feat: Cast elements to correct type when creating vectors

[aleksanderkatan]

No description provided.

[github-actions]

pkg.pr.new

packages

pnpm i https://pkg.pr.new/software-mansion/TypeGPU/typegpu@1311
pnpm i https://pkg.pr.new/software-mansion/TypeGPU/typegpu@89b285cdad7336f1c28714a83789a5e5266d1229

benchmark
view benchmark

commit
view commit

[aleksanderkatan]

There is a little side effect of the float cast we do with the Float32Array

[aleksanderkatan]

All those checks seem to have affected the performance a bit

[iwoplaza]

@aleksanderkatan Try the PR build, sometimes release builds are faster than local builds because of the transform that the bundler performs.

[aleksanderkatan]

@iwoplaza if this is what you meant, then sadly there is no significant change

[mhawryluk]

Well done!! 🚀

[reczkok]

I'm really on the fence about this — I get that the consistency between JS and WGSL is nice, but is it really worth the performance hit (which, as we've established, is kind of important)?

[reczkok]

I'm really on the fence about this — I get that the consistency between JS and WGSL is nice, but is it really worth the performance hit (which, as we've established, is kind of important)?

Maybe we could leave this in and figure out a way to provide users with a very fast mass vector constructor

[aleksanderkatan]

I think we could modify the schema casts for scalar values, so that f32 and f16 are identity, i32 is Math.floor(n), u32 is Math.max(0, Math.floor(n)), and bool is left as is (all plus the undefined check). This would probably increase the performance and leave the most important part of the cast in place. What do you think?

Edit: Another benefit is that we'd avoid the floating point error in parsed wgsl.

[reczkok]

I think we should keep the casting logic to stay compliant with WGSL's behavior. I’m actually curious how much of the performance hit comes from the casting itself versus the overhead of the function calls.

[aleksanderkatan]

@reczkok seems like most of the overhead is in Math functions and array writes.

Before any changes:

When calling schema constructors, but all they do is return the value passed in:

When calling schema constructors, and casting all values properly:

Edit: the results aren't too consistent. Sometimes I get latest with 3500 ops/s, and once it returned 1500 ops/s

[aleksanderkatan]

The benchmark results are now wildly inconsistent. For one test, I get 30 ops/s five times in a row, then I get 11 ops/s five times in a row...
What is consistent is that the PR version gets around 20% less ops/s on the biggest vector creation test than the latest.

I added the coercion to the vector element assignment (through [0] and through .x). I also removed the float value casts, and I simplified a bit the integer casts.