Hi,

I’m building a slab‑allocator‑style helper that pre‑allocates one big CudaSlice<u8> and hands out sub‑slices which I immediately CudaSlice::transmute() into typed CudaView<T> objects (f32, f64, etc.). I have a few concerns:

• What alignment requirements should I follow?
• In terms of coalescing / transaction size, should I add any padding?
• Are there architecture‑specific mis‑alignment or transaction‑width issues I should be aware of?

Thanks for your time !