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 #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 CeedCheck(impl->d_q_weight_1d, ceed, CEED_ERROR_BACKEND, "%s not supported; q_weight_1d not set", CeedEvalModes[e_mode]); 203 CeedInt elem_dofs_size = CeedIntPow(Q, dim); 204 CeedInt num_threads = 1; 205 CeedInt num_t_col = 1; 206 CeedInt shared_mem = 0; 207 208 switch (dim) { 209 case 1: 210 num_threads = Q; 211 num_t_col = MAGMA_BASIS_NTCOL(num_threads, MAGMA_MAXTHREADS_1D); 212 shared_mem += sizeof(CeedScalar) * Q; // for d_q_weight_1d 213 shared_mem += sizeof(CeedScalar) * num_t_col * Q; // for output 214 break; 215 case 2: 216 num_threads = Q; 217 num_t_col = MAGMA_BASIS_NTCOL(num_threads, MAGMA_MAXTHREADS_2D); 218 shared_mem += sizeof(CeedScalar) * Q; // for d_q_weight_1d 219 break; 220 case 3: 221 num_threads = Q * Q; 222 num_t_col = MAGMA_BASIS_NTCOL(num_threads, MAGMA_MAXTHREADS_3D); 223 shared_mem += sizeof(CeedScalar) * Q; // for d_q_weight_1d 224 break; 225 } 226 CeedInt grid = CeedDivUpInt(num_elem, num_t_col); 227 void *args[] = {&impl->d_q_weight_1d, &d_v, &elem_dofs_size, &num_elem}; 228 229 CeedCallBackend(CeedRunKernelDimSharedMagma(ceed, impl->Weight, grid, num_threads, num_t_col, 1, shared_mem, args)); 230 } break; 231 // LCOV_EXCL_START 232 case CEED_EVAL_DIV: 233 case CEED_EVAL_CURL: 234 return CeedError(ceed, CEED_ERROR_BACKEND, "%s not supported", CeedEvalModes[e_mode]); 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_source; 297 const char *basis_kernel_path, *weight_kernel_path; 298 char **file_paths = NULL; 299 CeedInt num_file_paths = 0; 300 magma_int_t arch = magma_getdevice_arch(); 301 302 // Tuning parameters for NB 303 CeedCallBackend(CeedBasisGetFESpace(basis, &fe_space)); 304 CeedCallBackend(CeedBasisGetNumQuadratureComponents(basis, CEED_EVAL_INTERP, &q_comp_interp)); 305 switch (fe_space) { 306 case CEED_FE_SPACE_H1: 307 CeedCallBackend(CeedBasisGetNumQuadratureComponents(basis, CEED_EVAL_GRAD, &q_comp_deriv)); 308 break; 309 case CEED_FE_SPACE_HDIV: 310 CeedCallBackend(CeedBasisGetNumQuadratureComponents(basis, CEED_EVAL_DIV, &q_comp_deriv)); 311 break; 312 case CEED_FE_SPACE_HCURL: 313 CeedCallBackend(CeedBasisGetNumQuadratureComponents(basis, CEED_EVAL_CURL, &q_comp_deriv)); 314 break; 315 } 316 impl->NB_interp[iN] = nontensor_rtc_get_nb(arch, 'n', q_comp_interp, P, Q, n_array[iN]); 317 impl->NB_interp_t[iN] = nontensor_rtc_get_nb(arch, 't', q_comp_interp, P, Q, n_array[iN]); 318 impl->NB_deriv[iN] = nontensor_rtc_get_nb(arch, 'n', q_comp_deriv, P, Q, n_array[iN]); 319 impl->NB_deriv_t[iN] = nontensor_rtc_get_nb(arch, 't', q_comp_deriv, P, Q, n_array[iN]); 320 321 // The RTC compilation code expects a Ceed with the common Ceed_Cuda or Ceed_Hip data 322 CeedCallBackend(CeedGetDelegate(ceed, &ceed_delegate)); 323 324 // Compile kernels 325 CeedCallBackend(CeedGetJitAbsolutePath(ceed, "ceed/jit-source/magma/magma-basis-interp-deriv-nontensor.h", &basis_kernel_path)); 326 CeedDebug256(ceed, CEED_DEBUG_COLOR_SUCCESS, "----- Loading Basis Kernel Source -----\n"); 327 CeedCallBackend(CeedLoadSourceAndInitializeBuffer(ceed, basis_kernel_path, &num_file_paths, &file_paths, &basis_kernel_source)); 328 if (!impl->Weight) { 329 