[ADOL-C] adolc-cuda vector mode

Kshitij Kulshreshtha kshitij at math.upb.de
Fri Aug 5 07:56:18 EDT 2016


Hi,

I think the computation partials separately in each thread was the
original intent of the code, since there really wasn't any computational
code that used NUMBER_DIRECTIONS at all and defining it caused
problematic code to be compiled.

I've now added the handling of NUMBER_DIRECTIONS in adoublecuda.h in the
git master branch of ADOL-C [1]. I don't know what side effects this
could have though.

Best wishes
Kshitij

[1] https://gitlab.com/adol-c/adol-c.git

On 29.07.2016 17:22, Robert wrote:
> Hello,
> 
> I want to use the vector mode in adol-c's cuda implementation, but I get
> this error when trying to compile:
> 
> 
> nvcc test.cu -std=c++11 -arch=compute_50 -maxrregcount=255 -o test
> /usr/include/adolc/adoublecuda.h(945): error: identifier "DIAG_OUT" is
> undefined
> 
> /usr/include/adolc/adoublecuda.h(949): error: expression must have
> pointer-to-object type
> 
> /usr/include/adolc/adoublecuda.h(954): error: identifier "DIAG_OUT" is
> undefined
> 
> /usr/include/adolc/adoublecuda.h(958): error: expression must have
> pointer-to-object type
> 
> /usr/include/adolc/adoublecuda.h(965): error: identifier "ADOLC_numDir"
> is undefined
> 
> /usr/include/adolc/adoublecuda.h(966): error: identifier "ADOLC_numDir"
> is undefined
> 
> /usr/include/adolc/adoublecuda.h(979): error: identifier "ADOLC_numDir"
> is undefined
> 
> 
> This is a simple sourcecode to replicate the problem (test.cu):
> 
> 
> #define NUMBER_DIRECTIONS 3
> #include <adolc/adoublecuda.h>
> 
> int main() {
>     return 0;
> }
> 
> 
> As a workaround I can calculate each partial derivative in its own cuda
> thread, but I wanted to analyse the performance difference between these
> methods.
> 
> Can someone please help me solve this problem?
> 
> 
> Best regards,
> 
> Robert
> 
> _______________________________________________
> ADOL-C mailing list
> ADOL-C at list.coin-or.org
> http://list.coin-or.org/mailman/listinfo/adol-c

-- 
Dr. Kshitij Kulshreshtha

Institut für Mathematik,
Universität Paderborn,
Warburger Straße 100,
33098 Paderborn.

Büro: TP21.1.21
Besucheradresse:
Technologiepark 21
33098 Paderborn.

Privatanschrift:
Arnikaweg 62
33100 Paderborn.


More information about the ADOL-C mailing list