Fix ensemble dataset loading
It's row-major, not column-major (i.e. the LAST dimension is contiguous, not the first). RIP muh cpu cache
- Id
- 01024158e0df30933a30ef58d3fed68eaf7fa18f
- Author
- Caio
- Commit time
- 2025-11-28T13:45:11+01:00
Modified chuva/src/lib.rs
.expect("valid extents spec");
precip.get_values_into(&mut buf, selector)?;
- // FIXME Is this not laid out how I think it is?
- // I'm finding it difficult to compare with the
- // smaller model
// FIXME getfattr zomgwtfbbq
// https://github.com/Unidata/netcdf-c/blob/6038ed2c4b8f53fbe38792d65cfca983c6c08907/libdispatch/dinfermodel.c#L1619
- //
- // Each chunk contains the 20 different model predictions for
- // the current `time` slice
- let (chunks, remainder) = buf[..].as_chunks_mut::<ENS_SIZE>();
- assert!(remainder.is_empty());
- for (idx, chunk) in chunks.iter_mut().enumerate() {
- // The simplest way to choose which of the outputs to use
- // is getting the median value. My VUA decided to bias
- // for rain (i.e.: I'd rather it tells me that it's gonna
- // rain but it doesn't instead of the other way around),
- // so I'm using the 70th percentile
- chunk.sort_unstable();
- let offset = (idx * STEPS) + time;
- data[offset] = f32::from(chunk[13]) * 0.01;
+ let mut ens_members = [0u16; ENS_SIZE];
+ for x in 0..WIDTH {
+ for y in 0..HEIGHT {
+ // Passing this lint makes the code look worse. z is
+ // used to compute the offset, not just for indexing
+ #[expect(clippy::needless_range_loop)]
+ for z in 0..ENS_SIZE {
+ let offset = z * WIDTH * HEIGHT + y * WIDTH + x;
+ ens_members[z] = buf[offset];
+ }
+ ens_members.sort_unstable();
+ let offset = (x * WIDTH + y) * STEPS + time;
+ data[offset] = f32::from(ens_members[13]) * 0.01;
+ }
}
}