Skip to content

Latest commit

 

History

History
302 lines (227 loc) · 14.7 KB

creating_tests.md

File metadata and controls

302 lines (227 loc) · 14.7 KB

Creating Tests

This is an "up-to-speed" document for writing tests to validate the Validation Layers

Information how to run the tests

Rule #1

The first rule is to make sure you are actually running the tests on the built version of the Validation Layers you want. If you have the Vulkan SDK installed, then you will have a pre-built version of the Validation Layers set in your path and those are probably not the version you want to test.

Make sure you have the correct VK_LAYER_PATH set on Windows or Linux (on Android the layers are baked into the APK so there is nothing to worry about)

export VK_LAYER_PATH=/path/to/Vulkan-ValidationLayers/build/layers/

There is nothing worse than debugging why your layers are not reporting the VUID you added due to not actually testing that code!

Google Test Overview

The tests take advantage of the Google Test (gtest) Framework which breaks each test into a TEST_F(VkLayerTest, TestName) "Test Fixture". This just means that for every test there will be a class that holds many useful member variables.

To run a test labeled TEST_F(VkLayerTest, Foo) is as simple as going --gtest_filter=VkLayerTest.Foo

VkRenderFramework

The VkRenderFramework class is "base class" that abstract most things in order to allow a test writer to focus on the small part of coded needed for the test.

For most tests, it is as simple as going

RETURN_IF_SKIP(Init());

// or

RETURN_IF_SKIP(InitFramework());
RETURN_IF_SKIP(InitState());

// For Best Practices tests
RETURN_IF_SKIP(InitBestPracticesFramework());
RETURN_IF_SKIP(InitState());

to set it up. This will create the VkInstance and VkDevice for you.

There are other useful helper functions such as InitSurface(), InitSwapchain(), InitRenderTarget(), and more. Please view the class source for more of an overview of what it currently supports

Extensions

Adding extension support is quite easy.

Here is an example of adding VK_KHR_sampler_ycbcr_conversion with all the extensions it requires as well. Note, most extensions will have only 1 or 2 dependency extensions needed

// Setup extensions, including dependent instance and device extensions. This call should be made before any call to InitFramework
AddRequiredExtensions(VK_KHR_SAMPLER_YCBCR_CONVERSION_EXTENSION_NAME);
AddRequiredFeature(vkt::Feature:: samplerYcbcrConversion);

// Among other things, this will create the VkInstance and VkPhysicalDevice that will be used for the test.
// Also will check that all extensions and their dependencies were enabled successfully
RETURN_IF_SKIP(InitFramework());

// Finish initializing state, including creating the VkDevice (whith extensions added) that will be used for the test
RETURN_IF_SKIP(InitState());

The pattern breaks down to

  • Check and add Instance extensions to list
  • Init Framework which creates VkInstance
  • Check and add Device extensions to list
  • Init State which creates the VkDevice
  • Optional: skip if test is not worth moving out without extension support (more below)

Pattern for optional extensions

Sometimes it is worth checking for an extension, but still running the parts of a test if the extension is not supported

AddRequiredExtensions(VK_KHR_GET_PHYSICAL_DEVICE_PROPERTIES_2_EXTENSION_NAME);
AddOptionalExtensions(VK_KHR_COPY_COMMANDS_2_EXTENSION_NAME);
// Check required (not optional) extensions are still supported
RETURN_IF_SKIP(Init());

// need to wait until after phyiscal device creation to know if it was enabled
const bool copy_commands2 = IsExtensionsEnabled(VK_KHR_COPY_COMMANDS_2_EXTENSION_NAME);

// Validate core copy command
m_errorMonitor->SetDesiredFailureMsg(kErrorBit, vuid);
vk::CmdCopyBuffer( /* */ );
m_errorMonitor->VerifyFound();

// optional test using VK_KHR_copy_commands2
if (copy_commands2) {
    m_errorMonitor->SetDesiredFailureMsg(kErrorBit, vuid);
    vk::CmdCopyBuffer2KHR( /* */  );
    m_errorMonitor->VerifyFound();
}

