Toochain update: new download url and MPICH/OpenMPI version (#5088)

* Add url download function * add -O option for wget * using url downloader in stage4 * update mpich version * Update README.md * disable rapidjson by default related to #5069 * add information for abacus_env.sh * modify info of abacus_env.sh * add information for gcc toolchain * update OpenMPI version to 5.0.5 * Update README.md * change the download url for openmpi * Update README.md * fix link of openmpi * fix openmpi url * Update install_libtorch.sh * update date of update --------- Co-authored-by: kirk0830 <[email protected]>
deepmodeling · Sep 14, 2024 · c8c0a10 · c8c0a10
1 parent 03c2582
commit c8c0a10
Show file tree

Hide file tree

Showing 14 changed files with 174 additions and 89 deletions.
diff --git a/toolchain/README.md b/toolchain/README.md
@@ -1,7 +1,9 @@
 # The ABACUS Toolchain
+
 Version 2024.2
 
 ## Author
+
 [QuantumMisaka](https://github.com/QuantumMisaka) 
 (Zhaoqing Liu) @PKU @AISI
 
@@ -25,7 +27,7 @@ and give setup files that you can use to compile ABACUS.
 - [x] Support for [LibRI](https://github.com/abacusmodeling/LibRI) by submodule or automatic installation from github.com (but installed LibRI via `wget` seems to have some problem, please be cautious)
 - [x] A mirror station by Bohrium database, which can download CEREAL, LibNPY, LibRI and LibComm by `wget` in China Internet. 
 - [x] Support for GPU compilation, users can add `-DUSE_CUDA=1` in builder scripts.
-- [ ] A better mirror station for all packages.
+- [ ] Change the downloading url from cp2k mirror to other mirror or directly downloading from official website. (doing)
 - [ ] A better README and Detail markdown file.
 - [ ] Automatic installation of [DEEPMD](https://github.com/deepmodeling/deepmd-kit).
 - [ ] Better compliation method for ABACUS-DEEPMD and ABACUS-DEEPKS.
@@ -35,11 +37,11 @@ and give setup files that you can use to compile ABACUS.
 
 
 ## Usage Online & Offline
+
 Main script is *install_abacus_toolchain.sh*, 
 which will use scripts in *scripts* directory 
 to compile install dependencies of ABACUS.
-
-You can just `./install_abacus_toolchain.sh -h` to get more help message.
+It can be directly used, but not recommended.
 
 There are also well-modified script to run *install_abacus_toolchain.sh* for `gnu-openblas` and `intel-mkl` toolchains dependencies.
 
@@ -51,6 +53,7 @@ There are also well-modified script to run *install_abacus_toolchain.sh* for `gn
 # for intel-mkl-mpich
 > ./toolchain_intel-mpich.sh
 ```
+
 It is recommended to run them first to get a fast installation of ABACUS under certain environments.
 
 If you have a fresh environments and you have `sudo` permission, you can use *install_requirements.sh* to install system libraries and dependencies needed by toolchain.
@@ -66,6 +69,7 @@ If you have a fresh environments and you have `sudo` permission, you can use *in
 ```
 
 All packages will be downloaded from [cp2k-static/download](https://www.cp2k.org/static/downloads). by  `wget` , and will be detailedly compiled and installed in `install` directory by toolchain scripts, despite of:
+
 - `CEREAL` which will be downloaded from [CEREAL](https://github.com/USCiLab/cereal)  
 - `Libnpy` which will be downloaded from [LIBNPY](https://github.com/llohse/libnpy)
 - `LibRI` which will be downloaded from [LibRI](https://github.com/abacusmodeling/LibRI)
@@ -77,12 +81,17 @@ Instead of github.com, we offer other package station, you can use it by:
 ```shell
 wget https://bohrium-api.dp.tech/ds-dl/abacus-deps-93wi-v3 -O abacus-deps-v3.zip
 ```
-`unzip` it ,and you can do offline installation of these packages above after rename. The above station will be updated handly but one should notice that the version will always lower than github repo.
+`unzip` it ,and you can do offline installation of these packages above after rename. 
+```shell
+# packages downloaded from github.com
+mv v1.3.2.tar.gz build/cereal-1.3.2.tar.gz
+```
+The above station will be updated handly but one should notice that the version will always lower than github repo.
 
