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 #include <ceed.h> 9 #include <ceed/backend.h> 10 #include <ceed/jit-tools.h> 11 #include <string.h> 12 13 #ifdef CEED_MAGMA_USE_HIP 14 #include "../hip/ceed-hip-common.h" 15 #include "../hip/ceed-hip-compile.h" 16 #else 17 #include "../cuda/ceed-cuda-common.h" 18 #include "../cuda/ceed-cuda-compile.h" 19 #endif 20 #include "ceed-magma-common.h" 21 #include "ceed-magma.h" 22 23 #include "ceed-magma-gemm-nontensor.h" 24 #include "ceed-magma-gemm-selector.h" 25 26 //------------------------------------------------------------------------------ 27 // Basis apply - tensor 28 //------------------------------------------------------------------------------ 29 static int CeedBasisApply_Magma(CeedBasis basis, CeedInt num_elem, CeedTransposeMode t_mode, CeedEvalMode e_mode, CeedVector u, CeedVector v) { 30 Ceed ceed; 31 Ceed_Magma *data; 32 CeedInt dim, num_comp, num_nodes, P_1d, Q_1d, P, Q; 33 const CeedScalar *d_u; 34 CeedScalar *d_v; 35 CeedBasis_Magma *impl; 36 37 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 38 CeedCallBackend(CeedGetData(ceed, &data)); 39 CeedCallBackend(CeedBasisGetData(basis, &impl)); 40 CeedCallBackend(CeedBasisGetDimension(basis, &dim)); 41 CeedCallBackend(CeedBasisGetNumComponents(basis, &num_comp)); 42 CeedCallBackend(CeedBasisGetNumNodes(basis, &num_nodes)); 43 CeedCallBackend(CeedBasisGetNumNodes1D(basis, &P_1d)); 44 CeedCallBackend(CeedBasisGetNumQuadraturePoints1D(basis, &Q_1d)); 45 P = P_1d; 46 Q = Q_1d; 47 if (t_mode == CEED_TRANSPOSE) { 48 P = Q_1d; 49 Q = P_1d; 50 } 51 52 // Read vectors 53 if (u != CEED_VECTOR_NONE) CeedCallBackend(CeedVectorGetArrayRead(u, CEED_MEM_DEVICE, &d_u)); 54 else CeedCheck(e_mode == CEED_EVAL_WEIGHT, ceed, CEED_ERROR_BACKEND, "An input vector is required for this CeedEvalMode"); 55 CeedCallBackend(CeedVectorGetArrayWrite(v, CEED_MEM_DEVICE, &d_v)); 56 57 // Apply basis operation 58 switch (e_mode) { 59 case CEED_EVAL_INTERP: { 60 // Define element sizes for dofs/quad 61 CeedInt elem_qpts_size = CeedIntPow(Q_1d, dim); 62 CeedInt elem_dofs_size = CeedIntPow(P_1d, dim); 63 64 // E-vector ordering -------------- Q-vector ordering 65 // component component 66 // elem elem 67 // node node 68 69 // --- Define strides for NOTRANSPOSE mode: --- 70 // Input (d_u) is E-vector, output (d_v) is Q-vector 71 72 // Element strides 73 CeedInt u_elem_stride = elem_dofs_size; 74 CeedInt v_elem_stride = elem_qpts_size; 75 // Component strides 76 CeedInt u_comp_stride = num_elem * elem_dofs_size; 77 CeedInt v_comp_stride = num_elem * elem_qpts_size; 78 if (t_mode == CEED_TRANSPOSE) { 79 // Input (d_u) is Q-vector, output (d_v) is E-vector 80 // Element strides 81 v_elem_stride = elem_dofs_size; 82 u_elem_stride = elem_qpts_size; 83 // Component strides 84 v_comp_stride = num_elem * elem_dofs_size; 85 u_comp_stride = num_elem * elem_qpts_size; 86 } 87 CeedInt num_threads = 1; 88 CeedInt num_t_col = 1; 89 CeedInt shared_mem = 0; 90 CeedInt max_P_Q = CeedIntMax(P, Q); 91 92 switch (dim) { 93 case 1: 94 num_threads = max_P_Q; 95 num_t_col = MAGMA_BASIS_NTCOL(num_threads, MAGMA_MAXTHREADS_1D); 96 shared_mem += sizeof(CeedScalar) * num_t_col * (num_comp * (1 * P + 1 * Q)); 97 shared_mem += sizeof(CeedScalar) * (P * Q); 98 break; 99 case 2: 100 num_threads = max_P_Q; 101 num_t_col = MAGMA_BASIS_NTCOL(num_threads, MAGMA_MAXTHREADS_2D); 102 shared_mem += P * Q * sizeof(CeedScalar); // for sT 103 // for reforming rU we need P x P, and for the intermediate output we need P x Q 104 shared_mem += num_t_col * (P * max_P_Q * sizeof(CeedScalar)); 105 break; 106 case 3: 107 num_threads = max_P_Q * max_P_Q; 108 num_t_col = MAGMA_BASIS_NTCOL(num_threads, MAGMA_MAXTHREADS_3D); 109 shared_mem += sizeof(CeedScalar) * (P * Q); // for sT 110 // rU needs P^2 x P, the intermediate output needs max(P^2 x Q, P x Q^2) 111 shared_mem += sizeof(CeedScalar) * num_t_col * (CeedIntMax(P * P * max_P_Q, P * Q * Q)); 112 break; 113 } 114 CeedInt grid = CeedDivUpInt(num_elem, num_t_col); 115 void *args[] = {&impl->d_interp_1d, &d_u, &u_elem_stride, &u_comp_stride, &d_v, &v_elem_stride, &v_comp_stride, &num_elem}; 116 117 if (t_mode == CEED_TRANSPOSE) { 118 CeedCallBackend(CeedRunKernelDimSharedMagma(ceed, impl->InterpTranspose, grid, num_threads, num_t_col, 1, shared_mem, args)); 119 } else { 120 CeedCallBackend(CeedRunKernelDimSharedMagma(ceed, impl->Interp, grid, num_threads, num_t_col, 1, shared_mem, args)); 121 } 122 } break; 123 case CEED_EVAL_GRAD: { 124 // Define element sizes for dofs/quad 125 CeedInt elem_qpts_size = CeedIntPow(Q_1d, dim); 126 CeedInt elem_dofs_size = CeedIntPow(P_1d, dim); 127 128 // In CEED_NOTRANSPOSE mode: 129 // d_u is (P^dim x nc), column-major layout (nc = num_comp) 130 // d_v is (Q^dim x nc x dim), column-major layout (nc = num_comp) 131 // In CEED_TRANSPOSE mode, the sizes of d_u and d_v are switched. 