Vulkan Version

As raised in a previous issue, when a Vulkan version is enabled, all extensions that are required are exposed. (For example, if the test is created as a Vulkan 1.1 application, then the VK_KHR_16bit_storage extension would be supported regardless as it was promoted to Vulkan 1.1 core)

If a certain version of Vulkan is needed a test writer can call

SetTargetApiVersion(VK_API_VERSION_1_1);
// Will skip if the supported instance version is not high enough, but actual device supported version may be lower
RETURN_IF_SKIP(InitFramework());

Later in the test the actual Vulkan version supported by the device can be checked

if (DeviceValidationVersion() >= VK_API_VERSION_1_1) {
    // Only can be ran on Vulkan 1.0 or 1.1
}

Promoted extensions

The test framework now automatically handles extensions promoted to core versions by not enabling the extensions if the target instance Vulkan version (set through SetTargetApiVersion) or the Vulkan version supported by the device (queriable using DeviceValidationVersion) already includes the instance or device extension's functionality, respectively. This applies to both requirements requested using AddRequiredExtensions or AddOptionalExtensions, as well as optional requirements checked using IsExtensionsEnabled.

In order to enforce enabling extensions even when they are included in the effective Vulkan version (which should only be necessary for very specific test cases), the test case can call the AllowPromotedExtensions function.

Getting Function Pointers

When using a version that has promoted the function, one can just directly use the call.

In the case of enabling the extensions, all the functions for those extensions will call vkGetDeviceProcAddr automatically.

// If VK_API_VERSION_1_1 or later is set
vk::BindImageMemory2(...);

// If VK_KHR_bind_memory2 is enabled
vk::BindImageMemory2KHR(...);

Error Monitor

The ErrorMonitor *m_errorMonitor in the VkRenderFramework becomes your best friend when you write tests

This small class handles checking all things to VUID and are ultimately what will "pass or fail" a test

The few common patterns that will cover 99% of cases are:

  • By default, all Vulkan API calls are expected to succeed. In the past, one would have to "wrap" API calls in ExpectSuccess/VerifyNotFound to ensure an API call did not trigger any errors. This is no longer the case. e.g.,
// m_errorMonitor->ExpectSuccess(); <- implicit
vk::CreateSampler(device(), &sci, nullptr, &samplers[0]);
// m_errorMonitor->VerifyNoutFound(); <- implicit

The ExpectSuccess and VerifyNotFound calls are now implicit.

  • For checking a call that invokes a VUID error
m_errorMonitor->SetDesiredFailureMsg(kErrorBit, "VUID-VkSamplerCreateInfo-addressModeU-01646");
// The following API call is expected to trigger 01646 and _only_ 01646
vk::CreateSampler(device(), &sci, NULL, &BadSampler);
m_errorMonitor->VerifyFound();