 If one want to install ABACUS by toolchain OFFLINE, 
 one can manually download all the packages from [cp2k-static/download](https://www.cp2k.org/static/downloads) or official website
 and put them in *build* directory by formatted name
-like *fftw-3.3.10.tar.gz*, or *openmpi-5.0.3.tar.gz*, 
+like *fftw-3.3.10.tar.gz*, or *openmpi-5.0.5.tar.bz2*, 
 then run this toolchain. 
 All package will be detected and installed automatically. 
 Also, one can install parts of packages OFFLINE and parts of packages ONLINE
@@ -96,10 +105,11 @@ just by using this toolchain
 ```
 
 The needed dependencies version default:
+
 - `cmake` 3.30.0
 - `gcc` 13.2.0 (which will always NOT be installed, But use system)
-- `OpenMPI` 5.0.3
-- `MPICH` 4.1.2
+- `OpenMPI` 5.0.5
+- `MPICH` 4.2.2
 - `OpenBLAS` 0.3.27 (Intel toolchain need `get_vars.sh` tool from it)
 - `ScaLAPACK` 2.2.1
 - `FFTW` 3.3.10
@@ -111,28 +121,21 @@ And Intel-oneAPI need user or server manager to manually install from Intel.
 [Intel-oneAPI](https://www.intel.cn/content/www/cn/zh/developer/tools/oneapi/toolkits.html)
 
 Dependencies below are optional， which is NOT installed by default:
+
 - `LibTorch` 2.1.2
 - `Libnpy` 1.0.1
 - `LibRI` 0.2.0
 - `LibComm` 0.1.1
-Users can install them by using `--with-*=install` in toolchain*.sh, which is `no` in default.
-> Notice: LibRI, LibComm and Libnpy is on actively development, you should check-out the package version when using this toolchain. Also, LibRI and LibComm can be installed by github submodule, which is also work for libnpy, which is more recommended.
 
-Notice: for `CEREAL`,`RapidJSON`, `Libnpy`, `LibRI` and `LibComm`, 
-you need to download them from github.com, 
-rename it as formatted, and put them in `build` directory at the same time
-e.g.:
-```shell
-# packages downloaded from github.com
-mv v1.3.2.tar.gz build/cereal-1.3.2.tar.gz
-```
+Users can install them by using `--with-*=install` in toolchain*.sh, which is `no` in default.
+> Notice: LibRI, LibComm and Libnpy is on actively development, you should check-out the package version when using this toolchain. Also, LibRI and LibComm can be installed by github submodule, that is also work for libnpy, which is more recommended.
 
 Users can easily compile and install dependencies of ABACUS
 by running these scripts after loading `gcc` or `intel-mkl-mpi`
 environment. 
 
 The toolchain installation process can be interrupted at anytime.
-just re-run *install_abacus_toolchain.sh*, toolchain itself may fix it
+just re-run *toolchain_\*.sh*, toolchain itself may fix it
 
 If compliation is successful, a message will be shown like this:
 
@@ -147,6 +150,7 @@ If compliation is successful, a message will be shown like this:
 >     ./build_abacus_intel.sh
 > or you can modify the builder scripts to suit your needs.
 ```
+
 You can run *build_abacus_gnu.sh* or *build_abacus_intel.sh* to build ABACUS 
 by gnu-toolchain or intel-toolchain respectively, the builder scripts will
 automatically locate the environment and compile ABACUS.
@@ -160,17 +164,28 @@ If users want to use toolchain but lack of some system library
 dependencies, *install_requirements.sh* scripts will help.
 
 If users want to re-install all the package, just do:
+
 ```shell
 > rm -rf install
 ```
+
 or you can also do it in a more completely way:
+
 ```shell
 > rm -rf install build/*/* build/OpenBLAS*/ build/setup_*
 ```
 
