Make kExpected file-specific, rewrite float tests to avoid epsilon.

SingleSource/UnitTests/Atomic/big_test.cpp

@@ -23,6 +23,14 @@
 
 #include "util.h"
 
+// V >> 56 = 66, so to prevent 32-bit overflow, kExpected must be less than
+// 2^31 / 66 = 32 x 10^6.
+#ifdef SMALL_PROBLEM_SIZE
+static constexpr int kIterations = 1'000'000;
+#else
+static constexpr int kIterations = 3'000'000;
+#endif
+static constexpr int kExpected = kThreads * kIterations;
 static constexpr int kBigSize = 10;
 struct big_t {
   int v[kBigSize];
@@ -44,7 +52,6 @@ void looper_big_cmpxchg(big_t *abig, int success_model, int fail_model) {
 
 void test_big_cmpxchg() {
   std::vector<std::thread> pool;
-
   for (int success_model : atomic_compare_exchange_models) {
     for (int fail_model : atomic_compare_exchange_models) {
       big_t abig = {};
@@ -66,9 +73,6 @@ void test_big() {
 }
 
 int main() {
-  std::cout << kThreads << " threads; "
-            << kIterations << " iterations each; total of "
-            << kExpected << "\n";
   test_big();
   std::cout << "PASSED\n";
 }

SingleSource/UnitTests/Atomic/big_test.reference_output

@@ -1,4 +1,3 @@
-10 threads; 1000000 iterations each; total of 10000000
 Testing big
 PASSED
 exit 0

SingleSource/UnitTests/Atomic/float_test.cpp

@@ -9,7 +9,7 @@
 // This file tests atomic operations on floating point types with aligned
 // memory addresses.
 //
-// The types tested are: float, double.
+// The types tested are: float, double, float128.
 // The ops tested are: xchg, cmpxchg.
 //
 // Please read the README before contributing.
@@ -22,47 +22,68 @@
 #include <thread>
 #include <vector>
 
-#include "numeric.h"
 #include "util.h"
 
-// See numeric.h for an explanation of numeric xchg tests.
+// There are 23-bits in the mantissa of a single-precision float.
+// Therefore, kExpected cannot exceed 2^24.
+static constexpr int kIterations = 1'500'000;
+static constexpr int kExpected = kThreads * kIterations;
+
+// See int_aligned_test.cc for an explanation of xchg tests.
+template <typename T>
+void looper_float_scalar_xchg(T *afloat, int model) {
+  __int128_t error = 0;
+  T next = *afloat + 1;
+  T result;
+  for (int n = 0; n < kIterations; ++n) {
+    __atomic_exchange(afloat, &next, &result, model);
+    error +=
+        static_cast<__int128_t>(next) - static_cast<__int128_t>(result + 1);
+    next = result + 1;
+  }
+  __atomic_fetch_sub(afloat, static_cast<T>(error), model);
+}
+
 template <typename T>
 void test_float_scalar_xchg() {
-  static constexpr T val = V >> right_shift<T>();
-  static constexpr T expected = val * kExpected;
   std::vector<std::thread> pool;
-
   for (int model : atomic_exchange_models) {
     T afloat = 0;
     for (int n = 0; n < kThreads; ++n)
-      pool.emplace_back(looper_numeric_xchg_atomic<T>, &afloat, model);
+      pool.emplace_back(looper_float_scalar_xchg<T>, &afloat, model);
     for (int n = 0; n < kThreads; ++n)
       pool[n].join();
     pool.clear();
-    if (afloat < expected * (1 - kEpsilon) ||
-        afloat > expected * (1 + kEpsilon))
+    if (afloat != kExpected)
       fail();
   }
 }
 
-// See numeric.h for an explanation of numeric cmpxchg tests.
+// See int_aligned_test.cc for an explanation of cmpxchg tests.
+template <typename T>
+void looper_float_scalar_cmpxchg(T *afloat, int success_model, int fail_model) {
+  for (int n = 0; n < kIterations; ++n) {
+    T desired, expected = 0;
+    do {
+      desired = expected + 1;
+    } while (!__atomic_compare_exchange(afloat, &expected, &desired, true,
+                                        success_model, fail_model));
+  }
+}
+
 template <typename T>
 void test_float_scalar_cmpxchg() {
-  static constexpr T val = V >> right_shift<T>();
-  static constexpr T expected = val * kExpected;
   std::vector<std::thread> pool;
-
   for (int success_model : atomic_compare_exchange_models) {
     for (int fail_model : atomic_compare_exchange_models) {
       T afloat = 0;
       for (int n = 0; n < kThreads; ++n)
-        pool.emplace_back(looper_numeric_cmpxchg<T>, &afloat, success_model,
-                          fail_model);
+        pool.emplace_back(looper_float_scalar_cmpxchg<T>, &afloat,
+                          success_model, fail_model);
       for (int n = 0; n < kThreads; ++n)
         pool[n].join();
       pool.clear();
-      if (afloat < expected * (1 - kEpsilon) ||
-          afloat > expected * (1 + kEpsilon))
+      if (afloat != kExpected)
         fail();
     }
   }
@@ -83,9 +104,6 @@ void test_floating_point() {
 }
 
