xref: /libCEED/include/ceed/jit-source/cuda/cuda-ref-basis-tensor.h (revision c0b5abf0f23b15c4f0ada76f8abe9f8d2b6fa247)
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 CUDA tensor product basis
10 
11 #include <ceed/types.h>
12 
13 //------------------------------------------------------------------------------
14 // Tensor Basis Kernels
15 //------------------------------------------------------------------------------
16 
17 //------------------------------------------------------------------------------
18 // Interp
19 //------------------------------------------------------------------------------
20 extern "C" __global__ void Interp(const CeedInt num_elem, const CeedInt is_transpose, const CeedScalar *__restrict__ interp_1d,
21                                   const CeedScalar *__restrict__ u, CeedScalar *__restrict__ v) {
22   const CeedInt i = threadIdx.x;
23 
24   __shared__ CeedScalar s_mem[BASIS_Q_1D * BASIS_P_1D + 2 * BASIS_BUF_LEN];
25   CeedScalar           *s_interp_1d = s_mem;
26   CeedScalar           *s_buffer_1  = s_mem + BASIS_Q_1D * BASIS_P_1D;
27   CeedScalar           *s_buffer_2  = s_buffer_1 + BASIS_BUF_LEN;
28   for (CeedInt k = i; k < BASIS_Q_1D * BASIS_P_1D; k += blockDim.x) {
29     s_interp_1d[k] = interp_1d[k];
30   }
31 
32   const CeedInt P             = is_transpose ? BASIS_Q_1D : BASIS_P_1D;
33   const CeedInt Q             = is_transpose ? BASIS_P_1D : BASIS_Q_1D;
34   const CeedInt stride_0      = is_transpose ? 1 : BASIS_P_1D;
35   const CeedInt stride_1      = is_transpose ? BASIS_P_1D : 1;
36   const CeedInt u_stride      = is_transpose ? BASIS_NUM_QPTS : BASIS_NUM_NODES;
37   const CeedInt v_stride      = is_transpose ? BASIS_NUM_NODES : BASIS_NUM_QPTS;
38   const CeedInt u_comp_stride = num_elem * (is_transpose ? BASIS_NUM_QPTS : BASIS_NUM_NODES);
39   const CeedInt v_comp_stride = num_elem * (is_transpose ? BASIS_NUM_NODES : BASIS_NUM_QPTS);
40   const CeedInt u_size        = is_transpose ? BASIS_NUM_QPTS : BASIS_NUM_NODES;
41 
42   // Apply basis element by element
43   for (CeedInt elem = blockIdx.x; elem < num_elem; elem += gridDim.x) {
44     for (CeedInt comp = 0; comp < BASIS_NUM_COMP; comp++) {
45       const CeedScalar *cur_u = &u[elem * u_stride + comp * u_comp_stride];
46       CeedScalar       *cur_v = &v[elem * v_stride + comp * v_comp_stride];
47       CeedInt           pre   = u_size;
48       CeedInt           post  = 1;
49 
50       for (CeedInt d = 0; d < BASIS_DIM; d++) {
51         __syncthreads();
52         // Update buffers used
53         pre /= P;
54         const CeedScalar *in       = d == 0 ? cur_u : (d % 2 ? s_buffer_2 : s_buffer_1);
55         CeedScalar       *out      = d == BASIS_DIM - 1 ? cur_v : (d % 2 ? s_buffer_1 : s_buffer_2);
56         const CeedInt     writeLen = pre * post * Q;
57 
58         // Contract along middle index
59         for (CeedInt k = i; k < writeLen; k += blockDim.x) {
60           const CeedInt c   = k % post;
61           const CeedInt j   = (k / post) % Q;
62           const CeedInt a   = k / (post * Q);
63           CeedScalar    v_k = 0;
64 
65           for (CeedInt b = 0; b < P; b++) v_k += s_interp_1d[j * stride_0 + b * stride_1] * in[(a * P + b) * post + c];
66           if (is_transpose && d == BASIS_DIM - 1) out[k] += v_k;
67           else out[k] = v_k;
68         }
69         post *= Q;
70       }
71     }
72   }
73 }
74 
75 //------------------------------------------------------------------------------
76 // Grad
77 //------------------------------------------------------------------------------
78 extern "C" __global__ void Grad(const CeedInt num_elem, const CeedInt is_transpose, const CeedScalar *__restrict__ interp_1d,
79                                 const CeedScalar *__restrict__ grad_1d, const CeedScalar *__restrict__ u, CeedScalar *__restrict__ v) {
80   const CeedInt i = threadIdx.x;
81 
82   __shared__ CeedScalar s_mem[2 * (BASIS_Q_1D * BASIS_P_1D + BASIS_BUF_LEN)];
83   CeedScalar           *s_interp_1d = s_mem;
84   CeedScalar           *s_grad_1d   = s_interp_1d + BASIS_Q_1D * BASIS_P_1D;
85   CeedScalar           *s_buffer_1  = s_grad_1d + BASIS_Q_1D * BASIS_P_1D;
86   CeedScalar           *s_buffer_2  = s_buffer_1 + BASIS_BUF_LEN;
87   for (CeedInt k = i; k < BASIS_Q_1D * BASIS_P_1D; k += blockDim.x) {
88     s_interp_1d[k] = interp_1d[k];
89     s_grad_1d[k]   = grad_1d[k];
90   }
91 
92   const CeedInt P             = is_transpose ? BASIS_Q_1D : BASIS_P_1D;
93   const CeedInt Q             = is_transpose ? BASIS_P_1D : BASIS_Q_1D;
94   const CeedInt stride_0      = is_transpose ? 1 : BASIS_P_1D;
95   const CeedInt stride_1      = is_transpose ? BASIS_P_1D : 1;
96   const CeedInt u_stride      = is_transpose ? BASIS_NUM_QPTS : BASIS_NUM_NODES;
