Skip to content

Latest commit

 

History

History
433 lines (303 loc) · 19.4 KB

ROCmInstall.md

File metadata and controls

433 lines (303 loc) · 19.4 KB
layout title
default
ROCm Install

ROCm Platform Installation Guide for Linux

Current ROCm Version: 1.9.0

Hardware Support

ROCm is focused on using AMD GPUs to accelerate computational tasks, such as machine learning, engineering workloads, and scientific computing. In order to focus our development efforts on these domains of interest, ROCm

Supported GPUs

Because the ROCm Platform has a focus on particular computational domains, we offer official support for a selection of AMD GPUs that are designed to offer good performance and price in these domains.

ROCm officially supports AMD GPUs that have use following chips:

  • GFX8 GPUs
    • "Fiji" chips, such as on the the AMD Radeon R9 Fury X and Radeon Instinct MI8
    • "Polaris 10" chips, such as on the AMD Radeon RX 580 and Radeon Instinct MI6
    • "Polaris 11" chips, such as on the AMD Radeon RX 570 and Radeon Pro WX 4100
  • GFX9 GPUs
    • "Vega 10" chips, such as on the AMD Radeon Radeon RX Vega 64 and Radeon Instinct MI25

ROCm is a collection of software ranging from drivers and runtimnes to libraries and developer tools. Some of this software may work with more GPUs than the "officially supported" list above, though AMD does not make any official claims of support for these devices on the ROCm software platform. The following list of GPUs are likely to work within ROCm, though full support is not guaranteed:

  • GFX7 GPUs
    • "Hawaii" chips, such as the AMD Radeon R9 390X and FirePro W9100

As described in the next section, GFX8 GPUs require PCIe gen 3 with support for PCIe atomics. This requires both CPU and motherboard support. GFX9 GPUs, by default, also require PCIe gen 3 with support for PCIe atomics; if you want to avoid using PCIe atomics, please set the environment variable HSA_ENABLE_SDMA=0. GFX7 GPUs do not require PCIe atomics.

At this time, the integrated GPUs in AMD APUs are not officially supported targets for ROCm.

For a more detailed list of hardware support, please see the following documentation.

Supported CPUs

As described above, GFX8 and GFX9 GPUs require PCI Express 3.0 with PCIe atomics in the default ROCm configuration. In particular, the CPU and every active PCIe point between the CPU and GPU require support for PCIe gen 3 and PCIe atomics. The CPU root must indicate PCIe AtomicOp Completion capabilities and any intermediate switch must indicate PCIe AtomicOp Routing capabilities.

Current CPUs which support PCIe Gen3 + PCIe Atomics are:

  • AMD Ryzen CPUs;
  • AMD Ryzen APUs;
  • AMD Ryzen Threadripper CPUs
  • AMD EPYC CPUs;
  • Intel Xeon E7 v3 or newer CPUs;
  • Intel Xeon E5 v3 or newer CPUs;
  • Intel Xeon E3 v3 or newer CPUs;
  • Intel Core i7 v4, Core i5 v4, Core i3 v4 or newer CPUs (i.e. Haswell family or newer).

Beginning with ROCm 1.8, we have relaxed the requirements for PCIe Atomics on GFX9 GPUs such as Vega 10. We have similarly opened up more options for number of PCIe lanes. GFX9 GPUs can now be run on CPUs without PCIe atomics and on older PCIe generations such as gen 2. To enable this option, please set the environment variable HSA_ENABLE_SDMA=0. This is not supported on GPUs below GFX9, e.g. GFX8 cards in Fiji and Polaris families.

If you are using any PCIe switches in your system, please note that PCIe Atomics are only supported on some switches, such as Boradcom PLX. When you install your GPUs, make sure you install them in a fully PCIe Gen3 x16 or x8, x4 or x1 slot attached either directly to the CPU's Root I/O controller or via a PCIe switch directly attached to the CPU's Root I/O controller.

In our experience, many issues stem from trying to use consumer motherboards which provide physical x16 connectors that are electrically connected as e.g. PCIe Gen2 x4, PCIe slots connected via the Southbridge PCIe I/O controller, or PCIe slots connected through a PCIe switch that does not support PCIe atomics.

