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 <stdbool.h> 12 #include <stddef.h> 13 #include <string.h> 14 #include <hip/hip_runtime.h> 15 16 #include "../hip/ceed-hip-common.h" 17 #include "../hip/ceed-hip-compile.h" 18 #include "ceed-hip-shared.h" 19 20 //------------------------------------------------------------------------------ 21 // Compute a block size based on required minimum threads 22 //------------------------------------------------------------------------------ 23 static CeedInt ComputeBlockSizeFromRequirement(const CeedInt required) { 24 CeedInt maxSize = 1024; // Max total threads per block 25 CeedInt currentSize = 64; // Start with one group 26 27 while (currentSize < maxSize) { 28 if (currentSize > required) break; 29 else currentSize = currentSize * 2; 30 } 31 return currentSize; 32 } 33 34 //------------------------------------------------------------------------------ 35 // Compute required thread block sizes for basis kernels given P, Q, dim, and 36 // num_comp (num_comp not currently used, but may be again in other basis 37 // parallelization options) 38 //------------------------------------------------------------------------------ 39 static int ComputeBasisThreadBlockSizes(const CeedInt dim, const CeedInt P_1d, const CeedInt Q_1d, const CeedInt num_comp, CeedInt *block_sizes) { 40 // Note that this will use the same block sizes for all dimensions when compiling, 41 // but as each basis object is defined for a particular dimension, we will never 42 // call any kernels except the ones for the dimension for which we have computed the 43 // block sizes. 44 const CeedInt thread_1d = CeedIntMax(P_1d, Q_1d); 45 46 switch (dim) { 47 case 1: { 48 // Interp kernels: 49 block_sizes[0] = 256; 50 51 // Grad kernels: 52 block_sizes[1] = 256; 53 54 // Weight kernels: 55 block_sizes[2] = 256; 56 } break; 57 case 2: { 58 // Interp kernels: 59 CeedInt required = thread_1d * thread_1d; 60 61 block_sizes[0] = CeedIntMax(256, ComputeBlockSizeFromRequirement(required)); 62 63 // Grad kernels: currently use same required minimum threads 64 block_sizes[1] = CeedIntMax(256, ComputeBlockSizeFromRequirement(required)); 65 66 // Weight kernels: 67 required = CeedIntMax(64, Q_1d * Q_1d); 68 block_sizes[2] = CeedIntMax(256, ComputeBlockSizeFromRequirement(required)); 69 70 } break; 71 case 3: { 72 // Interp kernels: 73 CeedInt required = thread_1d * thread_1d; 74 75 block_sizes[0] = CeedIntMax(256, ComputeBlockSizeFromRequirement(required)); 76 77 // Grad kernels: currently use same required minimum threads 78 block_sizes[1] = CeedIntMax(256, ComputeBlockSizeFromRequirement(required)); 79 80 // Weight kernels: 81 required = Q_1d * Q_1d * Q_1d; 82 block_sizes[2] = CeedIntMax(256, ComputeBlockSizeFromRequirement(required)); 83 } 84 } 85 return CEED_ERROR_SUCCESS; 86 } 87 88 //------------------------------------------------------------------------------ 89 // Apply basis 90 //------------------------------------------------------------------------------ 91 static int CeedBasisApplyTensorCore_Hip_shared(CeedBasis basis, bool apply_add, const CeedInt num_elem, CeedTransposeMode t_mode, 92 CeedEvalMode eval_mode, CeedVector u, CeedVector v) { 93 Ceed ceed; 94 Ceed_Hip *ceed_Hip; 95 CeedInt dim, num_comp; 96 const CeedScalar *d_u; 97 CeedScalar *d_v; 98 CeedBasis_Hip_shared *data; 99 100 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 101 CeedCallBackend(CeedGetData(ceed, &ceed_Hip)); 102 CeedCallBackend(CeedBasisGetData(basis, &data)); 103 CeedCallBackend(CeedBasisGetDimension(basis, &dim)); 104 CeedCallBackend(CeedBasisGetNumComponents(basis, &num_comp)); 105 106 // Get read/write access to u, v 107 if (u != CEED_VECTOR_NONE) CeedCallBackend(CeedVectorGetArrayRead(u, CEED_MEM_DEVICE, &d_u)); 108 else CeedCheck(eval_mode == CEED_EVAL_WEIGHT, ceed, CEED_ERROR_BACKEND, "An input vector is required for this CeedEvalMode"); 109 if (apply_add) CeedCallBackend(CeedVectorGetArray(v, CEED_MEM_DEVICE, &d_v)); 110 else CeedCallBackend(CeedVectorGetArrayWrite(v, CEED_MEM_DEVICE, &d_v)); 111 112 // Apply basis operation 113 switch (eval_mode) { 114 case CEED_EVAL_INTERP: { 115 CeedInt P_1d, Q_1d; 116 CeedInt block_size = data->block_sizes[0]; 117 118 CeedCheck(data->d_interp_1d, ceed, CEED_ERROR_BACKEND, "%s not supported; interp_1d not set", CeedEvalModes[eval_mode]); 119 CeedCallBackend(CeedBasisGetNumNodes1D(basis, &P_1d)); 120 CeedCallBackend(CeedBasisGetNumQuadraturePoints1D(basis, &Q_1d)); 121 CeedInt thread_1d = CeedIntMax(Q_1d, P_1d); 122 void *interp_args[] = {(void *)&num_elem, &data->d_interp_1d, &d_u, &d_v}; 123 124 if (dim == 1) { 125 CeedInt elems_per_block = 64 * thread_1d > 256 ? 