1 // Copyright (c) 2017-2026, Lawrence Livermore National Security, LLC and other CEED contributors. 2 // All Rights Reserved. See the top-level LICENSE and NOTICE files for details. 3 // 4 // SPDX-License-Identifier: BSD-2-Clause 5 // 6 // This file is part of CEED: http://github.com/ceed 7 8 /// @file 9 /// Internal header for CUDA operator diagonal assembly 10 #include <ceed/types.h> 11 12 #if USE_CEEDSIZE 13 typedef CeedSize IndexType; 14 #else 15 typedef CeedInt IndexType; 16 #endif 17 18 //------------------------------------------------------------------------------ 19 // Get basis pointer 20 //------------------------------------------------------------------------------ 21 static __device__ __inline__ void GetBasisPointer(const CeedScalar **basis_ptr, CeedEvalMode eval_modes, const CeedScalar *identity, 22 const CeedScalar *interp, const CeedScalar *grad, const CeedScalar *div, const CeedScalar *curl) { 23 switch (eval_modes) { 24 case CEED_EVAL_NONE: 25 *basis_ptr = identity; 26 break; 27 case CEED_EVAL_INTERP: 28 *basis_ptr = interp; 29 break; 30 case CEED_EVAL_GRAD: 31 *basis_ptr = grad; 32 break; 33 case CEED_EVAL_DIV: 34 *basis_ptr = div; 35 break; 36 case CEED_EVAL_CURL: 37 *basis_ptr = curl; 38 break; 39 case CEED_EVAL_WEIGHT: 40 break; // Caught by QF assembly 41 } 42 } 43 44 //------------------------------------------------------------------------------ 45 // Core code for diagonal assembly 46 //------------------------------------------------------------------------------ 47 extern "C" __launch_bounds__(BLOCK_SIZE) __global__ 48 void LinearDiagonal(const CeedInt num_elem, const CeedScalar *identity, const CeedScalar *interp_in, const CeedScalar *grad_in, 49 const CeedScalar *div_in, const CeedScalar *curl_in, const CeedScalar *interp_out, const CeedScalar *grad_out, 50 const CeedScalar *div_out, const CeedScalar *curl_out, const CeedEvalMode *eval_modes_in, const CeedEvalMode *eval_modes_out, 51 const CeedScalar *__restrict__ assembled_qf_array, CeedScalar *__restrict__ elem_diag_array) { 52 const int tid = threadIdx.x; // Running with P threads 53 54 if (tid >= NUM_NODES) return; 55 56 // Compute the diagonal of B^T D B 57 // Each element 58 for (IndexType e = blockIdx.x * blockDim.z + threadIdx.z; e < num_elem; e += gridDim.x * blockDim.z) { 59 // Each basis eval mode pair 60 IndexType d_out = 0; 61 CeedEvalMode eval_modes_out_prev = CEED_EVAL_NONE; 62 63 for (IndexType e_out = 0; e_out < NUM_EVAL_MODES_OUT; e_out++) { 64 IndexType d_in = 0; 65 CeedEvalMode eval_modes_in_prev = CEED_EVAL_NONE; 66 const CeedScalar *b_t = NULL; 67 68 GetBasisPointer(&b_t, eval_modes_out[e_out], identity, interp_out, grad_out, div_out, curl_out); 69 if (e_out == 0 || eval_modes_out[e_out] != eval_modes_out_prev) d_out = 0; 70 else b_t = &b_t[(++d_out) * NUM_QPTS * NUM_NODES]; 71 eval_modes_out_prev = eval_modes_out[e_out]; 72 73 for (IndexType e_in = 0; e_in < NUM_EVAL_MODES_IN; e_in++) { 74 const CeedScalar *b = NULL; 75 76 GetBasisPointer(&b, eval_modes_in[e_in], identity, interp_in, grad_in, div_in, curl_in); 77 if (e_in == 0 || eval_modes_in[e_in] != eval_modes_in_prev) d_in = 0; 78 else b = &b[(++d_in) * NUM_QPTS * NUM_NODES]; 79 eval_modes_in_prev = eval_modes_in[e_in]; 80 81 // Each component 82 for (IndexType comp_out = 0; comp_out < NUM_COMP; comp_out++) { 83 #if USE_POINT_BLOCK 84 // Point block diagonal 85 for (IndexType comp_in = 0; comp_in < NUM_COMP; comp_in++) { 86 CeedScalar e_value = 0.; 87 88 // Each qpoint/node pair 89 for (IndexType q = 0; q < NUM_QPTS; q++) { 90 const CeedScalar qf_value = 91 assembled_qf_array[((((e_in * NUM_COMP + comp_in) * NUM_EVAL_MODES_OUT + e_out) * NUM_COMP + comp_out) * num_elem + e) * NUM_QPTS + 92 q]; 93 94 e_value += b_t[q * NUM_NODES + tid] * qf_value * b[q * NUM_NODES + tid]; 95 } 96 elem_diag_array[((comp_out * NUM_COMP + comp_in) * num_elem + e) * NUM_NODES + tid] += e_value; 97 } 98 #else 99 // Diagonal only 100 CeedScalar e_value = 0.; 101 102 // Each qpoint/node pair 103 for (IndexType q = 0; q < NUM_QPTS; q++) { 104 const CeedScalar qf_value = 105 assembled_qf_array[((((e_in * NUM_COMP + comp_out) * NUM_EVAL_MODES_OUT + e_out) * NUM_COMP + comp_out) * num_elem + e) * NUM_QPTS + 106 q]; 107 108 e_value += b_t[q * NUM_NODES + tid] * qf_value * b[q * NUM_NODES + tid]; 109 } 110 elem_diag_array[(comp_out * num_elem + e) * NUM_NODES + tid] += e_value; 111 #endif 112 } 113 } 114 } 115 } 116 } 117