Vulkan CTS README

This document describes how to build and run Vulkan Conformance Test suite.

Vulkan CTS is built on dEQP framework. dEQP documentation is available at http://source.android.com/devices/graphics/testing.html

Requirements

Common

The following tools must be installed and present in the PATH variable:

  • Git (for checking out sources)
  • Python 3.x (for the build related scripts, some other scripts still use Python 2.7.x)
  • CMake 2.8 (3.6 for Android NDK r17+ builds) or newer

Win32

  • Visual Studio 2015 or newer (glslang uses several C++11 features)

Linux

  • Standard toolchain (make, gcc/clang)

Android

  • Android NDK r15c or later.
  • Android SDK with: SDK Tools, SDK Platform-tools, SDK Build-tools, and API 28
  • Java Development Kit (JDK)
  • Windows: either NMake or Ninja in PATH

If you have downloaded Android SDK tools, you can install necessary components by running:

tools/android update sdk --no-ui --all --filter tools,platform-tools,build-tools-25.0.2,android-28

Building CTS

To build dEQP, you need first to download sources for zlib, libpng, glslang, spirv-headers, and spirv-tools.

To download sources, run:

python external/fetch_sources.py

You may need to re-run fetch_sources.py to update to the latest glslang and spirv-tools revisions occasionally.

With CMake out-of-source builds are always recommended. Create a build directory of your choosing, and in that directory generate Makefiles or IDE project using cmake.

Windows x86-32

cmake <path to vulkancts> -G"Visual Studio 14"
start dEQP-Core-default.sln

Windows x86-64

cmake <path to vulkancts> -G"Visual Studio 14 Win64"
start dEQP-Core-default.sln

Linux 32-bit Debug

cmake <path to vulkancts> -DCMAKE_BUILD_TYPE=Debug -DCMAKE_C_FLAGS=-m32 -DCMAKE_CXX_FLAGS=-m32
make -j

Release build can be done by using -DCMAKE_BUILD_TYPE=Release

Linux 64-bit Debug

cmake <path to vulkancts> -DCMAKE_BUILD_TYPE=Debug -DCMAKE_C_FLAGS=-m64 -DCMAKE_CXX_FLAGS=-m64
make -j

Android

Following command will build dEQP.apk:

python scripts/android/build_apk.py --sdk <path to Android SDK> --ndk <path to Android NDK>

By default the CTS package will be built for the Android API level 28. Another API level may be supplied using --native-api command line option.

The package can be installed by either running:

python scripts/android/install_apk.py

By default the CTS package will contain libdeqp.so built for armeabi-v7a, arm64-v8a, x86, and x86_64 ABIs, but that can be changed using --abis command line option.

To pick which ABI to use at install time, following commands must be used instead:

adb install --abi <ABI name> <build-root>/package/dEQP.apk /data/local/tmp/dEQP-debug.apk

Building Mustpass

Current mustpass is checked into repository and can be found at:

external/vulkancts/mustpass/1.1.4/vk-default.txt

Vulkan CTS mustpass can be re-generated by running:

python <vulkancts>/external/vulkancts/scripts/build_mustpass.py

Pre-compiling SPIR-V binaries

For distribution, and platforms that don't support GLSL to SPIR-V compilation, SPIR-V binaries can be pre-built with following command:

python external/vulkancts/scripts/build_spirv_binaries.py

By default the script builds SPIR-V binaries for Vulkan 1.1. Binaries for other Vulkan versions can be requested by supplying an extra command line option:

python external/vulkancts/scripts/build_spirv_binaries.py --target-vulkan-version <Vulkan version>

Binaries will be written to external/vulkancts/data/vulkan/prebuilt/.

Test modules (or in case of Android, the APK) must be re-built after building SPIR-V programs in order for the binaries to be available.

Running CTS

Following command line options MUST be used when running CTS:

--deqp-caselist-file=<vulkancts>/external/vulkancts/mustpass/1.1.4/vk-default.txt
--deqp-log-images=disable
--deqp-log-shader-sources=disable

In addition on multi-device systems the device for which conformance is claimed can be selected with:

--deqp-vk-device-id=<value>

To speed up the conformance run on some platforms the following command line option may be used to disable frequent fflush() calls to the output logs:

--deqp-log-flush=disable

By default, the test log will be written into the path “TestResults.qpa”. If the platform requires a different path, it can be specified with:

--deqp-log-filename=<path>

By default, the shader cache will be written into the path “shadercache.bin”. If the platform requires a different path, it can be specified with:

--deqp-shadercache-filename=<path>

If the shader cache is not desired, it can be disabled with:

--deqp-shadercache=disable

No other command line options are allowed.

Win32

cd <builddir>/external/vulkancts/modules/vulkan
Debug/deqp-vk.exe --deqp-caselist-file=...

Test log will be written into TestResults.qpa

Linux

cd <builddir>/external/vulkancts/modules/vulkan
./deqp-vk --deqp-vk-caselist-file=...

