folly: simplify the stats avgHelper() function

Summary:
When the input type is a long double, perform division using long
double.  In all other cases, divide using double precision.

Also fix the EXPECT_EQ() usage in the test case.

Test Plan: fbconfig -r common/stats folly && fbmake runtests

Reviewed By: andrei.alexandrescu@fb.com

FB internal diff: D555433

folly/detail/Stats.h

@@ -18,53 +18,35 @@
 #define FOLLY_DETAIL_STATS_H_
 
 #include <cstdint>
+#include <type_traits>
 
 namespace folly { namespace detail {
 
 /*
- * Helper functions for how to perform division based on the desired
+ * Helper function to compute the average, given a specified input type and
  * return type.
  */
 
-// For floating point input types, do floating point division
-template <typename ReturnType, typename ValueType>
-typename std::enable_if<std::is_floating_point<ValueType>::value,
-                        ReturnType>::type
-avgHelper(ValueType sum, uint64_t count) {
-  if (count == 0) { return ReturnType(0); }
-  return static_cast<ReturnType>(sum / count);
-}
-
-// For floating point return types, do floating point division
-template <typename ReturnType, typename ValueType>
-typename std::enable_if<std::is_floating_point<ReturnType>::value &&
-                        !std::is_floating_point<ValueType>::value,
-                        ReturnType>::type
-avgHelper(ValueType sum, uint64_t count) {
-  if (count == 0) { return ReturnType(0); }
-  return static_cast<ReturnType>(sum) / count;
-}
-
-// For signed integer input types, do signed division
-template <typename ReturnType, typename ValueType>
-typename std::enable_if<!std::is_floating_point<ReturnType>::value &&
-                        !std::is_floating_point<ValueType>::value &&
-                        std::is_signed<ValueType>::value,
-                        ReturnType>::type
-avgHelper(ValueType sum, uint64_t count) {
+// If the input is long double, divide using long double to avoid losing
+// precision.
+template <typename ReturnType>
+ReturnType avgHelper(long double sum, uint64_t count) {
   if (count == 0) { return ReturnType(0); }
-  return sum / static_cast<int64_t>(count);
+  const long double countf = count;
+  return static_cast<ReturnType>(sum / countf);
 }
 
-// For unsigned integer input types, do unsigned division
+// In all other cases divide using double precision.
+// This should be relatively fast, and accurate enough for most use cases.
 template <typename ReturnType, typename ValueType>
-typename std::enable_if<!std::is_floating_point<ReturnType>::value &&
-                        !std::is_floating_point<ValueType>::value &&
-                        std::is_unsigned<ValueType>::value,
+typename std::enable_if<!std::is_same<typename std::remove_cv<ValueType>::type,
+                                      long double>::value,
                         ReturnType>::type
 avgHelper(ValueType sum, uint64_t count) {
   if (count == 0) { return ReturnType(0); }
-  return sum / count;
+  const double sumf = sum;
+  const double countf = count;
+  return static_cast<ReturnType>(sumf / countf);
 }
 
 

folly/test/TimeseriesTest.cpp

@@ -322,9 +322,8 @@ TEST(BucketedTimeSeries, rate) {
 }
 
 TEST(BucketedTimeSeries, avgTypeConversion) {
-  // The average code has many different code paths to decide what type of
-  // division to perform (floating point, signed integer, unsigned integer).
-  // Test the various code paths.
+  // Make sure the computed average values are accurate regardless
+  // of the input type and return type.
 
