[libc][math] Optimize copysign{,f,f16} and fabs{,f,f16} with builtins when available #99037
Conversation
…able Remove support for __builtin_frexpf16 as it decreases performance.
|
@llvm/pr-subscribers-libc Author: OverMighty (overmighty) ChangesPatch is 26.77 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/99037.diff 21 Files Affected:
diff --git a/libc/cmake/modules/CheckCompilerFeatures.cmake b/libc/cmake/modules/CheckCompilerFeatures.cmake
index a6d793d495c45..2bab968f901eb 100644
--- a/libc/cmake/modules/CheckCompilerFeatures.cmake
+++ b/libc/cmake/modules/CheckCompilerFeatures.cmake
@@ -5,6 +5,7 @@
set(
ALL_COMPILER_FEATURES
"builtin_ceil_floor_rint_trunc"
+ "builtin_fmax_fmin"
"builtin_round"
"builtin_roundeven"
"float16"
@@ -82,6 +83,8 @@ foreach(feature IN LISTS ALL_COMPILER_FEATURES)
set(LIBC_COMPILER_HAS_FIXED_POINT TRUE)
elseif(${feature} STREQUAL "builtin_ceil_floor_rint_trunc")
set(LIBC_COMPILER_HAS_BUILTIN_CEIL_FLOOR_RINT_TRUNC TRUE)
+ elseif(${feature} STREQUAL "builtin_fmax_fmin")
+ set(LIBC_COMPILER_HAS_BUILTIN_FMAX_FMIN TRUE)
elseif(${feature} STREQUAL "builtin_round")
set(LIBC_COMPILER_HAS_BUILTIN_ROUND TRUE)
elseif(${feature} STREQUAL "builtin_roundeven")
diff --git a/libc/cmake/modules/LLVMLibCCompileOptionRules.cmake b/libc/cmake/modules/LLVMLibCCompileOptionRules.cmake
index 253da4ae890e5..ead578f95ac72 100644
--- a/libc/cmake/modules/LLVMLibCCompileOptionRules.cmake
+++ b/libc/cmake/modules/LLVMLibCCompileOptionRules.cmake
@@ -6,6 +6,7 @@ function(_get_compile_options_from_flags output_var)
endif()
check_flag(ADD_ROUND_OPT_FLAG ${ROUND_OPT_FLAG} ${ARGN})
check_flag(ADD_EXPLICIT_SIMD_OPT_FLAG ${EXPLICIT_SIMD_OPT_FLAG} ${ARGN})
+ check_flag(ADD_MISC_MATH_BASIC_OPS_OPT_FLAG ${MISC_MATH_BASIC_OPS_OPT_FLAG} ${ARGN})
if(LLVM_COMPILER_IS_GCC_COMPATIBLE)
if(ADD_FMA_FLAG)
@@ -37,6 +38,12 @@ function(_get_compile_options_from_flags output_var)
if(ADD_EXPLICIT_SIMD_OPT_FLAG)
list(APPEND compile_options "-D__LIBC_EXPLICIT_SIMD_OPT")
endif()
+ if(ADD_MISC_MATH_BASIC_OPS_OPT_FLAG)
+ list(APPEND compile_options "-D__LIBC_MISC_MATH_BASIC_OPS_OPT")
+ if(LIBC_COMPILER_HAS_BUILTIN_FMAX_FMIN)
+ list(APPEND compile_options "-D__LIBC_USE_BUILTIN_FMAX_FMIN")
+ endif()
+ endif()
elseif(MSVC)
if(ADD_FMA_FLAG)
list(APPEND compile_options "/arch:AVX2")
diff --git a/libc/cmake/modules/LLVMLibCFlagRules.cmake b/libc/cmake/modules/LLVMLibCFlagRules.cmake
index eca7ba8d183e6..4398fe55db5aa 100644
--- a/libc/cmake/modules/LLVMLibCFlagRules.cmake
+++ b/libc/cmake/modules/LLVMLibCFlagRules.cmake
@@ -263,6 +263,9 @@ set(FMA_OPT_FLAG "FMA_OPT")
set(ROUND_OPT_FLAG "ROUND_OPT")
# This flag controls whether we use explicit SIMD instructions or not.
set(EXPLICIT_SIMD_OPT_FLAG "EXPLICIT_SIMD_OPT")
+# This flag controls whether we use compiler builtin functions to implement
+# various basic math operations or not.
