[libc][math] Improve performance test framework #134501

meltq · 2025-04-05T19:13:00Z

Merges BinaryOpSingleOutputPerf.h and SingleInputSingleOutputPerf.h files into a unified PerfTest.h and update all performance tests to use this.
Improve the output printed to log file for tests.
Removes unused run_diff method and redundant run_perf call in BINARY_INPUT_SINGLE_OUTPUT_PERF_EX (previously BINARY_OP_SINGLE_OUTPUT_PERF_EX)
Change BINARY_INPUT_SINGLE_OUTPUT_PERF_EX and SINGLE_INPUT_SINGLE_OUTPUT_PERF to not define main

llvmbot · 2025-04-05T19:13:34Z

@llvm/pr-subscribers-libc

Author: Tejas Vipin (meltq)

Changes

Switches away from "Mine/Other function" to "Function A/B" and prints the names of the actual functions passed on the first line. Also uses ratio of ops per second instead of ratio of average runtime.

Full diff: https://github.com/llvm/llvm-project/pull/134501.diff

2 Files Affected:

(modified) libc/test/src/math/performance_testing/BinaryOpSingleOutputPerf.h (+26-25)
(modified) libc/test/src/math/performance_testing/SingleInputSingleOutputPerf.h (+22-21)

diff --git a/libc/test/src/math/performance_testing/BinaryOpSingleOutputPerf.h b/libc/test/src/math/performance_testing/BinaryOpSingleOutputPerf.h
index 98a1813bd7b54..8001710d83f5e 100644
--- a/libc/test/src/math/performance_testing/BinaryOpSingleOutputPerf.h
+++ b/libc/test/src/math/performance_testing/BinaryOpSingleOutputPerf.h
@@ -26,7 +26,7 @@ class BinaryOpSingleOutputPerf {
 public:
   typedef OutputType Func(InputType, InputType);
 
-  static void run_perf_in_range(Func myFunc, Func otherFunc,
+  static void run_perf_in_range(Func FuncA, Func FuncB,
                                 StorageType startingBit, StorageType endingBit,
                                 size_t N, size_t rounds, std::ofstream &log) {
     if (sizeof(StorageType) <= sizeof(size_t))
@@ -54,48 +54,48 @@ class BinaryOpSingleOutputPerf {
 
     Timer timer;
     timer.start();
-    runner(myFunc);
+    runner(FuncA);
     timer.stop();
 
-    double my_average = static_cast<double>(timer.nanoseconds()) / N / rounds;
-    log << "-- My function --\n";
+    double a_average = static_cast<double>(timer.nanoseconds()) / N / rounds;
+    log << "-- Function A --\n";
     log << "     Total time      : " << timer.nanoseconds() << " ns \n";
-    log << "     Average runtime : " << my_average << " ns/op \n";
+    log << "     Average runtime : " << a_average << " ns/op \n";
     log << "     Ops per second  : "
-        << static_cast<uint64_t>(1'000'000'000.0 / my_average) << " op/s \n";
+        << static_cast<uint64_t>(1'000'000'000.0 / a_average) << " op/s \n";
 
     timer.start();
-    runner(otherFunc);
+    runner(FuncB);
     timer.stop();
 
-    double other_average =
-        static_cast<double>(timer.nanoseconds()) / N / rounds;
-    log << "-- Other function --\n";
+    double b_average = static_cast<double>(timer.nanoseconds()) / N / rounds;
+    log << "-- Function B --\n";
     log << "     Total time      : " << timer.nanoseconds() << " ns \n";
-    log << "     Average runtime : " << other_average << " ns/op \n";
+    log << "     Average runtime : " << b_average << " ns/op \n";
     log << "     Ops per second  : "
-        << static_cast<uint64_t>(1'000'000'000.0 / other_average) << " op/s \n";
+        << static_cast<uint64_t>(1'000'000'000.0 / b_average) << " op/s \n";
 
-    log << "-- Average runtime ratio --\n";
-    log << "     Mine / Other's  : " << my_average / other_average << " \n";
+    log << "-- Average ops per second ratio --\n";
+    log << "     A / B  : " << b_average / a_average << " \n";
   }
 
