Histograms: optimize floatBucketIterator for common case

Signed-off-by: Linas Medziunas <linas.medziunas@gmail.com>
This commit is contained in:
Linas Medziunas 2023-11-29 08:54:05 +02:00
parent 97f9ad9d31
commit 048886ae8a
1 changed files with 69 additions and 51 deletions

View File

@ -767,6 +767,7 @@ func (h *FloatHistogram) floatBucketIterator(
}, },
targetSchema: targetSchema, targetSchema: targetSchema,
absoluteStartValue: absoluteStartValue, absoluteStartValue: absoluteStartValue,
boundReachedStartValue: absoluteStartValue == 0,
} }
if positive { if positive {
i.spans = h.PositiveSpans i.spans = h.PositiveSpans
@ -824,6 +825,32 @@ func (i *floatBucketIterator) Next() bool {
return false return false
} }
if i.schema == i.targetSchema {
// Fast path for the common case.
span := i.spans[i.spansIdx]
if i.bucketsIdx == 0 {
// Seed origIdx for the first bucket.
i.currIdx = span.Offset
} else {
i.currIdx++
}
for i.idxInSpan >= span.Length {
// We have exhausted the current span and have to find a new
// one. We even handle pathologic spans of length 0 here.
i.idxInSpan = 0
i.spansIdx++
if i.spansIdx >= len(i.spans) {
return false
}
span = i.spans[i.spansIdx]
i.currIdx += span.Offset
}
i.currCount = i.buckets[i.bucketsIdx]
i.idxInSpan++
i.bucketsIdx++
} else {
// Copy all of these into local variables so that we can forward to the // Copy all of these into local variables so that we can forward to the
// next bucket and then roll back if needed. // next bucket and then roll back if needed.
origIdx, spansIdx, idxInSpan := i.origIdx, i.spansIdx, i.idxInSpan origIdx, spansIdx, idxInSpan := i.origIdx, i.spansIdx, i.idxInSpan
@ -831,7 +858,7 @@ func (i *floatBucketIterator) Next() bool {
firstPass := true firstPass := true
i.currCount = 0 i.currCount = 0
mergeLoop: // Merge together all buckets from the original schema that fall into one bucket in the targetSchema. mergeLoop: // Merge together all buckets from the original schema that fall into one bucket in the targetSchema.
for { for {
if i.bucketsIdx == 0 { if i.bucketsIdx == 0 {
// Seed origIdx for the first bucket. // Seed origIdx for the first bucket.
@ -853,7 +880,7 @@ mergeLoop: // Merge together all buckets from the original schema that fall into
span = i.spans[spansIdx] span = i.spans[spansIdx]
origIdx += span.Offset origIdx += span.Offset
} }
currIdx := i.targetIdx(origIdx) currIdx := targetIdx(origIdx, i.schema, i.targetSchema)
switch { switch {
case firstPass: case firstPass:
i.currIdx = currIdx i.currIdx = currIdx
@ -873,6 +900,8 @@ mergeLoop: // Merge together all buckets from the original schema that fall into
break mergeLoop break mergeLoop
} }
} }
}
// Skip buckets before absoluteStartValue. // Skip buckets before absoluteStartValue.
// TODO(beorn7): Maybe do something more efficient than this recursive call. // TODO(beorn7): Maybe do something more efficient than this recursive call.
if !i.boundReachedStartValue && getBound(i.currIdx, i.targetSchema) <= i.absoluteStartValue { if !i.boundReachedStartValue && getBound(i.currIdx, i.targetSchema) <= i.absoluteStartValue {
@ -882,17 +911,6 @@ mergeLoop: // Merge together all buckets from the original schema that fall into
return true return true
} }
// targetIdx returns the bucket index within i.targetSchema for the given bucket
// index within i.schema.
func (i *floatBucketIterator) targetIdx(idx int32) int32 {
if i.schema == i.targetSchema {
// Fast path for the common case. The below would yield the same
// result, just with more effort.
return idx
}
return ((idx - 1) >> (i.schema - i.targetSchema)) + 1
}
type reverseFloatBucketIterator struct { type reverseFloatBucketIterator struct {
baseBucketIterator[float64, float64] baseBucketIterator[float64, float64]
idxInSpan int32 // Changed from uint32 to allow negative values for exhaustion detection. idxInSpan int32 // Changed from uint32 to allow negative values for exhaustion detection.