Faster floating point comparisons #814

mborland · 2022-08-05T19:29:12Z

No description provided.

mborland · 2022-08-07T15:42:34Z

@NAThompson Since you got this to work previously for 128 bit float do you see anything obvious I am missing? Error for incomplete type here.

NAThompson · 2022-08-07T20:57:33Z

Since you got this to work previously for 128 bit float do you see anything obvious I am missing?

IIRC, this was something I looked up in the quadmath manual. I couldn't find it though . . .

mborland · 2022-08-23T14:19:40Z

@NAThompson This is good for review. I pulled 128 bit support into it's own header otherwise you would have to link quadmath anytime you used next.hpp.

AreaZR · 2022-10-13T23:40:12Z

Any updates on this?

include/boost/math/special_functions/fast_float_distance.hpp

mborland · 2022-10-15T22:35:09Z

Here are the before and after benchmarks @NAThompson :

Original performance (Boost 1.80.0):

 Unable to determine clock rate from sysctl: hw.cpufrequency: No such file or directory
 This does not affect benchmark measurements, only the metadata output.
 2022-10-15T15:24:07-07:00
 Running ./new_next_performance
 Run on (10 X 24.0916 MHz CPU s)
 CPU Caches:
   L1 Data 64 KiB
   L1 Instruction 128 KiB
   L2 Unified 4096 KiB (x10)
 Load Average: 1.86, 2.53, 5.83
 ---------------------------------------------------------------------------------
 Benchmark                                       Time             CPU   Iterations
 ---------------------------------------------------------------------------------
 float_distance<float>/2/real_time            61.4 ns         61.4 ns      9074469
 float_distance<float>/4/real_time            61.7 ns         61.7 ns     11384150
 float_distance<float>/8/real_time            61.4 ns         61.4 ns     10814604
 float_distance<float>/16/real_time           61.7 ns         61.7 ns     11348376
 float_distance<float>/32/real_time           61.4 ns         61.4 ns     11387167
 float_distance<float>/64/real_time           61.6 ns         61.6 ns     11131932
 float_distance<float>/128/real_time          61.4 ns         61.4 ns     11382029
 float_distance<float>/256/real_time          61.4 ns         61.4 ns     11307649
 float_distance<float>/512/real_time          61.4 ns         61.4 ns     11376048
 float_distance<float>/1024/real_time         61.4 ns         61.4 ns     11355748
 float_distance<float>/2048/real_time         61.8 ns         61.8 ns     11373776
 float_distance<float>/4096/real_time         61.4 ns         61.4 ns     11382368
 float_distance<float>/8192/real_time         61.4 ns         61.4 ns     11353453
 float_distance<float>/16384/real_time        61.4 ns         61.4 ns     11378298
 float_distance<float>/real_time_BigO        61.48 (1)       61.47 (1)
 float_distance<float>/real_time_RMS             0 %             0 %
 float_distance<double>/2/real_time           55.6 ns         55.6 ns     12580218
 float_distance<double>/4/real_time           55.6 ns         55.6 ns     12577835
 float_distance<double>/8/real_time           55.6 ns         55.6 ns     12564909
 float_distance<double>/16/real_time          56.2 ns         56.2 ns     12554909
 float_distance<double>/32/real_time          56.0 ns         56.0 ns     12544381
 float_distance<double>/64/real_time          55.6 ns         55.6 ns     12566488
 float_distance<double>/128/real_time         55.6 ns         55.6 ns     12499581
 float_distance<double>/256/real_time         55.6 ns         55.6 ns     12565661
 float_distance<double>/512/real_time         56.1 ns         56.1 ns     12550023
 float_distance<double>/1024/real_time        55.8 ns         55.8 ns     12568603
 float_distance<double>/2048/real_time        55.6 ns         55.6 ns     12546049
 float_distance<double>/4096/real_time        55.6 ns         55.6 ns     12528525
 float_distance<double>/8192/real_time        55.9 ns         55.9 ns     12563030
 float_distance<double>/16384/real_time       56.0 ns         56.0 ns     12447644
 float_distance<double>/real_time_BigO       55.78 (1)       55.78 (1)
 float_distance<double>/real_time_RMS            0 %             0 %

 New performance:

 Unable to determine clock rate from sysctl: hw.cpufrequency: No such file or directory
 This does not affect benchmark measurements, only the metadata output.
 2022-10-15T15:31:37-07:00
 Running ./new_next_performance
 Run on (10 X 24.122 MHz CPU s)
 CPU Caches:
   L1 Data 64 KiB
   L1 Instruction 128 KiB
   L2 Unified 4096 KiB (x10)
 Load Average: 2.12, 2.17, 4.26
 ---------------------------------------------------------------------------------
 Benchmark                                       Time             CPU   Iterations
 ---------------------------------------------------------------------------------
 float_distance<float>/2/real_time            15.8 ns         15.8 ns     42162717
 float_distance<float>/4/real_time            15.9 ns         15.9 ns     44213877
 float_distance<float>/8/real_time            15.8 ns         15.8 ns     43972542
 float_distance<float>/16/real_time           15.8 ns         15.8 ns     44209456
 float_distance<float>/32/real_time           15.8 ns         15.8 ns     44200244
 float_distance<float>/64/real_time           15.8 ns         15.8 ns     44239293
 float_distance<float>/128/real_time          15.8 ns         15.8 ns     44171202
 float_distance<float>/256/real_time          15.8 ns         15.8 ns     44241507
 float_distance<float>/512/real_time          15.9 ns         15.8 ns     44230034
 float_distance<float>/1024/real_time         15.8 ns         15.8 ns     44241554
 float_distance<float>/2048/real_time         15.8 ns         15.8 ns     44220802
 float_distance<float>/4096/real_time         15.8 ns         15.8 ns     44220441
 float_distance<float>/8192/real_time         15.9 ns         15.9 ns     44213994
 float_distance<float>/16384/real_time        15.8 ns         15.8 ns     44215413
 float_distance<float>/real_time_BigO        15.83 (1)       15.83 (1)
 float_distance<float>/real_time_RMS             0 %             0 %
 float_distance<double>/2/real_time           15.5 ns         15.5 ns     45098165
 float_distance<double>/4/real_time           15.6 ns         15.6 ns     45065465
 float_distance<double>/8/real_time           15.5 ns         15.5 ns     45058733
 float_distance<double>/16/real_time          15.8 ns         15.7 ns     45078404
 float_distance<double>/32/real_time          15.5 ns         15.5 ns     44832734
 float_distance<double>/64/real_time          15.5 ns         15.5 ns     45077303
 float_distance<double>/128/real_time         15.5 ns         15.5 ns     45067255
 float_distance<double>/256/real_time         15.5 ns         15.5 ns     45073844
 float_distance<double>/512/real_time         15.6 ns         15.6 ns     45109342
 float_distance<double>/1024/real_time        15.5 ns         15.5 ns     44845180
 float_distance<double>/2048/real_time        15.5 ns         15.5 ns     45051846
 float_distance<double>/4096/real_time        15.5 ns         15.5 ns     45064317
 float_distance<double>/8192/real_time        15.5 ns         15.5 ns     45115653
 float_distance<double>/16384/real_time       15.5 ns         15.5 ns     45067642
 float_distance<double>/real_time_BigO       15.54 (1)       15.54 (1)
 float_distance<double>/real_time_RMS            0 %             0 %

NAThompson · 2022-10-15T22:35:58Z

test/test_fast_float_distance.cpp

+
+    BOOST_MATH_ASSERT(fast_float_distance(float_advance(val, 4), val) == -4);
+    BOOST_MATH_ASSERT(fast_float_distance(float_advance(val, -4), val) == 4);
+    if(std::numeric_limits<T>::is_specialized && (std::numeric_limits<T>::has_denorm == std::denorm_present))


Doesn't the fast float distance unconditionally require denorm support?

It's just type punning so as long as the floating point and fixed width integer are the same number of bits you should be fine. I don't know of a platform without denorm support to test on.

NAThompson · 2022-10-15T22:37:10Z

@mborland : Does it need a fast_float_distance header? Could we not just improve the performance of the current implementation?

@jzmaddock : This is looking like it's about ready to go; might want to take a look.

mborland · 2022-10-15T22:40:03Z

@mborland : Does it need a fast_float_distance header? Could we not just improve the performance of the current implementation?