-## Common Problem and Solution
+## Common Problems and Solutions
+
+### LibRI and LibComm for EXX
+
+- GCC toolchain with OpenMPI cannot compile LibComm v0.1.1 due to the different MPI variable type from MPICH and IntelMPI, see discussion here [#5033](https://github.com/deepmodeling/abacus-develop/issues/5033), you can switch to GCC-MPICH or Intel toolchain
+- It is recommended to use Intel toolchain if one wants to include EXX feature in ABACUS, which can have much better performance and can use more than 16 threads in OpenMP parallelization to accelerate the EXX process.
+
 ### GPU version of ABACUS
-add following options in build*.sh:
+
+For GPU version of ABACUS (do not GPU version installer of ELPA, which is still doing work), add following options in build*.sh:
+
 ```shell
 cmake -B $BUILD_DIR -DCMAKE_INSTALL_PREFIX=$PREFIX \
         -DCMAKE_CXX_COMPILER=icpx \
@@ -180,46 +195,59 @@ cmake -B $BUILD_DIR -DCMAKE_INSTALL_PREFIX=$PREFIX \
         -DCMAKE_CUDA_COMPILER=${path to cuda toolkit}/bin/nvcc \
         ......
 ```
-Notice: You CANNOT use `icpx` compiler for GPU version of ABACUS for now
 
