### What changes were proposed in this pull request?
Use `SpecificInternalRow` instead of `GenericInternalRow` to avoid boxing / unboxing cost.
### Why are the changes needed?
Since it doesn't know the input row schema, `GenericInternalRow` potentially need to apply boxing for input arguments. It's better to use `SpecificInternalRow` instead since we know input data types.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
Existing tests.
Closes#32647 from sunchao/specific-input-row.
Authored-by: Chao Sun <sunchao@apple.com>
Signed-off-by: Hyukjin Kwon <gurwls223@apache.org>
### What changes were proposed in this pull request?
Move hash map lookup operation out of `InvokeLike.invoke` since it doesn't depend on the input.
### Why are the changes needed?
We shouldn't need to look up the hash map for every input row evaluated by `InvokeLike.invoke` since it doesn't depend on input. This could speed up the performance a bit.
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
Existing tests.
Closes#32532 from sunchao/SPARK-35384-follow-up.
Authored-by: Chao Sun <sunchao@apple.com>
Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
### What changes were proposed in this pull request?
Change `map` in `InvokeLike.invoke` to a while loop to improve performance, following Spark [style guide](https://github.com/databricks/scala-style-guide#traversal-and-zipwithindex).
### Why are the changes needed?
`InvokeLike.invoke`, which is used in non-codegen path for `Invoke` and `StaticInvoke`, currently uses `map` to evaluate arguments:
```scala
val args = arguments.map(e => e.eval(input).asInstanceOf[Object])
if (needNullCheck && args.exists(_ == null)) {
// return null if one of arguments is null
null
} else {
...
```
which is pretty expensive if the method itself is trivial. We can change it to a plain while loop.
<img width="871" alt="Screen Shot 2021-05-12 at 12 19 59 AM" src="https://user-images.githubusercontent.com/506679/118055719-7f985a00-b33d-11eb-943b-cf85eab35f44.png">
Benchmark results show this can improve as much as 3x from `V2FunctionBenchmark`:
Before
```
OpenJDK 64-Bit Server VM 1.8.0_292-b10 on Linux 5.4.0-1046-azure
Intel(R) Xeon(R) CPU E5-2673 v3 2.40GHz
scalar function (long + long) -> long, result_nullable = false codegen = false: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
--------------------------------------------------------------------------------------------------------------------------------------------------------------
native_long_add 36506 36656 251 13.7 73.0 1.0X
java_long_add_default 47151 47540 370 10.6 94.3 0.8X
java_long_add_magic 178691 182457 1327 2.8 357.4 0.2X
java_long_add_static_magic 177151 178258 1151 2.8 354.3 0.2X
```
After
```
OpenJDK 64-Bit Server VM 1.8.0_292-b10 on Linux 5.4.0-1046-azure
Intel(R) Xeon(R) CPU E5-2673 v3 2.40GHz
scalar function (long + long) -> long, result_nullable = false codegen = false: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
--------------------------------------------------------------------------------------------------------------------------------------------------------------
native_long_add 29897 30342 568 16.7 59.8 1.0X
java_long_add_default 40628 41075 664 12.3 81.3 0.7X
java_long_add_magic 54553 54755 182 9.2 109.1 0.5X
java_long_add_static_magic 55410 55532 127 9.0 110.8 0.5X
```
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
Existing tests.
Closes#32527 from sunchao/SPARK-35384.
Authored-by: Chao Sun <sunchao@apple.com>
Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
### What changes were proposed in this pull request?
Switch to plain `while` loop following Spark [style guide](https://github.com/databricks/scala-style-guide#traversal-and-zipwithindex).
### Why are the changes needed?
`while` loop may yield better performance comparing to `foreach`.
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
N/A
Closes#32522 from sunchao/SPARK-35361-follow-up.
Authored-by: Chao Sun <sunchao@apple.com>
Signed-off-by: Dongjoon Hyun <dongjoon@apache.org>
### What changes were proposed in this pull request?
In `ApplyFunctionExpression`, move `zipWithIndex` out of the loop for each input row.
### Why are the changes needed?
When the `ScalarFunction` is trivial, `zipWithIndex` could incur significant costs, as shown below:
<img width="899" alt="Screen Shot 2021-05-11 at 10 03 42 AM" src="https://user-images.githubusercontent.com/506679/117866421-fb19de80-b24b-11eb-8c94-d5e8c8b1eda9.png">
By removing it out of the loop, I'm seeing sometimes 2x speedup from `V2FunctionBenchmark`. For instance:
Before:
```
scalar function (long + long) -> long, result_nullable = false codegen = false: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
native_long_add 32437 32896 434 15.4 64.9 1.0X
java_long_add_default 85675 97045 NaN 5.8 171.3 0.4X
```
After:
```
scalar function (long + long) -> long, result_nullable = false codegen = false: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
native_long_add 30182 30387 279 16.6 60.4 1.0X
java_long_add_default 42862 43009 209 11.7 85.7 0.7X
```
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
Existing tests
Closes#32507 from sunchao/SPARK-35361.
Authored-by: Chao Sun <sunchao@apple.com>
Signed-off-by: Hyukjin Kwon <gurwls223@apache.org>
### What changes were proposed in this pull request?
Change `failOnError` to false for `NativeAdd` in `V2FunctionBenchmark`.
### Why are the changes needed?
Since `NativeAdd` is simply doing addition on long it's better to set `failOnError` to false so it will use native long addition instead of `Math.addExact`.
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
N/A
Closes#32481 from sunchao/SPARK-35261-follow-up.
Authored-by: Chao Sun <sunchao@apache.org>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
### What changes were proposed in this pull request?
This allows `ScalarFunction` implemented in Java to optionally specify the magic method `invoke` to be static, which can be used if the UDF is stateless. Comparing to the non-static method, it can potentially give better performance due to elimination of dynamic dispatch, etc.
Also added a benchmark to measure performance of: the default `produceResult`, non-static magic method and static magic method.
### Why are the changes needed?
For UDFs that are stateless (e.g., no need to maintain intermediate state between each function call), it's better to allow users to implement the UDF function as static method which could potentially give better performance.
### Does this PR introduce _any_ user-facing change?
Yes. Spark users can now have the choice to define static magic method for `ScalarFunction` when it is written in Java and when the UDF is stateless.
### How was this patch tested?
Added new UT.
Closes#32407 from sunchao/SPARK-35261.
Authored-by: Chao Sun <sunchao@apache.org>
Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
