28   data.t_id_x = threadIdx.x;  in InterpAtPoints()
31   data.t_id   = threadIdx.x + threadIdx.y * blockDim.x + threadIdx.z * blockDim.y * blockDim.x;  in InterpAtPoints()
45 …for (CeedInt elem = blockIdx.x * blockDim.z + threadIdx.z; elem < num_elem; elem += gridDim.x * bl…  in InterpAtPoints()
60 …const CeedInt point_loop_bound = (blockDim.x * blockDim.y) * ceil(1.0 * BASIS_NUM_PTS / (blockDim.…  in InterpAtPoints()
62 …for (CeedInt i = threadIdx.x + threadIdx.y * blockDim.x; i < point_loop_bound; i += blockDim.x * b…  in InterpAtPoints()
84   data.t_id_x = threadIdx.x;  in InterpTransposeAtPoints()
87   data.t_id   = threadIdx.x + threadIdx.y * blockDim.x + threadIdx.z * blockDim.y * blockDim.x;  in InterpTransposeAtPoints()
101 …for (CeedInt elem = blockIdx.x * blockDim.z + threadIdx.z; elem < num_elem; elem += gridDim.x * bl…  in InterpTransposeAtPoints()
117 …const CeedInt point_loop_bound = (blockDim.x * blockDim.y) * ceil(1.0 * BASIS_NUM_PTS / (blockDim.…  in InterpTransposeAtPoints()
119 …for (CeedInt i = threadIdx.x + threadIdx.y * blockDim.x; i < point_loop_bound; i += blockDim.x * b…  in InterpTransposeAtPoints()
154   data.t_id_x = threadIdx.x;  in InterpTransposeAddAtPoints()
157   data.t_id   = threadIdx.x + threadIdx.y * blockDim.x + threadIdx.z * blockDim.y * blockDim.x;  in InterpTransposeAddAtPoints()
171 …for (CeedInt elem = blockIdx.x * blockDim.z + threadIdx.z; elem < num_elem; elem += gridDim.x * bl…  in InterpTransposeAddAtPoints()
176 …const CeedInt point_loop_bound = (blockDim.x * blockDim.y) * ceil(1.0 * BASIS_NUM_PTS / (blockDim.…  in InterpTransposeAddAtPoints()
178 …for (CeedInt i = threadIdx.x + threadIdx.y * blockDim.x; i < point_loop_bound; i += blockDim.x * b…  in InterpTransposeAddAtPoints()
215   data.t_id_x = threadIdx.x;  in GradAtPoints()
218   data.t_id   = threadIdx.x + threadIdx.y * blockDim.x + threadIdx.z * blockDim.y * blockDim.x;  in GradAtPoints()
232 …for (CeedInt elem = blockIdx.x * blockDim.z + threadIdx.z; elem < num_elem; elem += gridDim.x * bl…  in GradAtPoints()
247 …const CeedInt point_loop_bound = (blockDim.x * blockDim.y) * ceil(1.0 * BASIS_NUM_PTS / (blockDim.…  in GradAtPoints()
249 …for (CeedInt i = threadIdx.x + threadIdx.y * blockDim.x; i < point_loop_bound; i += blockDim.x * b…  in GradAtPoints()
271   data.t_id_x = threadIdx.x;  in GradTransposeAtPoints()
274   data.t_id   = threadIdx.x + threadIdx.y * blockDim.x + threadIdx.z * blockDim.y * blockDim.x;  in GradTransposeAtPoints()
288 …for (CeedInt elem = blockIdx.x * blockDim.z + threadIdx.z; elem < num_elem; elem += gridDim.x * bl…  in GradTransposeAtPoints()
304 …const CeedInt point_loop_bound = (blockDim.x * blockDim.y) * ceil(1.0 * BASIS_NUM_PTS / (blockDim.…  in GradTransposeAtPoints()
306 …for (CeedInt i = threadIdx.x + threadIdx.y * blockDim.x; i < point_loop_bound; i += blockDim.x * b…  in GradTransposeAtPoints()
342   data.t_id_x = threadIdx.x;  in GradTransposeAddAtPoints()
345   data.t_id   = threadIdx.x + threadIdx.y * blockDim.x + threadIdx.z * blockDim.y * blockDim.x;  in GradTransposeAddAtPoints()
359 …for (CeedInt elem = blockIdx.x * blockDim.z + threadIdx.z; elem < num_elem; elem += gridDim.x * bl…  in GradTransposeAddAtPoints()
364 …const CeedInt point_loop_bound = (blockDim.x * blockDim.y) * ceil(1.0 * BASIS_NUM_PTS / (blockDim.…  in GradTransposeAddAtPoints()
366 …for (CeedInt i = threadIdx.x + threadIdx.y * blockDim.x; i < point_loop_bound; i += blockDim.x * b…  in GradTransposeAddAtPoints()