132 133 // E-vector ordering -------------- Q-vector ordering 134 // dim 135 // component component 136 // elem elem 137 // node node 138 139 // --- Define strides for NOTRANSPOSE mode: --- 140 // Input (d_u) is E-vector, output (d_v) is Q-vector 141 142 // Element strides 143 CeedInt u_elem_stride = elem_dofs_size; 144 CeedInt v_elem_stride = elem_qpts_size; 145 // Component strides 146 CeedInt u_comp_stride = num_elem * elem_dofs_size; 147 CeedInt v_comp_stride = num_elem * elem_qpts_size; 148 // Dimension strides 149 CeedInt u_dim_stride = 0; 150 CeedInt v_dim_stride = num_elem * elem_qpts_size * num_comp; 151 if (t_mode == CEED_TRANSPOSE) { 152 // Input (d_u) is Q-vector, output (d_v) is E-vector 153 // Element strides 154 v_elem_stride = elem_dofs_size; 155 u_elem_stride = elem_qpts_size; 156 // Component strides 157 v_comp_stride = num_elem * elem_dofs_size; 158 u_comp_stride = num_elem * elem_qpts_size; 159 // Dimension strides 160 v_dim_stride = 0; 161 u_dim_stride = num_elem * elem_qpts_size * num_comp; 162 } 163 CeedInt num_threads = 1; 164 CeedInt num_t_col = 1; 165 CeedInt shared_mem = 0; 166 CeedInt max_P_Q = CeedIntMax(P, Q); 167 168 switch (dim) { 169 case 1: 170 num_threads = max_P_Q; 171 num_t_col = MAGMA_BASIS_NTCOL(num_threads, MAGMA_MAXTHREADS_1D); 172 shared_mem += sizeof(CeedScalar) * num_t_col * (num_comp * (1 * P + 1 * Q)); 173 shared_mem += sizeof(CeedScalar) * (P * Q); 174 break; 175 case 2: 176 num_threads = max_P_Q; 177 num_t_col = MAGMA_BASIS_NTCOL(num_threads, MAGMA_MAXTHREADS_2D); 178 shared_mem += sizeof(CeedScalar) * 2 * P * Q; // for sTinterp and sTgrad 179 // for reforming rU we need P x P, and for the intermediate output we need P x Q 180 shared_mem += sizeof(CeedScalar) * num_t_col * (P * max_P_Q); 181 break; 182 case 3: 183 num_threads = max_P_Q * max_P_Q; 184 num_t_col = MAGMA_BASIS_NTCOL(num_threads, MAGMA_MAXTHREADS_3D); 185 shared_mem += sizeof(CeedScalar) * 2 * P * Q; // for sTinterp and sTgrad 186 // rU needs P^2 x P, the intermediate outputs need (P^2 x Q + P x Q^2) 187 shared_mem += sizeof(CeedScalar) * num_t_col * CeedIntMax(P * P * P, (P * P * Q) + (P * Q * Q)); 188 break; 189 } 190 CeedInt grid = CeedDivUpInt(num_elem, num_t_col); 191 void *args[] = {&impl->d_interp_1d, &impl->d_grad_1d, &d_u, &u_elem_stride, &u_comp_stride, &u_dim_stride, &d_v, 192 &v_elem_stride, &v_comp_stride, &v_dim_stride, &num_elem}; 193 194 if (t_mode == CEED_TRANSPOSE) { 195 CeedCallBackend(CeedRunKernelDimSharedMagma(ceed, impl->GradTranspose, grid, num_threads, num_t_col, 1, shared_mem, args)); 196 } else { 197 CeedCallBackend(CeedRunKernelDimSharedMagma(ceed, impl->Grad, grid, num_threads, num_t_col, 1, shared_mem, args)); 198 } 199 } break; 200 case CEED_EVAL_WEIGHT: { 201 CeedCheck(t_mode != CEED_TRANSPOSE, ceed, CEED_ERROR_BACKEND, "CEED_EVAL_WEIGHT incompatible with CEED_TRANSPOSE"); 202 CeedInt elem_dofs_size = CeedIntPow(Q, dim); 203 CeedInt num_threads = 1; 204 CeedInt num_t_col = 1; 205 CeedInt shared_mem = 0; 206 207 switch (dim) { 208 case 1: 209 num_threads = Q; 210 num_t_col = MAGMA_BASIS_NTCOL(num_threads, MAGMA_MAXTHREADS_1D); 211 shared_mem += sizeof(CeedScalar) * Q; // for d_q_weight_1d 212 shared_mem += sizeof(CeedScalar) * num_t_col * Q; // for output 213 break; 214 case 2: 215 num_threads = Q; 216 num_t_col = MAGMA_BASIS_NTCOL(num_threads, MAGMA_MAXTHREADS_2D); 217 shared_mem += sizeof(CeedScalar) * Q; // for d_q_weight_1d 218 break; 219 case 3: 220 num_threads = Q * Q; 221 num_t_col = MAGMA_BASIS_NTCOL(num_threads, MAGMA_MAXTHREADS_3D); 222 shared_mem += sizeof(CeedScalar) * Q; // for d_q_weight_1d 223 break; 224 } 225 CeedInt grid = CeedDivUpInt(num_elem, num_t_col); 226 void *args[] = {&impl->d_q_weight_1d, &d_v, &elem_dofs_size, &num_elem}; 227 228 CeedCallBackend(CeedRunKernelDimSharedMagma(ceed, impl->Weight, grid, num_threads, num_t_col, 1, shared_mem, args)); 229 } break; 230 // LCOV_EXCL_START 231 case CEED_EVAL_DIV: 232 