xref: /libCEED/rust/libceed-sys/c-src/include/ceed/jit-source/magma/magma-basis-grad-2d.h (revision 9e0c01fa54c8586fe77be0891a986d6b8bea4638)
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 gradient in 2D
10 #ifndef CEED_MAGMA_BASIS_GRAD_2D_H
11 #define CEED_MAGMA_BASIS_GRAD_2D_H
12 
13 #include "magma-common-tensor.h"
14 
15 // macros to abstract access of shared memory and reg. file
16 #define sT(i, j) sT[(j)*P + (i)]
17 #define sTmp(i, j, ldw) sTmp[(j) * (ldw) + (i)]
18 
19 ////////////////////////////////////////////////////////////////////////////////
20 // grad basis action (2D)
21 // This function is called two times at a higher level for 2D
22 // DIM_U   -- for the size of rU[DIM_U * NUM_COMP * MAX_P_Q]
23 // DIM_V   -- for the size of rV[DIM_V * NUM_COMP * MAX_P_Q]
24 // i_DIM   -- the index of the outermost loop over dimensions in grad
25 // i_DIM_U -- which dim index of rU is accessed (always 0 for notrans, 0 or 1 for trans)
26 // i_DIM_V -- which dim index of rV is accessed (0 or 1 for notrans, always 0 for trans)
27 // the scalar beta is used to specify whether to accumulate to rV, or overwrite it
28 template <typename T, int DIM_U, int DIM_V, int NUM_COMP, int P, int Q, int rU_SIZE, int rV_SIZE, int i_DIM, int i_DIM_U, int i_DIM_V>
29 static __device__ __inline__ void magma_grad_2d_device(const T *sTinterp, const T *sTgrad, T rU[DIM_U][NUM_COMP][rU_SIZE],
30                                                        T rV[DIM_V][NUM_COMP][rV_SIZE], T beta, const int tx, T rTmp, T *swork) {
31   // Assumptions
32   // 0. This device routine applies grad for one dim only (i_DIM), so it should be called twice for 2D
33   // 1. 1D threads of size max(P,Q)
34   // 2. input:  rU[DIM_U x NUM_COMP x P] in registers (per thread)
35   // 3. output: rV[DIM_V x NUM_COMP x Q] in registers (per thread)
36   // 4. Two products per each (dim,component) pair
37   //  4.1 Batch P of (1xP) matrices times (PxQ) matrix => Batch P of (1xQ) matrices
38   //  4.2 Batch 1 of (QxP) matrix   times (PxQ) matrix => (QxQ) matrix
39   // 6. Each thread computes one row of the output of each product
40   // 7. Sync is recommended before and after the call
41 
42   for (int comp = 0; comp < NUM_COMP; comp++) {
43     // 1st product -- Batch P of (1xP) matrices [reg] x (PxQ) [shmem] => Batch P of (1xQ) matrices
44     // the batch output P x (1xQ) is written on the fly to shmem
45     if (tx < P) {
46       const int batchid = tx;
47       const int sld     = 1;
48       const T  *sT      = (i_DIM == 0) ? sTgrad : sTinterp;
49       T        *sTmp    = swork + batchid * (1 * Q);
50       for (int j = 0; j < Q; j++) {
51         rTmp = 0.0;
52         for (int i = 0; i < P; i++) {
53           rTmp += rU[i_DIM_U][comp][i] * sT(i, j);
54         }
55         sTmp(0, j, sld) = rTmp;
56       }
57     }  // end of: if (tx < P)
58     __syncthreads();
59 
60     // 2nd product -- Batch 1 of a (QxP) matrix [shmem] x (PxQ) [shmem] => (QxQ) matrix [reg]
61     if (tx < Q) {
62       const int batchid = 0;
63       const int sld     = Q;
64       const T  *sT      = (i_DIM == 1) ? sTgrad : sTinterp;
