The Side-by-Side install feature does not behave as expected, and multiple release installations of AMD ROCm fail. Attempts to install AMD ROCm v3.5.0 and AMD ROCm v3.5.1 side-by-side can result in breaking the existing ROCm v3.5.0 installation.
This known issue is currently under investigation.
AMD ROCm released a maintenance patch release v3.5.1. For more information, see
https://github.com/RadeonOpenCompute/ROCm/blob/master/AMD_ROCm_Release_Notes_v3.5.1.pdf
This page describes the features, fixed issues, and information about downloading and installing the ROCm software. It also covers known issues and deprecated features in the ROCm v3.5.0 release.
Machine Learning and High Performance Computing Software Stack for AMD GPU
The AMD ROCm v3.5.x platform is designed to support the following operating systems:
Access the following links for more information on:
For AMD ROCm documentation, see
For installation instructions on supported platforms, see
https://rocmdocs.amd.com/en/latest/Installation_Guide/Installation-Guide.html
For AMD ROCm binary structure, see
For AMD ROCm Release History, see
You must perform a fresh and a clean AMD ROCm install to successfully upgrade from v3.3 to v3.5. The following changes apply in this release:
SQLite3 is a required Python module for the rocprof command-line tool. You can install the SQLite3 Python module using the pip utility and set env var ROCP_PYTHON_VERSION to the Python version, which includes the SQLite3 module.
In this release, the Heterogeneous Compute Compiler (HCC) compiler is deprecated and the HIP-Clang compiler is introduced for compiling Heterogeneous-Compute Interface for Portability (HIP) programs.
NOTE: The HCC environment variables will be gradually deprecated in subsequent releases.
The majority of the codebase for the HIP-Clang compiler has been upstreamed to the Clang trunk. The HIP-Clang implementation has undergone a strict code review by the LLVM/Clang community and comprehensive tests consisting of LLVM/Clang build bots. These reviews and tests resulted in higher productivity, code quality, and lower cost of maintenance.
For most HIP applications, the transition from HCC to HIP-Clang is transparent and efficient as the HIPCC and HIP cmake files automatically choose compilation options for HIP-Clang and hide the difference between the HCC and HIP-Clang code. However, minor changes may be required as HIP-Clang has a stricter syntax and semantic checks compared to HCC.
NOTE: Native HCC language features are no longer supported.
In this release, the HIP runtime API is implemented on top of Radeon Open Compute Common Language Runtime (ROCclr). ROCclr is an abstraction layer that provides the ability to interact with different runtime backends such as ROCr.
The following OpenCL runtime changes are made in this release:
The AMD ROCm Debugger (ROCgdb) is the AMD ROCm source-level debugger for Linux based on the GNU Debugger (GDB). It enables heterogeneous debugging on the AMD ROCm platform of an x86-based host architecture along with AMD GPU architectures and supported by the AMD Debugger API Library (ROCdbgapi).
The AMD ROCm Debugger is installed by the rocm-gdb package. The rocm-gdb package is part of the rocm-dev meta-package, which is in the rocm-dkms package.
The current AMD ROCm Debugger (ROCgdb) is an initial prototype that focuses on source line debugging. Note, symbolic variable debugging capabilities are not currently supported.
You can use the standard GDB commands for both CPU and GPU code debugging. For more information about ROCgdb, refer to the ROCgdb User Guide, which is installed at:
The AMD ROCm Debugger User Guide is available as a PDF at:
https://github.com/RadeonOpenCompute/ROCm/blob/master/gdb.pdf
For more information about GNU Debugger (GDB), refer to the GNU Debugger (GDB) web site at: http://www.gnu.org/software/gdb
The AMD ROCm Debugger API Library (ROCdbgapi) implements an AMD GPU debugger application programming interface (API) that provides the support necessary for a client of the library to control the execution and inspect the state of AMD GPU devices.
The following AMD GPU architectures are supported:
The AMD ROCm Debugger API Library is installed by the rocm-dbgapi package. The rocm-gdb package is part of the rocm-dev meta-package, which is in the rocm-dkms package. The AMD ROCm Debugger API Specification is available as a PDF at:
https://github.com/RadeonOpenCompute/ROCm/blob/master/amd-dbgapi.pdf
In this release, a new rocprofiler start/stop API is added to enable/disable GPU kernel HSA dispatch callbacks. The callback can be registered with the 'rocprofiler_set_hsa_callbacks' API. The API helps you eliminate some profiling performance impact by invoking the profiler only for kernel dispatches of interest. This optimization will result in significant performance gains.
The API provides the following functions:
For more information on kernel dispatches, see the HSA Platform System Architecture Specification guide at http://www.hsafoundation.com/standards/.
