How to convert vector to GPU compatible array?

[akashshinde1418]

Dear Team,

I have a problem that demands a vector to be passed inside the ParallelFor loop for GPU computations. Since I do not know the size of the vector at compile time, I cannot use Array1D/2D/3D or GpuArray. How can I solve this issue?

[atmyers]

Hi Akash,

You need to use amrex::DeviceVector, and capture the data pointer in your lambda function, as here.

[akashshinde1418]

Hi Andrew,
Thank you for the quick reply. I have checked your reference and see that the dataPtr() is used only for 1D vectors. Let us imagine a case where I'm dealing with a 3D vector. I have attached the code for the solution I could develop for this problem. I believe it is an inefficient way to tackle this problem. Can we have a solution where I don't have to explicitly copy the values from host to device rather just use a pointer to access the data from host?

` {
Vector< Vector< Vector< Real>>> Temp( 5, Vector< Vector< Real>>( 5, Vector< Real>( 5, 10.0)));

 for ( MFIter mfi( phi_old); mfi. isValid(); ++ mfi)
 {
 const Box& bx = mfi. validbox();

 amrex:: Gpu:: AsyncVector < Real> flag ( 125, 0.0);

 /*Operation to copy values from 3d Vector Temp to a 1d AsyncVector flag*/

 auto ptr = flag. dataPtr();

 amrex:: ParallelFor( bx, [=] AMREX_GPU_DEVICE( int i, int j, int k)
 {

      /*Do computations with ptr here*/

 });
 }

}`

[WeiqunZhang]

We can add that in the future. For now, you can just do like what's in the example above. That is t3d(r,s,t) is equivalent to Temp[t][s][r].

[WeiqunZhang]

Row-major order is added in #3973

[akashshinde1418]

Thank you for adding the support!

[akashshinde1418]

Dear Weiqun Zhang,
I appreciate your efforts to add the row-major support in "AMReX_TableData.H"; however, the addition has yet to be merged in your development branch as it is pending review. I was looking forward to using Table2D and Table3D in my codes. Can you shed some light on when the addition will be reflected in the development branch?

[WeiqunZhang]

It has been merged.