Skip to content

Fix benchmark checksums, enable whitespace skipping, minor cleanups #9

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
perazz merged 5 commits into from
Mar 26, 2026
Merged

Fix benchmark checksums, enable whitespace skipping, minor cleanups #9

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
084c6db
Fix benchmark: strip leading spaces, add meaningful XOR checksums
perazz Mar 26, 2026
4bd4124
Update README benchmarks with corrected results
perazz Mar 26, 2026
c220f02
Enable FMT_SKIP_WS in DEFAULT_PARSING
perazz Mar 26, 2026
6abe438
Simplify bigint_compare with sign intrinsic
perazz Mar 26, 2026
5c7501c
Remove get_double_bits/get_float_bits wrappers, use transfer directly
perazz Mar 26, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
92 changes: 48 additions & 44 deletions README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -65,66 +65,70 @@ fpm build
fpm test
```

## Profiling

Use `profile_benchmark.sh` to profile the `benchmark_compare` benchmark on macOS:

```bash
./profile_benchmark.sh xctrace # Time Profiler trace
./profile_benchmark.sh sample # CLI sampling report
```

Environment overrides: `REPEAT_COUNT`, `DATA_DIR`, `TRACE_OUT`.

## Benchmarks

Run with `fpm test --profile release --target benchmark_compare` on Apple Silicon (M1 Max).
Data files from [simple_fastfloat_benchmark](https://github.com/lemire/simple_fastfloat_benchmark).
Use `./run_benchmarks.sh` to run the full suite (random + data files) with C++ comparison.

### Random uniform [0,1) -- 100k floats, 2.19 MB
### Random uniform [0,1) -- 100k floats, 2.10 MB

```
netlib (C) : 403.76 MB/s (+/- 1.5 %) 19.24 Mfloat/s
strtod (C) : 786.63 MB/s (+/- 1.3 %) 37.49 Mfloat/s
abseil (C++) : 917.36 MB/s (+/- 0.9 %) 43.72 Mfloat/s
fastfloat (C++) : 1586.85 MB/s (+/- 1.3 %) 75.63 Mfloat/s
ffc (C) : 1650.41 MB/s (+/- 1.3 %) 78.66 Mfloat/s
fortran (fast_float) : 876.76 MB/s (+/- 0.7 %) 41.79 Mfloat/s
fortran (stdlib to_num) : 1066.10 MB/s (+/- 1.0 %) 50.81 Mfloat/s
fortran (str2real) : 609.20 MB/s (+/- 0.7 %) 29.04 Mfloat/s
fortran (read *) : 56.90 MB/s (+/- 0.4 %) 2.71 Mfloat/s
ffc via fortran interop : 1493.30 MB/s (+/- 1.4 %) 71.17 Mfloat/s
```

### canada_short.txt -- 111k lines, 0.70 MB

```
netlib (C) : 405.14 MB/s (+/- 1.2 %) 19.31 Mfloat/s
strtod (C) : 787.20 MB/s (+/- 1.0 %) 37.52 Mfloat/s
abseil (C++) : 919.15 MB/s (+/- 1.0 %) 43.81 Mfloat/s
fastfloat (C++) : 1586.45 MB/s (+/- 1.3 %) 75.61 Mfloat/s
ffc (C) : 1653.72 MB/s (+/- 1.6 %) 78.82 Mfloat/s
fortran (fast_float) : 2394.60 MB/s (+/- 8.8 %) 109.17 Mfloat/s
fortran (stdlib to_num) : 1078.39 MB/s (+/- 2.7 %) 49.16 Mfloat/s
fortran (str2real) : 666.91 MB/s (+/- 0.5 %) 30.40 Mfloat/s
fortran (read *) : 58.58 MB/s (+/- 0.5 %) 2.67 Mfloat/s
ffc via fortran interop : 4284.08 MB/s (+/- 11.0 %) 195.31 Mfloat/s
netlib (C) : 546.31 MB/s (+/- 4.0 %) 101.50 Mfloat/s
strtod (C) : 367.00 MB/s (+/- 0.9 %) 68.19 Mfloat/s
abseil (C++) : 361.36 MB/s (+/- 1.2 %) 67.14 Mfloat/s
fastfloat (C++) : 566.31 MB/s (+/- 1.0 %) 105.22 Mfloat/s
ffc (C) : 704.02 MB/s (+/- 1.1 %) 130.81 Mfloat/s
fortran (fast_float) : 638.99 MB/s (+/- 1.4 %) 118.72 Mfloat/s
fortran (stdlib to_num) : 710.33 MB/s (+/- 3.1 %) 131.98 Mfloat/s
fortran (str2real) : 243.62 MB/s (+/- 0.4 %) 45.27 Mfloat/s
fortran (read *) : 22.14 MB/s (+/- 0.6 %) 4.11 Mfloat/s
ffc via fortran interop : 598.70 MB/s (+/- 4.8 %) 111.24 Mfloat/s
```

### canada.txt -- 111k lines, 1.93 MB
### canada.txt -- 111k lines, 2.04 MB

