double-float performance

[sinandredemption]

As soon as my current PR goes green, I will present a PR in which cpp_double_fp_backend is tested against cpp_bin_float of the appropriate size.

In the preliminary tests, it seems to me just a thousand operations is too few to accurately judge how the two types compare. Is their any way we can increase the number of test cases per operator?

CC @ckormanyos @cosurgi

[ckormanyos]

way we can increase the number of test cases per operator?

Great idea. I believe the number of testsis hard-coded to 1000 in a bunch of lines like this in the file performance_test.hpp.

Since you are doing tests relative one type to another, simply jump those lines up to 10,000 or 100,000 I think. If that does not work, please get back to me.

[sinandredemption]

Here are the preliminary results for performance of double-float.

Operator	cpp_bin_float	cpp_double_fp_backend
*	11.114 (0.362158s)	[*1] (0.0325857s)
*(int)	7.27636 (0.242081s)	[*1] (0.0332695s)
*(unsigned long long)	4.3046 (0.306152s)	[*1] (0.0711219s)
*=(unsigned long long)	7.6058 (0.303511s)	[*1] (0.0399052s)
+	5.60299 (0.377083s)	[*1] (0.0673002s)
+(int)	8.00336 (0.440066s)	[*1] (0.0549851s)
+(unsigned long long)	7.48374 (0.466146s)	[*1] (0.0622879s)
+=(unsigned long long)	9.05229 (0.440637s)	[*1] (0.0486769s)
-	7.12818 (0.39518s)	[*1] (0.0554391s)
-(int)	7.69794 (0.447368s)	[*1] (0.0581152s)
-(unsigned long long)	6.90328 (0.455248s)	[*1] (0.0659466s)
-=(unsigned long long)	8.97643 (0.496401s)	[*1] (0.0553005s)
/	4.35787 (4.21081s)	[*1] (0.966254s)
/(int)	1.38088 (0.632865s)	[*1] (0.458306s)
/(unsigned long long)	3.60605 (1.64141s)	[*1] (0.455182s)
/=(unsigned long long)	4.03905 (1.73374s)	[*1] (0.429245s)
construct	3.10884 (0.188684s)	[*1] (0.0606926s)
construct(unsigned long long)	9.86198 (1.43813s)	[*1] (0.145825s)
construct(unsigned)	21.2099 (1.39659s)	[*1] (0.0658462s)
str	1.23162 (0.0016888s)	[*1] (0.0013712s)

A few platform details:

Platform: Windows x64 
Compiler: Microsoft Visual C++ version 14.2
Boost: 107500

EDIT: cpp_bin_float was running at 31 decimal digits of precision

[ckormanyos]

Really nice numbers Fahad. We concentrated so much on quad recently, I was actually thinking that one of the real benefits of this GSoC's results will be the good 'ol cpp_double_fp_backend instantiated for either double or long double.

Great job! @sinandredemption

Cc: @cosurgi and @jzmaddock

[ckormanyos]

@sinandredemption I confirm your benchmark methodology for double-float.

On GCC 9 running on Ubuntu, I got the following raw data with factors of 5-10 speedup on arithmetic ops:

operation	cpp_bin_float	cpp_double_fp_backend
+	0.244161	0.030418
-	0.241686	0.0284256
*	0.125933	0.0163531
/	2.30554	0.473302
str	0.001701	0.002918
+(int)	0.229878	0.0281554
-(int)	0.242622	0.0269989
*(int)	0.118462	0.0134481
/(int)	0.487138	0.218203
construct	0.336057	0.0470241
construct(unsigned)	0.237162	0.0841882
construct(unsigned long long)	0.211154	0.11853
+(unsigned long long)	0.27494	0.027713
-(unsigned long long)	0.250832	0.0278098
*(unsigned long long)	0.126704	0.0154193
/(unsigned long long)	0.552782	0.175357
+=(unsigned long long)	0.261626	0.0279107
-=(unsigned long long)	0.242469	0.0279352
*=(unsigned long long)	0.125763	0.0165767
/=(unsigned long long)	0.53905	0.184863

[ckormanyos]

On GCC 9 running on Ubuntu,

Note to self: Remember to test on MSVC with /fp:fast and/or GCC with -ffast-math. I will integrate some -ffast-math runs in CI and see if that also influences perf numbers.

Cc: @sinandredemption and @cosurgi

[ckormanyos]

This issue is (for the GSoC'21 purposes) done.