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 1D 10 #include "magma-common-tensor.h" 11 12 // macros to abstract access of shared memory and reg. file 13 #define sT(i, j) sT[(j) * P + (i)] 14 #define sTmp(i, j, ldw) sTmp[(j) * (ldw) + (i)] 15 16 //////////////////////////////////////////////////////////////////////////////// 17 // interp basis action (2D) 18 template <typename T, int DIM_U, int DIM_V, int NUM_COMP, int P, int Q, int rU_SIZE, int rV_SIZE> 19 static __device__ __inline__ void magma_interp_2d_device(const T *sT, T rU[DIM_U][NUM_COMP][rU_SIZE], T rV[DIM_V][NUM_COMP][rV_SIZE], const int tx, 20 T rTmp, T *swork) { 21 // Assumptions 22 // 1. 1D threads of size max(P,Q) 23 // 2. input: rU[DIM_U x NUM_COMP x rU_SIZE] in registers (per thread) 24 // 3. output: rV[DIM_V x NUM_COMP x rV_SIZE] in registers (per thread) 25 // 4. Two products per component 26 // 4.1 Batch P of (1xP) matrices times (PxQ) matrix => Batch P of (1xQ) matrices 27 // 4.2 Batch 1 of (QxP) matrix times (PxQ) matrix => (QxQ) matrix 28 // 5. Each thread computes one row of the output of each product 29 // 6. Sync is recommended before and after the call 30 31 for (int comp = 0; comp < NUM_COMP; comp++) { 32 // 1st product -- Batch P of (1xP) matrices [reg] x (PxQ) [shmem] => Batch P of (1xQ) matrices 33 // the batch output P x (1xQ) is written on the fly to shmem 34 if (tx < P) { 35 const int batchid = tx; 36 const int sld = 1; 37 T *sTmp = swork + batchid * (1 * Q); 38 for (int j = 0; j < Q; j++) { 39 rTmp = 0.0; 40 for (int i = 0; i < P; i++) { 41 rTmp += rU[0][comp][i] * sT(i, j); 42 } 43 sTmp(0, j, sld) = rTmp; 44 } 45 } // end of: if (tx < P) 46 __syncthreads(); 47 48 // 2nd product -- Batch 1 of a (QxP) matrix [shmem] x (PxQ) [shmem] => (QxQ) matrix [reg] 49 if (tx < Q) { 50 const int batchid = 0; 51 const int sld = Q; 52 T *sTmp = swork + batchid * (Q * P); 53 for (int j = 0; j < Q; j++) { 54 rTmp = 0.0; 55 for (int i = 0; i < P; i++) { 56 rTmp += sTmp(tx, i, sld) * sT(i, j); 57 } 58 rV[0][comp][j] += rTmp; 59 } 60 } 61 __syncthreads(); 62 } 63 } 64 65 //////////////////////////////////////////////////////////////////////////////// 66 extern "C" __launch_bounds__(MAGMA_BASIS_BOUNDS(BASIS_MAX_P_Q, MAGMA_MAXTHREADS_2D)) __global__ 67 void magma_interpn_2d_kernel(const CeedScalar *dT, const CeedScalar *dU, const int estrdU, const int cstrdU, CeedScalar *dV, const int estrdV, 68 const int cstrdV, const int nelem) { 69 MAGMA_DEVICE_SHARED(CeedScalar, shared_data) 70 71 const int tx = threadIdx.x; 72 const int ty = threadIdx.y; 73 const int elem_id = (blockIdx.x * blockDim.y) + ty; 74 75 if (elem_id >= nelem) return; 76 77 CeedScalar rU[1][BASIS_NUM_COMP][BASIS_P] = {0.0}; // for a non-fused operator BASIS_DIM is always 1 78 CeedScalar rV[1][BASIS_NUM_COMP][BASIS_Q] = {0.0}; // for a non-fused operator BASIS_DIM is always 1 79 CeedScalar rTmp = 0.0; 80 81 // shift global memory pointers by elem stride 82 dU += elem_id * estrdU; 83 dV += elem_id * estrdV; 84 85 // assign shared memory pointers 86 CeedScalar *sT = (CeedScalar *)shared_data; 87 CeedScalar *sTmp = sT + BASIS_P * BASIS_Q; 88 sTmp += ty * (BASIS_P * BASIS_MAX_P_Q); 89 90 // read T 91 if (ty == 0) { 92 read_T_notrans_gm2sm<BASIS_P, BASIS_Q>(tx, dT, sT); 93 } 94 95 // read U -- there is a sync at the end of this function 96 read_U_2d<CeedScalar, BASIS_P, 1, BASIS_NUM_COMP, BASIS_P, 0>(dU, cstrdU, rU, sTmp, tx); 97 98 // no sync needed here -- read_U_2d already syncs at the end 99 magma_interp_2d_device<CeedScalar, 1, 1, BASIS_NUM_COMP, BASIS_P, BASIS_Q, BASIS_P, BASIS_Q>(sT, rU, rV, tx, rTmp, sTmp); 