Skip to content

Commit

Permalink
Update documentation for V2.1
Browse files Browse the repository at this point in the history
  • Loading branch information
@chesik-amd
chesik-amd committed Apr 24, 2024
1 parent 91e7f56 commit 97d1fcc
Show file tree
Hide file tree
Showing 36 changed files with 175 additions and 65 deletions.
88 changes: 46 additions & 42 deletions RELEASE_NOTES.txt
View file
@@ -1,66 +1,69 @@
Radeon™ GPU Profiler V2.0 12-07-2023
Radeon™ GPU Profiler V2.1 04-24-2024
-------------------------------------

V2.0 Changes
V2.1 Changes
-------------------------------------

* Radeon GPU Profiler

1) Support for additional AMD RDNA™ 3 hardware
2) Redesigned Wavefront occupancy user interface, allowing for user customization and improved usage of available screen real estate
3) Dark mode user interface support, allowing the user to choose between a light and dark theme (or have RGP follow the OS theme setting)
4) The ray tracing shader table displayed in the Pipeline state pane can now display data on thread divergence for individual shader functions (DirectX® 12 only for now, requires AMD Software: Adrenalin Edition™ 23.12.1 or newer)
5) Allow opening .rgp files which contain a large number of events
6) HIP kernels that contain calls to other functions now support the Call Targets table in the Instruction timing pane's side panel
7) PIX3 marker support updated for latest version of WinPixEventRuntime
8) Bug/stability fixes

* Radeon Developer Panel

1) To support collecting data for thread divergence in the ray tracing shader table, the Profiling workflow user interface has a new "Enable shader instrumentation" checkbox (affects DirectX 12 only for now, requires AMD Software: Adrenalin Edition™ 23.12.1 or newer)
1) Interoperability with the Radeon GPU Analyzer: binary pipelines can now be extracted from a loaded profile data set in RGP and automatically loaded into a new instance of RGA for analysis
2) Rows in Wavefront occupancy pane can now be resized, allowing for additional user customization
3) New "Color by limiting factor" coloring mode in the Wavefront occupancy and Event timing panes. This will highlight events whose theoretical occupancy is limited by VGPR usage, LDS usage or thread group dimensions
4) New "Color by context rolls" coloring mode in the Wavefront occupancy and Event timing panes. This will highlight events where a context roll occurred since the previous event
5) Latency visualization in the Instruction timing pane will now show which part of the total latency represents a "pre-issue" stall
6) Fixed issue with incorrect LDS usage reported on RDNA™-based GPUs
7) Bug/stability fixes

Known Issues
-------------------------------------

* All platforms

1) Radeon Developer Panel can only capture a profile on a single AMD GPU at a time.
2) Radeon Developer Panel cannot capture profiles from non-AMD GPUs.
3) Applications that call Present() from the async compute queue are not supported on pre-RDNA hardware. Incomplete profile data may result on RDNA-based hardware.
4) When using RGP with RenderDoc, please make sure that RenderDoc is terminated between RenderDoc capture sessions (generating a RenderDoc capture file or loading a RenderDoc capture file is considered a session for the purpose here). While it is possible to take multiple RGP profiles of a RenderDoc capture file, it is not possible to take RGP profiles between RenderDoc sessions. If this is attempted, RenderDoc will show an error dialog box indicating that an RGP profile can't be taken and to restart RenderDoc
5) If an instance of Radeon GPU Profiler is spawned from RenderDoc, it must be closed before restarting RenderDoc. The menu option to create new RGP profiles will not be enabled otherwise.
6) OpenCL™ captures may include an extra DMA command buffer in the Profile summary.
7) In some rare cases on RDNA 2 hardware, all counter data may be missing from a captured RGP profile. When this happens, Radeon Developer Panel will prompt the user to recapture.
8) In some rare cases, data for one or more cache counters may be missing. Usually, recapturing will allow the missing data to show up.
9) It is recommended to use at least 1080p display resolution (1920 x 1080) with the RGP user interface. Some minor user interface issues may appear when using a lower resolution.
10) Cache and ray tracing counter data collection is not currently supported on RDNA based APUs.
11) For systems consisting of an AMD APU and an AMD discrete GPU, capturing profiles should work, but an error may be logged in the Radeon Developer Panel regarding not being able to set peak clock mode. It is recommended that the GPU in the APU be disabled in the BIOS.
1) When using RGP with RenderDoc, please make sure that RenderDoc is terminated between RenderDoc capture sessions (generating a RenderDoc capture file or loading a RenderDoc capture file is considered a session for the purpose here). While it is possible to take multiple RGP profiles of a RenderDoc capture file, it is not possible to take RGP profiles between RenderDoc sessions. If this is attempted, RenderDoc will show an error dialog box indicating that an RGP profile can't be taken and to restart RenderDoc
2) If an instance of Radeon GPU Profiler is spawned from RenderDoc, it must be closed before restarting RenderDoc. The menu option to create new RGP profiles will not be enabled otherwise.
3) OpenCL™ captures may include an extra DMA command buffer in the Profile summary.
4) It is recommended to use at least 1080p display resolution (1920 x 1080) with the RGP user interface. Some minor user interface issues may appear when using a lower resolution.