 int main() {
-  std::cout << kThreads << " threads; "
-            << kIterations << " iterations each; total of "
-            << kExpected << "\n";
   test_floating_point();
   std::cout << "PASSED\n";
 }

SingleSource/UnitTests/Atomic/float_test.reference_output

@@ -1,4 +1,3 @@
-10 threads; 1000000 iterations each; total of 10000000
 Testing float
 Testing double
 Testing float128

SingleSource/UnitTests/Atomic/int_aligned_test.cpp

@@ -24,9 +24,17 @@
 #include <thread>
 #include <vector>
 
-#include "numeric.h"
 #include "util.h"
 
+// V >> 56 = 66, so to prevent 32-bit overflow, kExpected must be less than
+// 2^31 / 66 = 32 x 10^6.
+#ifdef SMALL_PROBLEM_SIZE
+static constexpr int kIterations = 1'000'000;
+#else
+static constexpr int kIterations = 3'000'000;
+#endif
+static constexpr int kExpected = kThreads * kIterations;
+
 template <typename T>
 void looper_int_fetch_add(T *aint, int model) {
   static constexpr T val = V >> right_shift<T>();
@@ -167,14 +175,41 @@ void test_int_fetch_xor() {
   }
 }
 
+// The xchg tests work as follows:
+//
+// Each thread increments a local copy of the shared variable, and exchanges it
+// with the shared value. Most of the time, the value moved from shared -> local
+// is one less than the value moved from local -> shared. Other times, the
+// difference is much bigger (or smaller). When this occurs, the thread
+// accumulates the difference in a local error variable. Upon completion, the
+// thread subtracts the error from the shared value, all at once.
+//
+// Like many tests, this test increments by more than 1 -- specifically, a
+// number that scales with the width of the type is picked.
+//
+template <typename T>
+void looper_int_xchg(T *aint, int model) {
+  static constexpr T val = V >> right_shift<T>();
+  __int128_t error = 0;
+  T next = *aint + val;
+  T result;
+  for (int n = 0; n < kIterations; ++n) {
+    __atomic_exchange(aint, &next, &result, model);
+    error +=
+        static_cast<__int128_t>(next) - static_cast<__int128_t>(result + val);
+    next = result + val;
+  }
+  __atomic_fetch_sub(aint, static_cast<T>(error), model);
+}
+
 template <typename T>
 void test_int_xchg() {
   static constexpr T val = V >> right_shift<T>();
   std::vector<std::thread> pool;
   for (int model : atomic_exchange_models) {
     T aint = 0;
     for (int n = 0; n < kThreads; ++n)
-      pool.emplace_back(looper_numeric_xchg_atomic<T>, &aint, model);
+      pool.emplace_back(looper_int_xchg<T>, &aint, model);
     for (int n = 0; n < kThreads; ++n)
       pool[n].join();
     pool.clear();
@@ -183,7 +218,6 @@ void test_int_xchg() {
   }
 }
 
-// See numeric.h for an explanation of numeric xchg tests.
 template <typename T>
 void looper_int_xchg_n(T *aint, int model) {
   static constexpr T val = V >> right_shift<T>();
@@ -213,6 +247,19 @@ void test_int_xchg_n() {
   }
 }
 
+// The cmpxchg tests act similar to fetch_add tests.
+template <typename T>
+void looper_int_cmpxchg(T *aint, int success_model, int fail_model) {
+  static constexpr T val = V >> right_shift<T>();
+  for (int n = 0; n < kIterations; ++n) {
+    T desired, expected = 0;
+    do {
+      desired = expected + val;
+    } while (!__atomic_compare_exchange(aint, &expected, &desired, true,
+                                        success_model, fail_model));
+  }
+}
+
 template <typename T>
 void test_int_cmpxchg() {
   static constexpr T val = V >> right_shift<T>();
@@ -221,7 +268,7 @@ void test_int_cmpxchg() {
     for (int fail_model : atomic_compare_exchange_models) {
       T aint = 0;
       for (int n = 0; n < kThreads; ++n)
-        pool.emplace_back(looper_numeric_cmpxchg<T>, &aint, success_model,
+        pool.emplace_back(looper_int_cmpxchg<T>, &aint, success_model,
                           fail_model);
       for (int n = 0; n < kThreads; ++n) pool[n].join();
       pool.clear();
@@ -231,7 +278,6 @@ void test_int_cmpxchg() {
   }
 }
 
