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) { 110 CeedCallBackend(CeedVectorGetArray(v, CEED_MEM_DEVICE, &d_v)); 111 } else { 112 CeedCallBackend(CeedVectorGetArrayWrite(v, CEED_MEM_DEVICE, &d_v)); 113 } 114 115 // Apply basis operation 116 switch (eval_mode) { 117 case CEED_EVAL_INTERP: { 118 CeedInt P_1d, Q_1d; 119 CeedInt block_size = data->block_sizes[0]; 120 121 CeedCheck(data->d_interp_1d, ceed, CEED_ERROR_BACKEND, "%s not supported; interp_1d not set", CeedEvalModes[eval_mode]); 122 CeedCallBackend(CeedBasisGetNumNodes1D(basis, &P_1d)); 123 CeedCallBackend(CeedBasisGetNumQuadraturePoints1D(basis, &Q_1d)); 124 CeedInt thread_1d = CeedIntMax(Q_1d, P_1d); 125 void *interp_args[] = {(void *)&num_elem, &data->d_interp_1d, &d_u, &d_v}; 126 127 if (dim == 1) { 128 CeedInt elems_per_block = 64 * thread_1d > 256 ? 256 / thread_1d : 64; 129 elems_per_block = elems_per_block > 0 ? elems_per_block : 1; 130 CeedInt grid = num_elem / elems_per_block + (num_elem % elems_per_block > 0); 131 CeedInt shared_mem = elems_per_block * thread_1d * sizeof(CeedScalar); 132 133 if (t_mode == CEED_TRANSPOSE) { 134 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, apply_add ? data->InterpTransposeAdd : data->InterpTranspose, grid, thread_1d, 1, 135 elems_per_block, shared_mem, interp_args)); 136 } else { 137 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, data->Interp, grid, thread_1d, 1, elems_per_block, shared_mem, interp_args)); 138 } 139 } else if (dim == 2) { 140 // Check if required threads is small enough to do multiple elems 141 const CeedInt elems_per_block = CeedIntMax(block_size / (thread_1d * thread_1d), 1); 142 CeedInt grid = num_elem / elems_per_block + (num_elem % elems_per_block > 0); 143 CeedInt shared_mem = elems_per_block * thread_1d * thread_1d * sizeof(CeedScalar); 144 145 if (t_mode == CEED_TRANSPOSE) { 146 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, apply_add ? data->InterpTransposeAdd : data->InterpTranspose, grid, thread_1d, thread_1d, 147 elems_per_block, shared_mem, interp_args)); 148 } else { 149 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, data->Interp, grid, thread_1d, thread_1d, elems_per_block, shared_mem, interp_args)); 150 } 151 } else if (dim == 3) { 152 const CeedInt elems_per_block = CeedIntMax(block_size / (thread_1d * thread_1d), 1); 153 CeedInt grid = num_elem / elems_per_block + (num_elem % elems_per_block > 0); 154 CeedInt shared_mem = elems_per_block * thread_1d * thread_1d * sizeof(CeedScalar); 155 156 if (t_mode == CEED_TRANSPOSE) { 157 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, apply_add ? data->InterpTransposeAdd : data->InterpTranspose, grid, thread_1d, thread_1d, 158 elems_per_block, shared_mem, interp_args)); 159 } else { 160 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, data->Interp, grid, thread_1d, thread_1d, elems_per_block, shared_mem, interp_args)); 161 } 162 } 163 } break; 164 case CEED_EVAL_GRAD: { 165 CeedInt P_1d, Q_1d; 166 CeedInt block_size = data->block_sizes[1]; 167 168 CeedCheck(data->d_grad_1d, ceed, CEED_ERROR_BACKEND, "%s not supported; grad_1d not set", CeedEvalModes[eval_mode]); 169 CeedCallBackend(CeedBasisGetNumNodes1D(basis, &P_1d)); 170 CeedCallBackend(CeedBasisGetNumQuadraturePoints1D(basis, &Q_1d)); 171 CeedInt thread_1d = CeedIntMax(Q_1d, P_1d); 172 CeedScalar *d_grad_1d = data->d_grad_1d; 173 174 if (data->d_collo_grad_1d) { 175 d_grad_1d = data->d_collo_grad_1d; 176 } 177 void *grad_args[] = {(void *)&num_elem, &data->d_interp_1d, &d_grad_1d, &d_u, &d_v}; 178 179 if (dim == 1) { 180 CeedInt elems_per_block = 64 * thread_1d > 256 ? 256 / thread_1d : 64; 181 elems_per_block = elems_per_block > 0 ? elems_per_block : 1; 182 CeedInt grid = num_elem / elems_per_block + (num_elem % elems_per_block > 0); 183 CeedInt shared_mem = elems_per_block * thread_1d * sizeof(CeedScalar); 184 185 if (t_mode == CEED_TRANSPOSE) { 186 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, apply_add ? data->GradTransposeAdd : data->GradTranspose, grid, thread_1d, 1, 187 elems_per_block, shared_mem, grad_args)); 188 } else { 189 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, data->Grad, grid, thread_1d, 1, elems_per_block, shared_mem, grad_args)); 190 } 191 } else if (dim == 2) { 192 // Check if required threads is small enough to do multiple elems 193 const CeedInt elems_per_block = CeedIntMax(block_size / (thread_1d * thread_1d), 1); 194 CeedInt grid = num_elem / elems_per_block + (num_elem % elems_per_block > 0); 195 CeedInt shared_mem = elems_per_block * thread_1d * thread_1d * sizeof(CeedScalar); 196 197 if (t_mode == CEED_TRANSPOSE) { 198 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, apply_add ? data->GradTransposeAdd : data->GradTranspose, grid, thread_1d, thread_1d, 199 elems_per_block, shared_mem, grad_args)); 200 } else { 201 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, data->Grad, grid, thread_1d, thread_1d, elems_per_block, shared_mem, grad_args)); 202 } 203 } else if (dim == 3) { 204 const CeedInt elems_per_block = CeedIntMax(block_size / (thread_1d * thread_1d), 1); 205 CeedInt grid = num_elem / elems_per_block + (num_elem % elems_per_block > 0); 206 CeedInt shared_mem = elems_per_block * thread_1d * thread_1d * sizeof(CeedScalar); 207 208 if (t_mode == CEED_TRANSPOSE) { 209 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, apply_add ? data->GradTransposeAdd : data->GradTranspose, grid, thread_1d, thread_1d, 210 elems_per_block, shared_mem, grad_args)); 211 } else { 212 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, data->Grad, grid, thread_1d, thread_1d, elems_per_block, shared_mem, grad_args)); 213 } 214 } 215 } break; 216 case CEED_EVAL_WEIGHT: { 217 CeedInt Q_1d; 218 CeedInt block_size = data->block_sizes[2]; 219 220 CeedCheck(data->d_q_weight_1d, ceed, CEED_ERROR_BACKEND, "%s not supported; q_weights_1d not set", CeedEvalModes[eval_mode]); 221 CeedCallBackend(CeedBasisGetNumQuadraturePoints1D(basis, &Q_1d)); 222 void *weight_args[] = {(void *)&num_elem, (void *)&data->d_q_weight_1d, &d_v}; 223 224 if (dim == 1) { 225 const CeedInt opt_elems = block_size / Q_1d; 226 const CeedInt elems_per_block = opt_elems > 0 ? opt_elems : 1; 227 const CeedInt grid_size = num_elem / elems_per_block + (num_elem % elems_per_block > 0); 228 229 CeedCallBackend(CeedRunKernelDim_Hip(ceed, data->Weight, grid_size, Q_1d, elems_per_block, 1, weight_args)); 230 } else if (dim == 2) { 231 const CeedInt opt_elems = block_size / (Q_1d * Q_1d); 232 const CeedInt elems_per_block = opt_elems > 0 ? opt_elems : 1; 233 const CeedInt grid_size = num_elem / elems_per_block + (num_elem % elems_per_block > 0); 234 235 CeedCallBackend(CeedRunKernelDim_Hip(ceed, data->Weight, grid_size, Q_1d, Q_1d, elems_per_block, weight_args)); 236 } else if (dim == 3) { 237 const CeedInt opt_elems = block_size / (Q_1d * Q_1d); 238 const CeedInt elems_per_block = opt_elems > 0 ? opt_elems : 1; 239 const CeedInt grid_size = num_elem / elems_per_block + (num_elem % elems_per_block > 0); 240 241 CeedCallBackend(CeedRunKernelDim_Hip(ceed, data->Weight, grid_size, Q_1d, Q_1d, elems_per_block, weight_args)); 242 } 243 } break; 244 case CEED_EVAL_NONE: /* handled separately below */ 245 break; 246 // LCOV_EXCL_START 247 case CEED_EVAL_DIV: 248 case CEED_EVAL_CURL: 249 return CeedError(ceed, CEED_ERROR_BACKEND, "%s not supported", CeedEvalModes[eval_mode]); 250 // LCOV_EXCL_STOP 251 } 252 253 // Restore vectors, cover CEED_EVAL_NONE 254 CeedCallBackend(CeedVectorRestoreArray(v, &d_v)); 255 if (eval_mode == CEED_EVAL_NONE) CeedCallBackend(CeedVectorSetArray(v, CEED_MEM_DEVICE, CEED_COPY_VALUES, (CeedScalar *)d_u)); 256 if (eval_mode != CEED_EVAL_WEIGHT) CeedCallBackend(CeedVectorRestoreArrayRead(u, &d_u)); 257 CeedCallBackend(CeedDestroy(&ceed)); 258 return CEED_ERROR_SUCCESS; 259 } 260 261 int CeedBasisApplyTensor_Hip_shared(CeedBasis basis, const CeedInt num_elem, CeedTransposeMode t_mode, CeedEvalMode eval_mode, CeedVector u, 262 CeedVector v) { 263 CeedCallBackend(CeedBasisApplyTensorCore_Hip_shared(basis, false, num_elem, t_mode, eval_mode, u, v)); 264 return CEED_ERROR_SUCCESS; 265 } 266 267 int CeedBasisApplyAddTensor_Hip_shared(CeedBasis basis, const CeedInt num_elem, CeedTransposeMode t_mode, CeedEvalMode eval_mode, CeedVector u, 268 CeedVector v) { 269 CeedCallBackend(CeedBasisApplyTensorCore_Hip_shared(basis, true, num_elem, t_mode, eval_mode, u, v)); 270 return CEED_ERROR_SUCCESS; 271 } 272 273 //------------------------------------------------------------------------------ 274 // Basis apply - tensor AtPoints 275 //------------------------------------------------------------------------------ 276 static int CeedBasisApplyAtPointsCore_Hip_shared(CeedBasis basis, bool apply_add, const CeedInt