xref: /libCEED/include/ceed/jit-source/magma/magma-basis-weight-2d.h (revision e6aaec067db7781c9cc5c3b72972e2661403e34f)
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 /// @file
9 /// Internal header for MAGMA tensor basis weight in 2D
10 #ifndef CEED_MAGMA_BASIS_WEIGHT_2D_H
11 #define CEED_MAGMA_BASIS_WEIGHT_2D_H
12 
13 #include "magma-common-tensor.h"
14 
15 //////////////////////////////////////////////////////////////////////////////////////////
16 // weight basis action -- 2D
17 template <typename T, int DIM, int NUM_COMP, int Q, int i_DIM, int i_COMP>
18 static __device__ __inline__ void magma_weight_2d_device(const T *sTweight, T rV[DIM][NUM_COMP][Q], const int tx) {
19   // Assumptions
20   // 1. 1D thread configuration of size Q
21   // 2. rV[][][] matches the storage used in other actions (interp, grad, ... etc)
22   // 3. i_DIM and i_COMP specify which indexes to use in rV,
23   //    since the output per thread is a register array of size Q
24   // 4. Sync is recommended after the call (to make sure sTweight can be overwritten)
25 
26   if (tx < Q) {
27     // x sTweight[j]  for first update
28     // x sTweight[tx] for second update
29     for (int j = 0; j < Q; j++) {
30       rV[i_DIM][i_COMP][j] = sTweight[j] * sTweight[tx];
31     }
32   }
33 }
34 
35 //////////////////////////////////////////////////////////////////////////////////////////
36 extern "C" __launch_bounds__(MAGMA_BASIS_BOUNDS(BASIS_Q, MAGMA_MAXTHREADS_2D)) __global__
37     void magma_weight_2d_kernel(const CeedScalar *dqweight1d, CeedScalar *dV, const int v_stride, const int nelem) {
38   MAGMA_DEVICE_SHARED(CeedScalar, shared_data)
39 
40   const int tx      = threadIdx.x;
41   const int ty      = threadIdx.y;
42   const int elem_id = (blockIdx.x * blockDim.y) + ty;
43 
44   if (elem_id >= nelem) return;
45 
46   CeedScalar rV[1][1][BASIS_Q];  // allocate with BASIS_DIM=BASIS_NUM_COMP=1, but sizes may differ for a fused operator
47   // global memory pointers
48   dV += elem_id * v_stride;
49 
50   // shared memory pointers
51   CeedScalar *sTweight = (CeedScalar *)shared_data;
52 
53   // read dqweight_1d
54   if (ty == 0 && tx < BASIS_Q) {
55     sTweight[tx] = dqweight1d[tx];
56   }
57 
58   __syncthreads();
59   magma_weight_2d_device<CeedScalar, 1, 1, BASIS_Q, 0, 0>(sTweight, rV, tx);
60 
61   // write V
62   if (tx < BASIS_Q) {
63     for (int j = 0; j < BASIS_Q; j++) {
64       dV[j * BASIS_Q + tx] = rV[0][0][j];
65     }
66   }
67 }
68 
69 #endif  // CEED_MAGMA_BASIS_WEIGHT_2D_H
70