-// See numeric.h for an explanation of numeric cmpxchg tests.
 template <typename T>
 void looper_int_cmpxchg_n(T *aint, int success_model, int fail_model) {
   static constexpr T val = V >> right_shift<T>();
@@ -298,8 +344,6 @@ void test_aligned_int() {
 }
 
 int main() {
-  std::cout << kThreads << " threads; " << kIterations
-            << " iterations each; total of " << kExpected << "\n";
   test_aligned_int();
   std::cout << "PASSED\n";
 }

SingleSource/UnitTests/Atomic/int_aligned_test.reference_output

@@ -1,4 +1,3 @@
-10 threads; 1000000 iterations each; total of 10000000
 Testing aligned uint32_t
 Testing aligned uint64_t
 Testing aligned int32_t

SingleSource/UnitTests/Atomic/int_misaligned_test.cpp

@@ -27,6 +27,13 @@
 
 #include "util.h"
 
+#ifdef SMALL_PROBLEM_SIZE
+static constexpr int kIterations = 50'000;
+#else
+static constexpr int kIterations = 500'000;
+#endif
+static constexpr int kExpected = kThreads * kIterations;
+
 template <typename T>
 struct __attribute__((packed)) misaligned {
   char byte;
@@ -179,9 +186,9 @@ void test_int_misaligned_fetch_xor() {
   }
 }
 
-// See numeric.h for an explanation of numeric xchg tests.
+// See int_aligned_test.cc for an explanation of xchg tests.
 template <typename T>
-void looper_int_misaligned_xchg_atomic(misaligned<T> &astruct, int model) {
+void looper_int_misaligned_xchg(misaligned<T> &astruct, int model) {
   static constexpr T val = V >> right_shift<T>();
   __int128_t error = 0;
   T next = astruct.data + val;
@@ -203,7 +210,7 @@ void test_int_misaligned_xchg() {
   for (int model : atomic_exchange_models) {
     astruct.data = 0;
     for (int n = 0; n < kThreads; ++n)
-      pool.emplace_back(looper_int_misaligned_xchg_atomic<T>, std::ref(astruct),
+      pool.emplace_back(looper_int_misaligned_xchg<T>, std::ref(astruct),
                         model);
     for (int n = 0; n < kThreads; ++n)
       pool[n].join();
@@ -213,7 +220,7 @@ void test_int_misaligned_xchg() {
   }
 }
 
-// See numeric.h for an explanation of numeric xchg tests.
+// See int_aligned_test.cc for an explanation of xchg tests.
 template <typename T>
 void looper_int_misaligned_xchg_n(misaligned<T> &astruct, int model) {
   static constexpr T val = V >> right_shift<T>();
@@ -246,7 +253,7 @@ void test_int_misaligned_xchg_n() {
   }
 }
 
-// See numeric.h for an explanation of numeric cmpxchg tests.
+// See int_aligned_test.cc for an explanation of cmpxchg tests.
 template <typename T>
 void looper_int_misaligned_cmpxchg(misaligned<T> &astruct, int success_model,
                                    int fail_model) {
@@ -280,7 +287,7 @@ void test_int_misaligned_cmpxchg() {
   }
 }
 
-// See numeric.h for an explanation of numeric cmpxchg tests.
+// See int_aligned_test.cc for an explanation of cmpxchg tests.
 template <typename T>
 void looper_int_misaligned_cmpxchg_n(misaligned<T> &astruct, int success_model,
                                      int fail_model) {
@@ -350,8 +357,6 @@ void test_misaligned_int() {
 }
 
 int main() {
-  std::cout << kThreads << " threads; " << kIterations
-            << " iterations each; total of " << kExpected << "\n";
   test_misaligned_int();
   std::cout << "PASSED\n";
 }

SingleSource/UnitTests/Atomic/int_misaligned_test.reference_output

@@ -1,4 +1,3 @@
-10 threads; 1000000 iterations each; total of 10000000
 Testing misaligned uint32_t
 Testing misaligned uint64_t
 Testing misaligned int32_t