Add ROCm 6.4 support

.github/workflows/build-image.yml

@@ -80,6 +80,12 @@ jobs:
           platforms: linux/amd64
           runner: ubuntu-latest
           build_args: "NUM_MAKE_JOBS=8"
+        - name: rocm6.4
+          dockerfile: rocm6.4.x
+          tags: superbench/main:rocm6.4
+          platforms: linux/amd64
+          runner: [self-hosted, linux/amd64, rocm]
+          build_args: "NUM_MAKE_JOBS=16"
         # - name: rocm6.2
         #   dockerfile: rocm6.2.x
         #   tags: superbench/main:rocm6.2

dockerfile/rocm6.2.x.dockerfile

@@ -130,7 +130,7 @@ RUN cd /tmp && \
 
 # Install Intel MLC
 RUN cd /tmp && \
-    wget -q https://downloadmirror.intel.com/763324/mlc_v3.10.tgz -O mlc.tgz && \
+    wget -q https://downloadmirror.intel.com/866182/mlc_v3.12.tgz -O mlc.tgz && \
     tar xzf mlc.tgz Linux/mlc && \
     cp ./Linux/mlc /usr/local/bin/ && \
     rm -rf ./Linux mlc.tgz

dockerfile/rocm6.4.x.dockerfile

@@ -0,0 +1,215 @@
+ARG BASE_IMAGE=rocm/pytorch:rocm6.4.4_ubuntu24.04_py3.12_pytorch_release_2.7.1
+
+FROM ${BASE_IMAGE}
+
+# OS:
+#   - Ubuntu: 24.04
+#   - Docker Client: 20.10.8
+# ROCm:
+#   - ROCm: 6.4
+# Lib:
+#   - torch: 2.7.1
+#   - rccl: release/rocm-rel-6.4
+#   - hipblaslt: release-staging/rocm-rel-6.4
+#   - rocblas: release-staging/rocm-rel-6.4
+#   - openmpi: 4.1.x
+# Intel:
+#   - mlc: v3.12
+
+LABEL maintainer="SuperBench"
+
+ENV DEBIAN_FRONTEND=noninteractive
+RUN apt-get update && \
+    apt-get -q install -y --no-install-recommends  \
+    autoconf \
+    automake \
+    bc \
+    build-essential \
+    curl \
+    dmidecode \
+    git \
+    hipify-clang \
+    iproute2 \
+    jq \
+    libaio-dev \
+    libboost-program-options-dev \
+    libcap2 \
+    libcurl4-openssl-dev \
+    libnuma-dev \
+    libpci-dev \
+    libssl-dev \
+    libtinfo6 \
+    libtool \
+    lshw \
+    net-tools \
+    numactl \
+    openssh-client \
+    openssh-server \
+    pciutils \
+    python3-mpi4py \
+    rsync \
+    sudo \
+    util-linux \
+    vim \
+    wget \
+    && \
+    rm -rf /tmp/*
+
+ARG NUM_MAKE_JOBS=64
+
+# Check if CMake is installed and its version
+RUN cmake_version=$(cmake --version 2>/dev/null | grep -oP "(?<=cmake version )(\d+\.\d+)" || echo "0.0") && \
+    required_version="3.24.1" && \
+    if [ "$(printf "%s\n" "$required_version" "$cmake_version" | sort -V | head -n 1)" != "$required_version" ]; then \
+    echo "existing cmake version is ${cmake_version}" && \
+    cd /tmp && \
+    wget -q https://github.com/Kitware/CMake/releases/download/v${required_version}/cmake-${required_version}.tar.gz && \
+    tar xzf cmake-${required_version}.tar.gz && \
+    cd cmake-${required_version} && \
+    ./bootstrap --prefix=/usr --no-system-curl --parallel=16 && \
+    make -j ${NUM_MAKE_JOBS} && \
+    make install && \
+    rm -rf /tmp/cmake-${required_version}* \
+    else \
+    echo "CMake version ${cmake_version} is greater than or equal to ${required_version}"; \
+    fi
+
+# Install Docker
+ENV DOCKER_VERSION=20.10.8
+RUN cd /tmp && \
+    wget -q https://download.docker.com/linux/static/stable/x86_64/docker-${DOCKER_VERSION}.tgz -O docker.tgz && \
+    tar --extract --file docker.tgz --strip-components 1 --directory /usr/local/bin/ && \
+    rm docker.tgz
+
+# Update system config