num_elem, const CeedInt *num_points, 277 CeedTransposeMode t_mode, CeedEvalMode eval_mode, CeedVector x_ref, CeedVector u, CeedVector v) { 278 Ceed ceed; 279 CeedInt Q_1d, dim, max_num_points = num_points[0]; 280 const CeedInt is_transpose = t_mode == CEED_TRANSPOSE; 281 const CeedScalar *d_x, *d_u; 282 CeedScalar *d_v; 283 CeedBasis_Hip_shared *data; 284 285 CeedCallBackend(CeedBasisGetData(basis, &data)); 286 CeedCallBackend(CeedBasisGetNumQuadraturePoints1D(basis, &Q_1d)); 287 CeedCallBackend(CeedBasisGetDimension(basis, &dim)); 288 289 // Weight handled separately 290 if (eval_mode == CEED_EVAL_WEIGHT) { 291 CeedCallBackend(CeedVectorSetValue(v, 1.0)); 292 return CEED_ERROR_SUCCESS; 293 } 294 295 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 296 297 // Check padded to uniform number of points per elem 298 for (CeedInt i = 1; i < num_elem; i++) max_num_points = CeedIntMax(max_num_points, num_points[i]); 299 { 300 CeedInt num_comp, q_comp; 301 CeedSize len, len_required; 302 303 CeedCallBackend(CeedBasisGetNumComponents(basis, &num_comp)); 304 CeedCallBackend(CeedBasisGetNumQuadratureComponents(basis, eval_mode, &q_comp)); 305 CeedCallBackend(CeedVectorGetLength(is_transpose ? u : v, &len)); 306 len_required = (CeedSize)num_comp * (CeedSize)q_comp * (CeedSize)num_elem * (CeedSize)max_num_points; 307 CeedCheck(len >= len_required, ceed, CEED_ERROR_BACKEND, 308 "Vector at points must be padded to the same number of points in each element for BasisApplyAtPoints on GPU backends." 309 " Found %" CeedSize_FMT ", Required %" CeedSize_FMT, 310 len, len_required); 311 } 312 313 // Move num_points array to device 314 if (is_transpose) { 315 const CeedInt num_bytes = num_elem * sizeof(CeedInt); 316 317 if (num_elem != data->num_elem_at_points) { 318 data->num_elem_at_points = num_elem; 319 320 if (data->d_points_per_elem) CeedCallHip(ceed, hipFree(data->d_points_per_elem)); 321 CeedCallHip(ceed, hipMalloc((void **)&data->d_points_per_elem, num_bytes)); 322 CeedCallBackend(CeedFree(&data->h_points_per_elem)); 323 CeedCallBackend(CeedCalloc(num_elem, &data->h_points_per_elem)); 324 } 325 if (memcmp(data->h_points_per_elem, num_points, num_bytes)) { 326 memcpy(data->h_points_per_elem, num_points, num_bytes); 327 CeedCallHip(ceed, hipMemcpy(data->d_points_per_elem, num_points, num_bytes, hipMemcpyHostToDevice)); 328 } 329 } 330 331 // Build kernels if needed 332 if (data->num_points != max_num_points) { 333 CeedInt P_1d; 334 335 CeedCallBackend(CeedBasisGetNumNodes1D(basis, &P_1d)); 336 data->num_points = max_num_points; 337 338 // -- Create interp matrix to Chebyshev coefficients 339 if (!data->d_chebyshev_interp_1d) { 340 CeedSize interp_bytes; 341 CeedScalar *chebyshev_interp_1d; 342 343 interp_bytes = P_1d * Q_1d * sizeof(CeedScalar); 344 CeedCallBackend(CeedCalloc(P_1d * Q_1d, &chebyshev_interp_1d)); 345 CeedCallBackend(CeedBasisGetChebyshevInterp1D(basis, chebyshev_interp_1d)); 346 CeedCallHip(ceed, hipMalloc((void **)&data->d_chebyshev_interp_1d, interp_bytes)); 347 CeedCallHip(ceed, hipMemcpy(data->d_chebyshev_interp_1d, chebyshev_interp_1d, interp_bytes, hipMemcpyHostToDevice)); 348 CeedCallBackend(CeedFree(&chebyshev_interp_1d)); 349 } 350 351 // -- Compile kernels 352 const char basis_kernel_source[] = "// AtPoints basis source\n#include <ceed/jit-source/hip/hip-shared-basis-tensor-at-points.h>\n"; 353 CeedInt num_comp; 354 355 if (data->moduleAtPoints) CeedCallHip(ceed, hipModuleUnload(data->moduleAtPoints)); 356 CeedCallBackend(CeedBasisGetNumComponents(basis, &num_comp)); 357 CeedCallBackend(CeedCompile_Hip(ceed, basis_kernel_source, &data->moduleAtPoints, 9, "BASIS_Q_1D", Q_1d, "BASIS_P_1D", P_1d, "T_1D", 358 CeedIntMax(Q_1d, P_1d), "BASIS_DIM", dim, "BASIS_NUM_COMP", num_comp, "BASIS_NUM_NODES", CeedIntPow(P_1d, dim), 359 "BASIS_NUM_QPTS", CeedIntPow(Q_1d, dim), "BASIS_NUM_PTS", max_num_points, "BASIS_INTERP_BLOCK_SIZE", 360 data->block_sizes[0])); 361 CeedCallBackend(CeedGetKernel_Hip(ceed, data->moduleAtPoints, "InterpAtPoints", &data->InterpAtPoints)); 362 CeedCallBackend(CeedGetKernel_Hip(ceed, data->moduleAtPoints, "InterpTransposeAtPoints", &data->InterpTransposeAtPoints)); 363 CeedCallBackend(CeedGetKernel_Hip(ceed, data->moduleAtPoints, "GradAtPoints", &data->GradAtPoints)); 364 CeedCallBackend(CeedGetKernel_Hip(ceed, data->moduleAtPoints, "GradTransposeAtPoints", &data->GradTransposeAtPoints)); 