GPU-AV Post Process

The goal behind Post Process is to have a very lightweight way to get information from the GPU so that it can be used on the CPU.

In order to be lightweight, Post Process makes assumptions that it can run the GPU without crashing/hanging.

Descriptor Indexing

For descriptor indexing we use Post Process to do the same validation on the CPU, but one at draw time and one after queue submission time.

Take the following two descriptors in a GLSL shader:

layout(set = 0, binding = 0) buffer SSBO_0 { int data; } descriptor_a;
layout(set = 0, binding = 1) buffer SSBO_1 { int data; } descriptor_b[];

Normally we will validate descriptor_a at draw time as the Vulkan spec requires any static usage to be valid.

Once we start having bindless (or dynamic array) of descriptor, it is important to only check the index accessed. If descriptor_b has a million items, but only two are accessed, we use GPU-AV Post Process to detect that.

We create a buffer that hold information about each descriptor and when the GPU is ran, we quickly just mark what has been access. The instrumented GLSL would look like

inst_post_process_descriptor_index(0 /*binding*/, 1 /*binding*/, index);
descriptor_b[index].data = 0;

Reusing the same code twice

Currently, we have a DescriptorValidator class that is created at draw time with Core Checks and also at post-queue submission with GPU-AV.

The goal is that the ValidateDescriptor() function can be called the same from both variations.

Tracking Descriptors and SPIR-V

When tracking data for Post Process, we need both information from vkCmdBindDescriptorSets and vkCmdBindPipeline (or shader objects), but not all are present at the same time.

The VkDescriptorSet object contains a buffer with a PostProcessDescriptorIndexSlot for each descriptor in it. This “descriptor index slot” holds information that can only be known at GPU execution time (which index, OpVariable, etc... are used). For each action command, we create a LUT mapping 32 (current max limit) addresses to the VkDescriptorSet buffer.

The pipelines contain the OpVariable SPIR-V information, such as which type of image is used. This is wrapped up in a BindingVariableMap pointer. Using this we can map the accessed descriptor on the GPU back to all the state tracking on the CPU.

The following diagram aims to illustrate the flow of data: