[libc-commits] [libc] [libc][math] Optimize maximum and minimum functions using builtins when available (PR #100002)
via libc-commits
libc-commits at lists.llvm.org
Mon Jul 22 16:18:52 PDT 2024
https://github.com/overmighty created https://github.com/llvm/llvm-project/pull/100002
None
>From 441993ff267ce7192e3ca88e309fef425b483f0d Mon Sep 17 00:00:00 2001
From: OverMighty <its.overmighty at gmail.com>
Date: Mon, 22 Jul 2024 20:11:49 +0200
Subject: [PATCH] [libc][math] Optimize maximum and minimum functions using
builtins when available
---
.../cmake/modules/CheckCompilerFeatures.cmake | 6 +
.../modules/LLVMLibCCheckCpuFeatures.cmake | 1 +
.../modules/LLVMLibCCompileOptionRules.cmake | 11 ++
.../check_builtin_fmax_fmin.cpp | 7 ++
.../check_builtin_fmaxf16_fminf16.cpp | 9 ++
.../modules/cpu_features/check_FullFP16.cpp | 5 +
libc/src/__support/FPUtil/BasicOperations.h | 116 ++++++++++++++----
libc/src/__support/FPUtil/CMakeLists.txt | 2 +
.../macros/properties/cpu_features.h | 4 +
libc/src/math/generic/CMakeLists.txt | 72 ++++++++---
.../math/performance_testing/CMakeLists.txt | 22 ++++
.../max_min_funcs_perf.cpp | 75 +++++++++++
12 files changed, 288 insertions(+), 42 deletions(-)
create mode 100644 libc/cmake/modules/compiler_features/check_builtin_fmax_fmin.cpp
create mode 100644 libc/cmake/modules/compiler_features/check_builtin_fmaxf16_fminf16.cpp
create mode 100644 libc/cmake/modules/cpu_features/check_FullFP16.cpp
create mode 100644 libc/test/src/math/performance_testing/max_min_funcs_perf.cpp
diff --git a/libc/cmake/modules/CheckCompilerFeatures.cmake b/libc/cmake/modules/CheckCompilerFeatures.cmake
index a6d793d495c45..8b02c590f11d0 100644
--- a/libc/cmake/modules/CheckCompilerFeatures.cmake
+++ b/libc/cmake/modules/CheckCompilerFeatures.cmake
@@ -5,6 +5,8 @@
set(
ALL_COMPILER_FEATURES
"builtin_ceil_floor_rint_trunc"
+ "builtin_fmax_fmin"
+ "builtin_fmaxf16_fminf16"
"builtin_round"
"builtin_roundeven"
"float16"
@@ -82,6 +84,10 @@ 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(${feature} STREQUAL "builtin_fmaxf16_fminf16")
+ set(LIBC_COMPILER_HAS_BUILTIN_FMAXF16_FMINF16 TRUE)
elseif(${feature} STREQUAL "builtin_round")
set(LIBC_COMPILER_HAS_BUILTIN_ROUND TRUE)
elseif(${feature} STREQUAL "builtin_roundeven")
diff --git a/libc/cmake/modules/LLVMLibCCheckCpuFeatures.cmake b/libc/cmake/modules/LLVMLibCCheckCpuFeatures.cmake
index 73b249374a066..be569f6f9cabf 100644
--- a/libc/cmake/modules/LLVMLibCCheckCpuFeatures.cmake
+++ b/libc/cmake/modules/LLVMLibCCheckCpuFeatures.cmake
@@ -9,6 +9,7 @@ if(${LIBC_TARGET_ARCHITECTURE_IS_X86})
set(ALL_CPU_FEATURES SSE2 SSE4_2 AVX AVX2 AVX512F AVX512BW FMA)
set(LIBC_COMPILE_OPTIONS_NATIVE -march=native)
elseif(${LIBC_TARGET_ARCHITECTURE_IS_AARCH64})
+ set(ALL_CPU_FEATURES "FullFP16")
set(LIBC_COMPILE_OPTIONS_NATIVE -mcpu=native)
endif()
diff --git a/libc/cmake/modules/LLVMLibCCompileOptionRules.cmake b/libc/cmake/modules/LLVMLibCCompileOptionRules.cmake