365 } 366 367 // Get read/write access to u, v 368 CeedCallBackend(CeedVectorGetArrayRead(x_ref, CEED_MEM_DEVICE, &d_x)); 369 if (u != CEED_VECTOR_NONE) CeedCallBackend(CeedVectorGetArrayRead(u, CEED_MEM_DEVICE, &d_u)); 370 else CeedCheck(eval_mode == CEED_EVAL_WEIGHT, ceed, CEED_ERROR_BACKEND, "An input vector is required for this CeedEvalMode"); 371 if (apply_add) { 372 CeedCallBackend(CeedVectorGetArray(v, CEED_MEM_DEVICE, &d_v)); 373 } else { 374 CeedCallBackend(CeedVectorGetArrayWrite(v, CEED_MEM_DEVICE, &d_v)); 375 } 376 377 // Basis action 378 switch (eval_mode) { 379 case CEED_EVAL_INTERP: { 380 CeedInt P_1d, Q_1d; 381 CeedInt block_size = data->block_sizes[0]; 382 383 CeedCallBackend(CeedBasisGetNumNodes1D(basis, &P_1d)); 384 CeedCallBackend(CeedBasisGetNumQuadraturePoints1D(basis, &Q_1d)); 385 CeedInt thread_1d = CeedIntMax(Q_1d, P_1d); 386 void *interp_args[] = {(void *)&num_elem, &data->d_chebyshev_interp_1d, &data->d_points_per_elem, &d_x, &d_u, &d_v}; 387 388 if (dim == 1) { 389 CeedInt elems_per_block = 64 * thread_1d > 256 ? 256 / thread_1d : 64; 390 elems_per_block = elems_per_block > 0 ? elems_per_block : 1; 391 CeedInt grid = num_elem / elems_per_block + (num_elem % elems_per_block > 0); 392 CeedInt shared_mem = elems_per_block * thread_1d * sizeof(CeedScalar); 393 394 if (is_transpose) { 395 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, apply_add ? data->InterpTransposeAddAtPoints : data->InterpTransposeAtPoints, grid, 396 thread_1d, 1, elems_per_block, shared_mem, interp_args)); 397 } else { 398 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, data->InterpAtPoints, grid, thread_1d, 1, elems_per_block, shared_mem, interp_args)); 399 } 400 } else if (dim == 2) { 401 // Check if required threads is small enough to do multiple elems 402 const CeedInt elems_per_block = CeedIntMax(block_size / (thread_1d * thread_1d), 1); 403 CeedInt grid = num_elem / elems_per_block + (num_elem % elems_per_block > 0); 404 CeedInt shared_mem = elems_per_block * thread_1d * thread_1d * sizeof(CeedScalar); 405 406 if (is_transpose) { 407 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, apply_add ? data->InterpTransposeAddAtPoints : data->InterpTransposeAtPoints, grid, 408 thread_1d, thread_1d, elems_per_block, shared_mem, interp_args)); 409 } else { 410 CeedCallBackend( 411 CeedRunKernelDimShared_Hip(ceed, data->InterpAtPoints, grid, thread_1d, thread_1d, elems_per_block, shared_mem, interp_args)); 412 } 413 } else if (dim == 3) { 414 const CeedInt elems_per_block = 1; 415 CeedInt grid = num_elem / elems_per_block + (num_elem % elems_per_block > 0); 416 CeedInt shared_mem = elems_per_block * thread_1d * thread_1d * sizeof(CeedScalar); 417 418 if (is_transpose) { 419 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, apply_add ? data->InterpTransposeAddAtPoints : data->InterpTransposeAtPoints, grid, 420 thread_1d, thread_1d, elems_per_block, shared_mem, interp_args)); 421 } else { 422 CeedCallBackend( 423 CeedRunKernelDimShared_Hip(ceed, data->InterpAtPoints, grid, thread_1d, thread_1d, elems_per_block, shared_mem, interp_args)); 424 } 425 } 426 } break; 427 case CEED_EVAL_GRAD: { 428 CeedInt P_1d, Q_1d; 429 CeedInt block_size = data->block_sizes[0]; 430 431 CeedCallBackend(CeedBasisGetNumNodes1D(basis, &P_1d)); 432 CeedCallBackend(CeedBasisGetNumQuadraturePoints1D(basis, &Q_1d)); 433 CeedInt thread_1d = CeedIntMax(Q_1d, P_1d); 434 void *grad_args[] = {(void *)&num_elem, &data->d_chebyshev_interp_1d, &data->d_points_per_elem, &d_x, &d_u, &d_v}; 435 436 if (dim == 1) { 437 CeedInt elems_per_block = 64 * thread_1d > 256 ? 256 / thread_1d : 64; 438 elems_per_block = elems_per_block > 0 ? 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 * sizeof(CeedScalar); 441 442 if (is_transpose) { 443 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, apply_add ? data->GradTransposeAddAtPoints : data->GradTransposeAtPoints, grid, thread_1d, 444 1, elems_per_block, shared_mem, grad_args)); 445 } else { 446 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, data->GradAtPoints, grid, thread_1d, 1, elems_per_block, shared_mem, grad_args)); 447 } 448 } else if (dim == 2) { 449 // Check if required threads is small enough to do multiple elems 450 const CeedInt elems_per_block = CeedIntMax(block_size / (thread_1d * thread_1d), 1); 451 CeedInt grid = num_elem / elems_per_block + (num_elem % elems_per_block > 0); 452 CeedInt shared_mem = elems_per_block * thread_1d * thread_1d * sizeof(CeedScalar); 453 454 if (is_transpose) { 455 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, apply_add ? data->GradTransposeAddAtPoints : data->GradTransposeAtPoints, grid, thread_1d, 456 thread_1d, elems_per_block, shared_mem, grad_args)); 457 } else { 458 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, data->GradAtPoints, grid, thread_1d, thread_1d, elems_per_block, shared_mem, grad_args)); 459 } 460 } else if (dim == 3) { 461 const CeedInt elems_per_block = 1; 462 CeedInt grid = num_elem / elems_per_block + (num_elem % elems_per_block > 0); 463 CeedInt shared_mem = elems_per_block * thread_1d * thread_1d * sizeof(CeedScalar); 464 465 if (is_transpose) { 466 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, apply_add ? data->GradTransposeAddAtPoints : data->GradTransposeAtPoints, grid, thread_1d, 467 thread_1d, elems_per_block, shared_mem, grad_args)); 468 } else { 469 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, data->GradAtPoints, grid, thread_1d, thread_1d, elems_per_block, shared_mem, grad_args)); 470 } 471 } 472 } break; 473 case CEED_EVAL_WEIGHT: 474 case CEED_EVAL_NONE: /* handled separately below */ 475 break; 476 // LCOV_EXCL_START 477 case CEED_EVAL_DIV: 478 case CEED_EVAL_CURL: 479 return CeedError(ceed, CEED_ERROR_BACKEND, "%s not supported", CeedEvalModes[eval_mode]); 480 // LCOV_EXCL_STOP 481 } 482 483 // Restore vectors, cover CEED_EVAL_NONE 484 CeedCallBackend(CeedVectorRestoreArrayRead(x_ref, &d_x)); 485 CeedCallBackend(CeedVectorRestoreArray(v, &d_v)); 486 if (eval_mode == CEED_EVAL_NONE) CeedCallBackend(CeedVectorSetArray(v, CEED_MEM_DEVICE, CEED_COPY_VALUES, (CeedScalar *)d_u)); 487 if (eval_mode != CEED_EVAL_WEIGHT) CeedCallBackend(CeedVectorRestoreArrayRead(u, &d_u)); 488 return CEED_ERROR_SUCCESS; 489 } 490 491 static int CeedBasisApplyAtPoints_Hip_shared(CeedBasis basis, const CeedInt num_elem, const CeedInt *num_points, CeedTransposeMode t_mode, 492 CeedEvalMode eval_mode, CeedVector x_ref, CeedVector u, CeedVector v) { 493 CeedCallBackend(CeedBasisApplyAtPointsCore_Hip_shared(basis, false, num_elem, num_points, t_mode, eval_mode, x_ref, u, v)); 494 return CEED_ERROR_SUCCESS; 495 } 496 497 static int CeedBasisApplyAddAtPoints_Hip_shared(CeedBasis basis, const CeedInt num_elem, const CeedInt *num_points, CeedTransposeMode t_mode, 498 CeedEvalMode eval_mode, CeedVector x_ref, CeedVector u, CeedVector v) { 499 CeedCallBackend(CeedBasisApplyAtPointsCore_Hip_shared(basis, true, num_elem, num_points, t_mode, eval_mode, x_ref, u, v)); 500 return CEED_ERROR_SUCCESS; 501 } 502 503 //------------------------------------------------------------------------------ 504 // Apply basis 505 //------------------------------------------------------------------------------ 506 static int CeedBasisApplyNonTensorCore_Hip_shared(CeedBasis basis, bool apply_add, const CeedInt num_elem, CeedTransposeMode t_mode, 507 CeedEvalMode eval_mode, CeedVector u, CeedVector v) { 508 Ceed ceed; 509 Ceed_Hip *ceed_Hip; 510 CeedInt dim, num_comp; 511 const CeedScalar *d_u; 512 CeedScalar *d_v; 513 CeedBasis_Hip_shared *data; 514 515 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 516 CeedCallBackend(CeedGetData(ceed, &ceed_Hip)); 517 CeedCallBackend(CeedBasisGetData(basis, &data)); 518 CeedCallBackend(CeedBasisGetDimension(basis, &dim)); 519 CeedCallBackend(CeedBasisGetNumComponents(basis, &num_comp)); 520 521 // Get read/write access to u, v 522 if (u != CEED_VECTOR_NONE) CeedCallBackend(CeedVectorGetArrayRead(u, CEED_MEM_DEVICE, &d_u)); 523 else CeedCheck(eval_mode == CEED_EVAL_WEIGHT, ceed, CEED_ERROR_BACKEND, "An input vector is required for this CeedEvalMode"); 524 if (apply_add) { 525 CeedCallBackend(CeedVectorGetArray(v, CEED_MEM_DEVICE, &d_v)); 526 } else { 527 CeedCallBackend(CeedVectorGetArrayWrite(v, CEED_MEM_DEVICE, &d_v)); 528 } 529 530 // Apply basis operation 531 switch (eval_mode) { 532 case CEED_EVAL_INTERP: { 533 CeedInt P, Q; 534 535 CeedCheck(data->d_interp_1d, ceed, CEED_ERROR_BACKEND, "%s not supported; interp not set", CeedEvalModes[eval_mode]); 536 CeedCallBackend(CeedBasisGetNumNodes(basis, &P)); 537 CeedCallBackend(CeedBasisGetNumQuadraturePoints(basis, &Q)); 538 CeedInt thread = CeedIntMax(Q, P); 539 void *interp_args[] = {(void *)&num_elem, &data->d_interp_1d, &d_u, &d_v}; 540 541 { 542 CeedInt elems_per_block = 64 * thread > 256 ? 