```
netlib (C) : 385.05 MB/s (+/- 1.3 %) 22.13 Mfloat/s
strtod (C) : 686.85 MB/s (+/- 1.4 %) 39.47 Mfloat/s
abseil (C++) : 868.24 MB/s (+/- 1.6 %) 49.89 Mfloat/s
fastfloat (C++) : 1095.24 MB/s (+/- 1.4 %) 62.94 Mfloat/s
ffc (C) : 1169.13 MB/s (+/- 1.8 %) 67.19 Mfloat/s
fortran (fast_float) : 1011.37 MB/s (+/- 1.3 %) 58.12 Mfloat/s
fortran (stdlib to_num) : 1074.30 MB/s (+/- 2.3 %) 61.74 Mfloat/s
fortran (str2real) : 463.62 MB/s (+/- 1.5 %) 26.64 Mfloat/s
fortran (read *) : 49.77 MB/s (+/- 0.5 %) 2.86 Mfloat/s
ffc via fortran interop : 1089.43 MB/s (+/- 0.9 %) 62.61 Mfloat/s
netlib (C) : 384.11 MB/s (+/- 0.9 %) 22.07 Mfloat/s
strtod (C) : 686.30 MB/s (+/- 1.0 %) 39.44 Mfloat/s
abseil (C++) : 868.22 MB/s (+/- 1.2 %) 49.89 Mfloat/s
fastfloat (C++) : 1095.50 MB/s (+/- 1.1 %) 62.95 Mfloat/s
ffc (C) : 1169.13 MB/s (+/- 1.5 %) 67.19 Mfloat/s
fortran (fast_float) : 1036.86 MB/s (+/- 0.7 %) 59.58 Mfloat/s
fortran (stdlib to_num) : 1038.53 MB/s (+/- 1.0 %) 59.68 Mfloat/s
fortran (str2real) : 452.02 MB/s (+/- 0.4 %) 25.98 Mfloat/s
fortran (read *) : 49.86 MB/s (+/- 0.7 %) 2.87 Mfloat/s
ffc via fortran interop : 1048.10 MB/s (+/- 1.6 %) 60.23 Mfloat/s
```

### mesh.txt -- 73k lines, 0.54 MB
### mesh.txt -- 73k lines, 0.61 MB