return CeedError(ceed, CEED_ERROR_BACKEND, "CEED_EVAL_DIV not supported"); 233 case CEED_EVAL_CURL: 234 return CeedError(ceed, CEED_ERROR_BACKEND, "CEED_EVAL_CURL not supported"); 235 case CEED_EVAL_NONE: 236 return CeedError(ceed, CEED_ERROR_BACKEND, "CEED_EVAL_NONE does not make sense in this context"); 237 // LCOV_EXCL_STOP 238 } 239 240 // Must sync to ensure completeness 241 ceed_magma_queue_sync(data->queue); 242 243 // Restore vectors 244 if (e_mode != CEED_EVAL_WEIGHT) { 245 CeedCallBackend(CeedVectorRestoreArrayRead(u, &d_u)); 246 } 247 CeedCallBackend(CeedVectorRestoreArray(v, &d_v)); 248 return CEED_ERROR_SUCCESS; 249 } 250 251 //------------------------------------------------------------------------------ 252 // Basis apply - non-tensor 253 //------------------------------------------------------------------------------ 254 static int CeedBasisApplyNonTensor_Magma(CeedBasis basis, CeedInt num_elem, CeedTransposeMode t_mode, CeedEvalMode e_mode, CeedVector u, 255 CeedVector v) { 256 Ceed ceed; 257 Ceed_Magma *data; 258 CeedInt num_comp, num_nodes, num_qpts, P, Q, N; 259 const CeedScalar *d_u; 260 CeedScalar *d_v; 261 CeedBasisNonTensor_Magma *impl; 262 263 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 264 CeedCallBackend(CeedGetData(ceed, &data)); 265 CeedCallBackend(CeedBasisGetData(basis, &impl)); 266 CeedCallBackend(CeedBasisGetNumComponents(basis, &num_comp)); 267 CeedCallBackend(CeedBasisGetNumNodes(basis, &num_nodes)); 268 CeedCallBackend(CeedBasisGetNumQuadraturePoints(basis, &num_qpts)); 269 P = num_nodes; 270 Q = num_qpts; 271 N = num_elem * num_comp; 272 273 // Read vectors 274 if (u != CEED_VECTOR_NONE) CeedCallBackend(CeedVectorGetArrayRead(u, CEED_MEM_DEVICE, &d_u)); 275 else CeedCheck(e_mode == CEED_EVAL_WEIGHT, ceed, CEED_ERROR_BACKEND, "An input vector is required for this CeedEvalMode"); 276 CeedCallBackend(CeedVectorGetArrayWrite(v, CEED_MEM_DEVICE, &d_v)); 277 278 // Compile kernels for N as needed 279 CeedInt iN = 0; 280 if (P <= MAGMA_NONTENSOR_CUSTOM_KERNEL_MAX_P && Q <= MAGMA_NONTENSOR_CUSTOM_KERNEL_MAX_Q && (e_mode != CEED_EVAL_WEIGHT || !impl->Weight)) { 281 CeedInt n_array[MAGMA_NONTENSOR_KERNEL_INSTANCES] = {MAGMA_NONTENSOR_KERNEL_N_VALUES}; 282 CeedInt diff = abs(n_array[iN] - N), idiff; 283 284 for (CeedInt in = iN + 1; in < MAGMA_NONTENSOR_KERNEL_INSTANCES; in++) { 285 idiff = abs(n_array[in] - N); 286 if (idiff < diff) { 287 iN = in; 288 diff = idiff; 289 } 290 } 291 292 if (!impl->NB_interp[iN]) { 293 CeedFESpace fe_space; 294 CeedInt q_comp_interp, q_comp_deriv; 295 Ceed ceed_delegate; 296 char *basis_kernel_path, *weight_kernel_path, *basis_kernel_source; 297 magma_int_t arch = magma_getdevice_arch(); 298 299 // Tuning parameters for NB 300 CeedCallBackend(CeedBasisGetFESpace(basis, &fe_space)); 301 CeedCallBackend(CeedBasisGetNumQuadratureComponents(basis, CEED_EVAL_INTERP, &q_comp_interp)); 302 switch (fe_space) { 303 case CEED_FE_SPACE_H1: 304 CeedCallBackend(CeedBasisGetNumQuadratureComponents(basis, CEED_EVAL_GRAD, &q_comp_deriv)); 305 break; 306 case CEED_FE_SPACE_HDIV: 307 CeedCallBackend(CeedBasisGetNumQuadratureComponents(basis, CEED_EVAL_DIV, &q_comp_deriv)); 308 break; 309 case CEED_FE_SPACE_HCURL: 310 CeedCallBackend(CeedBasisGetNumQuadratureComponents(basis, CEED_EVAL_CURL, &q_comp_deriv)); 311 break; 312 } 313 impl->NB_interp[iN] = nontensor_rtc_get_nb(arch, 'n', q_comp_interp, P, Q, n_array[iN]); 314 impl->NB_interp_t[iN] = nontensor_rtc_get_nb(arch, 't', q_comp_interp, P, Q, n_array[iN]); 315 impl->NB_deriv[iN] = nontensor_rtc_get_nb(arch, 'n', q_comp_deriv, P, Q, n_array[iN]); 316 impl->NB_deriv_t[iN] = nontensor_rtc_get_nb(arch, 't', q_comp_deriv, P, Q, n_array[iN]); 317 318 // The RTC compilation code expects a Ceed with the common Ceed_Cuda or Ceed_Hip data 319 CeedCallBackend(CeedGetDelegate(ceed, &ceed_delegate)); 320 321 // Compile kernels 322 CeedCallBackend(CeedGetJitAbsolutePath(ceed, "ceed/jit-source/magma/magma-basis-interp-deriv-nontensor.h", &basis_kernel_path)); 323 CeedDebug256(ceed, CEED_DEBUG_COLOR_SUCCESS, "----- Loading Basis Kernel Source -----\n"); 324 CeedCallBackend(CeedLoadSourceToBuffer(ceed, basis_kernel_path, &basis_kernel_source)); 325 if (!impl->Weight) { 326 CeedCallBackend(CeedGetJitAbsolutePath(ceed, "ceed/jit-source/magma/magma-basis-weight-nontensor.h", &weight_kernel_path)); 327 CeedCallBackend(CeedLoadSourceToInitializedBuffer(ceed, weight_kernel_path, &basis_kernel_source)); 