Re: [AMBER] Question regarding QUICK support with HIP/AMD GPUs in Amber26

From: Bauer, Paul via AMBER <amber.ambermd.org>
Date: Fri, 13 Nov 2026 15:22:54 +0000

AMD General

Hello,

ok, the issue is with your ROCm version. There had been issues with the ROCm 6 series (before ROCm 7 and after some earlier version) where the code in QUICK caused problems during kernel code generation, so we agreed with the QUICK developers to disable it for those ROCm versions. Can you use a newer ROCm on the system you have access to?

Otherwise your configuration looks good.

Cheers

Paul

-----Original Message-----
From: jourdain.cines.fr <jourdain.cines.fr>
Sent: Friday, November 13, 2026 4:18 PM
To: Bauer, Paul <Paul.Bauer.amd.com>
Cc: AMBER Mailing List <amber.ambermd.org>
Subject: Re: [AMBER] Question regarding QUICK support with HIP/AMD GPUs in Amber26

Caution: This message originated from an External Source. Use proper caution when opening attachments, clicking links, or responding.


Dear Paul,

Thanks for your quick answer. I compile Amber26, with ambertools26.tar (it seems to be QUICK_VERSION='QUICK-26.03'). I used rocm/6.4.3. If you have a fonctionnal compilations script for an Amber26 compiled with Quick for AMD GPUs, I would be happy to have a look.


Here the main line of my configuration, you'll find my compilations scripts with patch in attachment:

module load cpe/25.09
module load craype-x86-trento
module load PrgEnv-gnu/8.6.0
module load rocm/6.4.3
module load cray-fftw
module load python/3.12 cmake
export MKL_PATH=/path/to/intel/oneapi/mkl/2022.1.0

export HIPCC_LINK_FLAGS_APPEND="--offload-arch=gfx90a --offload-arch=gfx942 --no-default-config"
export HIPCC_COMPILE_FLAGS_APPEND="--offload-arch=gfx90a --offload-arch=gfx942 --no-default-config"
export GPU_ARCHS="gfx90a;gfx942"

cmake \
    -D CMAKE_BUILD_TYPE=Release \
    -D COMPILER=AUTO \
    -D CMAKE_C_COMPILER=gcc \
    -D CMAKE_CXX_COMPILER=g++ \
    -D CMAKE_Fortran_COMPILER=gfortran \
    -D CMAKE_CXX_FLAGS="-O3" \
    -D CMAKE_C_FLAGS="-O3" \
    -D CMAKE_Fortran_FLAGS="-O3" \
    -D DISABLE_WARNINGS=ON \
    -D USE_FFT=ON \
    -D BUILD_PYTHON=OFF \
    -D CHECK_UPDATES=OFF \
    -D DOWNLOAD_MINICONDA=OFF \
    -D HIP=ON \
    -D GTI=TRUE \
    -D VKFFT=ON \
    -D MPI=ON \
    -D HIP_RDC=ON \
    -D HIP_TOOLKIT_ROOT_DIR="$ROCM_PATH" \
    -D HIPCUDA_EMULATE_VERSION="10.1" \
    -D BUILD_HOST_TOOLS=ON \
    -D CMAKE_INSTALL_PREFIX="$PREFIX_DIR/tools" \
    -D CMAKE_PROJECT_INCLUDE_BEFORE="$(realpath "$PATCH_SCRIPT")" \
    -S . \
    -B build/tools


