Struct SideInfo 

pub struct SideInfo {
    pub blocks_wide: usize,
    pub blocks_tall: usize,
    /* private fields */
}

Per-coefficient rounding errors from quantization.

Each error is in [-0.5, +0.5] and represents how far the continuous (unquantized) DCT coefficient was from its rounded integer value. Positive error means the pre-quantization value was above the integer; negative means below.

Stored as i8 [-127, +127] via error * 254, giving ~0.004 resolution. This is 8x smaller than f64 and 4x smaller than f32, saving 85 MB (12MP) or 341 MB (48MP) compared to the original f64 representation.

Fields

blocks_wide: usize

Number of 8x8 blocks horizontally.

blocks_tall: usize

Number of 8x8 blocks vertically.

Implementations

impl SideInfo

pub fn compute( raw_rgb: &[u8], pixel_width: u32, pixel_height: u32, cover_grid: &DctGrid, qt_values: &[u16; 64], ) -> Self

Compute side information from raw RGB pixels and the cover JPEG.

For each Y-channel 8x8 block:

1. Forward DCT on the original (pre-JPEG) pixels
2. Divide by quantization table (without rounding)
3. error = unquantized_value - cover_integer_coefficient

Errors are clamped to [-0.5, +0.5] for robustness against minor floating-point differences between the platform’s JPEG encoder and our forward DCT implementation.

Luma blocks are computed in strips of 50 block-rows to limit transient memory (~12.9 MB per strip instead of ~97.5 MB for all blocks at once on a 12MP image).

pub fn error_at(&self, flat_idx: usize) -> f32

Get the rounding error at a flat index, decoded from i8 to f32.