2021-04-07 15:59:07 +08:00

23 KiB

Raw Blame History

AMD Open Source Driver for Vulkan®

The AMD Open Source Driver for Vulkan® is an open-source Vulkan driver for Radeon™ graphics adapters on Linux®. It is built on top of AMD's Platform Abstraction Library (PAL), a shared component that is designed to encapsulate certain hardware and OS-specific programming details for many of AMD's 3D and compute drivers. Leveraging PAL can help provide a consistent experience across platforms, including support for recently released GPUs and compatibility with AMD developer tools.

Shaders that compose a particular VkPipeline object are compiled as a single entity using the LLVM-Based Pipeline Compiler (LLPC) library. LLPC builds on LLVM's existing shader compilation infrastructure for AMD GPUs to generate code objects compatible with PAL's pipeline ABI. Notably, AMD's closed-source Vulkan driver currently uses a different pipeline compiler, which is the major difference between AMD's open-source and closed-source Vulkan drivers.

Product Support

The AMD Open Source Driver for Vulkan is designed to support the following AMD GPUs:

Radeon™ RX 6900/6800/6700 Series
Radeon™ RX 5700/5600/5500 Series
Radeon™ RX Vega Series
Radeon™ R5/R7/R9 200/300 Series
Radeon™ RX 400/500 Series
Radeon™ M200/M300/M400 Series
Radeon™ HD 8000M Series
Radeon™ HD 7000 Series
AMD FirePro™ Workstation Wx000/Wx100/Wx300 Series
Radeon™ Pro WX x100/x200 Series
Radeon™ Pro 400/500 Series
Radeon™ W5700/W5500 Series

Operating System Support

The AMD Open Source Driver for Vulkan is designed to support following distros on both the AMDGPU upstream driver stack and the AMDGPU Pro driver stack:

Ubuntu 20.04 (64-bit version)
Ubuntu 18.04 (64-bit version)
RedHat 8.2 (64-bit version)
RedHat 7.8 (64-bit version)

The driver has not been tested on other distros. You may try it out on other distros of your choice.

Note: To run the Vulkan driver with AMDGPU upstream driver stack on SI and CI generation GPUs, amdgpu.si_support and amdgpu.cik_support need to be enabled in kernel

Feature Support and Performance

The AMD Open Source Driver for Vulkan is designed to support the following features:

Vulkan 1.2
More than 30 extensions
Radeon™ GPUProfiler tracing
Built-in debug and profiling tools
Mid-command buffer preemption and SR-IOV virtualization

The following features and improvements are planned in future releases (Please refer to Release Notes for update of each release):

Upcoming versions of the Vulkan API
Hardware performance counter collection through RenderDoc
LLPC optimizations to improve GPU-limited performance and compile time
Optimizations to improve CPU-limited performance

Known Issues

CTS may hang in VK.synchronization.internally_synchronized_objects.pipeline_cache_compute with Linux kernel versions lower than 4.13
The driver can only work with firmware of ME feature version >= 25 (you can check the version with command "sudo cat /sys/kernel/debug/dri/0/amdgpu_firmware_info"). If you are using upstream stack with GPUs of SI or CI family, you may need to upgrade the kernel to 4.19 or later version and firmware (under /lib/firmware/amdgpu/) to the right version from https://git.kernel.org/pub/scm/linux/kernel/git/firmware/linux-firmware.git/tree/amdgpu, and then update ramfs (sudo mkinitramfs -o /boot/initrd.img-`uname -r` `uname -r`)
Timeline semaphore is not fully supported in Linux kernel until version 5.5. You can install Vulkan timeline semaphore layer to enable the extension if you are using earlier version of Linux kernel

How to Contribute

You are welcome to submit contributions of code to the AMD Open Source Driver for Vulkan.

The driver is built from source code in four repositories: LLVM, XGL, LLPC and PAL.

For changes to LLVM, you should submit contribution to the LLVM trunk. Commits there will be evaluated to merge into the amd-gfx-gpuopen-master branch periodically.

For changes to XGL, LLPC and PAL, please create a pull request against the dev branch. After your change is reviewed and if it is accepted, it will be evaluated to merge into the master branch in a subsequent regular promotion.

IMPORTANT: By creating a pull request, you agree to allow your contribution to be licensed by the project owners under the terms of the MIT License.