256 / thread : 64; 543 elems_per_block = elems_per_block > 0 ? elems_per_block : 1; 544 CeedInt grid = num_elem / elems_per_block + (num_elem % elems_per_block > 0); 545 CeedInt shared_mem = elems_per_block * thread * sizeof(CeedScalar); 546 547 if (t_mode == CEED_TRANSPOSE) { 548 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, apply_add ? data->InterpTransposeAdd : data->InterpTranspose, grid, thread, 1, 549 elems_per_block, shared_mem, interp_args)); 550 } else { 551 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, data->Interp, grid, thread, 1, elems_per_block, shared_mem, interp_args)); 552 } 553 } 554 } break; 555 case CEED_EVAL_GRAD: { 556 CeedInt P, Q; 557 558 CeedCheck(data->d_grad_1d, ceed, CEED_ERROR_BACKEND, "%s not supported; grad not set", CeedEvalModes[eval_mode]); 559 CeedCallBackend(CeedBasisGetNumNodes(basis, &P)); 560 CeedCallBackend(CeedBasisGetNumQuadraturePoints(basis, &Q)); 561 CeedInt thread = CeedIntMax(Q, P); 562 void *grad_args[] = {(void *)&num_elem, &data->d_grad_1d, &d_u, &d_v}; 563 564 { 565 CeedInt elems_per_block = 64 * thread > 256 ? 256 / thread : 64; 566 elems_per_block = elems_per_block > 0 ? elems_per_block : 1; 567 CeedInt grid = num_elem / elems_per_block + (num_elem % elems_per_block > 0); 568 CeedInt shared_mem = elems_per_block * thread * sizeof(CeedScalar); 569 570 if (t_mode == CEED_TRANSPOSE) { 571 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, apply_add ? data->GradTransposeAdd : data->GradTranspose, grid, thread, 1, elems_per_block, 572 shared_mem, grad_args)); 573 } else { 574 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, data->Grad, grid, thread, 1, elems_per_block, shared_mem, grad_args)); 575 } 576 } 577 } break; 578 case CEED_EVAL_WEIGHT: { 579 CeedInt P, Q; 580 581 CeedCheck(data->d_q_weight_1d, ceed, CEED_ERROR_BACKEND, "%s not supported; q_weights not set", CeedEvalModes[eval_mode]); 582 CeedCallBackend(CeedBasisGetNumNodes(basis, &P)); 583 CeedCallBackend(CeedBasisGetNumQuadraturePoints(basis, &Q)); 584 CeedInt thread = CeedIntMax(Q, P); 585 void *weight_args[] = {(void *)&num_elem, (void *)&data->d_q_weight_1d, &d_v}; 586 587 { 588 CeedInt elems_per_block = 64 * thread > 256 ? 256 / thread : 64; 589 elems_per_block = elems_per_block > 0 ? elems_per_block : 1; 590 const CeedInt grid_size = num_elem / elems_per_block + (num_elem % elems_per_block > 0); 591 592 CeedCallBackend(CeedRunKernelDim_Hip(ceed, data->Weight, grid_size, thread, elems_per_block, 1, weight_args)); 593 } 594 } break; 595 case CEED_EVAL_NONE: /* handled separately below */ 596 break; 597 // LCOV_EXCL_START 598 case CEED_EVAL_DIV: 599 case CEED_EVAL_CURL: 600 return CeedError(ceed, CEED_ERROR_BACKEND, "%s not supported", CeedEvalModes[eval_mode]); 601 // LCOV_EXCL_STOP 602 } 603 604 // Restore vectors, cover CEED_EVAL_NONE 605 CeedCallBackend(CeedVectorRestoreArray(v, &d_v)); 606 if (eval_mode == CEED_EVAL_NONE) CeedCallBackend(CeedVectorSetArray(v, CEED_MEM_DEVICE, CEED_COPY_VALUES, (CeedScalar *)d_u)); 607 if (eval_mode != CEED_EVAL_WEIGHT) CeedCallBackend(CeedVectorRestoreArrayRead(u, &d_u)); 608 CeedCallBackend(CeedDestroy(&ceed)); 609 return CEED_ERROR_SUCCESS; 610 } 611 612 int CeedBasisApplyNonTensor_Hip_shared(CeedBasis basis, const CeedInt num_elem, CeedTransposeMode t_mode, CeedEvalMode eval_mode, CeedVector u, 613 CeedVector v) { 614 CeedCallBackend(CeedBasisApplyNonTensorCore_Hip_shared(basis, false, num_elem, t_mode, eval_mode, u, v)); 615 return CEED_ERROR_SUCCESS; 616 } 617 618 int CeedBasisApplyAddNonTensor_Hip_shared(CeedBasis basis, const CeedInt num_elem, CeedTransposeMode t_mode, CeedEvalMode eval_mode, CeedVector u, 619 CeedVector v) { 620 CeedCallBackend(CeedBasisApplyNonTensorCore_Hip_shared(basis, true, num_elem, t_mode, eval_mode, u, v)); 621 return CEED_ERROR_SUCCESS; 622 } 623 624 //------------------------------------------------------------------------------ 625 // Destroy basis 626 //------------------------------------------------------------------------------ 627 static int CeedBasisDestroy_Hip_shared(CeedBasis basis) { 628 Ceed ceed; 629 CeedBasis_Hip_shared *data; 630 631 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 632 CeedCallBackend(CeedBasisGetData(basis, &data)); 633 CeedCallHip(ceed, hipModuleUnload(data->module)); 634 if (data->moduleAtPoints) CeedCallHip(ceed, hipModuleUnload(data->moduleAtPoints)); 635 if (data->d_q_weight_1d) CeedCallHip(ceed, hipFree(data->d_q_weight_1d)); 636 CeedCallBackend(CeedFree(&data->h_points_per_elem)); 637 if (data->d_points_per_elem) CeedCallHip(ceed, hipFree(data->d_points_per_elem)); 638 CeedCallHip(ceed, hipFree(data->d_interp_1d)); 639 CeedCallHip(ceed, hipFree(data->d_grad_1d)); 640 CeedCallHip(ceed, hipFree(data->d_collo_grad_1d)); 641 CeedCallHip(ceed, hipFree(data->d_chebyshev_interp_1d)); 642 CeedCallBackend(CeedFree(&data)); 643 return CEED_ERROR_SUCCESS; 644 } 645 646 //------------------------------------------------------------------------------ 647 // Create tensor basis 648 //------------------------------------------------------------------------------ 649 int CeedBasisCreateTensorH1_Hip_shared(CeedInt dim, CeedInt P_1d, CeedInt Q_1d, const CeedScalar *interp_1d, const CeedScalar *grad_1d, 650 const CeedScalar *q_ref_1d, const CeedScalar *q_weight_1d, CeedBasis basis) { 651 Ceed ceed; 652 CeedInt num_comp; 653 const CeedInt q_bytes = Q_1d * sizeof(CeedScalar); 654 const CeedInt interp_bytes = q_bytes * P_1d; 655 CeedBasis_Hip_shared *data; 656 657 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 658 CeedCallBackend(CeedCalloc(1, &data)); 659 660 // Copy basis data to GPU 661 if (q_weight_1d) { 662 CeedCallHip(ceed, hipMalloc((void **)&data->d_q_weight_1d, q_bytes)); 663 CeedCallHip(ceed, hipMemcpy(data->d_q_weight_1d, q_weight_1d, q_bytes, hipMemcpyHostToDevice)); 664 } 665 CeedCallHip(ceed, hipMalloc((void **)&data->d_interp_1d, interp_bytes)); 666 CeedCallHip(ceed, hipMemcpy(data->d_interp_1d, interp_1d, interp_bytes, hipMemcpyHostToDevice)); 667 CeedCallHip(ceed, hipMalloc((void **)&data->d_grad_1d, interp_bytes)); 668 CeedCallHip(ceed, hipMemcpy(data->d_grad_1d, grad_1d, interp_bytes, hipMemcpyHostToDevice)); 669 670 // Compute collocated gradient and copy to GPU 671 data->d_collo_grad_1d = NULL; 672 bool has_collocated_grad = dim == 3 && Q_1d >= P_1d; 673 674 if (has_collocated_grad) { 675 CeedScalar *collo_grad_1d; 676 677 CeedCallBackend(CeedMalloc(Q_1d * Q_1d, &collo_grad_1d)); 678 CeedCallBackend(CeedBasisGetCollocatedGrad(basis, collo_grad_1d)); 679 CeedCallHip(ceed, hipMalloc((void **)&data->d_collo_grad_1d, q_bytes * Q_1d)); 680 CeedCallHip(ceed, hipMemcpy(data->d_collo_grad_1d, collo_grad_1d, q_bytes * Q_1d, hipMemcpyHostToDevice)); 681 CeedCallBackend(CeedFree(&collo_grad_1d)); 682 } 683 684 // Set number of threads per block for basis kernels 685 CeedCallBackend(CeedBasisGetNumComponents(basis, &num_comp)); 686 CeedCallBackend(ComputeBasisThreadBlockSizes(dim, P_1d, Q_1d, num_comp, data->block_sizes)); 687 688 // Compile basis kernels 689 const char basis_kernel_source[] = "// Tensor basis source\n#include <ceed/jit-source/hip/hip-shared-basis-tensor.h>\n"; 690 691 CeedCallBackend(CeedCompile_Hip(ceed, basis_kernel_source, &data->module, 11, "BASIS_Q_1D", Q_1d, "BASIS_P_1D", P_1d, "T_1D", 692 CeedIntMax(Q_1d, P_1d), "BASIS_DIM", dim, "BASIS_NUM_COMP", num_comp, "BASIS_NUM_NODES", CeedIntPow(P_1d, dim), 693 "BASIS_NUM_QPTS", CeedIntPow(Q_1d, dim), "BASIS_INTERP_BLOCK_SIZE", data->block_sizes[0], "BASIS_GRAD_BLOCK_SIZE", 694 data->block_sizes[1], "BASIS_WEIGHT_BLOCK_SIZE", data->block_sizes[2], "BASIS_HAS_COLLOCATED_GRAD", 695 has_collocated_grad)); 696 CeedCallBackend(CeedGetKernel_Hip(ceed, data->module, "Interp", &data->Interp)); 697 CeedCallBackend(CeedGetKernel_Hip(ceed, data->module, "InterpTranspose", &data->InterpTranspose)); 698 CeedCallBackend(CeedGetKernel_Hip(ceed, data->module, "InterpTransposeAdd", &data->InterpTransposeAdd)); 699 CeedCallBackend(CeedGetKernel_Hip(ceed, data->module, "Grad", &data->Grad)); 700 CeedCallBackend(CeedGetKernel_Hip(ceed, data->module, "GradTranspose", &data->GradTranspose)); 701 CeedCallBackend(CeedGetKernel_Hip(ceed, data->module, "GradTransposeAdd", &data->GradTransposeAdd)); 702 CeedCallBackend(CeedGetKernel_Hip(ceed, data->module, "Weight", &data->Weight)); 703 704 CeedCallBackend(CeedBasisSetData(basis, data)); 705 706 // Register backend functions 707 CeedCallBackend(CeedSetBackendFunction(ceed, "Basis", basis, "Apply", CeedBasisApplyTensor_Hip_shared)); 708 CeedCallBackend(CeedSetBackendFunction(ceed, "Basis", basis, "ApplyAdd", CeedBasisApplyAddTensor_Hip_shared)); 709 