-### shell problem
+Notice: You CANNOT use `icpx` compiler for GPU version of ABACUS for now, see discussion here [#2906](https://github.com/deepmodeling/abacus-develop/issues/2906) and [#4976](https://github.com/deepmodeling/abacus-develop/issues/4976)
+
+
+### Shell problem
+
 If you encounter problem like:
+
 ```shell
 /bin/bash^M: bad interpreter: No such file or directory
 ```
+
 or   `permission denied` problem, you can simply run:
+
 ```shell
 ./pre_set.sh
 ```
+
 And also, you can fix `permission denied` problem via `chmod +x`
 if *pre_set.sh* have no execution permission; 
 if the *pre_set.sh* also have `/bin/bash^M` problem, you can run:
-```
+
+```shell
 > dos2unix pre_set.sh
 ```
+
 to fix it
 
-### libtorch and deepks problem
+### Libtorch and DeePKS problem
+
 If deepks feature have problem, you can manually change libtorch version
-from 2.0.1 to 1.12.0 in `toolchain/scripts/stage4/install_libtorch.sh`.
+from 2.1.2 to 2.0.1 or 1.12.0 in `toolchain/scripts/stage4/install_libtorch.sh`.
+
 Also, you can install ABACUS without deepks by removing all the deepks and related options.
 
 NOTICE: if you want deepks feature, your intel-mkl environment should be accessible in building process. you can check it in `build_abacus_gnu.sh`
 
-### deepmd feature problem
+### DeePMD feature problem
+
 When you encounter problem like `GLIBCXX_3.4.29 not found`, it is sure that your `gcc` version is lower than the requirement of `libdeepmd`.
 
 After my test, you need `gcc`>11.3.1 to enable deepmd feature in ABACUS.
 
 ### ELPA problem via Intel-oneAPI toolchain in AMD server
+
 The default compiler for Intel-oneAPI is `icpx` and `icx`, which will cause problem when compling ELPA in AMD server. (Which is a problem and needed to have more check-out)
 
 The best way is to change `icpx` to `icpc`, `icx` to `icc`. user can manually change it in toolchain*.sh via `--with-intel-classic=yes`
 
-Notice: `icc` and `icpc` from Intel Classic Compiler of Intel-oneAPI is not supported for 2024.0 and newer version.
-
+Notice: `icc` and `icpc` from Intel Classic Compiler of Intel-oneAPI is not supported for 2024.0 and newer version. And Intel-OneAPI 2023.2.0 can be found in website. See discussion here [#4976](https://github.com/deepmodeling/abacus-develop/issues/4976)
 
 ### Intel-oneAPI problem
+
 Sometimes Intel-oneAPI have problem to link `mpirun`, 
 which will always show in 2023.2.0 version of MPI in Intel-oneAPI. 
 Try `source /path/to/setvars.sh` or install another version of IntelMPI may help.
@@ -229,7 +257,6 @@ And will not occur in Intel-MPI before 2021.10.0 (Intel-oneAPI before 2023.2.0)
 
 More problem and possible solution can be accessed via [#2928](https://github.com/deepmodeling/abacus-develop/issues/2928)
 
-
 ## Advanced Installation Usage
 
 1. Users can move toolchain directory to anywhere you like, 
@@ -243,4 +270,5 @@ of each packages, which may let the installation more fiexible.
 
 
 ## More
+
 More infomation can be read from `Details.md`.
diff --git a/toolchain/build_abacus_gnu.sh b/toolchain/build_abacus_gnu.sh
@@ -47,7 +47,7 @@ cmake -B $BUILD_DIR -DCMAKE_INSTALL_PREFIX=$PREFIX \
         -DENABLE_LIBXC=ON \
         -DUSE_OPENMP=ON \
         -DUSE_ELPA=ON \
-        -DENABLE_RAPIDJSON=ON \
+        -DENABLE_RAPIDJSON=OFF \
 #         -DENABLE_DEEPKS=1 \
 #         -DTorch_DIR=$LIBTORCH \
 #         -Dlibnpy_INCLUDE_DIR=$LIBNPY \
@@ -74,3 +74,12 @@ cat << EOF > "${TOOL}/abacus_env.sh"
 source $INSTALL_DIR/setup
 export PATH="${PREFIX}/bin":\${PATH}
 EOF
+
+# generate information
+cat << EOF
+========================== usage =========================
+Done!
+To use the installed ABACUS version
+You need to source $(pwd)/abacus_env.sh first !
+"""
+EOF
diff --git a/toolchain/build_abacus_intel-mpich.sh b/toolchain/build_abacus_intel-mpich.sh
@@ -43,7 +43,7 @@ cmake -B $BUILD_DIR -DCMAKE_INSTALL_PREFIX=$PREFIX \
         -DENABLE_LIBXC=ON \
         -DUSE_OPENMP=ON \
         -DUSE_ELPA=ON \
-        -DENABLE_RAPIDJSON=ON \
+        -DENABLE_RAPIDJSON=OFF \
 #       -DENABLE_DEEPKS=1 \
 #       -DTorch_DIR=$LIBTORCH \
 #       -Dlibnpy_INCLUDE_DIR=$LIBNPY \
@@ -64,4 +64,13 @@ cat << EOF > "${TOOL}/abacus_env.sh"
 #!/bin/bash
 source $INSTALL_DIR/setup
 export PATH="${PREFIX}/bin":\${PATH}
+EOF
+
+# generate information
+cat << EOF
+========================== usage =========================
+Done!
+To use the installed ABACUS version
+You need to source $(pwd)/abacus_env.sh first !
+"""
 EOF
