There are several questions on this site about implementing fast CSV/TSV export for machine numbers without loss of precision. Recently I discovered that Export merely applies ToString[#, InputForm] & to every number in the table for conversion it into a string:

Trace[ExportString[{-7756.0224337393065}, "TSV"], ToString] // Flatten
{ToString[-7756.0224337393065, InputForm], "-7756.0224337393065"}

Unfortunately there seems to be no built-in alternative to InputForm in Mathematica for obtaining the correct digits of machine numbers, even RealDigits isn't precise:

x = 44802.395880518656;
ToExpression@ToString[x, InputForm] - x
FromDigits@RealDigits[x] - x    


InputForm works correctly but very slow. From the other hand, it shouldn't be too difficult to re-implement the algorithm behind this conversion as a compilable function in order to get huge speed-up.

Does anybody know what the algorithm is? Probably it isn't proprietary and can be found in the literature.

  • $\begingroup$ Strongly related: stackoverflow.com/q/7153979/590388 $\endgroup$ Commented May 16, 2018 at 7:55
  • $\begingroup$ RealDigits appears to drop the digit which is not precise. x = 44802.395880518658; InputForm@x returns 44802.395880518656 $\endgroup$
    – LLlAMnYP
    Commented May 16, 2018 at 12:38
  • $\begingroup$ Consider using InternalDoubleToString. It offers a ~3-4x speedup: arr = RandomReal[{0, 50000}, 10^6]; AbsoluteTiming[ToString[#, InputForm] & /@ arr] // First AbsoluteTiming[InternalDoubleToString /@ arr] // First but it also loses the last digit. $\endgroup$
    – LLlAMnYP
    Commented May 16, 2018 at 13:04
  • 2
    $\begingroup$ You may find interesting the algorithms referenced in this answer. $\endgroup$
    – ilian
    Commented May 18, 2018 at 19:07
  • $\begingroup$ Tangentially related: "Is NumberForm double rounding numbers?" $\endgroup$ Commented Sep 8, 2019 at 11:25

1 Answer 1


This section of my answer previously showed that Export[fname, data, "TSV"] is incredibly slow (~500 seconds for exporting 10^5 * 10 reals), but this is easily worked around by WriteString[fname, ExportString[data, "TSV"]].

All of this is less relevant for CSV export, as I demonstrated below and the foolishly missed while going on to ramble about TSV export. But even for CSV a small speedup is possible with this workaround.

I would assume that what you mean by "the algorithm behind InputForm" is defined by low-level code, something like

#include <stdio.h>
#include <float.h>


printf("%.*f", DBL_DIG, x);

Where the implementation of that is defined, I do not know, but I would expect that to follow the IEEE-754 standard (Wiki, Official paywall).

Any (finite) binary float can be represented by a terminating sequence of decimal digits, of approximately 52 decimal digits of precision, but (provably) 17 digits are always enough. However...

N @ MachinePrecision
(* 15.9546 *)

1 + $MachineEpsilon
(* 1. *)

InputForm @ %
(* 1.0000000000000002 *)

SetPrecision[%, 16]
(* 1.000000000000000 *)

The second paragraph of section 5.12 of 2008 IEEE-754 clearly states

Implementations shall provide conversions between each supported binary format and external decimal character sequences such that, under roundTiesToEven, conversion from the supported format to external decimal character sequence and back recovers the original floating-point representation, except that a signaling NaN might be converted to a quiet NaN. See 5.12.1 and 5.12.2 for details.

  • InputForm is compliant with this, though I would expect 44802.39588051865 to be represented as 44802.395880518648 and not to remain as ...865.

  • RealDigits is not a true conversion to an "external decimal representation", though one could expect it to behave as such. It discards the digit of least precision. Annoyingly, FromDigits converts back to a rational number.

  • Internal`DoubleToString is quite fast (but not orders of magnitude faster) and it suffers from the same problem as RealDigits -- it discards the ULP. I'd say, this is blatant non-compliance with the standard. I can no longer reproduce this on my machine. Perhaps DoubleToString is compliant after all.

  • ToString@SetPrecision[x, 17] should give always correct results and is not too bad in terms of speed. In fact, you don't even need ToString here.


  SetPrecision[RandomReal[{0, 1000}, {10^5, 10}], 17], "CSV"] // AbsoluteTiming
(* {20.1076, "test.csv"} *)
Export["test.csv", RandomReal[{0, 1000}, {10^5, 10}], "CSV"] // AbsoluteTiming
(* {15.307, "test.csv"} *)
With[{arr = RandomReal[{0, 1000}, {10^5, 10}]}, 
  Export["test.csv", arr, "CSV"] // AbsoluteTiming]
(* {15.2228, "test.csv"} *)
With[{arr = SetPrecision[RandomReal[{0, 1000}, {10^5, 10}], 17]}, 
  Export["test.csv", arr, "CSV"] // AbsoluteTiming]
(* {19.3445, "test.csv"} *)
RandomReal[{0, 1000}, {10^5, 10}] // AbsoluteTiming // First
With[{arr = RandomReal[{0, 1000}, {10^5, 10}]}, 
  SetPrecision[arr, 17] // AbsoluteTiming // First]
(* 0.026734 *)
(* 0.840021 *)

Basically, generating 10^6 numbers is fast (its timing is negligible), converting them to 17 digits of precision is also very fast, and frankly, I don't think that the unpacked array causes the slowdown in the CSV export. What's more likely, is that there is simply more bytes to write.