+RUN mkdir -p /root/.ssh && \
+    touch /root/.ssh/authorized_keys && \
+    mkdir -p /var/run/sshd && \
+    sed -i "s/[# ]*PermitRootLogin prohibit-password/PermitRootLogin yes/" /etc/ssh/sshd_config && \
+    sed -i "s/[# ]*PermitUserEnvironment no/PermitUserEnvironment yes/" /etc/ssh/sshd_config && \
+    sed -i "s/[# ]*Port.*/Port 22/" /etc/ssh/sshd_config && \
+    echo "* soft nofile 1048576\n* hard nofile 1048576" >> /etc/security/limits.conf && \
+    echo "root soft nofile 1048576\nroot hard nofile 1048576" >> /etc/security/limits.conf
+
+
+# Set Ubuntu version
+ENV UBUNTU_VERSION=24.04
+
+# Install OFED
+ENV OFED_VERSION=24.10-1.1.4.0
+# Check if ofed_info is present and has a version
+RUN if ! command -v ofed_info >/dev/null 2>&1; then \
+    echo "OFED not found. Installing OFED..."; \
+    cd /tmp && \
+    wget -q http://content.mellanox.com/ofed/MLNX_OFED-${OFED_VERSION}/MLNX_OFED_LINUX-${OFED_VERSION}-ubuntu${UBUNTU_VERSION}-x86_64.tgz && \
+    tar xzf MLNX_OFED_LINUX-${OFED_VERSION}-ubuntu${UBUNTU_VERSION}-x86_64.tgz && \
+    PATH=/usr/bin:${PATH} MLNX_OFED_LINUX-${OFED_VERSION}-ubuntu${UBUNTU_VERSION}-x86_64/mlnxofedinstall --user-space-only --without-fw-update --force --all && \
+    rm -rf MLNX_OFED_LINUX-${OFED_VERSION}* ; \
+    fi
+
+ENV ROCM_PATH=/opt/rocm
+
+# Install OpenMPI
+ENV OPENMPI_VERSION=4.1.x
+ENV MPI_HOME=/usr/local/mpi
+# Check if Open MPI is installed
+RUN cd /tmp && \
+    git clone --recursive https://github.com/open-mpi/ompi.git -b v${OPENMPI_VERSION}  && \
+    cd ompi && \
+    ./autogen.pl && \
+    mkdir build && \
+    cd build && \
+    ../configure --prefix=/usr/local/mpi  --enable-orterun-prefix-by-default --enable-mpirun-prefix-by-default  --enable-prte-prefix-by-default --with-rocm=/opt/rocm && \
+    make -j $(nproc) && \
+    make -j $(nproc) install && \
+    ldconfig && \
+    cd / && \
+    rm -rf /tmp/openmpi-${OPENMPI_VERSION}*
+
+# Install Intel MLC
+RUN cd /tmp && \
+    wget -q https://downloadmirror.intel.com/866182/mlc_v3.12.tgz -O mlc.tgz && \
+    tar xzf mlc.tgz Linux/mlc && \
+    cp ./Linux/mlc /usr/local/bin/ && \
+    rm -rf ./Linux mlc.tgz
+
+# Install RCCL
+RUN cd /opt/ &&  \
+    git clone -b release/rocm-rel-6.4 https://github.com/ROCmSoftwarePlatform/rccl.git && \
+    cd rccl && \
+    mkdir build && \
+    cd build && \
+    CXX=/opt/rocm/bin/hipcc cmake -DHIP_COMPILER=clang -DCMAKE_BUILD_TYPE=Release -DCMAKE_VERBOSE_MAKEFILE=1 \
+    -DCMAKE_PREFIX_PATH="${ROCM_PATH}/hsa;${ROCM_PATH}/hip;${ROCM_PATH}/share/rocm/cmake/;${ROCM_PATH}" \
+    .. && \
+    make -j${NUM_MAKE_JOBS}
+
+# Install AMD SMI Python Library
+RUN apt install amd-smi-lib -y && \
+    cd /opt/rocm/share/amd_smi && \
+    python3 -m pip install .
+
+ENV PATH="/usr/local/mpi/bin:/opt/superbench/bin:/usr/local/bin/:/opt/rocm/hip/bin/:/opt/rocm/bin/:${PATH}" \
+    LD_LIBRARY_PATH="/opt/rccl/build:/usr/local/mpi/lib:/opt/rocm/lib:/usr/local/lib/:${LD_LIBRARY_PATH}" \
+    SB_HOME=/opt/superbench \
+    SB_MICRO_PATH=/opt/superbench \
+    ANSIBLE_DEPRECATION_WARNINGS=FALSE \
+    ANSIBLE_COLLECTIONS_PATH=/usr/share/ansible/collections
+
+RUN echo PATH="$PATH" > /etc/environment && \
+    echo LD_LIBRARY_PATH="$LD_LIBRARY_PATH" >> /etc/environment && \
+    echo SB_MICRO_PATH="$SB_MICRO_PATH" >> /etc/environment
