trex-coe / qmckl Goto Github PK

Some compilers/systems are limiting stack size (beyond "ulimit -s").
About Intel compilers, limitation is 4 MB. That can be changed via the KMP_STACKSIZE env. variable, for instance using "10m" value for a 10 Megabytes limit.

source: https://community.intel.com/t5/Intel-Fortran-Compiler/Segmentation-fault-when-using-large-array-with-OpenMP/m-p/758829

bench_jastrow is segfaulting on the Alz_large dataset, more precisely on the een_rescaled_e_ij local stack array defined in the parallel region of the qmckl_compute_een_rescaled_e_hpc function.

A naive workaround consist in dynamically (heap) allocating that array (and then freeing it at the end of the scope). Maybe you could find a smarter fix...

Instead of
double een_rescaled_e_ij[len_een_ij];
I have succesfuly tested
double *een_rescaled_e_ij = malloc (len_een_ij * sizeof (een_rescaled_e_ij[0]));
...
free (een_rescaled_e_ij);

Probably better to use qmckl_malloc/free, if applicable.

Begin the Chameleon build of QMCkl. This implies extensive work to make it easy for the user to install QMCkl with Chameleon support.

• Build with chameleon
• Use chameleon via define flags
• Example version of dgemm with chameleon

Functions for rescaled distances.

The function was not updated after transition to the new TREXIO basis set format and so it currently returns nucleus_index array of size nucleus_num instead of ao_basis_shell_num.

I can fix that one but now the tests have to be updated too.

With Nvidia compilers, fortran generates an internal compiler error. Dependencies in configure need to be checked.

• README: out of source build
• configure without swig

Write en-distances based on ee-distances already started.

• Make sure build is working
• Make sure testing is working
• en distance implementation
• Write tests

TODO list for MOs

• Implement native qmckl_dgemm
• Implement MOs
• Add tests

The c and fortran inteface of qmckl can be set up to be autogenerated from the documentation of the functions.

There are multiple advantages of this strategy which are as follows:

• Consistent interface generation
• Consistent conventions of function names
• Consistent conventions for arguments
• Documentation is by construction up to date with the code
• Users and developers are not required to write the interface thus reducing hand written lines of code

Some important points that need to be stressed while adopting this strategy are as follows:

• New users and developers need to be introduced to the documentation procedure
• A standard has to be clearly documented and defined

TODO for calculating Kinetic Energy:

• Implement determinant, adjoint, and inverse of the matrix
• VGL of Determinant
• Adjoint matrix
• Inverse of Determinant
• Kinetic energy
• Potential energy
• Local energy
• Drift vector
• Move
• Acceptance probability

The qmckl_context_get_range function is not found in the libqmckl.so library.

nm libqmckl.so does not show the qmckl_context_get_range function.

Currently there are two principal branches: main and master.
It's not clear which one is the authoritative one. I suggest we pick one and stick with it.

The main branch seems to lag behind but some of the links still point to it. For example, the README.md in master branch documentation has links that point to the main branch instead.

What do you think @scemama ?

Thanks !

Files in tools` directory should be put in org-mode files with documentation around.

While compiling fortran files with ifort the following FLAGS are needed:

FCFLAGS='-nofor-main -assume nounderscore'

The call to configure including the above options for developers would look like:

QMCKL_DEVEL=1 ./configure --enable-silent-rules --enable-maintainer-mode FC='ifort' FCFLAGS='-nofor-main -assume nounderscore'

TLDR: Local energy and Determinant tests failing with icc and ifort

Configure command used:

./configure --prefix=$CONDA_PREFIX FC=ifort FCFLAGS='-g -xCORE-AVX2 -O3 ' CC='icc' CFLAGS='-g -xCORE-AVX2 -O3 -fno-omit-frame-pointer'

Summary of configuration:

CC..............: icc
CPPFLAGS........: 
CFLAGS..........: -g -xCORE-AVX2 -O3 -fno-omit-frame-pointer  
FC..............: ifort
FCLAGS..........: -g -xCORE-AVX2 -O3  -nofor-main
LDFLAGS:........: 
LIBS............: -lopenblas    -ltrexio -lpthread -lm 
USE CHAMELEON...: 
HPC version.....: no

Package features:
   trexio

Output running make check:

============================================================================
Testsuite summary for qmckl 0.1.1
============================================================================
# TOTAL: 16
# PASS:  14
# SKIP:  0
# XFAIL: 0
# FAIL:  2
# XPASS: 0
# ERROR: 0

Failing tests:

