More benchmarks work

Chowdhury-DSP · Nov 23, 2023 · efcf6ad · efcf6ad
1 parent f0ca995
commit efcf6ad
Show file tree

Hide file tree

Showing 5 changed files with 191 additions and 23 deletions.
diff --git a/include/math_approx/src/log_approx.hpp b/include/math_approx/src/log_approx.hpp
@@ -56,11 +56,11 @@ namespace log_detail
 template <typename Base, int order>
 float log (float x)
 {
-    const auto vi = reinterpret_cast<int32_t&> (x); // NOSONAR
+    const auto vi = reinterpret_cast<int32_t&> (x);
     const auto ex = vi & 0x7f800000;
     const auto e = (ex >> 23) - 127;
     const auto vfi = (vi - ex) | 0x3f800000;
-    const auto vf = reinterpret_cast<const float&> (vfi); // NOSONAR
+    const auto vf = reinterpret_cast<const float&> (vfi);
 
     static constexpr auto log2_base_r = 1.0f / Base::log2_base;
     return log2_base_r * ((float) e + log_detail::log2_approx<float, order> (vf));
@@ -70,16 +70,46 @@ float log (float x)
 template <typename Base, int order>
 double log (double x)
 {
-    const auto vi = reinterpret_cast<int64_t&> (x); // NOSONAR
+    const auto vi = reinterpret_cast<int64_t&> (x);
     const auto ex = vi & 0x7ff0000000000000;
     const auto e = (ex >> 52) - 1023;
     const auto vfi = (vi - ex) | 0x3ff0000000000000;
-    const auto vf = reinterpret_cast<const double&> (vfi); // NOSONAR
+    const auto vf = reinterpret_cast<const double&> (vfi);
 
     static constexpr auto log2_base_r = 1.0 / Base::log2_base;
     return log2_base_r * ((double) e + log_detail::log2_approx<double, order> (vf));
 }
 
+#if defined(XSIMD_HPP)
+/** approximation for pow(Base, x) (32-bit SIMD) */
+template <typename Base, int order>
+xsimd::batch<float> log (xsimd::batch<float> x)
+{
+    const auto vi = reinterpret_cast<int32_t&> (x);
+    const auto ex = vi & 0x7f800000;
+    const auto e = (ex >> 23) - 127;
+    const auto vfi = (vi - ex) | 0x3f800000;
+    const auto vf = reinterpret_cast<const float&> (vfi);
+
+    static constexpr auto log2_base_r = 1.0f / Base::log2_base;
+    return log2_base_r * ((float) e + log_detail::log2_approx<xsimd::batch<float>, order> (vf));
+}
+
+/** approximation for pow(Base, x) (64-bit SIMD) */
+template <typename Base, int order>
+xsimd::batch<double> log (xsimd::batch<double> x)
+{
+    const auto vi = reinterpret_cast<int64_t&> (x);
+    const auto ex = vi & 0x7ff0000000000000;
+    const auto e = (ex >> 52) - 1023;
+    const auto vfi = (vi - ex) | 0x3ff0000000000000;
+    const auto vf = reinterpret_cast<const double&> (vfi);
+
+    static constexpr auto log2_base_r = 1.0 / Base::log2_base;
+    return log2_base_r * ((double) e + log_detail::log2_approx<xsimd::batch<double>, order> (vf));
+}
+#endif
+
 template <int order, typename T>
 T log (T x)
 {

diff --git a/include/math_approx/src/pow_approx.hpp b/include/math_approx/src/pow_approx.hpp
@@ -97,6 +97,36 @@ double pow (double x)
     return reinterpret_cast<const double&> (vi) * pow_detail::pow2_approx<double, order> (d);
 }
 
+#if defined(XSIMD_HPP)
+/** approximation for pow(Base, x) (32-bit SIMD) */
+template <typename Base, int order>
+xsimd::batch<float> pow (xsimd::batch<float> x)
+{
+    x = xsimd::max (xsimd::broadcast (-126.0f), Base::log2_base * x);
+
+    const auto xi = xsimd::to_int (x);
+    const auto l = xsimd::select (xsimd::batch_bool_cast<int32_t> (x < 0.0f), xi - 1, xi);
+    const auto f = x - xsimd::to_float (l);
+    const auto vi = (l + 127) << 23;
+
+    return reinterpret_cast<const xsimd::batch<float>&> (vi) * pow_detail::pow2_approx<xsimd::batch<float>, order> (f);
+}
+
+/** approximation for pow(Base, x) (64-bit SIMD) */
+template <typename Base, int order>
+xsimd::batch<double> pow (xsimd::batch<double> x)
+{
+    x = xsimd::max (-1022.0, Base::log2_base * x);
+
+    const auto xi = xsimd::to_int (x);
+    const auto l = xsimd::select (xsimd::batch_bool_cast<int64_t> (x < 0.0), xi - 1, xi);
+    const auto d = x - xsimd::to_float (l);
+    const auto vi = (l + 1023) << 52;
+
+    return reinterpret_cast<const xsimd::batch<double>&> (vi) * pow_detail::pow2_approx<xsimd::batch<double>, order> (d);
+}
+#endif
+
 template <int order, typename T>
 T exp (T x)
 {