CeedCallBackend(CeedSetBackendFunction(ceed, "Basis", basis, "ApplyAtPoints", CeedBasisApplyAtPoints_Hip_shared)); 710 CeedCallBackend(CeedSetBackendFunction(ceed, "Basis", basis, "ApplyAddAtPoints", CeedBasisApplyAddAtPoints_Hip_shared)); 711 CeedCallBackend(CeedSetBackendFunction(ceed, "Basis", basis, "Destroy", CeedBasisDestroy_Hip_shared)); 712 CeedCallBackend(CeedDestroy(&ceed)); 713 return CEED_ERROR_SUCCESS; 714 } 715 716 //------------------------------------------------------------------------------ 717 // Create non-tensor basis 718 //------------------------------------------------------------------------------ 719 int CeedBasisCreateH1_Hip_shared(CeedElemTopology topo, CeedInt dim, CeedInt num_nodes, CeedInt num_qpts, const CeedScalar *interp, 720 const CeedScalar *grad, const CeedScalar *q_ref, const CeedScalar *q_weight, CeedBasis basis) { 721 Ceed ceed; 722 CeedInt num_comp, q_comp_interp, q_comp_grad; 723 const CeedInt q_bytes = num_qpts * sizeof(CeedScalar); 724 CeedBasis_Hip_shared *data; 725 726 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 727 728 // Check shared memory size 729 { 730 Ceed_Hip *hip_data; 731 732 CeedCallBackend(CeedGetData(ceed, &hip_data)); 733 if (((size_t)num_nodes * (size_t)num_qpts * (size_t)dim + (size_t)CeedIntMax(num_nodes, num_qpts)) * sizeof(CeedScalar) > 734 hip_data->device_prop.sharedMemPerBlock) { 735 CeedCallBackend(CeedBasisCreateH1Fallback(ceed, topo, dim, num_nodes, num_qpts, interp, grad, q_ref, q_weight, basis)); 736 return CEED_ERROR_SUCCESS; 737 } 738 } 739 740 CeedCallBackend(CeedCalloc(1, &data)); 741 742 // Copy basis data to GPU 743 CeedCallBackend(CeedBasisGetNumQuadratureComponents(basis, CEED_EVAL_INTERP, &q_comp_interp)); 744 CeedCallBackend(CeedBasisGetNumQuadratureComponents(basis, CEED_EVAL_GRAD, &q_comp_grad)); 745 if (q_weight) { 746 CeedCallHip(ceed, hipMalloc((void **)&data->d_q_weight_1d, q_bytes)); 747 CeedCallHip(ceed, hipMemcpy(data->d_q_weight_1d, q_weight, q_bytes, hipMemcpyHostToDevice)); 748 } 749 if (interp) { 750 const CeedInt interp_bytes = q_bytes * num_nodes * q_comp_interp; 751 752 CeedCallHip(ceed, hipMalloc((void **)&data->d_interp_1d, interp_bytes)); 753 CeedCallHip(ceed, hipMemcpy(data->d_interp_1d, interp, interp_bytes, hipMemcpyHostToDevice)); 754 } 755 if (grad) { 756 const CeedInt grad_bytes = q_bytes * num_nodes * q_comp_grad; 757 758 CeedCallHip(ceed, hipMalloc((void **)&data->d_grad_1d, grad_bytes)); 759 CeedCallHip(ceed, hipMemcpy(data->d_grad_1d, grad, grad_bytes, hipMemcpyHostToDevice)); 760 } 761 762 // Compile basis kernels 763 const char basis_kernel_source[] = "// Non-tensor basis source\n#include <ceed/jit-source/hip/hip-shared-basis-nontensor.h>\n"; 764 765 CeedCallBackend(CeedBasisGetNumComponents(basis, &num_comp)); 766 CeedCallBackend(ComputeBasisThreadBlockSizes(dim, num_nodes, num_qpts, num_comp, data->block_sizes)); 767 CeedCallBackend(CeedCompile_Hip(ceed, basis_kernel_source, &data->module, 6, "BASIS_Q", num_qpts, "BASIS_P", num_nodes, "T_1D", 768 CeedIntMax(num_qpts, num_nodes), "BASIS_DIM", dim, "BASIS_NUM_COMP", num_comp, "BASIS_INTERP_BLOCK_SIZE", 769 data->block_sizes[0])); 770 CeedCallBackend(CeedGetKernel_Hip(ceed, data->module, "Interp", &data->Interp)); 771 CeedCallBackend(CeedGetKernel_Hip(ceed, data->module, "InterpTranspose", &data->InterpTranspose)); 772 CeedCallBackend(CeedGetKernel_Hip(ceed, data->module, "InterpTransposeAdd", &data->InterpTransposeAdd)); 773 CeedCallBackend(CeedGetKernel_Hip(ceed, data->module, "Grad", &data->Grad)); 774 CeedCallBackend(CeedGetKernel_Hip(ceed, data->module, "GradTranspose", &data->GradTranspose)); 775 CeedCallBackend(CeedGetKernel_Hip(ceed, data->module, "GradTransposeAdd", &data->GradTransposeAdd)); 776 CeedCallBackend(CeedGetKernel_Hip(ceed, data->module, "Weight", &data->Weight)); 777 778 CeedCallBackend(CeedBasisSetData(basis, data)); 779 780 // Register backend functions 781 CeedCallBackend(CeedSetBackendFunction(ceed, "Basis", basis, "Apply", CeedBasisApplyNonTensor_Hip_shared)); 782 CeedCallBackend(CeedSetBackendFunction(ceed, "Basis", basis, "ApplyAdd", CeedBasisApplyAddNonTensor_Hip_shared)); 783 CeedCallBackend(CeedSetBackendFunction(ceed, "Basis", basis, "Destroy", CeedBasisDestroy_Hip_shared)); 784 CeedCallBackend(CeedDestroy(&ceed)); 785 return CEED_ERROR_SUCCESS; 786 } 787 788 //------------------------------------------------------------------------------ 789