Android

adb push <vulkancts>/external/vulkancts/mustpass/1.1.4/vk-default.txt /sdcard/vk-default.txt
adb shell

In device shell:

am start -n com.drawelements.deqp/android.app.NativeActivity -e cmdLine "deqp --deqp-caselist-file=/sdcard/vk-default.txt --deqp-log-images=disable --deqp-log-shader-sources=disable --deqp-log-filename=/sdcard/TestResults.qpa"

Test progress will be written to device log and can be displayed with:

adb logcat -s dEQP

Test log will be written into /sdcard/TestResults.qpa.

Conformance Submission Package Requirements

The conformance submission package must contain the following:

  1. Full test logs (TestResults.qpa) from CTS runs against all driver builds
  2. Result of git status and git log from CTS source directory
  3. Any patches used on top of release tag
  4. Conformance statement

Test logs (1) should be named <submission pkg dir>/TestResults-<driver build type>.qpa, for example TestResults-armeabi-v7a.qpa. On platforms where multiple different driver builds (for example 64-bit and 32-bit) are present, CTS logs must be provided for each driver build as part of the submission package.

Test logs generated on a system which exposes more than one physical device in a device group can be used for products that expose one or more physical devices in their device group.

The CTS build must always be done from clean git repository that doesn't have any uncommitted changes. Thus it is necessary to run and capture output of git status and git log (2) in the source directory:

git status > <submission pkg dir>/git-status.txt
git log --first-parent <release tag>^..HEAD > <submission pkg dir>/git-log.txt

Any changes made to CTS must be committed to the local repository, and provided as part of the submission package (3). This can be done by running:

git format-patch -o <submission pkg dir> <release tag>..HEAD

In general, bugfixes and changes to platform-specific code (mostly under framework/platform) are allowed. The commit message for each change must include a clear description of the change and why it is necessary. Non-porting related changes must be accompanied by a waiver (see below).

NOTE: When cherry-picking patches on top of release tag, please use git cherry-pick -x to include original commit hash in the commit message.

Conformance statement (4) must be included in a file called STATEMENT-<adopter> and must contain following:

CONFORM_VERSION:         <git tag of CTS release>
PRODUCT:                 <string-value>
CPU:                     <string-value>
OS:                      <string-value>

Note that product/cpu/os information is also captured in dEQP-VK.info.* tests if vk::Platform::describePlatform() is implemented.

If the submission package covers multiple products, you can list them by appending additional PRODUCT: lines to the conformance statement. For example:

CONFORM_VERSION:         vulkan-cts-1.1.4.0
PRODUCT:                 Product A
PRODUCT:                 Product B
...

The actual submission package consists of the above set of files which must be bundled into a gzipped tar file named VK<API major><API minor>_<adopter><_info>.tgz. <API major> is the major version of the Vulkan API specification, <API minor>is the minor version of the Vulkan API specification. <adopter> is the name of the Adopting member company, or some recognizable abbreviation. The <_info> field is optional. It may be used to uniquely identify a submission by OS, platform, date, or other criteria when making multiple submissions. For example, a company XYZ may make a submission for a Vulkan 1.1 implementation named VK11_XYZ_PRODUCTA_Windows10.tgz

One way to create a suiteable gzipped tar file is to execute the command:

tar -cvzf <filename.tgz> -C <submission pkg dir> .

where <submission pkg dir> is the directory containing the files from (1)-(4) from above. A submission package must contain all of the files listed above, and only those files.

As an example submission package could contain:

STATEMENT-Khronos
git-log.txt
git-status.txt
0001-Remove-Waived-Filtering-Tests.patch
0002-Fix-Pipeline-Parameters.patch
TestResults-armeabi-v7a.qpa
TestResults-arm64-v8a.qpa

Waivers

The process for requesting a waiver is to report the issue by filing a bug report in the Gitlab VulkanCTS project (TODO Github?). When creating the submission package, include references to the waiver in the commit message of the relevant change. Including as much information as possible in your bug report (including a unified diff or a merge request of suggested file changes) will ensure the issue can be progressed as rapidly as possible. Issues must be labeled “Waiver” (TODO!) and identify the version of the CTS and affected tests.

Conformance Criteria

Conformance run is considered passing if all tests finish with allowed result codes. Test results are contained in the TestResults.qpa log. Each test case section contains XML tag Result, for example:

<Result StatusCode="Pass">Not validated</Result>

The result code is the value of the StatusCode attribute. Following status codes are allowed:

Pass
NotSupported
QualityWarning
CompatibilityWarning

Submission package can be verified using external/vulkancts/scripts/verify_submission.py script. The script takes two arguments: path to extracted submission package and path to current mustpass list. For example:

python external/vulkancts/scripts/verify_submission.py VK_11_Khronos_1/ external/vulkancts/mustpass/1.1.4/vk-default.txt

Please note that the script reports a warning even for a correctly generated git-log.txt If your git-log.txt contains only head commit of the release tag then the warning can be ignored.

Vulkan platform port

Vulkan support from Platform implementation requires providing getVulkanPlatform() method in tcu::Platform class implementation.