When contributing to XGL, LLPC and PAL, your code should:

Match the style of nearby existing code. Your code may be edited to comply with our coding standards when it is merged into the master branch.
Avoid adding new dependencies, including dependencies on STL.

Please make each contribution reasonably small. If you would like to make a big contribution, like a new feature or extension, please raise an issue first to allow planning to evaluate and review your work.

Note: Since PAL is a shared component that must support other APIs, other operating systems, and pre-production hardware, you might be asked to revise your PAL change for reasons that may not be obvious from a pure Linux Vulkan driver perspective.

Build Instructions

System Requirements

It is recommended to install 16GB RAM in your build system.

Build System

CMake 3.13.4 or newer is required. Download and install proper one if the cmake is older than 3.13.4.
Ninja is requred.

Install Dev and Tools Packages

Ubuntu

sudo apt-get install build-essential curl g++-multilib gcc-multilib git pkg-config python3

64-bit

sudo apt-get install libssl-dev libx11-dev libxcb1-dev x11proto-dri2-dev libxcb-dri3-dev libxcb-dri2-0-dev libxcb-present-dev libxshmfence-dev libxrandr-dev libwayland-dev

32-bit

dpkg --add-architecture i386
sudo apt-get install libssl-dev:i386 libx11-dev:i386 libxcb1-dev:i386 libxcb-dri3-dev:i386 libxcb-dri2-0-dev:i386 libxcb-present-dev:i386 libxshmfence-dev:i386 libwayland-dev libwayland-dev:i386 libxrandr-dev:i386

RedHat

wget https://dl.fedoraproject.org/pub/epel/epel-release-latest-7.noarch.rpm

sudo yum localinstall epel-release-latest-7.noarch.rpm

sudo yum update

64-bit

sudo yum -y install openssl-devel gcc-c++ python3 curl glibc-devel libstdc++-devel libxcb-devel libX11-devel libxshmfence-devel libXrandr-devel wayland-devel

32-bit

sudo yum -y install openssl-devel.i686 gcc-c++ python3 curl glibc-devel.i686 libstdc++-devel.i686 libxcb-devel.i686 libX11-devel.i686 libxshmfence-devel.i686 libXrandr-devel.i686 wayland-devel.i686

Get Repo Tools

Ubuntu 18.04

sudo apt-get install repo

Ubuntu 20.04, RedHat 7.8, 8.2

mkdir ~/bin
curl https://storage.googleapis.com/git-repo-downloads/repo > ~/bin/repo
# Replacing python with python3 is only needed on Ubuntu 20.04 if the 'python' executable is not available
sed -i s/python/python3/ ~/bin/repo
chmod a+x ~/bin/repo
export PATH=~/bin:"$PATH"

Get Source Code

mkdir vulkandriver
cd vulkandriver
repo init -u https://github.com/GPUOpen-Drivers/AMDVLK.git -b master
repo sync

Note: Source code in dev branch can be gotten by using "-b dev" in the "repo init" command

Build Driver and Generate JSON Files

Ubuntu

cmake -G Ninja -S drivers/xgl -B drivers/xgl/builds/Release64 -DCMAKE_INSTALL_PREFIX=/usr/lib/ -DCMAKE_INSTALL_LIBDIR=x86_64-linux-gnu
ninja -C drivers/xgl/builds/Release64

cmake -G Ninja -S drivers/xgl -B drivers/xgl/builds/Release32 -DCMAKE_C_FLAGS=-m32 -DCMAKE_CXX_FLAGS=-m32 -DCMAKE_INSTALL_PREFIX=/usr/lib/ -DCMAKE_INSTALL_LIBDIR=i386-linux-gnu
ninja -C drivers/xgl/builds/Release32

RedHat

cmake -G Ninja -S drivers/xgl -B drivers/xgl/builds/Release64 -DCMAKE_INSTALL_PREFIX=/usr/ -DCMAKE_INSTALL_LIBDIR=lib64
ninja -C drivers/xgl/builds/Release64

cmake -G Ninja -S drivers/xgl -B drivers/xgl/builds/Release32 -DCMAKE_C_FLAGS=-m32 -DCMAKE_CXX_FLAGS=-m32 -DCMAKE_INSTALL_PREFIX=/usr/ -DCMAKE_INSTALL_LIBDIR=lib
ninja -C drivers/xgl/builds/Release32

Note:

For RedHat 7.x, please use cmake3(>= 3.13.4) instead of cmake.
For debug build, use -DCMAKE_BUILD_TYPE=Debug -DLLVM_PARALLEL_LINK_JOBS=2 (Linking a debug build of llvm is very memory intensive, so we use only two parallel jobs).
To enable Wayland support, use -DBUILD_WAYLAND_SUPPORT=ON.

