testreport.txt

CIRRUS 93XX系列windows mobile 6.0 BSP

                        
        <6> ceilf       The exception behavor is the same as soft-float version.
        <7> floorf      The exception behavor is the same as soft-float version.
        <8> fabsf       The exception behavor is the same as soft-float version.
        <9> sqrtf       The exceptions sent include inexact results and invalid operations.
                        The soft-float version sents the same invalid operations and nearly
                        double inexact results.
        <10> fmodf      The exception behavor is the same as soft-float version.
                        
        <11> pow        The exceptions sent are all inexact results. While, soft-float 
                        version sents overflow or underflow exceptions, instead.
        <12> log        The exception behavor is almost the same as that of soft-float
                        version. Both of them send division by zero, inexact result, and 
                        invalid operation exceptions. The only difference is one less 
                        invalid operation exception from soft-float version.
        <13> log10      The exception behavor is almost the same as that of soft-float
                        version. Both of them send division by zero, inexact result, and 
                        invalid operation exceptions. The only difference is one less 
                        invalid operation exception from soft-float version.
        <14> exp        The exception behavor is almost the same as that of soft-float
                        version. Both of them send inexact result, overflow and underflow 
                        exceptions. The only difference is one more inexact result exceptions
                        from soft-float version.

        <15> sin        The exception behavor is the same as soft-float version.
        <16> cos        The exception behavor is the same as soft-float version.
        <17> tan        The exception behavor is the same as soft-float version.
        <18> asin       The exception behavor is the same as soft-float version.
        <19> acos       The exception behavor is almost the same as that of soft-float 
                        version. Both of them send invalid operation and inexact result
                        exceptions. The only difference is one more inexact result 
                        exception in soft-float.
        <20> atan       The exception behavor is the same as soft-float version.
        <21> atan2      The exception behavor is the same as soft-float version.
                        
        <22> sinh       The exception behavor is the same as soft-float version.
        <23> cosh       The exception behavor is the same as soft-float version.
        <24> tanh       The exception behavor is almost the same as that of soft-float
                        version. Both of them send inexact result exceptions only. The
                        only difference is one more inexact result exceptions from
                        soft-float version.
                        
        <25> _y0        The exceptions sent include division by zero, inexact result and
                        invalid operation. The difference with soft-float version is that
                        soft-float sends much more inexact result exceptions.
        <26> _y1        The exceptions sent include division by zero, inexact result and
                        invalid operation. The soft-float version sends one more division
                        by zero exception, one more invalid operation exception and much 
                        more inexact result exceptions.
        <27> _yn        The exceptions sent include division by zero and inexact result.
                        The soft-float version also sends invalid operation exceptions.
                        And the soft-float version sends more division by zero and much
                        more inexact result exceptions.
        

6. Thread Security Test
        Thread security test is to verify that functions in crunchlib give correct results
        in a multi-task environment that there are more than one task running on the crunch
        coprocessor. The test is based on the accuracy test. That is two accruacy tests run
        in two independent threads simultaneously. The calculation results are compared with
        that of single-thread accuracy test.
        
        The following is the summarized test result of accuracytest.exe, the thread security
        test program.         
        
        <1> ceil        The results of both threads are the same as that of single thread.
        <2> floor       The results of both threads are the same as that of single thread.
        <3> fabs        The results of both threads are the same as that of single thread.
        <4> sqrt        The results of both threads are the same as that of single thread. 
        <5> fmod        The results of both threads are the same as that of single thread.
        
        <6> ceilf       The results of both threads are the same as that of single thread.
        <7> floorf      The results of both threads are the same as that of single thread.
        <8> fabsf       The results of both threads are the same as that of single thread.
        <9> sqrtf       The results of both threads are the same as that of single thread. 
        <10> fmodf      The results of both threads are the same as that of single thread.
        
