Files
tippecanoe/fpfmt/bench.cpp
T
Claude 7127e49c86 Replace the Grisu2 float formatter with a C++ port of rsc/fpfmt
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
2026-08-31 00:22:35 +00:00

279 lines
7.5 KiB
C++

// Head-to-head comparison of fpfmt::dtoa() against the Grisu2 implementation
// (milo::dtoa_milo) that it replaced.
//
// make fpfmt-bench && ./fpfmt-bench # timings
// ./fpfmt-bench -check [n] # correctness sweep
//
// -check verifies, over structured edge cases plus n random doubles, that
// every fpfmt::dtoa() result parses back to the value it came from, and
// reports how its digit strings compare with Grisu2's.
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <algorithm>
#include <chrono>
#include <cmath>
#include <cstdint>
#include <random>
#include <string>
#include <vector>
#include "fpfmt/fpfmt.hpp"
#include "milo/dtoa_milo.h"
using clk = std::chrono::steady_clock;
static double from_bits(uint64_t b) {
double d;
memcpy(&d, &b, sizeof(d));
return d;
}
static uint64_t bits_of(double d) {
uint64_t b;
memcpy(&b, &d, sizeof(b));
return b;
}
// significant_digits counts the significant decimal digits in a formatted
// number: leading zeros and zeros trailing the last nonzero digit don't count.
static int significant_digits(const std::string &str) {
std::string s = str;
size_t e = s.find('e');
if (e != std::string::npos) {
s = s.substr(0, e);
}
int n = 0, trailing = 0;
bool seen = false;
for (char c : s) {
if (c >= '1' && c <= '9') {
n += trailing + 1;
trailing = 0;
seen = true;
} else if (c == '0' && seen) {
trailing++;
}
}
return n == 0 ? 1 : n;
}
struct dataset {
const char *name;
std::vector<double> v;
};
static std::vector<dataset> make_datasets(size_t n, std::mt19937_64 &rng) {
std::vector<dataset> sets;
{
// Uniform random bit patterns: the worst case, nearly all 17 digits.
dataset s{"random bit patterns", {}};
while (s.v.size() < n) {
double d = from_bits(rng());
if (std::isfinite(d) && d != 0) {
s.v.push_back(d);
}
}
sets.push_back(std::move(s));
}
{
// What tippecanoe mostly formats: longitudes and latitudes.
dataset s{"geo coordinates", {}};
std::uniform_real_distribution<double> u(-180, 180);
while (s.v.size() < n) {
s.v.push_back(u(rng));
}
sets.push_back(std::move(s));
}
{
// Attribute values with only a few significant digits.
dataset s{"short decimals", {}};
std::uniform_int_distribution<int> m(1, 999999);
std::uniform_int_distribution<int> e(-4, 4);
while (s.v.size() < n) {
s.v.push_back(m(rng) * pow(10.0, e(rng)));
}
sets.push_back(std::move(s));
}
{
// IDs, counts, zoom levels.
dataset s{"small integers", {}};
std::uniform_int_distribution<int> u(0, 1000000);
while (s.v.size() < n) {
s.v.push_back((double) u(rng));
}
sets.push_back(std::move(s));
}
return sets;
}
static volatile size_t sink;
template <typename F>
static double best_ns(const std::vector<double> &v, int reps, F f) {
double best = 1e300;
for (int r = 0; r < reps; r++) {
auto t0 = clk::now();
size_t acc = 0;
for (double d : v) {
acc += f(d);
}
auto t1 = clk::now();
sink += acc;
double ns = std::chrono::duration<double, std::nano>(t1 - t0).count() / (double) v.size();
best = std::min(best, ns);
}
return best;
}
static int check(long n) {
uint64_t tested = 0, roundtrip_fail = 0, differ = 0, fpfmt_shorter = 0, milo_shorter = 0, same_length = 0;
std::string first_fail;
auto one = [&](double d) {
tested++;
std::string a = fpfmt::dtoa(d);
std::string b = milo::dtoa_milo(d);
if (strtod(a.c_str(), NULL) != d) {
roundtrip_fail++;
if (first_fail.empty()) {
char buf[128];
snprintf(buf, sizeof(buf), "bits=%016llx -> %s", (unsigned long long) bits_of(d), a.c_str());
first_fail = buf;
}
}
if (a != b) {
differ++;
int na = significant_digits(a), nb = significant_digits(b);
if (na < nb) {
fpfmt_shorter++;
} else if (na > nb) {
milo_shorter++;
} else {
same_length++;
}
}
};
// Non-finite and zero cases have to agree exactly.
