Struct BucketData 

pub struct BucketData {

    pub base_indices: Vec<u32>,
    pub bucket_pointers: Vec<u32>,
    pub bucket_sizes: Vec<u32>,
    pub bucket_values: Vec<u32>,
    pub window_starts: Vec<u32>,
    pub window_counts: Vec<u32>,
    pub num_windows: u32,
    pub num_active_buckets: u32,
    pub num_dispatched: u32,
    pub orig_bucket_values: Vec<u32>,
    pub orig_window_starts: Vec<u32>,
    pub orig_window_counts: Vec<u32>,
    pub reduce_starts: Vec<u32>,
    pub reduce_counts: Vec<u32>,
    pub has_chunks: bool,
    pub bucket_width: usize,
}

Bucket sorting result for GPU MSM dispatch.

Uses a Structure-of-Arrays layout: each array is uploaded as a separate storage<read_only> GPU buffer. This avoids struct padding issues in WGSL and allows independent buffer bindings per kernel.

When sub-bucket chunking is active (has_chunks == true), the parallel arrays (bucket_pointers, bucket_sizes, bucket_values, window_starts, window_counts) describe sub-buckets (dispatched units), not logical buckets. The reduce_starts/reduce_counts arrays map original bucket indices to their sub-bucket ranges for a post-aggregation reduction pass.

Invariants:

• bucket_pointers[i] is the starting index in base_indices for sub-bucket i
• bucket_sizes[i] is the count of points in sub-bucket i
• bucket_values[i] is the scalar weight for sub-bucket i (in [1, 2^(c-1)])
• window_starts[w] is the first sub-bucket index belonging to window w
• window_counts[w] is the number of sub-buckets in window w
• reduce_starts[j] is the first sub-bucket index for original bucket j
• reduce_counts[j] is the number of sub-buckets for original bucket j

Fields

base_indices: Vec<u32>

bucket_pointers: Vec<u32>

Sub-bucket pointers into base_indices (length = num_dispatched).

bucket_sizes: Vec<u32>

Sub-bucket sizes (length = num_dispatched).

bucket_values: Vec<u32>

Sub-bucket values, same as parent’s value (length = num_dispatched).

window_starts: Vec<u32>

Sub-bucket window starts (length = num_windows).

window_counts: Vec<u32>

Sub-bucket counts per window (length = num_windows).

num_windows: u32

num_active_buckets: u32

Number of original (logical) buckets.

num_dispatched: u32

Number of dispatched sub-buckets (>= num_active_buckets when chunking occurs).

orig_bucket_values: Vec<u32>

Original bucket values for weight/subsum passes (length = num_active_buckets).

orig_window_starts: Vec<u32>

Original window starts for weight/subsum passes (length = num_windows).

orig_window_counts: Vec<u32>

Original window counts for weight/subsum passes (length = num_windows).

reduce_starts: Vec<u32>

Start offset in the dispatch buffer for each original bucket.

reduce_counts: Vec<u32>

Number of sub-buckets for each original bucket.

has_chunks: bool

Whether any bucket was split into sub-buckets.

bucket_width: usize