1 #include <typeinfo>
2 #include <iostream>
3 #include <Eigen/Core>
4 #include "BenchTimer.h"
5 using namespace Eigen;
6 using namespace std;
7 
8 template<typename T>
sqsumNorm(const T & v)9 EIGEN_DONT_INLINE typename T::Scalar sqsumNorm(const T& v)
10 {
11   return v.norm();
12 }
13 
14 template<typename T>
hypotNorm(const T & v)15 EIGEN_DONT_INLINE typename T::Scalar hypotNorm(const T& v)
16 {
17   return v.hypotNorm();
18 }
19 
20 template<typename T>
blueNorm(const T & v)21 EIGEN_DONT_INLINE typename T::Scalar blueNorm(const T& v)
22 {
23   return v.blueNorm();
24 }
25 
26 template<typename T>
lapackNorm(T & v)27 EIGEN_DONT_INLINE typename T::Scalar lapackNorm(T& v)
28 {
29   typedef typename T::Scalar Scalar;
30   int n = v.size();
31   Scalar scale = 0;
32   Scalar ssq = 1;
33   for (int i=0;i<n;++i)
34   {
35     Scalar ax = internal::abs(v.coeff(i));
36     if (scale >= ax)
37     {
38       ssq += internal::abs2(ax/scale);
39     }
40     else
41     {
42       ssq = Scalar(1) + ssq * internal::abs2(scale/ax);
43       scale = ax;
44     }
45   }
46   return scale * internal::sqrt(ssq);
47 }
48 
49 template<typename T>
twopassNorm(T & v)50 EIGEN_DONT_INLINE typename T::Scalar twopassNorm(T& v)
51 {
52   typedef typename T::Scalar Scalar;
53   Scalar s = v.cwise().abs().maxCoeff();
54   return s*(v/s).norm();
55 }
56 
57 template<typename T>
bl2passNorm(T & v)58 EIGEN_DONT_INLINE typename T::Scalar bl2passNorm(T& v)
59 {
60   return v.stableNorm();
61 }
62 
63 template<typename T>
divacNorm(T & v)64 EIGEN_DONT_INLINE typename T::Scalar divacNorm(T& v)
65 {
66   int n =v.size() / 2;
67   for (int i=0;i<n;++i)
68     v(i) = v(2*i)*v(2*i) + v(2*i+1)*v(2*i+1);
69   n = n/2;
70   while (n>0)
71   {
72     for (int i=0;i<n;++i)
73       v(i) = v(2*i) + v(2*i+1);
74     n = n/2;
75   }
76   return internal::sqrt(v(0));
77 }
78 
79 #ifdef EIGEN_VECTORIZE
plt(const Packet4f & a,Packet4f & b)80 Packet4f internal::plt(const Packet4f& a, Packet4f& b) { return _mm_cmplt_ps(a,b); }
plt(const Packet2d & a,Packet2d & b)81 Packet2d internal::plt(const Packet2d& a, Packet2d& b) { return _mm_cmplt_pd(a,b); }
82 
pandnot(const Packet4f & a,Packet4f & b)83 Packet4f internal::pandnot(const Packet4f& a, Packet4f& b) { return _mm_andnot_ps(a,b); }
pandnot(const Packet2d & a,Packet2d & b)84 Packet2d internal::pandnot(const Packet2d& a, Packet2d& b) { return _mm_andnot_pd(a,b); }
85 #endif
86 
87 template<typename T>
pblueNorm(const T & v)88 EIGEN_DONT_INLINE typename T::Scalar pblueNorm(const T& v)
89 {
90   #ifndef EIGEN_VECTORIZE
91   return v.blueNorm();
92   #else
93   typedef typename T::Scalar Scalar;
94 
95   static int nmax = 0;
96   static Scalar b1, b2, s1m, s2m, overfl, rbig, relerr;
97   int n;
98 
99   if(nmax <= 0)
100   {
101     int nbig, ibeta, it, iemin, iemax, iexp;
102     Scalar abig, eps;
103 
104     nbig  = std::numeric_limits<int>::max();            // largest integer
105     ibeta = std::numeric_limits<Scalar>::radix; //NumTraits<Scalar>::Base;                    // base for floating-point numbers
106     it    = std::numeric_limits<Scalar>::digits; //NumTraits<Scalar>::Mantissa;                // number of base-beta digits in mantissa
107     iemin = std::numeric_limits<Scalar>::min_exponent;  // minimum exponent
108     iemax = std::numeric_limits<Scalar>::max_exponent;  // maximum exponent
109     rbig  = std::numeric_limits<Scalar>::max();         // largest floating-point number
110 
111     // Check the basic machine-dependent constants.
