module Fpu:Access to low level floating point functions. THIS LIBRARY ONLY WORKS FOR INTEL PROCESSORS.sig
..end
Almost all low level functions are implemented using the x87 functions and x87 rounding modes. There are unfortunately a few problems to understand. The x87 is supposed to be able to return a nearest value and a upper and a lower bound for each elementary operation it can perform. This is not always true. Some functions such as cos(), sin() or tan() are not properly implemented everywhere.
For example, for the angle a= 1.570 796 326 794 896 557 998 981 734 272 092 580 795 288 085 937 5 the following values are computed for cos(a), by (1) the MPFI library (with 128 bits precision), (2) the x87 in low mode, (3) the x87 in nearest mode (default value for the C and Ocaml library on 32 bits linux), (4) the x87 in high mode, (5) the SSE2 implementation (default value for the C and Ocaml library on 64 bits linux):
(1) 6.123 233 995 736 765 886 130 329 661 375 001 464 640 377 798 836e-17
(2) 6.123 031 769 111 885 058 461 925 285 082 049 859 451 216 355 021e-17
(3) 6.123 031 769 111 886 291 057 089 692 912 995 815 277 099 609 375e-17
(4) 6.123 031 769 111 886 291 057 089 692 912 995 815 277 099 609 375e-17
(5) 6.123 233 995 736 766 035 868 820 147 291 983 023 128 460 623 387e-17
The upper bound (4) computed by the x87 is clearly incorrect, as it is lower than the correct value computed by the MPFI library.
The value computed by the SSE2 (5) is much more precise than the one computed by the x87. Unfortunately, there is no way to get an upper and lower bound value, and we are thus stuck with the x87 for computing these (sometimes incorrect) bounds.
The problem here is that the value computed by the standard, C-lib (or ocaml) cos function doesn't always lie in the lower/upper bound interval returned by the x87 functions, and this can be a very serious problem when executing Branch and Bound algorithms which expect the mid-value to be inside the lower/upper interval.
We solved the problem by rewritting the trigonometric functions in order to make them both consistant and correct. We used the following property: when -pi/4<=a<=pi/4 the rounding in 64 bits of the 80 bits low/std/high value returned by the x87 are correct. Moreover, when 0<a<2**53 then (a mod (2Pi_low)) and (a mod (2Pi_high)) are in the same quadrant. Last, (a mod Pi/2_High) <= (a mod Pi/2) <= (a mod Pi/2_Low). With this implementation, the lower and upper bounds are properly set and they are always lower (resp. higher) than the value computed by the standard cos functions on 32 and 64 bits architecture. This rewritting has been done in assembly language and is quite efficient.
Keep in mind that values returned by the standard (C-lib or Ocaml) cos(),
sin() or tan() functions are still
different on 32 and 64 bits architecture. If you want to have a program which
behaves exactly in the same way on both architectures, you can use the Fpu
module fcos
, fsin
or ftan
functions which always return the same values on all
architectures, or even use the Fpu_rename
or Fpu_rename_all
modules to transparently
rename the floating point functions.
The functions are quite efficient (see below). However, they have a serious disadvantage compared to their standard counterparts. When the compiler compiles instruction ''a+.b'', the code of the operation is inlined, while when it compiles ''(fadd a b)'', the compiler generates a function call, which is expensive.
Intel Atom 230 Linux 32 bits
Intel 980X Linux 64 bits
val ffloat : int -> float
val ffloat_high : int -> float
val ffloat_low : int -> float
val fadd : float -> float -> float
val fadd_low : float -> float -> float
val fadd_high : float -> float -> float
val fsub : float -> float -> float
val fsub_low : float -> float -> float
val fsub_high : float -> float -> float
val fmul : float -> float -> float
val fmul_low : float -> float -> float
val fmul_high : float -> float -> float
val fdiv : float -> float -> float
val fdiv_low : float -> float -> float
val fdiv_high : float -> float -> float
val fmod : float -> float -> float
val fsqrt : float -> float
val fsqrt_low : float -> float
val fsqrt_high : float -> float
val fexp : float -> float
val fexp_low : float -> float
val fexp_high : float -> float
val flog : float -> float
val flog_low : float -> float
val flog_high : float -> float
val flog_pow : float -> float -> float
val flog_pow_low : float -> float -> float
val flog_pow_high : float -> float -> float
val fpow : float -> float -> float
val fpow_low : float -> float -> float
val fpow_high : float -> float -> float
val fsin : float -> float
val fsin_low : float -> float
val fsin_high : float -> float
val fcos : float -> float
val fcos_low : float -> float
val fcos_high : float -> float
val ftan : float -> float
val ftan_low : float -> float
val ftan_high : float -> float
val fatan : float -> float -> float
val fatan_low : float -> float -> float
val fatan_high : float -> float -> float
val facos : float -> float
val facos_low : float -> float
val facos_high : float -> float
val fasin : float -> float
val fasin_low : float -> float
val fasin_high : float -> float
val fsinh : float -> float
val fsinh_low : float -> float
val fsinh_high : float -> float
val fcosh : float -> float
val fcosh_low : float -> float
val fcosh_high : float -> float
val ftanh : float -> float
val ftanh_low : float -> float
val ftanh_high : float -> float
val is_neg : float -> bool
BE VERY CAREFUL: using these functions unwisely can ruin all your computations. Remember also that on 64 bits machine these functions won't change the behaviour of the SSE instructions.
When setting the rounding mode to UPWARD or DOWNWARD, it is better to set it immediately back to NEAREST. However we have no guarantee on how the compiler will reorder the instructions generated. It is ALWAYS better to write:
let a = set_high(); let res = 1./.3. in set_nearest (); res;;
The above code will NOT work on linux-x64 where many floating point functions are implemented using SSE instructions. These three functions should only be used when there is no other solution, and you really know what tou are doing, and this should never happen. Please use the regular functions of the fpu module for computations. For example prefer:
let a = fdiv_high 1. 3.;;
PS: The Interval module and the fpu module functions correctly set and restore the rounding mode for all interval computations, so you don't really need these functions.
PPS: Please, don't use them...
val set_low : unit -> unit
val set_high : unit -> unit
val set_nearest : unit -> unit