You absolutely cannot implement stream compaction “at the speed of native” as WebGPU is missing the wave/subgroup intrinsics and globally coherent memory necessary to do that efficiently as possible.
It's possible you might not need direct access to wave/subgroup ops to implement efficient stream compaction. There's a great old Nvidia blog post on "warp-aggregated atomics"
where they show that their compiler is sometimes able to automatically convert global atomic operations into the warp local versions, and achieve the same performance as manually written intrinsics.
I was recently curious if 10 years later these same optimizations had made it into other GPUs and platforms besides cuda, so I put together a simple atomics benchmark in WebGPU.
The results seem to indicate that these optimizations are accessible through webgpu on chrome on both MacOS and Linux (with nvidia gpu).
Note that I'm not directly testing stream compaction, just incrementing a single global atomic counter. So that would need to be tested to know for sure if the optimization still holds there.
If you see any issues with the benchmark or this reasoning please let me know! I am hoping to solidify my knowledge in this area :)
I think compilers should be smart enough to substitute group-shared atomics with horizontal ops. If it's not already doing it, it should be!
But anyways, Histogram Pyramids is a more efficient algorithm for implementing parallel scan anyways. It essentially builds a series of 3D buffers, each having half the dimension of the previous level, and each value containing the sum of the amounts in each underlying cells, with the top cube being just a single value, the total amount of cells.
Then instead of doing the second pass where you figure out what index thread is supposed to write to, and writing it to a buffer, you just simply drill down into said cubes and figure out the index at the invocation of the meshing part by looking at your thread index (lets say 1526), and looking at the 8 smaller cubes (okay, cube 1 has 516 entries, so 1100 to go, cube 2 has 1031 entries, so 69 to go, cube 3 has 225 entries, so we go to cube 3), and recursively repeat until you find the index. Since all threads in a group tend go into the same cubes, all threads tend to read the same bits of memory until getting down to the bottom levels, making it very GPU cache friendly (divergent reads kill GPGPU perf).
Forgive me if I got the technical terminology wrong, I haven't actually worked on GPGPU in more than a decade, but it's fun to not that something that I did cca 2011 as an undergrad is suddenly relevant again (in which I implemented HistoPyramids from a 2007ish paper, and Marching Cubes, an 1980s algorithm). Everything old is new again.
You seem knowledgeable, and I’m possibly going back into a GPGPU project after many years out of the game, so: overall do you see a good future for filling these compute-related gaps in the WebGPU API? Really I’m wondering whether wgpu is an okay choice versus raw Vulkan for native GPGPU outside the browser.
The answer to that for any given feature is "can untrusted code be trusted with that?". Wave intrinsics are probably doable. Bindless maybe, but expect a bunch of bounds checking overhead. Pointers/BDA, absolutely not.
Native libraries like wgpu can do whatever they want in extensions, safety be damned, but you're stepping outside of the WebGPU spec in that case.
Don't know about GPGPU, but can give you a probably correct answer: Compared to "native" APIs you trade features for compatibility. It's always going to lag behind Vulkan/DX/Metal. Are you ok with excluding platforms? Vulkan/Metal/DX. If not, then I'd give wgpu a chance. Wgpu is also higher-level than Vulkan, which is borh a pro and a con.
The demo doesn't work on mobile Chrome. Worse, the blog post crashes the embedded browser in the HN app. May I suggest just linking to the demo instead?
tehsauce|1 year ago
https://developer.nvidia.com/blog/cuda-pro-tip-optimized-fil...
where they show that their compiler is sometimes able to automatically convert global atomic operations into the warp local versions, and achieve the same performance as manually written intrinsics. I was recently curious if 10 years later these same optimizations had made it into other GPUs and platforms besides cuda, so I put together a simple atomics benchmark in WebGPU.
https://github.com/PWhiddy/webgpu-atomics-benchmark
The results seem to indicate that these optimizations are accessible through webgpu on chrome on both MacOS and Linux (with nvidia gpu). Note that I'm not directly testing stream compaction, just incrementing a single global atomic counter. So that would need to be tested to know for sure if the optimization still holds there. If you see any issues with the benchmark or this reasoning please let me know! I am hoping to solidify my knowledge in this area :)
FL33TW00D|1 year ago
jsheard|1 year ago
https://github.com/gpuweb/gpuweb/blob/main/proposals/subgrou...
There is a proposal for supporting subgroups in WebGPU proper but it's still in the draft stage.
pjmlp|1 year ago
torginus|1 year ago
But anyways, Histogram Pyramids is a more efficient algorithm for implementing parallel scan anyways. It essentially builds a series of 3D buffers, each having half the dimension of the previous level, and each value containing the sum of the amounts in each underlying cells, with the top cube being just a single value, the total amount of cells.
Then instead of doing the second pass where you figure out what index thread is supposed to write to, and writing it to a buffer, you just simply drill down into said cubes and figure out the index at the invocation of the meshing part by looking at your thread index (lets say 1526), and looking at the 8 smaller cubes (okay, cube 1 has 516 entries, so 1100 to go, cube 2 has 1031 entries, so 69 to go, cube 3 has 225 entries, so we go to cube 3), and recursively repeat until you find the index. Since all threads in a group tend go into the same cubes, all threads tend to read the same bits of memory until getting down to the bottom levels, making it very GPU cache friendly (divergent reads kill GPGPU perf).
Forgive me if I got the technical terminology wrong, I haven't actually worked on GPGPU in more than a decade, but it's fun to not that something that I did cca 2011 as an undergrad is suddenly relevant again (in which I implemented HistoPyramids from a 2007ish paper, and Marching Cubes, an 1980s algorithm). Everything old is new again.
masspro|1 year ago
jsheard|1 year ago
Native libraries like wgpu can do whatever they want in extensions, safety be damned, but you're stepping outside of the WebGPU spec in that case.
tormeh|1 year ago
dekhn|1 year ago
Archit3ch|1 year ago
spintin|1 year ago
The browser is dead, the only thing you can use it for is filling out HTML forms and maybe some light inventory management.
The final app is C+Java where you put the right stuff where it is needed. Just like the browser used to be before Oracle did it's magic on the applet.
worik|1 year ago
Yea. Nah!
That obit is a bit premature
teaearlgraycold|1 year ago