Private prover input on float numbers

[fattaneh88]

Hi,
I understand that to work with float numbers I need to write a code on SFDL. I also know pepper has a feature for private prover input like what we have in "apps_sfdl/genome_snp_freq.c". My question is how can I used the private prover input feature, and at the same type work with float numbers. I wrote a simple code:

program LR {
//Constants
//Dimention of inputs
const SIZE = 3;
//An upper bound for number of multiplication in exponentiation operations defined by a FOR LOOP
const UPPER_B = 10000;
//Precision
const L = 3;
//These two are defined based on L
const TEN_P_L = 1000;
const E_P_TEN_P_ML = 1.0010005;

//Types
type Param = struct {float<32,32>[SIZE] W, float<32,32>[1] B};
type Input = struct {float<32,32>[SIZE] X, hash_t WB};
type Output = struct {float P};

//Main method, called "output"
function Output output (Input In){
    var Param p;
hashget(&p, &(In.WB));

    var int<32> i;
    var float sum;
var float exp;
sum = 0;
exp = 1;

//Computing WX+B
    for (i = 0 to SIZE-1)
	sum = sum + In.X[i] * p.W[i];
    sum = sum + p.B[0];

//Appling precision L 
    sum = sum * TEN_P_L;

//Computing e^{WX+B}
    for (i = 1 to UPPER_B)
	if (sum >= i)
		exp = exp * E_P_TEN_P_ML;
    //Returning e^{WX+B} as the output
    output.P = exp;        
}

}

but I got the following error:
LOG: Building executables

• compile apps_sfdl/LR_S.sfdl --metrics --cql -b 0 -w 10240 -t ZAATAR -db-hash-func ggh -db-num-addresses 1024 -ram-cell-num-bits 32 -db-hash-num-bits 1216 -db-thr-num-addresses-naive 32768 -fast-ram-address-width 32 -fast-ram-word-width 64
  metric_num_lines_in_sfdl LR_S 38
  metric_num_lines_in_source LR_S 38
  make[1]: Entering directory `/home/fara/pepper/compiler/frontend'
  ant compile
  Buildfile: /home/fara/pepper/compiler/frontend/build.xml

compile:

BUILD SUCCESSFUL
Total time: 0 seconds
make[1]: Leaving directory `/home/fara/pepper/compiler/frontend'
WARNING: --cstdarithtruncate is disabled, so type errors will warn and arithmetic is not ANSI C compliant
Compiling ../pepper/apps_sfdl/LR_S.sfdl
java.text.ParseException: Unknown type hash_t
at SFE.Compiler.SFECompiler.compileDataType(SFECompiler.java:670)
at SFE.Compiler.SFECompiler.compileStructFields(SFECompiler.java:1041)
at SFE.Compiler.SFECompiler.compileKnownType(SFECompiler.java:776)
at SFE.Compiler.SFECompiler.compileDataType(SFECompiler.java:662)
at SFE.Compiler.SFECompiler.compileType(SFECompiler.java:618)
at SFE.Compiler.SFECompiler.compileTypeDeclarations(SFECompiler.java:575)
at SFE.Compiler.SFECompiler.compileProgram(SFECompiler.java:334)
at zcc.ZCC.compile(ZCC.java:207)
at zcc.ZCC.main(ZCC.java:127)
Error in line 15: Unknown type hash_t
make: *** [apps_sfdl_gen/LR_S.cpp] Error 1

I appreciate if you can advise me on that.
Thanks,
Fattaneh

[maxhowald]

Unfortunately, constructs such as hashget and exo_compute which enable private prover input are only supported by the C compiler right now. If you need both, you'll need to port floats to C or these functions to SFDL.

I don't have much time to look at this myself, but if you or someone else are interested in taking it on, my recommendation would be to port floats to C in the pequin repository.

To do this, you'll need to modify the frontend part of the compiler in compiler/frontend, to parse float specifiers in .c files and generate the appropriate .circuit file. Probably, you'll want to write a simple floating-point program in both C and SFDL, compile the SFDL version with pepper, and then hack on the compiler in pequin until it can produce a matching .circuit file from the C version.

If you're interested in this and want more info or run in to problems, feel free to open more issues!

[fattaneh88]

Thank you very much for your advice. - Fattaneh

…

On Wed, Dec 26, 2018 at 1:10 PM Max Howald ***@***.***> wrote: Unfortunately, constructs such as hashget and exo_compute which enable private prover input are only supported by the C compiler right now. If you need both, you'll need to port floats to C or these functions to SFDL. I don't have much time to look at this myself, but if you or someone else are interested in taking it on, my recommendation would be to port floats to C in the pequin repository. To do this, you'll need to modify the frontend part of the compiler in compiler/frontend, to parse float specifiers in .c files and generate the appropriate .circuit file. Probably, you'll want to write a simple floating-point program in both C and SFDL, compile the SFDL version with pepper, and then hack on the compiler in pequin until it can produce a matching .circuit file from the C version. If you're interested in this and want more info or run in to problems, feel free to open more issues! — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#4 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AHfWZYsDTnasCai_zmuFCNNK4DbrMOdyks5u8-XbgaJpZM4ZdPCL> .

-- F.B

[fattaneh88]

Hi Max, Thanks again for all your suggestion till know. I decided to add floating point implementation to C in the pequin repository. Based on my exploration, I think this is the place that I need to start adding my updates: https://github.com/pepper-project/pequin/blob/master/compiler/frontend/src/ccomp/parser_hw/CCompiler.java I appreciate if I can get your comment on that. Regards, Fattaneh On Thu, Dec 27, 2018 at 6:50 PM Fattaneh Bayat <[email protected]> wrote:

…

Thank you very much for your advice. - Fattaneh On Wed, Dec 26, 2018 at 1:10 PM Max Howald ***@***.***> wrote: > Unfortunately, constructs such as hashget and exo_compute which enable > private prover input are only supported by the C compiler right now. If you > need both, you'll need to port floats to C or these functions to SFDL. > > I don't have much time to look at this myself, but if you or someone else > are interested in taking it on, my recommendation would be to port floats > to C in the pequin repository. > > To do this, you'll need to modify the frontend part of the compiler in > compiler/frontend, to parse float specifiers in .c files and generate > the appropriate .circuit file. Probably, you'll want to write a simple > floating-point program in both C and SFDL, compile the SFDL version with > pepper, and then hack on the compiler in pequin until it can produce a > matching .circuit file from the C version. > > If you're interested in this and want more info or run in to problems, > feel free to open more issues! > > — > You are receiving this because you authored the thread. > Reply to this email directly, view it on GitHub > <#4 (comment)>, > or mute the thread > <https://github.com/notifications/unsubscribe-auth/AHfWZYsDTnasCai_zmuFCNNK4DbrMOdyks5u8-XbgaJpZM4ZdPCL> > . > -- F.B

-- F.B

[maxhowald]

Yes, that should be the right place to start. The float type specifier and float constant specifications should already by part of the grammar. You'll probably want to compare *.circuit files generated by the existing Pepper release from SFDL files with your implementation.