The ROCm Communications Collective Library (RCCL) consists of the following enhancements:
AMD RCCL is now compatible with NVIDIA Communications Collective Library (NCCL) v2.6.4 and provides the following features:
MIOpen provides an optional pre-compiled kernel package to reduce startup latency.
NOTE: The installation of this package is optional. MIOpen will continue to function as expected even if you choose to not install the pre-compiled kernel package. This is because MIOpen compiles the kernels on the target machine once the kernel is run. However, the compilation step may significantly increase the startup time for different operations.
To install the kernel package for your GPU architecture, use the following command:
apt-get install miopen-kernels--
A Systems Management Interface (SMI) event interface is added to the kernel and ROCm SMI lib for system administrators to get notified when specific events occur. On the kernel side, AMDKFD_IOC_SMI_EVENTS input/output control is enhanced to allow notifications propagation to user mode through the event channel.
On the ROCm SMI lib side, APIs are added to set an event mask and receive event notifications with a timeout option. Further, ROCm SMI API details can be found in the PDF generated by Doxygen from source or by referring to the rocm_smi.h header file (see the rsmi_event_notification_* functions).
For the more details about ROCm SMI API, see
https://github.com/RadeonOpenCompute/ROCm/blob/master/ROCm_SMI_Manual.pdf
A new API is introduced for aiding applications to select the appropriate memory node for a given accelerator(GPU).
The API for CPU affinity has the following signature:
*rsmi_status_t rsmi_topo_numa_affinity_get(uint32_t dv_ind, uint32_t numa_node);
This API takes as input, device index (dv_ind), and returns the NUMA node (CPU affinity), stored at the location pointed by numa_node pointer, associated with the device.
Non-Uniform Memory Access (NUMA) is a computer memory design used in multiprocessing, where the memory access time depends on the memory location relative to the processor.
The new Radeon Performance Primitives (RPP) library is a comprehensive high-performance computer vision library for AMD (CPU and GPU) with the HIP and OpenCL backend. The target operating system is Linux.
For more information about prerequisites and library functions, see
https://github.com/GPUOpen-ProfessionalCompute-Libraries/MIVisionX/tree/master/docs
HIP now supports the use of printf in the device code. The parameters and return value for the device-side printf follow the POSIX.1 standard, with the exception that the "%n" specifier is not supported. A call to printf blocks the calling wavefront until the operation is completely processed by the host.
No host-side runtime calls by the application are needed to cause the output to appear. There is also no limit on the number of device-side calls to printf or the amount of data that is printed.
For more details, refer the HIP Programming Guide at: https://rocmdocs.amd.com/en/latest/Programming_Guides/HIP-GUIDE.html#hip-guide
Previously, a failing assertion caused early termination of kernels and the application to exit with a line number, file, and failing condition printed to the screen. This issue is now fixed and the assert() and abort() functions are implemented for HIP device code. NOTE: There may be a performance impact in the use of device assertions in its current form.
You may choose to disable the assertion in the production code. For example, to disable an assertion of:
assert(foo != 0);
you may comment it out as:
//assert(foo != 0);
NOTE: Assertions are currently enabled by default.
The following are the known issues in the v3.5 release.
HIPify-Clang fails to install on CentOS/RHEL with the following error:
file from install of hipify-clang conflicts with file from package hip-base
Workaround: This is a known issue and the following workaround is recommended for a successful installation of HIPify-Clang on CentOS/RHEL:
Download HIPify-Clang RPM. For example, hipify-clang-11.0.0.x86_64.rpm
Perform a force install using the following command:
sudo rpm -ivh --force hipify-clang-11.0.0.x86_64.rpm
When ROCgdb is in non-stop mode with an application that rapidly creates and destroys queues, a breakpoint may be reported that is not processed by the debugger before the queue is deleted. In some cases, this can result in the following error that prevents further debugging:
[amd-dbgapi]: fatal error: kfd_queue_id 2 should have been reported as a NEW_QUEUE before next_pending_event failed (rc=-2)
There are no known workarounds at this time.
When the ROCdbgapi library is used with an application that rapidly creates and destroys queues, a breakpoint may be reported that is not processed by the client before the queue is deleted. In some cases, this can result in a fatal error and the following error log message is produced:
[amd-dbgapi]: fatal error: kfd_queue_id 2 should have been reported as a NEW_QUEUE before next_pending_event failed (rc=-2)
There are no known workarounds at this time.
The following unit test failures have been observed due to known issues in the ROCclr runtime.
rocThrust
hipCUB
There are no known workarounds in the current release.
A random freeze has been observed with Imagenet training and tf_cnn_benchmark on TensorFlow when multiple GPU configurations are involved.
