[SDK] Add base2 exponential histogram indexer

sdk/include/opentelemetry/sdk/metrics/aggregation/base2_exponential_histogram_indexer.h

@@ -0,0 +1,46 @@
+// Copyright The OpenTelemetry Authors
+// SPDX-License-Identifier: Apache-2.0
+
+#pragma once
+
+#include "opentelemetry/version.h"
+
+#include <cstdint>
+
+OPENTELEMETRY_BEGIN_NAMESPACE
+namespace sdk
+{
+namespace metrics
+{
+
+/*
+ * An indexer for base2 exponential histograms. It is used to calculate index for a given value and
+ * scale.
+ */
+class Base2ExponentialHistogramIndexer
+{
+public:
+  /*
+   * Construct a new indexer for a given scale.
+   */
+  explicit Base2ExponentialHistogramIndexer(int32_t scale = 0);
+  Base2ExponentialHistogramIndexer(const Base2ExponentialHistogramIndexer &other) = default;
+  Base2ExponentialHistogramIndexer &operator=(const Base2ExponentialHistogramIndexer &other) =
+      default;
+
+  /**
+   * Compute the index for the given value.
+   *
+   * @param value Measured value (must be non-zero).
+   * @return the index of the bucket which the value maps to.
+   */
+  int32_t ComputeIndex(double value) const;
+
+private:
+  int32_t scale_;
+  double scale_factor_;
+};
+
+}  // namespace metrics
+}  // namespace sdk
+OPENTELEMETRY_END_NAMESPACE

sdk/src/metrics/CMakeLists.txt

@@ -14,6 +14,7 @@ add_library(
   state/observable_registry.cc
   state/sync_metric_storage.cc
   state/temporal_metric_storage.cc
+  aggregation/base2_exponential_histogram_indexer.cc
   aggregation/histogram_aggregation.cc
   aggregation/lastvalue_aggregation.cc
   aggregation/sum_aggregation.cc

sdk/src/metrics/aggregation/base2_exponential_histogram_indexer.cc

@@ -0,0 +1,71 @@
+// Copyright The OpenTelemetry Authors
+// SPDX-License-Identifier: Apache-2.0
+
+#include "opentelemetry/sdk/metrics/aggregation/base2_exponential_histogram_indexer.h"
+
+#include <cmath>
+
+OPENTELEMETRY_BEGIN_NAMESPACE
+namespace sdk
+{
+namespace metrics
+{
+
+namespace
+{
+
+const double kLogBase2E = 1.0 / std::log(2.0);
+
+double ComputeScaleFactor(int32_t scale)
+{
+  return std::scalbn(kLogBase2E, scale);
+}
+
+// Compute the bucket index using a logarithm based approach.
+int32_t GetIndexByLogarithm(double value, double scale_factor)
+{
+  return static_cast<int32_t>(std::ceil(std::log(value) * scale_factor)) - 1;
+}
+
+// Compute the bucket index at scale 0.
+int32_t MapToIndexScaleZero(double value)
+{
+  // Note: std::frexp() rounds submnormal values to the smallest normal value and returns an
+  // exponent corresponding to fractions in the range [0.5, 1), whereas an exponent for the range
+  // [1, 2), so subtract 1 from the exponent immediately.
+  int exp;
+  double frac = std::frexp(value, &exp);
+  exp--;
+
+  if (frac == 0.5)
+  {
+    // Special case for powers of two: they fall into the bucket numbered one less.
+    exp--;
+  }
+  return exp;
+}
+
+}  // namespace
+
+Base2ExponentialHistogramIndexer::Base2ExponentialHistogramIndexer(int32_t scale)
+    : scale_(scale), scale_factor_(scale > 0 ? ComputeScaleFactor(scale) : 0)
+{}
+
+int32_t Base2ExponentialHistogramIndexer::ComputeIndex(double value) const
+{
+  const double abs_value = std::fabs(value);
+  // For positive scales, compute the index by logarithm, which is simpler but may be
+  // inaccurate near bucket boundaries
+  if (scale_ > 0)
+  {
+    return GetIndexByLogarithm(abs_value, scale_factor_);
+  }
+  // For scale zero, compute the exact index by extracting the exponent.
+  // For negative scales, compute the exact index by extracting the exponent and shifting it to
+  // the right by -scale
+  return MapToIndexScaleZero(abs_value) >> -scale_;
+}
+
+}  // namespace metrics
+}  // namespace sdk
+OPENTELEMETRY_END_NAMESPACE

sdk/test/metrics/BUILD

@@ -285,6 +285,21 @@ cc_test(
     ],
 )
 