112     if(iemin > 1 - 2*it || 1+it>iemax || (it==2 && ibeta<5)
113       || (it<=4 && ibeta <= 3 ) || it<2)
114     {
115       eigen_assert(false && "the algorithm cannot be guaranteed on this computer");
116     }
117     iexp  = -((1-iemin)/2);
118     b1    = std::pow(ibeta, iexp);  // lower boundary of midrange
119     iexp  = (iemax + 1 - it)/2;
120     b2    = std::pow(ibeta,iexp);   // upper boundary of midrange
121 
122     iexp  = (2-iemin)/2;
123     s1m   = std::pow(ibeta,iexp);   // scaling factor for lower range
124     iexp  = - ((iemax+it)/2);
125     s2m   = std::pow(ibeta,iexp);   // scaling factor for upper range
126 
127     overfl  = rbig*s2m;          // overfow boundary for abig
128     eps     = std::pow(ibeta, 1-it);
129     relerr  = internal::sqrt(eps);      // tolerance for neglecting asml
130     abig    = 1.0/eps - 1.0;
131     if (Scalar(nbig)>abig)  nmax = abig;  // largest safe n
132     else                    nmax = nbig;
133   }
134 
135   typedef typename internal::packet_traits<Scalar>::type Packet;
136   const int ps = internal::packet_traits<Scalar>::size;
137   Packet pasml = internal::pset1(Scalar(0));
138   Packet pamed = internal::pset1(Scalar(0));
139   Packet pabig = internal::pset1(Scalar(0));
140   Packet ps2m = internal::pset1(s2m);
141   Packet ps1m = internal::pset1(s1m);
142   Packet pb2  = internal::pset1(b2);
143   Packet pb1  = internal::pset1(b1);
144   for(int j=0; j<v.size(); j+=ps)
145   {
146     Packet ax = internal::pabs(v.template packet<Aligned>(j));
147     Packet ax_s2m = internal::pmul(ax,ps2m);
148     Packet ax_s1m = internal::pmul(ax,ps1m);
149     Packet maskBig = internal::plt(pb2,ax);
150     Packet maskSml = internal::plt(ax,pb1);
151 
152 //     Packet maskMed = internal::pand(maskSml,maskBig);
153 //     Packet scale = internal::pset1(Scalar(0));
154 //     scale = internal::por(scale, internal::pand(maskBig,ps2m));
155 //     scale = internal::por(scale, internal::pand(maskSml,ps1m));
156 //     scale = internal::por(scale, internal::pandnot(internal::pset1(Scalar(1)),maskMed));
157 //     ax = internal::pmul(ax,scale);
158 //     ax = internal::pmul(ax,ax);
159 //     pabig = internal::padd(pabig, internal::pand(maskBig, ax));
160 //     pasml = internal::padd(pasml, internal::pand(maskSml, ax));
161 //     pamed = internal::padd(pamed, internal::pandnot(ax,maskMed));
162 
163 
164     pabig = internal::padd(pabig, internal::pand(maskBig, internal::pmul(ax_s2m,ax_s2m)));
165     pasml = internal::padd(pasml, internal::pand(maskSml, internal::pmul(ax_s1m,ax_s1m)));
166     pamed = internal::padd(pamed, internal::pandnot(internal::pmul(ax,ax),internal::pand(maskSml,maskBig)));
167   }
168   Scalar abig = internal::predux(pabig);
169   Scalar asml = internal::predux(pasml);
170   Scalar amed = internal::predux(pamed);
171   if(abig > Scalar(0))
172   {
173     abig = internal::sqrt(abig);
174     if(abig > overfl)
175     {
176       eigen_assert(false && "overflow");
177       return rbig;
178     }
179     if(amed > Scalar(0))
180     {
181       abig = abig/s2m;
182       amed = internal::sqrt(amed);
183     }
184     else
185     {
186       return abig/s2m;
187     }
188 