```
netlib (C) : 537.08 MB/s (+/- 2.7 %) 73.17 Mfloat/s
strtod (C) : 523.55 MB/s (+/- 1.5 %) 71.32 Mfloat/s
abseil (C++) : 415.44 MB/s (+/- 1.3 %) 56.59 Mfloat/s
fastfloat (C++) : 825.16 MB/s (+/- 1.5 %) 112.41 Mfloat/s
ffc (C) : 948.34 MB/s (+/- 2.5 %) 129.19 Mfloat/s
fortran (fast_float) : 846.78 MB/s (+/- 2.4 %) 115.35 Mfloat/s
fortran (stdlib to_num) : 802.41 MB/s (+/- 1.0 %) 109.31 Mfloat/s
fortran (str2real) : 301.81 MB/s (+/- 0.4 %) 41.11 Mfloat/s
fortran (read *) : 28.43 MB/s (+/- 1.2 %) 3.87 Mfloat/s
ffc via fortran interop : 853.52 MB/s (+/- 0.9 %) 116.27 Mfloat/s
netlib (C) : 537.91 MB/s (+/- 2.1 %) 73.28 Mfloat/s
strtod (C) : 522.81 MB/s (+/- 1.2 %) 71.22 Mfloat/s
abseil (C++) : 415.05 MB/s (+/- 1.3 %) 56.54 Mfloat/s
fastfloat (C++) : 825.10 MB/s (+/- 1.1 %) 112.40 Mfloat/s
ffc (C) : 947.22 MB/s (+/- 1.3 %) 129.04 Mfloat/s
fortran (fast_float) : 885.97 MB/s (+/- 0.9 %) 120.69 Mfloat/s
fortran (stdlib to_num) : 796.45 MB/s (+/- 1.2 %) 108.50 Mfloat/s
fortran (str2real) : 298.61 MB/s (+/- 2.4 %) 40.68 Mfloat/s
fortran (read *) : 28.21 MB/s (+/- 0.6 %) 3.84 Mfloat/s
ffc via fortran interop : 838.82 MB/s (+/- 2.1 %) 114.27 Mfloat/s
```

## Acknowledgements
Expand Down
34 changes: 32 additions & 2 deletions benchmark/benchmark_compare.f90
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -135,6 +135,7 @@ subroutine generate_random_lines(howmany, lines, packed_text, packed_data, offse
do i = 1, nlines
call random_number(x)
write(buf, '(es23.16)') x
buf = adjustl(buf)
n = len_trim(buf)
lines(i)%text = trim(buf)
allocate(lines(i)%c_text(n))
Expand Down Expand Up @@ -267,7 +268,9 @@ subroutine run_benchmark(name, lines, packed_text, packed_data, offsets, lengths
"ffc interop bits = " ]

real(real64) :: min_ns(NCASES), avg_ns(NCASES), checksum(NCASES)
real(real64) :: answer, elapsed_ns, volume_mb, x_f
integer(int64) :: xor_checksum(NCASES)
real(real64), volatile :: answer
real(real64) :: elapsed_ns, volume_mb, x_f
real(real64) :: x_stdlib, x_str2real, x_read
real(c_double) :: x_c
integer :: i, k, r, ios_read
Expand All @@ -280,6 +283,33 @@ subroutine run_benchmark(name, lines, packed_text, packed_data, offsets, lengths
avg_ns = 0.0_real64
checksum = 0.0_real64

! Compute XOR checksums in a single untimed pass
xor_checksum = 0_int64
do i = 1, nlines
call parse_double_range_sub( &
packed_text, &
int(offsets(i), kind=kind(i)) + 1, &
int(offsets(i) + lengths(i), kind=kind(i)), &
x_f, f_result, DEFAULT_PARSING)
if (f_result%outcome == outcomes%OK) &
xor_checksum(B_RSUB) = ieor(xor_checksum(B_RSUB), transfer(x_f, 0_int64))

x_stdlib = 0.0_real64
x_stdlib = to_num(lines(i)%text, x_stdlib)
xor_checksum(B_STDLIB) = ieor(xor_checksum(B_STDLIB), transfer(x_stdlib, 0_int64))

x_str2real = str2real(lines(i)%text)
xor_checksum(B_S2R) = ieor(xor_checksum(B_S2R), transfer(x_str2real, 0_int64))

read(lines(i)%text, *, iostat=ios_read) x_read
if (ios_read == 0) &
xor_checksum(B_READ) = ieor(xor_checksum(B_READ), transfer(x_read, 0_int64))

call ffc_parse_double_c(lines(i)%c_text, len(lines(i)%text), x_c, c_outcome)
if (c_outcome == outcomes%OK%state) &
xor_checksum(B_CLOOP) = ieor(xor_checksum(B_CLOOP), transfer(real(x_c, real64), 0_int64))
end do

call system_clock(count_rate=count_rate)
do r = 1, repeat_count
answer = 0.0_real64
Expand Down Expand Up @@ -358,7 +388,7 @@ subroutine run_benchmark(name, lines, packed_text, packed_data, offsets, lengths
call print_result(labels(k), volume_mb, nlines, min_ns(k), avg_ns(k))
end do
do k = 1, NCASES
write(output_unit, "(a,z16.16)") cksum_labels(k), transfer(checksum(k), 0_int64)
write(output_unit, "(a,z16.16)") cksum_labels(k), xor_checksum(k)
end do
end subroutine run_benchmark

Expand Down
24 changes: 5 additions & 19 deletions src/fast_float_module.F90
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -57,7 +57,7 @@ module fast_float_module
integer(i8), parameter :: FMT_ALLOW_PLUS = int(b'010000000', i8)
integer(i8), parameter :: FMT_SKIP_WS = int(b'100000000', i8)

integer(i8), parameter :: PRESET_GENERAL = ior(FMT_FIXED, FMT_SCIENTIFIC)
integer(i8), parameter :: PRESET_GENERAL = ior(ior(FMT_FIXED, FMT_SCIENTIFIC), FMT_SKIP_WS)
integer(i8), parameter :: PRESET_JSON = ior(ior(FMT_JSON, PRESET_GENERAL), FMT_NO_INFNAN)
integer(i8), parameter :: PRESET_FORTRAN = ior(FMT_FORTRAN, PRESET_GENERAL)

Expand Down Expand Up @@ -868,18 +868,6 @@ end subroutine mul_u64

! ===== Character and digit utilities =====

!> Reinterpret a double as its 64-bit integer representation.
pure elemental integer(i8) function get_double_bits(d)
real(dp), intent(in) :: d
get_double_bits = transfer(d, 0_i8)
end function get_double_bits

!> Reinterpret a float as its 32-bit integer representation.
pure elemental integer(i4) function get_float_bits(f)
real(sp), intent(in) :: f
get_float_bits = transfer(f, 0_i4)
end function get_float_bits

!> Count leading zero bits in a 64-bit integer.
pure elemental integer function count_leading_zeros(x)
integer(i8), intent(in) :: x
Expand Down Expand Up @@ -1886,11 +1874,9 @@ end subroutine bigint_hi64
pure elemental integer function bigint_compare(a, b)
type(bigint), intent(in) :: a, b
integer :: j
if (a%vec%ln > b%vec%ln) then
bigint_compare = 1
return
else if (a%vec%ln < b%vec%ln) then
bigint_compare = -1
j = a%vec%ln - b%vec%ln
if (j/=0) then
bigint_compare = sign(1,j)
return
end if
bigint_compare = 0
Expand Down Expand Up @@ -1942,7 +1928,7 @@ pure elemental subroutine to_extended(vd, f, res)
integer(i8) :: b
integer(i4) :: bi
bi = int(f%mantissa_bits - f%min_exponent, i4)
b = get_double_bits(vd)
b = transfer(vd, b)
if (iand(b, f%exponent_mask) == 0) then
res = adjusted_mantissa( iand(b, f%mantissa_mask), 1 - bi)
else
Expand Down
Loading