i remember doing a project years ago where i had to manually write cuda & it was genuinely unbearable. you had to reason about shit like: - host vs device memory - explicit allocation & copies - grids, blocks, warps, & threads - shared vs. global vs local memory - synchronization - bizarre race conditions - kernel launch configuration more stuff i can’t even remember anymore. & then after getting it to compile, you could still write something technically correct that performed like absolute…