CeedCallBackend(CeedGetJitAbsolutePath(ceed, "ceed/jit-source/magma/magma-basis-weight-nontensor.h", &weight_kernel_path)); 330 CeedCallBackend(CeedLoadSourceToInitializedBuffer(ceed, weight_kernel_path, &num_file_paths, &file_paths, &basis_kernel_source)); 331 } 332 CeedDebug256(ceed, CEED_DEBUG_COLOR_SUCCESS, "----- Loading Basis Kernel Source Complete! -----\n"); 333 CeedCallBackend(CeedCompileMagma(ceed_delegate, basis_kernel_source, &impl->module[iN], 8, "BASIS_Q_COMP_INTERP", q_comp_interp, 334 "BASIS_Q_COMP_DERIV", q_comp_deriv, "BASIS_P", P, "BASIS_Q", Q, "BASIS_NB_INTERP_N", impl->NB_interp[iN], 335 "BASIS_NB_INTERP_T", impl->NB_interp_t[iN], "BASIS_NB_DERIV_N", impl->NB_deriv[iN], "BASIS_NB_DERIV_T", 336 impl->NB_deriv_t[iN])); 337 CeedCallBackend(CeedGetKernelMagma(ceed, impl->module[iN], "magma_interp_nontensor_n", &impl->Interp[iN])); 338 CeedCallBackend(CeedGetKernelMagma(ceed, impl->module[iN], "magma_interp_nontensor_t", &impl->InterpTranspose[iN])); 339 CeedCallBackend(CeedGetKernelMagma(ceed, impl->module[iN], "magma_deriv_nontensor_n", &impl->Deriv[iN])); 340 CeedCallBackend(CeedGetKernelMagma(ceed, impl->module[iN], "magma_deriv_nontensor_t", &impl->DerivTranspose[iN])); 341 if (!impl->Weight) { 342 CeedCallBackend(CeedGetKernelMagma(ceed, impl->module[iN], "magma_weight_nontensor", &impl->Weight)); 343 CeedCallBackend(CeedFree(&weight_kernel_path)); 344 } 345 CeedCallBackend(CeedFree(&basis_kernel_path)); 346 CeedCallBackend(CeedFree(&basis_kernel_source)); 347 for (CeedInt i = 0; i < num_file_paths; i++) CeedCall(CeedFree(&file_paths[i])); 348 CeedCall(CeedFree(&file_paths)); 349 } 350 } 351 352 // Apply basis operation 353 if (e_mode != CEED_EVAL_WEIGHT) { 354 const CeedScalar *d_b = NULL; 355 CeedInt q_comp, NB, M, K; 356 CeedMagmaFunction Kernel; 357 358 switch (e_mode) { 359 case CEED_EVAL_INTERP: 360 d_b = impl->d_interp; 361 break; 362 case CEED_EVAL_GRAD: 363 d_b = impl->d_grad; 364 break; 365 case CEED_EVAL_DIV: 366 d_b = impl->d_div; 367 break; 368 case CEED_EVAL_CURL: 369 d_b = impl->d_curl; 370 break; 371 // LCOV_EXCL_START 372 case CEED_EVAL_WEIGHT: 373 case CEED_EVAL_NONE: 374 return CeedError(ceed, CEED_ERROR_BACKEND, "%s does not make sense in this context", CeedEvalModes[e_mode]); 375 // LCOV_EXCL_STOP 376 } 377 CeedCallBackend(CeedBasisGetNumQuadratureComponents(basis, e_mode, &q_comp)); 378 M = (t_mode == CEED_TRANSPOSE) ? P : Q, K = (t_mode == CEED_TRANSPOSE) ? Q : P; 379 380 if (P <= MAGMA_NONTENSOR_CUSTOM_KERNEL_MAX_P && Q <= MAGMA_NONTENSOR_CUSTOM_KERNEL_MAX_Q) { 381 if (e_mode == CEED_EVAL_INTERP) { 382 if (t_mode == CEED_TRANSPOSE) { 383 Kernel = impl->InterpTranspose[iN]; 384 NB = impl->NB_interp_t[iN]; 385 } else { 386 Kernel = impl->Interp[iN]; 387 NB = impl->NB_interp[iN]; 388 } 389 } else { 390 if (t_mode == CEED_TRANSPOSE) { 391 Kernel = impl->DerivTranspose[iN]; 392 NB = impl->NB_deriv_t[iN]; 393 } else { 394 Kernel = impl->Deriv[iN]; 395 NB = impl->NB_deriv[iN]; 396 } 397 } 398 CeedInt num_t_col = MAGMA_BASIS_NTCOL(M, MAGMA_MAXTHREADS_1D); 399 CeedInt grid = CeedDivUpInt(N, num_t_col * NB); 400 CeedInt shared_mem_A = P * Q * sizeof(CeedScalar); 401 CeedInt shared_mem_B = num_t_col * K * NB * sizeof(CeedScalar); 402 CeedInt shared_mem = (t_mode != CEED_TRANSPOSE && q_comp > 1) ? (shared_mem_A + shared_mem_B) : CeedIntMax(shared_mem_A, shared_mem_B); 403 void *args[] = {&N, &d_b, &d_u, &d_v}; 404 405 CeedCallBackend(CeedRunKernelDimSharedMagma(ceed, Kernel, grid, M, num_t_col, 1, shared_mem, args)); 406 } else { 407 for (CeedInt d = 0; d < q_comp; d++) { 408 if (t_mode == CEED_TRANSPOSE) { 409 const CeedScalar beta = (d > 0) ? 