static const double special[] = {0.0, -0.0, INFINITY, -INFINITY, NAN};
for (double d : special) {
std::string a = fpfmt::dtoa(d), b = milo::dtoa_milo(d);
if (a != b) {
printf("MISMATCH on special value: fpfmt=%s milo=%s\n", a.c_str(), b.c_str());
return 1;
}
}
static const double edge[] = {
1.0, -1.0, 0.1, 0.2, 0.3, 0.5, 1.5, 2.0, 3.0,
1e-7, 1e-6, 1e-5, 1e-1, 1e20, 1e21, 1e22, 1e23,
5e-324, 1e-323, 2.2250738585072014e-308, 2.2250738585072011e-308,
1.7976931348623157e308, 123456789.0, 3.14159265358979,
9007199254740992.0, 9007199254740993.0, 1e100, 1e-100};
for (double d : edge) {
one(d);
}
for (int i = -320; i <= 308; i++) {
double d = pow(10.0, i);
if (std::isfinite(d) && d != 0) {
one(d);
}
}
for (int i = 1; i < 100000; i++) {
one((double) i);
one(1.0 / i);
one(-(double) i);
}
// Every float32 bit pattern, subsampled, promoted to double.
for (uint64_t u = 0; u < (1ull << 32); u += 521) {
float f;
uint32_t b = (uint32_t) u;
memcpy(&f, &b, sizeof(f));
if (std::isfinite(f) && f != 0) {
one((double) f);
}
}
std::mt19937_64 rng(12345);
for (long i = 0; i < n; i++) {
double d = from_bits(rng());
if (std::isfinite(d) && d != 0) {
one(d);
}
}
std::uniform_real_distribution<double> lat(-90, 90), lon(-180, 180);
for (long i = 0; i < n / 4; i++) {
one(lat(rng));
one(lon(rng));
}
printf("values tested %llu\n", (unsigned long long) tested);
printf("fpfmt failed to round trip %llu\n", (unsigned long long) roundtrip_fail);
printf("output differs from Grisu2 %llu\n", (unsigned long long) differ);
printf(" fpfmt used fewer digits %llu\n", (unsigned long long) fpfmt_shorter);
printf(" Grisu2 used fewer digits %llu\n", (unsigned long long) milo_shorter);
printf(" same digit count, last differs %llu\n", (unsigned long long) same_length);
if (!first_fail.empty()) {
printf("first round trip failure: %s\n", first_fail.c_str());
}
return (roundtrip_fail != 0 || milo_shorter != 0) ? 1 : 0;
}
int main(int argc, char **argv) {
if (argc > 1 && strcmp(argv[1], "-check") == 0) {
return check(argc > 2 ? atol(argv[2]) : 20000000L);
}
size_t n = (argc > 1) ? (size_t) atol(argv[1]) : 2000000;
int reps = (argc > 2) ? atoi(argv[2]) : 5;
std::mt19937_64 rng(2024);
std::vector<dataset> sets = make_datasets(n, rng);
printf("%zu values per dataset, best of %d runs\n\n", n, reps * 2);
printf("full std::string formatting\n");
printf("%-22s %14s %14s %9s\n", "dataset", "Grisu2 ns", "fpfmt ns", "speedup");
for (auto &s : sets) {
// Interleave the two so they share cache and clock conditions.
double tm = best_ns(s.v, reps, [](double d) { return milo::dtoa_milo(d).size(); });
double tf = best_ns(s.v, reps, [](double d) { return fpfmt::dtoa(d).size(); });
tm = std::min(tm, best_ns(s.v, reps, [](double d) { return milo::dtoa_milo(d).size(); }));
tf = std::min(tf, best_ns(s.v, reps, [](double d) { return fpfmt::dtoa(d).size(); }));
printf("%-22s %14.2f %14.2f %8.2fx\n", s.name, tm, tf, tm / tf);
}
printf("\ndigit generation only, without the std::string\n");
printf("%-22s %14s %14s %9s\n", "dataset", "Grisu2 ns", "fpfmt ns", "speedup");
for (auto &s : sets) {
double tm = best_ns(s.v, reps, [](double d) {
static std::string b;
b.clear();
if (d == 0) {
return (size_t) 1;
}
if (d < 0) {
d = -d;
}
int length, K;
milo::Grisu2(d, b, &length, &K);
return (size_t) (length + K);
});
double tf = best_ns(s.v, reps, [](double d) {
if (d == 0) {
return (size_t) 1;
}
uint64_t dd;
int p;
fpfmt::shortest(d, &dd, &p);
return (size_t) (fpfmt::digits(dd) + p);
});
printf("%-22s %14.2f %14.2f %8.2fx\n", s.name, tm, tf, tm / tf);
}
return 0;
}