+set(MISC_MATH_BASIC_OPS_OPT_FLAG "MISC_MATH_BASIC_OPS_OPT")
# Skip FMA_OPT flag for targets that don't support fma.
if(NOT((LIBC_TARGET_ARCHITECTURE_IS_X86 AND (LIBC_CPU_FEATURES MATCHES "FMA")) OR
diff --git a/libc/cmake/modules/compiler_features/check_builtin_fmax_fmin.cpp b/libc/cmake/modules/compiler_features/check_builtin_fmax_fmin.cpp
new file mode 100644
index 0000000000000..a962df33e31c4
--- /dev/null
+++ b/libc/cmake/modules/compiler_features/check_builtin_fmax_fmin.cpp
@@ -0,0 +1,14 @@
+_Float16 try_builtin_fmaxf16(_Float16 x, _Float16 y) {
+ return __builtin_fmaxf16(x, y);
+}
+_Float16 try_builtin_fminf16(_Float16 x, _Float16 y) {
+ return __builtin_fminf16(x, y);
+}
+
+float try_builtin_fmaxf(float x, float y) { return __builtin_fmaxf(x, y); }
+float try_builtin_fminf(float x, float y) { return __builtin_fminf(x, y); }
+
+double try_builtin_fmax(double x, double y) { return __builtin_fmax(x, y); }
+double try_builtin_fmin(double x, double y) { return __builtin_fmin(x, y); }
+
+extern "C" void _start() {}
diff --git a/libc/src/math/generic/CMakeLists.txt b/libc/src/math/generic/CMakeLists.txt
index c2f58fb1a4f71..3d713368251f6 100644
--- a/libc/src/math/generic/CMakeLists.txt
+++ b/libc/src/math/generic/CMakeLists.txt
@@ -400,6 +400,8 @@ add_entrypoint_object(
libc.src.__support.FPUtil.basic_operations
COMPILE_OPTIONS
-O2
+ FLAGS
+ MISC_MATH_BASIC_OPS_OPT
)
add_entrypoint_object(
@@ -412,6 +414,8 @@ add_entrypoint_object(
libc.src.__support.FPUtil.basic_operations
COMPILE_OPTIONS
-O2
+ FLAGS
+ MISC_MATH_BASIC_OPS_OPT
)
add_entrypoint_object(
@@ -437,6 +441,8 @@ add_entrypoint_object(
libc.src.__support.FPUtil.basic_operations
COMPILE_OPTIONS
-O3
+ FLAGS
+ MISC_MATH_BASIC_OPS_OPT
)
add_entrypoint_object(
@@ -1406,6 +1412,8 @@ add_entrypoint_object(
libc.src.__support.FPUtil.manipulation_functions
COMPILE_OPTIONS
-O3
+ FLAGS
+ MISC_MATH_BASIC_OPS_OPT
)
add_entrypoint_object(
@@ -1418,6 +1426,8 @@ add_entrypoint_object(
libc.src.__support.FPUtil.manipulation_functions
COMPILE_OPTIONS
-O3
+ FLAGS
+ MISC_MATH_BASIC_OPS_OPT
)
add_entrypoint_object(
@@ -1443,6 +1453,8 @@ add_entrypoint_object(
libc.src.__support.FPUtil.manipulation_functions
COMPILE_OPTIONS
-O3
+ FLAGS
+ MISC_MATH_BASIC_OPS_OPT
)
add_entrypoint_object(
@@ -2202,6 +2214,8 @@ add_entrypoint_object(
libc.src.__support.FPUtil.basic_operations
COMPILE_OPTIONS
-O2
+ FLAGS
+ MISC_MATH_BASIC_OPS_OPT
)
add_entrypoint_object(
@@ -2214,6 +2228,8 @@ add_entrypoint_object(
libc.src.__support.FPUtil.basic_operations
COMPILE_OPTIONS
-O2
+ FLAGS
+ MISC_MATH_BASIC_OPS_OPT
)
add_entrypoint_object(
@@ -2239,6 +2255,8 @@ add_entrypoint_object(
libc.src.__support.FPUtil.basic_operations
COMPILE_OPTIONS
-O3
+ FLAGS
+ MISC_MATH_BASIC_OPS_OPT
)
add_entrypoint_object(
@@ -2450,6 +2468,8 @@ add_entrypoint_object(
libc.src.__support.FPUtil.basic_operations
COMPILE_OPTIONS
-O2
+ FLAGS
+ MISC_MATH_BASIC_OPS_OPT
)
add_entrypoint_object(
@@ -2462,6 +2482,8 @@ add_entrypoint_object(