328 } 329 CeedDebug256(ceed, CEED_DEBUG_COLOR_SUCCESS, "----- Loading Basis Kernel Source Complete! -----\n"); 330 CeedCallBackend(CeedCompileMagma(ceed_delegate, basis_kernel_source, &impl->module[iN], 8, "BASIS_Q_COMP_INTERP", q_comp_interp, 331 "BASIS_Q_COMP_DERIV", q_comp_deriv, "BASIS_P", P, "BASIS_Q", Q, "BASIS_NB_INTERP_N", impl->NB_interp[iN], 332 "BASIS_NB_INTERP_T", impl->NB_interp_t[iN], "BASIS_NB_DERIV_N", impl->NB_deriv[iN], "BASIS_NB_DERIV_T", 333 impl->NB_deriv_t[iN])); 334 CeedCallBackend(CeedGetKernelMagma(ceed, impl->module[iN], "magma_interp_nontensor_n", &impl->Interp[iN])); 335 CeedCallBackend(CeedGetKernelMagma(ceed, impl->module[iN], "magma_interp_nontensor_t", &impl->InterpTranspose[iN])); 336 CeedCallBackend(CeedGetKernelMagma(ceed, impl->module[iN], "magma_deriv_nontensor_n", &impl->Deriv[iN])); 337 CeedCallBackend(CeedGetKernelMagma(ceed, impl->module[iN], "magma_deriv_nontensor_t", &impl->DerivTranspose[iN])); 338 if (!impl->Weight) { 339 CeedCallBackend(CeedGetKernelMagma(ceed, impl->module[iN], "magma_weight_nontensor", &impl->Weight)); 340 CeedCallBackend(CeedFree(&weight_kernel_path)); 341 } 342 CeedCallBackend(CeedFree(&basis_kernel_path)); 343 CeedCallBackend(CeedFree(&basis_kernel_source)); 344 } 345 } 346 347 // Apply basis operation 348 if (e_mode != CEED_EVAL_WEIGHT) { 349 const CeedScalar *d_b = NULL; 350 CeedInt q_comp, NB, M, K; 351 CeedMagmaFunction Kernel; 352 353 switch (e_mode) { 354 case CEED_EVAL_INTERP: 355 d_b = impl->d_interp; 356 break; 357 case CEED_EVAL_GRAD: 358 d_b = impl->d_grad; 359 break; 360 case CEED_EVAL_DIV: 361 d_b = impl->d_div; 362 break; 363 case CEED_EVAL_CURL: 364 d_b = impl->d_curl; 365 break; 366 // LCOV_EXCL_START 367 case CEED_EVAL_WEIGHT: 368 return CeedError(ceed, CEED_ERROR_BACKEND, "CEED_EVAL_WEIGHT does not make sense in this context"); 369 case CEED_EVAL_NONE: 370 return CeedError(ceed, CEED_ERROR_BACKEND, "CEED_EVAL_NONE does not make sense in this context"); 371 // LCOV_EXCL_STOP 372 } 373 CeedCallBackend(CeedBasisGetNumQuadratureComponents(basis, e_mode, &q_comp)); 374 M = (t_mode == CEED_TRANSPOSE) ? P : Q, K = (t_mode == CEED_TRANSPOSE) ? Q : P; 375 376 if (P <= MAGMA_NONTENSOR_CUSTOM_KERNEL_MAX_P && Q <= MAGMA_NONTENSOR_CUSTOM_KERNEL_MAX_Q) { 377 if (e_mode == CEED_EVAL_INTERP) { 378 if (t_mode == CEED_TRANSPOSE) { 379 Kernel = impl->InterpTranspose[iN]; 380 NB = impl->NB_interp_t[iN]; 381 } else { 382 Kernel = impl->Interp[iN]; 383 NB = impl->NB_interp[iN]; 384 } 385 } else { 386 if (t_mode == CEED_TRANSPOSE) { 387 Kernel = impl->DerivTranspose[iN]; 388 NB = impl->NB_deriv_t[iN]; 389 } else { 390 Kernel = impl->Deriv[iN]; 391 NB = impl->NB_deriv[iN]; 392 } 393 } 394 CeedInt num_t_col = MAGMA_BASIS_NTCOL(M, MAGMA_MAXTHREADS_1D); 395 CeedInt grid = CeedDivUpInt(N, num_t_col * NB); 396 CeedInt shared_mem_A = P * Q * sizeof(CeedScalar); 397 CeedInt shared_mem_B = num_t_col * K * NB * sizeof(CeedScalar); 398 CeedInt shared_mem = (t_mode != CEED_TRANSPOSE && q_comp > 1) ? (shared_mem_A + shared_mem_B) : CeedIntMax(shared_mem_A, shared_mem_B); 399 void *args[] = {&N, &d_b, &d_u, &d_v}; 400 401 CeedCallBackend(CeedRunKernelDimSharedMagma(ceed, Kernel, grid, M, num_t_col, 1, shared_mem, args)); 402 } else { 403 for (CeedInt d = 0; d < q_comp; d++) { 404 if (t_mode == CEED_TRANSPOSE) { 405 const CeedScalar beta = (d > 0) ? 1.0 : 0.0; 406 magma_gemm_nontensor(MagmaNoTrans, MagmaNoTrans, P, N, Q, 1.0, d_b + d * P * Q, P, d_u + d * N * Q, Q, beta, d_v, P, data->queue); 407 } else { 408 magma_gemm_nontensor(MagmaTrans, MagmaNoTrans, Q, N, P, 1.0, d_b + d * P * Q, P, d_u, P, 0.0, d_v + d * N * Q, Q, data->queue); 409 } 410 } 411 } 412 } else { 413 CeedCheck(t_mode != CEED_TRANSPOSE, ceed, CEED_ERROR_BACKEND, "CEED_EVAL_WEIGHT incompatible with CEED_TRANSPOSE"); 414 CeedInt num_t_col = MAGMA_BASIS_NTCOL(Q, MAGMA_MAXTHREADS_1D); 415 CeedInt grid = CeedDivUpInt(num_elem, num_t_col); 416 CeedInt shared_mem = Q * sizeof(CeedScalar) + num_t_col * Q * sizeof(CeedScalar); 417 void *args[] = {&num_elem, &impl->d_q_weight, &d_v}; 418 419 CeedCallBackend(CeedRunKernelDimSharedMagma(ceed, impl->Weight, grid, Q, num_t_col, 1, shared_mem, args)); 420 } 421 422 // Must sync to ensure completeness 423 ceed_magma_queue_sync(data->queue); 424 425 // Restore vectors 426 if (e_mode != CEED_EVAL_WEIGHT) { 427 CeedCallBackend(CeedVectorRestoreArrayRead(u, &d_u)); 