openjpeg fails test suite with fused-multiply-add instructions enabled

Hi

The test suite of OpenJPEG fails in the following cases:

         26 - NR-C1P0-p0_04.j2k-compare2base (Failed)                     
         29 - NR-C1P0-p0_05.j2k-compare2base (Failed)            
         32 - NR-C1P0-p0_06.j2k-compare2base (Failed)            
         68 - NR-C1P1-p1_02.j2k-compare2base (Failed)              
         71 - NR-C1P1-p1_03.j2k-compare2base (Failed)            
         74 - NR-C1P1-p1_04.j2k-compare2base (Failed)                                           
         77 - NR-C1P1-p1_05.j2k-compare2base (Failed)               
         89 - NR-JP2-file2.jp2-compare2base (Failed)                                     
        176 - NR-RIC-subsampling_1.jp2-compare2base (Failed)                                                                  
        179 - NR-RIC-subsampling_2.jp2-compare2base (Failed)                                                       
        182 - NR-RIC-zoo1.jp2-compare2base (Failed)                                              
        185 - NR-RIC-zoo2.jp2-compare2base (Failed)                                   
        929 - NR-DEC-_00042.j2k-2-decode-md5 (Failed)                           
        939 - NR-DEC-buxI.j2k-9-decode-md5 (Failed)                                      
        947 - NR-DEC-CT_Phillips_JPEG2K_Decompr_Problem.j2k-13-decode-md5 (Failed)
        957 - NR-DEC-Marrin.jp2-18-decode-md5 (Failed)                           
        983 - NR-DEC-kodak_2layers_lrcp.j2c-31-decode-md5 (Failed)    
        985 - NR-DEC-kodak_2layers_lrcp.j2c-32-decode-md5 (Failed)                              
        987 - NR-DEC-issue104_jpxstream.jp2-33-decode-md5 (Failed)                             
        996 - NR-DEC-file409752.jp2-40-decode-md5 (Failed)                                                     
        998 - NR-DEC-issue188_beach_64bitsbox.jp2-41-decode-md5 (Failed)                                                 
        1000 - NR-DEC-issue206_image-000.jp2-42-decode-md5 (Failed)                                        
        1002 - NR-DEC-issue205.jp2-43-decode-md5 (Failed)                                              
        1020 - NR-DEC-issue228.j2k-60-decode-md5 (Failed)                 
        1028 - NR-DEC-issue142.j2k-66-decode-md5 (Failed)           
        1030 - NR-DEC-issue134.jp2-67-decode-md5 (Failed)
        1032 - NR-DEC-issue135.j2k-68-decode-md5 (Failed)
        1034 - NR-DEC-issue208.jp2-69-decode-md5 (Failed)
        1036 - NR-DEC-issue211.jp2-70-decode-md5 (Failed)
        1040 - NR-DEC-broken.jpc-73-decode (Failed)
        1041 - NR-DEC-issue226.j2k-74-decode (Failed)
        1042 - NR-DEC-issue226.j2k-74-decode-md5 (Failed)
        1101 - NR-DEC-p1_04.j2k-124-decode-md5 (Failed)
        1103 - NR-DEC-p1_04.j2k-125-decode-md5 (Failed)
        1105 - NR-DEC-p1_04.j2k-126-decode-md5 (Failed)
        1107 - NR-DEC-p1_04.j2k-127-decode-md5 (Failed)
        1109 - NR-DEC-p1_04.j2k-128-decode-md5 (Failed)
        1111 - NR-DEC-p1_04.j2k-129-decode-md5 (Failed)
        1115 - NR-DEC-p1_04.j2k-131-decode-md5 (Failed)
        1129 - NR-DEC-p1_04.j2k-138-decode-md5 (Failed)
        1133 - NR-DEC-p1_04.j2k-140-decode-md5 (Failed)
        1164 - NR-DEC-p1_06.j2k-156-decode (Failed)
        1165 - NR-DEC-p1_06.j2k-156-decode-md5 (Failed)
        1180 - NR-DEC-p1_06.j2k-164-decode (Failed)
        1181 - NR-DEC-p1_06.j2k-164-decode-md5 (Failed)
        1185 - NR-DEC-p0_04.j2k-166-decode-md5 (Failed)
        1187 - NR-DEC-p0_04.j2k-167-decode-md5 (Failed)
        1189 - NR-DEC-p0_04.j2k-168-decode-md5 (Failed)
        1197 - NR-DEC-p0_04.j2k-172-decode-md5 (Failed)
        1358 - NR-DEC-issue205.jp2-253-decode-md5 (Failed)
        1360 - NR-DEC-issue236-ESYCC-CDEF.jp2-254-decode-md5 (Failed)
        1362 - NR-DEC-issue559-eci-090-CIELab.jp2-255-decode-md5 (Failed)
        1364 - NR-DEC-issue559-eci-091-CIELab.jp2-256-decode-md5 (Failed)

These failures seems to be related to the usage of fused-multiply-add (FMA)
instructions in the compiled code. Using the compiler flags

-O2 -march=broadwell

triggers above failures. The flags

-O2 -march=broadwell -ffp-contract=off

(which is disabling FMA) lets all test succeed except for two which are listed
in the known-to-fail file of Travis CI.

The GCC compiler enables FMA if -O2 is given and the/all target architectures
support FMA (if you have a different processor than broadwell, try -march=native
or change broadwell to the architecture you are using). The reason that these
tests fail seems to be in the changed floating point precision of a FMA
instruction (which is a*b + c) in contrast to the normal a*b + c (with
standard plus and multiply). Though I'm not an expert on that, so I might be
wrong.

I don't know if this directly counts as bug/issue since it is essential "the
hardware does not calculate correctly". However, it might be worth documenting
this and maybe also emit a warning in the configuration phase if FMA usage is
detected -- or maybe even disable it "-tfp-contract" is supported by GCC and clang
not sure about other compilers.

Alternatively but far more cumbersome and laborious, adapt the test suite to
be lenient to different floating point precisions.

oedipus

[rouault]

This list is slightly comparable to the known failures with Intel 32 bit mode (see https://github.com/uclouvain/openjpeg/blob/master/tools/travis-ci/knownfailures-Ubuntu12.04-gcc4.6.3-i386-Release-3rdP.txt ) which uses FP387 unit instead of SSE2 for floating point computations. So yes the test suite is quite sensitive currently. Might be good to add some tolerance in tests. I guess this affects only irreversible encoding / decoding.