256 / thread_1d : 64; 126 elems_per_block = elems_per_block > 0 ? elems_per_block : 1; 127 CeedInt grid = num_elem / elems_per_block + (num_elem % elems_per_block > 0); 128 CeedInt shared_mem = elems_per_block * thread_1d * sizeof(CeedScalar); 129 130 if (t_mode == CEED_TRANSPOSE) { 131 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, apply_add ? data->InterpTransposeAdd : data->InterpTranspose, grid, thread_1d, 1, 132 elems_per_block, shared_mem, interp_args)); 133 } else { 134 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, data->Interp, grid, thread_1d, 1, elems_per_block, shared_mem, interp_args)); 135 } 136 } else if (dim == 2) { 137 // Check if required threads is small enough to do multiple elems 138 const CeedInt elems_per_block = CeedIntMax(block_size / (thread_1d * thread_1d), 1); 139 CeedInt grid = num_elem / elems_per_block + (num_elem % elems_per_block > 0); 140 CeedInt shared_mem = elems_per_block * thread_1d * thread_1d * sizeof(CeedScalar); 141 142 if (t_mode == CEED_TRANSPOSE) { 143 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, apply_add ? data->InterpTransposeAdd : data->InterpTranspose, grid, thread_1d, thread_1d, 144 elems_per_block, shared_mem, interp_args)); 145 } else { 146 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, data->Interp, grid, thread_1d, thread_1d, elems_per_block, shared_mem, interp_args)); 147 } 148 } else if (dim == 3) { 149 const CeedInt elems_per_block = CeedIntMax(block_size / (thread_1d * thread_1d), 1); 150 CeedInt grid = num_elem / elems_per_block + (num_elem % elems_per_block > 0); 151 CeedInt shared_mem = elems_per_block * thread_1d * thread_1d * sizeof(CeedScalar); 152 153 if (t_mode == CEED_TRANSPOSE) { 154 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, apply_add ? data->InterpTransposeAdd : data->InterpTranspose, grid, thread_1d, thread_1d, 155 elems_per_block, shared_mem, interp_args)); 156 } else { 157 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, data->Interp, grid, thread_1d, thread_1d, elems_per_block, shared_mem, interp_args)); 158 } 159 } 160 } break; 161 case CEED_EVAL_GRAD: { 162 CeedInt P_1d, Q_1d; 163 CeedInt block_size = data->block_sizes[1]; 164 165 CeedCheck(data->d_grad_1d, ceed, CEED_ERROR_BACKEND, "%s not supported; grad_1d not set", CeedEvalModes[eval_mode]); 166 CeedCallBackend(CeedBasisGetNumNodes1D(basis, &P_1d)); 167 CeedCallBackend(CeedBasisGetNumQuadraturePoints1D(basis, &Q_1d)); 168 CeedInt thread_1d = CeedIntMax(Q_1d, P_1d); 169 CeedScalar *d_grad_1d = data->d_grad_1d; 170 171 if (data->d_collo_grad_1d) { 172 d_grad_1d = data->d_collo_grad_1d; 173 } 174 void *grad_args[] = {(void *)&num_elem, &data->d_interp_1d, &d_grad_1d, &d_u, &d_v}; 175 176 if (dim == 1) { 177 CeedInt elems_per_block = 64 * thread_1d > 256 ? 256 / thread_1d : 64; 178 elems_per_block = elems_per_block > 0 ? elems_per_block : 1; 179 CeedInt grid = num_elem / elems_per_block + (num_elem % elems_per_block > 0); 180 CeedInt shared_mem = elems_per_block * thread_1d * sizeof(CeedScalar); 181 182 if (t_mode == CEED_TRANSPOSE) { 183 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, apply_add ? data->GradTransposeAdd : data->GradTranspose, grid, thread_1d, 1, 184 elems_per_block, shared_mem, grad_args)); 185 } else { 186 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, data->Grad, grid, thread_1d, 1, elems_per_block, shared_mem, grad_args)); 187 } 188 } else if (dim == 2) { 189 // Check if required threads is small enough to do multiple elems 190 const CeedInt elems_per_block = CeedIntMax(block_size / (thread_1d * thread_1d), 1); 191 CeedInt grid = num_elem / elems_per_block + (num_elem % elems_per_block > 0); 192 CeedInt shared_mem = elems_per_block * thread_1d * thread_1d * sizeof(CeedScalar); 193 194 if (t_mode == CEED_TRANSPOSE) { 195 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, apply_add ? data->GradTransposeAdd : data->GradTranspose, grid, thread_1d, thread_1d, 196 elems_per_block, shared_mem, grad_args)); 197 } else { 198 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, data->Grad, grid, thread_1d, thread_1d, elems_per_block, shared_mem, grad_args)); 199 } 200 } else if (dim == 3) { 201 const CeedInt elems_per_block = CeedIntMax(block_size / (thread_1d * thread_1d), 1); 202 CeedInt grid = num_elem / elems_per_block + (num_elem % elems_per_block > 0); 203 CeedInt shared_mem = elems_per_block * thread_1d * thread_1d * sizeof(CeedScalar); 204 205 if (t_mode == CEED_TRANSPOSE) { 206 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, apply_add ? data->GradTransposeAdd : data->GradTranspose, grid, thread_1d, thread_1d, 207 elems_per_block, shared_mem, grad_args)); 208 } else { 209 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, data->Grad, grid, thread_1d, thread_1d, elems_per_block, shared_mem, grad_args)); 210 } 211 } 