-  static void run_perf(Func myFunc, Func otherFunc, int rounds,
-                       const char *logFile) {
+  static void run_perf(Func FuncA, Func FuncB, int rounds, const char *name_a,
+                       const char *name_b, const char *logFile) {
     std::ofstream log(logFile);
+    log << "Function A - " << name_a << " Function B - " << name_b << "\n";
     log << " Performance tests with inputs in denormal range:\n";
-    run_perf_in_range(myFunc, otherFunc, /* startingBit= */ StorageType(0),
+    run_perf_in_range(FuncA, FuncB, /* startingBit= */ StorageType(0),
                       /* endingBit= */ FPBits::max_subnormal().uintval(),
                       1'000'001, rounds, log);
     log << "\n Performance tests with inputs in normal range:\n";
-    run_perf_in_range(myFunc, otherFunc,
+    run_perf_in_range(FuncA, FuncB,
                       /* startingBit= */ FPBits::min_normal().uintval(),
                       /* endingBit= */ FPBits::max_normal().uintval(),
                       1'000'001, rounds, log);
     log << "\n Performance tests with inputs in normal range with exponents "
            "close to each other:\n";
     run_perf_in_range(
-        myFunc, otherFunc,
+        FuncA, FuncB,
         /* startingBit= */ FPBits(OutputType(0x1.0p-10)).uintval(),
         /* endingBit= */ FPBits(OutputType(0x1.0p+10)).uintval(), 1'000'001,
         rounds, log);
@@ -128,21 +128,22 @@ class BinaryOpSingleOutputPerf {
 } // namespace testing
 } // namespace LIBC_NAMESPACE_DECL
 
-#define BINARY_OP_SINGLE_OUTPUT_PERF(OutputType, InputType, myFunc, otherFunc, \
+#define BINARY_OP_SINGLE_OUTPUT_PERF(OutputType, InputType, FuncA, FuncB,      \
                                      filename)                                 \
   int main() {                                                                 \
     LIBC_NAMESPACE::testing::BinaryOpSingleOutputPerf<                         \
-        OutputType, InputType>::run_perf(&myFunc, &otherFunc, 1, filename);    \
+        OutputType, InputType>::run_perf(&FuncA, &FuncB, 1, #FuncA, #FuncB,    \
+                                         filename);                            \
     return 0;                                                                  \
   }
 
-#define BINARY_OP_SINGLE_OUTPUT_PERF_EX(OutputType, InputType, myFunc,         \
-                                        otherFunc, rounds, filename)           \
+#define BINARY_OP_SINGLE_OUTPUT_PERF_EX(OutputType, InputType, FuncA, FuncB,   \
+                                        rounds, filename)                      \
   {                                                                            \
     LIBC_NAMESPACE::testing::BinaryOpSingleOutputPerf<                         \
-        OutputType, InputType>::run_perf(&myFunc, &otherFunc, rounds,          \
+        OutputType, InputType>::run_perf(&FuncA, &FuncB, 1, #FuncA, #FuncB,    \
                                          filename);                            \
     LIBC_NAMESPACE::testing::BinaryOpSingleOutputPerf<                         \
-        OutputType, InputType>::run_perf(&myFunc, &otherFunc, rounds,          \
+        OutputType, InputType>::run_perf(&FuncA, &FuncB, 1, #FuncA, #FuncB,    \
                                          filename);                            \
   }
diff --git a/libc/test/src/math/performance_testing/SingleInputSingleOutputPerf.h b/libc/test/src/math/performance_testing/SingleInputSingleOutputPerf.h
index efad1259d6bf1..93c217de250e6 100644
--- a/libc/test/src/math/performance_testing/SingleInputSingleOutputPerf.h
+++ b/libc/test/src/math/performance_testing/SingleInputSingleOutputPerf.h
@@ -25,7 +25,7 @@ template <typename T> class SingleInputSingleOutputPerf {
 public:
   typedef T Func(T);
 
-  static void runPerfInRange(Func myFunc, Func otherFunc,
+  static void runPerfInRange(Func FuncA, Func FuncB,
                              StorageType startingBit, StorageType endingBit,
                              size_t rounds, std::ofstream &log) {
     size_t n = 10'010'001;
@@ -47,40 +47,41 @@ template <typename T> class SingleInputSingleOutputPerf {
 
     Timer timer;
     timer.start();
-    runner(myFunc);
+    runner(FuncA);
     timer.stop();
 
-    double myAverage = static_cast<double>(timer.nanoseconds()) / n / rounds;
-    log << "-- My function --\n";
+    double a_average = static_cast<double>(timer.nanoseconds()) / n / rounds;
+    log << "-- Function A --\n";
     log << "     Total time      : " << timer.nanoseconds() << " ns \n";
-    log << "     Average runtime : " << myAverage << " ns/op \n";
+    log << "     Average runtime : " << a_average << " ns/op \n";
     log << "     Ops per second  : "
-        << static_cast<uint64_t>(1'000'000'000.0 / myAverage) << " op/s \n";
+        << static_cast<uint64_t>(1'000'000'000.0 / a_average) << " op/s \n";
 
     timer.start();
-    runner(otherFunc);
+    runner(FuncB);
     timer.stop();
 
