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 3D 10 #ifndef CEED_MAGMA_BASIS_GRAD_3D_H 11 #define CEED_MAGMA_BASIS_GRAD_3D_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 #define sTmp2(i, j, ldw) sTmp2[(j) * (ldw) + (i)] 19 20 //////////////////////////////////////////////////////////////////////////////// 21 // Helper function to add or set into V 22 template <typename T, bool Add> 23 struct magma_grad_3d_device_accumulate; 24 25 template <typename T> 26 struct magma_grad_3d_device_accumulate<T, true> { 27 static __device__ __inline__ void op(T &rV, const T &rTmp) { rV += rTmp; } 28 }; 29 30 template <typename T> 31 struct magma_grad_3d_device_accumulate<T, false> { 32 static __device__ __inline__ void op(T &rV, const T &rTmp) { rV = rTmp; } 33 }; 34 35 //////////////////////////////////////////////////////////////////////////////// 36 // grad basis action (3D) 37 // This function is called three times at a higher level for 3D 38 // DIM_U -- for the size of rU[DIM_U * NUM_COMP * MAX_P_Q] 39 // DIM_V -- for the size of rV[DIM_V * NUM_COMP * MAX_P_Q] 40 // i_DIM -- the index of the outermost loop over dimensions in grad 41 // i_DIM_U -- which dim index of rU is accessed (always 0 for notrans, 0, 1, or 2 for trans) 42 // i_DIM_V -- which dim index of rV is accessed (0, 1, or 2 for notrans, always 0 for trans) 43 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, bool ADD> 44 static __device__ __inline__ void magma_grad_3d_device(const T *sTinterp, const T *sTgrad, T rU[DIM_U][NUM_COMP][rU_SIZE], 45 T rV[DIM_V][NUM_COMP][rV_SIZE], const int tx, T rTmp, T *swork) { 46 // Assumptions 47 // 0. This device routine applies grad for one dim only (i_DIM), so it should be thrice for 3D 48 // 1. 1D threads of size max(P,Q)^2 49 // 2. input: rU[DIM_U x NUM_COMP x rU_SIZE] in registers (per thread) 50 // 3. output: rV[DIM_V x NUM_COMP x rV_SIZE] in registers (per thread) 51 // 4. Three products per each (dim,component) pair 52 // 4.1 Batch P^2 of (1xP) matrices times (PxQ) matrix => Batch P^2 of (1xQ) matrices 53 // 4.2 Batch P of (QxP) matrices times (PxQ) matrix => Batch P of (QxQ) matrices 54 // 4.3 Batch 1 of (Q^2xP_) matrix times (PxQ) matrix => (Q^2xQ_) matrix 55 // 6. Each thread computes one row of the output of each product 56 // 7. Sync is recommended before and after the call 57 58 T *sW1 = swork; 59 T *sW2 = sW1 + P * P * Q; 60 for (int comp = 0; comp < NUM_COMP; comp++) { 61 // Batch P^2 of (1xP) matrices [reg] times (PxQ) matrix [shmem] => Batch P^2 of (1xQ) matrices [shmem] 62 if (tx < (P * P)) { 63 const int batchid = tx; 64 const int sld = 1; 65 const T *sT = (i_DIM == 0) ? sTgrad : sTinterp; 66 T *sTmp = sW1 + batchid * (1 * Q); 67 for (int j = 0; j < Q; j++) { 68 rTmp = 0.0; 69 for (int i = 0; i < P; i++) { 70 rTmp += rU[i_DIM_U][comp][i] * sT(i, j); 71 } 72 sTmp(0, j, sld) = rTmp; 73 } 74 } // end of: if (tx < P*P) 75 __syncthreads(); 76 77 // Batch P of (QxP) matrices [shmem] times (PxQ) matrix [shmem] => Batch P of (QxQ) matrices [reg] 78 if (tx < (P * Q)) { 79 const int batchid = tx / Q; 80 const int tx_ = tx % Q; 81 const int sld = Q; 82 const T *sT = (i_DIM == 1) ? sTgrad : sTinterp; 83 T *sTmp = sW1 + batchid * (Q * P); // sTmp is input 84 T *sTmp2 = sW2 + batchid * (Q * Q); // sTmp2 is output 85 for (int j = 0; j < Q; j++) { 86 rTmp = 0.0; 87 for (int i = 0; i < P; i++) { 88 rTmp += sTmp(tx_, i, sld) * sT(i, j); 89 } 90 sTmp2(tx_, j, sld) = rTmp; 91 } 92 } 93 __syncthreads(); 