There is no freeze observed with single GPUs.
There are no known workarounds at this time.
Running AMD ROCm v3.3 in the user mode with the AMD ROCm v3.5 DKMS kernel module will cause the following features to be broken:
Resolution: Install ROCm v3.5 Thunk (Hsakmt) when using ROCm 3.5 Kernel Fusion Driver (KFD).
The ROCProfiler tool appears to be broken when the SQLite3 library is not found.
Resolution: Install the SQLite3 Python module separately and ensure the environment variable is set to ROCP_PYTHON_VERSION to confirm the Python version, which includes the SQLite3 module.
Install ROCm v3.5 Thunk (Hsakmt) when using ROCm 3.5 Kernel Fusion Driver (KFD). You can access the Thunk package at:
https://github.com/RadeonOpenCompute/ROCT-Thunk-Interface
In this release, the Heterogeneous Compute Compiler (HCC) compiler is deprecated and the HIP-Clang compiler is introduced for compiling Heterogeneous-Compute Interface for Portability (HIP) programs.
For more information, see HIP documentation at: https://rocmdocs.amd.com/en/latest/Programming_Guides/Programming-Guides.html
AMD hosts both Debian and RPM repositories for the ROCm v3.5.x packages.
For more information on ROCM installation on all platforms, see
https://rocmdocs.amd.com/en/latest/Installation_Guide/Installation-Guide.html
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 supports a targeted set of hardware configurations which are detailed further in this section.
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 use following chips:
ROCm is a collection of software ranging from drivers and runtimes 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 enabled in the ROCm software, though full support is not guaranteed:
As described in the next section, GFX8 GPUs require PCI Express 3.0 (PCIe 3.0) with support for PCIe atomics. This requires both CPU and motherboard support. GFX9 GPUs require PCIe 3.0 with support for PCIe atomics by default, but they can operate in most cases without this capability.
The integrated GPUs in AMD APUs are not officially supported targets for ROCm. As described below, "Carrizo", "Bristol Ridge", and "Raven Ridge" APUs are enabled in our upstream drivers and the ROCm OpenCL runtime. However, they are not enabled in our HCC or HIP runtimes, and may not work due to motherboard or OEM hardware limitations. As such, they are not yet officially supported targets for ROCm.
For a more detailed list of hardware support, please see the following documentation.
As described above, GFX8 GPUs require PCIe 3.0 with PCIe atomics in order to run ROCm. In particular, the CPU and every active PCIe point between the CPU and GPU require support for PCIe 3.0 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:
Beginning with ROCm 1.8, GFX9 GPUs (such as Vega 10) no longer require PCIe atomics. 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 PCIe 2.0. This is not supported on GPUs below GFX9, e.g. GFX8 cards in the 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 Broadcom PLX. When you install your GPUs, make sure you install them in a PCIe 3.1.0 x16, 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 2.0 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.
As of ROCm 1.9.0, the ROCm user-level software is compatible with the AMD drivers in certain upstream Linux kernels. As such, users have the option of either using the ROCK kernel driver that are part of AMD's ROCm repositories or using the upstream driver and only installing ROCm user-level utilities from AMD's ROCm repositories.
These releases of the upstream Linux kernel support the following GPUs in ROCm:
The upstream driver may be useful for running ROCm software on systems that are not compatible with the kernel driver available in AMD's repositories. For users that have the option of using either AMD's or the upstreamed driver, there are various tradeoffs to take into consideration:
Using AMD's rock-dkms package |
Using the upstream kernel driver | |
---|---|---|
Pros | More GPU features, and they are enabled earlier | Includes the latest Linux kernel features |
Tested by AMD on supported distributions | May work on other distributions and with custom kernels | |
Supported GPUs enabled regardless of kernel version | ||
Includes the latest GPU firmware | ||
Cons | May not work on all Linux distributions or versions | Features and hardware support varies depending on kernel version |
Not currently supported on kernels newer than 5.4 | Limits GPU's usage of system memory to 3/8 of system memory (before 5.6). For 5.6 and beyond, both DKMS and upstream kernels allow use of 15/16 of system memory. | |
IPC and RDMA capabilities are not yet enabled | ||
Not tested by AMD to the same level as rock-dkms package |
||
Does not include most up-to-date firmware |
For an updated version of the software stack for AMD GPU, see
此处可能存在不合适展示的内容,页面不予展示。您可通过相关编辑功能自查并修改。
如您确认内容无涉及 不当用语 / 纯广告导流 / 暴力 / 低俗色情 / 侵权 / 盗版 / 虚假 / 无价值内容或违法国家有关法律法规的内容,可点击提交进行申诉,我们将尽快为您处理。