The float and double cases improve upon the current implementation. I could not get the __float128 case to work without forcing the user to link -lquadmath if using GCC which would be an unwelcome breaking change.

NAThompson · 2022-10-16T05:03:08Z

I could not get the __float128 case to work without forcing the user to link -lquadmath if using GCC which would be an unwelcome breaking change.

Wait, I thought we had to link libquadmath to use __float128 . . .

include/boost/math/special_functions/fast_float_distance.hpp

Co-authored-by: Alexander Grund <Flamefire@users.noreply.github.com>

NAThompson · 2023-03-02T02:57:29Z

I could not get the __float128 case to work without forcing the user to link -lquadmath if using GCC which would be an unwelcome breaking change.

Could we use a judicious __has_include to workaround?

Also, 99% of the value will be from float and double . . .

mborland · 2023-03-02T03:00:04Z

Having the include isn't enough in this case. You have to explicitly link to -lquadmath.

mborland · 2023-04-13T12:15:05Z

@jzmaddock Can you please take a look at this one? @NAThompson hit me up about merging this.

This was linked to issues Aug 5, 2022

Floating-Point Comparison Performance #775

Open

Can boost::math::float_distance be sped up? #359

Open

mborland force-pushed the fast_next branch 2 times, most recently from 6c3977f to d5baadc Compare August 7, 2022 00:12

mborland force-pushed the fast_next branch 4 times, most recently from 2c2b13a to 86b0dbd Compare August 14, 2022 00:20

mborland force-pushed the fast_next branch from 86b0dbd to 08d2c34 Compare August 22, 2022 02:55

NAThompson reviewed Oct 14, 2022

View reviewed changes

include/boost/math/special_functions/fast_float_distance.hpp Show resolved Hide resolved

mborland added 9 commits October 15, 2022 15:11

Add specialized float distance for type float

fb37db7

Reinterpret cast breaks strict aliasing rules use memcpy

5fddf08

Add specialized float distance for type double

4660844

Improve error handling

cc0d6e9

Offer 128 bit support

00d2875

Add standalone support

0a333eb

Fix GCC 5 error for static_assert without message

0a6b08c

Fix type used in standalone mode

a481675

Include performance file and results

a1fe973

mborland force-pushed the fast_next branch from f1444c2 to a1fe973 Compare October 15, 2022 22:33

NAThompson reviewed Oct 15, 2022

View reviewed changes

Flamefire reviewed Dec 31, 2022

View reviewed changes

include/boost/math/special_functions/fast_float_distance.hpp Outdated Show resolved Hide resolved

Update include/boost/math/special_functions/fast_float_distance.hpp

73bb192

Co-authored-by: Alexander Grund <Flamefire@users.noreply.github.com>

ryanelandt mentioned this pull request Aug 23, 2023

More efficient bisection for 1D Newton root finder #1012

Open

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Faster floating point comparisons #814

Faster floating point comparisons #814

mborland commented Aug 5, 2022

mborland commented Aug 7, 2022

NAThompson commented Aug 7, 2022

mborland commented Aug 23, 2022

AreaZR commented Oct 13, 2022

mborland commented Oct 15, 2022

NAThompson Oct 15, 2022

mborland Oct 15, 2022

NAThompson commented Oct 15, 2022

mborland commented Oct 15, 2022

NAThompson commented Oct 16, 2022

NAThompson commented Mar 2, 2023

mborland commented Mar 2, 2023

mborland commented Apr 13, 2023

Faster floating point comparisons #814

Are you sure you want to change the base?

Faster floating point comparisons #814

Conversation

mborland commented Aug 5, 2022

mborland commented Aug 7, 2022

NAThompson commented Aug 7, 2022

mborland commented Aug 23, 2022

AreaZR commented Oct 13, 2022

mborland commented Oct 15, 2022

NAThompson Oct 15, 2022

Choose a reason for hiding this comment

mborland Oct 15, 2022

Choose a reason for hiding this comment

NAThompson commented Oct 15, 2022

mborland commented Oct 15, 2022

NAThompson commented Oct 16, 2022

NAThompson commented Mar 2, 2023

mborland commented Mar 2, 2023

mborland commented Apr 13, 2023