212 } break; 213 case CEED_EVAL_WEIGHT: { 214 CeedInt Q_1d; 215 CeedInt block_size = data->block_sizes[2]; 216 217 CeedCheck(data->d_q_weight_1d, ceed, CEED_ERROR_BACKEND, "%s not supported; q_weights_1d not set", CeedEvalModes[eval_mode]); 218 CeedCallBackend(CeedBasisGetNumQuadraturePoints1D(basis, &Q_1d)); 219 void *weight_args[] = {(void *)&num_elem, (void *)&data->d_q_weight_1d, &d_v}; 220 221 if (dim == 1) { 222 const CeedInt opt_elems = block_size / Q_1d; 223 const CeedInt elems_per_block = opt_elems > 0 ? opt_elems : 1; 224 const CeedInt grid_size = num_elem / elems_per_block + (num_elem % elems_per_block > 0); 225 226 CeedCallBackend(CeedRunKernelDim_Hip(ceed, data->Weight, grid_size, Q_1d, elems_per_block, 1, weight_args)); 227 } else if (dim == 2) { 228 const CeedInt opt_elems = block_size / (Q_1d * Q_1d); 229 const CeedInt elems_per_block = opt_elems > 0 ? opt_elems : 1; 230 const CeedInt grid_size = num_elem / elems_per_block + (num_elem % elems_per_block > 0); 231 232 CeedCallBackend(CeedRunKernelDim_Hip(ceed, data->Weight, grid_size, Q_1d, Q_1d, elems_per_block, weight_args)); 233 } else if (dim == 3) { 234 const CeedInt opt_elems = block_size / (Q_1d * Q_1d); 235 const CeedInt elems_per_block = opt_elems > 0 ? opt_elems : 1; 236 const CeedInt grid_size = num_elem / elems_per_block + (num_elem % elems_per_block > 0); 237 238 CeedCallBackend(CeedRunKernelDim_Hip(ceed, data->Weight, grid_size, Q_1d, Q_1d, elems_per_block, weight_args)); 239 } 240 } break; 241 case CEED_EVAL_NONE: /* handled separately below */ 242 break; 243 // LCOV_EXCL_START 244 case CEED_EVAL_DIV: 245 case CEED_EVAL_CURL: 246 return CeedError(ceed, CEED_ERROR_BACKEND, "%s not supported", CeedEvalModes[eval_mode]); 247 // LCOV_EXCL_STOP 248 } 249 250 // Restore vectors, cover CEED_EVAL_NONE 251 CeedCallBackend(CeedVectorRestoreArray(v, &d_v)); 252 if (eval_mode == CEED_EVAL_NONE) CeedCallBackend(CeedVectorSetArray(v, CEED_MEM_DEVICE, CEED_COPY_VALUES, (CeedScalar *)d_u)); 253 if (eval_mode != CEED_EVAL_WEIGHT) CeedCallBackend(CeedVectorRestoreArrayRead(u, &d_u)); 254 CeedCallBackend(CeedDestroy(&ceed)); 255 return CEED_ERROR_SUCCESS; 256 } 257 258 int CeedBasisApplyTensor_Hip_shared(CeedBasis basis, const CeedInt num_elem, CeedTransposeMode t_mode, CeedEvalMode eval_mode, CeedVector u, 259 CeedVector v) { 260 CeedCallBackend(CeedBasisApplyTensorCore_Hip_shared(basis, false, num_elem, t_mode, eval_mode, u, v)); 261 return CEED_ERROR_SUCCESS; 262 } 263 264 int CeedBasisApplyAddTensor_Hip_shared(CeedBasis basis, const CeedInt num_elem, CeedTransposeMode t_mode, CeedEvalMode eval_mode, CeedVector u, 265 CeedVector v) { 266 CeedCallBackend(CeedBasisApplyTensorCore_Hip_shared(basis, true, num_elem, t_mode, eval_mode, u, v)); 267 return CEED_ERROR_SUCCESS; 268 } 269 270 //------------------------------------------------------------------------------ 271 // Basis apply - tensor AtPoints 272 //------------------------------------------------------------------------------ 273 static int CeedBasisApplyAtPointsCore_Hip_shared(CeedBasis basis, bool apply_add, const CeedInt num_elem, const CeedInt *num_points, 274 CeedTransposeMode t_mode, CeedEvalMode eval_mode, CeedVector x_ref, CeedVector u, CeedVector v) { 275 Ceed ceed; 276 CeedInt Q_1d, dim, max_num_points = num_points[0]; 277 const CeedInt is_transpose = t_mode == CEED_TRANSPOSE; 278 const CeedScalar *d_x, *d_u; 279 CeedScalar *d_v; 280 CeedBasis_Hip_shared *data; 281 282 CeedCallBackend(CeedBasisGetData(basis, &data)); 283 CeedCallBackend(CeedBasisGetNumQuadraturePoints1D(basis, &Q_1d)); 284 CeedCallBackend(CeedBasisGetDimension(basis, &dim)); 285 286 // Weight handled separately 287 if (eval_mode == CEED_EVAL_WEIGHT) { 288 CeedCallBackend(CeedVectorSetValue(v, 1.0)); 289 return CEED_ERROR_SUCCESS; 290 } 291 292 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 293 294 // Check padded to uniform number of points per elem 295 for (CeedInt i = 1; i < num_elem; i++) max_num_points = CeedIntMax(max_num_points, num_points[i]); 296 { 297 CeedInt num_comp, q_comp; 298 CeedSize len, len_required; 299 300 CeedCallBackend(CeedBasisGetNumComponents(basis, &num_comp)); 301 CeedCallBackend(CeedBasisGetNumQuadratureComponents(basis, eval_mode, &q_comp)); 302 CeedCallBackend(CeedVectorGetLength(is_transpose ? u : v, &len)); 303 len_required = (CeedSize)num_comp * (CeedSize)q_comp * (CeedSize)num_elem * (CeedSize)max_num_points; 304 CeedCheck(len >= len_required, ceed, CEED_ERROR_BACKEND, 305 "Vector at points must be padded to the same number of points in each element for BasisApplyAtPoints on GPU backends." 