Jacob He 3775c7f5a1 Update README.md:

update AMDVLK README.md for some GPUProfiler setting names are changed

2018-08-22 15:01:14 +08:00

19 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™ HD 7000 Series
Radeon™ HD 8000M Series
Radeon™ R5/R7/R9 200/300 Series
Radeon™ RX 400/500 Series
Radeon™ M200/M300/M400 Series
Radeon™ RX Vega Series
AMD FirePro™ Workstation Wx000/Wx100/Wx300 Series
Radeon™ Pro WX x100 Series
Radeon™ Pro 400/500 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 16.04.3 (64-bit version)
RedHat 7.4 (64-bit version)

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

Feature Support and Performance

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

Vulkan 1.1
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

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-vulkan-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.

Install Dev and Tools Packages

Ubuntu

sudo apt-get install build-essential python3 cmake curl g++-multilib gcc-multilib

sudo apt-get install libx11-dev libxcb1-dev x11proto-dri2-dev libxcb-dri3-dev libxcb-dri2-0-dev libxcb-present-dev libxshmfence-dev libx11-dev:i386 libxcb1-dev:i386 x11proto-dri2-dev:i386 libxcb-dri3-dev:i386 libxcb-dri2-0-dev:i386 libxcb-present-dev:i386 libxshmfence-dev:i386 libwayland-dev libwayland-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

sudo yum -y install gcc-c++ cmake3 python34 curl glibc-devel glibc-devel.i686 libstdc++-devel libstdc++-devel.i686 libxcb-devel libxcb-devel.i686 libX11-devel libX11-devel.i686 libxshmfence-devel libxshmfence-devel.i686

Get Repo Tools

mkdir ~/bin
curl https://storage.googleapis.com/git-repo-downloads/repo > ~/bin/repo
chmod a+x ~/bin/repo

Get Source Code

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

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

64-bit Build

Ubuntu

cd <vulkandriver_path>/drivers/xgl

cmake -H. -Bbuilds/Release64

cd builds/Release64

make -j$(nproc)

RedHat

cd <vulkandriver_path>/drivers/xgl

cmake3 -H. -Bbuilds/Release64

cd builds/Release64

make -j$(nproc)

32-bit Build

Ubuntu

cd <vulkandriver_path>/drivers/xgl

cmake -H. -Bbuilds/Release -DCMAKE_C_FLAGS=-m32 -DCMAKE_CXX_FLAGS=-m32

cd builds/Release

make -j$(nproc)

RedHat

cd <vulkandriver_path>/drivers/xgl

cmake3 -H. -Bbuilds/Release -DCMAKE_C_FLAGS=-m32 -DCMAKE_CXX_FLAGS=-m32

cd builds/Release

make -j$(nproc)

Note:

If the build runs into errors like "collect2: fatal error: ld terminated with signal 9 [Killed]" due to out of memory, you could try with reducing the number of threads in "make" command.
Debug build can be done by using -DCMAKE_BUILD_TYPE=Debug.
To enable Wayland support, you need to build the driver by using -DBUILD_WAYLAND_SUPPORT=ON and install the Wayland WSA library.

Installation Instructions

Install Vulkan SDK

Refer to installation instructions here.

Uninstall Previously Installed JSON Files

Please make sure all JSON files for AMD GPUs under below folders are uninstalled:

/etc/vulkan/icd.d
/usr/share/vulkan/icd.d

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/Release/icd/amdvlk32.so /usr/lib/i386-linux-gnu/
sudo cp <vulkandriver_path>/drivers/AMDVLK/json/Ubuntu/* /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/Release/icd/amdvlk32.so /usr/lib/
sudo cp <vulkandriver_path>/drivers/AMDVLK/json/Redhat/* /etc/vulkan/icd.d/

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

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
`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.

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>.
`GpuProfilerPerfCounterConfig_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.
`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.
`PrintFrameNumber`	0, 1	Reports the current frame number. Useful when determining a good frame range for profiling with the GpuProfiler layer.
`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 terms of the University of Illinois/NCSA Open Source License. See LICENSE.TXT file in the top directory of the LLVM 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.

19 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

Install Dev and Tools Packages

Ubuntu

RedHat

Get Repo Tools

Get Source Code

64-bit Build

Ubuntu

RedHat

32-bit Build

Ubuntu

RedHat

Installation Instructions

Install Vulkan SDK

Uninstall Previously Installed JSON Files

Copy Driver and JSON Files

Ubuntu

RedHat

Turn on DRI3 and disable modesetting X driver

Required Settings

Runtime Settings

PAL GpuProfiler Layer

Dump Pipelines and Shaders

PAL Debug Overlay

Third Party Software

DISCLAIMER

19 KiB

Raw Blame History