Vulkan Subgroup Tutorial

[JeGX]

Subgroups are an important new feature in Vulkan 1.1 because they enable highly-efficient sharing and manipulation of data between multiple tasks running in parallel on a GPU. In this tutorial, we will cover how to use the new subgroup functionality.

Modern heterogeneous hardware like GPUs gain performance by using parallel hardware and exposing a parallel programming model to target this hardware. When a user wants to run N parallel tasks for their algorithm, a GPU would divide this N-sized workload between the compute units of that GPU. Each compute unit of the GPU is then capable of running one or more of these parallel tasks concurrently. In Vulkan, we refer to the data that runs on a single compute unit of a GPU as the local workgroup, and an individual parallel task as an invocation.

Vulkan 1.0 already exposes a method to share data between the invocations in a local workgroup via shared memory, which is exposed only in compute shaders. Shared memory allows for invocations within the local workgroup to share some data via memory that is faster to access than reading and writing to buffer memory, providing a mechanism to share data in a performance sensitive context.

Vulkan 1.1 goes further and introduces a mechanism to share data between the invocations that run in parallel on a single compute unit. These concurrently running invocations are named the subgroup. This subgroup allows for the sharing of data between a much smaller set of invocations than the local workgroup could, but at a significantly higher performance.

While shared memory is only available in compute shaders, sharing data via subgroup operations is allowed in all shader stages via optionally supported stages as we'll explain below.

Link: https://www.khronos.org/blog/vulkan-subgroup-tutorial