[FEA] Support TruncDate expression

[kuhushukla]

Is your feature request related to a problem? Please describe.

   ! <TruncDate> trunc(cast(my_field#860 as date), MM) cannot run on GPU because GPU does not currently support the operator class org.apache.spark.sql.catalyst.expressions.TruncDate

Describe the solution you'd like
Customer query shows this fallback, timezone is UTC and GMT

Describe alternatives you've considered
None
Additional context
None

[revans2]

@kuhushukla So is it only 'MM' ('MONTH') that is needed?

https://spark.apache.org/docs/latest/api/sql/index.html#date_trunc supports

• YEAR
• QUARTER
• MONTH
• WEEK
• DAY
• HOUR
• MINUTE
• SECOND
• MILLISECOND
• MICROSECOND

DAY and below we should be able to support with existing CUDF operators. Beyond that it is probably best if we write a custom kernel for most of it. We could come up with a way to put together a string representation of the timestamp we need by extracting year, month, day, etc and parsing it back with CUDF, but that feels really wasteful. I would rather just do it in a kernel ourselves based on the same kind of logic.

[kuhushukla]

Currently I have only encountered month. I can update further in the next couple days, but MM would be a good start for the customer query in question.

[ttnghia]

Fun facts:

• trunc(date, fmt) is used for date type, while date_trunc(fmt, timestamp) is used for timestamp type.
• They have opposite order of parameters.

I'm going to implement both functions, and support most formats for them except QUARTER and WEEK since these are not trivially computed.

[ttnghia]

I changed my plan: will try to support all. This is a binary expression, and fmt can only be checked at run time, not planning time. As such, if we don't fully support it we won't be able to fallback but will throw exception.

[revans2]

@ttnghia it is perfectly acceptable for us to only support a scalar fmt, and in those cases we can check it a planning time. That is what we do for all of the regular expression commands.

In the type checks you can mark a type as TypeSig.lit

spark-rapids/sql-plugin/src/main/scala/com/nvidia/spark/rapids/GpuOverrides.scala

Lines 3352 to 3361 in 738c8e3

	expr[RegExpReplace](
	"String replace using a regular expression pattern",
	ExprChecks.projectOnly(TypeSig.STRING, TypeSig.STRING,
	Seq(ParamCheck("str", TypeSig.STRING, TypeSig.STRING),
	ParamCheck("regex", TypeSig.lit(TypeEnum.STRING), TypeSig.STRING),
	ParamCheck("rep", TypeSig.lit(TypeEnum.STRING), TypeSig.STRING),
	ParamCheck("pos", TypeSig.lit(TypeEnum.INT)
	.withPsNote(TypeEnum.INT, "only a value of 1 is supported"),
	TypeSig.lit(TypeEnum.INT)))),
	(a, conf, p, r) => new GpuRegExpReplaceMeta(a, conf, p, r)),

That makes it so we would fall back for anything that is not a literal value. Then you can use some pattern matching APIs to pull out the value.

spark-rapids/sql-plugin/src/main/scala/com/nvidia/spark/rapids/GpuOverrides.scala

Lines 582 to 587 in 738c8e3

	def extractStringLit(exp: Expression): Option[String] = extractLit(exp) match {
	case Some(Literal(v: UTF8String, StringType)) =>
	val s = if (v == null) null else v.toString
	Some(s)
	case _ => None
	}

[ttnghia]

That is fine. I've completed the implementation to support all, as it is fairly simple.