+
+RUN apt install rocm-cmake -y && \
+    python3 -m pip install --upgrade pip wheel "setuptools>=69.0"
+
+WORKDIR ${SB_HOME}
+
+ADD third_party third_party
+# perftest_rocm6.patch changes are already upstream in the submodule version
+# rocm_megatron_lm: broken upstream (pretrain_deepseek.py missing in rocm_dev branch)
+# apex_rocm: skipped — all apex imports in Megatron-DeepSpeed are guarded with try/except,
+#   superbench has zero direct apex usage, and PyTorch 2.7 has native fused optimizers/AMP.
+RUN make RCCL_HOME=/opt/rccl/build/ ROCBLAS_BRANCH=release-staging/rocm-rel-6.4 HIPBLASLT_BRANCH=release-staging/rocm-rel-6.4 ROCM_VER=rocm-5.5.0 -C third_party rocm -o cpu_hpl -o cpu_stream -o megatron_lm -o rocm_hipblaslt -o rocm_megatron_lm -o apex_rocm
+# Build hipblaslt separately with Tensile target-triple fix for ROCm 6.4 clang
+# Also fix joblib race condition (github.com/joblib/joblib/issues/1788) in Python 3.12:
+# joblib's _retrieve() iterates _jobs_set while callbacks modify it.
+# Fix: copy the set before iterating. Patch all joblib instances system-wide.
+RUN pip install "joblib>=1.4.2" && \
+    find / -path '*/joblib/parallel.py' -not -path '*/.git/*' -exec sed -i \
+        's/timeout_control_job = next(iter(self\._jobs_set), None)/timeout_control_job = next(iter(set(self._jobs_set)), None)/' {} +
+RUN cd third_party && \
+    git clone -b release-staging/rocm-rel-6.4 https://github.com/ROCmSoftwarePlatform/hipBLASLt.git && \
+    sed -i 's/host-x86_64-unknown-linux,/host-x86_64-unknown-linux-gnu,/' \
+        hipBLASLt/tensilelite/Tensile/BuildCommands/SharedCommands.py && \
+    cd hipBLASLt && ./install.sh -dc && \
+    find /opt -path '*/joblib/parallel.py' -not -path '*/.git/*' -exec sed -i \
+        's/timeout_control_job = next(iter(self\._jobs_set), None)/timeout_control_job = next(iter(set(self._jobs_set)), None)/' {} + && \
+    cp -v build/release/clients/staging/hipblaslt-bench /opt/superbench/bin/
+RUN cp -r /opt/superbench/third_party/hipBLASLt/build/release/hipblaslt-install/lib/*  /opt/rocm/lib/ && \
+    cp -r /opt/superbench/third_party/hipBLASLt/build/release/hipblaslt-install/include/*  /opt/rocm/include/
+RUN cd third_party/Megatron/Megatron-DeepSpeed && \
+    git apply ../megatron_deepspeed_rocm6.patch
+
+# Install TransformerEngine - pin to 386bd316 (before NVFP4/hip_fp4.h which needs ROCm 7.0+).
+# Disable CK fused attention (aiter submodule has gfx950-only code); aotriton stays enabled.
+RUN git clone --recursive https://github.com/ROCm/TransformerEngine.git && \
+    cd TransformerEngine && \
+    git checkout 386bd316 && \
+    git submodule update --init --recursive && \
+    NVTE_FRAMEWORK=pytorch \
+    NVTE_FUSED_ATTN_CK=0 \
+    NVTE_ROCM_ARCH=gfx942 \
+    python3 setup.py install
+RUN python3 -c "import transformer_engine.pytorch; print('TE installed successfully')"
+
+ADD . .
+ENV USE_HIP_DATATYPE=1
+ENV USE_HIPBLAS_COMPUTETYPE=1
+RUN python3 -m pip install .[amdworker]  && \
+    CXX=/opt/rocm/bin/hipcc make cppbuild  && \
+    make postinstall
+
+# Fix stale hypothesis plugin from base image (imports removed pkg_resources)
+# and add test dependencies missing from the base image.
+RUN python3 -m pip install --upgrade hypothesis setuptools pytest-timeout vcrpy