428 } 429 CeedCallBackend(CeedVectorRestoreArray(v, &d_v)); 430 return CEED_ERROR_SUCCESS; 431 } 432 433 //------------------------------------------------------------------------------ 434 // Destroy tensor basis 435 //------------------------------------------------------------------------------ 436 static int CeedBasisDestroy_Magma(CeedBasis basis) { 437 Ceed ceed; 438 CeedBasis_Magma *impl; 439 440 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 441 CeedCallBackend(CeedBasisGetData(basis, &impl)); 442 #ifdef CEED_MAGMA_USE_HIP 443 CeedCallHip(ceed, hipModuleUnload(impl->module)); 444 #else 445 CeedCallCuda(ceed, cuModuleUnload(impl->module)); 446 #endif 447 CeedCallBackend(magma_free(impl->d_interp_1d)); 448 CeedCallBackend(magma_free(impl->d_grad_1d)); 449 CeedCallBackend(magma_free(impl->d_q_weight_1d)); 450 CeedCallBackend(CeedFree(&impl)); 451 return CEED_ERROR_SUCCESS; 452 } 453 454 //------------------------------------------------------------------------------ 455 // Destroy non-tensor basis 456 //------------------------------------------------------------------------------ 457 static int CeedBasisDestroyNonTensor_Magma(CeedBasis basis) { 458 Ceed ceed; 459 CeedBasisNonTensor_Magma *impl; 460 461 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 462 CeedCallBackend(CeedBasisGetData(basis, &impl)); 463 for (CeedInt in = 0; in < MAGMA_NONTENSOR_KERNEL_INSTANCES; in++) { 464 if (impl->module[in]) { 465 #ifdef CEED_MAGMA_USE_HIP 466 CeedCallHip(ceed, hipModuleUnload(impl->module[in])); 467 #else 468 CeedCallCuda(ceed, cuModuleUnload(impl->module[in])); 469 #endif 470 } 471 } 472 CeedCallBackend(magma_free(impl->d_interp)); 473 CeedCallBackend(magma_free(impl->d_grad)); 474 CeedCallBackend(magma_free(impl->d_div)); 475 CeedCallBackend(magma_free(impl->d_curl)); 476 CeedCallBackend(magma_free(impl->d_q_weight)); 477 CeedCallBackend(CeedFree(&impl)); 478 return CEED_ERROR_SUCCESS; 479 } 480 481 //------------------------------------------------------------------------------ 482 // Create tensor 483 //------------------------------------------------------------------------------ 484 int CeedBasisCreateTensorH1_Magma(CeedInt dim, CeedInt P_1d, CeedInt Q_1d, const CeedScalar *interp_1d, const CeedScalar *grad_1d, 485 const CeedScalar *q_ref_1d, const CeedScalar *q_weight_1d, CeedBasis basis) { 486 Ceed ceed, ceed_delegate; 487 Ceed_Magma *data; 488 char *interp_kernel_path, *grad_kernel_path, *weight_kernel_path, *basis_kernel_source; 489 CeedInt num_comp; 490 CeedBasis_Magma *impl; 491 492 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 493 CeedCallBackend(CeedGetData(ceed, &data)); 494 CeedCallBackend(CeedCalloc(1, &impl)); 495 496 // Copy basis data to GPU 497 CeedCallBackend(magma_malloc((void **)&impl->d_q_weight_1d, Q_1d * sizeof(q_weight_1d[0]))); 498 magma_setvector(Q_1d, sizeof(q_weight_1d[0]), q_weight_1d, 1, impl->d_q_weight_1d, 1, data->queue); 499 CeedCallBackend(magma_malloc((void **)&impl->d_interp_1d, Q_1d * P_1d * sizeof(interp_1d[0]))); 500 magma_setvector(Q_1d * P_1d, sizeof(interp_1d[0]), interp_1d, 1, impl->d_interp_1d, 1, data->queue); 501 CeedCallBackend(magma_malloc((void **)&impl->d_grad_1d, Q_1d * P_1d * sizeof(grad_1d[0]))); 502 magma_setvector(Q_1d * P_1d, sizeof(grad_1d[0]), grad_1d, 1, impl->d_grad_1d, 1, data->queue); 503 504 // The RTC compilation code expects a Ceed with the common Ceed_Cuda or Ceed_Hip data 505 CeedCallBackend(CeedGetDelegate(ceed, &ceed_delegate)); 506 507 // Compile kernels 508 CeedCallBackend(CeedBasisGetNumComponents(basis, &num_comp)); 509 { 510 char *interp_kernel_name_base = "ceed/jit-source/magma/magma-basis-interp"; 511 CeedInt interp_kernel_name_len = strlen(interp_kernel_name_base) + 6; 512 char interp_kernel_name[interp_kernel_name_len]; 513 514 snprintf(interp_kernel_name, interp_kernel_name_len, "%s-%" CeedInt_FMT "d.h", interp_kernel_name_base, dim); 515 CeedCallBackend(CeedGetJitAbsolutePath(ceed, interp_kernel_name, &interp_kernel_path)); 516 } 517 CeedDebug256(ceed, CEED_DEBUG_COLOR_SUCCESS, "----- Loading Basis Kernel Source -----\n"); 518 CeedCallBackend(CeedLoadSourceToBuffer(ceed, interp_kernel_path, &basis_kernel_source)); 519 { 520 char *grad_kernel_name_base = "ceed/jit-source/magma/magma-basis-grad"; 521 CeedInt grad_kernel_name_len = strlen(grad_kernel_name_base) + 6; 522 char grad_kernel_name[grad_kernel_name_len]; 523 