Bottom line: Export is simply that slow.

Here's a not-very-thoroughly-though-out alternative which is already quite a bit faster:

With[{rand = RandomReal[{0, 1000}, {10^5, 10}]},
 Block[{arr = SetPrecision[rand, 17], str},
   str = Map[ToString, arr, {-1}];
   Export["test.dat", StringRiffle[str], "String"]] // AbsoluteTiming]
(* {7.75655, "test.dat"} *)

This can all be much faster if you're prepared to downgrade from 17 to 16 digits of precision - the SetPrecision and ToString can be replaced by a single Internal`DoubleToString.

You may likely be interested in studying Put, PutAppend (which, by the way, generate output in InputForm) as well as Write and related functions. They'll likely be somewhat complicated to get to work properly, but can offer large speedups.

LibraryLink solution

This is a work-in-progress which I'll fine-tune a bit over the next few days. In principle, Henrik Schumacher has already made a solution here, but OP apparently had some troubles getting it to work, so I'll show a bare-bones example to get started.

I'm assuming, that a C-compiler is set up and <<CCompilerDriver` has been executed. Start by creating a directory and creating a file, ExportTable.c in it with the following code:

 #include "WolframLibrary.h"
 #include <float.h>
 #include <stdio.h>

 DLLEXPORT int fastExportCSV(WolframLibraryData libData, mint Argc, MArgument *Args, MArgument Res) {

        double* data;
        char* fname = MArgument_getUTF8String(Args[0]);
        mint i = 0;
        mint j = 0;
//        mint len;  // correct handling of array dimensions still to be implemented

        MTensor in = MArgument_getMTensor(Args[1]);
//      len = libData->MTensor_getFlattenedLength(in);
        data = libData->MTensor_getRealData(in);

        FILE * ofile;
        ofile = fopen(fname, "w");

        for(i = 0; i<100000; i++) {
            for(j = 0; j<10; j++) {
                fprintf(ofile, "%.*f,", DBL_DIG, data[10*i + j]);
            fprintf(ofile, "\n");

        return LIBRARY_NO_ERROR;

Now go over to Mathematica to set everything up there:

lib = CreateLibrary[{"ExportTable.c"}, "ExportTable"]
fastExportCSV = LibraryFunctionLoad[
    {"UTF8String", {Real, 2, "Constant"}}, 

Because the dimensions of the array are hard-coded right now, you have to feed a 2-dimensional Real array with at least 10^6 elements and the program will write them as a 10^5*10 CSV table. Like so:

  RandomReal[{-100, 100}, {10^5, 10}]] // AbsoluteTiming
(* {0.958084, "D:\\Larkin\\Dev\\LLTutorial\\test2.dat"} *)

Of course, this is duplication of the linked answer, but in simpler, purely C code.

  • $\begingroup$ In the linked question I've already found a faster alternative to Internal`DoubleToString: it is DecimalForm combined with SetPrecision. The problem is that it doesn't allow to recover the original decimal representation of the numbers exactly (as well as Internal`DoubleToString). $\endgroup$ Commented May 17, 2018 at 8:02
  • $\begingroup$ @AlexeyPopkov That faster alternative is quite a fine example of what in Russia we call костыли. Anyway, your words "the original decimal representation" suggest some confusion about they way, numbers are handled, at least on an intuitive level. In the case of your linked question, the numbers cannot and will not be "exactly" represented (your complaint about the changing last digit). If you want an exact decimal representation, you should input numbers with arbitrary precision, e.g. -56705.907687414737`17. (cont'd) $\endgroup$
    – LLlAMnYP
    Commented May 17, 2018 at 8:14
  • $\begingroup$ Note, then, that x = -56705.907687414737; y = -56705.907687414737`17; ExportString[{x,y},"TSV"] will return -56705.907687414736 -56705.907687414737. Besides that, you say that DecimalForm[-38741.90768741473, Infinity] changes the last digit to 2. I, unfortunately, don't have v11.3 to test with, but both InputForm and DoubleToString show me the exact input, regardless if the last digit is 2 or 3. This seems to be a bug of DecimalForm $\endgroup$
    – LLlAMnYP
    Commented May 17, 2018 at 8:20
  • $\begingroup$ @AlexeyPopkov anyway, it would be helpful if you specify your problem more completely (including the background). Why exactly do you care specifically about the decimal representation of your numbers when they, likely have been obtained through calculations with binary double-precision floats? If that is not the case, then calculations ought to have been done in arbitrary precision, in which case the numbers should be stored as such. In this case, their InputForm and/or ToString, etc conserves their decimal representation to the arbitrary precision that was given. $\endgroup$
    – LLlAMnYP
    Commented May 17, 2018 at 8:24
  • $\begingroup$ The numbers are generated by another application and then Imported by Mathematica. Export allows to generate exactly the same file, but very slow. My goal is to add two columns into that TSV file. $\endgroup$ Commented May 17, 2018 at 8:27

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.