// All calls after m_errorMonitor->VerifyFound() are expected to not trigger any errors. e.g., the following API call should succeed with no validation errors being triggered.
vk::CreateImage(device(), &ci, nullptr, &mp_image);
  • When it is possible another VUID will be triggered that you are not testing. This usually happens due to making something invalid can cause a chain effect causing other things to be invalid as well.
    • Note: If the SetUnexpectedError is never called it will not fail the test
m_errorMonitor->SetUnexpectedError("VUID-VkImageMemoryRequirementsInfo2-image-01590");
m_errorMonitor->SetDesiredFailureMsg(kErrorBit, "VUID-VkImageMemoryRequirementsInfo2-image-02280");
vkGetImageMemoryRequirements2Function(device(), &mem_req_info2, &mem_req2);
m_errorMonitor->VerifyFound();
  • When you expect multpile VUID to be triggered. This is also be a case if you expect the same VUID to be called twice.
    • Note: If both VUID are not found the test will fail
m_errorMonitor->SetDesiredFailureMsg(kErrorBit, "VUID-VkDeviceGroupRenderPassBeginInfo-deviceMask-00905");
m_errorMonitor->SetDesiredFailureMsg(kErrorBit, "VUID-VkDeviceGroupRenderPassBeginInfo-deviceMask-00907");
vk::CmdBeginRenderPass(m_commandBuffer->handle(), &m_renderPassBeginInfo, VK_SUBPASS_CONTENTS_INLINE);
m_errorMonitor->VerifyFound();
  • When a VUID is dependent on an extension being present
    • Note: The start of the test might already have a boolean that checks for extension support
const char* vuid = IsExtensionsEnabled(VK_KHR_SAMPLER_YCBCR_CONVERSION_EXTENSION_NAME) ? "VUID-vkCmdCopyImage-dstImage-01733" : "VUID-vkCmdCopyImage-dstImage-01733";
m_errorMonitor->SetDesiredFailureMsg(kErrorBit, vuid);
m_commandBuffer->CopyImage(image_2.image(), VK_IMAGE_LAYOUT_GENERAL, image_1.image(), VK_IMAGE_LAYOUT_GENERAL, 1, &copy_region);
m_errorMonitor->VerifyFound();
  • There should be a single Vulkan function between m_errorMontior calls to ensure the test is only applied to the single function call being tested.
  • Keep it simple. Try to make each test as small and concise as possible.
  • Avoid testing VUIDs in "batches" such as:
m_errorMonitor->SetDesiredFailureMsg(kErrorBit, "VUID-VkCommandBufferBeginInfo-flags-06003");
m_errorMonitor->SetDesiredFailureMsg(kErrorBit, "VUID-VkCommandBufferInheritanceRenderingInfo-colorAttachmentCount-06004");
m_errorMonitor->SetDesiredFailureMsg(kErrorBit, "VUID-VkCommandBufferInheritanceRenderingInfo-variableMultisampleRate-06005");
m_errorMonitor->SetDesiredFailureMsg(kErrorBit, "VUID-VkCommandBufferInheritanceRenderingInfo-depthAttachmentFormat-06007");
m_errorMonitor->SetDesiredFailureMsg(kErrorBit, "VUID-VkCommandBufferInheritanceRenderingInfo-multiview-06008");
m_errorMonitor->SetDesiredFailureMsg(kErrorBit, "VUID-VkCommandBufferInheritanceRenderingInfo-viewMask-06009");
...
vk::BeginCommandBuffer(secondary_cmd_buffer, &cmd_buffer_begin_info);
m_errorMonitor->VerifyFound();

If there is a problem with one of these checks later on, this method makes it difficult and more time-consuming to figure out which check is problematic. It also makes it difficult to understand which Vulkan parameter/setting triggered which error. It should be relatively obvious to tell which line(s) of code caused a validation error to be triggered (and if it isn't, comments should be used to make it obvious).

  • Make it clear how the VUID you're testing is triggered. e.g.,
// Try to get layout for plane 3 when we only have 2
VkImageSubresource subresource{};
subresource.aspectMask = VK_IMAGE_ASPECT_MEMORY_PLANE_3_BIT_EXT;
VkSubresourceLayout layout{};
m_errorMonitor->SetDesiredFailureMsg(kErrorBit, "VUID-vkGetImageSubresourceLayout-tiling-09433");
vk::GetImageSubresourceLayout(m_device->handle(), image.handle(), &subresource, &layout);
m_errorMonitor->VerifyFound();

Here it is obvious that the aspectMask parameter is the cause of 02271.

Viewing VU Messages

When SetDesiredFailureMsg is used, nothing is displayed if the test is successful. To see the messages regardless use --print-vu

./tests/vk_layer_validation_tests --print-vu --gtest_filter=Tests

Device Profiles API

There are times a test writer will want to test a case where an implementation returns a certain support for a format feature or limit. Instead of just hoping to find a device that supports a certain case, there is the Device Profiles API layer. This layer will allow a test writer to inject certain values for format features and/or limits.

