Support from_utc_timestamp on the GPU for non-UTC timezones (non-DST)

integration_tests/src/main/python/date_time_test.py

@@ -281,9 +281,9 @@ def test_from_utc_timestamp_unsupported_timezone_fallback(data_gen, time_zone):
 @pytest.mark.parametrize('time_zone', ["UTC", "Asia/Shanghai", "EST", "MST", "VST"], ids=idfn)
 @pytest.mark.parametrize('data_gen', [timestamp_gen], ids=idfn)
 def test_from_utc_timestamp_supported_timezones(data_gen, time_zone):
-    # Remove spark.rapids.test.CPU.timezone configuration when GPU kernel is ready to really test on GPU
+    # TODO: Remove spark.rapids.sql.nonUTC.enabled configuration 
     assert_gpu_and_cpu_are_equal_collect(
-        lambda spark: unary_op_df(spark, data_gen).select(f.from_utc_timestamp(f.col('a'), time_zone)), conf = {"spark.rapids.test.CPU.timezone": "true"})
+        lambda spark: unary_op_df(spark, data_gen).select(f.from_utc_timestamp(f.col('a'), time_zone)), conf = {"spark.rapids.sql.nonUTC.enabled": "true"})
 
 
 @allow_non_gpu('ProjectExec')

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

@@ -621,6 +621,10 @@ object GpuOverrides extends Logging {
     }
   }
 
+  def isUTCTimezone(timezoneId: ZoneId): Boolean = {
+    timezoneId.normalized() == UTC_TIMEZONE_ID
+  }
+
   def areAllSupportedTypes(types: DataType*): Boolean = types.forall(isSupportedType(_))
 
   /**

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

@@ -26,6 +26,7 @@ import scala.util.Try
 
 import ai.rapids.cudf.{Cuda, CudaException, CudaFatalException, CudfException, MemoryCleaner}
 import com.nvidia.spark.rapids.filecache.{FileCache, FileCacheLocalityManager, FileCacheLocalityMsg}
+import com.nvidia.spark.rapids.jni.GpuTimeZoneDB
 import com.nvidia.spark.rapids.python.PythonWorkerSemaphore
 import org.apache.commons.lang3.exception.ExceptionUtils
 
@@ -503,6 +504,7 @@ class RapidsExecutorPlugin extends ExecutorPlugin with Logging {
   }
 
   override def shutdown(): Unit = {
+    GpuTimeZoneDB.shutdown()
     GpuSemaphore.shutdown()
     PythonWorkerSemaphore.shutdown()
     GpuDeviceManager.shutdown()

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

@@ -2044,6 +2044,13 @@ object RapidsConf {
         "The gpu to disk spill bounce buffer must have a positive size")
       .createWithDefault(128L * 1024 * 1024)
 
+  val NON_UTC_TIME_ZONE_ENABLED = 
+    conf("spark.rapids.sql.nonUTC.enabled")
+      .doc("An option to enable/disable non-UTC time zone support.")
+      .internal()
+      .booleanConf
+      .createWithDefault(false)
+
   val SPLIT_UNTIL_SIZE_OVERRIDE = conf("spark.rapids.sql.test.overrides.splitUntilSize")
       .doc("Only for tests: override the value of GpuDeviceManager.splitUntilSize")
       .internal()
@@ -2056,12 +2063,6 @@ object RapidsConf {
     .booleanConf
     .createOptional
 
-  val TEST_USE_TIMEZONE_CPU_BACKEND = conf("spark.rapids.test.CPU.timezone")
-    .doc("Only for tests: verify for timezone related functions")
-    .internal()
-    .booleanConf
-    .createOptional
-
   private def printSectionHeader(category: String): Unit =
     println(s"\n### $category")
 
@@ -2754,6 +2755,8 @@ class RapidsConf(conf: Map[String, String]) extends Logging {
 
   lazy val splitUntilSizeOverride: Option[Long] = get(SPLIT_UNTIL_SIZE_OVERRIDE)
 
+  lazy val nonUTCTimeZoneEnabled: Boolean = get(NON_UTC_TIME_ZONE_ENABLED)
+
   private val optimizerDefaults = Map(
     // this is not accurate because CPU projections do have a cost due to appending values
     // to each row that is produced, but this needs to be a really small number because

sql-plugin/src/main/scala/org/apache/spark/sql/rapids/TimeZoneDB.scala

@@ -20,15 +20,10 @@ import java.time.ZoneId
 
 import ai.rapids.cudf.{ColumnVector, DType, HostColumnVector}
 import com.nvidia.spark.rapids.Arm.withResource
-import com.nvidia.spark.rapids.GpuOverrides
 
 import org.apache.spark.sql.catalyst.util.DateTimeUtils
 
 object TimeZoneDB {
-  def isUTCTimezone(timezoneId: ZoneId): Boolean = {
-    timezoneId.normalized() == GpuOverrides.UTC_TIMEZONE_ID
-  }
-
   // Copied from Spark. Used to format time zone ID string with (+|-)h:mm and (+|-)hh:m
   def getZoneId(timezoneId: String): ZoneId = {
     val formattedZoneId = timezoneId
@@ -40,7 +35,7 @@ object TimeZoneDB {
   }
 