See framework/common/tcuPlatform.hpp and examples in framework/platform/win32/tcuWin32Platform.cpp and framework/platform/android/tcuAndroidPlatform.cpp.

If any WSI extensions are supported, platform port must also implement methods for creating native display (vk::Platform::createWsiDisplay) and window handles (vk::wsi::Display::createWindow). Otherwise tests under dEQP-VK.wsi will fail.

Null (dummy) driver

For testing and development purposes it might be useful to be able to run tests on dummy Vulkan implementation. One such implementation is provided in vkNullDriver.cpp. To use that, implement vk::Platform::createLibrary() with vk::createNullDriver().

Validation Layers

Vulkan CTS framework includes first-party support for validation layers, that can be turned on with --deqp-validation=enable command line option.

When validation is turned on, default instance and device will be created with validation layers enabled and debug callback is registered to record any messages. Debug messages collected during test execution will be included at the end of the test case log.

If any validation errors are found, test result will be set to InternalError.

By default VK_DEBUG_REPORT_INFORMATION_BIT_EXT and _DEBUG_BIT_EXT messages are excluded from the log, but that can be customized by modifying vkt::TestCaseExecutor::deinit() in vktTestPackage.cpp.

On the Android target, layers can be added to the APK during the build process by setting the --layers-path command line option to point into the NDK or to a locally-built layers tree. The layers are expected to be found under $abi/ under the layers path.

Cherry GUI

Vulkan test module can be used with Cherry (GUI for test execution and analysis). Cherry is available at https://android.googlesource.com/platform/external/cherry. Please follow instructions in README to get started.

Before first launch, and every time test hierarchy has been modified, test case list must be refreshed by running:

python scripts/build_caselists.py path/to/cherry/data

Cherry must be restarted for the case list update to take effect.

Shader Optimizer

Vulkan CTS can be optionally run with the shader optimizer enabled. This is an experimental feature which can be used to further stress both the drivers as well as the optimizer itself. The shader optimizer is disabled by default.

The following command line options can be used to configure the shader optimizer:

--deqp-optimization-recipe=<number>

The list of the optimization recipes can be found and customized in the optimizeCompiledBinary() function in vkPrograms.cpp.

As of this writing, there are 8 recipes to choose from:

0. Disabled (default)
1. The example recipe from spir-v opt 1.0 whitepaper
2. RegisterPerformancePasses from commandline optimizer tool october 2017
3. RegisterSizePasses from commandline optimizer tool october 2017
4. RegisterLegalizationPasses from commandline optimizer tool April 2018
5. RegisterPerformancePasses from commandline optimizer tool April 2018
6. RegisterPerformancePasses from commandline optimizer tool April 2018 with CreateCommonUniformElimPass
7. RegisterSizePasses from commandline optimizer tool April 2018
8. RegisterSizePasses from commandline optimizer tool April 2018 with CreateCommonUniformElimPass

The Register...Passes() calls change in the SPIR-V optimizer tool from time to time. Since different sets of passes may result in different shaders, having several fixed sets is useful for issue discovery.

--deqp-optimize-spirv=enable

This option is not required to run the optimizer. By default, the shader optimizer only optimizes shaders generated from GLSL or HLSL, and leaves hand-written SPIR-V shaders alone.

Many of the hand-written SPIR-V tests stress specific features of the SPIR-V which might get optimized out. Using this option will enable the optimizer on the hand-written SPIR-V as well, which may be useful in finding new bugs in drivers or the optimizer itself, but will likely invalidate the tests themselves.

Shader Cache

The Vulkan CTS framework contains a shader cache for speeding up the running of the CTS. Skipping shader compilation can significantly reduce runtime, especially for repeated runs.

Default behavior is to have the shader cache enabled, but truncated at the start of the CTS run. This still gives the benefit of skipping shader compilation for identical shaders in different tests (which there are many), while making sure that the shader cache file does not grow indefinitely.

The shader cache identifies the shaders by hashing the shader source code along with various bits of information that may affect the shader compilation (such as shader stage, CTS version, possible compilation flags, etc). If a cached shader with matching hash is found, a byte-by-byte comparison of the shader sources is made to make sure that the correct shader is being retrieved from the cache.

The behavior of the shader cache can be modified with the following command line options:

--deqp-shadercache=disable

Disable the shader cache. All shaders will be compiled every time.

--deqp-shadercache-filename=<filename>

Set the name of the file where the cached shaders will be stored. This option may be required for the shader cache to work at all on Android targets.

--deqp-shadercache-truncate=disable

Do not truncate the shader cache file at startup. No shader compilation will occur on repeated runs of the CTS.

RenderDoc

The RenderDoc (https://renderdoc.org/) graphics debugger may be used to debug Vulkan tests.

Following command line option should be used when launching tests from RenderDoc UI:

--deqp-renderdoc=enable

This causes the framework to interface with the debugger and mark each dEQP test case as a separate ‘frame’, just for the purpose of capturing. The frames are added using RenderDoc ‘In-Application API’, instead of swapchain operations.