Device Profile Format Feature

Here is an example of how To enable it to allow overriding format features (limits are the same idea, just different function names):

RETURN_IF_SKIP(Init());

// Load required functions
PFN_vkSetPhysicalDeviceFormatPropertiesEXT fpvkSetPhysicalDeviceFormatPropertiesEXT = nullptr;
PFN_vkGetOriginalPhysicalDeviceFormatPropertiesEXT fpvkGetOriginalPhysicalDeviceFormatPropertiesEXT = nullptr;
if (!LoadDeviceProfileLayer(fpvkSetPhysicalDeviceFormatPropertiesEXT, fpvkGetOriginalPhysicalDeviceFormatPropertiesEXT)) {
    GTEST_SKIP() << "Failed to load device profile layer.";
}

This is an example of injecting the VK_FORMAT_FEATURE_STORAGE_IMAGE_BIT feature for the VK_FORMAT_R32G32B32A32_UINT format. This will force the Validations Layers to act as if the implementation had support for this feature later in the test's code.

VkFormatProperties fmt_props;
fpvkGetOriginalPhysicalDeviceFormatPropertiesEXT(gpu(), VK_FORMAT_R32G32B32A32_UINT, &fmt_props);
fmt_props.optimalTilingFeatures |= VK_FORMAT_FEATURE_STORAGE_IMAGE_BIT;
fpvkSetPhysicalDeviceFormatPropertiesEXT(gpu(), VK_FORMAT_R32G32B32A32_UINT, fmt_props);

If you are in need of VkFormatProperties3 the following is an example how to use the layer

PFN_vkSetPhysicalDeviceFormatProperties2EXT fpvkSetPhysicalDeviceFormatProperties2EXT = nullptr;
PFN_vkGetOriginalPhysicalDeviceFormatProperties2EXT fpvkGetOriginalPhysicalDeviceFormatProperties2EXT = nullptr;
if (!LoadDeviceProfileLayer(fpvkSetPhysicalDeviceFormatProperties2EXT, fpvkGetOriginalPhysicalDeviceFormatProperties2EXT)) {
    GTEST_SKIP() << "Failed to load device profile layer.";
}

VkFormatProperties3 fmt_props_3 = vku::InitStructHelper();
VkFormatProperties2 fmt_props = vku::InitStructHelper(&fmt_props_3);

// Removes unwanted support
fpvkGetOriginalPhysicalDeviceFormatProperties2EXT(gpu(), image_format, &fmt_props);
// VK_FORMAT_FEATURE_2_STORAGE_IMAGE_BIT == VK_FORMAT_FEATURE_STORAGE_IMAGE_BIT
// Need to edit both VkFormatFeatureFlags/VkFormatFeatureFlags2
fmt_props.formatProperties.optimalTilingFeatures = (fmt_props.formatProperties.optimalTilingFeatures & ~VK_FORMAT_FEATURE_2_STORAGE_IMAGE_BIT);
fmt_props_3.optimalTilingFeatures = (fmt_props_3.optimalTilingFeatures & ~VK_FORMAT_FEATURE_2_STORAGE_IMAGE_BIT);
// Was added with VkFormatFeatureFlags2 so only need to edit here
fmt_props_3.optimalTilingFeatures = (fmt_props_3.optimalTilingFeatures & ~VK_FORMAT_FEATURE_2_STORAGE_WRITE_WITHOUT_FORMAT_BIT);
fpvkSetPhysicalDeviceFormatProperties2EXT(gpu(), image_format, fmt_props);

Device Profile Limits

When using the device profile layer for limits, the test maybe need to call vkSetPhysicalDeviceLimitsEXT prior to creating the VkDevice for some validation state tracking

RETURN_IF_SKIP(InitFramework());

// Load required functions
PFN_vkSetPhysicalDeviceLimitsEXT fpvkSetPhysicalDeviceLimitsEXT = nullptr;
PFN_vkGetOriginalPhysicalDeviceLimitsEXT fpvkGetOriginalPhysicalDeviceLimitsEXT = nullptr;
if (!LoadDeviceProfileLayer(fpvkSetPhysicalDeviceLimitsEXT, fpvkGetOriginalPhysicalDeviceLimitsEXT)) {
    GTEST_SKIP() << "Failed to load device profile layer.";
}

VkPhysicalDeviceProperties props;
fpvkGetOriginalPhysicalDeviceLimitsEXT(gpu(), &props.limits);
props.limits.maxPushConstantsSize = 16; // example
fpvkSetPhysicalDeviceLimitsEXT(gpu(), &props.limits);

RETURN_IF_SKIP(InitState());