+cc_test(
+    name = "base2_exponential_histogram_indexer_test",
+    srcs = [
+        "base2_exponential_histogram_indexer_test.cc",
+    ],
+    tags = [
+        "metrics",
+        "test",
+    ],
+    deps = [
+        "metrics_common_test_utils",
+        "@com_google_googletest//:gtest_main",
+    ],
+)
+
 cc_test(
     name = "histogram_aggregation_test",
     srcs = [

sdk/test/metrics/CMakeLists.txt

@@ -14,6 +14,7 @@ foreach(
   histogram_aggregation_test
   attributes_processor_test
   attributes_hashmap_test
+  base2_exponential_histogram_indexer_test
   circular_buffer_counter_test
   histogram_test
   sync_metric_storage_counter_test

sdk/test/metrics/base2_exponential_histogram_indexer_test.cc

@@ -0,0 +1,178 @@
+// Copyright The OpenTelemetry Authors
+// SPDX-License-Identifier: Apache-2.0
+
+#include "opentelemetry/sdk/metrics/aggregation/base2_exponential_histogram_indexer.h"
+
+#include <gtest/gtest.h>
+
+using namespace opentelemetry::sdk::metrics;
+
+TEST(Base2ExponentialHistogramIndexerTest, ScaleOne)
+{
+  const Base2ExponentialHistogramIndexer indexer{1};
+  auto compute_index = [indexer](double value) { return indexer.ComputeIndex(value); };
+
+  EXPECT_EQ(compute_index(std::numeric_limits<double>::max()), 2047);
+  EXPECT_EQ(compute_index(strtod("0x1p1023", nullptr)), 2045);
+  EXPECT_EQ(compute_index(strtod("0x1.1p1023", nullptr)), 2046);
+  EXPECT_EQ(compute_index(strtod("0x1p1022", nullptr)), 2043);
+  EXPECT_EQ(compute_index(strtod("0x1.1p1022", nullptr)), 2044);
+  EXPECT_EQ(compute_index(strtod("0x1p-1022", nullptr)), -2045);
+  EXPECT_EQ(compute_index(strtod("0x1.1p-1022", nullptr)), -2044);
+  EXPECT_EQ(compute_index(strtod("0x1p-1021", nullptr)), -2043);
+  EXPECT_EQ(compute_index(strtod("0x1.1p-1021", nullptr)), -2042);
+  EXPECT_EQ(compute_index(std::numeric_limits<double>::min()), -2045);
+  EXPECT_EQ(compute_index(std::numeric_limits<double>::denorm_min()), -2149);
+  EXPECT_EQ(compute_index(15.0), 7);
+  EXPECT_EQ(compute_index(9.0), 6);
+  EXPECT_EQ(compute_index(7.0), 5);
+  EXPECT_EQ(compute_index(5.0), 4);
+  EXPECT_EQ(compute_index(3.0), 3);
+  EXPECT_EQ(compute_index(2.5), 2);
+  EXPECT_EQ(compute_index(1.5), 1);
+  EXPECT_EQ(compute_index(1.2), 0);
+  EXPECT_EQ(compute_index(1.0), -1);
+  EXPECT_EQ(compute_index(0.75), -1);
+  EXPECT_EQ(compute_index(0.55), -2);
+  EXPECT_EQ(compute_index(0.45), -3);
+}
+
+TEST(Base2ExponentialHistogramIndexerTest, ScaleZero)
+{
+  const Base2ExponentialHistogramIndexer indexer{0};
+  auto compute_index = [indexer](double value) { return indexer.ComputeIndex(value); };
+
+  // Near +Inf.
+  // Use constant for exp, as max_exponent constant includes bias.
+  EXPECT_EQ(compute_index(std::numeric_limits<double>::max()), 1023);
+  EXPECT_EQ(compute_index(strtod("0x1p1023", nullptr)), 1022);
+  EXPECT_EQ(compute_index(strtod("0x1.1p1023", nullptr)), 1023);
+  EXPECT_EQ(compute_index(strtod("0x1p1022", nullptr)), 1021);
+  EXPECT_EQ(compute_index(strtod("0x1.1p1022", nullptr)), 1022);
+  // Near 0.
+  EXPECT_EQ(compute_index(strtod("0x1p-1022", nullptr)), -1023);
+  EXPECT_EQ(compute_index(strtod("0x1.1p-1022", nullptr)), -1022);
+  EXPECT_EQ(compute_index(strtod("0x1p-1021", nullptr)), -1022);
+  EXPECT_EQ(compute_index(strtod("0x1.1p-1021", nullptr)), -1021);