superbench/benchmarks/micro_benchmarks/gpu_stream.py

@@ -117,3 +117,4 @@ def _process_raw_result(self, cmd_idx, raw_output):
 
 
 BenchmarkRegistry.register_benchmark('gpu-stream', GpuStreamBenchmark, platform=Platform.CUDA)
+BenchmarkRegistry.register_benchmark('gpu-stream', GpuStreamBenchmark, platform=Platform.ROCM)

superbench/benchmarks/micro_benchmarks/gpu_stream/CMakeLists.txt

@@ -11,34 +11,114 @@ set(CMAKE_CXX_STANDARD_REQUIRED ON)
 set(CMAKE_CUDA_STANDARD 17)
 set(CMAKE_CUDA_STANDARD_REQUIRED ON)
 
+# Source files (CUDA-style names; hipify-perl translates them on ROCm path)
+set(GPU_STREAM_SOURCES
+    gpu_stream_test.cpp
+    gpu_stream_utils.cpp
+    gpu_stream.cu
+    gpu_stream_kernels.cu
+)
+set(GPU_STREAM_HEADERS
+    gpu_stream.hpp
+    gpu_stream_utils.hpp
+    gpu_stream_kernels.hpp
+)
+
 find_package(CUDAToolkit QUIET)
 
-if(NOT CUDAToolkit_FOUND)
-    message(WARNING "gpu_stream: CUDA not found, skipping build (requires NVIDIA GPU with NVML)")
+if(CUDAToolkit_FOUND)
+    # NVIDIA / CUDA path
+    find_library(NVML_LIBRARY nvidia-ml PATHS ${CUDAToolkit_LIBRARY_DIR} PATH_SUFFIXES stubs)
+    if(NOT NVML_LIBRARY)
+        message(WARNING "gpu_stream: NVML (nvidia-ml) not found, skipping build")
+        return()
+    endif()
+
+    message(STATUS "Found CUDA: ${CUDAToolkit_VERSION}")
+    include(../cuda_common.cmake)
+    add_executable(gpu_stream ${GPU_STREAM_SOURCES})
+    set_property(TARGET gpu_stream PROPERTY CUDA_ARCHITECTURES ${NVCC_ARCHS_SUPPORTED})
+    target_include_directories(gpu_stream PRIVATE ${CUDAToolkit_INCLUDE_DIRS})
+    target_link_libraries(gpu_stream numa ${NVML_LIBRARY})
+    install(TARGETS gpu_stream RUNTIME DESTINATION bin)
     return()
 endif()
 
-# Check for NVML (nvidia-ml) library, required for querying memory clock rates
-find_library(NVML_LIBRARY nvidia-ml PATHS ${CUDAToolkit_LIBRARY_DIR} PATH_SUFFIXES stubs)
-if(NOT NVML_LIBRARY)
-    message(WARNING "gpu_stream: NVML (nvidia-ml) not found, skipping build")
+# ROCm / HIP path
+include(../rocm_common.cmake)
+find_package(hip QUIET)
+
+if(NOT hip_FOUND)
+    message(WARNING "gpu_stream: neither CUDA nor HIP found, skipping build")
     return()
 endif()
 