1.0 : 0.0; 410 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); 411 } else { 412 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); 413 } 414 } 415 } 416 } else { 417 CeedCheck(t_mode != CEED_TRANSPOSE, ceed, CEED_ERROR_BACKEND, "CEED_EVAL_WEIGHT incompatible with CEED_TRANSPOSE"); 418 CeedCheck(impl->d_q_weight, ceed, CEED_ERROR_BACKEND, "%s not supported; q_weight not set", CeedEvalModes[e_mode]); 419 CeedInt num_t_col = MAGMA_BASIS_NTCOL(Q, MAGMA_MAXTHREADS_1D); 420 CeedInt grid = CeedDivUpInt(num_elem, num_t_col); 421 CeedInt shared_mem = Q * sizeof(CeedScalar) + num_t_col * Q * sizeof(CeedScalar); 422 void *args[] = {&num_elem, &impl->d_q_weight, &d_v}; 423 424 CeedCallBackend(CeedRunKernelDimSharedMagma(ceed, impl->Weight, grid, Q, num_t_col, 1, shared_mem, args)); 425 } 426 427 // Must sync to ensure completeness 428 ceed_magma_queue_sync(data->queue); 429 430 // Restore vectors 431 if (e_mode != CEED_EVAL_WEIGHT) { 432 CeedCallBackend(CeedVectorRestoreArrayRead(u, &d_u)); 433 } 434 CeedCallBackend(CeedVectorRestoreArray(v, &d_v)); 435 return CEED_ERROR_SUCCESS; 436 } 437 438 //------------------------------------------------------------------------------ 439 // Destroy tensor basis 440 //------------------------------------------------------------------------------ 441 static int CeedBasisDestroy_Magma(CeedBasis basis) { 442 Ceed ceed; 443 CeedBasis_Magma *impl; 444 445 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 446 CeedCallBackend(CeedBasisGetData(basis, &impl)); 447 #ifdef CEED_MAGMA_USE_HIP 448 CeedCallHip(ceed, hipModuleUnload(impl->module)); 449 #else 450 CeedCallCuda(ceed, cuModuleUnload(impl->module)); 451 #endif 452 CeedCallBackend(magma_free(impl->d_interp_1d)); 453 CeedCallBackend(magma_free(impl->d_grad_1d)); 454 if (impl->d_q_weight_1d) CeedCallBackend(magma_free(impl->d_q_weight_1d)); 455 CeedCallBackend(CeedFree(&impl)); 456 return CEED_ERROR_SUCCESS; 457 } 458 459 //------------------------------------------------------------------------------ 460 // Destroy non-tensor basis 461 //------------------------------------------------------------------------------ 462 static int CeedBasisDestroyNonTensor_Magma(CeedBasis basis) { 463 Ceed ceed; 464 CeedBasisNonTensor_Magma *impl; 465 466 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 467 CeedCallBackend(CeedBasisGetData(basis, &impl)); 468 for (CeedInt in = 0; in < MAGMA_NONTENSOR_KERNEL_INSTANCES; in++) { 469 if (impl->module[in]) { 470 #ifdef CEED_MAGMA_USE_HIP 471 CeedCallHip(ceed, hipModuleUnload(impl->module[in])); 472 #else 473 CeedCallCuda(ceed, cuModuleUnload(impl->module[in])); 474 #endif 475 } 476 } 477 CeedCallBackend(magma_free(impl->d_interp)); 478 CeedCallBackend(magma_free(impl->d_grad)); 479 CeedCallBackend(magma_free(impl->d_div)); 480 CeedCallBackend(magma_free(impl->d_curl)); 481 if (impl->d_q_weight) CeedCallBackend(magma_free(impl->d_q_weight)); 