189   }
190   else if(asml > Scalar(0))
191   {
192     if (amed > Scalar(0))
193     {
194       abig = internal::sqrt(amed);
195       amed = internal::sqrt(asml) / s1m;
196     }
197     else
198     {
199       return internal::sqrt(asml)/s1m;
200     }
201   }
202   else
203   {
204     return internal::sqrt(amed);
205   }
206   asml = std::min(abig, amed);
207   abig = std::max(abig, amed);
208   if(asml <= abig*relerr)
209     return abig;
210   else
211     return abig * internal::sqrt(Scalar(1) + internal::abs2(asml/abig));
212   #endif
213 }
214 
215 #define BENCH_PERF(NRM) { \
216   Eigen::BenchTimer tf, td, tcf; tf.reset(); td.reset(); tcf.reset();\
217   for (int k=0; k<tries; ++k) { \
218     tf.start(); \
219     for (int i=0; i<iters; ++i) NRM(vf); \
220     tf.stop(); \
221   } \
222   for (int k=0; k<tries; ++k) { \
223     td.start(); \
224     for (int i=0; i<iters; ++i) NRM(vd); \
225     td.stop(); \
226   } \
227   for (int k=0; k<std::max(1,tries/3); ++k) { \
228     tcf.start(); \
229     for (int i=0; i<iters; ++i) NRM(vcf); \
230     tcf.stop(); \
231   } \
232   std::cout << #NRM << "\t" << tf.value() << "   " << td.value() <<  "    " << tcf.value() << "\n"; \
233 }
234 
check_accuracy(double basef,double based,int s)235 void check_accuracy(double basef, double based, int s)
236 {
237   double yf = basef * internal::abs(internal::random<double>());
238   double yd = based * internal::abs(internal::random<double>());
239   VectorXf vf = VectorXf::Ones(s) * yf;
240   VectorXd vd = VectorXd::Ones(s) * yd;
241 
242   std::cout << "reference\t" << internal::sqrt(double(s))*yf << "\t" << internal::sqrt(double(s))*yd << "\n";
243   std::cout << "sqsumNorm\t" << sqsumNorm(vf) << "\t" << sqsumNorm(vd) << "\n";
244   std::cout << "hypotNorm\t" << hypotNorm(vf) << "\t" << hypotNorm(vd) << "\n";
245   std::cout << "blueNorm\t" << blueNorm(vf) << "\t" << blueNorm(vd) << "\n";
246   std::cout << "pblueNorm\t" << pblueNorm(vf) << "\t" << pblueNorm(vd) << "\n";
247   std::cout << "lapackNorm\t" << lapackNorm(vf) << "\t" << lapackNorm(vd) << "\n";
248   std::cout << "twopassNorm\t" << twopassNorm(vf) << "\t" << twopassNorm(vd) << "\n";
249   std::cout << "bl2passNorm\t" << bl2passNorm(vf) << "\t" << bl2passNorm(vd) << "\n";
250 }
251 
check_accuracy_var(int ef0,int ef1,int ed0,int ed1,int s)252 void check_accuracy_var(int ef0, int ef1, int ed0, int ed1, int s)
253 {
254   VectorXf vf(s);
255   VectorXd vd(s);
256   for (int i=0; i<s; ++i)
257   {
258     vf[i] = internal::abs(internal::random<double>()) * std::pow(double(10), internal::random<int>(ef0,ef1));
259     vd[i] = internal::abs(internal::random<double>()) * std::pow(double(10), internal::random<int>(ed0,ed1));
260   }
261 
262   //std::cout << "reference\t" << internal::sqrt(double(s))*yf << "\t" << internal::sqrt(double(s))*yd << "\n";
263   std::cout << "sqsumNorm\t"  << sqsumNorm(vf)  << "\t" << sqsumNorm(vd)  << "\t" << sqsumNorm(vf.cast<long double>()) << "\t" << sqsumNorm(vd.cast<long double>()) << "\n";
264   std::cout << "hypotNorm\t"  << hypotNorm(vf)  << "\t" << hypotNorm(vd)  << "\t" << hypotNorm(vf.cast<long double>()) << "\t" << hypotNorm(vd.cast<long double>()) << "\n";
