DataFormats/Math/interface/FastMath.h - master - githubmirror/cmssw

DataFormats/Math/interface/FastMath.h
170 строк4 KB
Cms Build
Clang-Format
11 июн 2019, 11:29
11 июн 2019, 11:29146bc98
#ifndef DataFormats_Math_FastMath_h
#define DataFormats_Math_FastMath_h
// faster function will a limited precision

#include <cmath>
#include <utility>
#ifdef __SSE2__
#include <emmintrin.h>
#endif
namespace fastmath {
  inline float invSqrt(float in) {
#ifndef __SSE2__
    return 1.f / std::sqrt(in);
#else
    float out;
    _mm_store_ss(&out, _mm_rsqrt_ss(_mm_load_ss(&in)));  // compiles to movss, rsqrtss, movss
    // return out; // already good enough!
    return out * (1.5f - 0.5f * in * out * out);  // One (more?) round of Newton's method
#endif
  }

  inline double invSqrt(double in) { return 1. / std::sqrt(in); }

}  // namespace fastmath

namespace fastmath_details {
  const double _2pi = (2.0 * 3.1415926535897932384626434);
  const float _2pif = float(_2pi);
  extern float atanbuf_[257 * 2];
  extern double datanbuf_[513 * 2];
}  // namespace fastmath_details

namespace fastmath {

  // =====================================================================
  // arctan, single-precision; returns phi and r (or 1/r if overR=true)
  // =====================================================================
  inline std::pair<float, float> atan2r(float y_, float x_, bool overR = false) {
    using namespace fastmath_details;
    float mag2 = x_ * x_ + y_ * y_;
    if (!(mag2 > 0)) {
      return std::pair<float, float>(0.f, 0.f);
    }  // degenerate case

    // float r_ = std::sqrt(mag2);
    float rinv = invSqrt(mag2);
    unsigned int flags = 0;
    float x, y;
    union {
      float f;
      int i;
    } yp;
    yp.f = 32768.f;
    if (y_ < 0) {
      flags |= 4;
      y_ = -y_;
    }
    if (x_ < 0) {
      flags |= 2;
      x_ = -x_;
    }
    if (y_ > x_) {
      flags |= 1;
      x = rinv * y_;
      y = rinv * x_;
      yp.f += y;
    } else {
      x = rinv * x_;
      y = rinv * y_;
      yp.f += y;
    }
    int ind = (yp.i & 0x01FF) * 2;

    float* asbuf = (float*)(atanbuf_ + ind);
    float sv = yp.f - 32768.f;
    float cv = asbuf[0];
    float asv = asbuf[1];
    sv = y * cv - x * sv;  // delta sin value
    // ____ compute arcsin directly
    float asvd = 6.f + sv * sv;
    sv *= float(1.0f / 6.0f);
    float th = asv + asvd * sv;
    if (flags & 1) {
      th = (_2pif / 4.f) - th;
    }
    if (flags & 2) {
      th = (_2pif / 2.f) - th;
    }
    if (flags & 4) {
      th = -th;
    }
    return std::pair<float, float>(th, overR ? rinv : rinv * mag2);
  }

  // =====================================================================
  // arctan, double-precision; returns phi and r (or 1/r if overR=true)
  // =====================================================================
  inline std::pair<double, double> atan2r(double y_, double x_, bool overR = false) {
    using namespace fastmath_details;
    // assert(ataninited);
    double mag2 = x_ * x_ + y_ * y_;
    if (!(mag2 > 0)) {
      return std::pair<double, double>(0., 0.);
    }  // degenerate case

    double r_ = std::sqrt(mag2);
    double rinv = 1. / r_;
    unsigned int flags = 0;
    double x, y;
    const double _2p43 = 65536.0 * 65536.0 * 2048.0;
    union {
      double d;
      int i[2];
    } yp;

    yp.d = _2p43;
    if (y_ < 0) {
      flags |= 4;
      y_ = -y_;
    }
    if (x_ < 0) {
      flags |= 2;
      x_ = -x_;
    }
    if (y_ > x_) {
      flags |= 1;
      x = rinv * y_;
      y = rinv * x_;
      yp.d += y;
    } else {
      x = rinv * x_;
      y = rinv * y_;
      yp.d += y;
    }

    int ind = (yp.i[0] & 0x03FF) * 2;  // 0 for little indian

    double* dasbuf = (double*)(datanbuf_ + ind);
    double sv = yp.d - _2p43;  // index fraction
    double cv = dasbuf[0];
    double asv = dasbuf[1];
    sv = y * cv - x * sv;  // delta sin value
    // double sv = y *(cv-x);
    // ____ compute arcsin directly
    double asvd = 6 + sv * sv;
    sv *= double(1.0 / 6.0);
    double th = asv + asvd * sv;
    if (flags & 1) {
      th = (_2pi / 4) - th;
    }
    if (flags & 2) {
      th = (_2pi / 2) - th;
    }
    if (flags & 4) {
      th = -th;
    }
    return std::pair<double, double>(th, overR ? rinv : r_);
  }

  // return eta phi saving some computation
  template <typename T>
  inline std::pair<T, T> etaphi(T x, T y, T z) {
    std::pair<T, T> por = atan2r(y, x, true);
    x = z * por.second;
    return std::pair<float, float>(std::log(x + std::sqrt(x * x + T(1))), por.first);
  }

}  // namespace fastmath

#endif