libc.src.__support.FPUtil.basic_operations
COMPILE_OPTIONS
-O2
+ FLAGS
+ MISC_MATH_BASIC_OPS_OPT
)
add_entrypoint_object(
@@ -2487,6 +2509,8 @@ add_entrypoint_object(
libc.src.__support.FPUtil.basic_operations
COMPILE_OPTIONS
-O3
+ FLAGS
+ MISC_MATH_BASIC_OPS_OPT
)
add_entrypoint_object(
diff --git a/libc/src/math/generic/copysign.cpp b/libc/src/math/generic/copysign.cpp
index 149d725af08e2..186bb2c5983f4 100644
--- a/libc/src/math/generic/copysign.cpp
+++ b/libc/src/math/generic/copysign.cpp
@@ -14,7 +14,11 @@
namespace LIBC_NAMESPACE_DECL {
LLVM_LIBC_FUNCTION(double, copysign, (double x, double y)) {
+#ifdef __LIBC_MISC_MATH_BASIC_OPS_OPT
+ return __builtin_copysign(x, y);
+#else
return fputil::copysign(x, y);
+#endif
}
} // namespace LIBC_NAMESPACE_DECL
diff --git a/libc/src/math/generic/copysignf.cpp b/libc/src/math/generic/copysignf.cpp
index 17cd70d37c308..c79e50b61ebda 100644
--- a/libc/src/math/generic/copysignf.cpp
+++ b/libc/src/math/generic/copysignf.cpp
@@ -14,7 +14,11 @@
namespace LIBC_NAMESPACE_DECL {
LLVM_LIBC_FUNCTION(float, copysignf, (float x, float y)) {
+#ifdef __LIBC_MISC_MATH_BASIC_OPS_OPT
+ return __builtin_copysignf(x, y);
+#else
return fputil::copysign(x, y);
+#endif
}
} // namespace LIBC_NAMESPACE_DECL
diff --git a/libc/src/math/generic/copysignf16.cpp b/libc/src/math/generic/copysignf16.cpp
index 42695b3b4a6de..546622f049ebe 100644
--- a/libc/src/math/generic/copysignf16.cpp
+++ b/libc/src/math/generic/copysignf16.cpp
@@ -14,7 +14,11 @@
namespace LIBC_NAMESPACE_DECL {
LLVM_LIBC_FUNCTION(float16, copysignf16, (float16 x, float16 y)) {
+#ifdef __LIBC_MISC_MATH_BASIC_OPS_OPT
+ return __builtin_copysignf16(x, y);
+#else
return fputil::copysign(x, y);
+#endif
}
} // namespace LIBC_NAMESPACE_DECL
diff --git a/libc/src/math/generic/fabs.cpp b/libc/src/math/generic/fabs.cpp
index 472297aecb2f7..55fa958cd7c00 100644
--- a/libc/src/math/generic/fabs.cpp
+++ b/libc/src/math/generic/fabs.cpp
@@ -13,6 +13,12 @@
namespace LIBC_NAMESPACE_DECL {
-LLVM_LIBC_FUNCTION(double, fabs, (double x)) { return fputil::abs(x); }
+LLVM_LIBC_FUNCTION(double, fabs, (double x)) {
+#ifdef __LIBC_MISC_MATH_BASIC_OPS_OPT
+ return __builtin_fabs(x);
+#else
+ return fputil::abs(x);
+#endif
+}
} // namespace LIBC_NAMESPACE_DECL
diff --git a/libc/src/math/generic/fabsf.cpp b/libc/src/math/generic/fabsf.cpp
index ad4fcb30c795d..2ba18d09bbd5b 100644
--- a/libc/src/math/generic/fabsf.cpp
+++ b/libc/src/math/generic/fabsf.cpp
@@ -13,6 +13,12 @@
namespace LIBC_NAMESPACE_DECL {
-LLVM_LIBC_FUNCTION(float, fabsf, (float x)) { return fputil::abs(x); }
+LLVM_LIBC_FUNCTION(float, fabsf, (float x)) {
+#ifdef __LIBC_MISC_MATH_BASIC_OPS_OPT
+ return __builtin_fabsf(x);
+#else
+ return fputil::abs(x);
+#endif
+}
} // namespace LIBC_NAMESPACE_DECL
diff --git a/libc/src/math/generic/fabsf16.cpp b/libc/src/math/generic/fabsf16.cpp
index 57671fb6067e2..2f982517614c4 100644
--- a/libc/src/math/generic/fabsf16.cpp
+++ b/libc/src/math/generic/fabsf16.cpp