265   std::cout << "blueNorm\t"   << blueNorm(vf)   << "\t" << blueNorm(vd)   << "\t" << blueNorm(vf.cast<long double>()) << "\t" << blueNorm(vd.cast<long double>()) << "\n";
266   std::cout << "pblueNorm\t"  << pblueNorm(vf)  << "\t" << pblueNorm(vd)  << "\t" << blueNorm(vf.cast<long double>()) << "\t" << blueNorm(vd.cast<long double>()) << "\n";
267   std::cout << "lapackNorm\t" << lapackNorm(vf) << "\t" << lapackNorm(vd) << "\t" << lapackNorm(vf.cast<long double>()) << "\t" << lapackNorm(vd.cast<long double>()) << "\n";
268   std::cout << "twopassNorm\t" << twopassNorm(vf) << "\t" << twopassNorm(vd) << "\t" << twopassNorm(vf.cast<long double>()) << "\t" << twopassNorm(vd.cast<long double>()) << "\n";
269 //   std::cout << "bl2passNorm\t" << bl2passNorm(vf) << "\t" << bl2passNorm(vd) << "\t" << bl2passNorm(vf.cast<long double>()) << "\t" << bl2passNorm(vd.cast<long double>()) << "\n";
270 }
271 
main(int argc,char ** argv)272 int main(int argc, char** argv)
273 {
274   int tries = 10;
275   int iters = 100000;
276   double y = 1.1345743233455785456788e12 * internal::random<double>();
277   VectorXf v = VectorXf::Ones(1024) * y;
278 
279 // return 0;
280   int s = 10000;
281   double basef_ok = 1.1345743233455785456788e15;
282   double based_ok = 1.1345743233455785456788e95;
283 
284   double basef_under = 1.1345743233455785456788e-27;
285   double based_under = 1.1345743233455785456788e-303;
286 
287   double basef_over = 1.1345743233455785456788e+27;
288   double based_over = 1.1345743233455785456788e+302;
289 
290   std::cout.precision(20);
291 
292   std::cerr << "\nNo under/overflow:\n";
293   check_accuracy(basef_ok, based_ok, s);
294 
295   std::cerr << "\nUnderflow:\n";
296   check_accuracy(basef_under, based_under, s);
297 
298   std::cerr << "\nOverflow:\n";
299   check_accuracy(basef_over, based_over, s);
300 
301   std::cerr << "\nVarying (over):\n";
302   for (int k=0; k<1; ++k)
303   {
304     check_accuracy_var(20,27,190,302,s);
305     std::cout << "\n";
306   }
307 
308   std::cerr << "\nVarying (under):\n";
309   for (int k=0; k<1; ++k)
310   {
311     check_accuracy_var(-27,20,-302,-190,s);
312     std::cout << "\n";
313   }
314 
315   std::cout.precision(4);
316   std::cerr << "Performance (out of cache):\n";
317   {
318     int iters = 1;
319     VectorXf vf = VectorXf::Random(1024*1024*32) * y;
320     VectorXd vd = VectorXd::Random(1024*1024*32) * y;
321     VectorXcf vcf = VectorXcf::Random(1024*1024*32) * y;
322     BENCH_PERF(sqsumNorm);
323     BENCH_PERF(blueNorm);
324 //     BENCH_PERF(pblueNorm);
325 //     BENCH_PERF(lapackNorm);
326 //     BENCH_PERF(hypotNorm);
327 //     BENCH_PERF(twopassNorm);
328     BENCH_PERF(bl2passNorm);
329   }
330 
331   std::cerr << "\nPerformance (in cache):\n";
332   {
333     int iters = 100000;
334     VectorXf vf = VectorXf::Random(512) * y;
335     VectorXd vd = VectorXd::Random(512) * y;
336     VectorXcf vcf = VectorXcf::Random(512) * y;
337     BENCH_PERF(sqsumNorm);
338     BENCH_PERF(blueNorm);
339 //     BENCH_PERF(pblueNorm);
340 //     BENCH_PERF(lapackNorm);
341 //     BENCH_PERF(hypotNorm);
342 //     BENCH_PERF(twopassNorm);
343     BENCH_PERF(bl2passNorm);
344   }
345 }
346