index 97d1c7262d24d..0f1ef6a575277 100644
--- a/libc/cmake/modules/LLVMLibCCompileOptionRules.cmake
+++ b/libc/cmake/modules/LLVMLibCCompileOptionRules.cmake
@@ -40,6 +40,17 @@ function(_get_compile_options_from_flags output_var)
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()
+ if(LIBC_COMPILER_HAS_BUILTIN_FMAXF16_FMINF16)
+ list(APPEND compile_options "-D__LIBC_USE_BUILTIN_FMAXF16_FMINF16")
+ endif()
+ if("FullFP16" IN_LIST LIBC_CPU_FEATURES AND
+ CMAKE_CXX_COMPILER_ID STREQUAL "Clang")
+ list(APPEND compile_options
+ "SHELL:-Xclang -target-feature -Xclang +fullfp16")
+ endif()
endif()
elseif(MSVC)
if(ADD_FMA_FLAG)
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..594d839f174bc
--- /dev/null
+++ b/libc/cmake/modules/compiler_features/check_builtin_fmax_fmin.cpp
@@ -0,0 +1,7 @@
+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_fmaxf(double x, double y) { return __builtin_fmax(x, y); }
+double try_builtin_fminf(double x, double y) { return __builtin_fmin(x, y); }
+
+extern "C" void _start() {}
diff --git a/libc/cmake/modules/compiler_features/check_builtin_fmaxf16_fminf16.cpp b/libc/cmake/modules/compiler_features/check_builtin_fmaxf16_fminf16.cpp
new file mode 100644
index 0000000000000..4ce377782800c
--- /dev/null
+++ b/libc/cmake/modules/compiler_features/check_builtin_fmaxf16_fminf16.cpp
@@ -0,0 +1,9 @@
+_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);
+}
+
+extern "C" void _start() {}
diff --git a/libc/cmake/modules/cpu_features/check_FullFP16.cpp b/libc/cmake/modules/cpu_features/check_FullFP16.cpp
new file mode 100644
index 0000000000000..b757fccc86e4c
--- /dev/null
+++ b/libc/cmake/modules/cpu_features/check_FullFP16.cpp
@@ -0,0 +1,5 @@
+#include "src/__support/macros/properties/cpu_features.h"
+
+#ifndef LIBC_TARGET_CPU_HAS_FULLFP16
+#error unsupported
+#endif
diff --git a/libc/src/__support/FPUtil/BasicOperations.h b/libc/src/__support/FPUtil/BasicOperations.h
index 3b7d7a5c249ae..83c319e151c61 100644
--- a/libc/src/__support/FPUtil/BasicOperations.h
+++ b/libc/src/__support/FPUtil/BasicOperations.h
@@ -17,6 +17,8 @@
#include "src/__support/common.h"
#include "src/__support/macros/config.h"
#include "src/__support/macros/optimization.h" // LIBC_UNLIKELY
+#include "src/__support/macros/properties/architectures.h"
+#include "src/__support/macros/properties/types.h"
#include "src/__support/uint128.h"
namespace LIBC_NAMESPACE_DECL {
@@ -27,6 +29,90 @@ LIBC_INLINE T abs(T x) {
return FPBits<T>(x).abs().get_val();
}
+namespace internal {
+
+template <typename T>
+LIBC_INLINE cpp::enable_if_t<cpp::is_floating_point_v<T>, T> max(T x, T y) {
+ FPBits<T> x_bits(x);
+ FPBits<T> y_bits(y);
+
+ // To make sure that fmax(+0, -0) == +0 == fmax(-0, +0), whenever x and y
+ // have different signs and both are not NaNs, we return the number with
+ // positive sign.
+ if (x_bits.sign() != y_bits.sign())
+ return x_bits.is_pos() ? x : y;
+ return x > y ? x : y;
+}
+
+#if defined(__LIBC_USE_BUILTIN_FMAXF16_FMINF16)
+template <> LIBC_INLINE float16 max(float16 x, float16 y) {
+ return __builtin_fmaxf16(x, y);
+}
+#elif !defined(LIBC_TARGET_ARCH_IS_AARCH64)
+template <> LIBC_INLINE float16 max(float16 x, float16 y) {
+ FPBits<float16> x_bits(x);
+ FPBits<float16> y_bits(y);
+
+ // If x and y have different signs, then the one with the lesser uintval is
+ // positive (has the sign bit set to 0), and therefore the maximum.
+ if (x_bits.sign() != y_bits.sign())
+ return x_bits.uintval() < y_bits.uintval() ? x : y;
+ return ((x_bits.uintval() > y_bits.uintval()) == x_bits.is_pos()) ? x : y;
+}
+#endif
+
+#if defined(__LIBC_USE_BUILTIN_FMAX_FMIN) && !defined(LIBC_TARGET_ARCH_IS_X86)
+template <> LIBC_INLINE float max(float x, float y) {
+ return __builtin_fmaxf(x, y);
+}
+
+template <> LIBC_INLINE double max(double x, double y) {
+ return __builtin_fmax(x, y);
+}
+#endif
+
+template <typename T>
+LIBC_INLINE cpp::enable_if_t<cpp::is_floating_point_v<T>, T> min(T x, T y) {
+ FPBits<T> x_bits(x);
+ FPBits<T> y_bits(y);
+
+ // To make sure that fmin(+0, -0) == -0 == fmin(-0, +0), whenever x and y have
+ // different signs and both are not NaNs, we return the number with negative
+ // sign.
+ if (x_bits.sign() != y_bits.sign())
+ return x_bits.is_neg() ? x : y;
+ return x < y ? x : y;
+}
+
+#if defined(__LIBC_USE_BUILTIN_FMAXF16_FMINF16)
+template <> LIBC_INLINE float16 min(float16 x, float16 y) {
+ return __builtin_fminf16(x, y);
+}
+#elif !defined(LIBC_TARGET_ARCH_IS_AARCH64)
+template <> LIBC_INLINE float16 min(float16 x, float16 y) {
+ FPBits<float16> x_bits(x);
+ FPBits<float16> y_bits(y);
+
+ // If x and y have different signs, then the one with the greater uintval is
+ // negative (has the sign bit set to 1), and therefore the minimum.
+ if (x_bits.sign() != y_bits.sign())
+ return x_bits.uintval() > y_bits.uintval() ? x : y;
+ return ((x_bits.uintval() < y_bits.uintval()) == x_bits.is_pos()) ? x : y;
+}
+#endif
+
+#if defined(__LIBC_USE_BUILTIN_FMAX_FMIN) && !defined(LIBC_TARGET_ARCH_IS_X86)
+template <> LIBC_INLINE float min(float x, float y) {
+ return __builtin_fminf(x, y);
+}
+
+template <> LIBC_INLINE double min(double x, double y) {
+ return __builtin_fmin(x, y);
+}
+#endif
+
+} // namespace internal
+
template <typename T, cpp::enable_if_t<cpp::is_floating_point_v<T>, int> = 0>
LIBC_INLINE T fmin(T x, T y) {
const FPBits<T> bitx(x), bity(y);
@@ -35,12 +121,7 @@ LIBC_INLINE T fmin(T x, T y) {
return y;
if (bity.is_nan())
return x;
- if (bitx.sign() != bity.sign())
- // To make sure that fmin(+0, -0) == -0 == fmin(-0, +0), whenever x and
- // y has different signs and both are not NaNs, we return the number
- // with negative sign.
- return bitx.is_neg() ? x : y;
- return x < y ? x : y;
+ return internal::min(x, y);
}
template <typename T, cpp::enable_if_t<cpp::is_floating_point_v<T>, int> = 0>
@@ -51,12 +132,7 @@ LIBC_INLINE T fmax(T x, T y) {
return y;
if (bity.is_nan())
return x;
- if (bitx.sign() != bity.sign())
- // To make sure that fmax(+0, -0) == +0 == fmax(-0, +0), whenever x and
- // y has different signs and both are not NaNs, we return the number
- // with positive sign.