@@ -13,6 +13,12 @@
namespace LIBC_NAMESPACE_DECL {
-LLVM_LIBC_FUNCTION(float16, fabsf16, (float16 x)) { return fputil::abs(x); }
+LLVM_LIBC_FUNCTION(float16, fabsf16, (float16 x)) {
+#ifdef __LIBC_MISC_MATH_BASIC_OPS_OPT
+ return __builtin_fabsf16(x);
+#else
+ return fputil::abs(x);
+#endif
+}
} // namespace LIBC_NAMESPACE_DECL
diff --git a/libc/src/math/generic/fmaximum_num.cpp b/libc/src/math/generic/fmaximum_num.cpp
index 33df7daa380df..1bfc1514393ee 100644
--- a/libc/src/math/generic/fmaximum_num.cpp
+++ b/libc/src/math/generic/fmaximum_num.cpp
@@ -1,4 +1,4 @@
-//===-- Implementation of fmaximum_num function----------------------------===//
+//===-- Implementation of fmaximum_num function ---------------------------===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
@@ -14,7 +14,11 @@
namespace LIBC_NAMESPACE_DECL {
LLVM_LIBC_FUNCTION(double, fmaximum_num, (double x, double y)) {
+#ifdef __LIBC_USE_BUILTIN_FMAX_FMIN
+ return __builtin_fmax(x, y);
+#else
return fputil::fmaximum_num(x, y);
+#endif
}
} // namespace LIBC_NAMESPACE_DECL
diff --git a/libc/src/math/generic/fmaximum_numf.cpp b/libc/src/math/generic/fmaximum_numf.cpp
index 1577080ba2c25..f8c69fa78be3d 100644
--- a/libc/src/math/generic/fmaximum_numf.cpp
+++ b/libc/src/math/generic/fmaximum_numf.cpp
@@ -1,4 +1,4 @@
-//===-- Implementation of fmaximum_numf function---------------------------===//
+//===-- Implementation of fmaximum_numf function --------------------------===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
@@ -14,7 +14,11 @@
namespace LIBC_NAMESPACE_DECL {
LLVM_LIBC_FUNCTION(float, fmaximum_numf, (float x, float y)) {
+#ifdef __LIBC_USE_BUILTIN_FMAX_FMIN
+ return __builtin_fmaxf(x, y);
+#else
return fputil::fmaximum_num(x, y);
+#endif
}
} // namespace LIBC_NAMESPACE_DECL
diff --git a/libc/src/math/generic/fmaximum_numf16.cpp b/libc/src/math/generic/fmaximum_numf16.cpp
index 394ce8b5fe4f3..6a012d38abea4 100644
--- a/libc/src/math/generic/fmaximum_numf16.cpp
+++ b/libc/src/math/generic/fmaximum_numf16.cpp
@@ -14,7 +14,11 @@
namespace LIBC_NAMESPACE_DECL {
LLVM_LIBC_FUNCTION(float16, fmaximum_numf16, (float16 x, float16 y)) {
+#ifdef __LIBC_USE_BUILTIN_FMAX_FMIN
+ return __builtin_fmaxf16(x, y);
+#else
return fputil::fmaximum_num(x, y);
+#endif
}
} // namespace LIBC_NAMESPACE_DECL
diff --git a/libc/src/math/generic/fminimum_num.cpp b/libc/src/math/generic/fminimum_num.cpp
index 3ff79def58075..5b9c426ca50c2 100644
--- a/libc/src/math/generic/fminimum_num.cpp
+++ b/libc/src/math/generic/fminimum_num.cpp
@@ -1,4 +1,4 @@
-//===-- Implementation of fminimum_num function----------------------------===//
+//===-- Implementation of fminimum_num function ---------------------------===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
@@ -14,7 +14,11 @@
namespace LIBC_NAMESPACE_DECL {
LLVM_LIBC_FUNCTION(double, fminimum_num, (double x, double y)) {
+#ifdef __LIBC_USE_BUILTIN_FMAX_FMIN
+ return __builtin_fmin(x, y);
+#else