If you attempt to run ROCm on a system without proper PCIe atomic support, you may see an error in the kernel log (dmesg):

kfd: skipped device 1002:7300, PCI rejects atomics

Experimental support for our Hawaii (GFX7) GPUs (Radeon R9 290, R9 390, FirePro W9100, S9150, S9170) does not require or take advantage of PCIe Atomics. However, we still recommend that you use a CPU from the list provided above for compatibility purposes.

Reminder: if you are using gfx9 GPUs, you can bypass this requirement by setting the environment variable HSA_ENABLE_SDMA=0. However, this disables the use of DMA engines to move data between the CPU and GPU memory. This can reduce performance.

Not supported or very limited support under ROCm

Limited support
  • ROCm 1.9 and Vega10 should support PCIe Gen2 enabled CPUs such as the AMD Opteron, Phenom, Phenom II, Athlon, Athlon X2, Athlon II and older Intel Xeon and Intel Core Architecture and Pentium CPUs. However, we have done very limited testing on these configurations, since our test farm has been catering to CPU listed above. This is where we need community support; if you find problems on such setups, please report these issues.
  • Thunderbolt 1, 2, and 3 enabled breakout boxes should now be able to work with ROCm. Thunderbolt 1 and 2 are PCIe Gen2 based, and thus are only supported with GPUs that do not require PCIe Gen 3 atomics (i.e. Vega 10). However, we have done no testing on this configuration and would need comunity support due to limited access to this type of equipment
Not supported
  • "Tonga", "Iceland", "Polaris 12", and "Vega M" GPUs are not supported in ROCm 1.9.0
  • We do not support GFX8-class GPUs (Fiji, Polaris, etc.) on CPUs that do not have PCIe Gen 3 with PCIe atomics.
    • As such, do not support AMD Carrizo and Kaveri APUs as hosts for such GPUs..
    • Thunderbolt 1 and 2 enabled GPUs are not supported by GFX8 GPUs on ROCm. Thunderbolt 1 & 2 are PCIe Gen2 based.
  • AMD Carrizo based APUs have limited support due to OEM & ODM's choices when it comes to some key configuration parameters. In particular, we have observed that Carrizo laptops, AIOs, and desktop systems showed inconsistencies in exposing and enabling the System BIOS parameters required by the ROCm stack. Before purchasing a Carrizo system for ROCm, please verify that the BIOS provides an option for enabling IOMMUv2 and that the system BIOS properly exposes the correct CRAT table - please inquire with the OEM about the latter.
  • AMD Merlin/Falcon Embedded System is not currently supported by the public repo.
  • AMD Raven Ridge APU are currently not supported

Software Support

The latest tested version of the drivers, tools, libraries and source code for the ROCm platform have been released and are available under the roc-1.9.0 or rocm-1.9.x tag of the following GitHub repositories:

Additionally, the following mirror repositories that support the HCC compiler are also available on GitHub, and frozen for the rocm-1.9.0 release:

Supported Operating Systems - New operating systems available

The ROCm 1.9.0 platform has been tested on the following operating systems:

  • Ubuntu 16.04 &. 18.04
  • CentOS 7.4 &. 7.5 (Using devetoolset-7 runtime support)
  • RHEL 7.4. &. 7.5 (Using devetoolset-7 runtime support)

Installing from AMD ROCm repositories

AMD is hosting both Debian and RPM repositories for the ROCm 1.9.0 packages at this time.

The packages in the Debian repository have been signed to ensure package integrity.

Ubuntu Support - installing from a Debian repository

First make sure your system is up to date
sudo apt update
sudo apt dist-upgrade
sudo apt install libnuma-dev
sudo reboot
Add the ROCm apt repository

For Debian based systems, like Ubuntu, configure the Debian ROCm repository as follows:

wget -qO - http://repo.radeon.com/rocm/apt/debian/rocm.gpg.key | sudo apt-key add -
echo 'deb [arch=amd64] http://repo.radeon.com/rocm/apt/debian/ xenial main' | sudo tee /etc/apt/sources.list.d/rocm.list

The gpg key might change, so it may need to be updated when installing a new release. If the key signature verification fails when you attempt to update, please re-add the key from ROCm apt repository. The current rocm.gpg.key is not avialable in a standard key ring distribution, but has the following sha1sum hash:

f7f8147431c75e505c58a6f3a3548510869357a6 rocm.gpg.key

Install

Next, update the apt repository list and install the ROCm package:

Warning: Before proceeding, make sure to completely uninstall any previous ROCm package:

sudo apt update
sudo apt install rocm-dkms
Next set your permissions

With move to upstreaming the KFD driver and the support of DKMS, for all Console aka headless user, you will need to add all your users to the 'video" group by setting the Unix permissions

Configure Ensure that your user account is a member of the "video" group prior to using the ROCm driver. You can find which groups you are a member of with the following command:

groups

To add yourself to the video group you will need the sudo password and can use the following command:

sudo usermod -a -G video $LOGNAME 

You may want to ensure that any future users you add to your system are put into the "video" group by default. To do that, you can run the following commands:

echo 'ADD_EXTRA_GROUPS=1' | sudo tee -a /etc/adduser.conf
echo 'EXTRA_GROUPS=video' | sudo tee -a /etc/adduser.conf

Once complete, reboot your system.

Upon Reboot run the following commands to verify that the ROCm installation waas successful. If you see your GPUs listed by both of these commands, you should be ready to go!

/opt/rocm/bin/rocminfo 
/opt/rocm/opencl/bin/x86_64/clinfo 

Note that, to make running ROCm programs easier, you may wish to put the ROCm libraries in your LD_LIBRARY_PATH environment variable and the ROCm binaries in your PATH.

echo 'export LD_LIBRARY_PATH=/opt/rocm/opencl/lib/x86_64:/opt/rocm/hsa/lib:$LD_LIBRARY_PATH' | sudo tee -a /etc/profile.d/rocm.sh
echo 'export PATH=$PATH:/opt/rocm/bin:/opt/rocm/profiler/bin:/opt/rocm/opencl/bin/x86_64' | sudo tee -a /etc/profile.d/rocm.sh

If you have an Install Issue please read this FAQ .

Vega10 users who want to run ROCm on a system that does not support PCIe atomics must set HSA_ENABLE_SDMA=0

Currently, if you want to run ROCm on a Vega10 GPU (GFX9) on a system without PCIe atomics, you must turn off SDMA functionality.

export HSA_ENABLE_SDMA=0
Performing an OpenCL-only Installation of ROCm

Some users may want to install a subset of the full ROCm installation. In particular, if you are trying to install on a system with a limited amount of storage space, or which will only run a small collection of known applications, you may want to install only the packages that are required to run OpenCL applications. To do that, you can run the following installation command instead of the command to install rocm-dkms.

sudo apt-get install dkms rock-dkms rocm-opencl
Upon restart, to test your OpenCL instance

Build and run Hello World OCL app.

HelloWorld sample:

 wget https://raw.githubusercontent.com/bgaster/opencl-book-samples/master/src/Chapter_2/HelloWorld/HelloWorld.cpp
 wget https://raw.githubusercontent.com/bgaster/opencl-book-samples/master/src/Chapter_2/HelloWorld/HelloWorld.cl

Build it using the default ROCm OpenCL include and library locations:

g++ -I /opt/rocm/opencl/include/ ./HelloWorld.cpp -o HelloWorld -L/opt/rocm/opencl/lib/x86_64 -lOpenCL

Run it:

./HelloWorld
How to un-install from Ubuntu 16.04 or Ubuntu 18.04

To un-install the entire rocm development package execute:

sudo apt autoremove rocm-dkms
Installing development packages for cross compilation

It is often useful to develop and test on different systems. In this scenario, you may prefer to avoid installing the ROCm Kernel to your development system.

In this case, install the development subset of packages:

sudo apt update
sudo apt install rocm-dev

Note: To execute ROCm enabled apps you will require a system with the full ROCm driver stack installed

Removing pre-release packages

It is recommended to remove previous rocm installations before installing the latest version to ensure a smooth installation.

If you installed any of the ROCm pre-release packages from github, they will need to be manually un-installed:

sudo apt purge hsakmt-roct
sudo apt purge hsakmt-roct-dev
sudo apt purge compute-firmware
sudo apt purge $(dpkg -l | grep 'kfd\|rocm' | grep linux | grep -v libc | awk '{print $2}')

If possible, we would recommend starting with a fresh OS install.

CentOS/RHEL 7 (both 7.4 and 7.5) Support

Support for CentOS/RHEL 7 has been added in ROCm 1.8, but requires a special runtime environment provided by the RHEL Software Collections and additional dkms support packages to properly install in run.

