1 // Copyright (c) 2017-2022, 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 ////////////////////////////////////////////////////////////////////////////////////////// 9 // weight basis action -- 2D 10 template <typename T, int DIM_, int NCOMP_, int Q_, int iDIM, int iCOMP> 11 __device__ __inline__ void magma_weight_2d_device(const T *sTweight, T rV[DIM_][NCOMP_][Q_], const int tx) { 12 // Assumptions 13 // 1. 1D thread configuration of size Q_ 14 // 2. rV[][][] matches the storage used in other actions (interp, grad, ... etc) 15 // 3. iDIM and iCOMP specify which indexes to use in rV, 16 // since the output per thread is a register array of size Q_ 17 // 4. Sync is recommended after the call (to make sure sTweight can be overwritten) 18 19 if (tx < Q_) { 20 // x sTweight[j] for first update 21 // x sTweight[tx] for second update 22 for (int j = 0; j < Q_; j++) { 23 rV[iDIM][iCOMP][j] = sTweight[j] * sTweight[tx]; 24 } 25 } 26 } 27 28 ////////////////////////////////////////////////////////////////////////////////////////// 29 extern "C" __launch_bounds__(MAGMA_BASIS_BOUNDS(Q, MAGMA_MAXTHREADS_2D)) __global__ 30 void magma_weight_2d_kernel(const CeedScalar *dqweight1d, CeedScalar *dV, const int v_stride, const int nelem) { 31 MAGMA_DEVICE_SHARED(CeedScalar, shared_data) 32 33 const int tx = threadIdx.x; 34 const int ty = threadIdx.y; 35 const int elem_id = (blockIdx.x * blockDim.y) + ty; 36 37 if (elem_id >= nelem) return; 38 39 CeedScalar rV[1][1][Q]; // allocate with DIM=NCOMP=1, but sizes may differ for a fused operator 40 // global memory pointers 41 dV += elem_id * v_stride; 42 43 // shared memory pointers 44 CeedScalar *sTweight = (CeedScalar *)shared_data; 45 46 // read dqweight_1d 47 if (ty == 0 && tx < Q) { 48 sTweight[tx] = dqweight1d[tx]; 49 } 50 51 __syncthreads(); 52 magma_weight_2d_device<CeedScalar, 1, 1, Q, 0, 0>(sTweight, rV, tx); 53 54 // write V 55 if (tx < Q) { 56 for (int j = 0; j < Q; j++) { 57 dV[j * Q + tx] = rV[0][0][j]; 58 } 59 } 60 } 61