return fputil::fminimum_num(x, y);
+#endif
}
} // namespace LIBC_NAMESPACE_DECL
diff --git a/libc/src/math/generic/fminimum_numf.cpp b/libc/src/math/generic/fminimum_numf.cpp
index c7ac99b14bd5a..6b6f905e63de3 100644
--- a/libc/src/math/generic/fminimum_numf.cpp
+++ b/libc/src/math/generic/fminimum_numf.cpp
@@ -1,4 +1,4 @@
-//===-- Implementation of fminimum_numf function---------------------------===//
+//===-- Implementation of fminimum_numf function --------------------------===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
@@ -14,7 +14,11 @@
namespace LIBC_NAMESPACE_DECL {
LLVM_LIBC_FUNCTION(float, fminimum_numf, (float x, float y)) {
+#ifdef __LIBC_USE_BUILTIN_FMAX_FMIN
+ return __builtin_fminf(x, y);
+#else
return fputil::fminimum_num(x, y);
+#endif
}
} // namespace LIBC_NAMESPACE_DECL
diff --git a/libc/src/math/generic/fminimum_numf16.cpp b/libc/src/math/generic/fminimum_numf16.cpp
index 0af7205713c10..8e48aaf27070f 100644
--- a/libc/src/math/generic/fminimum_numf16.cpp
+++ b/libc/src/math/generic/fminimum_numf16.cpp
@@ -14,7 +14,11 @@
namespace LIBC_NAMESPACE_DECL {
LLVM_LIBC_FUNCTION(float16, fminimum_numf16, (float16 x, float16 y)) {
+#ifdef __LIBC_USE_BUILTIN_FMAX_FMIN
+ return __builtin_fminf16(x, y);
+#else
return fputil::fminimum_num(x, y);
+#endif
}
} // namespace LIBC_NAMESPACE_DECL
diff --git a/libc/test/src/math/performance_testing/BinaryOpSingleOutputPerf.h b/libc/test/src/math/performance_testing/BinaryOpSingleOutputPerf.h
index 995e41ba84b03..1ab0afbc9cbe8 100644
--- a/libc/test/src/math/performance_testing/BinaryOpSingleOutputPerf.h
+++ b/libc/test/src/math/performance_testing/BinaryOpSingleOutputPerf.h
@@ -10,6 +10,7 @@
#include "src/__support/macros/config.h"
#include "test/src/math/performance_testing/Timer.h"
+#include <algorithm>
#include <cstddef>
#include <fstream>
@@ -28,11 +29,11 @@ template <typename T> class BinaryOpSingleOutputPerf {
static void run_perf_in_range(Func myFunc, Func otherFunc,
StorageType startingBit, StorageType endingBit,
size_t N, size_t rounds, std::ofstream &log) {
- if (endingBit - startingBit < N)
- N = endingBit - startingBit;
+ if (sizeof(StorageType) <= sizeof(size_t))
+ N = std::min(N, static_cast<size_t>(endingBit - startingBit));
auto runner = [=](Func func) {
- volatile T result;
+ [[maybe_unused]] volatile T result;
if (endingBit < startingBit) {
return;
}
diff --git a/libc/test/src/math/performance_testing/CMakeLists.txt b/libc/test/src/math/performance_testing/CMakeLists.txt
index a75becba04d07..a4059c8ff4dd8 100644
--- a/libc/test/src/math/performance_testing/CMakeLists.txt
+++ b/libc/test/src/math/performance_testing/CMakeLists.txt
@@ -402,3 +402,23 @@ add_perf_binary(
LINK_LIBRARIES
LibcFPTestHelpers
)
+
+add_perf_binary(
+ misc_basic_ops_perf
+ SRCS
+ misc_basic_ops_perf.cpp
+ DEPENDS
+ .binary_op_single_output_diff
+ .single_input_single_output_diff
+ libc.src.math.copysignf
+ libc.src.math.copysignf16
+ libc.src.math.fabsf
+ libc.src.math.fabsf16