97   const CeedInt v_stride      = is_transpose ? BASIS_NUM_NODES : BASIS_NUM_QPTS;
98   const CeedInt u_comp_stride = num_elem * (is_transpose ? BASIS_NUM_QPTS : BASIS_NUM_NODES);
99   const CeedInt v_comp_stride = num_elem * (is_transpose ? BASIS_NUM_NODES : BASIS_NUM_QPTS);
100   const CeedInt u_dim_stride  = is_transpose ? num_elem * BASIS_NUM_QPTS * BASIS_NUM_COMP : 0;
101   const CeedInt v_dim_stride  = is_transpose ? 0 : num_elem * BASIS_NUM_QPTS * BASIS_NUM_COMP;
102 
103   // Apply basis element by element
104   for (CeedInt elem = blockIdx.x; elem < num_elem; elem += gridDim.x) {
105     for (CeedInt comp = 0; comp < BASIS_NUM_COMP; comp++) {
106       // dim*dim contractions for grad
107       for (CeedInt dim_1 = 0; dim_1 < BASIS_DIM; dim_1++) {
108         CeedInt           pre   = is_transpose ? BASIS_NUM_QPTS : BASIS_NUM_NODES;
109         CeedInt           post  = 1;
110         const CeedScalar *cur_u = &u[elem * u_stride + dim_1 * u_dim_stride + comp * u_comp_stride];
111         CeedScalar       *cur_v = &v[elem * v_stride + dim_1 * v_dim_stride + comp * v_comp_stride];
112 
113         for (CeedInt dim_2 = 0; dim_2 < BASIS_DIM; dim_2++) {
114           __syncthreads();
115           // Update buffers used
116           pre /= P;
117           const CeedScalar *op       = dim_1 == dim_2 ? s_grad_1d : s_interp_1d;
118           const CeedScalar *in       = dim_2 == 0 ? cur_u : (dim_2 % 2 ? s_buffer_2 : s_buffer_1);
119           CeedScalar       *out      = dim_2 == BASIS_DIM - 1 ? cur_v : (dim_2 % 2 ? s_buffer_1 : s_buffer_2);
120           const CeedInt     writeLen = pre * post * Q;
121 
122           // Contract along middle index
123           for (CeedInt k = i; k < writeLen; k += blockDim.x) {
124             const CeedInt c   = k % post;
125             const CeedInt j   = (k / post) % Q;
126             const CeedInt a   = k / (post * Q);
127             CeedScalar    v_k = 0;
128 
129             for (CeedInt b = 0; b < P; b++) v_k += op[j * stride_0 + b * stride_1] * in[(a * P + b) * post + c];
130             if (is_transpose && dim_2 == BASIS_DIM - 1) out[k] += v_k;
131             else out[k] = v_k;
132           }
133           post *= Q;
134         }
135       }
136     }
137   }
138 }
139 
140 //------------------------------------------------------------------------------
141 // 1D quadrature weights
142 //------------------------------------------------------------------------------
143 __device__ void Weight1d(const CeedInt num_elem, const CeedScalar *q_weight_1d, CeedScalar *w) {
144   const CeedInt i = threadIdx.x;
145 
146   if (i < BASIS_Q_1D) {
147     const size_t elem = blockIdx.x;
148 
149     if (elem < num_elem) w[elem * BASIS_Q_1D + i] = q_weight_1d[i];
150   }
151 }
152 
153 //------------------------------------------------------------------------------
154 // 2D quadrature weights
155 //------------------------------------------------------------------------------
156 __device__ void Weight2d(const CeedInt num_elem, const CeedScalar *q_weight_1d, CeedScalar *w) {
157   const CeedInt i = threadIdx.x;
158   const CeedInt j = threadIdx.y;
159 
160   if (i < BASIS_Q_1D && j < BASIS_Q_1D) {
161     const size_t elem = blockIdx.x;
162 
163     if (elem < num_elem) {
164       const size_t ind = (elem * BASIS_Q_1D + j) * BASIS_Q_1D + i;
165 
166       w[ind] = q_weight_1d[i] * q_weight_1d[j];
167     }
168   }
169 }
170 
171 //------------------------------------------------------------------------------
172 // 3D quadrature weights
173 //------------------------------------------------------------------------------
174 __device__ void Weight3d(const CeedInt num_elem, const CeedScalar *q_weight_1d, CeedScalar *w) {
175   const CeedInt i = threadIdx.x;
176   const CeedInt j = threadIdx.y;
177 
178   if (i < BASIS_Q_1D && j < BASIS_Q_1D) {
179     const size_t elem = blockIdx.x;
180 
181     if (elem < num_elem) {
182       for (CeedInt k = 0; k < BASIS_Q_1D; k++) {
183         const size_t ind = ((elem * BASIS_Q_1D + k) * BASIS_Q_1D + j) * BASIS_Q_1D + i;
184 
185         w[ind] = q_weight_1d[i] * q_weight_1d[j] * q_weight_1d[k];
186       }
187     }
188   }
189 }
190 
191 //------------------------------------------------------------------------------
192 // Quadrature weights
193 //------------------------------------------------------------------------------
194 extern "C" __global__ void Weight(const CeedInt num_elem, const CeedScalar *__restrict__ q_weight_1d, CeedScalar *__restrict__ v) {
195   if (BASIS_DIM == 1) Weight1d(num_elem, q_weight_1d, v);
196   else if (BASIS_DIM == 2) Weight2d(num_elem, q_weight_1d, v);
197   else if (BASIS_DIM == 3) Weight3d(num_elem, q_weight_1d, v);
198 }
199