diff --git a/tools/bench/CMakeLists.txt b/tools/bench/CMakeLists.txt
@@ -16,3 +16,6 @@ target_link_libraries(trig_approx_bench PRIVATE benchmark::benchmark math_approx
 
 add_executable(pow_approx_bench pow_bench.cpp)
 target_link_libraries(pow_approx_bench PRIVATE benchmark::benchmark math_approx)
+
+add_executable(log_approx_bench log_bench.cpp)
+target_link_libraries(log_approx_bench PRIVATE benchmark::benchmark math_approx)
diff --git a/tools/bench/log_bench.cpp b/tools/bench/log_bench.cpp
@@ -0,0 +1,77 @@
+#include <math_approx/math_approx.hpp>
+#include <benchmark/benchmark.h>
+
+static constexpr size_t N = 2000;
+const auto data = []
+{
+    std::vector<float> x;
+    x.resize (N, 0.0f);
+    for (size_t i = 0; i < N; ++i)
+        x[i] = -10.0f + 20.0f * (float) i / (float) N;
+    return x;
+}();
+
+#define LOG_BENCH(name, func) \
+void name (benchmark::State& state) \
+{ \
+for (auto _ : state) \
+{ \
+for (auto& x : data) \
+{ \
+auto y = func (x); \
+benchmark::DoNotOptimize (y); \
+} \
+} \
+} \
+BENCHMARK (name);
+LOG_BENCH (log_std, std::log)
+LOG_BENCH (log_approx6, math_approx::log<6>)
+LOG_BENCH (log_approx5, math_approx::log<5>)
+LOG_BENCH (log_approx4, math_approx::log<4>)
+LOG_BENCH (log_approx3, math_approx::log<3>)
+
+LOG_BENCH (log2_std, std::log2)
+LOG_BENCH (log2_approx6, math_approx::log2<6>)
+LOG_BENCH (log2_approx5, math_approx::log2<5>)
+LOG_BENCH (log2_approx4, math_approx::log2<4>)
+LOG_BENCH (log2_approx3, math_approx::log2<3>)
+
+LOG_BENCH (log10_std, std::log10)
+LOG_BENCH (log10_approx6, math_approx::log10<6>)
+LOG_BENCH (log10_approx5, math_approx::log10<5>)
+LOG_BENCH (log10_approx4, math_approx::log10<4>)
+LOG_BENCH (log10_approx3, math_approx::log10<3>)
+
+#define LOG_SIMD_BENCH(name, func) \
+void name (benchmark::State& state) \
+{ \
+for (auto _ : state) \
+{ \
+for (auto& x : data) \
+{ \
+auto y = func (xsimd::broadcast (x)); \
+static_assert (std::is_same_v<xsimd::batch<float>, decltype(y)>); \
+benchmark::DoNotOptimize (y); \
+} \
+} \
+} \
+BENCHMARK (name);
+LOG_SIMD_BENCH (log_xsimd, xsimd::log)
+LOG_SIMD_BENCH (log_simd_approx6, math_approx::log<6>)
+LOG_SIMD_BENCH (log_simd_approx5, math_approx::log<5>)
+LOG_SIMD_BENCH (log_simd_approx4, math_approx::log<4>)
+LOG_SIMD_BENCH (log_simd_approx3, math_approx::log<3>)
+
+LOG_SIMD_BENCH (log2_xsimd, xsimd::log2)
+LOG_SIMD_BENCH (log2_simd_approx6, math_approx::log2<6>)
+LOG_SIMD_BENCH (log2_simd_approx5, math_approx::log2<5>)
+LOG_SIMD_BENCH (log2_simd_approx4, math_approx::log2<4>)
+LOG_SIMD_BENCH (log2_simd_approx3, math_approx::log2<3>)
+
+LOG_SIMD_BENCH (log10_xsimd, xsimd::log10)
+LOG_SIMD_BENCH (log10_simd_approx6, math_approx::log10<6>)
+LOG_SIMD_BENCH (log10_simd_approx5, math_approx::log10<5>)
+LOG_SIMD_BENCH (log10_simd_approx4, math_approx::log10<4>)
+LOG_SIMD_BENCH (log10_simd_approx3, math_approx::log10<3>)
+
+BENCHMARK_MAIN();
diff --git a/tools/bench/pow_bench.cpp b/tools/bench/pow_bench.cpp
@@ -30,24 +30,52 @@ POW_BENCH (exp_approx5, math_approx::exp<5>)
 POW_BENCH (exp_approx4, math_approx::exp<4>)
 POW_BENCH (exp_approx3, math_approx::exp<3>)
 