306 " Found %" CeedSize_FMT ", Required %" CeedSize_FMT, 307 len, len_required); 308 } 309 310 // Move num_points array to device 311 if (is_transpose) { 312 const CeedInt num_bytes = num_elem * sizeof(CeedInt); 313 314 if (num_elem != data->num_elem_at_points) { 315 data->num_elem_at_points = num_elem; 316 317 if (data->d_points_per_elem) CeedCallHip(ceed, hipFree(data->d_points_per_elem)); 318 CeedCallHip(ceed, hipMalloc((void **)&data->d_points_per_elem, num_bytes)); 319 CeedCallBackend(CeedFree(&data->h_points_per_elem)); 320 CeedCallBackend(CeedCalloc(num_elem, &data->h_points_per_elem)); 321 } 322 if (memcmp(data->h_points_per_elem, num_points, num_bytes)) { 323 memcpy(data->h_points_per_elem, num_points, num_bytes); 324 CeedCallHip(ceed, hipMemcpy(data->d_points_per_elem, num_points, num_bytes, hipMemcpyHostToDevice)); 325 } 326 } 327 328 // Build kernels if needed 329 if (data->num_points != max_num_points) { 330 CeedInt P_1d; 331 332 CeedCallBackend(CeedBasisGetNumNodes1D(basis, &P_1d)); 333 data->num_points = max_num_points; 334 335 // -- Create interp matrix to Chebyshev coefficients 336 if (!data->d_chebyshev_interp_1d) { 337 CeedSize interp_bytes; 338 CeedScalar *chebyshev_interp_1d; 339 340 interp_bytes = P_1d * Q_1d * sizeof(CeedScalar); 341 CeedCallBackend(CeedCalloc(P_1d * Q_1d, &chebyshev_interp_1d)); 342 CeedCallBackend(CeedBasisGetChebyshevInterp1D(basis, chebyshev_interp_1d)); 343 CeedCallHip(ceed, hipMalloc((void **)&data->d_chebyshev_interp_1d, interp_bytes)); 344 CeedCallHip(ceed, hipMemcpy(data->d_chebyshev_interp_1d, chebyshev_interp_1d, interp_bytes, hipMemcpyHostToDevice)); 345 CeedCallBackend(CeedFree(&chebyshev_interp_1d)); 346 } 347 348 // -- Compile kernels 349 const char basis_kernel_source[] = "// AtPoints basis source\n#include <ceed/jit-source/hip/hip-shared-basis-tensor-at-points.h>\n"; 350 CeedInt num_comp; 351 352 if (data->moduleAtPoints) CeedCallHip(ceed, hipModuleUnload(data->moduleAtPoints)); 353 CeedCallBackend(CeedBasisGetNumComponents(basis, &num_comp)); 354 CeedCallBackend(CeedCompile_Hip(ceed, basis_kernel_source, &data->moduleAtPoints, 9, "BASIS_Q_1D", Q_1d, "BASIS_P_1D", P_1d, "T_1D", 355 CeedIntMax(Q_1d, P_1d), "BASIS_DIM", dim, "BASIS_NUM_COMP", num_comp, "BASIS_NUM_NODES", CeedIntPow(P_1d, dim), 356 "BASIS_NUM_QPTS", CeedIntPow(Q_1d, dim), "BASIS_NUM_PTS", max_num_points, "BASIS_INTERP_BLOCK_SIZE", 357 data->block_sizes[0])); 358 CeedCallBackend(CeedGetKernel_Hip(ceed, data->moduleAtPoints, "InterpAtPoints", &data->InterpAtPoints)); 359 CeedCallBackend(CeedGetKernel_Hip(ceed, data->moduleAtPoints, "InterpTransposeAtPoints", &data->InterpTransposeAtPoints)); 360 CeedCallBackend(CeedGetKernel_Hip(ceed, data->moduleAtPoints, "GradAtPoints", &data->GradAtPoints)); 361 CeedCallBackend(CeedGetKernel_Hip(ceed, data->moduleAtPoints, "GradTransposeAtPoints", &data->GradTransposeAtPoints)); 362 } 363 364 // Get read/write access to u, v 365 CeedCallBackend(CeedVectorGetArrayRead(x_ref, CEED_MEM_DEVICE, &d_x)); 366 if (u != CEED_VECTOR_NONE) CeedCallBackend(CeedVectorGetArrayRead(u, CEED_MEM_DEVICE, &d_u)); 367 else CeedCheck(eval_mode == CEED_EVAL_WEIGHT, ceed, CEED_ERROR_BACKEND, "An input vector is required for this CeedEvalMode"); 368 if (apply_add) { 369 CeedCallBackend(CeedVectorGetArray(v, CEED_MEM_DEVICE, &d_v)); 370 } else { 371 // Clear v for transpose operation 372 if (is_transpose) CeedCallBackend(CeedVectorSetValue(v, 0.0)); 373 CeedCallBackend(CeedVectorGetArrayWrite(v, CEED_MEM_DEVICE, &d_v)); 374 } 375 376 // Basis action 377 switch (eval_mode) { 378 case CEED_EVAL_INTERP: { 379 CeedInt P_1d, Q_1d; 380 CeedInt block_size = data->block_sizes[0]; 381 382 CeedCallBackend(CeedBasisGetNumNodes1D(basis, &P_1d)); 383 CeedCallBackend(CeedBasisGetNumQuadraturePoints1D(basis, &Q_1d)); 384 CeedInt thread_1d = CeedIntMax(Q_1d, P_1d); 385 void *interp_args[] = {(void *)&num_elem, &data->d_chebyshev_interp_1d, &data->d_points_per_elem, &d_x, &d_u, &d_v}; 386 387 if (dim == 1) { 388 CeedInt elems_per_block = 64 * thread_1d > 256 ? 256 / thread_1d : 64; 389 elems_per_block = elems_per_block > 0 ? elems_per_block : 1; 390 CeedInt grid = num_elem / elems_per_block + (num_elem % elems_per_block > 0); 391 CeedInt shared_mem = elems_per_block * thread_1d * sizeof(CeedScalar); 392 393 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, is_transpose ? data->InterpTransposeAtPoints : data->InterpAtPoints, grid, thread_1d, 1, 394 elems_per_block, shared_mem, interp_args)); 395 } else if (dim == 2) { 396 // Check if required threads is small enough to do multiple elems 397 const CeedInt elems_per_block = CeedIntMax(block_size / (thread_1d * thread_1d), 1); 398 CeedInt grid = num_elem / elems_per_block + (num_elem % elems_per_block > 0); 399 CeedInt shared_mem = elems_per_block * thread_1d * thread_1d * sizeof(CeedScalar); 400 401 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, is_transpose ? data->InterpTransposeAtPoints : data->InterpAtPoints, grid, thread_1d, 402 thread_1d, elems_per_block, shared_mem, interp_args)); 403 } else if (dim == 3) { 404 const CeedInt elems_per_block = 1; 405 CeedInt grid = num_elem / elems_per_block + (num_elem % elems_per_block > 0); 406 CeedInt shared_mem = elems_per_block * thread_1d * thread_1d * sizeof(CeedScalar); 407 408 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, is_transpose ? data->InterpTransposeAtPoints : data->InterpAtPoints, grid, thread_1d, 409 thread_1d, elems_per_block, shared_mem, interp_args)); 410 } 411 } break; 412 case CEED_EVAL_GRAD: { 413 CeedInt P_1d, Q_1d; 414 CeedInt block_size = data->block_sizes[0]; 415 416 CeedCallBackend(CeedBasisGetNumNodes1D(basis, &P_1d)); 417 CeedCallBackend(CeedBasisGetNumQuadraturePoints1D(basis, &Q_1d)); 418 CeedInt thread_1d = CeedIntMax(Q_1d, P_1d); 419 void *grad_args[] = {(void *)&num_elem, &data->d_chebyshev_interp_1d, &data->d_points_per_elem, &d_x, &d_u, &d_v}; 420 421 if (dim == 1) { 422 CeedInt elems_per_block = 64 * thread_1d > 256 ? 256 / thread_1d : 64; 423 elems_per_block = elems_per_block > 0 ? elems_per_block : 1; 424 CeedInt grid = num_elem / elems_per_block + (num_elem % elems_per_block > 0); 425 CeedInt shared_mem = elems_per_block * thread_1d * sizeof(CeedScalar); 426 427 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, is_transpose ? data->GradTransposeAtPoints : data->GradAtPoints, grid, thread_1d, 1, 428 elems_per_block, shared_mem, grad_args)); 429 } else if (dim == 2) { 430 // Check if required threads is small enough to do multiple elems 431 const CeedInt elems_per_block = CeedIntMax(block_size / (thread_1d * thread_1d), 1); 432 CeedInt grid = num_elem / elems_per_block + (num_elem % elems_per_block > 0); 433 CeedInt shared_mem = elems_per_block * thread_1d * thread_1d * sizeof(CeedScalar); 434 435 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, is_transpose ? data->GradTransposeAtPoints : data->GradAtPoints, grid, thread_1d, thread_1d, 436 elems_per_block, shared_mem, grad_args)); 437 } else if (dim == 3) { 438 const CeedInt elems_per_block = 1; 439 CeedInt grid = num_elem / elems_per_block + (num_elem % elems_per_block > 0); 440 CeedInt shared_mem = elems_per_block * thread_1d * thread_1d * sizeof(CeedScalar); 441 442 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, is_transpose ? data->GradTransposeAtPoints : data->GradAtPoints, grid, thread_1d, thread_1d, 443 elems_per_block, shared_mem, grad_args)); 444 } 445 } break; 446 case CEED_EVAL_WEIGHT: 447 case CEED_EVAL_NONE: /* handled separately below */ 448 break; 449 // LCOV_EXCL_START 450 case CEED_EVAL_DIV: 451 case CEED_EVAL_CURL: 452 return CeedError(ceed, CEED_ERROR_BACKEND, "%s not supported", CeedEvalModes[eval_mode]); 453 // LCOV_EXCL_STOP 454 } 455 456 // Restore vectors, cover CEED_EVAL_NONE 457 CeedCallBackend(CeedVectorRestoreArrayRead(x_ref, &d_x)); 458 CeedCallBackend(CeedVectorRestoreArray(v, &d_v)); 459 if (eval_mode == CEED_EVAL_NONE) CeedCallBackend(CeedVectorSetArray(v, CEED_MEM_DEVICE, CEED_COPY_VALUES, (CeedScalar *)d_u)); 460 if (eval_mode != CEED_EVAL_WEIGHT) CeedCallBackend(CeedVectorRestoreArrayRead(u, &d_u)); 461 return CEED_ERROR_SUCCESS; 462 } 463 464 static int CeedBasisApplyAtPoints_Hip_shared(CeedBasis basis, const CeedInt num_elem, const CeedInt *num_points, CeedTransposeMode t_mode, 465 CeedEvalMode eval_mode, CeedVector x_ref, CeedVector u, CeedVector v) { 466 CeedCallBackend(CeedBasisApplyAtPointsCore_Hip_shared(basis, false, num_elem, num_points, t_mode, eval_mode, x_ref, u, v)); 467 return CEED_ERROR_SUCCESS; 468 } 469 470 static int CeedBasisApplyAddAtPoints_Hip_shared(CeedBasis basis, const CeedInt num_elem, const CeedInt *num_points, CeedTransposeMode t_mode, 471 CeedEvalMode eval_mode, CeedVector x_ref, CeedVector u, CeedVector v) { 472 CeedCallBackend(CeedBasisApplyAtPointsCore_Hip_shared(basis, true, num_elem, num_points, t_mode, eval_mode, x_ref, u, v)); 473 return CEED_ERROR_SUCCESS; 474 } 475 476 //------------------------------------------------------------------------------ 477 // Apply basis 478 //------------------------------------------------------------------------------ 479 static int CeedBasisApplyNonTensorCore_Hip_shared(CeedBasis basis, bool apply_add, const CeedInt num_elem, CeedTransposeMode t_mode, 480 CeedEvalMode eval_mode, CeedVector u, CeedVector v) { 481 Ceed ceed; 482 Ceed_Hip *ceed_Hip; 483 CeedInt dim, num_comp; 484 const CeedScalar *d_u; 485 CeedScalar *d_v; 486 CeedBasis_Hip_shared *data; 487 488 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 489 CeedCallBackend(CeedGetData(ceed, &ceed_Hip)); 490 CeedCallBackend(CeedBasisGetData(basis, &data)); 491 