./tools/test-driver: line 107: 226043 Segmentation fault      (core dumped) "$@" > $log_file 2>&1
FAIL: tests/test_qmckl_determinant
...
./tools/test-driver: line 107: 226277 Segmentation fault      (core dumped) "$@" > $log_file 2>&1
FAIL: tests/test_qmckl_local_energy

We should not require the make all to depend on building the html documentation files using htmlize.el. The reason is that it happens every now and then that the htmlize.el is not present and so each make attempt will try to download it (e.g. after make maintainer-clean). Several issues that I observed:

1. No internet connection at build time: this is quite common (maybe remove htmlize.el from the files to be cleaned?)
2. No access to git: in principal, it would be nice to decouple the build process from the git commands, see one of my last commits #85a0d485575e55b0a04a192331f37538e0ab5900 . For example. building Guix package from the GitHub repo fails exactly because of that (internal issue to clone the git submodule).

Calculate the Jastrow factor and derivatives using a simple DGEMM formulation.

• Build factor_ee
• Build factor_en
• Build factor_een
• Build factor_ee_deriv_e
• Build factor_en_deriv_e
• Build factor_een_deriv_e

On latest master commit while using icc/ifort 2021.3 + -xCORE-AVX2 + fno-omit-frame-pointer
I tried to run test-driver directly but didn't manage to have it to work (was always missing one argument apparently ...).

Configure command used :
./configure --prefix=$PWD/_install CC=icc FC=ifort CFLAGS="-g -O2 -xCORE-AVX2 -fno-omit-frame-pointer" FCFLAGS="-cpp -g -O2 -xCORE-AVX2 -fno-omit-frame-pointer"

Resume :

qmckl Version 0.1.1  -- Developer mode

Prefix: '/home/kcamus/Trex/qmckl/qmckl/_install'.

CC..............: icc
CPPFLAGS........: 
CFLAGS..........: -g -O2 -xCORE-AVX2 -fno-omit-frame-pointer  
FC..............: ifort
FCLAGS..........: -cpp -g -O2 -xCORE-AVX2 -fno-omit-frame-pointer -nofor-main
LDFLAGS:........: 
LIBS............: -lopenblas -llapack -lopenblas   -ltrexio -lpthread -lm 
USE CHAMELEON...: 

Package features:
   trexio

Output extract when running make check (all others tests pass):

PASS: tests/test_qmckl_jastrow
./tools/test-driver: line 109: 541134 Segmentation fault      (core dumped) "$@" > $log_file 2>&1
FAIL: tests/test_qmckl_local_energy

Let me know if you need anything specific or if I need to try something.

Derivatives of the following

• electron - nuclear rescaled distances
• electron - electron rescaled distances
• Tests

An extra pointer argument to an abstract struct is needed to enable the possibility to pass infos to allocate memory on CPU or GPU, alignment, etc...

Try to fix build install fails in 4a63473

• First make sure make install works
• Make sure the make dist check works

Same issue as the one reported for TREXIO: TREX-CoE/trexio#62

• Issue: Compilation fails with the following message

make[2]: *** No rule to make target 'share/qmckl/fortran/qmckl_f.f90', needed by 'src/qmckl_f.f90'. Stop.

• Steps to reproduce

Clone the repository
git clone https://github.com/TREX-CoE/qmckl
configure
./autogen
./configure --without-trexio
Issue make command
make

Required for Python interface

A cutoff of 1.e-12 in champ is too high and breaks.

TODO: Create a new repo containing all GPU functions. This will allow to configure and compile teh GPU functions with the best GPU compiler and keep the best CPU compiler for CPU functions.

Integrate Sherman-Morrison-Woodbury into QMCkl

• Basic integration of kernels (QMCkl compiles and kernels are callable from outside)
• Complete basic requirements documentation
• Add real tests
• Add and complete documentation
• Take care of compiler warnings

In the generated qmckl.h header file the lists of arguments follow 3 different patterns (see below). It adds additional complexity to processing the header file by external tools (e.g. SWIG which parses the header file looking for certain patterns in the arguments lists).

Having arguments in one line is also good for grep-ing but might decrease the readability of the header file.

<func_name> (<type1>* <arg1>, ...)

<func_name> (<type1> * <arg1>, ...) (space between pointer and type declaration)

<func_name> (<type1> <arg1>,
             <type2> <arg2>,
                       ...
                      )

The following define statement can be added before including munit.h
#define MUNIT_ENABLE_ASSERT_ALIASES
in order to get rid of the munit_ prefix in munit_assert calls, e.g. assert_int instead of munit_assert_int