Installation Instructions

Install Vulkan SDK

You can download and install the SDK package here.

Copy Driver and JSON Files

Ubuntu

sudo cp <vulkandriver_path>/drivers/xgl/builds/Release64/icd/amdvlk64.so /usr/lib/x86_64-linux-gnu/
sudo cp <vulkandriver_path>/drivers/xgl/builds/Release32/icd/amdvlk32.so /usr/lib/i386-linux-gnu/
sudo cp <vulkandriver_path>/drivers/xgl/builds/Release64/icd/amd-icd64.json /etc/vulkan/icd.d/
sudo cp <vulkandriver_path>/drivers/xgl/builds/Release32/icd/amd-icd32.json /etc/vulkan/icd.d/

RedHat

sudo cp <vulkandriver_path>/drivers/xgl/builds/Release64/icd/amdvlk64.so /usr/lib64/
sudo cp <vulkandriver_path>/drivers/xgl/builds/Release32/icd/amdvlk32.so /usr/lib/
sudo cp <vulkandriver_path>/drivers/xgl/builds/Release64/icd/amd-icd64.json /etc/vulkan/icd.d/
sudo cp <vulkandriver_path>/drivers/xgl/builds/Release32/icd/amd-icd32.json /etc/vulkan/icd.d/

NOTE: To make AMDVLK driver work correctly on system with both AMDVLK and RADV installed, AMD switchable graphics layer needs to be enabled by:

sudo ln -s /etc/vulkan/icd.d/amd_icd64.json /etc/vulkan/implicit_layer.d/amd_icd64.json
sudo ln -s /etc/vulkan/icd.d/amd_icd32.json /etc/vulkan/implicit_layer.d/amd_icd32.json

By default, AMDVLK driver is enabled. You can switch the driver between AMDVLK and RADV by environment variable AMD_VULKAN_ICD = AMDVLK or RADV.

Note: The remaining steps are only required when running the AMDGPU upstream driver stack.

Turn on DRI3 and disable modesetting X driver

Add following lines in /usr/share/X11/xorg.conf.d/10-amdgpu.conf:

Section "Device"

Identifier "AMDgpu"

Option  "DRI" "3"

EndSection

And make sure following line is NOT included in the section:

Driver      "modesetting"

Required Settings

On the AMDGPU upstream driver stack with libdrm version lower than 2.4.92, the max number of IB per submission MUST be limited to 4 (the default setting in AMD Open Source driver for Vulkan is 16). This can be accomplished via the Runtime Settings mechanism by adding the following line to amdPalSettings.cfg:

MaxNumCmdStreamsPerSubmit,4
CommandBufferCombineDePreambles,1

Install with pre-built driver

You could download and install pre-built package from https://github.com/GPUOpen-Drivers/AMDVLK/releases for each code promotion in master branch:

Ubuntu 18.04, 20.04

sudo dpkg -r amdvlk   /* If old version is installed on the machine, remove it first */
sudo dpkg -i amdvlk_x.x.x_amd64.deb
sudo apt-get -f install

RedHat 7.8, 8.2

sudo rpm -e amdvlk   /* If old version is installed on the machine, remove it first */
sudo rpm -i amdvlk-x.x.x-el.x86_64.rpm

For Ubuntu, you could also install the latest driver build from https://repo.radeon.com:

sudo wget -qO - http://repo.radeon.com/amdvlk/apt/debian/amdvlk.gpg.key | sudo apt-key add -
sudo sh -c 'echo deb [arch=amd64,i386] http://repo.radeon.com/amdvlk/apt/debian/ bionic main > /etc/apt/sources.list.d/amdvlk.list'
sudo apt-get remove amdvlk /* If old version is installed on the machine, remove it first */
sudo apt update
sudo apt-get install amdvlk

Runtime Settings

The driver exposes many settings that can customize the driver's behavior and facilitate debugging. You can add/edit settings in amdPalSettings.cfg file under one of below paths, formatted with one name,value pair per line:

/etc/amd
$XDG_CONFIG_HOME
$HOME/.config

Some example settings are listed below:

Setting Name	Valid Values	Comment
`AllowVkPipelineCachingToDisk`	0: disallow 1: default	1 is default value which enables Pal's archive-file based caching. The archive-file is stored under ~/.cache/AMD/VkCache.
`ShaderCacheMode`	0: disable cache 1: runtime cache 2: cache to disk	Runtime cache is the default mode. For "cache to disk", the cache file is generated under $AMD_SHADER_DISK_CACHE_PATH/AMD/LlpcCache or $XDG_CACHE_HOME/AMD/LlpcCache or $HOME/.cache/AMD/LlpcCache
`IFH`	0: default 1: drop all submits	Infinitely Fast Hardware. Submit calls are dropped before being sent to hardware. Useful for measuring CPU-limited performance.
`EnableVmAlwaysValid`	0: disable 1: default 2: force enable	1 is the default setting which enables the VM-always-valid feature for kernel 4.16 and above. The feature can reduce command buffer submission overhead related to virtual memory management.
`IdleAfterSubmitGpuMask`	Bitmask of GPUs (i.e., bit 0 is GPU0, etc.)	Forces the CPU to immediately wait for each GPU submission to complete on the specified set of GPUs.

All available settings can be determined by examining below source files that define them.

.../xgl/icd/settings/settings.cfg (API layer settings)
.../pal/src/core/settings_core.json (PAL hardware-independent settings)
.../pal/src/core/hw/gfxip/gfx6/settings_gfx6.json (PAL GFX6-8 settings)
.../pal/src/core/hw/gfxip/gfx9/settings_gfx9.json (PAL GFX9+ settings)

Runtime settings are only read at device initialization, and cannot be changed without restarting the application. If running on a system with multiple GPUs, the same settings will apply to all of them. Lines in the settings file that start with ; will be treated as comments.

Enable extensions under development

The extensions under development are not enabled by default in driver. You can enable them through environment variable:

export AMDVLK_ENABLE_DEVELOPING_EXT="<extension1-name> [<extension2-name>...]"

export AMDVLK_ENABLE_DEVELOPING_EXT="all"

The extension name is case-insensitive.

PAL GpuProfiler Layer

The GpuProfiler is an optional layer that is designed to intercept the PAL interface to provide basic GPU profiling support. Currently, this layer is controlled exclusively through runtime settings and outputs its results to file.

You can use the following Runtime Settings to generate .csv files with GPU timings of work performed during the designated frames of an application (one file for each frame):

Setting Name	Value	Comment
`GpuProfilerMode`	0: disable 1: enable with sqtt off 2: enable with sqtt for thread trace 3: enable with sqtt for RGP	Enables and sets the SQTT mode for the GPU performance profiler layer. Actual capture of performance data must be specified via frame number with GpuProfilerConfig_StartFrame or by pressing shift-F11.
`GpuProfilerConfig.LogDirectory`	<directory-path>	The directory path is relative to $AMD_DEBUG_DIR or $TMPDIR or /var/tmp/, default value is "amdpal/". Your application must have write permissions to the directory. The profiling logs are output to a subdirectory that is named in the format like <AppName><yyyy-MM-dd><HH:mm:ss>.
`GpuProfilerConfig.Granularity`	0: per-draw 1: per-cmdbuf	Defines what is measured/profiled. Per-draw times individual commands (such as draw, dispatch, etc.) inside command buffers, while per-cmdbuf only profiles entire command buffers in aggregate.
`GpuProfilerConfig.StartFrame`	Positive integer	First frame to capture data for. If StartFrame and FrameCount are not set, all frames will be profiled.
`GpuProfilerConfig.FrameCount`	Positive integer	Number of frames to capture data for.
`GpuProfilerConfig.RecordPipelineStats`	0, 1	Gathers pipeline statistic query data per entry if enabled.

