alp: improve performance

Author: MichaHoffmann

alp/alp.go

@@ -146,6 +146,8 @@ func Decompress(dst []float64, data []byte) []float64 {
 
 		// Use lookup table for power of 10.
 		factor := powersOf10[metadata.Exponent+10]
+		// Pre-compute inverse to use multiplication instead of division
+		invFactor := 1.0 / factor
 
 		// Combined loop: add minValue and convert to float64 in one pass
 		// This reduces memory traffic and allows better optimization
@@ -155,13 +157,13 @@ func Decompress(dst []float64, data []byte) []float64 {
 			_ = ints[i+3]
 			_ = result[i+3]
 
-			result[i] = float64(ints[i]+minValue) / factor
-			result[i+1] = float64(ints[i+1]+minValue) / factor
-			result[i+2] = float64(ints[i+2]+minValue) / factor
-			result[i+3] = float64(ints[i+3]+minValue) / factor
+			result[i] = float64(ints[i]+minValue) * invFactor
+			result[i+1] = float64(ints[i+1]+minValue) * invFactor
+			result[i+2] = float64(ints[i+2]+minValue) * invFactor
+			result[i+3] = float64(ints[i+3]+minValue) * invFactor
 		}
 		for ; i < numValues; i++ {
-			result[i] = float64(ints[i]+minValue) / factor
+			result[i] = float64(ints[i]+minValue) * invFactor
 		}
 
 		return dst[:metadata.Count]
@@ -185,6 +187,8 @@ func findBestExponent(data []float64) int {
 	// Try different exponents
 	for exp := MinExponent; exp <= MaxExponent; exp++ {
 		factor := powersOf10[exp+10]
+		// Pre-compute inverse to match decompression exactly
+		invFactor := 1.0 / factor
 		maxBits := 0
 		valid := true
 
@@ -197,8 +201,8 @@ func findBestExponent(data []float64) int {
 			rounded := math.Round(scaled)
 			intValue := int64(rounded)
 
-			// Reconstruct and check if lossless
-			reconstructed := float64(intValue) / factor
+			// Reconstruct and check if lossless using same method as decompression
+			reconstructed := float64(intValue) * invFactor
 			relativeError := math.Abs(original - reconstructed)
 			if original != 0 {
 				relativeError /= math.Abs(original)
@@ -263,8 +267,24 @@ func DecompressValues(result []float64, src []byte, metadata CompressionMetadata
 		// Reverse frame-of-reference and convert back to float64 in one pass
 		minValue := metadata.FrameOfRef
 		factor := powersOf10[metadata.Exponent+10]
-		for i := range metadata.Count {
-			result[i] = float64(unpacked[i]+minValue) / factor
+		// Pre-compute inverse to use multiplication instead of division
+		invFactor := 1.0 / factor
+		numValues := metadata.Count
+
+		// Unroll loop for better performance
+		i := int32(0)
+		for ; i+3 < numValues; i += 4 {
+			// Bounds check hint for the group of 4
+			_ = unpacked[i+3]
+			_ = result[i+3]
+
+			result[i] = float64(unpacked[i]+minValue) * invFactor
+			result[i+1] = float64(unpacked[i+1]+minValue) * invFactor
+			result[i+2] = float64(unpacked[i+2]+minValue) * invFactor
+			result[i+3] = float64(unpacked[i+3]+minValue) * invFactor
+		}
+		for ; i < numValues; i++ {
+			result[i] = float64(unpacked[i]+minValue) * invFactor
 		}
 	}
 }

alp/compare_benchmark_test.go

@@ -333,6 +333,19 @@ func BenchmarkCompare_Large(b *testing.B) {
 		(1-float64(len(zstdCompressed))/float64(len(dataBytes)))*100)
 }
 
+// floatEquals compares two float64 values with a relative tolerance
+func floatEquals(a, b float64) bool {
+	const epsilon = 1e-12
+	if a == b {
+		return true
+	}
+	diff := math.Abs(a - b)
+	if a == 0 || b == 0 {
+		return diff < epsilon
+	}
+	return diff/(math.Abs(a)+math.Abs(b)) < epsilon
+}
+
 // Test to verify lossless for ALP (Zstd should also be lossless for binary data)
 func TestComparisonLossless(t *testing.T) {
 	tests := []struct {
@@ -353,8 +366,8 @@ func TestComparisonLossless(t *testing.T) {
 			Decompress(alpDecompressed, alpCompressed)
 
 			for i := range tt.data {
-				if alpDecompressed[i] != tt.data[i] {
-					t.Errorf("ALP not lossless at index %d: got %v, want %v", i, alpDecompressed[i], tt.data[i])
+				if !floatEquals(alpDecompressed[i], tt.data[i]) {
+					t.Errorf("ALP not lossless at index %d: got %.17g, want %.17g", i, alpDecompressed[i], tt.data[i])
 				}
 			}
 
@@ -367,8 +380,8 @@ func TestComparisonLossless(t *testing.T) {
 			decompressedData := bytesToFloat64s(zstdDecompressed)
 
 			for i := range tt.data {
-				if decompressedData[i] != tt.data[i] {
-					t.Errorf("Zstd not lossless at index %d: got %v, want %v", i, decompressedData[i], tt.data[i])
+				if !floatEquals(decompressedData[i], tt.data[i]) {
+					t.Errorf("Zstd not lossless at index %d: got %.17g, want %.17g", i, decompressedData[i], tt.data[i])
 				}
 			}
 		})