482 CeedCallBackend(CeedFree(&impl)); 483 return CEED_ERROR_SUCCESS; 484 } 485 486 //------------------------------------------------------------------------------ 487 // Create tensor 488 //------------------------------------------------------------------------------ 489 int CeedBasisCreateTensorH1_Magma(CeedInt dim, CeedInt P_1d, CeedInt Q_1d, const CeedScalar *interp_1d, const CeedScalar *grad_1d, 490 const CeedScalar *q_ref_1d, const CeedScalar *q_weight_1d, CeedBasis basis) { 491 Ceed ceed, ceed_delegate; 492 Ceed_Magma *data; 493 char *basis_kernel_source; 494 const char *interp_kernel_path, *grad_kernel_path, *weight_kernel_path; 495 char **file_paths = NULL; 496 CeedInt num_file_paths = 0; 497 CeedInt num_comp; 498 CeedBasis_Magma *impl; 499 500 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 501 CeedCallBackend(CeedGetData(ceed, &data)); 502 CeedCallBackend(CeedCalloc(1, &impl)); 503 504 // Copy basis data to GPU 505 if (q_weight_1d) { 506 CeedCallBackend(magma_malloc((void **)&impl->d_q_weight_1d, Q_1d * sizeof(q_weight_1d[0]))); 507 magma_setvector(Q_1d, sizeof(q_weight_1d[0]), q_weight_1d, 1, impl->d_q_weight_1d, 1, data->queue); 508 } 509 CeedCallBackend(magma_malloc((void **)&impl->d_interp_1d, Q_1d * P_1d * sizeof(interp_1d[0]))); 510 magma_setvector(Q_1d * P_1d, sizeof(interp_1d[0]), interp_1d, 1, impl->d_interp_1d, 1, data->queue); 511 CeedCallBackend(magma_malloc((void **)&impl->d_grad_1d, Q_1d * P_1d * sizeof(grad_1d[0]))); 512 magma_setvector(Q_1d * P_1d, sizeof(grad_1d[0]), grad_1d, 1, impl->d_grad_1d, 1, data->queue); 513 514 // The RTC compilation code expects a Ceed with the common Ceed_Cuda or Ceed_Hip data 515 CeedCallBackend(CeedGetDelegate(ceed, &ceed_delegate)); 516 517 // Compile kernels 518 CeedCallBackend(CeedBasisGetNumComponents(basis, &num_comp)); 519 { 520 char *interp_kernel_name_base = "ceed/jit-source/magma/magma-basis-interp"; 521 CeedInt interp_kernel_name_len = strlen(interp_kernel_name_base) + 6; 522 char interp_kernel_name[interp_kernel_name_len]; 523 524 snprintf(interp_kernel_name, interp_kernel_name_len, "%s-%" CeedInt_FMT "d.h", interp_kernel_name_base, dim); 525 CeedCallBackend(CeedGetJitAbsolutePath(ceed, interp_kernel_name, &interp_kernel_path)); 526 } 527 CeedDebug256(ceed, CEED_DEBUG_COLOR_SUCCESS, "----- Loading Basis Kernel Source -----\n"); 528 CeedCallBackend(CeedLoadSourceAndInitializeBuffer(ceed, interp_kernel_path, &num_file_paths, &file_paths, &basis_kernel_source)); 529 { 530 char *grad_kernel_name_base = "ceed/jit-source/magma/magma-basis-grad"; 531 CeedInt grad_kernel_name_len = strlen(grad_kernel_name_base) + 6; 532 char grad_kernel_name[grad_kernel_name_len]; 533 534 snprintf(grad_kernel_name, grad_kernel_name_len, "%s-%" CeedInt_FMT "d.h", grad_kernel_name_base, dim); 535 CeedCallBackend(CeedGetJitAbsolutePath(ceed, grad_kernel_name, &grad_kernel_path)); 536 } 537 CeedCallBackend(CeedLoadSourceToInitializedBuffer(ceed, grad_kernel_path, &num_file_paths, &file_paths, &basis_kernel_source)); 538 { 539 char *weight_kernel_name_base = "ceed/jit-source/magma/magma-basis-weight"; 540 CeedInt weight_kernel_name_len = strlen(weight_kernel_name_base) + 6; 541 char weight_kernel_name[weight_kernel_name_len]; 