+  EXPECT_EQ(compute_index(std::numeric_limits<double>::min()), -1023);
+  EXPECT_EQ(compute_index(std::numeric_limits<double>::denorm_min()), -1075);
+  // Near 1.
+  EXPECT_EQ(compute_index(4.0), 1);
+  EXPECT_EQ(compute_index(3.0), 1);
+  EXPECT_EQ(compute_index(2.0), 0);
+  EXPECT_EQ(compute_index(1.5), 0);
+  EXPECT_EQ(compute_index(1.0), -1);
+  EXPECT_EQ(compute_index(0.75), -1);
+  EXPECT_EQ(compute_index(0.51), -1);
+  EXPECT_EQ(compute_index(0.5), -2);
+  EXPECT_EQ(compute_index(0.26), -2);
+  EXPECT_EQ(compute_index(0.25), -3);
+  EXPECT_EQ(compute_index(0.126), -3);
+  EXPECT_EQ(compute_index(0.125), -4);
+}
+
+TEST(Base2ExponentialHistogramIndexerTest, ScaleNegativeOne)
+{
+  const Base2ExponentialHistogramIndexer indexer{-1};
+  auto compute_index = [indexer](double value) { return indexer.ComputeIndex(value); };
+
+  EXPECT_EQ(compute_index(17.0), 2);
+  EXPECT_EQ(compute_index(16.0), 1);
+  EXPECT_EQ(compute_index(15.0), 1);
+  EXPECT_EQ(compute_index(9.0), 1);
+  EXPECT_EQ(compute_index(8.0), 1);
+  EXPECT_EQ(compute_index(5.0), 1);
+  EXPECT_EQ(compute_index(4.0), 0);
+  EXPECT_EQ(compute_index(3.0), 0);
+  EXPECT_EQ(compute_index(2.0), 0);
+  EXPECT_EQ(compute_index(1.5), 0);
+  EXPECT_EQ(compute_index(1.0), -1);
+  EXPECT_EQ(compute_index(0.75), -1);
+  EXPECT_EQ(compute_index(0.5), -1);
+  EXPECT_EQ(compute_index(0.25), -2);
+  EXPECT_EQ(compute_index(0.20), -2);
+  EXPECT_EQ(compute_index(0.13), -2);
+  EXPECT_EQ(compute_index(0.125), -2);
+  EXPECT_EQ(compute_index(0.10), -2);
+  EXPECT_EQ(compute_index(0.0625), -3);
+  EXPECT_EQ(compute_index(0.06), -3);
+}
+
+TEST(Base2ExponentialHistogramIndexerTest, ScaleNegativeFour)
+{
+  const Base2ExponentialHistogramIndexer indexer{-4};
+  auto compute_index = [indexer](double value) { return indexer.ComputeIndex(value); };
+
+  EXPECT_EQ(compute_index(strtod("0x1p0", nullptr)), -1);
+  EXPECT_EQ(compute_index(strtod("0x10p0", nullptr)), 0);
+  EXPECT_EQ(compute_index(strtod("0x100p0", nullptr)), 0);
+  EXPECT_EQ(compute_index(strtod("0x1000p0", nullptr)), 0);
+  EXPECT_EQ(compute_index(strtod("0x10000p0", nullptr)), 0);  // Base == 2**16
+  EXPECT_EQ(compute_index(strtod("0x100000p0", nullptr)), 1);
+  EXPECT_EQ(compute_index(strtod("0x1000000p0", nullptr)), 1);
+  EXPECT_EQ(compute_index(strtod("0x10000000p0", nullptr)), 1);
+  EXPECT_EQ(compute_index(strtod("0x100000000p0", nullptr)), 1);  // == 2**32
+  EXPECT_EQ(compute_index(strtod("0x1000000000p0", nullptr)), 2);
+  EXPECT_EQ(compute_index(strtod("0x10000000000p0", nullptr)), 2);
+  EXPECT_EQ(compute_index(strtod("0x100000000000p0", nullptr)), 2);
+  EXPECT_EQ(compute_index(strtod("0x1000000000000p0", nullptr)), 2);  // 2**48
+  EXPECT_EQ(compute_index(strtod("0x10000000000000p0", nullptr)), 3);
+  EXPECT_EQ(compute_index(strtod("0x100000000000000p0", nullptr)), 3);
+  EXPECT_EQ(compute_index(strtod("0x1000000000000000p0", nullptr)), 3);
+  EXPECT_EQ(compute_index(strtod("0x10000000000000000p0", nullptr)), 3);  // 2**64
+  EXPECT_EQ(compute_index(strtod("0x100000000000000000p0", nullptr)), 4);
+  EXPECT_EQ(compute_index(strtod("0x1000000000000000000p0", nullptr)), 4);
+  EXPECT_EQ(compute_index(strtod("0x10000000000000000000p0", nullptr)), 4);
+  EXPECT_EQ(compute_index(strtod("0x100000000000000000000p0", nullptr)), 4);