        <11> pow        The results of multi-thread is slightly different with that of single
                        thread. The largest difference still occurs at bit 54. The average
                        difference occurs at bit 60.10188679. The average difference for single
                        thread occurs at 60.1171875. That is perhaps because that the test of
                        pow is based on randomly generated double values.
        <12> log        The results of both threads are the same as that of single thread.
        <13> log10      The results of both threads are the same as that of single thread.
        <14> exp        The results of both threads are the same as that of single thread.
        <15> sin        The results of both threads are the same as that of single thread.
        <16> cos        The results of both threads are the same as that of single thread.
        <17> tan        The results of both threads are the same as that of single thread.
        <18> asin       The results of both threads are the same as that of single thread.
        <19> acos       The results of both threads are the same as that of single thread.
        <20> atan       The results of both threads are the same as that of single thread.
        <21> atan2      The results of both threads are the same as that of single thread.
        <22> sinh       The results of both threads are the same as that of single thread.
        <23> cosh       The results of both threads are the same as that of single thread.
        <24> tanh       The results of both threads are the same as that of single thread.
        <25> _y0        The results of both threads are the same as that of single thread.
        <26> _y1        The results of both threads are the same as that of single thread.
        <27> _yn        The results of both threads are the same as that of single thread.
                  
                  
                  
7. Performance Test
        Performance test is to verify that functions in crunchlib are improved on speed
        compared with their counterparts in Windows CE kernel. To do the test, each 
        function in crunchlib and its soft-float counterpart run many times. The time
        consumed is recorded and percentage of speed improvement is calculated.        
                  
        The following is the summarized test result of performance.exe, the performance
        test program.
                  
                  crunchlib soft-float improved      accerlated    repeating
                    (ms)     (ms)      by (%)        by (times)    count
        <1>  ceil    78      126       35            1.615384615   5210*10         
        <2>  floor   79      121       35.53719008   1.53164557    5210*10         
***     <3>  fabs    40      35        -14.28571429  0.875         5210*10         
        <4>  sqrt    190     774       72.71448664   4.073684211   5210*10         
        <5>  fmod    405     5928      92.7639807    14.63703704   5210*10         
        <6>  ceilf   51      98        47.52475248   1.921568627   4130*10         
        <7>  floorf  50      105       45.91836735   2.1           4130*10  
        <8>  fabsf   29      40        34.14634146   1.379310345   4130*10         
        <9>  sqrtf   68      616       87.82161235   9.058823529   4130*10         
        <10> fmodf   84      1660      94.35215947   19.76190476   4130*10         
        <11> pow     397     2781      84.30913349   7.005037783   5210*10
        <12> log     198     1104      81.17195005   5.575757576   5210*10         
        <13> log10   243     1189      77.43119266   4.893004115   5210*10         
        <14> exp     109     681       82.64984227   6.247706422   5210*10         
        <15> sin     515     695       20.52067381   1.349514563   5210*10         
        <16> cos     523     638       9.532062392   1.219885277   5210*10         
        <17> tan     606     690       6.288343558   1.138613861   5210*10         
        <18> asin    124     338       59.74440895   2.725806452   5210*10         
        <19> acos    124     400       65.58265583   3.225806452   5210*10         
        <20> atan    152     588       73.92795883   3.868421053   5210*10         
        <21> atan2   290     775       62.20779221   2.672413793   5210*10         
        <22> sinh    148     671       77.59643917   4.533783784   5210*10         
        <23> cosh    197     958       79.62577963   4.862944162   5210*10         
        <24> tanh    169     552       68.44036697   3.266272189   5210*10         
        <25> _y0     1159    3688      68.63648685   3.182053494   5210*10         
        <26> _y1     1145    3750      69.15008069   3.27510917    5210*10         
        <27> _yn     48439   129410    62.56935322   2.671607589   4130



8. Test Conclusion
        The test conclusion can be summarized as bellow.
        1> The calcation error between hard-float and soft-float is within a tiny scale,
           except that sin, cos, and tan have larger periodical error near zero output.
        2> The exceptions sent are similar to those of soft-float. 
        3> The crunchlib functions are thread secure.
        4> All functions, but fabs, are improved in speed. Therefore, fabs will not be
           exported.       
        
        What prevents the result of hard-float math functions from being exactly the
        same as those of soft-float ones is the difference between the algorithms that
        crunchlib adopts and Windows CE kernel adopts. However, these difference in
        results is understandable and will not bring any ill effect to users' applications.
        
9. Reference
        1>crunchlibspec.txt
        
Appendices
        No