46 …for (CeedInt elem = blockIdx.x * blockDim.z + threadIdx.z; elem < num_elem; elem += gridDim.x * bl…  in __launch_bounds__()  local
49 …ReadElementStrided1d<BASIS_NUM_COMP, BASIS_P_1D>(data, elem, 1, BASIS_P_1D * num_elem, BASIS_P_1D,…  in __launch_bounds__()
52 …ReadElementStrided2d<BASIS_NUM_COMP, BASIS_P_1D>(data, elem, 1, BASIS_P_1D * BASIS_P_1D * num_elem…  in __launch_bounds__()
55 …ReadElementStrided3d<BASIS_NUM_COMP, BASIS_P_1D>(data, elem, 1, BASIS_P_1D * BASIS_P_1D * BASIS_P_…  in __launch_bounds__()
66 …ReadPoint<BASIS_DIM, BASIS_NUM_PTS>(data, elem, p, BASIS_NUM_PTS, 1, num_elem * BASIS_NUM_PTS, BAS…  in __launch_bounds__()
74 …WritePoint<BASIS_NUM_COMP, BASIS_NUM_PTS>(data, elem, p, BASIS_NUM_PTS, 1, num_elem * BASIS_NUM_PT…  in __launch_bounds__()
102 …for (CeedInt elem = blockIdx.x * blockDim.z + threadIdx.z; elem < num_elem; elem += gridDim.x * bl…  in __launch_bounds__()  local
109 …WriteElementStrided1d<BASIS_NUM_COMP, BASIS_P_1D>(data, elem, 1, BASIS_P_1D * num_elem, BASIS_P_1D…  in __launch_bounds__()
111 …WriteElementStrided2d<BASIS_NUM_COMP, BASIS_P_1D>(data, elem, 1, BASIS_P_1D * BASIS_P_1D * num_ele…  in __launch_bounds__()
113 …WriteElementStrided3d<BASIS_NUM_COMP, BASIS_P_1D>(data, elem, 1, BASIS_P_1D * BASIS_P_1D * BASIS_P…  in __launch_bounds__()
123 …ReadPoint<BASIS_DIM, BASIS_NUM_PTS>(data, elem, p, BASIS_NUM_PTS, 1, num_elem * BASIS_NUM_PTS, BAS…  in __launch_bounds__()
124 …ReadPoint<BASIS_NUM_COMP, BASIS_NUM_PTS>(data, elem, i, points_per_elem[elem], 1, num_elem * BASIS…  in __launch_bounds__()
137 …SumElementStrided1d<BASIS_NUM_COMP, BASIS_P_1D>(data, elem, 1, BASIS_P_1D * num_elem, BASIS_P_1D, …  in __launch_bounds__()
140 …SumElementStrided2d<BASIS_NUM_COMP, BASIS_P_1D>(data, elem, 1, BASIS_P_1D * BASIS_P_1D * num_elem,…  in __launch_bounds__()
143 …SumElementStrided3d<BASIS_NUM_COMP, BASIS_P_1D>(data, elem, 1, BASIS_P_1D * BASIS_P_1D * BASIS_P_1…  in __launch_bounds__()
172 …for (CeedInt elem = blockIdx.x * blockDim.z + threadIdx.z; elem < num_elem; elem += gridDim.x * bl…  in __launch_bounds__()  local
182 …ReadPoint<BASIS_DIM, BASIS_NUM_PTS>(data, elem, p, BASIS_NUM_PTS, 1, num_elem * BASIS_NUM_PTS, BAS…  in __launch_bounds__()
183 …ReadPoint<BASIS_NUM_COMP, BASIS_NUM_PTS>(data, elem, i, points_per_elem[elem], 1, num_elem * BASIS…  in __launch_bounds__()
196 …SumElementStrided1d<BASIS_NUM_COMP, BASIS_P_1D>(data, elem, 1, BASIS_P_1D * num_elem, BASIS_P_1D, …  in __launch_bounds__()
199 …SumElementStrided2d<BASIS_NUM_COMP, BASIS_P_1D>(data, elem, 1, BASIS_P_1D * BASIS_P_1D * num_elem,…  in __launch_bounds__()
202 …SumElementStrided3d<BASIS_NUM_COMP, BASIS_P_1D>(data, elem, 1, BASIS_P_1D * BASIS_P_1D * BASIS_P_1…  in __launch_bounds__()