-    double otherAverage = static_cast<double>(timer.nanoseconds()) / n / rounds;
-    log << "-- Other function --\n";
+    double b_average = static_cast<double>(timer.nanoseconds()) / n / rounds;
+    log << "-- Function B --\n";
     log << "     Total time      : " << timer.nanoseconds() << " ns \n";
-    log << "     Average runtime : " << otherAverage << " ns/op \n";
+    log << "     Average runtime : " << b_average << " ns/op \n";
     log << "     Ops per second  : "
-        << static_cast<uint64_t>(1'000'000'000.0 / otherAverage) << " op/s \n";
+        << static_cast<uint64_t>(1'000'000'000.0 / b_average) << " op/s \n";
 
-    log << "-- Average runtime ratio --\n";
-    log << "     Mine / Other's  : " << myAverage / otherAverage << " \n";
+    log << "-- Average ops per second ratio --\n";
+    log << "     A / B  : " << b_average / a_average << " \n";
   }
 
-  static void runPerf(Func myFunc, Func otherFunc, size_t rounds,
-                      const char *logFile) {
+  static void runPerf(Func FuncA, Func FuncB, size_t rounds, const char *name_a,
+                      const char *name_b, const char *logFile) {
     std::ofstream log(logFile);
+    log << "Function A - " << name_a << " Function B - " << name_b << "\n";
     log << " Performance tests with inputs in denormal range:\n";
-    runPerfInRange(myFunc, otherFunc, /* startingBit= */ StorageType(0),
+    runPerfInRange(FuncA, FuncB, /* startingBit= */ StorageType(0),
                    /* endingBit= */ FPBits::max_subnormal().uintval(), rounds,
                    log);
     log << "\n Performance tests with inputs in normal range:\n";
-    runPerfInRange(myFunc, otherFunc,
+    runPerfInRange(FuncA, FuncB,
                    /* startingBit= */ FPBits::min_normal().uintval(),
                    /* endingBit= */ FPBits::max_normal().uintval(), rounds,
                    log);
@@ -90,16 +91,16 @@ template <typename T> class SingleInputSingleOutputPerf {
 } // namespace testing
 } // namespace LIBC_NAMESPACE_DECL
 
-#define SINGLE_INPUT_SINGLE_OUTPUT_PERF(T, myFunc, otherFunc, filename)        \
+#define SINGLE_INPUT_SINGLE_OUTPUT_PERF(T, FuncA, FuncB, filename)             \
   int main() {                                                                 \
     LIBC_NAMESPACE::testing::SingleInputSingleOutputPerf<T>::runPerf(          \
-        &myFunc, &otherFunc, 1, filename);                                     \
+        &FuncA, &FuncB, 1, #FuncA, #FuncB, filename);                          \
     return 0;                                                                  \
   }
 
-#define SINGLE_INPUT_SINGLE_OUTPUT_PERF_EX(T, myFunc, otherFunc, rounds,       \
+#define SINGLE_INPUT_SINGLE_OUTPUT_PERF_EX(T, FuncA, FuncB, rounds,            \
                                            filename)                           \
   {                                                                            \
     LIBC_NAMESPACE::testing::SingleInputSingleOutputPerf<T>::runPerf(          \
-        &myFunc, &otherFunc, rounds, filename);                                \
+        &FuncA, &FuncB, rounds, #FuncA, #FuncB, filename);                     \
   }

github-actions · 2025-04-05T19:15:11Z

✅ With the latest revision this PR passed the C/C++ code formatter.

meltq · 2025-04-05T19:15:14Z

Old output

 Performance tests with inputs in denormal range:
-- My function --
     Total time      : 4545310 ns 
     Average runtime : 4443.12 ns/op 
     Ops per second  : 225067 op/s 
-- Other function --
     Total time      : 654400 ns 
     Average runtime : 639.687 ns/op 
     Ops per second  : 1563264 op/s 
-- Average runtime ratio --
     Mine / Other's  : 6.94577