   {
     // Simple sanity tests for small positive integer values
@@ -333,14 +332,14 @@ TEST(BucketedTimeSeries, avgTypeConversion) {
     ts.addValue(seconds(0), 10, 200);
     ts.addValue(seconds(0), 16, 100);
 
-    EXPECT_DOUBLE_EQ(ts.avg(), 10.0);
-    EXPECT_DOUBLE_EQ(ts.avg<float>(), 10.0);
-    EXPECT_EQ(ts.avg<uint64_t>(), 10);
-    EXPECT_EQ(ts.avg<int64_t>(), 10);
-    EXPECT_EQ(ts.avg<int32_t>(), 10);
-    EXPECT_EQ(ts.avg<int16_t>(), 10);
-    EXPECT_EQ(ts.avg<int8_t>(), 10);
-    EXPECT_EQ(ts.avg<uint8_t>(), 10);
+    EXPECT_DOUBLE_EQ(10.0, ts.avg());
+    EXPECT_DOUBLE_EQ(10.0, ts.avg<float>());
+    EXPECT_EQ(10, ts.avg<uint64_t>());
+    EXPECT_EQ(10, ts.avg<int64_t>());
+    EXPECT_EQ(10, ts.avg<int32_t>());
+    EXPECT_EQ(10, ts.avg<int16_t>());
+    EXPECT_EQ(10, ts.avg<int8_t>());
+    EXPECT_EQ(10, ts.avg<uint8_t>());
   }
 
   {
@@ -351,11 +350,11 @@ TEST(BucketedTimeSeries, avgTypeConversion) {
     ts.addValue(seconds(0), -300);
     ts.addValue(seconds(0), -200, 65535);
 
-    EXPECT_DOUBLE_EQ(ts.avg(), -200.0);
-    EXPECT_DOUBLE_EQ(ts.avg<float>(), -200.0);
-    EXPECT_EQ(ts.avg<int64_t>(), -200);
-    EXPECT_EQ(ts.avg<int32_t>(), -200);
-    EXPECT_EQ(ts.avg<int16_t>(), -200);
+    EXPECT_DOUBLE_EQ(-200.0, ts.avg());
+    EXPECT_DOUBLE_EQ(-200.0, ts.avg<float>());
+    EXPECT_EQ(-200, ts.avg<int64_t>());
+    EXPECT_EQ(-200, ts.avg<int32_t>());
+    EXPECT_EQ(-200, ts.avg<int16_t>());
   }
 
   {
@@ -365,11 +364,11 @@ TEST(BucketedTimeSeries, avgTypeConversion) {
                           std::numeric_limits<uint64_t>::max(),
                           std::numeric_limits<uint64_t>::max());
 
-    EXPECT_DOUBLE_EQ(ts.avg(), 1.0);
-    EXPECT_DOUBLE_EQ(ts.avg<float>(), 1.0);
-    EXPECT_EQ(ts.avg<uint64_t>(), 1);
-    EXPECT_EQ(ts.avg<int64_t>(), 1);
-    EXPECT_EQ(ts.avg<int8_t>(), 1);
+    EXPECT_DOUBLE_EQ(1.0, ts.avg());
+    EXPECT_DOUBLE_EQ(1.0, ts.avg<float>());
+    EXPECT_EQ(1, ts.avg<uint64_t>());
+    EXPECT_EQ(1, ts.avg<int64_t>());
+    EXPECT_EQ(1, ts.avg<int8_t>());
   }
 
   {
@@ -379,14 +378,14 @@ TEST(BucketedTimeSeries, avgTypeConversion) {
     ts.addValue(seconds(0), 10.0, 200);
     ts.addValue(seconds(0), 16.0, 100);
 
-    EXPECT_DOUBLE_EQ(ts.avg(), 10.0);
-    EXPECT_DOUBLE_EQ(ts.avg<float>(), 10.0);
-    EXPECT_EQ(ts.avg<uint64_t>(), 10);
-    EXPECT_EQ(ts.avg<int64_t>(), 10);
-    EXPECT_EQ(ts.avg<int32_t>(), 10);
-    EXPECT_EQ(ts.avg<int16_t>(), 10);
-    EXPECT_EQ(ts.avg<int8_t>(), 10);
-    EXPECT_EQ(ts.avg<uint8_t>(), 10);
+    EXPECT_DOUBLE_EQ(10.0, ts.avg());
+    EXPECT_DOUBLE_EQ(10.0, ts.avg<float>());
+    EXPECT_EQ(10, ts.avg<uint64_t>());
+    EXPECT_EQ(10, ts.avg<int64_t>());
+    EXPECT_EQ(10, ts.avg<int32_t>());
+    EXPECT_EQ(10, ts.avg<int16_t>());
+    EXPECT_EQ(10, ts.avg<int8_t>());
+    EXPECT_EQ(10, ts.avg<uint8_t>());
   }
 
