Index: /issm/trunk/externalpackages/adic/install.sh =================================================================== --- /issm/trunk/externalpackages/adic/install.sh (revision 9707) +++ /issm/trunk/externalpackages/adic/install.sh (revision 9707) @@ -0,0 +1,33 @@ +#!/bin/bash + +#get number of CPUS +NUMCPUS=$1; + + +#Erase install +rm -rf source build install + +#Download code +hg clone http://mercurial.mcs.anl.gov/ad/ADIC -r 631 source + +#Then update to -r 631 +cd source +hg update -r 631 + +cd source +./autogen.sh +./aclocal +cd .. + +mkdir build +cd build +../source/configure --with-rose=$ISSM_TIER/externalpackages/rose/install \ + --with-openanalysis=$ISSM_TIER/externalpackages/openanalysis/openanalysis/x86_64-Linux \ + --with-boost=$ISSM_TIER/externalpackages/boost/install \ + --with-xerces=$ISSM_TIER/externalpackages/xerces/src \ + --with-xaifbooster=$ISSM_TIER/externalpackages/xaifbooster/xaifBooster \ + --with-colpack=$ISSM_TIER/externalpackages/colpack/install\ + --prefix=$ISSM_TIER/externalpackages/adic/install + +make -j $NUMCPUS +make -j $NUMCPUS install Index: /issm/trunk/externalpackages/adic/tests/notes =================================================================== --- /issm/trunk/externalpackages/adic/tests/notes (revision 9707) +++ /issm/trunk/externalpackages/adic/tests/notes (revision 9707) @@ -0,0 +1,143 @@ +mini1.c is: +7:37 PM +#include + +void mini1(double *y, double *x) +{ + int i; + for (i = 0; i < 2; i=i+1) { + y[i] = x[i] + sin(x[i]*x[i]); + } + } + Eric Larour 7:37 PM + ok + Sri Hari Krishna Narayanan 7:37 PM + If you cat any of mini1_grad_length.ad.c mini1_dense_forward.ad.c mini1_sparse_forward.ad.c + 7:37 PM + You should be able to see the forward mode output + Eric Larour 7:38 PM + ok + Sri Hari Krishna Narayanan 7:38 PM + First the makefile compiles mini1_grad_length.ad.c with a runtime library to estimate the number of columns in the Jacobian. + Eric Larour 7:39 PM + ok, makes sense + Sri Hari Krishna Narayanan 7:39 PM + That generates the output : ADIC_GRADVEC_LENGTH 2 + Eric Larour 7:39 PM + where? + 7:39 PM + ok, sorry, got it + Sri Hari Krishna Narayanan 7:40 PM + The value is also written to : runtime_dense/adic_gradvec_length.h + Eric Larour 7:40 PM + where is the main for this? + 7:40 PM + oh, it creates a .h directly + Sri Hari Krishna Narayanan 7:40 PM + mini1_driver.c + 7:40 PM + yes. + 7:40 PM + The runtime library takes care of that + 7:40 PM + There is alot or #defined compilation going on + Eric Larour 7:41 PM + wraps around mini1_forward_array.exe ? + Sri Hari Krishna Narayanan 7:41 PM + /*intermediary: */ creates mini1_grad_length.exe + Eric Larour 7:41 PM + Isn't that the Makefile? what is the runtime library ? + 7:42 PM + ok, the Makefile launches the runtime library + Sri Hari Krishna Narayanan 7:42 PM + Makefile copies versions of the runtime library (bunch of macros in most cases) + Eric Larour 7:42 PM + what is the name of the runtime library? + Sri Hari Krishna Narayanan 7:43 PM + form $adicsrc/runtime/ + 7:43 PM + There are multiple versions + Eric Larour 7:44 PM + ah, I see the grad_length dir + 7:44 PM + is all that stuff compiled into adic2? + Sri Hari Krishna Narayanan 7:44 PM + The one used to estimate the grad_lengh is : $adicinstall/grad_length + 7:44 PM + yes + Eric Larour 7:44 PM + ok, makes sense + Sri Hari Krishna Narayanan 7:45 PM + For the next step though + 7:45 PM + the 'dense' case + 7:45 PM + the runtime library cannot be precompiled + 7:45 PM + the .h file has been created in the previous step + 7:46 PM + So we copy the runtime/dense to the local directory + Eric Larour 7:46 PM + because we can't guess the size of the gradient right? + Sri Hari Krishna Narayanan 7:46 PM + and copy the .h into it + 7:46 PM + correct + Eric Larour 7:46 PM + nice + Sri Hari Krishna Narayanan 7:46 PM + Then we compile it and create a .la file + 7:46 PM + which is then linked to create mini1_dense_forward.exe + 7:47 PM + When we run this, it also happens to run the analyticla derivates (the driver is built this way) + 7:47 PM + Also diveded differences + 7:47 PM + So you get the output: + 7:47 PM + Analytic result is: [1.968912] + Analytic result is: [1.968912] + DD result is: [1.968997] error {-8.45e-05} + DD result is: [1.968997] error {-8.45e-05} + AD result is: [1.968912] error {0.000000} + AD result is: [1.968912] error {0.000000} + 7:48 PM + The next step is to detect sparsity in the Jacobian and possibly compress it using Colpack (create a seed matrix). + 7:49 PM + The gradient computation takes place on the compressed matrix. + 7:49 PM + The output : ADIC_GRADVEC_LENGTH 1 + 7:49 PM + means that the seed matrix has one column. + 7:50 PM + After the calculation of the elements of the compressed matrix, it is then extracted to form the 'full Jacobian' (sorry if these tems are inexact). + 7:51 PM + The extracted output is then printed out: + 7:51 PM + AD result is: [1.968912] + AD result is: [1.968912] + 7:51 PM + And we are done! + Eric Larour 7:51 PM + This is a lot for me to take in at once I'm going to have to look through the makefile, + 7:51 PM + and your explanations, and figure it out + Sri Hari Krishna Narayanan 7:51 PM + Sure. + Eric Larour 7:52 PM + I'll get back to you as soon as I have injested all of this + Sri Hari Krishna Narayanan 7:52 PM + No problem. + Eric Larour 7:52 PM + I'll probably have a million questions + Sri Hari Krishna Narayanan 7:52 PM + This is an exhaustive test + 7:52 PM + FOr m ost purposes, we wil not need all of this + Eric Larour 7:52 PM + this is really nice, I'm pretty exited ! + Sri Hari Krishna Narayanan 7:52 PM + Jean belives for example the the grad_length version is pointless. I find it interesting though. + 7:53 PM + Okay. There are more interesting examples we can try