65       T        *sTmp    = swork + batchid * (Q * P);
66       for (int j = 0; j < Q; j++) {
67         rTmp = 0.0;
68         for (int i = 0; i < P; i++) {
69           rTmp += sTmp(tx, i, sld) * sT(i, j);
70         }
71         rV[i_DIM_V][comp][j] *= beta;
72         rV[i_DIM_V][comp][j] += rTmp;
73       }
74     }
75     __syncthreads();
76   }  // loop over NUM_COMP
77 }
78 
79 ////////////////////////////////////////////////////////////////////////////////
80 extern "C" __launch_bounds__(MAGMA_BASIS_BOUNDS(BASIS_MAX_P_Q, MAGMA_MAXTHREADS_2D)) __global__
81     void magma_gradn_2d_kernel(const CeedScalar *dinterp1d, const CeedScalar *dgrad1d, const CeedScalar *dU, const int estrdU, const int cstrdU,
82                                const int dstrdU, CeedScalar *dV, const int estrdV, const int cstrdV, const int dstrdV, const int nelem) {
83   MAGMA_DEVICE_SHARED(CeedScalar, shared_data)
84 
85   const int     tx      = threadIdx.x;
86   const int     ty      = threadIdx.y;
87   const int     elem_id = (blockIdx.x * blockDim.y) + ty;
88   magma_trans_t transT  = MagmaNoTrans;
89 
90   if (elem_id >= nelem) return;
91 
92   CeedScalar rU[1][BASIS_NUM_COMP][BASIS_P] = {0.0};  // here DIM_U = 1, but might be different for a fused operator
93   CeedScalar rV[1][BASIS_NUM_COMP][BASIS_Q] = {0.0};  // here DIM_V = 1, but might be different for a fused operator
94   CeedScalar rTmp                           = 0.0;
95 
96   // shift global memory pointers by elem stride
97   dU += elem_id * estrdU;
98   dV += elem_id * estrdV;
99 
100   // assign shared memory pointers
101   CeedScalar *sTinterp = (CeedScalar *)shared_data;
102   CeedScalar *sTgrad   = sTinterp + BASIS_P * BASIS_Q;
103   CeedScalar *sTmp     = sTgrad + BASIS_P * BASIS_Q;
104   sTmp += ty * (BASIS_P * BASIS_MAX_P_Q);
105 
106   // read T
107   if (ty == 0) {
108     read_T_notrans_gm2sm<BASIS_P, BASIS_Q>(tx, dinterp1d, sTinterp);
109     read_T_notrans_gm2sm<BASIS_P, BASIS_Q>(tx, dgrad1d, sTgrad);
110   }
111 
112   // No need to read V ( required only in transposed grad )
113   const CeedScalar beta = 0.0;
114 
115   /* read U (idim = 0 for dU, i_DIM = 0 for rU) --
116      there is a sync at the end of this function */
117   read_U_2d<CeedScalar, BASIS_P, 1, BASIS_NUM_COMP, BASIS_P, 0>(dU + (0 * dstrdU), cstrdU, rU, sTmp, tx);
118 
119   /* first call (i_DIM = 0, i_DIM_U = 0, i_DIM_V = 0) --
120      output from rV[0][][] into dV (idim = 0) */
121   magma_grad_2d_device<CeedScalar, 1, 1, BASIS_NUM_COMP, BASIS_P, BASIS_Q, BASIS_P, BASIS_Q, 0, 0, 0>(sTinterp, sTgrad, rU, rV, beta, tx, rTmp, sTmp);
122   /* there is a sync at the end of magma_grad_2d_device */
123   write_V_2d<CeedScalar, BASIS_Q, 1, BASIS_NUM_COMP, BASIS_Q, 0>(dV + (0 * dstrdV), cstrdV, rV, tx);
124 
125   /* second call (i_DIM = 1, i_DIM_U = 0, i_DIM_V = 0) --
126   output from rV[0][][] into dV (idim = 1) */
127   magma_grad_2d_device<CeedScalar, 1, 1, BASIS_NUM_COMP, BASIS_P, BASIS_Q, BASIS_P, BASIS_Q, 1, 0, 0>(sTinterp, sTgrad, rU, rV, beta, tx, rTmp, sTmp);
128   /* there is a sync at the end of magma_grad_2d_device */
129   write_V_2d<CeedScalar, BASIS_Q, 1, BASIS_NUM_COMP, BASIS_Q, 0>(dV + (1 * dstrdV), cstrdV, rV, tx);