- return bitx.is_neg() ? y : x;
- return x > y ? x : y;
+ return internal::max(x, y);
}
template <typename T, cpp::enable_if_t<cpp::is_floating_point_v<T>, int> = 0>
@@ -67,9 +143,7 @@ LIBC_INLINE T fmaximum(T x, T y) {
return x;
if (bity.is_nan())
return y;
- if (bitx.sign() != bity.sign())
- return (bitx.is_neg() ? y : x);
- return x > y ? x : y;
+ return internal::max(x, y);
}
template <typename T, cpp::enable_if_t<cpp::is_floating_point_v<T>, int> = 0>
@@ -80,9 +154,7 @@ LIBC_INLINE T fminimum(T x, T y) {
return x;
if (bity.is_nan())
return y;
- if (bitx.sign() != bity.sign())
- return (bitx.is_neg()) ? x : y;
- return x < y ? x : y;
+ return internal::min(x, y);
}
template <typename T, cpp::enable_if_t<cpp::is_floating_point_v<T>, int> = 0>
@@ -97,9 +169,7 @@ LIBC_INLINE T fmaximum_num(T x, T y) {
return y;
if (bity.is_nan())
return x;
- if (bitx.sign() != bity.sign())
- return (bitx.is_neg() ? y : x);
- return x > y ? x : y;
+ return internal::max(x, y);
}
template <typename T, cpp::enable_if_t<cpp::is_floating_point_v<T>, int> = 0>
@@ -114,9 +184,7 @@ LIBC_INLINE T fminimum_num(T x, T y) {
return y;
if (bity.is_nan())
return x;
- if (bitx.sign() != bity.sign())
- return (bitx.is_neg() ? x : y);
- return x < y ? x : y;
+ return internal::min(x, y);
}
template <typename T, cpp::enable_if_t<cpp::is_floating_point_v<T>, int> = 0>
diff --git a/libc/src/__support/FPUtil/CMakeLists.txt b/libc/src/__support/FPUtil/CMakeLists.txt
index 8804f3a4d5e23..bfdfffb85c255 100644
--- a/libc/src/__support/FPUtil/CMakeLists.txt
+++ b/libc/src/__support/FPUtil/CMakeLists.txt
@@ -192,6 +192,8 @@ add_header_library(
libc.src.__support.uint128
libc.src.__support.common
libc.src.__support.macros.optimization
+ libc.src.__support.macros.properties.architectures
+ libc.src.__support.macros.properties.types
)
add_header_library(
diff --git a/libc/src/__support/macros/properties/cpu_features.h b/libc/src/__support/macros/properties/cpu_features.h
index 80d48be702070..8d431f203335f 100644
--- a/libc/src/__support/macros/properties/cpu_features.h
+++ b/libc/src/__support/macros/properties/cpu_features.h
@@ -14,6 +14,10 @@
#include "architectures.h"
+#if defined(__ARM_FEATURE_FP16_SCALAR_ARITHMETIC)
+#define LIBC_TARGET_CPU_HAS_FULLFP16
+#endif
+
#if defined(__SSE2__)
#define LIBC_TARGET_CPU_HAS_SSE2
#endif
diff --git a/libc/src/math/generic/CMakeLists.txt b/libc/src/math/generic/CMakeLists.txt
index 74360edff3f9a..ef96d6c4c2f37 100644
--- a/libc/src/math/generic/CMakeLists.txt
+++ b/libc/src/math/generic/CMakeLists.txt
@@ -2065,7 +2065,9 @@ add_entrypoint_object(
DEPENDS
libc.src.__support.FPUtil.basic_operations
COMPILE_OPTIONS
- -O2
+ -O3
+ FLAGS
+ MISC_MATH_BASIC_OPS_OPT
)
add_entrypoint_object(
@@ -2077,7 +2079,9 @@ add_entrypoint_object(
DEPENDS
libc.src.__support.FPUtil.basic_operations
COMPILE_OPTIONS
- -O2
+ -O3
