re-render threads with user handle of reply.
This commit is contained in:
1 parent
1ceb4b1583
commit
f8380db665
515 files changed
+2042
-1496
No files matched your search
+7
-7
@@ -25,34 +25,34 @@ in_reply_to: ""
|
||||
|
||||
## Thread
|
||||
|
||||
**1/**
|
||||
**1/** @rianflo
|
||||
|
||||
"The road to 16-bit floats GPU is paved with our blood"
|
||||
:-/
|
||||
|
||||
https://www.yosoygames.com.ar/wp/2022/01/the-road-to-16-bit-floats-gpu-is-paved-with-our-blood/
|
||||
|
||||
**2/**
|
||||
**2/** @NOTimothyLottes
|
||||
|
||||
@rianflo Explicit packed 16-bit works on AMD Vulkan Vega and up. I typically get up to 30% improvement on ALU bound stuff. Lots of occupancy wins. I don't use {HLSL, RenderDoc, Reflection, RADV, or VS/PS}. All constants are packed and aliased as UINT, so no coversion overheads.
|
||||
|
||||
**3/**
|
||||
**3/** @rianflo
|
||||
|
||||
@NOTimothyLottes Oh I know the benefits. Just no simple clear way to write it in GLSL for vulkan.
|
||||
|
||||
**4/**
|
||||
**4/** @NOTimothyLottes
|
||||
|
||||
@rianflo Sure there is. CAS/FSR1/etc all shipped with fantastic GLSL versions using 16-bit packed math (I wrote those), all which at the time got fantastic code generation using AMD's drivers.
|
||||
|
||||
**5/**
|
||||
**5/** @rianflo
|
||||
|
||||
@NOTimothyLottes What GLSL extension did you use?
|
||||
|
||||
**6/**
|
||||
**6/** @rianflo
|
||||
|
||||
@NOTimothyLottes Oh wait, you're saying you wrote the fp16 math manually?
|
||||
|
||||
**7/**
|
||||
**7/** @NOTimothyLottes
|
||||
|
||||
@rianflo Explicit packed 16-bit code. FSR1 example: https://github.com/GPUOpen-Effects/FidelityFX-FSR/blob/master/ffx-fsr/ffx_fsr1.h - There are different 'F' (32-bit) and 'H' and 'Hx2' (packed 16-bit) functions.
|
||||
|
||||
|
||||
+3
-3
@@ -26,15 +26,15 @@ parent_post_id: "2001106069238563204"
|
||||
|
||||
## Branch
|
||||
|
||||
**1/**
|
||||
**1/** @rianflo
|
||||
|
||||
@NOTimothyLottes question is how much overhead. of course even not knowing how it was implemented, its obviously complex. but i’d take a bit of overhead in favor of writing a “program” top down on the gpu.
|
||||
|
||||
**2/**
|
||||
**2/** @rianflo
|
||||
|
||||
@NOTimothyLottes you already know we’re on the same page wrt to rendering stuff differently, but it would be fun to show a proof of concept for rendering trad stuff this way, even just to proof a point. like load and render any gltf. :-)
|
||||
|
||||
**3/**
|
||||
**3/** @NOTimothyLottes
|
||||
|
||||
@rianflo I just need to get off the R&D train at some point and actually ship something :)
|
||||
|
||||
|
||||
@@ -25,42 +25,42 @@ in_reply_to: ""
|
||||
|
||||
## Thread
|
||||
|
||||
**1/**
|
||||
**1/** @rianflo
|
||||
|
||||
Significant amount of bloat dropped :-) 🎉🥳🎈
|
||||
|
||||
I’d move it all to the GPU though. A single kernel that launches your pipeline steps per frame or per whatever. Pretty sure NV can do it. AMD I have no deep knowledge. @NOTimothyLottes ?
|
||||
|
||||
**2/**
|
||||
**2/** @NOTimothyLottes
|
||||
|
||||
@rianflo From my perspective being compute-only for graphics, few reasons to launch kernels because I'd always use a fixed amount of VGPRs. The larger problem is compilers that cannot keep up with the large program sizes (in compile time or register allocation).
|
||||
|
||||
Branches: [[archive/threads/rianflo/2025-12-17-significant-amount-of-bloat-dropped/2025-12-17-rianflo-question-is-how-much-overhead-of-course-even-not]]
|
||||
|
||||
**3/**
|
||||
**3/** @NOTimothyLottes
|
||||
|
||||
@rianflo Huge amount of pitfalls with compute-side dynamic dispatch of work (NV had it back when I worked there BTW), from long latency chains, to overheads tracking work completion, preemption, state restore, etc. It's not the golden answer people think it is. It is a slow path.
|
||||
|
||||
**4/**
|
||||
**4/** @NOTimothyLottes
|
||||
|
||||
@rianflo There are however a bunch of things we could trivially do GPU side that are currently CPU side, like computing best tiling modes for images, computing image {size, alignment, etc}, building descriptors, etc. Lots of other driver stuff could be done on the GPU for sure.
|
||||
|
||||
**5/**
|
||||
**5/** @NOTimothyLottes
|
||||
|
||||
@rianflo My plan for 2026 is actually to build my open source AMD SteamOS/Linux driver for a future compute-graphics API, and it will be doing some radical stuff since I'm also doing shader binary code generation too :) Hopefully you all get live examples.
|
||||
|
||||
**6/**
|
||||
**6/** @bkaradzic
|
||||
|
||||
@NOTimothyLottes @rianflo Can't wait! This compute-graphics API is something like compute + meshlets + fragment shaders only graphics API?
|
||||
|
||||
**7/**
|
||||
**7/** @NOTimothyLottes
|
||||
|
||||
@bkaradzic @rianflo Compute as in only compute shader dispatches, no fixed function graphics (no fragment/etc). And graphics via pure compute.
|
||||
|
||||
**8/**
|
||||
**8/** @Nicolas_Lopez_
|
||||
|
||||
@NOTimothyLottes @bkaradzic @rianflo Considering how much we already software rasterize, we are not so far from it :)
|
||||
|
||||
**9/**
|
||||
**9/** @rianflo
|
||||
|
||||
@Nicolas_Lopez_ @NOTimothyLottes @bkaradzic yeah but then someone insists on a couple of large triangles and wants to access the rasterizer hw again, which is supposedly too small of die area to be even worth ripping it out.
|
||||
Reference in new issue
Block a user