On 31/08/2015 01:05, Joe Taylor wrote: > Hi Bill, Hi Joe, > > Thanks for your work on making a more rugged Fortran-to-C interface for > the nhash function. I'm sure it will make a good addition to our code > base. OK, I will test properly and commit to the wsjtx_exp branch when done. > > Here and elsewhere, as I'm sure you have long recognized, my approach > has often used quick-and-dirty approaches to test a new idea, make it > work, maybe tune it some for performance... and then, eventually, clean > up the code. Your efforts will be a big help, once again, in seeing > this sequence of steps through to the end. Inter language interoperation is a minefield and IMHO using any compiler based tools to specify formal interfaces saves time in the long run. The end is a long way off :( > > Simple answer to your question about why this hash algorithm is used: > it's handy, fast, has been used in WSPR mode for > 5 years, and is > already present in the WSJT-X code base. OK, that's fine. I mentioned it because Intel SSE4 has a crc32 instruction which will beat any software or table based CRC32 implementation hands down for performance at better than 5 bytes processed per CPU cycle. The zlib library is one common way of getting hash and CRC functions but I don't think it takes advantage of SSE SIMD instructions. All probably academic for our tiny messages. > > -- 73, Joe, K1JT 73 Bill G4WJS. > > On 8/30/2015 4:48 PM, Bill Somerville wrote: >> Hi Joe, >> >> I note the changes around the nhash C function and despite it not being >> the root cause of the initially reported JTMSK crashes it seems >> appropriate to suggest that we make use of the Fortran/C >> interoperability features of modern Fortran to ensure that >> interoperation is more robust and potentially a little more efficient. >> >> As an example I have prepared a module the defines the interface to the >> nhash C function for calling from Fortran which looks like: >> >> module hashing >> interface >> integer(c_int32_t) function nhash (key, length, initval) bind(C, >> name="nhash") >> use iso_c_binding, only: c_ptr, c_size_t, c_int32_t >> type(c_ptr), intent(in), value :: key >> integer(c_size_t), intent(in), value :: length >> integer(c_int32_t), intent(in), value :: initval >> end function nhash >> end interface >> end module hashing >> >> This can be called without any extra wrapping code, argument >> manipulation or, function renaming and allows any contiguous block of >> memory that is represented by a Fortran scalar or contiguous vector to >> be hashed. For example in the source lib/hash.f90: >> >> subroutine hash(string,len,ihash) >> use iso_c_binding, only: c_loc >> use hashing >> parameter (MASK15=32767) >> character*(*), target :: string >> i=nhash(c_loc(string),len,146) >> ihash=iand(i,MASK15) >> >> ! print*,'C',ihash,len,string >> return >> end subroutine hash >> >> rather than the current copy to an integer array. >> >> A more complete patch is attached that uses the module to define the >> interface in all places nhash is called from Fortran. >> >> I have tested this patch in ^/branches/wsjtx_exp in a trivial back to >> back JTMSK QSO more as a test of building that a thorough regression test. >> >> I also wonder why the nhash algorithm has been used for a CRC on JTMSK >> messages since there are probably simpler well tried and tested CRC >> algorithms in common usage. >> >> Comments? >> >> 73 >> Bill >> G4WJS. >>
------------------------------------------------------------------------------ _______________________________________________ wsjt-devel mailing list wsjt-devel@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/wsjt-devel
