v1.3.0

v1.3.0

Fixed and tested: an upstream test audit

We compared our tests with the eccodes, libaec and openjpeg test suites. That found 80 behaviours we never tested; 21 of them were live bugs. All are fixed and covered by tests.

Round-trips

• WriteMessage is now byte-exact on every fixture. It used to drop three things:
  • the Section 4 pv array (eccodes rejected the rewritten files)
  • Section 1 octets 22 onward
  • the reduced lat/lon point-count list
• Multi-field messages that repeat from Section 2, 3 or 4 decode and rewrite byte-exact. New fields: Field.SharesSection3 and Field.Section2.

Decoding

• Negative forecastTime is read as sign-magnitude.
• A missing level scale factor means the level is used unscaled (it gave 10^127).
• Constant fields (0 bits per value) decode to the reference value R, without the decimal scale factor, as eccodes does.
• PNG at 25–32 bits per value no longer loses data to alpha premultiplication.
• Legacy template numbers 5.40000 and 5.41000 are decoded.
• Bit-map indicator 254 reuses the previous bitmap.
• Inconsistent counts between sections are errors.

Coordinates and lookups

• One shared grid geometry for GridCoordinates, NearestValue, BilinearValue and Sampler:
  • every scanning mode
  • every shape of the Earth, including oblate Lambert
  • basic angle and subdivisions
  • reduced Gaussian sub-areas and reduced lat/lon grids
  • rotated grids' angle of rotation
  • the 0/360 seam and Gaussian rows
• Longitudes are in [0, 360). Inconsistent grid descriptions return errors.
• For projected grids scanned in a mode other than 0x40, the first grid point is the first point in data order, as WMO defines it. eccodes treats it as the south-west corner; the two agree for 0x40, the common case.

Encoding

• The reference value is never above the minimum. It used to give errors of up to 32.
• Encoders reject NaN, infinities, values outside float32 range, and 0 bits per value on varying data. Complex packing stores NaN as missing.
• Constant input encodes with 0 bits per value.
• All-NaN input to complex packing stays missing, and the complex encoder no longer drops an unstored bias.
• AEC/CCSDS: the encoder output now matches libaec byte for byte; it miscoded a run of exactly 5 zero blocks.
• JPEG2000:
  • the decoder is ported from OpenJPEG (tiles, layers, progression orders, all code-block styles);
  • 9/7 and multi-component streams return explicit errors;
  • the encoder writes valid tile-part lengths, so Jasper reads its output;
  • encoding is limited to 30 bits per value.

Reading streams

• Decoder skips junk and GRIB1 messages between messages; a GRIB1 message gives an error wrapping ErrNotEdition2.
• It checks each message's total length.
• A truncated message gives ErrTruncated; a clean end gives io.EOF.
• No more panics on malformed templates.

New API

• Field.SharesSection3, Field.Section2
• Section1.Reserved, Section1.ValidateReferenceTime
• Section6.PreviousBitmap
• MissingScaleFactor
• Section4.Describe(discipline)
• Code-table names now cover every code in WMO table version 35.

Compatibility notes

• Section1 holds a []byte, so it can no longer be compared with ==.
• Inputs that were silently accepted now return errors. That covers invalid encoder input, inconsistent grids, mismatched counts, and a message length that doesn't match the declared total.

Tests

Every test runs from committed data, so anyone can run the full suite:

• an eccodes-derived reference for 317 fields (exact value fingerprints, sampled coordinates, metadata, pv arrays, nearest-point answers)
• golden encoder output that eccodes decoded
• libaec's CCSDS test vectors
• JPEG2000 streams checked with OpenJPEG and Jasper

With eccodes built locally:

• GRIB2_WRITE_ECCODES_REFERENCE=1 go test -run TestEccodesReference_Write . regenerates the data.
• TestEccodesReference_UpToDate checks the committed data still matches eccodes.

v1.2.0

v1.2.0

New: Sampler for reusable bilinear weights

BilinearValue unpacks the whole field on every call (~408 µs per point on an RDPS rotated grid), which rules it out for rendering, regridding, or reading the same points from many fields on one grid.

• NewSampler(f) builds once per grid; Weights(lat, lon) returns the four corner indices and fractions, and Interpolate applies them to the Values() of any field on that grid.
• Supports templates 3.0, 3.1 and regular 3.40.
• BilinearWeights.Outside flags points beyond the grid edge (clamped to the edge cell, as before).
• NearestCorner() returns the heaviest corner for categorical fields such as precipitation type.
• Sampler.Matches(f) checks that a field shares the grid.
• BilinearValue is now built on the same weights/interpolate steps with the same arithmetic order, so its results are unchanged.

Verification

Sampler results are bit-identical to BilinearValue, errors included, across 69 decodable fixture grids (57 on 3.0, 4 on 3.1, 8 on 3.40). The existing suite passes unchanged.

lookup	cost per point
BilinearValue	408 µs
Sampler.Weights + Interpolate	0.21 µs

v1.1.1

v1.1.1

Bug fix

• Scale factor sign-magnitude encoding: Fixed the root cause of all RDPS isobaric level bugs. The WMO spec encodes scale factors as sign-magnitude (byte 0x84 = -4), not two s complement (byte 0x84 = -124). The standard formula now works for all producers without hacks.