-message(STATUS "Found CUDA: " ${CUDAToolkit_VERSION})
+message(STATUS "Found ROCm: ${HIP_VERSION}")
+enable_language(HIP)
 
-# Source files
-set(SOURCES
-    gpu_stream_test.cpp
-    gpu_stream_utils.cpp
-    gpu_stream.cu
-    gpu_stream_kernels.cu
+# rocm_smi is required for actual memory clock queries (NVML equivalent).
+find_path(ROCM_SMI_INCLUDE_DIR rocm_smi/rocm_smi.h
+    HINTS ${ROCM_PATH} /opt/rocm
+    PATH_SUFFIXES include
 )
+find_library(ROCM_SMI_LIBRARY
+    NAMES rocm_smi64
+    HINTS ${ROCM_PATH} /opt/rocm
+    PATH_SUFFIXES lib lib64
+)
+if(NOT ROCM_SMI_INCLUDE_DIR OR NOT ROCM_SMI_LIBRARY)
+    message(WARNING "gpu_stream: rocm_smi not found, skipping build")
+    return()
+endif()
+message(STATUS "Found rocm_smi: ${ROCM_SMI_LIBRARY}")
+
+# Translate the CUDA-named sources/headers to HIP via hipify-perl into the
+# build directory. hipify-perl translates cuda* identifiers and CUDA includes
+# but leaves NVML/rocm_smi symbols untouched (they are gated by #ifdef in the
+# source).
+find_program(HIPIFY_PERL hipify-perl REQUIRED)
+
+set(HIPIFY_OUTPUT_DIR ${CMAKE_CURRENT_BINARY_DIR}/hipified)
+file(MAKE_DIRECTORY ${HIPIFY_OUTPUT_DIR})
+
+set(HIP_SOURCES "")
+set(HIPIFY_OUTPUTS "")
+foreach(SRC IN LISTS GPU_STREAM_SOURCES GPU_STREAM_HEADERS)
+    set(SRC_IN ${CMAKE_CURRENT_SOURCE_DIR}/${SRC})
+    set(SRC_OUT ${HIPIFY_OUTPUT_DIR}/${SRC})
+    add_custom_command(
+        OUTPUT ${SRC_OUT}
+        COMMAND ${HIPIFY_PERL} -o ${SRC_OUT} ${SRC_IN}
+        DEPENDS ${SRC_IN}
+        COMMENT "hipify-perl ${SRC}"
+        VERBATIM
+    )
+    list(APPEND HIPIFY_OUTPUTS ${SRC_OUT})
+    if(SRC MATCHES "\\.(cpp|cu)$")
+        list(APPEND HIP_SOURCES ${SRC_OUT})
+    endif()
+endforeach()
 
-include(../cuda_common.cmake)
-add_executable(gpu_stream ${SOURCES})
-set_property(TARGET gpu_stream PROPERTY CUDA_ARCHITECTURES ${NVCC_ARCHS_SUPPORTED})
-target_include_directories(gpu_stream PRIVATE ${CUDAToolkit_INCLUDE_DIRS})
-target_link_libraries(gpu_stream numa ${NVML_LIBRARY})
+# Aggregate target so all hipified outputs (including headers) are produced
+# before any object is compiled.
+add_custom_target(gpu_stream_hipify DEPENDS ${HIPIFY_OUTPUTS})
+
+set_source_files_properties(${HIP_SOURCES} PROPERTIES LANGUAGE HIP)
+
+include(CheckSymbolExists)
+check_symbol_exists("hipDeviceMallocUncached" "hip/hip_runtime_api.h" HIP_UNCACHED_MEMORY)
+if(HIP_UNCACHED_MEMORY)
+    add_compile_definitions(HIP_UNCACHED_MEMORY)
+endif()
+
+set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -O2")
+
+add_executable(gpu_stream ${HIP_SOURCES})
+add_dependencies(gpu_stream gpu_stream_hipify)
+target_include_directories(gpu_stream PRIVATE
+    ${HIPIFY_OUTPUT_DIR}
+    ${ROCM_SMI_INCLUDE_DIR}
+)
+target_link_libraries(gpu_stream numa hip::device ${ROCM_SMI_LIBRARY})
 
 install(TARGETS gpu_stream RUNTIME DESTINATION bin)