Runtime Performance Optimization

CMakeLists.txt

@@ -18,7 +18,7 @@
 ##
 ################################################################################
 
-cmake_minimum_required(VERSION 3.1)
+cmake_minimum_required(VERSION 3.10)
 project(gcem)
 
 set(GCEM_INCLUDE_DIR ${PROJECT_SOURCE_DIR}/include)

include/gcem.hpp

@@ -70,9 +70,10 @@ namespace gcem
     #include "gcem_incl/gcd.hpp"
     #include "gcem_incl/lcm.hpp"
 
-    #include "gcem_incl/tan.hpp"
     #include "gcem_incl/cos.hpp"
     #include "gcem_incl/sin.hpp"
+    #include "gcem_incl/tan.hpp"
+
 
     #include "gcem_incl/atan.hpp"
     #include "gcem_incl/atan2.hpp"

include/gcem_incl/cos.hpp

@@ -19,7 +19,7 @@
   ################################################################################*/
 
 /*
- * compile-time cosine function using tan(x/2)
+ * compile-time cosine function using 30th degree Chebyshev approximation
  */
 
 #ifndef _gcem_cos_HPP
@@ -37,6 +37,24 @@ noexcept
     return( T(1) - x*x)/(T(1) + x*x );
 }
 
+template<typename T>
+constexpr
+T
+cos_compute_chebyshev_inner(const T x2)
+noexcept {
+    // Chebyshev approximation of sin(x) with degree 30
+    // weights are from: https://publik-void.github.io/sin-cos-approximations/#_cos_rel_error_minimized_degree_30
+    return 1. + x2*(-0.499999999999999999999999999999999997 + x2*(0.0416666666666666666666666666666665864 + x2*(-0.00138888888888888888888888888888792216 + x2*(0.0000248015873015873015873015872954247692 + x2*(-2.75573192239858906525573168323374212e-7 + x2*(2.08767569878680989792094774070130099e-9 + x2*(-1.147074559772972471374259582051505e-11 + x2*(4.77947733238738528351115798998742201e-14 + x2*(-1.56192069685862134697357839719554288e-16 + x2*(4.11031762331127151219297331433042247e-19 + x2*(-8.8967913919984472368244650875181655e-22 + x2*(1.61173755446230144849643763129805917e-24 + x2*(-2.47959193649582299714152529144837331e-27 + x2*(3.27913485904815466449434894827885198e-30 - 3.69081529290172836989264646507263976e-33*x2))))))))))))));
+}
+
+template<typename T>
+constexpr
+T
+cos_compute_chebyshev(const T x1)
+noexcept {
+    return cos_compute_chebyshev_inner(x1 * x1);
+}
+
 template<typename T>
 constexpr
 T
@@ -49,17 +67,7 @@ noexcept
             // indistinguishable from 0
             GCLIM<T>::min() > abs(x) ? 
                 T(1) :
-            // special cases: pi/2 and pi
-            GCLIM<T>::min() > abs(x - T(GCEM_HALF_PI)) ? \
-                T(0) :
-            GCLIM<T>::min() > abs(x + T(GCEM_HALF_PI)) ? \
-                T(0) :
-            GCLIM<T>::min() > abs(x - T(GCEM_PI)) ? \
-                - T(1) :
-            GCLIM<T>::min() > abs(x + T(GCEM_PI)) ? \
-                - T(1) :
-            // else
-                cos_compute( tan(x/T(2)) ) );
+                cos_compute_chebyshev(x - T(GCEM_2PI) * floor(x / T(GCEM_2PI))));
 }
 
 }

include/gcem_incl/gcem_options.hpp

@@ -98,6 +98,10 @@ namespace gcem
     #define GCEM_PI 3.1415926535897932384626433832795028841972L
 #endif
 
+#ifndef GCEM_2PI
+    #define GCEM_2PI 6.2831853071795864769252867665590057683943L
+#endif
+
 #ifndef GCEM_LOG_PI
     #define GCEM_LOG_PI 1.1447298858494001741434273513530587116473L
 #endif

include/gcem_incl/sin.hpp

@@ -19,9 +19,7 @@
   ################################################################################*/
 
 /*
- * compile-time sine function using tan(x/2)
- * 
- * see eq. 5.4.8 in Numerical Recipes
+ * compile-time sine function using 29th degree Chebyshev polynomial approximation
  */
 
