Add a chapter on Compute Shaders#366
Add a chapter on Compute Shaders#366spencer-lunarg wants to merge 1 commit intoKhronosGroup:mainfrom
Conversation
SaschaWillems
left a comment
SaschaWillems
left a comment
There was a problem hiding this comment.
Having a chapter explaining basic compute terms is a great addition 👍🏻
|
|
||
| Setting the `workgroup` size can be done in 3 ways | ||
|
|
||
| 1. Using the `WorkgroupSize` built-in (link:https://godbolt.org/z/ees83eT7x[example]) |
There was a problem hiding this comment.
Most developers wouldn't care about this detail and it would be more useful to show how it's declared in GLSL/HLSL/Slang.
There was a problem hiding this comment.
sure, part of this was the "pure" vulkan/spirv point-of-view, but can easily throw a "how to do it in GLSL/HLSL/Slang" example with it
| ---- | ||
| shared uint my_var; | ||
| void main() { | ||
| // All the invocations in the workgroup are going to try to write to the same memory. |
There was a problem hiding this comment.
I think it's confusing to lead with this example, and would be better to start with examples of cross-thread communication and using barrier() first.
There was a problem hiding this comment.
can you provide some GLSL snippets here (in the comment), I know you are more of an expert in this area and probably better at providing some better examples to provide
gpx1000
left a comment
gpx1000
left a comment
There was a problem hiding this comment.
The chapter is a solid addition and fills a clear gap in the guide!
|
|
||
| For those who are more familiar with graphics in Vulkan, compute will be a simple transition. Basically everything is the same except: | ||
|
|
||
| - No vertex buffers, render passes, swapchains needed |
There was a problem hiding this comment.
While technically correct that they aren't required to execute the pipeline, it's worth mentioning that compute shaders often interact with these resources (e.g., writing data into a vertex buffer or post-processing a swapchain image as a storage image). This clarifies that they aren't "isolated" from the rest of the graphics pipeline.
| For those who are more familiar with graphics in Vulkan, compute will be a simple transition. Basically everything is the same except: | ||
|
|
||
| - No vertex buffers, render passes, swapchains needed | ||
| - Call `vkCmdDispatch` instead of `vkCmdDraw` |
There was a problem hiding this comment.
I'd recommend mentioning vkCmdDispatchIndirect here as well. Indirect dispatch is a cornerstone of advanced compute usage and GPU-driven pipelines, and it fits perfectly in a list of "Vulkan-isms."
| ---- | ||
| shared uint my_var; | ||
| void main() { | ||
| atomicStore(temp, 0u, gl_ScopeWorkgroup, 0, 0); |
There was a problem hiding this comment.
There is a typo here—temp should be my_var to match the variable declared above.
4 participants
Rendered view
Kind review @jeffbolznv - mainly around the shared memory races and any other stuff you think would be useful