Testing overhaul

Added eccodes cross-validation and integration test suite (372 tests, up from 294).

Real-world fixtures from GFS and RDPS (previously 0% CMC data). Oracle tests comparing every parsed field against live eccodes output. Integration tests simulating real user workflows including RDPS NearestValue queries, isobaric level verification, 40-field GFS parsing, and write/re-read round-trips.

v1.1.0

v1.1.0

New features

• NearestValue haversine distance — all grid types now use great-circle distance to pick the closest point, matching the eccodes approach. Correct at grid edges where projection-space rounding fails.
• O(1) NearestValue for Template 3.1 (rotated lat/lon) — converts target to rotated coordinates, finds 4 surrounding points, picks nearest by haversine. Turns seconds into microseconds on 1.2M point RDPS grids.
• BilinearValue for Template 3.1 — bilinear interpolation in rotated coordinate space.
• GridCoordinates canonical ordering — now returns coordinates in the same order as Values(), regardless of scanning mode.
• Product templates 4.2, 4.11, 4.12 — derived ensemble forecast, ensemble + time interval, derived ensemble + time interval. Covers GEFS/ECMWF ensemble statistics.
• Grid template 3.90 — Space View / Satellite Perspective (GOES-R, Himawari).
• SurfaceLevel/SurfaceLevelHPa helpers — handle both GFS-style (raw Pa) and RDPS-style (compact hPa with non-standard scale factors).
• CMC parameter codes — SkyTransparencyIndex (0,19,38) and SeeingIndex (0,19,39).
• Fuzz tests — 9 fuzz targets covering all parsers (j2k, aec, complex, run-length, message, JPEG2000, CCSDS, simple, PNG).
• Hardened against malformed input — bounds-checked extractBits, 100M allocation caps, iteration limits on state machines, 256MB section size limit.

Bug fixes

• NearestValue returned wrong index for rotated grids with non-zero scanning mode (row inversion)
• GridCoordinates returned raw scanning order instead of canonical order
• SurfaceLevelHPa double-division for RDPS isobaric encoding
• pixelToUint negative shift panic when bits > 16
• Template330 longitude fields now correctly use sign-magnitude encoding
• gridScanParams now handles Template390
• Section0 reserved bytes preserved on round-trip
• WriteMessage patches TotalLength after writing
• Scale factor fields for surfaces changed to signed int8

Testing

• 294 tests, 0 failures
• Exhaustive brute-force agreement tests for NearestValue (59 points on rotated grids, 10 Canadian cities on real RDPS data)
• Cross-validated against eccodes 2.47.0

v1.0.4

v1.0.4

Bug fixes

• NearestValue index for rotated lat/lon grids (Template 3.1): Fixed canonical grid origin computation to use scanning mode flags instead of min/max heuristics. The O(1) path now matches brute-force search exactly on real RDPS data (1.2M point grid).
• GridCoordinates ordering: Now returns coordinates in canonical order (matching Values()), fixing the root cause where GridCoordinates[i] did not correspond to Values()[i] for non-zero scanning modes.
• SurfaceLevelHPa for RDPS isobaric encoding: Fixed double-division bug where CMC's compact hPa values were divided by 100 twice.

New features

• O(1) NearestValue and BilinearValue for Template 3.1 (rotated lat/lon)
• CMC local parameter codes: SkyTransparencyIndex (0,19,38), SeeingIndex (0,19,39)
• SurfaceLevel(), SurfaceLevelHPa(), SurfaceType() helpers with fallback for non-standard scale factors

Testing

• Regression test comparing O(1) vs brute-force search on Template31 grids
• Verified on real RDPS 1.2M point grid: NearestValue(32.715, -117.161) returns 291.59K (18.44C) matching GFS and Open-Meteo

v1.0.0

v1.0.0 — Initial Release

Pure Go GRIB2 reader/writer. No CGo, no external dependencies.

Compression formats (read + write)

• Simple packing (5.0)
• Complex packing with spatial differencing (5.2/5.3)
• IEEE floating point (5.4)
• JPEG2000 (5.40)
• PNG (5.41)
• CCSDS/AEC (5.42)
• Run-length (5.200, decode only)

Grid templates

Lat/lon, rotated lat/lon, Mercator, polar stereographic, Lambert conformal, Gaussian (regular + reduced), space view/satellite

High-level API

• Field.Values() with automatic bitmap application and scanning mode reordering
• Field.GridCoordinates() for all projection types
• Field.NearestValue() and Field.BilinearValue() for point queries
• WMO code table lookups for disciplines, parameters, centres, surface types
• Streaming multi-message decoder

Quality

• 274 tests, cross-validated against eccodes 2.47.0
• Fuzz tested (9 targets)
• Hardened against malformed input