* Windows®

1) D3D12 command list calls of ExecuteIndirect() may show in RGP as multiple compute events.
2) Some Radeon Software hotkeys may conflict with Radeon GPU Profiler shortcut keys. The Radeon Software hotkeys can be reconfigured by opening the Radeon Software panel (from the system tray), selecting the Hotkeys tab under Settings then changing or unbinding any conflicting hotkeys.
3) If a DirectX® 12 profile is missing GPU synchronization primitive data (i.e. signals and waits) on the Frame summary pane, please try running the included scripts\AddUserToGroup.bat batch file and then recapturing the profile. This batch file must be run as Administrator.
4) Current versions of the Radeon Developer Panel cannot profile Universal Windows Platform (UWP) applications. Please use Radeon Developer Panel v2.8 to profile a UWP application.

* Linux®

1) Installations of Ubuntu 20.04 or newer may have the RADV open source Vulkan® driver installed by default on the system. As a result, after an amdgpu-pro driver install, the default Vulkan ICD may be the RADV ICD. In order to capture a profile, Vulkan applications must be using the amdgpu-pro Vulkan ICD. The default Vulkan ICD can be overridden by setting the following environment variable before launching a Vulkan application: VK_ICD_FILENAMES=/etc/vulkan/icd.d/amd_icd64.json
2) After launching RGP from the Developer Panel to view a captured profile, the panel may fail to connect the next time it is launched. The workaround is to close RGP before relaunching the panel.
3) If the Developer Panel or the Developer Service crash while running with the root account, it may be necessary to restart/exit them again with the root account in order to cleanup shared memory.
4) When running with the root account, the Developer Panel may output error or warning messages to the terminal. These should not prevent the panel from functioning properly.
5) If the RadeonDeveloperServiceCLI application crashes, shared memory may need to be cleaned up by running the remove_shared_memory.sh script located in the script folder of the RGP release kit. Run the script with elevated privileges using sudo.
6) The Radeon Developer Panel may fail to start the Radeon Developer Service when the Connect button is clicked. If this occurs, manually start the Radeon Developer Service, select localhost from the the Recent connections list and click the Connect button again.
7) On some RDNA 3 hardware, detailed instruction timing data will not be available even when the user asks Radeon Developer Panel to collect it.
1) Installations of Ubuntu 20.04 or newer may have the RADV open source Vulkan driver installed by default on the system. As a result, after an amdgpu-pro driver install, the default Vulkan ICD may be the RADV ICD. In order to capture a profile, Vulkan applications must be using the amdgpu-pro Vulkan ICD. The default Vulkan ICD can be overridden by setting the following environment variable before launching a Vulkan application: VK_ICD_FILENAMES=/etc/vulkan/icd.d/amd_icd64.json

* RDNA

1) The Device configuration does not show the correct Work group processor per Shader engine for certain parts with harvested CUs.

* Radeon Developer Panel

1) See RDP_RELEASE_NOTES.txt for additional items that may affect profiling.


Release Notes History
-------------------------------------

V2.0 Changes
-------------------------------------

* Radeon GPU Profiler

1) Support for additional AMD RDNA 3 hardware
2) Redesigned Wavefront occupancy user interface, allowing for user customization and improved usage of available screen real estate
3) Dark mode user interface support, allowing the user to choose between a light and dark theme (or have RGP follow the OS theme setting)
4) The ray tracing shader table displayed in the Pipeline state pane can now display data on thread divergence for individual shader functions (DirectX 12 only for now, requires AMD Software: Adrenalin Edition™ 23.12.1 or newer)
5) Allow opening .rgp files which contain a large number of events
6) HIP kernels that contain calls to other functions now support the Call Targets table in the Instruction timing pane's side panel
7) PIX3 marker support updated for latest version of WinPixEventRuntime
8) Bug/stability fixes