CeedCallBackend(CeedBasisGetDimension(basis, &dim)); 492 CeedCallBackend(CeedBasisGetNumComponents(basis, &num_comp)); 493 494 // Get read/write access to u, v 495 if (u != CEED_VECTOR_NONE) CeedCallBackend(CeedVectorGetArrayRead(u, CEED_MEM_DEVICE, &d_u)); 496 else CeedCheck(eval_mode == CEED_EVAL_WEIGHT, ceed, CEED_ERROR_BACKEND, "An input vector is required for this CeedEvalMode"); 497 if (apply_add) { 498 CeedCallBackend(CeedVectorGetArray(v, CEED_MEM_DEVICE, &d_v)); 499 } else { 500 // Clear v for transpose operation 501 if (is_transpose) CeedCallBackend(CeedVectorSetValue(v, 0.0)); 502 CeedCallBackend(CeedVectorGetArrayWrite(v, CEED_MEM_DEVICE, &d_v)); 503 } 504 505 // Apply basis operation 506 switch (eval_mode) { 507 case CEED_EVAL_INTERP: { 508 CeedInt P, Q; 509 510 CeedCheck(data->d_interp_1d, ceed, CEED_ERROR_BACKEND, "%s not supported; interp not set", CeedEvalModes[eval_mode]); 511 CeedCallBackend(CeedBasisGetNumNodes(basis, &P)); 512 CeedCallBackend(CeedBasisGetNumQuadraturePoints(basis, &Q)); 513 CeedInt thread = CeedIntMax(Q, P); 514 void *interp_args[] = {(void *)&num_elem, &data->d_interp_1d, &d_u, &d_v}; 515 516 { 517 CeedInt elems_per_block = 64 * thread > 256 ? 256 / thread : 64; 518 elems_per_block = elems_per_block > 0 ? elems_per_block : 1; 519 CeedInt grid = num_elem / elems_per_block + (num_elem % elems_per_block > 0); 520 CeedInt shared_mem = elems_per_block * thread * sizeof(CeedScalar); 521 522 if (t_mode == CEED_TRANSPOSE) { 523 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, apply_add ? data->InterpTransposeAdd : data->InterpTranspose, grid, thread, 1, 524 elems_per_block, shared_mem, interp_args)); 525 } else { 526 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, data->Interp, grid, thread, 1, elems_per_block, shared_mem, interp_args)); 527 } 528 } 529 } break; 530 case CEED_EVAL_GRAD: { 531 CeedInt P, Q; 532 533 CeedCheck(data->d_grad_1d, ceed, CEED_ERROR_BACKEND, "%s not supported; grad not set", CeedEvalModes[eval_mode]); 534 CeedCallBackend(CeedBasisGetNumNodes(basis, &P)); 535 CeedCallBackend(CeedBasisGetNumQuadraturePoints(basis, &Q)); 536 CeedInt thread = CeedIntMax(Q, P); 537 void *grad_args[] = {(void *)&num_elem, &data->d_grad_1d, &d_u, &d_v}; 538 539 { 540 CeedInt elems_per_block = 64 * thread > 256 ? 256 / thread : 64; 541 elems_per_block = elems_per_block > 0 ? elems_per_block : 1; 542 CeedInt grid = num_elem / elems_per_block + (num_elem % elems_per_block > 0); 543 CeedInt shared_mem = elems_per_block * thread * sizeof(CeedScalar); 544 545 if (t_mode == CEED_TRANSPOSE) { 546 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, apply_add ? data->GradTransposeAdd : data->GradTranspose, grid, thread, 1, elems_per_block, 547 shared_mem, grad_args)); 548 } else { 549 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, data->Grad, grid, thread, 1, elems_per_block, shared_mem, grad_args)); 550 } 551 } 552 } break; 553 case CEED_EVAL_WEIGHT: { 554 CeedInt P, Q; 555 556 CeedCheck(data->d_q_weight_1d, ceed, CEED_ERROR_BACKEND, "%s not supported; q_weights not set", CeedEvalModes[eval_mode]); 557 CeedCallBackend(CeedBasisGetNumNodes(basis, &P)); 558 CeedCallBackend(CeedBasisGetNumQuadraturePoints(basis, &Q)); 559 CeedInt thread = CeedIntMax(Q, P); 560 void *weight_args[] = {(void *)&num_elem, (void *)&data->d_q_weight_1d, &d_v}; 561 562 { 563 CeedInt elems_per_block = 64 * thread > 256 ? 256 / thread : 64; 564 elems_per_block = elems_per_block > 0 ? elems_per_block : 1; 565 const CeedInt grid_size = num_elem / elems_per_block + (num_elem % elems_per_block > 0); 566 567 CeedCallBackend(CeedRunKernelDim_Hip(ceed, data->Weight, grid_size, thread, elems_per_block, 1, weight_args)); 568 } 569 } break; 570 case CEED_EVAL_NONE: /* handled separately below */ 571 break; 572 // LCOV_EXCL_START 573 case CEED_EVAL_DIV: 574 case CEED_EVAL_CURL: 575 return CeedError(ceed, CEED_ERROR_BACKEND, "%s not supported", CeedEvalModes[eval_mode]); 576 // LCOV_EXCL_STOP 577 } 578 579 // Restore vectors, cover CEED_EVAL_NONE 580 CeedCallBackend(CeedVectorRestoreArray(v, &d_v)); 581 if (eval_mode == CEED_EVAL_NONE) CeedCallBackend(CeedVectorSetArray(v, CEED_MEM_DEVICE, CEED_COPY_VALUES, (CeedScalar *)d_u)); 582 if (eval_mode != CEED_EVAL_WEIGHT) CeedCallBackend(CeedVectorRestoreArrayRead(u, &d_u)); 583 CeedCallBackend(CeedDestroy(&ceed)); 584 return CEED_ERROR_SUCCESS; 585 } 586 587 int CeedBasisApplyNonTensor_Hip_shared(CeedBasis basis, const CeedInt num_elem, CeedTransposeMode t_mode, CeedEvalMode eval_mode, CeedVector u, 588 CeedVector v) { 589 CeedCallBackend(CeedBasisApplyNonTensorCore_Hip_shared(basis, false, num_elem, t_mode, eval_mode, u, v)); 590 return CEED_ERROR_SUCCESS; 591 } 592 593 int CeedBasisApplyAddNonTensor_Hip_shared(CeedBasis basis, const