MaxGekk commented on a change in pull request #28406:
URL: https://github.com/apache/spark/pull/28406#discussion_r418217773
##########
File path:
sql/core/src/main/java/org/apache/spark/sql/execution/datasources/parquet/VectorizedPlainValuesReader.java
##########
@@ -96,6 +124,33 @@ public final void readLongs(int total, WritableColumnVector
c, int rowId) {
}
}
+ // A fork of `readLongs` to rebase the timestamp values. For performance
reasons, this method
+ // iterates the values twice: check if we need to rebase first, then go to
the optimized branch
+ // if rebase is not needed.
+ @Override
+ public final void readLongsWithRebase(int total, WritableColumnVector c, int
rowId) {
+ int requiredBytes = total * 8;
+ ByteBuffer buffer = getBuffer(requiredBytes);
+ boolean rebase = false;
+ for (int i = 0; i < total; i += 1) {
Review comment:
The byte code of the loop is:
```
LINENUMBER 136 L6
ILOAD 6
ALOAD 5
ALOAD 5
INVOKEVIRTUAL java/nio/ByteBuffer.position ()I
ILOAD 7
BIPUSH 8
IMUL
IADD
INVOKEVIRTUAL java/nio/ByteBuffer.getLong (I)J
INVOKESTATIC
org/apache/spark/sql/catalyst/util/RebaseDateTime.lastSwitchJulianTs ()J
LCMP
IFGE L7
ICONST_1
GOTO L8
```
We could avoid mul like
```java
int pos = buffer.position();
int endPos = pos + total * 8;
long threshold = RebaseDateTime.lastSwitchJulianTs();
while (pos < endPos) {
rebase |= buffer.getLong(pos) < threshold;
pos += 8;
}
```
Would it be faster?
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