542 543 snprintf(weight_kernel_name, weight_kernel_name_len, "%s-%" CeedInt_FMT "d.h", weight_kernel_name_base, dim); 544 CeedCallBackend(CeedGetJitAbsolutePath(ceed, weight_kernel_name, &weight_kernel_path)); 545 } 546 CeedCallBackend(CeedLoadSourceToInitializedBuffer(ceed, weight_kernel_path, &num_file_paths, &file_paths, &basis_kernel_source)); 547 CeedDebug256(ceed, CEED_DEBUG_COLOR_SUCCESS, "----- Loading Basis Kernel Source Complete! -----\n"); 548 CeedCallBackend(CeedCompileMagma(ceed_delegate, basis_kernel_source, &impl->module, 5, "BASIS_DIM", dim, "BASIS_NUM_COMP", num_comp, "BASIS_P", 549 P_1d, "BASIS_Q", Q_1d, "BASIS_MAX_P_Q", CeedIntMax(P_1d, Q_1d))); 550 switch (dim) { 551 case 1: 552 CeedCallBackend(CeedGetKernelMagma(ceed, impl->module, "magma_interpn_1d_kernel", &impl->Interp)); 553 CeedCallBackend(CeedGetKernelMagma(ceed, impl->module, "magma_interpt_1d_kernel", &impl->InterpTranspose)); 554 CeedCallBackend(CeedGetKernelMagma(ceed, impl->module, "magma_gradn_1d_kernel", &impl->Grad)); 555 CeedCallBackend(CeedGetKernelMagma(ceed, impl->module, "magma_gradt_1d_kernel", &impl->GradTranspose)); 556 CeedCallBackend(CeedGetKernelMagma(ceed, impl->module, "magma_weight_1d_kernel", &impl->Weight)); 557 break; 558 case 2: 559 CeedCallBackend(CeedGetKernelMagma(ceed, impl->module, "magma_interpn_2d_kernel", &impl->Interp)); 560 CeedCallBackend(CeedGetKernelMagma(ceed, impl->module, "magma_interpt_2d_kernel", &impl->InterpTranspose)); 561 CeedCallBackend(CeedGetKernelMagma(ceed, impl->module, "magma_gradn_2d_kernel", &impl->Grad)); 562 CeedCallBackend(CeedGetKernelMagma(ceed, impl->module, "magma_gradt_2d_kernel", &impl->GradTranspose)); 563 CeedCallBackend(CeedGetKernelMagma(ceed, impl->module, "magma_weight_2d_kernel", &impl->Weight)); 564 break; 565 case 3: 566 CeedCallBackend(CeedGetKernelMagma(ceed, impl->module, "magma_interpn_3d_kernel", &impl->Interp)); 567 CeedCallBackend(CeedGetKernelMagma(ceed, impl->module, "magma_interpt_3d_kernel", &impl->InterpTranspose)); 568 CeedCallBackend(CeedGetKernelMagma(ceed, impl->module, "magma_gradn_3d_kernel", &impl->Grad)); 569 CeedCallBackend(CeedGetKernelMagma(ceed, impl->module, "magma_gradt_3d_kernel", &impl->GradTranspose)); 570 CeedCallBackend(CeedGetKernelMagma(ceed, impl->module, "magma_weight_3d_kernel", &impl->Weight)); 571 break; 572 } 573 CeedCallBackend(CeedFree(&interp_kernel_path)); 574 CeedCallBackend(CeedFree(&grad_kernel_path)); 575 CeedCallBackend(CeedFree(&weight_kernel_path)); 576 CeedCallBackend(CeedFree(&basis_kernel_source)); 577 for (CeedInt i = 0; i < num_file_paths; i++) CeedCall(CeedFree(&file_paths[i])); 578 CeedCall(CeedFree(&file_paths)); 579 580 CeedCallBackend(CeedBasisSetData(basis, impl)); 581 582 CeedCallBackend(CeedSetBackendFunction(ceed, "Basis", basis, "Apply", CeedBasisApply_Magma)); 583 CeedCallBackend(CeedSetBackendFunction(ceed, "Basis", basis, "Destroy", CeedBasisDestroy_Magma)); 584 return CEED_ERROR_SUCCESS; 585 } 586 587 //------------------------------------------------------------------------------ 588 // Create non-tensor H^1 589 //------------------------------------------------------------------------------ 590 int CeedBasisCreateH1_Magma(CeedElemTopology topo, CeedInt dim, CeedInt