1038a8942SZach Atkins /// @file
2038a8942SZach Atkins /// Test gradient transpose in multiple dimensions at arbitrary points
3038a8942SZach Atkins /// \test Test gradient transpose in multiple dimensions at arbitrary points
4038a8942SZach Atkins #include <ceed.h>
5038a8942SZach Atkins #include <math.h>
6038a8942SZach Atkins #include <stdio.h>
7038a8942SZach Atkins
Eval(CeedInt dim,const CeedScalar x[])8038a8942SZach Atkins static CeedScalar Eval(CeedInt dim, const CeedScalar x[]) {
9038a8942SZach Atkins CeedScalar result = tanh(x[0] + 0.1);
10038a8942SZach Atkins if (dim > 1) result += atan(x[1] + 0.2);
11038a8942SZach Atkins if (dim > 2) result += exp(-(x[2] + 0.3) * (x[2] + 0.3));
12038a8942SZach Atkins return result;
13038a8942SZach Atkins }
14038a8942SZach Atkins
GetTolerance(CeedScalarType scalar_type,int dim)15038a8942SZach Atkins static CeedScalar GetTolerance(CeedScalarType scalar_type, int dim) {
16038a8942SZach Atkins CeedScalar tol;
17038a8942SZach Atkins if (scalar_type == CEED_SCALAR_FP32) {
18038a8942SZach Atkins if (dim == 3) tol = 0.005;
19038a8942SZach Atkins else tol = 1.e-4;
20038a8942SZach Atkins } else {
21038a8942SZach Atkins tol = 1.e-11;
22038a8942SZach Atkins }
23038a8942SZach Atkins return tol;
24038a8942SZach Atkins }
25038a8942SZach Atkins
main(int argc,char ** argv)26038a8942SZach Atkins int main(int argc, char **argv) {
27038a8942SZach Atkins Ceed ceed;
28038a8942SZach Atkins
29038a8942SZach Atkins CeedInit(argv[1], &ceed);
30038a8942SZach Atkins
31038a8942SZach Atkins for (CeedInt dim = 1; dim <= 3; dim++) {
32038a8942SZach Atkins CeedVector x, x_nodes, x_points, u, u_points, v, ones;
33038a8942SZach Atkins CeedBasis basis_x, basis_u;
34038a8942SZach Atkins const CeedInt p = 9, q = 9, num_points = 4, x_dim = CeedIntPow(2, dim), p_dim = CeedIntPow(p, dim);
35038a8942SZach Atkins CeedScalar sum_1 = 0, sum_2 = 0;
36038a8942SZach Atkins
37038a8942SZach Atkins CeedVectorCreate(ceed, x_dim * dim, &x);
38038a8942SZach Atkins CeedVectorCreate(ceed, p_dim * dim, &x_nodes);
39038a8942SZach Atkins CeedVectorCreate(ceed, num_points * dim, &x_points);
40038a8942SZach Atkins CeedVectorCreate(ceed, p_dim, &u);
41038a8942SZach Atkins CeedVectorCreate(ceed, num_points * dim, &u_points);
42038a8942SZach Atkins CeedVectorCreate(ceed, p_dim, &v);
43038a8942SZach Atkins CeedVectorCreate(ceed, num_points * dim, &ones);
44038a8942SZach Atkins
45038a8942SZach Atkins CeedVectorSetValue(ones, 1);
46038a8942SZach Atkins CeedVectorSetValue(v, 0);
47038a8942SZach Atkins
48038a8942SZach Atkins // Get nodal coordinates
49038a8942SZach Atkins CeedBasisCreateTensorH1Lagrange(ceed, dim, dim, 2, p, CEED_GAUSS_LOBATTO, &basis_x);
50038a8942SZach Atkins {
51038a8942SZach Atkins CeedScalar x_array[x_dim * dim];
52038a8942SZach Atkins
53038a8942SZach Atkins for (CeedInt d = 0; d < dim; d++) {
54038a8942SZach Atkins for (CeedInt i = 0; i < x_dim; i++) x_array[d * x_dim + i] = (i % CeedIntPow(2, d + 1)) / CeedIntPow(2, d) ? 1 : -1;
55038a8942SZach Atkins }
56038a8942SZach Atkins CeedVectorSetArray(x, CEED_MEM_HOST, CEED_COPY_VALUES, x_array);
57038a8942SZach Atkins }
58038a8942SZach Atkins CeedBasisApply(basis_x, 1, CEED_NOTRANSPOSE, CEED_EVAL_INTERP, x, x_nodes);
59038a8942SZach Atkins
60038a8942SZach Atkins // Set values of u at nodes
61038a8942SZach Atkins {
62038a8942SZach Atkins const CeedScalar *x_array;