 #ifndef _gcem_sin_HPP
@@ -39,6 +37,24 @@ noexcept
     return T(2)*x/(T(1) + x*x);
 }
 
+template<typename T>
+constexpr
+T
+sin_compute_chebyshev_inner(const T x1, const T x2)
+noexcept {
+    // Chebyshev approximation of sin(x) with degree 29
+    // weights are from: https://publik-void.github.io/sin-cos-approximations/#_sin_rel_error_minimized_degree_29
+    return -x1*(1. + x2*(-0.166666666666666666666666666666666629 + x2*(0.00833333333333333333333333333333222318 + x2*(-0.000198412698412698412698412698399723487 + x2*(2.75573192239858906525573184281026561e-6 + x2*(-2.50521083854417187750518138734779333e-8 + x2*(1.60590438368216145993211483359730933e-10 + x2*(-7.64716373181981646407639862566185217e-13 + x2*(2.81145725434551937513944561966331531e-15 + x2*(-8.22063524662315930664562316486611394e-18 + x2*(1.9572941062679701610307767721364446e-20 + x2*(-3.8681701397118386284141935821892017e-23 + x2*(6.44694091890808795934274904084719604e-26 + x2*(-9.18181103282828321298337267365314117e-29 + 1.10854165613066936680763885008011392e-31*x2))))))))))))));
+}
+
+template<typename T>
+constexpr
+T
+sin_compute_chebyshev(const T x1)
+noexcept {
+    return sin_compute_chebyshev_inner(x1, x1 * x1);
+}
+
 template<typename T>
 constexpr
 T
@@ -51,17 +67,7 @@ noexcept
             // indistinguishable from zero
             GCLIM<T>::min() > abs(x) ? \
                 T(0) :
-            // special cases: pi/2 and pi
-            GCLIM<T>::min() > abs(x - T(GCEM_HALF_PI)) ? \
-                T(1) :
-            GCLIM<T>::min() > abs(x + T(GCEM_HALF_PI)) ? \
-                - T(1) :
-            GCLIM<T>::min() > abs(x - T(GCEM_PI)) ? \
-                T(0) :
-            GCLIM<T>::min() > abs(x + T(GCEM_PI)) ? \
-                - T(0) :
-            // else
-                sin_compute( tan(x/T(2)) ) );
+                sin_compute_chebyshev(x - T(GCEM_2PI) * floor(x / T(GCEM_2PI)) - T(GCEM_PI)));
 }
 
 }

include/gcem_incl/sqrt.hpp

@@ -64,6 +64,31 @@ noexcept
                 m_val * sqrt_recur(x, x / T(2), 0) );
 }
 
+// template<typename T>
+// T
+// sqrt_babylonian(const T x)
+// noexcept
+// {
+//     // see here for more details: https://blogs.sas.com/content/iml/2016/05/16/babylonian-square-roots.html
+//     T xn = x;
+//     T xn1 = (xn + x/xn) / T(2);
+//     while(abs(xn1 - xn) > GCLIM<T>::epsilon()) {
+//         xn = xn1;
+//         xn1 = (xn + x/xn) / T(2);
+//     }
+//     return xn1;
+// }
+
+template<typename T>
+T
+sqrt_babylonian(const T x, const T xn, const T xn1)
+noexcept
+{
+    return ( abs(xn1 - xn) < GCLIM<T>::epsilon() ? \
+                xn1 :
+                sqrt_babylonian(x, xn1, (xn1 + x/xn1) / T(2)));
+}
+
 template<typename T>
 constexpr
 T
@@ -84,7 +109,8 @@ noexcept
             GCLIM<T>::min() > abs(T(1) - x) ? \
                 x :
             // else
-            sqrt_simplify(x, T(1)) );
+            // sqrt_simplify(x, T(1)) );
+            sqrt_babylonian(x, x, (x + T(1)) / T(2)) );
 }
 