* Radeon Developer Panel

1) To support collecting data for thread divergence in the ray tracing shader table, the Profiling workflow user interface has a new "Enable shader instrumentation" checkbox (affects DirectX 12 only for now, requires AMD Software: Adrenalin Edition 23.12.1 or newer)

V1.16 Changes
-------------------------------------

Expand Down Expand Up @@ -95,20 +98,21 @@ V1.15 Changes

* Radeon GPU Profiler

1) Support for additional AMD RDNA 3 hardware
2) Newly redesigned ISA disassembly views in the Pipeline state and Instruction timing panes
1) Support for additional AMD RDNA 3 hardware
2) Newly redesigned ISA disassembly views in the Pipeline state and Instruction timing panes
- Code blocks can now be collapsed/expanded
- Selected token highlighting allows you to quickly see other instances of the selected token (instruction opcodes, registers and constants)
- One-click navigation between branch instructions and their targets, along with tracked navigation history
- Customize the displayed columns
- Improved search result highlighting
3) Improved performance in the System activity timeline in the Frame summary pane when opening large profiles
4) Instruction timing side panel will now report the total number of WMMA (wave matrix multiply accumulate) instructions executed by a shader when running on RDNA 3 or newer hardware
5) Pipeline state pane will now report when conservative rasterization is enabled
6) Fixed issues with keyboard selection in the tree view in the Event timing and Pipeline state panes
7) DirectX 12 Mesh shader functions and Vulkan Mesh shader extension functions now are identified properly in RGP's event lists
8) Fixed incorrect tree hierarchy in the Event timing and Pipeline state pane when events are grouped by user events and event filtering is used
9) Bug/stability fixes
3) Improved performance in the System activity timeline in the Frame summary pane when opening large profiles
4) Instruction timing side panel will now report the total number of WMMA (wave matrix multiply accumulate) instructions executed by a shader when running on RDNA 3 or newer hardware
5) Pipeline state pane will now report when conservative rasterization is enabled
6) Fixed issues with keyboard selection in the tree view in the Event timing and Pipeline state panes
7) DirectX 12 Mesh shader functions and Vulkan Mesh shader extension functions now are identified properly in RGP's event lists
8) Fixed incorrect tree hierarchy in the Event timing and Pipeline state pane when events are grouped by user events and event filtering is used
9) Initial support for DirectX 12 Work Graphs
10) Bug/stability fixes

V1.14.1 Changes
-------------------------------------
Expand Down
8 changes: 5 additions & 3 deletions docs/source/conf.py
View file
@@ -1,5 +1,7 @@
# -*- coding: utf-8 -*-
#
# Copyright (c) 2017-2024 Advanced Micro Devices, Inc. All rights reserved.
#
# Radeon GPU Profiler documentation build configuration file, created by
# sphinx-quickstart on Fri Jun 30 12:01:48 2017.
#
Expand Down Expand Up @@ -46,17 +48,17 @@

# General information about the project.
project = u'Radeon GPU Profiler'
copyright = u'2017-2023, Advanced Micro Devices, Inc. All rights reserved.'
copyright = u'2017-2024, Advanced Micro Devices, Inc. All rights reserved.'
author = u'AMD Developer Tools'

# The version info for the project you're documenting, acts as replacement for
# |version| and |release|, also used in various other places throughout the
# built documents.
#
# The short X.Y version.
version = u'2.0.0'
version = u'2.1.0'
# The full version, including alpha/beta/rc tags.
release = u'2.0.0'
release = u'2.1.0'

# The language for content autogenerated by Sphinx. Refer to documentation
# for a list of supported languages.
Expand Down
74 changes: 65 additions & 9 deletions docs/source/index.rst
View file
Expand Up @@ -308,6 +308,13 @@ right-click context menu to jump between panes, the option to "View in
context rolls" will only be available if the selected event is currently
present in the events table on the context rolls pane.

In the events panes selecting the "context rolls" option from the "Color
By" drop down box in the Wavefront occupancy event timeline or the Event
timing pane shows all events that have had their context rolled from the
previous event.