130 }
131 
132 ////////////////////////////////////////////////////////////////////////////////
133 extern "C" __launch_bounds__(MAGMA_BASIS_BOUNDS(BASIS_MAX_P_Q, MAGMA_MAXTHREADS_2D)) __global__
134     void magma_gradt_2d_kernel(const CeedScalar *dinterp1d, const CeedScalar *dgrad1d, const CeedScalar *dU, const int estrdU, const int cstrdU,
135                                const int dstrdU, CeedScalar *dV, const int estrdV, const int cstrdV, const int dstrdV, const int nelem) {
136   MAGMA_DEVICE_SHARED(CeedScalar, shared_data)
137 
138   const int     tx      = threadIdx.x;
139   const int     ty      = threadIdx.y;
140   const int     elem_id = (blockIdx.x * blockDim.y) + ty;
141   magma_trans_t transT  = MagmaTrans;
142 
143   if (elem_id >= nelem) return;
144 
145   CeedScalar rU[1][BASIS_NUM_COMP][BASIS_Q] = {0.0};  // here DIM_U = 1, but might be different for a fused operator
146   CeedScalar rV[1][BASIS_NUM_COMP][BASIS_P] = {0.0};  // here DIM_V = 1, but might be different for a fused operator
147   CeedScalar rTmp                           = 0.0;
148 
149   // shift global memory pointers by elem stride
150   dU += elem_id * estrdU;
151   dV += elem_id * estrdV;
152 
153   // assign shared memory pointers
154   CeedScalar *sTinterp = (CeedScalar *)shared_data;
155   CeedScalar *sTgrad   = sTinterp + BASIS_Q * BASIS_P;
156   CeedScalar *sTmp     = sTgrad + BASIS_Q * BASIS_P;
157   sTmp += ty * (BASIS_Q * BASIS_MAX_P_Q);
158 
159   // read T
160   if (ty == 0) {
161     read_T_trans_gm2sm<BASIS_Q, BASIS_P>(tx, dinterp1d, sTinterp);
162     read_T_trans_gm2sm<BASIS_Q, BASIS_P>(tx, dgrad1d, sTgrad);
163   }
164   __syncthreads();
165 
166   /* read V (since this is transposed mode --
167      idim = 0 for dV, i_DIM = 0 for rV) */
168   const CeedScalar beta = 1.0;
169   read_V_2d<CeedScalar, BASIS_P, 1, BASIS_NUM_COMP, BASIS_P, 0>(dV + (0 * dstrdV), cstrdV, rV, tx);
170 
171   /* read U (idim = 0 for dU, i_DIM = 0 for rU) --
172      there is a sync at the end of this function */
173   read_U_2d<CeedScalar, BASIS_Q, 1, BASIS_NUM_COMP, BASIS_Q, 0>(dU + (0 * dstrdU), cstrdU, rU, sTmp, tx);
174   /* first call (i_DIM = 0, i_DIM_U = 0, i_DIM_V = 0) */
175   magma_grad_2d_device<CeedScalar, 1, 1, BASIS_NUM_COMP, BASIS_Q, BASIS_P, BASIS_Q, BASIS_P, 0, 0, 0>(sTinterp, sTgrad, rU, rV, beta, tx, rTmp, sTmp);
176   /* there is a sync at the end of magma_grad_2d_device */
177 
178   /* read U (idim = 1 for dU, i_DIM = 0 for rU) --
179      there is a sync at the end of this function */
180   read_U_2d<CeedScalar, BASIS_Q, 1, BASIS_NUM_COMP, BASIS_Q, 0>(dU + (1 * dstrdU), cstrdU, rU, sTmp, tx);
181   /* second call (i_DIM = 1, i_DIM_U = 0, i_DIM_V = 0) */
182   magma_grad_2d_device<CeedScalar, 1, 1, BASIS_NUM_COMP, BASIS_Q, BASIS_P, BASIS_Q, BASIS_P, 1, 0, 0>(sTinterp, sTgrad, rU, rV, beta, tx, rTmp, sTmp);
183   /* there is a sync at the end of magma_grad_2d_device */
184 
185   // write V
186   write_V_2d<CeedScalar, BASIS_P, 1, BASIS_NUM_COMP, BASIS_P, 0>(dV + (0 * dstrdV), cstrdV, rV, tx);
187 }
188 
189 #endif  // CEED_MAGMA_BASIS_GRAD_2D_H
190