[
https://issues.apache.org/jira/browse/PARQUET-2159?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17681672#comment-17681672
]
ASF GitHub Bot commented on PARQUET-2159:
-----------------------------------------
jiangjiguang commented on code in PR #1011:
URL: https://github.com/apache/parquet-mr/pull/1011#discussion_r1089861017
##########
parquet-column/src/main/java/org/apache/parquet/column/values/bitpacking/ParquetReadRouter.java:
##########
@@ -0,0 +1,132 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one
+ * or more contributor license agreements. See the NOTICE file
+ * distributed with this work for additional information
+ * regarding copyright ownership. The ASF licenses this file
+ * to you under the Apache License, Version 2.0 (the
+ * "License"); you may not use this file except in compliance
+ * with the License. You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+ * KIND, either express or implied. See the License for the
+ * specific language governing permissions and limitations
+ * under the License.
+ */
+package org.apache.parquet.column.values.bitpacking;
+
+import org.apache.parquet.bytes.ByteBufferInputStream;
+import org.slf4j.Logger;
+import org.slf4j.LoggerFactory;
+
+import java.io.EOFException;
+import java.io.IOException;
+import java.nio.ByteBuffer;
+import java.nio.charset.StandardCharsets;
+import java.nio.file.Files;
+import java.nio.file.Paths;
+import java.util.Arrays;
+import java.util.List;
+import java.util.Set;
+import java.util.stream.Collectors;
+
+/**
+ * This is a utils class which is used for big data applications(such as Spark
Flink).
+ * For Intel CPU, Flags containing avx512vbmi and avx512_vbmi2 can have better
performance gains.
+ */
+public class ParquetReadRouter {
+ private static final Logger LOG =
LoggerFactory.getLogger(ParquetReadRouter.class);
+
+ private static volatile Boolean vector;
+
+ public static void read(int bitWidth, ByteBufferInputStream in, int
currentCount, int[] currentBuffer) throws IOException {
+ if (supportVector()) {
+ readBatchVector(bitWidth, in, currentCount, currentBuffer);
+ } else {
+ readBatchVector(bitWidth, in, currentCount, currentBuffer);
+ }
+ }
+
+ public static void readBatchVector(int bitWidth, ByteBufferInputStream in,
int currentCount, int[] currentBuffer) throws IOException {
+ BytePacker packer = Packer.LITTLE_ENDIAN.newBytePacker(bitWidth);
+ BytePacker packerVector =
Packer.LITTLE_ENDIAN.newBytePackerVector(bitWidth);
+ int valueIndex = 0;
+ int byteIndex = 0;
+ int unpackCount = packerVector.getUnpackCount();
+ int inputByteCountPerVector = packerVector.getUnpackCount() / 8 * bitWidth;
+ int totalByteCount = currentCount * bitWidth / 8;
+
+ // register of avx512 are 512 bits, and can load up to 64 bytes
Review Comment:
@wgtmac It is very difficult to make 64 bytes to be an input parameter. The
PR using only AVX512 Vector to optimize parquet decode. As you said, sometimes
AVX2 outperforms better, but we should know the specific workloads for AVX2 so
that we can make a choice to use AVX2 or AVX512.
> Parquet bit-packing de/encode optimization
> ------------------------------------------
>
> Key: PARQUET-2159
> URL: https://issues.apache.org/jira/browse/PARQUET-2159
> Project: Parquet
> Issue Type: Improvement
> Components: parquet-mr
> Affects Versions: 1.13.0
> Reporter: Fang-Xie
> Assignee: Fang-Xie
> Priority: Major
> Fix For: 1.13.0
>
> Attachments: image-2022-06-15-22-56-08-396.png,
> image-2022-06-15-22-57-15-964.png, image-2022-06-15-22-58-01-442.png,
> image-2022-06-15-22-58-40-704.png
>
>
> Current Spark use Parquet-mr as parquet reader/writer library, but the
> built-in bit-packing en/decode is not efficient enough.
> Our optimization for Parquet bit-packing en/decode with jdk.incubator.vector
> in Open JDK18 brings prominent performance improvement.
> Due to Vector API is added to OpenJDK since 16, So this optimization request
> JDK16 or higher.
> *Below are our test results*
> Functional test is based on open-source parquet-mr Bit-pack decoding
> function: *_public final void unpack8Values(final byte[] in, final int inPos,
> final int[] out, final int outPos)_* __
> compared with our implementation with vector API *_public final void
> unpack8Values_vec(final byte[] in, final int inPos, final int[] out, final
> int outPos)_*
> We tested 10 pairs (open source parquet bit unpacking vs ours optimized
> vectorized SIMD implementation) decode function with bit
> width=\{1,2,3,4,5,6,7,8,9,10}, below are test results:
> !image-2022-06-15-22-56-08-396.png|width=437,height=223!
> We integrated our bit-packing decode implementation into parquet-mr, tested
> the parquet batch reader ability from Spark VectorizedParquetRecordReader
> which get parquet column data by the batch way. We construct parquet file
> with different row count and column count, the column data type is Int32, the
> maximum int value is 127 which satisfies bit pack encode with bit width=7,
> the count of the row is from 10k to 100 million and the count of the column
> is from 1 to 4.
> !image-2022-06-15-22-57-15-964.png|width=453,height=229!
> !image-2022-06-15-22-58-01-442.png|width=439,height=217!
> !image-2022-06-15-22-58-40-704.png|width=415,height=208!
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
