Fix LEAD/LAG failures in Spark 3.1.1

[mythrocks]

Fixes #999.

Fixes class-loader related errors of LEAD/LAG operators, which caused test failures like:

: java.lang.VerifyError: Bad type on operand stack
Exception Details:
Location:
com/nvidia/spark/rapids/OffsetWindowFunctionMeta.<init>(Lorg/apache/spark/sql/catalyst/expressions/OffsetWindowFunction;Lcom/nvidia/spark/rapids/RapidsConf;Lscala/Option;Lcom/nvidia/spark/rapids/ConfKeysAndIncompat;)V @11: invokespecial
Reason:
Type 'org/apache/spark/sql/catalyst/expressions/OffsetWindowFunction' (current frame, stack[1]) is not assignable to 'org/apache/spark/sql/catalyst/expressions/Expression'

Also corrects for the reversal in offset semantics for LAG() expressions in Spark 3.1.1, causing Lag.offset to be negative.

Signed-off-by: Mithun RK mythrocks@gmail.com

[mythrocks]

At the risk of blowing this up a little, we might consider a couple of changes more:

1. Add similarly robust offset checks in com.nvidia.spark.rapids.OffsetWindowFunctionMeta.
2. Add a tagSelfForGpu(), and fail GPU pushdown in case the offsets in the window are not suitable for acceleration.

[revans2]

At the risk of blowing this up a little, we might consider a couple of changes more:

1. Add similarly robust offset checks in `com.nvidia.spark.rapids.OffsetWindowFunctionMeta`.

2. Add a `tagSelfForGpu()`, and fail GPU pushdown in case the offsets in the window are not suitable for acceleration.

Do it

[mythrocks]

Hmm... I wonder if adding tagExprForGpu might not help.

@revans2, is there a guarantee that tagExprForGpu() will be called before OffsetWindowFunctionMeta::childExprs is evaluated? If not, the exceptions we throw when calculating the adjustedOffset would've been thrown anyway.

[gerashegalov]

LGTM

[mythrocks]

If not, the exceptions we throw when calculating the adjustedOffset would've been thrown anyway.

Apologies for the delay. I have now added tagExprForGpu(), and set it up so that the expressions aren't converted for GPU on construction.

Could I please bother @gerashegalov and @revans2 to see if this is agreeable?

[revans2]

build

[mythrocks]

Thanks for the reviews, @revans2, @gerashegalov.