234 …for (CeedInt elem = blockIdx.x * blockDim.z + threadIdx.z; elem < num_elem; elem += gridDim.x * bl…  in __launch_bounds__()  local
237 …ReadElementStrided1d<BASIS_NUM_COMP, BASIS_P_1D>(data, elem, 1, BASIS_P_1D * num_elem, BASIS_P_1D,…  in __launch_bounds__()
240 …ReadElementStrided2d<BASIS_NUM_COMP, BASIS_P_1D>(data, elem, 1, BASIS_P_1D * BASIS_P_1D * num_elem…  in __launch_bounds__()
243 …ReadElementStrided3d<BASIS_NUM_COMP, BASIS_P_1D>(data, elem, 1, BASIS_P_1D * BASIS_P_1D * BASIS_P_…  in __launch_bounds__()
254 …ReadPoint<BASIS_DIM, BASIS_NUM_PTS>(data, elem, p, BASIS_NUM_PTS, 1, num_elem * BASIS_NUM_PTS, BAS…  in __launch_bounds__()
262 …WritePoint<BASIS_NUM_COMP * BASIS_DIM, BASIS_NUM_PTS>(data, elem, p, BASIS_NUM_PTS, 1, num_elem * …  in __launch_bounds__()
290 …for (CeedInt elem = blockIdx.x * blockDim.z + threadIdx.z; elem < num_elem; elem += gridDim.x * bl…  in __launch_bounds__()  local
297 …WriteElementStrided1d<BASIS_NUM_COMP, BASIS_P_1D>(data, elem, 1, BASIS_P_1D * num_elem, BASIS_P_1D…  in __launch_bounds__()
299 …WriteElementStrided2d<BASIS_NUM_COMP, BASIS_P_1D>(data, elem, 1, BASIS_P_1D * BASIS_P_1D * num_ele…  in __launch_bounds__()
301 …WriteElementStrided3d<BASIS_NUM_COMP, BASIS_P_1D>(data, elem, 1, BASIS_P_1D * BASIS_P_1D * BASIS_P…  in __launch_bounds__()
311 …ReadPoint<BASIS_DIM, BASIS_NUM_PTS>(data, elem, p, BASIS_NUM_PTS, 1, num_elem * BASIS_NUM_PTS, BAS…  in __launch_bounds__()
312 …ReadPoint<BASIS_NUM_COMP * BASIS_DIM, BASIS_NUM_PTS>(data, elem, i, points_per_elem[elem], 1, num_…  in __launch_bounds__()
326 …SumElementStrided1d<BASIS_NUM_COMP, BASIS_P_1D>(data, elem, 1, BASIS_P_1D * num_elem, BASIS_P_1D, …  in __launch_bounds__()
329 …SumElementStrided2d<BASIS_NUM_COMP, BASIS_P_1D>(data, elem, 1, BASIS_P_1D * BASIS_P_1D * num_elem,…  in __launch_bounds__()
332 …SumElementStrided3d<BASIS_NUM_COMP, BASIS_P_1D>(data, elem, 1, BASIS_P_1D * BASIS_P_1D * BASIS_P_1…  in __launch_bounds__()
361 …for (CeedInt elem = blockIdx.x * blockDim.z + threadIdx.z; elem < num_elem; elem += gridDim.x * bl…  in __launch_bounds__()  local
371 …ReadPoint<BASIS_DIM, BASIS_NUM_PTS>(data, elem, p, BASIS_NUM_PTS, 1, num_elem * BASIS_NUM_PTS, BAS…  in __launch_bounds__()
372 …ReadPoint<BASIS_NUM_COMP * BASIS_DIM, BASIS_NUM_PTS>(data, elem, i, points_per_elem[elem], 1, num_…  in __launch_bounds__()
386 …SumElementStrided1d<BASIS_NUM_COMP, BASIS_P_1D>(data, elem, 1, BASIS_P_1D * num_elem, BASIS_P_1D, …  in __launch_bounds__()
389 …SumElementStrided2d<BASIS_NUM_COMP, BASIS_P_1D>(data, elem, 1, BASIS_P_1D * BASIS_P_1D * num_elem,…  in __launch_bounds__()
392 …SumElementStrided3d<BASIS_NUM_COMP, BASIS_P_1D>(data, elem, 1, BASIS_P_1D * BASIS_P_1D * BASIS_P_1…  in __launch_bounds__()