Preparing RHEL 7 for installation

RHEL is a subscription based operating system, and must enable several external repositories to enable installation of the devtoolset-7 environment and the DKMS support files. These steps are not required for CentOS.

First, the subscription for RHEL must be enabled and attached to a pool id. Please see Obtaining an RHEL image and license page for instructions on registering your system with the RHEL subscription server and attaching to a pool id.

Second, enable the following repositories:

sudo subscription-manager repos --enable rhel-server-rhscl-7-rpms
sudo subscription-manager repos --enable rhel-7-server-optional-rpms
sudo subscription-manager repos --enable rhel-7-server-extras-rpms

Third, enable additional repositories by downloading and installing the epel-release-latest-7 repository RPM:

sudo rpm -ivh https://dl.fedoraproject.org/pub/epel/epel-release-latest-7.noarch.rpm
Install and setup Devtoolset-7

To setup the Devtoolset-7 environment, follow the instructions on this page:

https://www.softwarecollections.org/en/scls/rhscl/devtoolset-7/

Note that devtoolset-7 is a Software Collections package, and is not supported by AMD.

Prepare CentOS/RHEL 7.4 or 7.5 for DKMS Install

Installing kernel drivers on CentOS/RHEL 7.4/7.5 requires dkms tool being installed:

sudo yum install -y epel-release
sudo yum install -y dkms kernel-headers-`uname -r` kernel-devel-`uname -r`
Installing ROCm on the system

It is recommended to remove previous rocm installations before installing the latest version to ensure a smooth installation.

At this point ROCm can be installed on the target system. Create a /etc/yum.repos.d/rocm.repo file with the following contents:

[ROCm]
name=ROCm
baseurl=http://repo.radeon.com/rocm/yum/rpm
enabled=1
gpgcheck=0

The repo's URL should point to the location of the repositories repodata database. Install ROCm components using these commands:

sudo yum install rocm-dkms

The rock-dkms component should be installed and the /dev/kfd device should be available on reboot.

Ensure that your user account is a member of the "video" or "wheel" group prior to using the ROCm driver. You can find which groups you are a member of with the following command:

groups

To add yourself to the video (or wheel) group you will need the sudo password and can use the following command:

sudo usermod -a -G video $LOGNAME 

Current release supports up to CentOS/RHEL 7.4 and 7.5. Users should update to the latest version of the OS:

sudo yum update
Vega10 users who want to run ROCm on a system that does not support PCIe atomics must set HSA_ENABLE_SDMA=0

Currently, if you want to run ROCm on a Vega10 GPU (GFX9) on a system without PCIe atomics, you must turn off SDMA functionality.

export HSA_ENABLE_SDMA=0
Performing an OpenCL-only Installation of ROCm

Some users may want to install a subset of the full ROCm installation. In particular, if you are trying to install on a system with a limited amount of storage space, or which will only run a small collection of known applications, you may want to install only the packages that are required to run OpenCL applications. To do that, you can run the following installation command instead of the command to install rocm-dkms.

sudo yum install rock-dkms rocm-opencl
Compiling applications using hcc, hip, etc.

To compile applications or samples, please use gcc-7.2 provided by the devtoolset-7 environment. To do this, compile all applications after running this command:

scl enable devtoolset-7 bash
How to un-install ROCm from CentOS/RHEL 7.4 and 7.5

To un-install the entire rocm development package execute:

sudo yum autoremove rocm-dkms

Known Issues / Workarounds

Radeon Compute Profiler does not run

rcprof -A <HSA_application> fails with error message: Radeon Compute Profiler could not be enabled. Version mismatch between HSA runtime and libhsa-runtime-tools64.so.1.

Running OCLPerfCounters test results in LLVM ERROR: out of memory

HipCaffe is supported on single GPU configurations

The ROCm SMI library calls to rsmi_dev_power_cap_set() and rsmi_dev_power_profile_set() will not work for all but the first gpu in multi-gpu set ups.

Vega10 users who want to run ROCm on a system that does not support PCIe atomics must set HSA_ENABLE_SDMA=0

Currently, if you want to run ROCm on a Vega10 GPU (GFX9) on a system without PCIe atomics, you must turn off SDMA functionality.

export HSA_ENABLE_SDMA=0