+  EXPECT_EQ(compute_index(strtod("0x1000000000000000000000p0", nullptr)), 5);
+  EXPECT_EQ(compute_index(1 / strtod("0x1p0", nullptr)), -1);
+  EXPECT_EQ(compute_index(1 / strtod("0x10p0", nullptr)), -1);
+  EXPECT_EQ(compute_index(1 / strtod("0x100p0", nullptr)), -1);
+  EXPECT_EQ(compute_index(1 / strtod("0x1000p0", nullptr)), -1);
+  EXPECT_EQ(compute_index(1 / strtod("0x10000p0", nullptr)), -2);  // 2**-16
+  EXPECT_EQ(compute_index(1 / strtod("0x100000p0", nullptr)), -2);
+  EXPECT_EQ(compute_index(1 / strtod("0x1000000p0", nullptr)), -2);
+  EXPECT_EQ(compute_index(1 / strtod("0x10000000p0", nullptr)), -2);
+  EXPECT_EQ(compute_index(1 / strtod("0x100000000p0", nullptr)), -3);  // 2**-32
+  EXPECT_EQ(compute_index(1 / strtod("0x1000000000p0", nullptr)), -3);
+  EXPECT_EQ(compute_index(1 / strtod("0x10000000000p0", nullptr)), -3);
+  EXPECT_EQ(compute_index(1 / strtod("0x100000000000p0", nullptr)), -3);
+  EXPECT_EQ(compute_index(1 / strtod("0x1000000000000p0", nullptr)), -4);  // 2**-48
+  EXPECT_EQ(compute_index(1 / strtod("0x10000000000000p0", nullptr)), -4);
+  EXPECT_EQ(compute_index(1 / strtod("0x100000000000000p0", nullptr)), -4);
+  EXPECT_EQ(compute_index(1 / strtod("0x1000000000000000p0", nullptr)), -4);
+  EXPECT_EQ(compute_index(1 / strtod("0x10000000000000000p0", nullptr)), -5);  // 2**-64
+  EXPECT_EQ(compute_index(1 / strtod("0x100000000000000000p0", nullptr)), -5);
+  // Max values.
+  EXPECT_EQ(compute_index(0x1.FFFFFFFFFFFFFp1023), 63);
+  EXPECT_EQ(compute_index(strtod("0x1p1023", nullptr)), 63);
+  EXPECT_EQ(compute_index(strtod("0x1p1019", nullptr)), 63);
+  EXPECT_EQ(compute_index(strtod("0x1p1009", nullptr)), 63);
+  EXPECT_EQ(compute_index(strtod("0x1p1008", nullptr)), 62);
+  EXPECT_EQ(compute_index(strtod("0x1p1007", nullptr)), 62);
+  EXPECT_EQ(compute_index(strtod("0x1p1000", nullptr)), 62);
+  EXPECT_EQ(compute_index(strtod("0x1p0993", nullptr)), 62);
+  EXPECT_EQ(compute_index(strtod("0x1p0992", nullptr)), 61);
+  EXPECT_EQ(compute_index(strtod("0x1p0991", nullptr)), 61);
+  // Min and subnormal values.
+  EXPECT_EQ(compute_index(strtod("0x1p-1074", nullptr)), -68);
+  EXPECT_EQ(compute_index(strtod("0x1p-1073", nullptr)), -68);
+  EXPECT_EQ(compute_index(strtod("0x1p-1072", nullptr)), -68);
+  EXPECT_EQ(compute_index(strtod("0x1p-1057", nullptr)), -67);
+  EXPECT_EQ(compute_index(strtod("0x1p-1056", nullptr)), -67);
+  EXPECT_EQ(compute_index(strtod("0x1p-1041", nullptr)), -66);
+  EXPECT_EQ(compute_index(strtod("0x1p-1040", nullptr)), -66);
+  EXPECT_EQ(compute_index(strtod("0x1p-1025", nullptr)), -65);
+  EXPECT_EQ(compute_index(strtod("0x1p-1024", nullptr)), -65);
+  EXPECT_EQ(compute_index(strtod("0x1p-1023", nullptr)), -64);
+  EXPECT_EQ(compute_index(strtod("0x1p-1022", nullptr)), -64);
+  EXPECT_EQ(compute_index(strtod("0x1p-1009", nullptr)), -64);
+  EXPECT_EQ(compute_index(strtod("0x1p-1008", nullptr)), -64);
+  EXPECT_EQ(compute_index(strtod("0x1p-1007", nullptr)), -63);
+  EXPECT_EQ(compute_index(strtod("0x1p-0993", nullptr)), -63);
+  EXPECT_EQ(compute_index(strtod("0x1p-0992", nullptr)), -63);
+  EXPECT_EQ(compute_index(strtod("0x1p-0991", nullptr)), -62);
+  EXPECT_EQ(compute_index(strtod("0x1p-0977", nullptr)), -62);
+  EXPECT_EQ(compute_index(strtod("0x1p-0976", nullptr)), -62);
+  EXPECT_EQ(compute_index(strtod("0x1p-0975", nullptr)), -61);
+}