+ FLAGS
+ MISC_MATH_BASIC_OPS_OPT
)
add_entrypoint_object(
@@ -2089,7 +2093,7 @@ add_entrypoint_object(
DEPENDS
libc.src.__support.FPUtil.basic_operations
COMPILE_OPTIONS
- -O2
+ -O3
)
add_entrypoint_object(
@@ -2116,6 +2120,8 @@ add_entrypoint_object(
libc.src.__support.FPUtil.basic_operations
COMPILE_OPTIONS
-O3
+ FLAGS
+ MISC_MATH_BASIC_OPS_OPT
)
@@ -2128,7 +2134,9 @@ add_entrypoint_object(
DEPENDS
libc.src.__support.FPUtil.basic_operations
COMPILE_OPTIONS
- -O2
+ -O3
+ FLAGS
+ MISC_MATH_BASIC_OPS_OPT
)
add_entrypoint_object(
@@ -2140,7 +2148,9 @@ add_entrypoint_object(
DEPENDS
libc.src.__support.FPUtil.basic_operations
COMPILE_OPTIONS
- -O2
+ -O3
+ FLAGS
+ MISC_MATH_BASIC_OPS_OPT
)
add_entrypoint_object(
@@ -2152,7 +2162,7 @@ add_entrypoint_object(
DEPENDS
libc.src.__support.FPUtil.basic_operations
COMPILE_OPTIONS
- -O2
+ -O3
)
add_entrypoint_object(
@@ -2179,6 +2189,8 @@ add_entrypoint_object(
libc.src.__support.FPUtil.basic_operations
COMPILE_OPTIONS
-O3
+ FLAGS
+ MISC_MATH_BASIC_OPS_OPT
)
add_entrypoint_object(
@@ -2190,7 +2202,9 @@ add_entrypoint_object(
DEPENDS
libc.src.__support.FPUtil.basic_operations
COMPILE_OPTIONS
- -O2
+ -O3
+ FLAGS
+ MISC_MATH_BASIC_OPS_OPT
)
add_entrypoint_object(
@@ -2202,7 +2216,9 @@ add_entrypoint_object(
DEPENDS
libc.src.__support.FPUtil.basic_operations
COMPILE_OPTIONS
- -O2
+ -O3
+ FLAGS
+ MISC_MATH_BASIC_OPS_OPT
)
add_entrypoint_object(
@@ -2214,7 +2230,7 @@ add_entrypoint_object(
DEPENDS
libc.src.__support.FPUtil.basic_operations
COMPILE_OPTIONS
- -O2
+ -O3
)
add_entrypoint_object(
@@ -2228,6 +2244,8 @@ add_entrypoint_object(
libc.src.__support.FPUtil.basic_operations
COMPILE_OPTIONS
-O3
+ FLAGS
+ MISC_MATH_BASIC_OPS_OPT
)
add_entrypoint_object(
@@ -2252,7 +2270,9 @@ add_entrypoint_object(
DEPENDS
libc.src.__support.FPUtil.basic_operations
COMPILE_OPTIONS
- -O2
+ -O3
+ FLAGS
+ MISC_MATH_BASIC_OPS_OPT
)
add_entrypoint_object(
@@ -2264,7 +2284,9 @@ add_entrypoint_object(
DEPENDS
libc.src.__support.FPUtil.basic_operations
COMPILE_OPTIONS
- -O2
+ -O3
+ FLAGS
+ MISC_MATH_BASIC_OPS_OPT
)
add_entrypoint_object(
@@ -2276,7 +2298,7 @@ add_entrypoint_object(
DEPENDS
libc.src.__support.FPUtil.basic_operations
COMPILE_OPTIONS
- -O2
+ -O3
)
add_entrypoint_object(
@@ -2290,6 +2312,8 @@ add_entrypoint_object(
libc.src.__support.FPUtil.basic_operations
COMPILE_OPTIONS
-O3
+ FLAGS
+ MISC_MATH_BASIC_OPS_OPT
)
add_entrypoint_object(
@@ -2438,7 +2462,9 @@ add_entrypoint_object(
DEPENDS
libc.src.__support.FPUtil.basic_operations
COMPILE_OPTIONS
- -O2
+ -O3
+ FLAGS
+ MISC_MATH_BASIC_OPS_OPT
)
add_entrypoint_object(
@@ -2450,7 +2476,9 @@ add_entrypoint_object(
DEPENDS
libc.src.__support.FPUtil.basic_operations
COMPILE_OPTIONS
- -O2
+ -O3
+ FLAGS
+ MISC_MATH_BASIC_OPS_OPT
)
add_entrypoint_object(
@@ -2462,7 +2490,7 @@ add_entrypoint_object(
DEPENDS
libc.src.__support.FPUtil.basic_operations