CeedInt num_elem, CeedTransposeMode t_mode, CeedEvalMode eval_mode, CeedVector u, 594 CeedVector v) { 595 CeedCallBackend(CeedBasisApplyNonTensorCore_Hip_shared(basis, true, num_elem, t_mode, eval_mode, u, v)); 596 return CEED_ERROR_SUCCESS; 597 } 598 599 //------------------------------------------------------------------------------ 600 // Destroy basis 601 //------------------------------------------------------------------------------ 602 static int CeedBasisDestroy_Hip_shared(CeedBasis basis) { 603 Ceed ceed; 604 CeedBasis_Hip_shared *data; 605 606 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 607 CeedCallBackend(CeedBasisGetData(basis, &data)); 608 CeedCallHip(ceed, hipModuleUnload(data->module)); 609 if (data->moduleAtPoints) CeedCallHip(ceed, hipModuleUnload(data->moduleAtPoints)); 610 if (data->d_q_weight_1d) CeedCallHip(ceed, hipFree(data->d_q_weight_1d)); 611 CeedCallBackend(CeedFree(&data->h_points_per_elem)); 612 if (data->d_points_per_elem) CeedCallHip(ceed, hipFree(data->d_points_per_elem)); 613 CeedCallHip(ceed, hipFree(data->d_interp_1d)); 614 CeedCallHip(ceed, hipFree(data->d_grad_1d)); 615 CeedCallHip(ceed, hipFree(data->d_collo_grad_1d)); 616 CeedCallHip(ceed, hipFree(data->d_chebyshev_interp_1d)); 617 CeedCallBackend(CeedFree(&data)); 618 return CEED_ERROR_SUCCESS; 619 } 620 621 //------------------------------------------------------------------------------ 622 // Create tensor basis 623 //------------------------------------------------------------------------------ 624 int CeedBasisCreateTensorH1_Hip_shared(CeedInt dim, CeedInt P_1d, CeedInt Q_1d, const CeedScalar *interp_1d, const CeedScalar *grad_1d, 625 const CeedScalar *q_ref_1d, const CeedScalar *q_weight_1d, CeedBasis basis) { 626 Ceed ceed; 627 CeedInt num_comp; 628 const CeedInt q_bytes = Q_1d * sizeof(CeedScalar); 629 const CeedInt interp_bytes = q_bytes * P_1d; 630 CeedBasis_Hip_shared *data; 631 632 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 633 CeedCallBackend(CeedCalloc(1, &data)); 634 635 // Copy basis data to GPU 636 if (q_weight_1d) { 637 CeedCallHip(ceed, hipMalloc((void **)&data->d_q_weight_1d, q_bytes)); 638 CeedCallHip(ceed, hipMemcpy(data->d_q_weight_1d, q_weight_1d, q_bytes, hipMemcpyHostToDevice)); 639 } 640 CeedCallHip(ceed, hipMalloc((void **)&data->d_interp_1d, interp_bytes)); 641 CeedCallHip(ceed, hipMemcpy(data->d_interp_1d, interp_1d, interp_bytes, hipMemcpyHostToDevice)); 642 CeedCallHip(ceed, hipMalloc((void **)&data->d_grad_1d, interp_bytes)); 643 CeedCallHip(ceed, hipMemcpy(data->d_grad_1d, grad_1d, interp_bytes, hipMemcpyHostToDevice)); 644 645 // Compute collocated gradient and copy to GPU 646 data->d_collo_grad_1d = NULL; 647 bool has_collocated_grad = dim == 3 && Q_1d >= P_1d; 648 649 if (has_collocated_grad) { 650 CeedScalar *collo_grad_1d; 651 652 CeedCallBackend(CeedMalloc(Q_1d * Q_1d, &collo_grad_1d)); 653 CeedCallBackend(CeedBasisGetCollocatedGrad(basis, collo_grad_1d)); 654 CeedCallHip(ceed, hipMalloc((void **)&data->d_collo_grad_1d, q_bytes * Q_1d)); 655 CeedCallHip(ceed, hipMemcpy(data->d_collo_grad_1d, collo_grad_1d, q_bytes * Q_1d, hipMemcpyHostToDevice)); 656 CeedCallBackend(CeedFree(&collo_grad_1d)); 657 } 658 659 // Set number of threads per block for basis kernels 660 CeedCallBackend(CeedBasisGetNumComponents(basis, &num_comp)); 661 CeedCallBackend(ComputeBasisThreadBlockSizes(dim, P_1d, Q_1d, num_comp, data->block_sizes)); 662 663 // Compile basis kernels 664 const char basis_kernel_source[] = "// Tensor basis source\n#include <ceed/jit-source/hip/hip-shared-basis-tensor.h>\n"; 665 666 CeedCallBackend(CeedCompile_Hip(ceed, basis_kernel_source, &data->module, 11, "BASIS_Q_1D", Q_1d, "BASIS_P_1D", P_1d, "T_1D", 667 CeedIntMax(Q_1d, P_1d), "BASIS_DIM", dim, "BASIS_NUM_COMP", num_comp, "BASIS_NUM_NODES", CeedIntPow(P_1d, dim), 668 "BASIS_NUM_QPTS", CeedIntPow(Q_1d, dim), "BASIS_INTERP_BLOCK_SIZE", data->block_sizes[0], "BASIS_GRAD_BLOCK_SIZE", 669 data->block_sizes[1], "BASIS_WEIGHT_BLOCK_SIZE", data->block_sizes[2], "BASIS_HAS_COLLOCATED_GRAD", 670 has_collocated_grad)); 671 CeedCallBackend(CeedGetKernel_Hip(ceed, data->module, "Interp", &data->Interp)); 672 CeedCallBackend(CeedGetKernel_Hip(ceed, data->module, "InterpTranspose", &data->InterpTranspose)); 673 CeedCallBackend(CeedGetKernel_Hip(ceed, data->module, "InterpTransposeAdd", &data->InterpTransposeAdd)); 674 CeedCallBackend(CeedGetKernel_Hip(ceed, data->module, "Grad", &data->Grad)); 675 CeedCallBackend(CeedGetKernel_Hip(ceed, data->module, "GradTranspose", &data->GradTranspose)); 676 CeedCallBackend(CeedGetKernel_Hip(ceed, data->module, "GradTransposeAdd", &data->GradTransposeAdd)); 677 CeedCallBackend(CeedGetKernel_Hip(ceed, data->module, "Weight", &data->Weight)); 