -4084,7 +4090,7 @@ private static NestedColumnVector createNestedColumnVector(DType type, long rows
         data.copyFromHostBuffer(dataBuffer, 0, dataLen);
       }
       if (validityBuffer != null) {
-        long validLen = getNativeValidPointerSize((int)rows);
+        long validLen = getValidityBufferSize((int)rows);
         valid = DeviceMemoryBuffer.allocate(validLen);
         valid.copyFromHostBuffer(validityBuffer, 0, validLen);
       }

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

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2019-2020, NVIDIA CORPORATION.
+ * Copyright (c) 2019-2022, NVIDIA CORPORATION.
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
@@ -307,7 +307,7 @@ public static DType fromJavaBigDecimal(BigDecimal dec) {
       return new DType(DTypeEnum.DECIMAL128, -dec.scale());
     }
     throw new IllegalArgumentException("Precision " + dec.precision() +
-        " exceeds max precision cuDF can support " + DECIMAL64_MAX_PRECISION);
+        " exceeds max precision cuDF can support " + DECIMAL128_MAX_PRECISION);
   }
 
   /**