+ libc.src.math.fmaximum_numf
+ libc.src.math.fmaximum_numf16
+ libc.src.math.fminimum_numf
+ libc.src.math.fminimum_numf16
+ libc.src.math.frexpf16
+ COMPILE_OPTIONS
+ -fno-builtin
+)
diff --git a/libc/test/src/math/performance_testing/SingleInputSingleOutputPerf.h b/libc/test/src/math/performance_testing/SingleInputSingleOutputPerf.h
index 48ae43d6315e3..e0beb729cb9f5 100644
--- a/libc/test/src/math/performance_testing/SingleInputSingleOutputPerf.h
+++ b/libc/test/src/math/performance_testing/SingleInputSingleOutputPerf.h
@@ -10,6 +10,7 @@
#include "src/__support/macros/config.h"
#include "test/src/math/performance_testing/Timer.h"
+#include <algorithm>
#include <fstream>
namespace LIBC_NAMESPACE_DECL {
@@ -26,16 +27,21 @@ template <typename T> class SingleInputSingleOutputPerf {
static void runPerfInRange(Func myFunc, Func otherFunc,
StorageType startingBit, StorageType endingBit,
- std::ofstream &log) {
+ size_t rounds, std::ofstream &log) {
+ size_t n = 10'010'001;
+ if (sizeof(StorageType) <= sizeof(size_t))
+ n = std::min(n, static_cast<size_t>(endingBit - startingBit));
+
auto runner = [=](Func func) {
- constexpr StorageType N = 10'010'001;
- StorageType step = (endingBit - startingBit) / N;
+ StorageType step = (endingBit - startingBit) / n;
if (step == 0)
step = 1;
- volatile T result;
- for (StorageType bits = startingBit; bits < endingBit; bits += step) {
- T x = FPBits(bits).get_val();
- result = func(x);
+ [[maybe_unused]] volatile T result;
+ for (size_t i = 0; i < rounds; i++) {
+ for (StorageType bits = startingBit; bits < endingBit; bits += step) {
+ T x = FPBits(bits).get_val();
+ result = func(x);
+ }
}
};
@@ -44,8 +50,7 @@ template <typename T> class SingleInputSingleOutputPerf {
runner(myFunc);
timer.stop();
- StorageType numberOfRuns = endingBit - startingBit + 1;
- double myAverage = static_cast<double>(timer.nanoseconds()) / numberOfRuns;
+ double myAverage = static_cast<double>(timer.nanoseconds()) / n / rounds;
log << "-- My function --\n";
log << " Total time : " << timer.nanoseconds() << " ns \n";
log << " Average runtime : " << myAverage << " ns/op \n";
@@ -56,8 +61,7 @@ template <typename T> class SingleInputSingleOutputPerf {
runner(otherFunc);
timer.stop();
- double otherAverage =
- static_cast<double>(timer.nanoseconds()) / numberOfRuns;
+ double otherAverage = static_cast<double>(timer.nanoseconds()) / n / rounds;
log << "-- Other function --\n";
log << " Total time : " << timer.nanoseconds() << " ns \n";
log << " Average runtime : " << otherAverage << " ns/op \n";
@@ -68,15 +72,18 @@ template <typename T> class SingleInputSingleOutputPerf {
log << " Mine / Other's : " << myAverage / otherAverage << " \n";
}
- static void runPerf(Func myFunc, Func otherFunc, const char *logFile) {
+ static void runPerf(Func myFunc, Func otherFunc, size_t rounds,
+ const char *logFile) {
std::ofstream log(logFile);
log << " Performance tests with inputs in denormal range:\n";
