xref: /libCEED/include/ceed/jit-source/magma/magma-basis-weight-3d.h (revision 07d5b98a8feba68a643190b8ea9bcdac5c3e6570)
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 weight in 3D
10 
11 #include "magma-common-tensor.h"
12 
13 ////////////////////////////////////////////////////////////////////////////////
14 // weight basis action -- 3D
15 template <typename T, int DIM, int NUM_COMP, int Q, int i_DIM, int i_COMP>
16 static __device__ __inline__ void magma_weight_3d_device(const T *sTweight, T rV[DIM][NUM_COMP][Q], const int tx) {
17   // Assumptions
18   // 1. 1D thread configuration of size Q^2
19   // 2. rV[][][] matches the storage used in other actions (interp, grad, ... etc)
20   // 3. i_DIM and i_COMP specify which indexes to use in rV,
21   //    since the output per thread is a register array of size Q
22   // 4. Sync is recommended after the call (to make sure sTweight can be overwritten)
23 
24   if (tx < Q * Q) {
25     // x sTweight[j]    for first update
26     // x sTweight[tx%Q] for second update
27     // x sTweight[tx/Q] for third update
28     for (int j = 0; j < Q; j++) {
29       rV[i_DIM][i_COMP][j] = sTweight[j] * sTweight[tx % Q] * sTweight[tx / Q];
30     }
31   }
32 }
33 
34 ////////////////////////////////////////////////////////////////////////////////
35 extern "C" __launch_bounds__(MAGMA_BASIS_BOUNDS(BASIS_Q *BASIS_Q, MAGMA_MAXTHREADS_3D)) __global__
36     void magma_weight_3d_kernel(const CeedScalar *dqweight1d, CeedScalar *dV, const int v_stride, const int nelem) {
37   MAGMA_DEVICE_SHARED(CeedScalar, shared_data)
38 
39   const int tx      = threadIdx.x;
40   const int ty      = threadIdx.y;
41   const int elem_id = (blockIdx.x * blockDim.y) + ty;
42 
43   if (elem_id >= nelem) return;
44 
45   CeedScalar rV[1][1][BASIS_Q];  // allocate with BASIS_DIM=BASIS_NUM_COMP=1, but sizes may differ for a fused operator
46   // global memory pointers
47   dV += elem_id * v_stride;
48 
49   // shared memory pointers
50   CeedScalar *sTweight = (CeedScalar *)shared_data;
51 
52   // read dqweight_1d
53   if (tx < BASIS_Q) {
54     sTweight[tx] = dqweight1d[tx];
55   }
56   __syncthreads();
57 
58   magma_weight_3d_device<CeedScalar, 1, 1, BASIS_Q, 0, 0>(sTweight, rV, tx);
59 
60   // write V
61   if (tx < (BASIS_Q * BASIS_Q)) {
62     for (int j = 0; j < BASIS_Q; j++) {
63       dV[j * (BASIS_Q * BASIS_Q) + tx] = rV[0][0][j];
64     }
65   }
66 }
67