678 679 CeedCallBackend(CeedBasisSetData(basis, data)); 680 681 // Register backend functions 682 CeedCallBackend(CeedSetBackendFunction(ceed, "Basis", basis, "Apply", CeedBasisApplyTensor_Hip_shared)); 683 CeedCallBackend(CeedSetBackendFunction(ceed, "Basis", basis, "ApplyAdd", CeedBasisApplyAddTensor_Hip_shared)); 684 CeedCallBackend(CeedSetBackendFunction(ceed, "Basis", basis, "ApplyAtPoints", CeedBasisApplyAtPoints_Hip_shared)); 685 CeedCallBackend(CeedSetBackendFunction(ceed, "Basis", basis, "ApplyAddAtPoints", CeedBasisApplyAddAtPoints_Hip_shared)); 686 CeedCallBackend(CeedSetBackendFunction(ceed, "Basis", basis, "Destroy", CeedBasisDestroy_Hip_shared)); 687 CeedCallBackend(CeedDestroy(&ceed)); 688 return CEED_ERROR_SUCCESS; 689 } 690 691 //------------------------------------------------------------------------------ 692 // Create non-tensor basis 693 //------------------------------------------------------------------------------ 694 int CeedBasisCreateH1_Hip_shared(CeedElemTopology topo, CeedInt dim, CeedInt num_nodes, CeedInt num_qpts, const CeedScalar *interp, 695 const CeedScalar *grad, const CeedScalar *q_ref, const CeedScalar *q_weight, CeedBasis basis) { 696 Ceed ceed; 697 CeedInt num_comp, q_comp_interp, q_comp_grad; 698 const CeedInt q_bytes = num_qpts * sizeof(CeedScalar); 699 CeedBasis_Hip_shared *data; 700 701 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 702 703 // Check shared memory size 704 { 705 Ceed_Hip *hip_data; 706 707 CeedCallBackend(CeedGetData(ceed, &hip_data)); 708 if (((size_t)num_nodes * (size_t)num_qpts * (size_t)dim + (size_t)CeedIntMax(num_nodes, num_qpts)) * sizeof(CeedScalar) > 709 hip_data->device_prop.sharedMemPerBlock) { 710 CeedCallBackend(CeedBasisCreateH1Fallback(ceed, topo, dim, num_nodes, num_qpts, interp, grad, q_ref, q_weight, basis)); 711 return CEED_ERROR_SUCCESS; 712 } 713 } 714 715 CeedCallBackend(CeedCalloc(1, &data)); 716 717 // Copy basis data to GPU 718 CeedCallBackend(CeedBasisGetNumQuadratureComponents(basis, CEED_EVAL_INTERP, &q_comp_interp)); 719 CeedCallBackend(CeedBasisGetNumQuadratureComponents(basis, CEED_EVAL_GRAD, &q_comp_grad)); 720 if (q_weight) { 721 CeedCallHip(ceed, hipMalloc((void **)&data->d_q_weight_1d, q_bytes)); 722 CeedCallHip(ceed, hipMemcpy(data->d_q_weight_1d, q_weight, q_bytes, hipMemcpyHostToDevice)); 723 } 724 if (interp) { 725 const CeedInt interp_bytes = q_bytes * num_nodes * q_comp_interp; 726 727 CeedCallHip(ceed, hipMalloc((void **)&data->d_interp_1d, interp_bytes)); 728 CeedCallHip(ceed, hipMemcpy(data->d_interp_1d, interp, interp_bytes, hipMemcpyHostToDevice)); 729 } 730 if (grad) { 731 const CeedInt grad_bytes = q_bytes * num_nodes * q_comp_grad; 732 733 CeedCallHip(ceed, hipMalloc((void **)&data->d_grad_1d, grad_bytes)); 734 CeedCallHip(ceed, hipMemcpy(data->d_grad_1d, grad, grad_bytes, hipMemcpyHostToDevice)); 735 } 736 737 // Compile basis kernels 738 const char basis_kernel_source[] = "// Non-tensor basis source\n#include <ceed/jit-source/hip/hip-shared-basis-nontensor.h>\n"; 739 740 CeedCallBackend(CeedBasisGetNumComponents(basis, &num_comp)); 741 CeedCallBackend(ComputeBasisThreadBlockSizes(dim, num_nodes, num_qpts, num_comp, data->block_sizes)); 742 CeedCallBackend(CeedCompile_Hip(ceed, basis_kernel_source, &data->module, 6, "BASIS_Q", num_qpts, "BASIS_P", num_nodes, "T_1D", 743 CeedIntMax(num_qpts, num_nodes), "BASIS_DIM", dim, "BASIS_NUM_COMP", num_comp, "BASIS_INTERP_BLOCK_SIZE", 744 data->block_sizes[0])); 745 CeedCallBackend(CeedGetKernel_Hip(ceed, data->module, "Interp", &data->Interp)); 746 CeedCallBackend(CeedGetKernel_Hip(ceed, data->module, "InterpTranspose", &data->InterpTranspose)); 747 CeedCallBackend(CeedGetKernel_Hip(ceed, data->module, "InterpTransposeAdd", &data->InterpTransposeAdd)); 748 CeedCallBackend(CeedGetKernel_Hip(ceed, data->module, "Grad", &data->Grad)); 749 CeedCallBackend(CeedGetKernel_Hip(ceed, data->module, "GradTranspose", &data->GradTranspose)); 750 CeedCallBackend(CeedGetKernel_Hip(ceed, data->module, "GradTransposeAdd", &data->GradTransposeAdd)); 751 CeedCallBackend(CeedGetKernel_Hip(ceed, data->module, "Weight", &data->Weight)); 752 753 CeedCallBackend(CeedBasisSetData(basis, data)); 754 755 // Register backend functions 756 CeedCallBackend(CeedSetBackendFunction(ceed, "Basis", basis, "Apply", CeedBasisApplyNonTensor_Hip_shared)); 757 CeedCallBackend(CeedSetBackendFunction(ceed, "Basis", basis, "ApplyAdd", CeedBasisApplyAddNonTensor_Hip_shared)); 758 CeedCallBackend(CeedSetBackendFunction(ceed, "Basis", basis, "Destroy", CeedBasisDestroy_Hip_shared)); 759 CeedCallBackend(CeedDestroy(&ceed)); 760 return CEED_ERROR_SUCCESS; 761 } 762 763 //------------------------------------------------------------------------------ 764