diff --git a/toolchain/build_abacus_intel.sh b/toolchain/build_abacus_intel.sh
@@ -44,7 +44,7 @@ cmake -B $BUILD_DIR -DCMAKE_INSTALL_PREFIX=$PREFIX \
         -DENABLE_LIBXC=ON \
         -DUSE_OPENMP=ON \
         -DUSE_ELPA=ON \
-        -DENABLE_RAPIDJSON=ON \
+        -DENABLE_RAPIDJSON=OFF \
 #         -DENABLE_DEEPKS=1 \
 #         -DTorch_DIR=$LIBTORCH \
 #         -Dlibnpy_INCLUDE_DIR=$LIBNPY \
@@ -66,3 +66,12 @@ cat << EOF > "${TOOL}/abacus_env.sh"
 source $INSTALL_DIR/setup
 export PATH="${PREFIX}/bin":\${PATH}
 EOF
+
+# generate information
+cat << EOF
+========================== usage =========================
+Done!
+To use the installed ABACUS version
+You need to source $(pwd)/abacus_env.sh first !
+"""
+EOF
diff --git a/toolchain/scripts/stage1/install_mpich.sh b/toolchain/scripts/stage1/install_mpich.sh
@@ -3,15 +3,17 @@
 # TODO: Review and if possible fix shellcheck errors.
 # shellcheck disable=all
 
-# Last Update in 2023-0918
+# Last Update in 2024-0912
 
 [ "${BASH_SOURCE[0]}" ] && SCRIPT_NAME="${BASH_SOURCE[0]}" || SCRIPT_NAME=$0
 SCRIPT_DIR="$(cd "$(dirname "$SCRIPT_NAME")/.." && pwd -P)"
 
 # mpich_ver="4.0.3"
 # mpich_sha256="17406ea90a6ed4ecd5be39c9ddcbfac9343e6ab4f77ac4e8c5ebe4a3e3b6c501"
-mpich_ver="4.1.2"
-mpich_sha256="3492e98adab62b597ef0d292fb2459b6123bc80070a8aa0a30be6962075a12f0"
+# mpich_ver="4.1.2"
+# mpich_sha256="3492e98adab62b597ef0d292fb2459b6123bc80070a8aa0a30be6962075a12f0"
+mpich_ver="4.2.2"
+mpich_sha256="883f5bb3aeabf627cb8492ca02a03b191d09836bbe0f599d8508351179781d41"
 mpich_pkg="mpich-${mpich_ver}.tar.gz"
 
 source "${SCRIPT_DIR}"/common_vars.sh
@@ -35,13 +37,15 @@ case "${with_mpich}" in
     pkg_install_dir="${INSTALLDIR}/mpich-${mpich_ver}"
     #pkg_install_dir="${HOME}/apps/mpich/${mpich_ver}-intel"
     install_lock_file="$pkg_install_dir/install_successful"
+    url="https://www.mpich.org/static/downloads/${mpich_ver}/${mpich_pkg}"
     if verify_checksums "${install_lock_file}"; then
       echo "mpich-${mpich_ver} is already installed, skipping it."
     else
       if [ -f ${mpich_pkg} ]; then
         echo "${mpich_pkg} is found"
       else
-        download_pkg_from_ABACUS_org "${mpich_sha256}" "${mpich_pkg}"
+        #download_pkg_from_ABACUS_org "${mpich_sha256}" "${mpich_pkg}"
+        download_pkg_from_url "${mpich_sha256}" "${mpich_pkg}" "${url}"
       fi
       echo "Installing from scratch into ${pkg_install_dir} for MPICH device ${MPICH_DEVICE}"
       [ -d mpich-${mpich_ver} ] && rm -rf mpich-${mpich_ver}

diff --git a/toolchain/scripts/stage1/install_openmpi.sh b/toolchain/scripts/stage1/install_openmpi.sh
@@ -3,13 +3,15 @@
 # TODO: Review and if possible fix shellcheck errors.
 # shellcheck disable=all
 
-# Last Update in 2024-0811
+# Last Update in 2024-0912
 
 [ "${BASH_SOURCE[0]}" ] && SCRIPT_NAME="${BASH_SOURCE[0]}" || SCRIPT_NAME=$0
 SCRIPT_DIR="$(cd "$(dirname "$SCRIPT_NAME")/.." && pwd -P)"
 
-openmpi_ver="5.0.3"
-openmpi_sha256="990582f206b3ab32e938aa31bbf07c639368e4405dca196fabe7f0f76eeda90b"
+openmpi_ver="5.0.5"
+openmpi_sha256="6588d57c0a4bd299a24103f4e196051b29e8b55fbda49e11d5b3d32030a32776"
+# openmpi_ver="4.1.6"
+# openmpi_sha256="f740994485516deb63b5311af122c265179f5328a0d857a567b85db00b11e415"
 openmpi_pkg="openmpi-${openmpi_ver}.tar.bz2"
 
 source "${SCRIPT_DIR}"/common_vars.sh
@@ -33,13 +35,14 @@ case "${with_openmpi}" in
     pkg_install_dir="${INSTALLDIR}/openmpi-${openmpi_ver}"
     #pkg_install_dir="${HOME}/apps/openmpi/${openmpi_ver}-gcc8"
     install_lock_file="$pkg_install_dir/install_successful"
+    url="https://download.open-mpi.org/release/open-mpi/v${openmpi_ver:0:3}/${openmpi_pkg}"
     if verify_checksums "${install_lock_file}"; then
       echo "openmpi-${openmpi_ver} is already installed, skipping it."
     else
       if [ -f ${openmpi_pkg} ]; then
         echo "${openmpi_pkg} is found"
       else
-        download_pkg_from_ABACUS_org "${openmpi_sha256}" "${openmpi_pkg}"
+        download_pkg_from_url "${openmpi_sha256}" "${openmpi_pkg}" "${url}"
       fi
       echo "Installing from scratch into ${pkg_install_dir}"
       [ -d openmpi-${openmpi_ver} ] && rm -rf openmpi-${openmpi_ver}
@@ -59,6 +62,8 @@ case "${with_openmpi}" in
         fi
       fi
     # OpenMPI 5.0 only supports PMIx
+    # PMI support is required for Slurm, but not for other schedulers
+    # default not use
     #   if [ $(command -v srun) ]; then
     #     echo "Slurm installation found. OpenMPI will be configured with --with-pmi."
     #     EXTRA_CONFIGURE_FLAGS="--with-pmi"