1 // Copyright (c) 2017-2024, 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 MAGMA tensor basis interpolation in 3D 10 11 #include "magma-common-tensor.h" 12 13 // macros to abstract access of shared memory and reg. file 14 #define sT(i, j) sT[(j) * P + (i)] 15 #define sTmp(i, j, ldw) sTmp[(j) * (ldw) + (i)] 16 17 //////////////////////////////////////////////////////////////////////////////// 18 // interp basis action (3D) 19 template <typename T, int DIM_U, int DIM_V, int NUM_COMP, int P, int Q, int rU_SIZE, int rV_SIZE> 20 static __device__ __inline__ void magma_interp_3d_device(const T *sT, T rU[DIM_U][NUM_COMP][rU_SIZE], T rV[DIM_V][NUM_COMP][rV_SIZE], const int tx, 21 T rTmp[Q], T *swork) { 22 // Assumptions 23 // 1. 1D threads of size max(P,Q)^2 24 // 2. input: rU[DIM_U x NUM_COMP x rU_SIZE] in registers (per thread) 25 // 3. output: rV[DIM_V x NUM_COMP x rV_SIZE] in registers (per thread) 26 // 4. Three products per component 27 // 4.1 Batch P^2 of (1xP) matrices times (PxQ) matrix => Batch P^2 of (1xQ) matrices 28 // 4.2 Batch P of (QxP) matrices times (PxQ) matrix => Batch P of (QxQ) matrices 29 // 4.3 Batch 1 of (Q^2xP_) matrix times (PxQ) matrix => (Q^2xQ_) matrix 30 // 5. Each thread computes one row of the output of each product 31 // 6. Sync is recommended before and after the call 32 33 for (int comp = 0; comp < NUM_COMP; comp++) { 34 // Batch P^2 of (1xP) matrices [reg] times (PxQ) matrix [shmem] => Batch P^2 of (1xQ) matrices [shmem] 35 if (tx < (P * P)) { 36 const int batchid = tx; 37 const int sld = 1; 38 T *sTmp = swork + batchid * (1 * Q); 39 for (int j = 0; j < Q; j++) { 40 rTmp[0] = 0.0; 41 for (int i = 0; i < P; i++) { 42 rTmp[0] += rU[0][comp][i] * sT(i, j); 43 } 44 sTmp(0, j, sld) = rTmp[0]; 45 } 46 } // end of: if (tx < P*P) 47 __syncthreads(); 48 49 // Batch P of (QxP) matrices [shmem] times (PxQ) matrix [shmem] => Batch P of (QxQ) matrices [reg] 50 if (tx < (P * Q)) { 51 const int batchid = tx / Q; 52 const int tx_ = tx % Q; 53 const int sld = Q; 54 T *sTmp = swork + batchid * (Q * P); // sTmp is input 55 for (int j = 0; j < Q; j++) { 56 rTmp[j] = 0.0; 57 for (int i = 0; i < P; i++) { 58 rTmp[j] += sTmp(tx_, i, sld) * sT(i, j); 59 } 60 } 61 } 62 __syncthreads(); 63 64 // write rTmp[] into shmem as batch P of QxQ matrices 65 if (tx < (P * Q)) { 66 const int batchid = tx / Q; 67 const int tx_ = tx % Q; 68 const int sld = Q; 69 T *sTmp = swork + batchid * (Q * Q); 70 for (int j = 0; j < Q; j++) { 71 sTmp(tx_, j, sld) = rTmp[j]; 72 } 73 } 74 __syncthreads(); 75 76 // Batch 1 of (Q^2xP_) matrices [shmem] times (PxQ) matrix [shmem] => Batch 1 of (Q^2xQ_) matrices [reg] 77 if (tx < (Q * Q)) { 78 // No need to declare batchid = (tx / Q^2) = always zero 79 // No need to declare tx_ = (tx_ % Q^2) = always tx 80 const int sld = Q * Q; 81 T *sTmp = swork; 82 for (int j = 0; j < Q; j++) { 83 rTmp[0] = 0.0; 84 for (int i = 0; i < P; i++) { 85 rTmp[0] += sTmp(tx, i, sld) * sT(i, j); 86 } 87 rV[0][comp][j] += rTmp[0]; 88 } 89 } 90 __syncthreads(); 91 } 92 } 93 94 //////////////////////////////////////////////////////////////////////////////// 95 extern "C" __launch_bounds__(MAGMA_BASIS_BOUNDS(BASIS_MAX_P_Q *BASIS_MAX_P_Q, MAGMA_MAXTHREADS_3D)) __global__ 96 void magma_interpn_3d_kernel(const CeedScalar *dT, const