num_nodes, CeedInt num_qpts, const CeedScalar *interp, const CeedScalar *grad, 591 const CeedScalar *q_ref, const CeedScalar *q_weight, CeedBasis basis) { 592 Ceed ceed; 593 Ceed_Magma *data; 594 CeedBasisNonTensor_Magma *impl; 595 596 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 597 CeedCallBackend(CeedGetData(ceed, &data)); 598 CeedCallBackend(CeedCalloc(1, &impl)); 599 600 // Copy basis data to GPU 601 if (q_weight) { 602 CeedCallBackend(magma_malloc((void **)&impl->d_q_weight, num_qpts * sizeof(q_weight[0]))); 603 magma_setvector(num_qpts, sizeof(q_weight[0]), q_weight, 1, impl->d_q_weight, 1, data->queue); 604 } 605 if (interp) { 606 CeedInt q_comp_interp; 607 608 CeedCallBackend(CeedBasisGetNumQuadratureComponents(basis, CEED_EVAL_INTERP, &q_comp_interp)); 609 CeedCallBackend(magma_malloc((void **)&impl->d_interp, num_qpts * num_nodes * q_comp_interp * sizeof(interp[0]))); 610 magma_setvector(num_qpts * num_nodes * q_comp_interp, sizeof(interp[0]), interp, 1, impl->d_interp, 1, data->queue); 611 } 612 if (grad) { 613 CeedInt q_comp_grad; 614 615 CeedCallBackend(CeedBasisGetNumQuadratureComponents(basis, CEED_EVAL_GRAD, &q_comp_grad)); 616 CeedCallBackend(magma_malloc((void **)&impl->d_grad, num_qpts * num_nodes * q_comp_grad * sizeof(grad[0]))); 617 magma_setvector(num_qpts * num_nodes * q_comp_grad, sizeof(grad[0]), grad, 1, impl->d_grad, 1, data->queue); 618 } 619 620 // Compile the weight kernel if it won't be compiled later on 621 if (num_nodes > MAGMA_NONTENSOR_CUSTOM_KERNEL_MAX_P || num_qpts > MAGMA_NONTENSOR_CUSTOM_KERNEL_MAX_Q) { 622 Ceed ceed_delegate; 623 char *basis_kernel_source; 624 const char *weight_kernel_path; 625 626 // The RTC compilation code expects a Ceed with the common Ceed_Cuda or Ceed_Hip data 627 CeedCallBackend(CeedGetDelegate(ceed, &ceed_delegate)); 628 629 // Compile weight kernel (the remainder of kernel compilation happens at first call to CeedBasisApply) 630 CeedCallBackend(CeedGetJitAbsolutePath(ceed, "ceed/jit-source/magma/magma-basis-weight-nontensor.h", &weight_kernel_path)); 631 CeedDebug256(ceed, CEED_DEBUG_COLOR_SUCCESS, "----- Loading Basis Kernel Source -----\n"); 632 CeedCallBackend(CeedLoadSourceToBuffer(ceed, weight_kernel_path, &basis_kernel_source)); 633 CeedDebug256(ceed, CEED_DEBUG_COLOR_SUCCESS, "----- Loading Basis Kernel Source Complete! -----\n"); 634 CeedCallBackend(CeedCompileMagma(ceed_delegate, basis_kernel_source, &impl->module[0], 1, "BASIS_Q", num_qpts)); 635 CeedCallBackend(CeedGetKernelMagma(ceed, impl->module[0], "magma_weight_nontensor", &impl->Weight)); 636 CeedCallBackend(CeedFree(&weight_kernel_path)); 637 CeedCallBackend(CeedFree(&basis_kernel_source)); 638 } 639 640 CeedCallBackend(CeedBasisSetData(basis, impl)); 641 642 // Register backend functions 643 CeedCallBackend(CeedSetBackendFunction(ceed, "Basis", basis, "Apply", CeedBasisApplyNonTensor_Magma)); 644 CeedCallBackend(CeedSetBackendFunction(ceed, "Basis", basis, "Destroy", CeedBasisDestroyNonTensor_Magma)); 645 return CEED_ERROR_SUCCESS; 646 } 647 648 //------------------------------------------------------------------------------ 649 // Create non-tensor H(div) 650 //------------------------------------------------------------------------------ 651 int