diegoripley/tippecanoe
1
0
Fork 0
mirror of https://github.com/felt/tippecanoe.git synced 2026-10-02 08:25:40 +02:00
Code Issues Packages Projects Releases Wiki Activity

Replace the Grisu2 float formatter with a C++ port of rsc/fpfmt

Browse Source 
Tippecanoe formatted every double it wrote through milo::dtoa_milo, a
vendored Grisu2. Grisu2 is fast, but it guarantees neither the shortest
digit string nor the correctly rounded one: it only guarantees that what
it prints parses back to the value it came from. In practice it prints a
digit more than necessary about 0.16% of the time, and picks a neighbor
of the correctly rounded digits about 32% of the time.

This ports Russ Cox's fpfmt (https://github.com/rsc/fpfmt) to C++ in
fpfmt/ and formats through it instead. fpfmt is both shortest and
correctly rounded, and it is faster:

  full std::string formatting     Grisu2      fpfmt   speedup
  random bit patterns          156.62 ns   66.83 ns     2.34x
  geo coordinates              124.07 ns   58.62 ns     2.12x
  short decimals                69.37 ns   49.16 ns     1.41x
  small integers                44.18 ns   38.06 ns     1.16x

  digit generation only           Grisu2      fpfmt   speedup
  random bit patterns           90.07 ns   20.81 ns     4.33x
  geo coordinates               80.64 ns   20.18 ns     4.00x
  short decimals                55.61 ns   21.90 ns     2.54x
  small integers                40.23 ns   22.50 ns     1.79x

(Intel Xeon @ 2.80GHz, g++ 13.3 -O3. `make fpfmt-bench` reproduces this,
and `./fpfmt-bench -check` reruns the correctness sweep, which is why
milo/dtoa_milo.h is kept even though nothing links it any more.)

The port is deliberately literal, so it can be diffed against fpfmt.go.
Its Short() agrees bit for bit with the Go original's on 445,640 values
covering powers of ten, small integers and reciprocals, subnormals, and
random bit patterns. Over 38.5 million values, fpfmt::dtoa always round
trips, is never longer than Grisu2's output, and is shorter 61,329 times.

Output is otherwise formatted exactly as before, including the choice
between plain and exponential notation, so 26 expected test outputs
change: some numbers lose digits (-26.170044999999999 becomes
-26.170045), and some have a corrected final digit (9.823748927348929e+55
becomes 9.823748927348928e+55). Every changed token was checked to parse
back to the identical double; none of the values themselves moved.

milo/milo.h, whose only job was to declare the C shim jsonpull calls, is
replaced by fpfmt/fpfmt.h, and the shim is renamed dtoa_shortest.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014wJRAuhMninQE4wK2TUfuZ
This commit is contained in:
Claude
2026-08-31 00:22:35 +00:00
parent 4f2621186a
commit 7127e49c86
58 changed files with 1813 additions and 219 deletions
Download Patch File Download Diff File Expand all files Collapse all files
CHANGELOG.md
+13
View File
@@ -1,3 +1,16 @@
# 2.83.0
* Replace the Grisu2 (`dtoa_milo`) float formatter with a C++ port of Russ
Cox's `fpfmt` (https://github.com/rsc/fpfmt), in `fpfmt/`. Grisu2 is fast but
neither always shortest nor always correctly rounded; `fpfmt` is both, and is
1.2x to 2.3x faster end to end (4.3x for digit generation alone) on this
hardware. Some numbers in tile output and tilestats therefore now print with
fewer digits (`-26.170044999999999` becomes `-26.170045`) or with a corrected
final digit (`9.823748927348929e+55` becomes `9.823748927348928e+55`). Every
such value still parses back to exactly the same double, so this changes only
the spelling, never the number. `make fpfmt-bench` rebuilds the head-to-head
comparison against the old implementation.
# 2.82.0
* Fix corruption of a JSON array when a non-final element was removed from it.