[skyler-ruiter]

Title: Fix 2-D/3-D compressed size missing the rate-array alignment padding

Problem

The 2-D and 3-D compress entry points report

*cmpSize = (size_t)glob_sync + blockNum;

but the kernels place the block payloads after the per-block rate array padded to a
multiple of 4 bytes:

const uint rate_ofs = (blockNum + 3) / 4 * 4;   // src/cuSZp_kernels_{2D,3D}_{f32,f64}.cu

So the reported size is short by (-blockNum) mod 4 bytes (0-3). Any caller that
stores or transfers exactly cmpSize bytes (the example CLI does both: it copies
cmpSize bytes to the host and writes them to the .cuszp file) drops the end of
the stream. Decompression derives offsets from the rate bytes, not from cmpSize, so
it reads the missing bytes as zero. Those final bytes are the most significant bit
plane of the last block with a payload, so up to 16-24 values in that block lose
their top quantization bit and miss the error bound by a power-of-two multiple of
2*eb.

The 1-D entry points are correct: they add the padded count, which matches the 1-D
kernels' offset.

Reproduction (cuSZp v3.0.0 and current master f581dcf)

• CESM-ATM 1800x3600, -m fixed 2-D, REL 1e-4: blockNum = 225*450 = 101250, 2 bytes
  lost; 16 values exceed the bound, the largest by about 1.3e5 x eb (the CLI's own check,
  which allows 1.1 x eb, reports them).
• HURR 100x500x500 and S3D 500^3 (3-D): blockNum is 390625 and 1953125, 3 bytes lost.
• Over a 186-field SDRBench corpus at REL 1e-2/1e-3/1e-4, every one of 362 severe bound
  violations in fixed, plain and outlier modes (2-D and 3-D, f32 and f64) has this cause.
  With the change below, all 362 cells satisfy the bound (worst 1.009 x eb; the only
  remaining overshoots are float64 rounding at <= 1.0000001 x eb). Compressed sizes grow by
  0-3 bytes.

Fix

Report the padded rate-array length in all 12 2-D/3-D compress entry points (fixed,
plain, outlier x 2-D, 3-D x f32, f64):

*cmpSize = (size_t)glob_sync + (blockNum + 3) / 4 * 4;

This matches rate_ofs in the corresponding kernels and the 1-D entry points.

[codecov]

Welcome to Codecov 🎉

Once you merge this PR into your default branch, you're all set! Codecov will compare coverage reports and display results in all future pull requests.

Thanks for integrating Codecov - We've got you covered ☂️

[hyfshishen]

Hi Skyler, I just merged the PR. Thanks for the great effort!