runPerfInRange(myFunc, otherFunc, /* startingBit= */ StorageType(0),
- /* endingBit= */ FPBits::max_subnormal().uintval(), log);
+ /* endingBit= */ FPBits::max_subnormal().uintval(), rounds,
+ log);
log << "\n Performance tests with inputs in normal range:\n";
runPerfInRange(myFunc, otherFunc,
/* startingBit= */ FPBits::min_normal().uintval(),
- /* endingBit= */ FPBits::max_normal().uintval(), log);
+ /* endingBit= */ FPBits::max_normal().uintval(), rounds,
+ log);
}
};
@@ -86,6 +93,13 @@ template <typename T> class SingleInputSingleOutputPerf {
#define SINGLE_INPUT_SINGLE_OUTPUT_PERF(T, myFunc, otherFunc, filename) \
int main() { \
LIBC_NAMESPACE::testing::SingleInputSingleOutputPerf<T>::runPerf( \
- &myFunc, &otherFunc, filename); \
+ &myFunc, &otherFunc, 1, filename); \
return 0; \
}
+
+#define SINGLE_INPUT_SINGLE_OUTPUT_PERF_EX(T, myFunc, otherFunc, rounds, \
+ filename) \
+ { \
+ LIBC_NAMESPACE::testing::SingleInputSingleOutputPerf<T>::runPerf( \
+ &myFunc, &otherFunc, rounds, filename); \
+ }
diff --git a/libc/test/src/m...
[truncated]
|
|
Before:
After:
|
libc/cmake/modules/compiler_features/check_builtin_fmax_fmin.cpp
Outdated
Show resolved
Hide resolved
|
Just FYI, if we get this working in general we can probably remove a lot of GPU implementations that are just |
…able
Split checks for __builtin_{fmax,fmin}f16 into separate feature.
|
For some reason, on x86, >>> f"{2141192192:>032b}"
'01111111101000000000000000000000'
>>>Code generated by Clang without 0000000000004dc0 <__llvm_libc_19_0_0_git::fmaximum_numf(float, float)>:
4dc0: 55 push rbp
4dc1: 48 89 e5 mov rbp, rsp
4dc4: 0f 28 d0 movaps xmm2, xmm0
4dc7: f3 0f c2 d0 03 cmpunordss xmm2, xmm0
4dcc: 0f 28 da movaps xmm3, xmm2
4dcf: 0f 54 d9 andps xmm3, xmm1
4dd2: f3 0f 5f c8 maxss xmm1, xmm0
4dd6: 0f 55 d1 andnps xmm2, xmm1
4dd9: 0f 56 d3 orps xmm2, xmm3
4ddc: 0f 28 c2 movaps xmm0, xmm2
4ddf: 5d pop rbp
4de0: c3 retCode generated by Clang with 0000000000004d50 <__llvm_libc_19_0_0_git::fmaximum_numf(float, float)>:
4d50: 55 push rbp
4d51: 48 89 e5 mov rbp, rsp
4d54: c5 f2 5f d0 vmaxss xmm2, xmm1, xmm0
4d58: c5 fa c2 c0 03 vcmpunordss xmm0, xmm0, xmm0
4d5d: c4 e3 69 4a c1 00 vblendvps xmm0, xmm2, xmm1, xmm0
4d63: 5d pop rbp
4d64: c3 ret |
|
The problem is that |
Let skip the builtin / hardware instructions for these functions. |
…able
Revert "Split checks for __builtin_{fmax,fmin}f16 into separate feature."
…able
Remove support for __builtin_{fmax,fmin}* due to incorrect results on x86-64 and AArch64.
…able Remove benchmark for frexpf16.
…able Disable __builtin_fabsf16 on GCC for x86.
libc/test/src/math/performance_testing/BinaryOpSingleOutputPerf.h
Outdated
Show resolved
Hide resolved
libc/test/src/math/performance_testing/SingleInputSingleOutputPerf.h
Outdated
Show resolved
Hide resolved
… when available (#99037) Summary: Test Plan: Reviewers: Subscribers: Tasks: Tags: Differential Revision: https://phabricator.intern.facebook.com/D60251098
No description provided.