Production de hpc tracer #594
Conversation
…BLE_OPENMP to in src/CMakeLists.txt
src/oce_tracer_mod.F90
Outdated
| do nz = 1, mesh%nl-1 | ||
| if (tracers%data(tr_num)%AB_order==2) then | ||
| tracers%data(tr_num)%valuesAB(nz, n) =-(0.5_WP+epsilon)*tracers%data(tr_num)%valuesold(1, nz, n)+(1.5_WP+epsilon)*tracers%data(tr_num)%values(nz, n) | ||
| elseif (tracers%data(tr_num)%AB_order==3) then | ||
| tracers%data(tr_num)%valuesAB(nz, n) =5.0_WP*tracers%data(tr_num)%valuesold(2, nz, n)-16.0_WP*tracers%data(tr_num)%valuesold(1, nz, n)+23.0_WP*tracers%data(tr_num)%values(nz, n) | ||
| tracers%data(tr_num)%valuesAB(nz, n) =tracers%data(tr_num)%valuesAB(nz, n)/12.0_WP | ||
| end if | ||
| end do |
There was a problem hiding this comment.
This change is where you place the loop out of the if condition, correct?
| do n=1, partit%myDim_nod2D+partit%eDim_nod2D | ||
| ! AB interpolation | ||
| if (tracers%data(tr_num)%AB_order==2) then | ||
| tracers%data(tr_num)%valuesAB(:, n) =-(0.5_WP+epsilon)*tracers%data(tr_num)%valuesold(1, :, n)+(1.5_WP+epsilon)*tracers%data(tr_num)%values(:, n) |
There was a problem hiding this comment.
Just to that I understand, can you let me know if the following is true @basava70?:
If we run this on GPUs one kernel is taking care of performing this operation for the whole vector, but instead we want that one kernel operates that formula for each element in the vector correct? (And that's what you are doing with your change).
@suvarchal and @basava70, would this change slow down the code for CPUs?
| !$OMP PARALLEL DO DEFAULT(SHARED) PRIVATE(elem, elnodes, nz, nzmin, nzmax) | ||
| #else | ||
| !$ACC UPDATE DEVICE(gradient_sca) | ||
| !$ACC parallel loop private(elnodes) |
There was a problem hiding this comment.
isn't this missing a collapse(2) or would that create some sort of conflict/race-condition related to the elnodes being defined private?
There was a problem hiding this comment.
I see... the nz loop depends on number of levels which varies between different elements so the loops cannot be collapsed. So you can ignore my question above.
| @@ -105,7 +144,12 @@ SUBROUTINE tracer_gradient_elements(ttf, partit, mesh) | |||
| #include "associate_mesh_def.h" | |||
| #include "associate_part_ass.h" | |||
| #include "associate_mesh_ass.h" | |||
| #ifndef ENABLE_OPENACC | |||
| !$OMP PARALLEL DO DEFAULT(SHARED) PRIVATE(elem, elnodes, nz, nzmin, nzmax) | |||
There was a problem hiding this comment.
Why are there so many variables private for OpenMP and only elnodes for OpenACC? (a question out of curiosity)
|
Why was this closed? |
See: #589 (comment) |
Creating a draft pull request for Tracer porting. I will keep updating in this pull request from now. When I reach a mile stone, I can make it a final pull request with updated computation timings.
Also, this is created on top of the changes from production_DE_openacc_test branch.