94 95 // Batch 1 of (Q^2xP_) matrices [shmem] times (PxQ) matrix [shmem] => Batch 1 of (Q^2xQ_) matrices [reg] 96 if (tx < (Q * Q)) { 97 // No need to declare batchid = (tx / Q^2) = always zero 98 // No need to declare tx_ = (tx_ % Q^2) = always tx 99 const int sld = Q * Q; 100 const T *sT = (i_DIM == 2) ? sTgrad : sTinterp; 101 T *sTmp = sW2; // sTmp is input 102 for (int j = 0; j < Q; j++) { 103 rTmp = 0.0; 104 for (int i = 0; i < P; i++) { 105 rTmp += sTmp(tx, i, sld) * sT(i, j); 106 } 107 magma_grad_3d_device_accumulate<T, ADD>::op(rV[i_DIM_V][comp][j], rTmp); 108 } 109 } 110 __syncthreads(); 111 } // loop over NUM_COMP 112 } 113 114 //////////////////////////////////////////////////////////////////////////////// 115 extern "C" __launch_bounds__(MAGMA_BASIS_BOUNDS(BASIS_MAX_P_Q *BASIS_MAX_P_Q, MAGMA_MAXTHREADS_3D)) __global__ 116 void magma_gradn_3d_kernel(const CeedScalar *dinterp1d, const CeedScalar *dgrad1d, const CeedScalar *dU, const int estrdU, const int cstrdU, 117 const int dstrdU, CeedScalar *dV, const int estrdV, const int cstrdV, const int dstrdV, const int nelem) { 118 MAGMA_DEVICE_SHARED(CeedScalar, shared_data) 119 120 const int tx = threadIdx.x; 121 const int ty = threadIdx.y; 122 const int elem_id = (blockIdx.x * blockDim.y) + ty; 123 124 if (elem_id >= nelem) return; 125 126 CeedScalar rU[1][BASIS_NUM_COMP][BASIS_P] = {0.0}; // here DIM_U = 1, but might be different for a fused operator 127 CeedScalar rV[1][BASIS_NUM_COMP][BASIS_Q] = {0.0}; // here DIM_V = 1, but might be different for a fused operator 128 CeedScalar rTmp = 0.0; 129 130 // shift global memory pointers by elem stride 131 dU += elem_id * estrdU; 132 dV += elem_id * estrdV; 133 134 // assign shared memory pointers 135 CeedScalar *sTinterp = (CeedScalar *)shared_data; 136 CeedScalar *sTgrad = sTinterp + BASIS_P * BASIS_Q; 137 CeedScalar *sTmp = sTgrad + BASIS_P * BASIS_Q; 138 sTmp += ty * (max(BASIS_P * BASIS_P * BASIS_P, (BASIS_P * BASIS_P * BASIS_Q) + (BASIS_P * BASIS_Q * BASIS_Q))); 139 140 // read T 141 if (ty == 0) { 142 read_T_notrans_gm2sm<BASIS_P, BASIS_Q>(tx, dinterp1d, sTinterp); 143 read_T_notrans_gm2sm<BASIS_P, BASIS_Q>(tx, dgrad1d, sTgrad); 144 } 145 __syncthreads(); 146 147 /* read U (idim = 0 for dU, i_DIM = 0 for rU) -- 148 there is a sync at the end of this function */ 149 read_U_3d<CeedScalar, BASIS_P, 1, BASIS_NUM_COMP, BASIS_P, 0>(dU + (0 * dstrdU), cstrdU, rU, sTmp, tx); 150 151 /* first call (i_DIM = 0, i_DIM_U = 0, i_DIM_V = 0) -- 152 output from rV[0][][] into dV (idim = 0) */ 153 magma_grad_3d_device<CeedScalar, 1, 1, BASIS_NUM_COMP, BASIS_P, BASIS_Q, BASIS_P, BASIS_Q, 0, 0, 0, false>(sTinterp, sTgrad, rU, rV, tx, rTmp, 154 sTmp); 155 /* there is a sync at the end of magma_grad_3d_device */ 156 write_V_3d<CeedScalar, BASIS_Q, 1, BASIS_NUM_COMP, BASIS_Q, 0>(dV + (0 * dstrdV), cstrdV, rV, tx); 157 158 /* second call (i_DIM = 1, i_DIM_U = 0, i_DIM_V = 0) -- 159 output from rV[0][][] into dV (idim = 1) */ 160 magma_grad_3d_device<CeedScalar, 1, 1, BASIS_NUM_COMP, BASIS_P, BASIS_Q, BASIS_P, BASIS_Q, 1, 0, 0, false>(sTinterp, sTgrad, rU, rV, tx, rTmp, 161 sTmp); 162 /* there is a sync at the end of magma_grad_3d_device */ 163 write_V_3d<CeedScalar, BASIS_Q, 1, BASIS_NUM_COMP, BASIS_Q, 0>(dV + (1 * dstrdV), cstrdV, rV, tx); 164 165 /* third call (i_DIM = 2, i_DIM_U = 0, i_DIM_V = 0) -- 166 output from rV[0][][] into dV (idim = 2) */ 167 magma_grad_3d_device<CeedScalar, 1, 1, BASIS_NUM_COMP, BASIS_P, BASIS_Q, BASIS_P, BASIS_Q, 