Prev Next whats_new_12

CppAD Changes and Additions During 2012

Introduction
This section contains a list of the changes to CppAD during 2012 (in reverse order by date). The purpose of this section is to assist you in learning about changes between various versions of CppAD.

05-24
Merged in changes from branches/sparse:
  1. A new interface was added to sparse_jacobian and sparse_hessian . This interface returns a sparse representation of the corresponding matrices using row and column index vectors.
  2. The examples sparse_jacobian.cpp and sparse_hessian.cpp were improved and extended to include the new interface.
  3. The definition of an AD function was improved to include definition of the corresponding n and m .


04-19
The BOOST_DIR configure command line value has been changed to be the corresponding prefix during the installation of boost. To be specific, it used to be that boost_dir/boost was the boost include directory, now boost_dir/include is the boost include directory. This make it the same as the other directory arguments on the configure command line. In addition, it fixes some bugs in the detection of the boost multi-threading library.

04-18
Add documentation and testing for not using free_all and user_atomic clear while in parallel mode.

04-17
Fix bug when using user_atomic functions with multi_threading .

04-10
Add control of the max_num_threads argument to the unix configure command.

04-06
  1. A change was made to the way that the tapes were managed to reduce false sharing during multi-threading . Because of this change, it is now suggest that the user call parallel_ad after the multi-threading section of the program.
  2. The routine ta_free_all was created to make it easier to manage memory and the routine memory_leak was deprecated.
  3. Add the -lteuchos flag to the link line for the speed_sacado tests. (This was not necessary for trilinos-10.8.3 but is necessary for trilinos-10.10.1)


04-05
The restriction was added that parallel_ad cannot be called while a tape is being recorded. This was necessary inorder to initialize some new statics in the tape.

04-01
Fixed a race condition when using CppAD with multi-threading . This has been fixed and the error message below no longer occurs. Suppose that you ran the CppAD configure command in the work directory. If you then edited the file work/multi_thread/makefile and changed 
 
	# AM_CXXFLAGS     = -g $(CXX_FLAGS)
	AM_CXXFLAGS = -DNDEBUG -O2 $(CXX_FLAGS)
to 
 
	AM_CXXFLAGS     = -g $(CXX_FLAGS)
	# AM_CXXFLAGS = -DNDEBUG -O2 $(CXX_FLAGS)
and then executed the commands 
 
	make clean
	make pthread_test
	valgrind --tool=helgrind ./pthread_test simple_ad
The following error message would result:
     ... snip ...
==7041== Possible data race during write of size 4 at 0x8077460 by thread #1
==7041==    at 0x804FE23: CppAD::AD<double>::tape_new() (tape_link.hpp:221)
     ... snip ...
03-27
Reduce the amount of memory allocation and copying of information during a Dependent operation or an ADFun sequence constructor .

03-26
Calling taylor_capacity, with to with capacity equal to zero, was not freeing memory . This has been fixed.

03-23
  1. Improve, the multi-threading examples simple_ad_openmp.cpp , simple_ad_bthread.cpp , and simple_ad_pthread.cpp . This includes separating generic code that can be used for all applications from problem specific code.
  2. Add initialization of statics in CheckSimpleVector during parallel_ad call. These statics are required to use CppAD::vector .
  3. Add a debugging check to make sure CheckSimpleVector is initialized in sequential mode.


03-21
Fix an incorrect error check in thread_alloc that did not allow ta_return_memory to return memory in sequential execution mode that was allocated by a different thread during parallel execution.

03-17
Debian recently converted the default shell from bash to dash (which caused multi_thread/test.sh to fail). This has been fixed. In general, Debian's policy is that bin/sh will be a Posix Shell .

03-11
There was a bug in thread_alloc where extra memory was held onto even if hold_memory was never called and only one thread was used by the program. This caused
        valgrind --leak-check=full --show-reachable=yes
         to generate an error message. If multiple threads are used, one should free this extra memory for threads other than thread zero. If hold_memory is used, one should call free_available for all threads.

03-03
  1. Add the examples simple_ad_openmp.cpp , simple_ad_bthread.cpp and simple_ad_pthread.cpp .
  2. Fix bug in finding boost multi-threading library (due to fact that boost_dir is not the prefix during the boost installation).