cmake \
    --build build/tools \
    --target install \
    -j48


and then

echo "======================="
echo " 3. Compile pmemd.hip (and pmemd.hip.MPI), the AMD GPU-accelerated molecular dynamics executable, as well as the GPU version of cpptraj, using the HIP macros."
echo "======================="
cmake \
    -D CMAKE_BUILD_TYPE=Release \
    -D COMPILER=AUTO \
    -D CMAKE_C_COMPILER=cc \
    -D CMAKE_CXX_COMPILER=CC \
    -D CMAKE_Fortran_COMPILER=ftn \
    -D MPI_C_COMPILER=cc\
    -D MPI_CXX_COMPILER=CC\
    -D MPI_Fortran_COMPILER=ftn\
    -D CMAKE_CXX_FLAGS="-O3" \
    -D CMAKE_C_FLAGS="-O3" \
    -D CMAKE_Fortran_FLAGS="-O3" \
    -D CMAKE_EXE_LINKER_FLAGS="-Wl,--disable-new-dtags -Wl,--unresolved-symbols=ignore-all" \
    -D CMAKE_SHARED_LINKER_FLAGS="-Wl,--disable-new-dtags -Wl,--unresolved-symbols=ignore-all" \
    -D DISABLE_WARNINGS=ON \
    -D USE_FFT=ON \
    -D FFTW_FOUND=TRUE \
    -D FFTW_HOME="$FFTW_ROOT" \
    -D FFTW_INCLUDE_DIR="$FFTW_INC" \
    -D FFTW_LIBRARY="$FFTW_DIR/libfftw3.so" \
    -D BUILD_PYTHON=OFF \
    -D CHECK_UPDATES=OFF \
    -D DOWNLOAD_MINICONDA=OFF \
    -D BUILD_QUICK=ON \
    -D QUICK_USER_ARCH=gfx90a \
    -D HIP=ON \
    -D USE_HIP=ON \
    -D GTI=TRUE \
    -D HIP_RDC=ON \
    -D CMAKE_HIP_ARCHITECTURES=gfx90a \
    -D VKFFT=ON \
    -D HIP_TOOLKIT_ROOT_DIR="$ROCM_PATH" \
    -D FORCE_EXTERNAL_LIBS=mkl \
    -D MKL_HOME="$MKL_PATH" \
    -D MKL_CORE_LIBRARY="$MKL_PATH/lib/intel64/libmkl_core.so" \
    -D MKL_SEQUENTIAL_THREADING_LIBRARY="$MKL_PATH/lib/intel64/libmkl_sequential.so" \
    -D MKL_INTERFACE_LIBRARY="$MKL_PATH/lib/intel64/libmkl_intel_lp64.so" \
    -D MKL_GFORTRAN_INTERFACE_LIBRARY="$MKL_PATH/lib/intel64/libmkl_gf_lp64.so" \
    -D MKL_GNU_THREADING_LIBRARY="$MKL_PATH/lib/intel64/libmkl_gnu_thread.so" \
    -D MKL_WORKS=YES \
    -D HIPCUDA_EMULATE_VERSION="10.1" \
    -D BUILD_HOST_TOOLS=OFF \
    -D USE_HOST_TOOLS=ON \
    -D HOST_TOOLS_DIR="$PREFIX_DIR/tools" \
    -D CMAKE_INSTALL_PREFIX="$PREFIX_DIR" \
    -D MPI_HOME="$MPICH_DIR" \
    -D MPI_mpi_LIBRARY="$MPICH_DIR/lib/libmpi.so" \
    -D MPI_C_INCLUDE_PATH="$MPICH_DIR/include" \
    -D MPI_CXX_INCLUDE_PATH="$MPICH_DIR/include" \
    -D MPI_C_LIBRARIES="$MPICH_DIR/lib" \
    -D MPI_C_LIB_NAMES=mpi \
    -D MPI_CXX_LIB_NAMES=mpi \
    -D MPI=TRUE \
    -D FORCE_INTERNAL_LIBS=xblas \
    -D NEED_xblas=ON \
    -D CUDA=FALSE \
    -D MVAPICH2GDR_GPU_DIRECT_COMM=TRUE \
    -D CMAKE_VERBOSE_MAKEFILE=TRUE \
    -D INSTALL_TESTS=TRUE \
    -D CMAKE_PROJECT_INCLUDE_BEFORE="$(realpath "$PATCH_SCRIPT")" \
    -S . \
    -B build/amber 2>&1

echo "======================="
echo "Build Amber xblas_build"
echo "======================="

cmake \
    --build build/amber \
      --target xblas_build 2>&1

cmake -S . -B build/amber

echo "======================="
echo "Install Amber"
echo "======================="
cmake \
    --build build/amber \
    --target install \
    -j100 2>&1


Cédric Jourdain

HPC Department | CINES
+33467141469
[ http://www.cines.fr/ | www.cines.fr ]

----- Mail original -----
De: "Paul Bauer" <Paul.Bauer.amd.com>
À: "jourdain" <jourdain.cines.fr>, "AMBER Mailing List" <amber.ambermd.org>
Envoyé: Vendredi 13 Novembre 2026 15:41:27
Objet: RE: [AMBER] Question regarding QUICK support with HIP/AMD GPUs in Amber26

AMD General

Hello (AMD person supporting AMBER and QUICK here)!

Can you tell me how exactly you have build the code, and which versions you are using (both of the ROCm stack and QUICK. I see that it is from AMBER26, but exact QUICK version would help as well). I have build and successfully run QUICK with HIP support in the past, so this might be a quick fix.

Cheers

Paul

-----Original Message-----
From: Jourdain via AMBER <amber.ambermd.org>
Sent: Friday, November 13, 2026 3:38 PM
To: amber.ambermd.org
Subject: [AMBER] Question regarding QUICK support with HIP/AMD GPUs in Amber26

Caution: This message originated from an External Source. Use proper caution when opening attachments, clicking links, or responding.


Hi Amber developers,

I am currently setting up Amber on an HPC system running RHEL 9 with AMD Trento nodes and 4x AMD MI250X GPUs.

One of our users would like to run calculations using sander with QUICK on these AMD GPUs. However, I noticed that the official documentation mentions QUICK explicitly under the CUDA section ( 2.2.2. Using CUDA ), but there is no mention of it in the HIP and AMD GPUs section ( 2.2.3 ).

Furthermore, while inspecting the source code, I found this comment in pmemd26_src/cmake/SanderConfig.cmake:

$ grep -rn QUICK pmemd26_src/cmake/SanderConfig.cmake
27:# QUICK
30:#disable QUICK for HIP for now
31:option(BUILD_QUICK "Enable building QUICK and its associated sander variant." FALSE) 33:option(BUILD_QUICK "Enable building QUICK and its associated sander variant." TRUE) By applying several patches to various CMakeLists.txt files, I managed to force the compilation, but the execution fails with this error:
```
[ERROR] GPU error: hipHostAlloc failure
Program received signal SIGSEGV: Segmentation fault – invalid memory reference.
```

Given the comment in the CMake file and this runtime failure, could you confirm whether sander/Amber with QUICK is currently unsupported or not yet functional on AMD GPUs (via HIP)?

Thank you for your guidance,
Cédric Jourdain
Ingénieur HPC

Centre Informatique National de l'Enseignement Supérieur
950 rue de Saint Priest, 34097 Montpellier
0467141469
[ http://www.cines.fr/ | www.cines.fr ]

_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Fri Nov 13 2026 - 07:30:02 PST
Custom Search