You can use the script timingReport.py to analyze the profiling log:

python timeReport.py <profiling_log_subdirectory>

Dump Pipelines and Shaders

The output of timeReport.py includes the information of top pipelines like below:

Top Pipelines (>= 1%)
Compiler Hash         | Type         | Avg. Call Count | Avg. GPU Time [us] | Avg. Frame %
1. 0xd91d15e42d62dcbb | VsPs         |              43 |          11,203.15 |      10.20 %
2. 0x724e9af55f2adf1b | Cs           |               1 |           9,347.50 |       8.51 %
3. 0x396e5ad6f7a789f7 | VsHsDsPs     |             468 |           8,401.35 |       7.65 %

You can add the following settings to amdPalSettings.cfg to dump the information of each pipeline:

EnablePipelineDump,1
PipelineDumpDir,<dump_dir_path>

PipelineDumpDir is a sub-path relative to $AMD_DEBUG_DIR or $TMPDIR or /var/tmp/, default value is "spvPipeline/". The pipeline dump file is named in the format like Pipeline<Type>_<Compiler_Hash>.pipe. For example, the above top 1 pipeline is dumped to PipelineVsFs_0xD91D15E42D62DCBB.pipe. The shaders referenced by each pipeline are also dumped to .spv files.

PAL Debug Overlay

PAL's debug overlay can be enabled to display real time statistics and information on top of a running application. This includes a rolling FPS average, CPU and GPU frame times, and a ledger tracking how much video memory has been allocated from each available heap. Benchmarking (i.e., "Benchmark (F11)") is currently unsupported.

Setting Name	Value	Comment
`DebugOverlayEnabled`	0, 1	Enables the debug overlay.
`DebugOverlayConfig.DebugOverlayLocation`	0: top-left 1: top-right 2: bottom-left 3: bottom-right	Determines where the overlay text should be displayed. Can be used to avoid collision with important rendering by the application.
`DebugOverlayConfig.PrintFrameNumber`	0, 1	Reports the current frame number. Useful when determining a good frame range for profiling with the GpuProfiler layer.
`DebugOverlayConfig.TimeGraphEnable`	0, 1	Enables rendering of a graph of recent CPU and GPU frame times.

Third Party Software

The AMD Open Source Driver for Vulkan contains code written by third parties.

LLVM is distributed under the Apache License v2.0 with LLVM Exceptions. See LICENSE.TXT file in the top directory of the LLVM repository.
MetroHash is distributed under the terms of Apache License 2.0. See LICENSE file in the top directory of the MetroHash repository.
CWPack is distributed under the terms of MITLicense. See LICENSE file in the top directory of the CWPack repository.
Please see the README.md file in the PAL, LLPC and XGL repositories for information on third party software used by those libraries.

DISCLAIMER

The information contained herein is for informational purposes only, and is subject to change without notice. This document may contain technical inaccuracies, omissions and typographical errors, and AMD is under no obligation to update or otherwise correct this information. Advanced Micro Devices, Inc. makes no representations or warranties with respect to the accuracy or completeness of the contents of this document, and assumes no liability of any kind, including the implied warranties of noninfringement, merchantability or fitness for particular purposes, with respect to the operation or use of AMD hardware, software or other products described herein. No license, including implied or arising by estoppel, to any intellectual property rights is granted by this document. Terms and limitations applicable to the purchase or use of AMD's products are as set forth in a signed agreement between the parties or in AMD's Standard Terms and Conditions of Sale.

AMD, the AMD Arrow logo, Radeon, FirePro, and combinations thereof are trademarks of Advanced Micro Devices, Inc. Other product names used in this publication are for identification purposes only and may be trademarks of their respective companies.

Vega is a codename for AMD architecture, and is not a product name.

Linux is the registered trademark of Linus Torvalds in the U.S. and other countries.

Vulkan and the Vulkan logo are registered trademarks of the Khronos Group, Inc.

23 KiB Raw Blame History

AMD Open Source Driver for Vulkan®

Product Support

Operating System Support

Feature Support and Performance

Known Issues

How to Contribute

Build Instructions

System Requirements

Build System

Install Dev and Tools Packages

Ubuntu

64-bit

32-bit

RedHat

64-bit

32-bit

Get Repo Tools

Ubuntu 18.04

Ubuntu 20.04, RedHat 7.8, 8.2

Get Source Code

Build Driver and Generate JSON Files

Ubuntu

RedHat

Installation Instructions

Install Vulkan SDK

Copy Driver and JSON Files

Ubuntu

RedHat

Turn on DRI3 and disable modesetting X driver

Required Settings

Install with pre-built driver

Ubuntu 18.04, 20.04

RedHat 7.8, 8.2

Runtime Settings

Enable extensions under development

PAL GpuProfiler Layer

Dump Pipelines and Shaders

PAL Debug Overlay

Third Party Software

DISCLAIMER

23 KiB

Raw Blame History