03-02
  1. Change the name simple_ad.cpp to team_example.cpp
  2. The multi-threading team_example.cpp example was changed to use f ( x ) = x 2 instead of the function atan 2 [ sin ( x ) , cos ( x ) ] (both functions should behave like the identity function f ( x ) = x ). This enabled the removal of multi_thread/arc_tan.cpp.
  3. In team_example.cpp check that all of the threads pass their individual test; i.e. work_all_[thread_num].ok is true for all thread_num .


02-11
  1. The requirements in base_member were missing from the base_require documentation. In addition, the base_require.cpp example has been added.
The specifications for memory_leak where changes so that calling routine specifies the amount of static memory to add. In addition, it is now possible to call memory_leak when num_threads is greater than one (still can't be in parallel mode).

02-10
  1. Add the missing Base class requirements in the entire base_member section and under the Output Operator in the base_require section.
  2. Add the base_alloc.hpp example.


02-09
  1. Add the set_static to memory_leak . This is necessary for testing base types that allocate memory for each element.
  2. Fix memory allocation bug in cppad/local/pod_vector.hpp when each element of the Base type allocated memory.


01-30
Make another attempt to fix linking with boost threads where the wrong version of the library is in the system include directory; i.e., to have boost_dir override the default library.

01-27
There were some problems with configure's automatic detection of the boost multi-threading library. These have been fixed.

01-24
It used to be that thread_alloc did not hold onto memory when num_threads was one in the previous call to parallel_setup . Holding onto memory is now controlled by the separate routine hold_memory . This give the user more control over the memory allocator and the ability to obtain a speed up even when there is only one thread. To convert old code to the new interface, after each call to
        thread_alloc::parallel_setup(num_threads, in_parallel, thread_num);
         put the following call
        thread_alloc::hold_memory(num_threads > 1);
        
01-23
Change variable notation and use optimize in mul_level.cpp .

01-20
  1. Add the example change_const.cpp which shows how to compute derivatives of functions that have parameters that can change (but derivatives are not computed with respect to these parameters).
  2. The documentation for machine epsilon has been improved. (The fact that it can be used for Base types was missing.)


01-19
  1. In cases where test.sh is trivial, put its operations in corresponding makefile.
  2. Fix problem compiling cppad/speed/sparse_evaluate.hpp under gcc on Fedora 17.
  3. Run multi_thread/test.sh from source directory (no need to copy to build directory).


01-16
The test program multi_thread/test.sh failed if the openmp_flags not present in the configure command. This has been fixed. In addition, this test.sh has been made faster by cycling through the available threading systems instead of doing every system for every test.

01-15
Fix make test so it works when configure is run in the distribution directory cppad-yyyymmdd (not just when it is run in a different directory).

01-12
The -lpthread library was missing from the multi_thread test program linker command. This has been fixed.

01-07
  1. A duplicated code block beginning with 
     
if( fabs( fcur ) <= epsilon_ )
    was removed from the routine multi_newton_worker, in the file multi_newton_work.cpp .
  2. The distance between solutions that are joined to one solution has been corrected from ( b - a ) / ( 2 n ) to ( b - a ) / n ; see xout . The correction was in the file multi_newton_work.cpp where sub_length_ / 2 was change to sub_length_.


01-02
  1. The thread_alloc memory allocator was changed to avoid certain false sharing situations (cases where two different thread were changing and using memory that is on the same page of cache). On one tests machine, the execution time for the 32 thread case for the test 
     
	   ./openmp_test multi_newton 1 32 1000 4800 10 true
    improved from 0.0302 seconds to 0.0135 seconds.
  2. There was a problem with the correctness test section of the multi_newton_time.cpp test. The convergence criteria, and correctness criteria, needed to be scaled by the largest argument values. This was a problem with over a hundred zeros were included in the test (and the largest argument value was 100 π or more).
  3. There was a problem with the way that multi_newton_combine joined two solutions into one. It is possible that one of the solutions that needs to be joined is on the boundary and very close to a solution in the next (or previous interval) that is not on the boundary. In this case, the one with the smaller function value is chosen.
for the previous
Input File: omh/whats_new_12.omh