COMPILE_OPTIONS
- -O2
+ -O3
)
add_entrypoint_object(
@@ -2476,6 +2504,8 @@ add_entrypoint_object(
libc.src.__support.FPUtil.basic_operations
COMPILE_OPTIONS
-O3
+ FLAGS
+ MISC_MATH_BASIC_OPS_OPT
)
add_entrypoint_object(
@@ -2500,7 +2530,9 @@ add_entrypoint_object(
DEPENDS
libc.src.__support.FPUtil.basic_operations
COMPILE_OPTIONS
- -O2
+ -O3
+ FLAGS
+ MISC_MATH_BASIC_OPS_OPT
)
add_entrypoint_object(
@@ -2512,7 +2544,9 @@ add_entrypoint_object(
DEPENDS
libc.src.__support.FPUtil.basic_operations
COMPILE_OPTIONS
- -O2
+ -O3
+ FLAGS
+ MISC_MATH_BASIC_OPS_OPT
)
add_entrypoint_object(
@@ -2524,7 +2558,7 @@ add_entrypoint_object(
DEPENDS
libc.src.__support.FPUtil.basic_operations
COMPILE_OPTIONS
- -O2
+ -O3
)
add_entrypoint_object(
@@ -2538,6 +2572,8 @@ 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/test/src/math/performance_testing/CMakeLists.txt b/libc/test/src/math/performance_testing/CMakeLists.txt
index 72e1a730ac756..536338bf2a56a 100644
--- a/libc/test/src/math/performance_testing/CMakeLists.txt
+++ b/libc/test/src/math/performance_testing/CMakeLists.txt
@@ -421,3 +421,25 @@ add_perf_binary(
COMPILE_OPTIONS
-fno-builtin
)
+
+add_perf_binary(
+ max_min_funcs_perf
+ SRCS
+ max_min_funcs_perf.cpp
+ DEPENDS
+ .binary_op_single_output_diff
+ libc.src.math.fmaxf
+ libc.src.math.fmaxf16
+ libc.src.math.fmaximumf
+ libc.src.math.fmaximumf16
+ libc.src.math.fmaximum_numf
+ libc.src.math.fmaximum_numf16
+ libc.src.math.fminf
+ libc.src.math.fminf16
+ libc.src.math.fminimumf
+ libc.src.math.fminimumf16
+ libc.src.math.fminimum_numf
+ libc.src.math.fminimum_numf16
+ COMPILE_OPTIONS
+ -fno-builtin
+)
diff --git a/libc/test/src/math/performance_testing/max_min_funcs_perf.cpp b/libc/test/src/math/performance_testing/max_min_funcs_perf.cpp
new file mode 100644
index 0000000000000..9540112e69ea6
--- /dev/null
+++ b/libc/test/src/math/performance_testing/max_min_funcs_perf.cpp
@@ -0,0 +1,75 @@
+//===-- Performance test for maximum and minimum functions ----------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+#include "BinaryOpSingleOutputPerf.h"
+#include "src/math/fmaxf.h"
+#include "src/math/fmaxf16.h"
+#include "src/math/fmaximum_numf.h"
+#include "src/math/fmaximum_numf16.h"
+#include "src/math/fmaximumf.h"
+#include "src/math/fmaximumf16.h"
+#include "src/math/fminf.h"
+#include "src/math/fminf16.h"
+#include "src/math/fminimum_numf.h"
+#include "src/math/fminimum_numf16.h"
+#include "src/math/fminimumf.h"
+#include "src/math/fminimumf16.h"
+
+#include <math.h>
+
+static constexpr size_t FLOAT16_ROUNDS = 20'000;
+static constexpr size_t FLOAT_ROUNDS = 40;
+
+// LLVM libc might be the only libc implementation with support for float16 math
+// functions currently. We can't compare our float16 functions against the
+// system libc, so we compare them against this placeholder function.