CeedScalar *dU, const int estrdU, const int cstrdU, CeedScalar *dV, const int estrdV, 97 const int cstrdV, const int nelem) { 98 MAGMA_DEVICE_SHARED(CeedScalar, shared_data) 99 100 const int tx = threadIdx.x; 101 const int ty = threadIdx.y; 102 const int elem_id = (blockIdx.x * blockDim.y) + ty; 103 104 if (elem_id >= nelem) return; 105 106 CeedScalar rU[1][BASIS_NUM_COMP][BASIS_P] = {0.0}; // for a non-fused operator BASIS_DIM is always 1 107 CeedScalar rV[1][BASIS_NUM_COMP][BASIS_Q] = {0.0}; // for a non-fused operator BASIS_DIM is always 1 108 CeedScalar rTmp[BASIS_Q] = {0.0}; 109 110 // shift global memory pointers by elem stride 111 dU += elem_id * estrdU; 112 dV += elem_id * estrdV; 113 114 // assign shared memory pointers 115 CeedScalar *sT = (CeedScalar *)shared_data; 116 CeedScalar *sTmp = sT + BASIS_P * BASIS_Q; 117 sTmp += ty * (max(BASIS_P * BASIS_P * BASIS_MAX_P_Q, BASIS_P * BASIS_Q * BASIS_Q)); 118 119 // read T 120 if (ty == 0) { 121 read_T_notrans_gm2sm<BASIS_P, BASIS_Q>(tx, dT, sT); 122 } 123 124 // read U (idim = 0 for dU, i_DIM = 0 for rU, u_dimstride is always 0) 125 read_U_3d<CeedScalar, BASIS_P, 1, BASIS_NUM_COMP, BASIS_P, 0>(dU, cstrdU, rU, sTmp, tx); 126 // there is a sync at the end of this function 127 128 magma_interp_3d_device<CeedScalar, 1, 1, BASIS_NUM_COMP, BASIS_P, BASIS_Q, BASIS_P, BASIS_Q>(sT, rU, rV, tx, rTmp, sTmp); 129 __syncthreads(); 130 131 // write V 132 write_V_3d<CeedScalar, BASIS_Q, 1, BASIS_NUM_COMP, BASIS_Q, 0>(dV, cstrdV, rV, tx); 133 } 134 135 //////////////////////////////////////////////////////////////////////////////// 136 extern "C" __launch_bounds__(MAGMA_BASIS_BOUNDS(BASIS_MAX_P_Q *BASIS_MAX_P_Q, MAGMA_MAXTHREADS_3D)) __global__ 137 void magma_interpt_3d_kernel(const CeedScalar *dT, const CeedScalar *dU, const int estrdU, const int cstrdU, CeedScalar *dV, const int estrdV, 138 const int cstrdV, const int nelem) { 139 MAGMA_DEVICE_SHARED(CeedScalar, shared_data) 140 141 const int tx = threadIdx.x; 142 const int ty = threadIdx.y; 143 const int elem_id = (blockIdx.x * blockDim.y) + ty; 144 145 if (elem_id >= nelem) return; 146 147 CeedScalar rU[1][BASIS_NUM_COMP][BASIS_Q] = {0.0}; // for a non-fused operator BASIS_DIM is always 1 148 CeedScalar rV[1][BASIS_NUM_COMP][BASIS_P] = {0.0}; // for a non-fused operator BASIS_DIM is always 1 149 CeedScalar rTmp[BASIS_P] = {0.0}; 150 151 // shift global memory pointers by elem stride 152 dU += elem_id * estrdU; 153 dV += elem_id * estrdV; 154 155 // assign shared memory pointers 156 CeedScalar *sT = (CeedScalar *)shared_data; 157 CeedScalar *sTmp = sT + BASIS_Q * BASIS_P; 158 sTmp += ty * (max(BASIS_Q * BASIS_Q * BASIS_MAX_P_Q, BASIS_Q * BASIS_P * BASIS_P)); 159 160 // read T 161 if (ty == 0) { 162 read_T_trans_gm2sm<BASIS_Q, BASIS_P>(tx, dT, sT); 163 } 164 165 // read U (idim = 0 for dU, i_DIM = 0 for rU, u_dimstride is always 0) 166 read_U_3d<CeedScalar, BASIS_Q, 1, BASIS_NUM_COMP, BASIS_Q, 0>(dU, cstrdU, rU, sTmp, tx); 167 // there is a sync at the end of this function 168 169 magma_interp_3d_device<CeedScalar, 1, 1, BASIS_NUM_COMP, BASIS_Q, BASIS_P, BASIS_Q, BASIS_P>(sT, rU, rV, tx, rTmp, sTmp); 170 __syncthreads(); 171 172 // write V 173 write_V_3d<CeedScalar, BASIS_P, 1, BASIS_NUM_COMP, BASIS_P, 0>(dV, cstrdV, rV, tx); 174 } 175