- zmij::to_shortest() keeps the last digit apart from the 15 or 16 digits
before it, and dtoa_impl::write_shortest() converts those at once: with
SSE2 on x86-64 and NEON on 64-bit Arm (no CPU check needed), else eight
digits at a time. The decimal point is inserted in the register; reading
the digits back right after storing them stalled store forwarding.
- json::dump() writes floats and integers straight into its write buffer
instead of copying them from number_buffer, and integers below 10^16
eight digits at a time.
- json_view's dump() writes doubles from their bits and tokens of up to 15
digits through the same code; its own NEON writer is removed.
to_chars() on canada.json: 35 -> 29 ns per double (x86-64), 18.8 -> 14.2 ns
(Apple M1); json::dump() of canada.json 16-29% faster, of citm_catalog.json
14-19%.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
dump() writes doubles with the conversion of Zmij by Victor Zverovich
(https://github.com/vitaut/zmij, MIT), ported to C++11 in
detail/conversions/zmij.hpp: the shortest decimal in the rounding
interval, the closest one if there are several. Grisu2 does not always
find the shortest digits; about 0.14% of random doubles are now written
differently (0.08% with fewer digits, 0.06% with the closest last
digit); short decimals such as 0.1 or 2555.56 are not affected. float
keeps Grisu2.
The layout of doubles is unchanged, but written differently: the digits
are converted eight at a time (the BCD conversion of Xiang JunBo, as in
Zmij) and stored with one byte swap per eight digits; leading and
trailing zeros are counted from those bytes; and the layouts of
format_buffer() are written with fixed-size moves instead of per-digit
loops and moves of the buffer (to_chars() uses a local buffer if the
caller's is shorter than the 41 bytes this may write).
The powers of ten come from the table for number parsing, adjusted
where it holds them rounded up, and from the compressed tables of Zmij
beyond 10^308. json::dump() gets faster on floats: canada -53%,
numbers -46%, mesh -37%, marine_ik -30%.
Tests: the powers of ten recomputed with a small big-integer; for random
doubles, all powers of two and of ten and their neighbors, and boundary
values: the output reads back as the same value, no decimal with one
digit fewer does, the layout equals that of format_buffer() for the same
digits, and (C++17) the digits equal those of std::to_chars.
The size ratios of canada.json in unit-binary_formats.cpp and one
expectation in unit-to_chars.cpp change with the shorter output.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
Niels Lohmann