+float16 placeholder_binaryf16(float16 x, float16 y) { return x; }
+
+// The system libc might not provide the fmaximum* and fminimum* C23 math
+// functions either.
+float placeholder_binaryf(float x, float y) { return x; }
+
+int main() {
+ BINARY_OP_SINGLE_OUTPUT_PERF_EX(float16, LIBC_NAMESPACE::fmaxf16,
+ placeholder_binaryf16, FLOAT16_ROUNDS,
+ "fmaxf16_perf.log")
+ BINARY_OP_SINGLE_OUTPUT_PERF_EX(float16, LIBC_NAMESPACE::fminf16,
+ placeholder_binaryf16, FLOAT16_ROUNDS,
+ "fminf16_perf.log")
+ BINARY_OP_SINGLE_OUTPUT_PERF_EX(float16, LIBC_NAMESPACE::fmaximumf16,
+ placeholder_binaryf16, FLOAT16_ROUNDS,
+ "fmaximumf16_perf.log")
+ BINARY_OP_SINGLE_OUTPUT_PERF_EX(float16, LIBC_NAMESPACE::fminimumf16,
+ placeholder_binaryf16, FLOAT16_ROUNDS,
+ "fminimumf16_perf.log")
+ BINARY_OP_SINGLE_OUTPUT_PERF_EX(float16, LIBC_NAMESPACE::fmaximum_numf16,
+ placeholder_binaryf16, FLOAT16_ROUNDS,
+ "fmaximum_numf16_perf.log")
+ BINARY_OP_SINGLE_OUTPUT_PERF_EX(float16, LIBC_NAMESPACE::fminimum_numf16,
+ placeholder_binaryf16, FLOAT16_ROUNDS,
+ "fminimum_numf16_perf.log")
+
+ BINARY_OP_SINGLE_OUTPUT_PERF_EX(float, LIBC_NAMESPACE::fmaxf, ::fmaxf,
+ FLOAT_ROUNDS, "fmaxf_perf.log")
+ BINARY_OP_SINGLE_OUTPUT_PERF_EX(float, LIBC_NAMESPACE::fminf, ::fminf,
+ FLOAT_ROUNDS, "fminf_perf.log")
+ BINARY_OP_SINGLE_OUTPUT_PERF_EX(float, LIBC_NAMESPACE::fmaximumf,
+ placeholder_binaryf, FLOAT_ROUNDS,
+ "fmaximumf_perf.log")
+ BINARY_OP_SINGLE_OUTPUT_PERF_EX(float, LIBC_NAMESPACE::fminimumf,
+ placeholder_binaryf, FLOAT_ROUNDS,
+ "fminimumf_perf.log")
+ BINARY_OP_SINGLE_OUTPUT_PERF_EX(float, LIBC_NAMESPACE::fmaximum_numf,
+ placeholder_binaryf, FLOAT_ROUNDS,
+ "fmaximum_numf_perf.log")
+ BINARY_OP_SINGLE_OUTPUT_PERF_EX(float, LIBC_NAMESPACE::fminimum_numf,
+ placeholder_binaryf, FLOAT_ROUNDS,
+ "fminimum_numf_perf.log")
+
+ return 0;
+}
More information about the libc-commits
mailing list