11f02d56fSJunchao Zhang static const char help[] = "Test PetscSF with integers and MPIU_2INT \n\n"; 21f02d56fSJunchao Zhang 31f02d56fSJunchao Zhang #include <petscvec.h> 41f02d56fSJunchao Zhang #include <petscsf.h> 51f02d56fSJunchao Zhang #include <petscdevice.h> 61f02d56fSJunchao Zhang 71f02d56fSJunchao Zhang int main(int argc, char *argv[]) 81f02d56fSJunchao Zhang { 91f02d56fSJunchao Zhang PetscInt n, n2, N = 12; 101f02d56fSJunchao Zhang PetscInt *indices; 111f02d56fSJunchao Zhang IS ix, iy; 121f02d56fSJunchao Zhang VecScatter vscat; 131f02d56fSJunchao Zhang Vec x, y; 141f02d56fSJunchao Zhang PetscInt rstart, rend; 151f02d56fSJunchao Zhang PetscInt *xh, *yh, *xd, *yd; 161f02d56fSJunchao Zhang PetscDeviceContext dctx; 171f02d56fSJunchao Zhang 181f02d56fSJunchao Zhang PetscFunctionBeginUser; 191f02d56fSJunchao Zhang PetscCall(PetscInitialize(&argc, &argv, NULL, help)); 201f02d56fSJunchao Zhang PetscCall(VecCreateFromOptions(PETSC_COMM_WORLD, NULL, 1, PETSC_DECIDE, N, &x)); 211f02d56fSJunchao Zhang PetscCall(VecDuplicate(x, &y)); 221f02d56fSJunchao Zhang PetscCall(VecGetLocalSize(x, &n)); 231f02d56fSJunchao Zhang 241f02d56fSJunchao Zhang PetscCall(VecGetOwnershipRange(x, &rstart, &rend)); 251f02d56fSJunchao Zhang PetscCall(ISCreateStride(PETSC_COMM_WORLD, n, rstart, 1, &ix)); 261f02d56fSJunchao Zhang PetscCall(PetscMalloc1(n, &indices)); 27*6497c311SBarry Smith for (PetscInt i = rstart; i < rend; i++) indices[i - rstart] = i / 2; 281f02d56fSJunchao Zhang PetscCall(ISCreateGeneral(PETSC_COMM_WORLD, n, indices, PETSC_OWN_POINTER, &iy)); 291f02d56fSJunchao Zhang // connect y[0] to x[0..1], y[1] to x[2..3], etc 301f02d56fSJunchao Zhang PetscCall(VecScatterCreate(y, iy, x, ix, &vscat)); // y has roots, x has leaves 311f02d56fSJunchao Zhang 321f02d56fSJunchao Zhang PetscCall(PetscDeviceContextGetCurrentContext(&dctx)); 331f02d56fSJunchao Zhang 341f02d56fSJunchao Zhang // double the allocation since we will use MPIU_2INT later 351f02d56fSJunchao Zhang n2 = 2 * n; 361f02d56fSJunchao Zhang PetscCall(PetscDeviceMalloc(dctx, PETSC_MEMTYPE_HOST, n2, &xh)); 371f02d56fSJunchao Zhang PetscCall(PetscDeviceMalloc(dctx, PETSC_MEMTYPE_HOST, n2, &yh)); 381f02d56fSJunchao Zhang PetscCall(PetscDeviceMalloc(dctx, PETSC_MEMTYPE_DEVICE, n2, &xd)); 391f02d56fSJunchao Zhang PetscCall(PetscDeviceMalloc(dctx, PETSC_MEMTYPE_DEVICE, n2, &yd)); 401f02d56fSJunchao Zhang 411f02d56fSJunchao Zhang for (PetscInt i = 0; i < n; i++) { 421f02d56fSJunchao Zhang xh[i] = xh[i + n] = i + rstart; 431f02d56fSJunchao Zhang yh[i] = yh[i + n] = i + rstart; 441f02d56fSJunchao Zhang } 451f02d56fSJunchao Zhang PetscCall(PetscDeviceMemcpy(dctx, xd, xh, sizeof(PetscInt) * n2)); 461f02d56fSJunchao Zhang PetscCall(PetscDeviceMemcpy(dctx, yd, yh, sizeof(PetscInt) * n2)); 471f02d56fSJunchao Zhang 481f02d56fSJunchao Zhang PetscCall(PetscSFReduceWithMemTypeBegin(vscat, MPIU_INT, PETSC_MEMTYPE_DEVICE, xd, PETSC_MEMTYPE_DEVICE, yd, MPI_SUM)); 491f02d56fSJunchao Zhang PetscCall(PetscSFReduceEnd(vscat, MPIU_INT, xd, yd, MPI_SUM)); 501f02d56fSJunchao Zhang PetscCall(PetscDeviceMemcpy(dctx, yh, yd, sizeof(PetscInt) * n)); 511f02d56fSJunchao Zhang PetscCall(PetscDeviceContextSynchronize(dctx)); // finish the async memcpy 521f02d56fSJunchao Zhang PetscCall(PetscIntView(n, yh, PETSC_VIEWER_STDOUT_WORLD)); 531f02d56fSJunchao Zhang 541f02d56fSJunchao Zhang PetscCall(PetscSFBcastWithMemTypeBegin(vscat, MPIU_2INT, PETSC_MEMTYPE_DEVICE, yd, PETSC_MEMTYPE_DEVICE, xd, MPI_MINLOC)); 551f02d56fSJunchao Zhang PetscCall(PetscSFBcastEnd(vscat, MPIU_2INT, yd, xd, MPI_MINLOC)); 561f02d56fSJunchao Zhang PetscCall(PetscDeviceMemcpy(dctx, xh, xd, sizeof(PetscInt) * n2)); 571f02d56fSJunchao Zhang PetscCall(PetscDeviceContextSynchronize(dctx)); // finish the async memcpy 581f02d56fSJunchao Zhang PetscCall(PetscIntView(n2, xh, PETSC_VIEWER_STDOUT_WORLD)); 591f02d56fSJunchao Zhang 601f02d56fSJunchao Zhang PetscCall(PetscDeviceFree(dctx, xh)); 611f02d56fSJunchao Zhang PetscCall(PetscDeviceFree(dctx, yh)); 621f02d56fSJunchao Zhang PetscCall(PetscDeviceFree(dctx, xd)); 631f02d56fSJunchao Zhang PetscCall(PetscDeviceFree(dctx, yd)); 641f02d56fSJunchao Zhang PetscCall(ISDestroy(&ix)); 651f02d56fSJunchao Zhang PetscCall(ISDestroy(&iy)); 661f02d56fSJunchao Zhang PetscCall(VecDestroy(&x)); 671f02d56fSJunchao Zhang PetscCall(VecDestroy(&y)); 681f02d56fSJunchao Zhang PetscCall(VecScatterDestroy(&vscat)); 691f02d56fSJunchao Zhang PetscCall(PetscFinalize()); 701f02d56fSJunchao Zhang } 711f02d56fSJunchao Zhang 721f02d56fSJunchao Zhang /*TEST 731f02d56fSJunchao Zhang testset: 741f02d56fSJunchao Zhang output_file: output/ex23.out 751f02d56fSJunchao Zhang nsize: 3 761f02d56fSJunchao Zhang 771f02d56fSJunchao Zhang test: 781f02d56fSJunchao Zhang suffix: 1 791f02d56fSJunchao Zhang requires: cuda 801f02d56fSJunchao Zhang 811f02d56fSJunchao Zhang test: 821f02d56fSJunchao Zhang suffix: 2 831f02d56fSJunchao Zhang requires: hip 841f02d56fSJunchao Zhang 851f02d56fSJunchao Zhang test: 861f02d56fSJunchao Zhang suffix: 3 871f02d56fSJunchao Zhang requires: sycl 881f02d56fSJunchao Zhang 891f02d56fSJunchao Zhang TEST*/ 90