New output

Function A - LIBC_NAMESPACE::hypotf16 Function B - LIBC_NAMESPACE::fputil::hypot<float16>
 Performance tests with inputs in denormal range:
-- Function A --
     Total time      : 2670736 ns 
     Average runtime : 2610.69 ns/op 
     Ops per second  : 383040 op/s 
-- Function B --
     Total time      : 499614 ns 
     Average runtime : 488.381 ns/op 
     Ops per second  : 2047580 op/s 
-- Average ops per second ratio --
     A / B  : 0.18707

meltq · 2025-04-05T19:18:44Z

@lntue @overmighty requesting review

meltq · 2025-04-15T07:44:34Z

Ping

overmighty · 2025-04-15T12:07:56Z

I would prefer function names to be printed like this personally:

 Performance tests with inputs in denormal range:
-- Function A: LIBC_NAMESPACE::hypotf16 --
     Total time      : 2670736 ns
     Average runtime : 2610.69 ns/op
     Ops per second  : 383040 op/s
-- Function B: LIBC_NAMESPACE::fputil::hypot<float16> --
     Total time      : 499614 ns
     Average runtime : 488.381 ns/op
     Ops per second  : 2047580 op/s
-- Average ops per second ratio --
     A / B  : 0.18707

libc/test/src/math/performance_testing/SingleInputSingleOutputPerf.h

meltq · 2025-04-18T18:45:58Z

Made all suggested changes.

michaelrj-google

LGTM, I'll let @lntue make the final call. Thanks for this cleanup!

meltq · 2025-04-24T04:46:39Z

I'll need help for this to be merged since I don't have write perms.

- Merges `BinaryOpSingleOutputPerf.h` and `SingleInputSingleOutputPerf.h` files into a unified `PerfTest.h` and update all performance tests to use this. - Improve the output printed to log file for tests. - Removes unused `run_diff` method and redundant `run_perf` call in `BINARY_INPUT_SINGLE_OUTPUT_PERF_EX` (previously `BINARY_OP_SINGLE_OUTPUT_PERF_EX`) - Change `BINARY_INPUT_SINGLE_OUTPUT_PERF_EX` and `SINGLE_INPUT_SINGLE_OUTPUT_PERF` to not define `main`

meltq added 2 commits April 5, 2025 16:51

Binary op perf output improvement

6d87fee

Single input perf output improvement

441efdf

llvmbot added the libc label Apr 5, 2025

Formatting changes

714f396

overmighty requested review from overmighty and lntue April 5, 2025 19:35

lntue requested review from jhuber6 and michaelrj-google April 15, 2025 09:38

michaelrj-google reviewed Apr 15, 2025

View reviewed changes

libc/test/src/math/performance_testing/SingleInputSingleOutputPerf.h Outdated Show resolved Hide resolved

Add PerfTest and remove old files

9bb2e8a

meltq changed the title ~~[libc][math] Improve performance test output~~ [libc][math] Improve performance test framework Apr 18, 2025

michaelrj-google reviewed Apr 22, 2025

View reviewed changes

lntue approved these changes Apr 23, 2025

View reviewed changes

lntue merged commit dde00f5 into llvm:main Apr 24, 2025
16 checks passed

meltq deleted the perf_cleanup branch April 26, 2025 05:06

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

[libc][math] Improve performance test framework #134501

[libc][math] Improve performance test framework #134501

Uh oh!

meltq commented Apr 5, 2025 •

edited

Loading

Uh oh!

llvmbot commented Apr 5, 2025

Uh oh!

github-actions bot commented Apr 5, 2025 •

edited

Loading

Uh oh!

meltq commented Apr 5, 2025

Uh oh!

meltq commented Apr 5, 2025

Uh oh!

meltq commented Apr 15, 2025

Uh oh!

overmighty commented Apr 15, 2025

Uh oh!

Uh oh!

meltq commented Apr 18, 2025

Uh oh!

michaelrj-google left a comment

Uh oh!

meltq commented Apr 24, 2025

Uh oh!

Uh oh!

Uh oh!

[libc][math] Improve performance test framework #134501

[libc][math] Improve performance test framework #134501

Uh oh!

Conversation

meltq commented Apr 5, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

llvmbot commented Apr 5, 2025

Uh oh!

github-actions bot commented Apr 5, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

meltq commented Apr 5, 2025

Uh oh!

meltq commented Apr 5, 2025

Uh oh!

meltq commented Apr 15, 2025

Uh oh!

overmighty commented Apr 15, 2025

Uh oh!

Uh oh!

meltq commented Apr 18, 2025

Uh oh!

michaelrj-google left a comment

Choose a reason for hiding this comment

Uh oh!

meltq commented Apr 24, 2025

Uh oh!

Uh oh!

Uh oh!

meltq commented Apr 5, 2025 •

edited

Loading

github-actions bot commented Apr 5, 2025 •

edited

Loading