.. image:: media_rgp/rgp_context_rolls_4.png

Most expensive events
---------------------

Expand Down Expand Up @@ -566,8 +573,8 @@ supports zooming:
.. |ZoomInRef| image:: media_rgp/rgp_zoom_in.png
.. |ZoomOutRef| image:: media_rgp/rgp_zoom_out.png

|ZoomSelectionRef| **Zoom to selection**
----------------------------------------
|ZoomSelectionRef| Zoom to selection
------------------------------------
When **Zoom to selection** is clicked, the zoom level is increased to a selected
region or selected event. A selection region is set by holding down the
left mouse button while the mouse is on the graph and dragging the mouse
Expand All @@ -580,14 +587,14 @@ to a selected event can be accomplished by simply double clicking the event.
Pressing the **Z** shortcut key while holding down the **CTRL** key activates
**Zoom to selection** as well.

|ZoomResetRef| **Zoom reset**
-----------------------------
|ZoomResetRef| Zoom reset
-------------------------
When **Zoom reset** is clicked, the zoom level is returned to the original level
to reveal the entire time span on the graph. The zoom level can also be reset
using the **H** shortcut key.

|ZoomInRef| **Zoom in**
-----------------------
|ZoomInRef| Zoom in
-------------------
Increases the zoom level incrementally to display a shorter time span on the
graph. The zoom level is increased each time this icon is clicked until the
maximum zoom level is reached. Alternatively, holding down the **CTRL** key
Expand All @@ -596,8 +603,8 @@ will also zoom in for a more detailed view. Zooming in can be activated with
the **A** shortcut key. To zoom in quickly at a 10x rate, press the **S**
shortcut key.

|ZoomOutRef| **Zoom out**
-------------------------
|ZoomOutRef| Zoom out
---------------------
Decreases the zoom level incrementally to display a longer time span on the
graph. The zoom level is decreased each time this icon is clicked until the
minimum zoom level is reached (i.e. the full available time region).
Expand Down Expand Up @@ -1046,6 +1053,55 @@ contain a user event hierarchy, nothing will be shown.
Events enclosed by user markers are colored in the wavefront occupancy
view. They are also visible in the side panel.

.. _rga_rgp_interop:

Radeon GPU Analyzer and Radeon GPU Profiler interop
===================================================

The Radeon GPU Analyzer now supports opening pipeline binary files in its binary analysis mode.
Users can create and open these binary files directly from RGP and view them in RGA. To export
a pipeline binary for analysis in RGA right-click an event in any RGP pane that contains events
and select "Analyze pipeline in Radeon GPU Analyzer" in the context menu options.

.. image:: media_rgp/rgp_analyze_pipeline_in_rga.png

Some events such as indirect raytracing events can have multiple pipeline binaries. To select
which pipeline binary to analyze in RGA right-click a specific binary in the pipeline state
shader table for ray tracing events and select the context menu option to analyze that pipeline
binary. Alternatively, right-click the event anywhere in RGP and select "Analyze pipeline in
Radeon GPU Analyzer" to open a window to pick from the full list of pipeline binaries in that event.
Pushing the "Analyze selected binaries" button will save and open all binaries that were checked in
the list in RGA. Keep in mind opening a large number of pipeline binaries in RGA may take some time.

.. image:: media_rgp/rgp_select_multiple_pipeline_binaries_for_rga_export.png

The location of the Radeon GPU Analyzer executable file as well as the location where pipeline
binaries are saved can be changed in the Radeon GPU Analyzer interop section of the general settings.

.. image:: media_rgp/rgp_rga_interop_settings.png

If either the executable file or pipeline binary file path cannot be found an error message will
be displayed next to the corresponding setting.

.. image:: media_rgp/rgp_rga_interop_settings_invalid.png

Selecting a pipeline binary for analysis while either of these file paths are invalid will open up
a message prompt to select a valid file path.

.. image:: media_rgp/rgp_rga_executable_not_found.png

If the pipeline binary being exported already exists on disk a message prompt will appear asking
if the file should be overwritten or directly opened in RGA without being overwritten. Select the
checkbox in the message prompt to save the selected option for the future and not see this message
again.

.. image:: media_rgp/rgp_pipeline_binary_file_already_exists.png

This setting can be changed at any time from the Radeon GPU Analyzer interop section of the
general settings.

.. image:: media_rgp/rgp_overwrite_existing_pipeline_binaries_options.png

RenderDoc & Radeon GPU Profiler interop BETA
============================================

Expand Down Expand Up @@ -1225,4 +1281,4 @@ Microsoft is a registered trademark of Microsoft Corporation in the US and other

Windows is a registered trademark of Microsoft Corporation in the US and other jurisdictions.

© 2016-2023 Advanced Micro Devices, Inc. All rights reserved.
© 2016-2024 Advanced Micro Devices, Inc. All rights reserved.

0 comments on commit 97d1fcc

Please sign in to comment.