 }

include/gcem_incl/tan.hpp

@@ -97,6 +97,32 @@ noexcept
                 tan_cf_main(x) );
 }
 
+template<typename T>
+constexpr
+T
+tan_from_sin_cos_inner(const T x1)
+noexcept
+{
+    // There should be faster ways to calculate tan(x) directly
+    // At least as fast as calculating sin(x) and cos(x) using Chebyshev polynomials
+    return (
+        x1 < T(GCEM_HALF_PI) ? sin(x1) / cos(x1) :
+        x1 > T(GCEM_HALF_PI) ? - sin(T(GCEM_PI) - x1) / cos(T(GCEM_PI) - x1) :
+        GCLIM<T>::quiet_NaN()   // x1 == 1/2 pi
+    );
+}
+
+
+template<typename T>
+constexpr
+T
+tan_from_sin_cos(const T x)
+noexcept
+{
+    // limit x to range(0, pi)
+    return tan_from_sin_cos_inner(x - T(GCEM_PI) * floor(x / T(GCEM_PI)));
+}
+
 template<typename T>
 constexpr
 T
@@ -110,9 +136,10 @@ noexcept
             GCLIM<T>::min() > abs(x) ? \
                 T(0) :
             // else
-                x < T(0) ? \
-                    - tan_begin(-x) : 
-                      tan_begin( x) );
+                tan_from_sin_cos(x));
+                // x < T(0) ? \
+                //     - tan_begin(-x) :
+                //       tan_begin( x) );
 }
 
 }

tests/cos.cpp

@@ -37,7 +37,10 @@ int main()
     GCEM_TEST_COMPARE_VALS(gcem::cos,std::cos,11.1L);
     GCEM_TEST_COMPARE_VALS(gcem::cos,std::cos,50.0L);
     GCEM_TEST_COMPARE_VALS(gcem::cos,std::cos,150.0L);
-
+    GCEM_TEST_COMPARE_VALS(gcem::cos,std::cos,GCEM_PI);
+    GCEM_TEST_COMPARE_VALS(gcem::cos,std::cos,GCEM_HALF_PI);
+    GCEM_TEST_COMPARE_VALS(gcem::cos,std::cos,-GCEM_PI);
+    GCEM_TEST_COMPARE_VALS(gcem::cos,std::cos,-GCEM_HALF_PI);
     GCEM_TEST_COMPARE_VALS(gcem::cos,std::cos,TEST_NAN);
 
     //

tests/sin.cpp

@@ -37,7 +37,10 @@ int main()
     GCEM_TEST_COMPARE_VALS(gcem::sin,std::sin,11.1L);
     GCEM_TEST_COMPARE_VALS(gcem::sin,std::sin,50.0L);
     GCEM_TEST_COMPARE_VALS(gcem::sin,std::sin,150.0L);
-
+    GCEM_TEST_COMPARE_VALS(gcem::sin,std::sin,GCEM_PI);
+    GCEM_TEST_COMPARE_VALS(gcem::sin,std::sin,GCEM_HALF_PI);
+    GCEM_TEST_COMPARE_VALS(gcem::sin,std::sin,-GCEM_PI);
+    GCEM_TEST_COMPARE_VALS(gcem::sin,std::sin,-GCEM_HALF_PI);
     GCEM_TEST_COMPARE_VALS(gcem::sin,std::sin,TEST_NAN);
 
     //

tests/tan.cpp

@@ -31,6 +31,7 @@ int main()
 
     // static constexpr long double lrgval = std::numeric_limits<int>::max()*100.0L;
 
+    std::cout << "tan(pi/2) = " << std::tan(GCEM_HALF_PI) << std::endl;
     GCEM_TEST_COMPARE_VALS(gcem::tan,std::tan, 0.0L);
     GCEM_TEST_COMPARE_VALS(gcem::tan,std::tan, 0.001L);
     GCEM_TEST_COMPARE_VALS(gcem::tan,std::tan, 1.001L);
@@ -39,7 +40,10 @@ int main()
     GCEM_TEST_COMPARE_VALS(gcem::tan,std::tan, 50.0L);
     GCEM_TEST_COMPARE_VALS(gcem::tan,std::tan, -1.5L);
     // GCEM_TEST_COMPARE_VALS(gcem::tan,std::tan, lrgval);
-
+    GCEM_TEST_COMPARE_VALS(gcem::tan,std::tan, GCEM_HALF_PI - 1e-6);
+    GCEM_TEST_COMPARE_VALS(gcem::tan,std::tan, GCEM_HALF_PI + 1e-6);
+    GCEM_TEST_COMPARE_VALS(gcem::tan,std::tan, -GCEM_HALF_PI - 1e-6);
+    GCEM_TEST_COMPARE_VALS(gcem::tan,std::tan, -GCEM_HALF_PI + 1e-6);
     GCEM_TEST_COMPARE_VALS(gcem::tan,std::tan, TEST_NAN);
 
     //