   {
@@ -409,10 +408,12 @@ TEST(BucketedTimeSeries, avgTypeConversion) {
       ts.addValue(seconds(0), value);
     }
 
-    EXPECT_DOUBLE_EQ(ts.avg(), value);
-    EXPECT_DOUBLE_EQ(ts.avg<float>(), value);
-    EXPECT_DOUBLE_EQ(ts.avg<uint64_t>(), value);
-    EXPECT_DOUBLE_EQ(ts.avg<int64_t>(), value);
+    EXPECT_DOUBLE_EQ(value, ts.avg());
+    EXPECT_DOUBLE_EQ(value, ts.avg<float>());
+    // Some precision is lost here due to the huge sum, so the
+    // integer average returned is off by one.
+    EXPECT_NEAR(value, ts.avg<uint64_t>(), 1);
+    EXPECT_NEAR(value, ts.avg<int64_t>(), 1);
   }
 
   {
@@ -422,12 +423,39 @@ TEST(BucketedTimeSeries, avgTypeConversion) {
       ts.addValue(seconds(0), i);
     }
 
-    EXPECT_DOUBLE_EQ(ts.avg(), 50.0);
-    EXPECT_DOUBLE_EQ(ts.avg<float>(), 50.0);
-    EXPECT_DOUBLE_EQ(ts.avg<uint64_t>(), 50);
-    EXPECT_DOUBLE_EQ(ts.avg<int64_t>(), 50);
-    EXPECT_DOUBLE_EQ(ts.avg<int16_t>(), 50);
-    EXPECT_DOUBLE_EQ(ts.avg<int8_t>(), 50);
+    EXPECT_DOUBLE_EQ(50.0, ts.avg());
+    EXPECT_DOUBLE_EQ(50.0, ts.avg<float>());
+    EXPECT_EQ(50, ts.avg<uint64_t>());
+    EXPECT_EQ(50, ts.avg<int64_t>());
+    EXPECT_EQ(50, ts.avg<int16_t>());
+    EXPECT_EQ(50, ts.avg<int8_t>());
+  }
+
+  {
+    // Test BucketedTimeSeries with long double input
+    BucketedTimeSeries<long double> ts(60, seconds(600));
+    ts.addValueAggregated(seconds(0), 1000.0L, 7);
+
+    long double expected = 1000.0L / 7.0L;
+    EXPECT_DOUBLE_EQ(static_cast<double>(expected), ts.avg());
+    EXPECT_DOUBLE_EQ(static_cast<float>(expected), ts.avg<float>());
+    EXPECT_DOUBLE_EQ(expected, ts.avg<long double>());
+    EXPECT_EQ(static_cast<uint64_t>(expected), ts.avg<uint64_t>());
+    EXPECT_EQ(static_cast<int64_t>(expected), ts.avg<int64_t>());
+  }
+
+  {
+    // Test BucketedTimeSeries with int64_t values,
+    // but using an average that requires a fair amount of precision.
+    BucketedTimeSeries<int64_t> ts(60, seconds(600));
+    ts.addValueAggregated(seconds(0), 1000, 7);
+
+    long double expected = 1000.0L / 7.0L;
+    EXPECT_DOUBLE_EQ(static_cast<double>(expected), ts.avg());
+    EXPECT_DOUBLE_EQ(static_cast<float>(expected), ts.avg<float>());
+    EXPECT_DOUBLE_EQ(expected, ts.avg<long double>());
+    EXPECT_EQ(static_cast<uint64_t>(expected), ts.avg<uint64_t>());
+    EXPECT_EQ(static_cast<int64_t>(expected), ts.avg<int64_t>());
   }
 }