CeedBasisCreateHdiv_Magma(CeedElemTopology topo, CeedInt dim, CeedInt num_nodes, CeedInt num_qpts, const CeedScalar *interp, 652 const CeedScalar *div, const CeedScalar *q_ref, const CeedScalar *q_weight, CeedBasis basis) { 653 Ceed ceed; 654 Ceed_Magma *data; 655 CeedBasisNonTensor_Magma *impl; 656 657 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 658 CeedCallBackend(CeedGetData(ceed, &data)); 659 CeedCallBackend(CeedCalloc(1, &impl)); 660 661 // Copy basis data to GPU 662 if (q_weight) { 663 CeedCallBackend(magma_malloc((void **)&impl->d_q_weight, num_qpts * sizeof(q_weight[0]))); 664 magma_setvector(num_qpts, sizeof(q_weight[0]), q_weight, 1, impl->d_q_weight, 1, data->queue); 665 } 666 if (interp) { 667 CeedInt q_comp_interp; 668 669 CeedCallBackend(CeedBasisGetNumQuadratureComponents(basis, CEED_EVAL_INTERP, &q_comp_interp)); 670 CeedCallBackend(magma_malloc((void **)&impl->d_interp, num_qpts * num_nodes * q_comp_interp * sizeof(interp[0]))); 671 magma_setvector(num_qpts * num_nodes * q_comp_interp, sizeof(interp[0]), interp, 1, impl->d_interp, 1, data->queue); 672 } 673 if (div) { 674 CeedInt q_comp_div; 675 676 CeedCallBackend(CeedBasisGetNumQuadratureComponents(basis, CEED_EVAL_DIV, &q_comp_div)); 677 CeedCallBackend(magma_malloc((void **)&impl->d_div, num_qpts * num_nodes * q_comp_div * sizeof(div[0]))); 678 magma_setvector(num_qpts * num_nodes * q_comp_div, sizeof(div[0]), div, 1, impl->d_div, 1, data->queue); 679 } 680 681 // Compile the weight kernel if it won't be compiled later on 682 if (num_nodes > MAGMA_NONTENSOR_CUSTOM_KERNEL_MAX_P || num_qpts > MAGMA_NONTENSOR_CUSTOM_KERNEL_MAX_Q) { 683 Ceed ceed_delegate; 684 char *basis_kernel_source; 685 const char *weight_kernel_path; 686 687 // The RTC compilation code expects a Ceed with the common Ceed_Cuda or Ceed_Hip data 688 CeedCallBackend(CeedGetDelegate(ceed, &ceed_delegate)); 689 690 // Compile weight kernel (the remainder of kernel compilation happens at first call to CeedBasisApply) 691 CeedCallBackend(CeedGetJitAbsolutePath(ceed, "ceed/jit-source/magma/magma-basis-weight-nontensor.h", &weight_kernel_path)); 692 CeedDebug256(ceed, CEED_DEBUG_COLOR_SUCCESS, "----- Loading Basis Kernel Source -----\n"); 693 CeedCallBackend(CeedLoadSourceToBuffer(ceed, weight_kernel_path, &basis_kernel_source)); 694 CeedDebug256(ceed, CEED_DEBUG_COLOR_SUCCESS, "----- Loading Basis Kernel Source Complete! -----\n"); 695 CeedCallBackend(CeedCompileMagma(ceed_delegate, basis_kernel_source, &impl->module[0], 1, "BASIS_Q", num_qpts)); 696 CeedCallBackend(CeedGetKernelMagma(ceed, impl->module[0], "magma_weight_nontensor", &impl->Weight)); 697 CeedCallBackend(CeedFree(&weight_kernel_path)); 698 CeedCallBackend(CeedFree(&basis_kernel_source)); 699 } 700 701 CeedCallBackend(CeedBasisSetData(basis, impl)); 702 703 // Register backend functions 704 CeedCallBackend(CeedSetBackendFunction(ceed, "Basis", basis, "Apply", CeedBasisApplyNonTensor_Magma)); 705 CeedCallBackend(CeedSetBackendFunction(ceed, "Basis", basis, "Destroy", CeedBasisDestroyNonTensor_Magma)); 706 return CEED_ERROR_SUCCESS; 707 } 708 709 //------------------------------------------------------------------------------ 710 // Create non-tensor H(curl) 711 //------------------------------------------------------------------------------ 712 int CeedBasisCreateHcurl_Magma(CeedElemTopology topo, CeedInt dim, CeedInt num_nodes, CeedInt num_qpts, const CeedScalar *interp, 713 const CeedScalar *curl, const CeedScalar *q_ref, const CeedScalar *q_weight, CeedBasis basis) { 714 Ceed ceed; 715 Ceed_Magma *data; 716 CeedBasisNonTensor_Magma *impl; 717 718 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 719 CeedCallBackend(CeedGetData(ceed, &data)); 720 CeedCallBackend(CeedCalloc(1, &impl)); 721 722 // Copy basis data to GPU 723 if (q_weight) { 724 CeedCallBackend(magma_malloc((void **)&impl->d_q_weight, num_qpts * sizeof(q_weight[0]))); 725 magma_setvector(num_qpts, sizeof(q_weight[0]), q_weight, 1, impl->d_q_weight, 1, data->queue); 726 } 727 if (interp) { 728 CeedInt q_comp_interp; 729 730 CeedCallBackend(CeedBasisGetNumQuadratureComponents(basis, CEED_EVAL_INTERP, &q_comp_interp)); 731 CeedCallBackend(magma_malloc((void **)&impl->d_interp, num_qpts * num_nodes * q_comp_interp * sizeof(interp[0]))); 732 magma_setvector(num_qpts * num_nodes * q_comp_interp, sizeof(interp[0]), interp, 1, impl->d_interp, 1, data->queue); 733 } 734 if (curl) { 735 CeedInt q_comp_curl; 736 737 CeedCallBackend(CeedBasisGetNumQuadratureComponents(basis, CEED_EVAL_CURL, &q_comp_curl)); 738 CeedCallBackend(magma_malloc((void **)&impl->d_curl, num_qpts * num_nodes * q_comp_curl * sizeof(curl[0]))); 739 magma_setvector(num_qpts * num_nodes * q_comp_curl, sizeof(curl[0]), curl, 1, impl->d_curl, 1, data->queue); 740 } 741 742 // Compile the weight kernel if it won't be compiled later on 743 if (num_nodes > MAGMA_NONTENSOR_CUSTOM_KERNEL_MAX_P || num_qpts > MAGMA_NONTENSOR_CUSTOM_KERNEL_MAX_Q) { 744 Ceed ceed_delegate; 745 char *basis_kernel_source; 746 const char *weight_kernel_path; 747 748 // The RTC compilation code expects a Ceed with the common Ceed_Cuda or Ceed_Hip data 749 CeedCallBackend(CeedGetDelegate(ceed, &ceed_delegate)); 750 751 // Compile weight kernel (the remainder of kernel compilation happens at first call to CeedBasisApply) 752 CeedCallBackend(CeedGetJitAbsolutePath(ceed, "ceed/jit-source/magma/magma-basis-weight-nontensor.h", &weight_kernel_path)); 753 CeedDebug256(ceed, CEED_DEBUG_COLOR_SUCCESS, "----- Loading Basis Kernel Source -----\n"); 754 CeedCallBackend(CeedLoadSourceToBuffer(ceed, weight_kernel_path, &basis_kernel_source)); 755 CeedDebug256(ceed, CEED_DEBUG_COLOR_SUCCESS, "----- Loading Basis Kernel Source Complete! -----\n"); 756 CeedCallBackend(CeedCompileMagma(ceed_delegate, basis_kernel_source, &impl->module[0], 1, "BASIS_Q", num_qpts)); 757 CeedCallBackend(CeedGetKernelMagma(ceed, impl->module[0], "magma_weight_nontensor", &impl->Weight)); 758 CeedCallBackend(CeedFree(&weight_kernel_path)); 759 CeedCallBackend(CeedFree(&basis_kernel_source)); 760 } 761 762 CeedCallBackend(CeedBasisSetData(basis, impl)); 763 764 // Register backend functions 765 CeedCallBackend(CeedSetBackendFunction(ceed, "Basis", basis, "Apply", CeedBasisApplyNonTensor_Magma)); 766 CeedCallBackend(CeedSetBackendFunction(ceed, "Basis", basis, "Destroy", CeedBasisDestroyNonTensor_Magma)); 767 return CEED_ERROR_SUCCESS; 768 } 769 770 //------------------------------------------------------------------------------ 771