   // Support fixed offset or no transition rule case
-  def isSupportedTimezone(timezoneId: String): Boolean = {
+  def isSupportedTimeZone(timezoneId: String): Boolean = {
     val rules = getZoneId(timezoneId).getRules
     rules.isFixedOffset || rules.getTransitionRules.isEmpty
   }
@@ -153,7 +148,7 @@ object TimeZoneDB {
     assert(inputVector.getType == DType.TIMESTAMP_DAYS)
     val rowCount = inputVector.getRowCount.toInt
     withResource(inputVector.copyToHost()) { input =>
-      withResource(HostColumnVector.builder(DType.INT64, rowCount)) { builder =>
+      withResource(HostColumnVector.builder(DType.TIMESTAMP_MICROSECONDS, rowCount)) { builder =>
         var currRow = 0
         while (currRow < rowCount) {
           if (input.isNull(currRow)) {

sql-plugin/src/main/scala/org/apache/spark/sql/rapids/datetimeExpressions.scala

@@ -20,11 +20,11 @@ import java.time.ZoneId
 import java.util.concurrent.TimeUnit
 
 import ai.rapids.cudf.{BinaryOp, CaptureGroups, ColumnVector, ColumnView, DType, RegexProgram, Scalar}
-import com.nvidia.spark.rapids.{BinaryExprMeta, BoolUtils, DataFromReplacementRule, DateUtils, GpuBinaryExpression, GpuBinaryExpressionArgsAnyScalar, GpuCast, GpuColumnVector, GpuExpression, GpuScalar, GpuUnaryExpression, RapidsConf, RapidsMeta}
+import com.nvidia.spark.rapids.{BinaryExprMeta, BoolUtils, DataFromReplacementRule, DateUtils, GpuBinaryExpression, GpuBinaryExpressionArgsAnyScalar, GpuCast, GpuColumnVector, GpuExpression, GpuOverrides, GpuScalar, GpuUnaryExpression, RapidsConf, RapidsMeta}
 import com.nvidia.spark.rapids.Arm._
 import com.nvidia.spark.rapids.GpuOverrides.{extractStringLit, getTimeParserPolicy}
-import com.nvidia.spark.rapids.RapidsConf.TEST_USE_TIMEZONE_CPU_BACKEND
 import com.nvidia.spark.rapids.RapidsPluginImplicits._
+import com.nvidia.spark.rapids.jni.GpuTimeZoneDB
 import com.nvidia.spark.rapids.shims.ShimBinaryExpression
 
 import org.apache.spark.sql.catalyst.expressions.{BinaryExpression, ExpectsInputTypes, Expression, FromUTCTimestamp, ImplicitCastInputTypes, NullIntolerant, TimeZoneAwareExpression}
@@ -1046,20 +1046,19 @@ class FromUTCTimestampExprMeta(
   extends BinaryExprMeta[FromUTCTimestamp](expr, conf, parent, rule) {
 
   private[this] var timezoneId: ZoneId = null
-  private[this] val isOnCPU: Boolean = conf.get(TEST_USE_TIMEZONE_CPU_BACKEND).getOrElse(false)
+  private[this] val nonUTCEnabled: Boolean = conf.nonUTCTimeZoneEnabled
 
   override def tagExprForGpu(): Unit = {
     extractStringLit(expr.right) match {
       case None =>
         willNotWorkOnGpu("timezone input must be a literal string")
       case Some(timezoneShortID) =>
         if (timezoneShortID != null) {
-          timezoneId = TimeZoneDB.getZoneId(timezoneShortID)
+          timezoneId = GpuTimeZoneDB.getZoneId(timezoneShortID)
           // Always pass for UTC timezone since it's no-op.
-          if (!TimeZoneDB.isUTCTimezone(timezoneId)) {
-            // Check CPU path, mostly for test purpose
-            if (isOnCPU) {
-              if(!TimeZoneDB.isSupportedTimezone(timezoneShortID)) {
+          if (!GpuOverrides.isUTCTimezone(timezoneId)) {
+            if (nonUTCEnabled) {
+              if(!GpuTimeZoneDB.isSupportedTimeZone(timezoneShortID)) {
                 willNotWorkOnGpu(s"Not supported timezone type $timezoneShortID.")
               }
             } else {
@@ -1072,11 +1071,11 @@ class FromUTCTimestampExprMeta(
   }
 
   override def convertToGpu(timestamp: Expression, timezone: Expression): GpuExpression =
-    GpuFromUTCTimestamp(timestamp, timezone, timezoneId, isOnCPU)
+    GpuFromUTCTimestamp(timestamp, timezone, timezoneId, nonUTCEnabled)
 }
 
 case class GpuFromUTCTimestamp(
-    timestamp: Expression, timezone: Expression, zoneId: ZoneId, isOnCPU: Boolean)
+    timestamp: Expression, timezone: Expression, zoneId: ZoneId, nonUTCEnabled: Boolean)
   extends GpuBinaryExpressionArgsAnyScalar
       with ImplicitCastInputTypes
       with NullIntolerant {
@@ -1088,12 +1087,12 @@ case class GpuFromUTCTimestamp(
 
   override def doColumnar(lhs: GpuColumnVector, rhs: GpuScalar): ColumnVector = {
     if (rhs.getBase.isValid) {
-      if (TimeZoneDB.isUTCTimezone(zoneId)) {
+      if (GpuOverrides.isUTCTimezone(zoneId)) {
         // For UTC timezone, just a no-op bypassing GPU computation.
         lhs.getBase.incRefCount()
       } else {
-        if (isOnCPU){
-          TimeZoneDB.fromUtcTimestampToTimestamp(lhs.getBase, zoneId)
+        if (nonUTCEnabled){
+          GpuTimeZoneDB.fromUtcTimestampToTimestamp(lhs.getBase, zoneId)
         } else {
           // TODO: remove this until GPU backend supported.
           throw new UnsupportedOperationException(