100 __syncthreads(); 101 102 // write V 103 write_V_2d<CeedScalar, BASIS_Q, 1, BASIS_NUM_COMP, BASIS_Q, 0>(dV, cstrdV, rV, tx); 104 } 105 106 //////////////////////////////////////////////////////////////////////////////// 107 extern "C" __launch_bounds__(MAGMA_BASIS_BOUNDS(BASIS_MAX_P_Q, MAGMA_MAXTHREADS_2D)) __global__ 108 void magma_interpt_2d_kernel(const CeedScalar *dT, const CeedScalar *dU, const int estrdU, const int cstrdU, CeedScalar *dV, const int estrdV, 109 const int cstrdV, const int nelem) { 110 MAGMA_DEVICE_SHARED(CeedScalar, shared_data) 111 112 const int tx = threadIdx.x; 113 const int ty = threadIdx.y; 114 const int elem_id = (blockIdx.x * blockDim.y) + ty; 115 116 if (elem_id >= nelem) return; 117 118 CeedScalar rU[1][BASIS_NUM_COMP][BASIS_Q] = {0.0}; // for a non-fused operator BASIS_DIM is always 1 119 CeedScalar rV[1][BASIS_NUM_COMP][BASIS_P] = {0.0}; // for a non-fused operator BASIS_DIM is always 1 120 CeedScalar rTmp = 0.0; 121 122 // shift global memory pointers by elem stride 123 dU += elem_id * estrdU; 124 dV += elem_id * estrdV; 125 126 // assign shared memory pointers 127 CeedScalar *sT = (CeedScalar *)shared_data; 128 CeedScalar *sTmp = sT + BASIS_Q * BASIS_P; 129 sTmp += ty * (BASIS_Q * BASIS_MAX_P_Q); 130 131 // read T 132 if (ty == 0) { 133 read_T_trans_gm2sm<BASIS_Q, BASIS_P>(tx, dT, sT); 134 } 135 136 // read U -- there is a sync at the end of this function 137 read_U_2d<CeedScalar, BASIS_Q, 1, BASIS_NUM_COMP, BASIS_Q, 0>(dU, cstrdU, rU, sTmp, tx); 138 139 // no sync needed here -- read_U_2d already syncs at the end 140 magma_interp_2d_device<CeedScalar, 1, 1, BASIS_NUM_COMP, BASIS_Q, BASIS_P, BASIS_Q, BASIS_P>(sT, rU, rV, tx, rTmp, sTmp); 141 __syncthreads(); 142 143 // write V 144 write_V_2d<CeedScalar, BASIS_P, 1, BASIS_NUM_COMP, BASIS_P, 0>(dV, cstrdV, rV, tx); 145 } 146 147 //////////////////////////////////////////////////////////////////////////////// 148 extern "C" __launch_bounds__(MAGMA_BASIS_BOUNDS(BASIS_MAX_P_Q, MAGMA_MAXTHREADS_2D)) __global__ 149 void magma_interpta_2d_kernel(const CeedScalar *dT, const CeedScalar *dU, const int estrdU, const int cstrdU, CeedScalar *dV, const int estrdV, 150 const int cstrdV, const int nelem) { 151 MAGMA_DEVICE_SHARED(CeedScalar, shared_data) 152 153 const int tx = threadIdx.x; 154 const int ty = threadIdx.y; 155 const int elem_id = (blockIdx.x * blockDim.y) + ty; 156 157 if (elem_id >= nelem) return; 158 159 CeedScalar rU[1][BASIS_NUM_COMP][BASIS_Q] = {0.0}; // for a non-fused operator BASIS_DIM is always 1 160 CeedScalar rV[1][BASIS_NUM_COMP][BASIS_P] = {0.0}; // for a non-fused operator BASIS_DIM is always 1 161 CeedScalar rTmp = 0.0; 162 163 // shift global memory pointers by elem stride 164 dU += elem_id * estrdU; 165 dV += elem_id * estrdV; 166 167 // assign shared memory pointers 168 CeedScalar *sT = (CeedScalar *)shared_data; 169 CeedScalar *sTmp = sT + BASIS_Q * BASIS_P; 170 sTmp += ty * (BASIS_Q * BASIS_MAX_P_Q); 171 172 // read T 173 if (ty == 0) { 174 read_T_trans_gm2sm<BASIS_Q, BASIS_P>(tx, dT, sT); 175 } 176 177 // read U -- there is a sync at the end of this function 178 read_U_2d<CeedScalar, BASIS_Q, 1, BASIS_NUM_COMP, BASIS_Q, 0>(dU, cstrdU, rU, sTmp, tx); 179 180 // no sync needed here -- read_U_2d already syncs at the end 181 magma_interp_2d_device<CeedScalar, 1, 1, BASIS_NUM_COMP, BASIS_Q, BASIS_P, BASIS_Q, BASIS_P>(sT, rU, rV, tx, rTmp, sTmp); 182 __syncthreads(); 183 184 // sum into V 185 sum_V_2d<CeedScalar, BASIS_P, 1, BASIS_NUM_COMP, BASIS_P, 0>(dV, cstrdV, rV, tx); 186 } 187