524 snprintf(grad_kernel_name, grad_kernel_name_len, "%s-%" CeedInt_FMT "d.h", grad_kernel_name_base, dim); 525 CeedCallBackend(CeedGetJitAbsolutePath(ceed, grad_kernel_name, &grad_kernel_path)); 526 } 527 CeedCallBackend(CeedLoadSourceToInitializedBuffer(ceed, grad_kernel_path, &basis_kernel_source)); 528 { 529 char *weight_kernel_name_base = "ceed/jit-source/magma/magma-basis-weight"; 530 CeedInt weight_kernel_name_len = strlen(weight_kernel_name_base) + 6; 531 char weight_kernel_name[weight_kernel_name_len]; 532 533 snprintf(weight_kernel_name, weight_kernel_name_len, "%s-%" CeedInt_FMT "d.h", weight_kernel_name_base, dim); 534 CeedCallBackend(CeedGetJitAbsolutePath(ceed, weight_kernel_name, &weight_kernel_path)); 535 } 536 CeedCallBackend(CeedLoadSourceToInitializedBuffer(ceed, weight_kernel_path, &basis_kernel_source)); 537 CeedDebug256(ceed, CEED_DEBUG_COLOR_SUCCESS, "----- Loading Basis Kernel Source Complete! -----\n"); 538 CeedCallBackend(CeedCompileMagma(ceed_delegate, basis_kernel_source, &impl->module, 5, "BASIS_DIM", dim, "BASIS_NUM_COMP", num_comp, "BASIS_P", 539 P_1d, "BASIS_Q", Q_1d, "BASIS_MAX_P_Q", CeedIntMax(P_1d, Q_1d))); 540 switch (dim) { 541 case 1: 542 CeedCallBackend(CeedGetKernelMagma(ceed, impl->module, "magma_interpn_1d_kernel", &impl->Interp)); 543 CeedCallBackend(CeedGetKernelMagma(ceed, impl->module, "magma_interpt_1d_kernel", &impl->InterpTranspose)); 544 CeedCallBackend(CeedGetKernelMagma(ceed, impl->module, "magma_gradn_1d_kernel", &impl->Grad)); 545 CeedCallBackend(CeedGetKernelMagma(ceed, impl->module, "magma_gradt_1d_kernel", &impl->GradTranspose)); 546 CeedCallBackend(CeedGetKernelMagma(ceed, impl->module, "magma_weight_1d_kernel", &impl->Weight)); 547 break; 548 case 2: 549 CeedCallBackend(CeedGetKernelMagma(ceed, impl->module, "magma_interpn_2d_kernel", &impl->Interp)); 550 CeedCallBackend(CeedGetKernelMagma(ceed, impl->module, "magma_interpt_2d_kernel", &impl->InterpTranspose)); 551 CeedCallBackend(CeedGetKernelMagma(ceed, impl->module, "magma_gradn_2d_kernel", &impl->Grad)); 552 CeedCallBackend(CeedGetKernelMagma(ceed, impl->module, "magma_gradt_2d_kernel", &impl->GradTranspose)); 553 CeedCallBackend(CeedGetKernelMagma(ceed, impl->module, "magma_weight_2d_kernel", &impl->Weight)); 554 break; 555 case 3: 556 CeedCallBackend(CeedGetKernelMagma(ceed, impl->module, "magma_interpn_3d_kernel", &impl->Interp)); 557 CeedCallBackend(CeedGetKernelMagma(ceed, impl->module, "magma_interpt_3d_kernel", &impl->InterpTranspose)); 558 CeedCallBackend(CeedGetKernelMagma(ceed, impl->module, "magma_gradn_3d_kernel", &impl->Grad)); 559 CeedCallBackend(CeedGetKernelMagma(ceed, impl->module, "magma_gradt_3d_kernel", &impl->GradTranspose)); 560 CeedCallBackend(CeedGetKernelMagma(ceed, impl->module, "magma_weight_3d_kernel", &impl->Weight)); 561 break; 562 } 563 CeedCallBackend(CeedFree(&interp_kernel_path)); 564 CeedCallBackend(CeedFree(&grad_kernel_path)); 565 CeedCallBackend(CeedFree(&weight_kernel_path)); 566 CeedCallBackend(CeedFree(&basis_kernel_source)); 567 568 CeedCallBackend(CeedBasisSetData(basis, impl)); 569 570 CeedCallBackend(CeedSetBackendFunction(ceed, "Basis", basis, "Apply", CeedBasisApply_Magma)); 571 CeedCallBackend(CeedSetBackendFunction(ceed, "Basis", basis, "Destroy", CeedBasisDestroy_Magma)); 572 return CEED_ERROR_SUCCESS; 573 } 574 575 //------------------------------------------------------------------------------ 576 // Create non-tensor H^1 577 //------------------------------------------------------------------------------ 578 int CeedBasisCreateH1_Magma(CeedElemTopology topo, CeedInt dim, CeedInt num_nodes, CeedInt num_qpts, const CeedScalar *interp, const CeedScalar *grad, 579 const CeedScalar *q_ref, const CeedScalar *q_weight, CeedBasis basis) { 580 Ceed ceed; 581 Ceed_Magma *data; 582 CeedBasisNonTensor_Magma *impl; 583 584 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 585 CeedCallBackend(CeedGetData(ceed, &data)); 586 CeedCallBackend(CeedCalloc(1, &impl)); 587 588 // Copy basis data to GPU 589 CeedCallBackend(magma_malloc((void **)&impl->d_q_weight, num_qpts * sizeof(q_weight[0]))); 590 magma_setvector(num_qpts, sizeof(q_weight[0]), q_weight, 1, impl->d_q_weight, 1, data->queue); 591 if (interp) { 592 CeedInt q_comp_interp; 593 594 CeedCallBackend(CeedBasisGetNumQuadratureComponents(basis, CEED_EVAL_INTERP, &q_comp_interp)); 595 CeedCallBackend(magma_malloc((void **)&impl->d_interp, num_qpts * num_nodes * q_comp_interp * sizeof(interp[0]))); 596 magma_setvector(num_qpts * num_nodes * q_comp_interp, sizeof(interp[0]), interp, 1, impl->d_interp, 1, data->queue); 597 } 598 if (grad) { 599 CeedInt q_comp_grad; 600 601 CeedCallBackend(CeedBasisGetNumQuadratureComponents(basis, CEED_EVAL_GRAD, &q_comp_grad)); 602 CeedCallBackend(magma_malloc((void **)&impl->d_grad, num_qpts * num_nodes * q_comp_grad * sizeof(grad[0]))); 603 magma_setvector(num_qpts * num_nodes * q_comp_grad, sizeof(grad[0]), grad, 1, impl->d_grad, 1, data->queue); 604 } 605 606 // Compile the weight kernel if it won't be compiled later on 607 if (num_nodes > MAGMA_NONTENSOR_CUSTOM_KERNEL_MAX_P || num_qpts > MAGMA_NONTENSOR_CUSTOM_KERNEL_MAX_Q) { 608 Ceed ceed_delegate; 609 char *weight_kernel_path, *basis_kernel_source; 610 611 // The RTC compilation code expects a Ceed with the common Ceed_Cuda or Ceed_Hip data 612 CeedCallBackend(CeedGetDelegate(ceed, &ceed_delegate)); 613 614 // Compile weight kernel (the remainder of kernel compilation happens at first call to CeedBasisApply) 615 CeedCallBackend(CeedGetJitAbsolutePath(ceed, "ceed/jit-source/magma/magma-basis-weight-nontensor.h", &weight_kernel_path)); 616 CeedDebug256(ceed, CEED_DEBUG_COLOR_SUCCESS, "----- Loading Basis Kernel Source -----\n"); 617 CeedCallBackend(CeedLoadSourceToBuffer(ceed, weight_kernel_path, &basis_kernel_source)); 618 CeedDebug256(ceed, CEED_DEBUG_COLOR_SUCCESS, "----- Loading Basis Kernel Source Complete! -----\n"); 619 CeedCallBackend(CeedCompileMagma(ceed_delegate, basis_kernel_source, &impl->module[0], 1, "BASIS_Q", num_qpts)); 620 CeedCallBackend(CeedGetKernelMagma(ceed, impl->module[0], "magma_weight_nontensor", &impl->Weight)); 621 CeedCallBackend(CeedFree(&weight_kernel_path)); 622 CeedCallBackend(CeedFree(&basis_kernel_source)); 623 } 624 625 CeedCallBackend(CeedBasisSetData(basis, impl)); 626 627 // Register backend functions 628 CeedCallBackend(CeedSetBackendFunction(ceed, "Basis", basis, "Apply", CeedBasisApplyNonTensor_Magma)); 629 CeedCallBackend(CeedSetBackendFunction(ceed, "Basis", basis, "Destroy", CeedBasisDestroyNonTensor_Magma)); 630 return CEED_ERROR_SUCCESS; 631 } 632 633 //------------------------------------------------------------------------------ 634 // Create non-tensor H(div) 635 //------------------------------------------------------------------------------ 636 int CeedBasisCreateHdiv_Magma(CeedElemTopology topo, CeedInt dim, CeedInt num_nodes, CeedInt num_qpts, const CeedScalar *interp, 637 const CeedScalar *div, const CeedScalar *q_ref, const CeedScalar *q_weight, CeedBasis basis) { 638 Ceed ceed; 639 Ceed_Magma *data; 640 CeedBasisNonTensor_Magma *impl; 641 642 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 643 CeedCallBackend(CeedGetData(ceed, &data)); 644 CeedCallBackend(CeedCalloc(1, &impl)); 645 646 // Copy basis data to GPU 647 CeedCallBackend(magma_malloc((void **)&impl->d_q_weight, num_qpts * sizeof(q_weight[0]))); 648 magma_setvector(num_qpts, sizeof(q_weight[0]), q_weight, 1, impl->d_q_weight, 1, data->queue); 649 if (interp) { 650 CeedInt q_comp_interp; 651 652 CeedCallBackend(CeedBasisGetNumQuadratureComponents(basis, CEED_EVAL_INTERP, &q_comp_interp)); 653 CeedCallBackend(magma_malloc((void **)&impl->d_interp, num_qpts * num_nodes * q_comp_interp * sizeof(interp[0]))); 654 magma_setvector(num_qpts * num_nodes * q_comp_interp, sizeof(interp[0]), interp, 1, impl->d_interp, 1, data->queue); 655 } 656 if (div) { 657 CeedInt q_comp_div; 658 659 CeedCallBackend(CeedBasisGetNumQuadratureComponents(basis, CEED_EVAL_DIV, &q_comp_div)); 660 CeedCallBackend(magma_malloc((void **)&impl->d_div, num_qpts * num_nodes * q_comp_div * sizeof(div[0]))); 661 magma_setvector(num_qpts * num_nodes * q_comp_div, sizeof(div[0]), div, 1, impl->d_div, 1, data->queue); 662 } 663 664 // Compile the weight kernel if it won't be compiled later on 665 if (num_nodes > MAGMA_NONTENSOR_CUSTOM_KERNEL_MAX_P || num_qpts > MAGMA_NONTENSOR_CUSTOM_KERNEL_MAX_Q) { 666 Ceed ceed_delegate; 667 char *weight_kernel_path, *basis_kernel_source; 668 669 // The RTC compilation code expects a Ceed with the common Ceed_Cuda or Ceed_Hip data 670 CeedCallBackend(CeedGetDelegate(ceed, &ceed_delegate)); 671 672 // Compile weight kernel (the remainder of kernel compilation happens at first call to CeedBasisApply) 673 