63038a8942SZach Atkins CeedScalar u_array[p_dim];
64038a8942SZach Atkins
65038a8942SZach Atkins CeedVectorGetArrayRead(x_nodes, CEED_MEM_HOST, &x_array);
66038a8942SZach Atkins for (CeedInt i = 0; i < p_dim; i++) {
67038a8942SZach Atkins CeedScalar coord[dim];
68038a8942SZach Atkins
69038a8942SZach Atkins for (CeedInt d = 0; d < dim; d++) coord[d] = x_array[d * p_dim + i];
70038a8942SZach Atkins u_array[i] = Eval(dim, coord);
71038a8942SZach Atkins }
72038a8942SZach Atkins CeedVectorRestoreArrayRead(x_nodes, &x_array);
73038a8942SZach Atkins CeedVectorSetArray(u, CEED_MEM_HOST, CEED_COPY_VALUES, (CeedScalar *)&u_array);
74038a8942SZach Atkins }
75038a8942SZach Atkins
76038a8942SZach Atkins // Interpolate to arbitrary points
77038a8942SZach Atkins CeedBasisCreateTensorH1Lagrange(ceed, dim, 1, p, q, CEED_GAUSS, &basis_u);
78038a8942SZach Atkins {
79038a8942SZach Atkins CeedScalar x_array[12] = {-0.33, -0.65, 0.16, 0.99, -0.65, 0.16, 0.99, -0.33, 0.16, 0.99, -0.33, -0.65};
80038a8942SZach Atkins
81038a8942SZach Atkins CeedVectorSetArray(x_points, CEED_MEM_HOST, CEED_COPY_VALUES, x_array);
82038a8942SZach Atkins }
83038a8942SZach Atkins
84038a8942SZach Atkins // Calculate G u at arbitrary points, G' * 1 at dofs
85*fc0f7cc6SJeremy L Thompson CeedBasisApplyAtPoints(basis_u, 1, &num_points, CEED_NOTRANSPOSE, CEED_EVAL_GRAD, x_points, u, u_points);
86*fc0f7cc6SJeremy L Thompson CeedBasisApplyAtPoints(basis_u, 1, &num_points, CEED_TRANSPOSE, CEED_EVAL_GRAD, x_points, ones, v);
87038a8942SZach Atkins {
88038a8942SZach Atkins const CeedScalar *u_array, *v_array, *u_points_array;
89038a8942SZach Atkins
90038a8942SZach Atkins CeedVectorGetArrayRead(u, CEED_MEM_HOST, &u_array);
91038a8942SZach Atkins CeedVectorGetArrayRead(v, CEED_MEM_HOST, &v_array);
92038a8942SZach Atkins CeedVectorGetArrayRead(u_points, CEED_MEM_HOST, &u_points_array);
93038a8942SZach Atkins for (CeedInt i = 0; i < p_dim; i++) sum_1 += v_array[i] * u_array[i];
94038a8942SZach Atkins for (CeedInt i = 0; i < num_points * dim; i++) sum_2 += u_points_array[i];
95038a8942SZach Atkins CeedVectorRestoreArrayRead(u, &u_array);
96038a8942SZach Atkins CeedVectorRestoreArrayRead(v, &v_array);
97038a8942SZach Atkins CeedVectorRestoreArrayRead(u_points, &u_points_array);
98038a8942SZach Atkins }
9999e754f0SJeremy L Thompson {
10099e754f0SJeremy L Thompson CeedScalarType scalar_type;
10199e754f0SJeremy L Thompson
10299e754f0SJeremy L Thompson CeedGetScalarType(&scalar_type);
10399e754f0SJeremy L Thompson
10499e754f0SJeremy L Thompson CeedScalar tol = GetTolerance(scalar_type, dim);
10599e754f0SJeremy L Thompson
106038a8942SZach Atkins if (fabs(sum_1 - sum_2) > tol) printf("[%" CeedInt_FMT "] %f != %f\n", dim, sum_1, sum_2);
10799e754f0SJeremy L Thompson }
108038a8942SZach Atkins
109038a8942SZach Atkins CeedVectorDestroy(&x);
110038a8942SZach Atkins CeedVectorDestroy(&x_nodes);
111038a8942SZach Atkins CeedVectorDestroy(&x_points);
112038a8942SZach Atkins CeedVectorDestroy(&u);
113038a8942SZach Atkins CeedVectorDestroy(&u_points);
114038a8942SZach Atkins CeedVectorDestroy(&ones);
115038a8942SZach Atkins CeedVectorDestroy(&v);
116038a8942SZach Atkins CeedBasisDestroy(&basis_x);
117038a8942SZach Atkins CeedBasisDestroy(&basis_u);
118038a8942SZach Atkins }
119038a8942SZach Atkins CeedDestroy(&ceed);
120038a8942SZach Atkins return 0;
121038a8942SZach Atkins }
122