-// #define POW_SIMD_BENCH(name, func) \
-// void name (benchmark::State& state) \
-// { \
-// for (auto _ : state) \
-// { \
-// for (auto& x : data) \
-// { \
-// auto y = func (xsimd::broadcast (x)); \
-// static_assert (std::is_same_v<xsimd::batch<float>, decltype(y)>); \
-// benchmark::DoNotOptimize (y); \
-// } \
-// } \
-// } \
-// BENCHMARK (name);
-// POW_SIMD_BENCH (exp_xsimd, xsimd::exp)
-// POW_SIMD_BENCH (exp_simd_approx6, math_approx::exp<6>)
-// POW_SIMD_BENCH (exp_simd_approx5, math_approx::exp<5>)
-// POW_SIMD_BENCH (exp_simd_approx4, math_approx::exp<4>)
-// POW_SIMD_BENCH (exp_simd_approx3, math_approx::exp<3>)
+POW_BENCH (exp2_std, std::exp2)
+POW_BENCH (exp2_approx6, math_approx::exp2<6>)
+POW_BENCH (exp2_approx5, math_approx::exp2<5>)
+POW_BENCH (exp2_approx4, math_approx::exp2<4>)
+POW_BENCH (exp2_approx3, math_approx::exp2<3>)
+
+float stdpow_exp10 (float x)
+{
+    return std::pow (10.0f, x);
+}
+POW_BENCH (exp10_std, stdpow_exp10)
+POW_BENCH (exp10_approx6, math_approx::exp10<6>)
+POW_BENCH (exp10_approx5, math_approx::exp10<5>)
+POW_BENCH (exp10_approx4, math_approx::exp10<4>)
+POW_BENCH (exp10_approx3, math_approx::exp10<3>)
+
+#define POW_SIMD_BENCH(name, func) \
+void name (benchmark::State& state) \
+{ \
+for (auto _ : state) \
+{ \
+for (auto& x : data) \
+{ \
+auto y = func (xsimd::broadcast (x)); \
+static_assert (std::is_same_v<xsimd::batch<float>, decltype(y)>); \
+benchmark::DoNotOptimize (y); \
+} \
+} \
+} \
+BENCHMARK (name);
+POW_SIMD_BENCH (exp_xsimd, xsimd::exp)
+POW_SIMD_BENCH (exp_simd_approx6, math_approx::exp<6>)
+POW_SIMD_BENCH (exp_simd_approx5, math_approx::exp<5>)
+POW_SIMD_BENCH (exp_simd_approx4, math_approx::exp<4>)
+POW_SIMD_BENCH (exp_simd_approx3, math_approx::exp<3>)
+
+POW_SIMD_BENCH (exp2_xsimd, xsimd::exp2)
+POW_SIMD_BENCH (exp2_simd_approx6, math_approx::exp2<6>)
+POW_SIMD_BENCH (exp2_simd_approx5, math_approx::exp2<5>)
+POW_SIMD_BENCH (exp2_simd_approx4, math_approx::exp2<4>)
+POW_SIMD_BENCH (exp2_simd_approx3, math_approx::exp2<3>)
+
+POW_SIMD_BENCH (exp10_xsimd, xsimd::exp10)
+POW_SIMD_BENCH (exp10_simd_approx6, math_approx::exp10<6>)
+POW_SIMD_BENCH (exp10_simd_approx5, math_approx::exp10<5>)
+POW_SIMD_BENCH (exp10_simd_approx4, math_approx::exp10<4>)
+POW_SIMD_BENCH (exp10_simd_approx3, math_approx::exp10<3>)
 
 BENCHMARK_MAIN();