2, 0, 0, false>(sTinterp, sTgrad, rU, rV, tx, rTmp, 168 sTmp); 169 /* there is a sync at the end of magma_grad_3d_device */ 170 write_V_3d<CeedScalar, BASIS_Q, 1, BASIS_NUM_COMP, BASIS_Q, 0>(dV + (2 * dstrdV), cstrdV, rV, tx); 171 } 172 173 //////////////////////////////////////////////////////////////////////////////// 174 extern "C" __launch_bounds__(MAGMA_BASIS_BOUNDS(BASIS_MAX_P_Q *BASIS_MAX_P_Q, MAGMA_MAXTHREADS_3D)) __global__ 175 void magma_gradt_3d_kernel(const CeedScalar *dinterp1d, const CeedScalar *dgrad1d, const CeedScalar *dU, const int estrdU, const int cstrdU, 176 const int dstrdU, CeedScalar *dV, const int estrdV, const int cstrdV, const int dstrdV, const int nelem) { 177 MAGMA_DEVICE_SHARED(CeedScalar, shared_data) 178 179 const int tx = threadIdx.x; 180 const int ty = threadIdx.y; 181 const int elem_id = (blockIdx.x * blockDim.y) + ty; 182 183 if (elem_id >= nelem) return; 184 185 CeedScalar rU[1][BASIS_NUM_COMP][BASIS_Q] = {0.0}; // here DIM_U = 1, but might be different for a fused operator 186 CeedScalar rV[1][BASIS_NUM_COMP][BASIS_P] = {0.0}; // here DIM_V = 1, but might be different for a fused operator 187 CeedScalar rTmp = 0.0; 188 189 // shift global memory pointers by elem stride 190 dU += elem_id * estrdU; 191 dV += elem_id * estrdV; 192 193 // assign shared memory pointers 194 CeedScalar *sTinterp = (CeedScalar *)shared_data; 195 CeedScalar *sTgrad = sTinterp + BASIS_Q * BASIS_P; 196 CeedScalar *sTmp = sTgrad + BASIS_Q * BASIS_P; 197 sTmp += ty * (max(BASIS_Q * BASIS_Q * BASIS_Q, (BASIS_Q * BASIS_Q * BASIS_P) + (BASIS_Q * BASIS_P * BASIS_P))); 198 199 // read T 200 if (ty == 0) { 201 read_T_trans_gm2sm<BASIS_Q, BASIS_P>(tx, dinterp1d, sTinterp); 202 read_T_trans_gm2sm<BASIS_Q, BASIS_P>(tx, dgrad1d, sTgrad); 203 } 204 __syncthreads(); 205 206 /* read U (idim = 0 for dU, i_DIM = 0 for rU) -- 207 there is a sync at the end of this function */ 208 read_U_3d<CeedScalar, BASIS_Q, 1, BASIS_NUM_COMP, BASIS_Q, 0>(dU + (0 * dstrdU), cstrdU, rU, sTmp, tx); 209 /* then first call (i_DIM = 0, i_DIM_U = 0, i_DIM_V = 0) */ 210 magma_grad_3d_device<CeedScalar, 1, 1, BASIS_NUM_COMP, BASIS_Q, BASIS_P, BASIS_Q, BASIS_P, 0, 0, 0, true>(sTinterp, sTgrad, rU, rV, tx, rTmp, sTmp); 211 /* there is a sync at the end of magma_grad_3d_device */ 212 213 /* read U (idim = 1 for dU, i_DIM = 0 for rU) -- 214 there is a sync at the end of this function */ 215 read_U_3d<CeedScalar, BASIS_Q, 1, BASIS_NUM_COMP, BASIS_Q, 0>(dU + (1 * dstrdU), cstrdU, rU, sTmp, tx); 216 /* then second call (i_DIM = 1, i_DIM_U = 0, i_DIM_V = 0) */ 217 magma_grad_3d_device<CeedScalar, 1, 1, BASIS_NUM_COMP, BASIS_Q, BASIS_P, BASIS_Q, BASIS_P, 1, 0, 0, true>(sTinterp, sTgrad, rU, rV, tx, rTmp, sTmp); 218 /* there is a sync at the end of magma_grad_3d_device */ 219 220 /* read U (idim = 2 for dU, i_DIM = 0 for rU) -- 221 there is a sync at the end of this function */ 222 read_U_3d<CeedScalar, BASIS_Q, 1, BASIS_NUM_COMP, BASIS_Q, 0>(dU + (2 * dstrdU), cstrdU, rU, sTmp, tx); 223 /* then third call (i_DIM = 2, i_DIM_U = 0, i_DIM_V = 0) */ 224 magma_grad_3d_device<CeedScalar, 1, 1, BASIS_NUM_COMP, BASIS_Q, BASIS_P, BASIS_Q, BASIS_P, 2, 0, 0, true>(sTinterp, sTgrad, rU, rV, tx, rTmp, sTmp); 225 /* there is a sync at the end of magma_grad_3d_device */ 226 227 // write V 228 write_V_3d<CeedScalar, BASIS_P, 1, BASIS_NUM_COMP, BASIS_P, 0>(dV + (0 * dstrdV), cstrdV, rV, tx); 229 } 230 231 #endif // CEED_MAGMA_BASIS_GRAD_3D_H 232