CeedCallBackend(CeedGetJitAbsolutePath(ceed, "ceed/jit-source/magma/magma-basis-weight-nontensor.h", &weight_kernel_path)); 674 CeedDebug256(ceed, CEED_DEBUG_COLOR_SUCCESS, "----- Loading Basis Kernel Source -----\n"); 675 CeedCallBackend(CeedLoadSourceToBuffer(ceed, weight_kernel_path, &basis_kernel_source)); 676 CeedDebug256(ceed, CEED_DEBUG_COLOR_SUCCESS, "----- Loading Basis Kernel Source Complete! -----\n"); 677 CeedCallBackend(CeedCompileMagma(ceed_delegate, basis_kernel_source, &impl->module[0], 1, "BASIS_Q", num_qpts)); 678 CeedCallBackend(CeedGetKernelMagma(ceed, impl->module[0], "magma_weight_nontensor", &impl->Weight)); 679 CeedCallBackend(CeedFree(&weight_kernel_path)); 680 CeedCallBackend(CeedFree(&basis_kernel_source)); 681 } 682 683 CeedCallBackend(CeedBasisSetData(basis, impl)); 684 685 // Register backend functions 686 CeedCallBackend(CeedSetBackendFunction(ceed, "Basis", basis, "Apply", CeedBasisApplyNonTensor_Magma)); 687 CeedCallBackend(CeedSetBackendFunction(ceed, "Basis", basis, "Destroy", CeedBasisDestroyNonTensor_Magma)); 688 return CEED_ERROR_SUCCESS; 689 } 690 691 //------------------------------------------------------------------------------ 692 // Create non-tensor H(curl) 693 //------------------------------------------------------------------------------ 694 int CeedBasisCreateHcurl_Magma(CeedElemTopology topo, CeedInt dim, CeedInt num_nodes, CeedInt num_qpts, const CeedScalar *interp, 695 const CeedScalar *curl, const CeedScalar *q_ref, const CeedScalar *q_weight, CeedBasis basis) { 696 Ceed ceed; 697 Ceed_Magma *data; 698 CeedBasisNonTensor_Magma *impl; 699 700 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 701 CeedCallBackend(CeedGetData(ceed, &data)); 702 CeedCallBackend(CeedCalloc(1, &impl)); 703 704 // Copy basis data to GPU 705 CeedCallBackend(magma_malloc((void **)&impl->d_q_weight, num_qpts * sizeof(q_weight[0]))); 706 magma_setvector(num_qpts, sizeof(q_weight[0]), q_weight, 1, impl->d_q_weight, 1, data->queue); 707 if (interp) { 708 CeedInt q_comp_interp; 709 710 CeedCallBackend(CeedBasisGetNumQuadratureComponents(basis, CEED_EVAL_INTERP, &q_comp_interp)); 711 CeedCallBackend(magma_malloc((void **)&impl->d_interp, num_qpts * num_nodes * q_comp_interp * sizeof(interp[0]))); 712 magma_setvector(num_qpts * num_nodes * q_comp_interp, sizeof(interp[0]), interp, 1, impl->d_interp, 1, data->queue); 713 } 714 if (curl) { 715 CeedInt q_comp_curl; 716 717 CeedCallBackend(CeedBasisGetNumQuadratureComponents(basis, CEED_EVAL_CURL, &q_comp_curl)); 718 CeedCallBackend(magma_malloc((void **)&impl->d_curl, num_qpts * num_nodes * q_comp_curl * sizeof(curl[0]))); 719 magma_setvector(num_qpts * num_nodes * q_comp_curl, sizeof(curl[0]), curl, 1, impl->d_curl, 1, data->queue); 720 } 721 722 // Compile the weight kernel if it won't be compiled later on 723 if (num_nodes > MAGMA_NONTENSOR_CUSTOM_KERNEL_MAX_P || num_qpts > MAGMA_NONTENSOR_CUSTOM_KERNEL_MAX_Q) { 724 Ceed ceed_delegate; 725 char *weight_kernel_path, *basis_kernel_source; 726 727 // The RTC compilation code expects a Ceed with the common Ceed_Cuda or Ceed_Hip data 728 CeedCallBackend(CeedGetDelegate(ceed, &ceed_delegate)); 729 730 // Compile weight kernel (the remainder of kernel compilation happens at first call to CeedBasisApply) 731 CeedCallBackend(CeedGetJitAbsolutePath(ceed, "ceed/jit-source/magma/magma-basis-weight-nontensor.h", &weight_kernel_path)); 732 CeedDebug256(ceed, CEED_DEBUG_COLOR_SUCCESS, "----- Loading Basis Kernel Source -----\n"); 733 CeedCallBackend(CeedLoadSourceToBuffer(ceed, weight_kernel_path, &basis_kernel_source)); 734 CeedDebug256(ceed, CEED_DEBUG_COLOR_SUCCESS, "----- Loading Basis Kernel Source Complete! -----\n"); 735 CeedCallBackend(CeedCompileMagma(ceed_delegate, basis_kernel_source, &impl->module[0], 1, "BASIS_Q", num_qpts)); 736 CeedCallBackend(CeedGetKernelMagma(ceed, impl->module[0], "magma_weight_nontensor", &impl->Weight)); 737 CeedCallBackend(CeedFree(&weight_kernel_path)); 738 CeedCallBackend(CeedFree(&basis_kernel_source)); 739 } 740 741 CeedCallBackend(CeedBasisSetData(basis, impl)); 742 743 // Register backend functions 744 CeedCallBackend(CeedSetBackendFunction(ceed, "Basis", basis, "Apply", CeedBasisApplyNonTensor_Magma)); 745 CeedCallBackend(CeedSetBackendFunction(ceed, "Basis", basis, "Destroy", CeedBasisDestroyNonTensor_Magma)); 746 return CEED_ERROR_SUCCESS; 747 } 748 749 //------------------------------------------------------------------------------ 750