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 <cuda.h> 12 #include <cuda_runtime.h> 13 #include <stdbool.h> 14 #include <stddef.h> 15 #include <string.h> 16 17 #include "../cuda/ceed-cuda-common.h" 18 #include "../cuda/ceed-cuda-compile.h" 19 #include "ceed-cuda-shared.h" 20 21 //------------------------------------------------------------------------------ 22 // Apply tensor basis 23 //------------------------------------------------------------------------------ 24 static int CeedBasisApplyTensorCore_Cuda_shared(CeedBasis basis, bool apply_add, const CeedInt num_elem, CeedTransposeMode t_mode, 25 CeedEvalMode eval_mode, CeedVector u, CeedVector v) { 26 Ceed ceed; 27 Ceed_Cuda *ceed_Cuda; 28 CeedInt dim, num_comp; 29 const CeedScalar *d_u; 30 CeedScalar *d_v; 31 CeedBasis_Cuda_shared *data; 32 33 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 34 CeedCallBackend(CeedGetData(ceed, &ceed_Cuda)); 35 CeedCallBackend(CeedBasisGetData(basis, &data)); 36 CeedCallBackend(CeedBasisGetDimension(basis, &dim)); 37 CeedCallBackend(CeedBasisGetNumComponents(basis, &num_comp)); 38 39 // Get read/write access to u, v 40 if (u != CEED_VECTOR_NONE) CeedCallBackend(CeedVectorGetArrayRead(u, CEED_MEM_DEVICE, &d_u)); 41 else CeedCheck(eval_mode == CEED_EVAL_WEIGHT, ceed, CEED_ERROR_BACKEND, "An input vector is required for this CeedEvalMode"); 42 if (apply_add) { 43 CeedCallBackend(CeedVectorGetArray(v, CEED_MEM_DEVICE, &d_v)); 44 } else { 45 CeedCallBackend(CeedVectorGetArrayWrite(v, CEED_MEM_DEVICE, &d_v)); 46 } 47 48 // Apply basis operation 49 switch (eval_mode) { 50 case CEED_EVAL_INTERP: { 51 CeedInt P_1d, Q_1d; 52 53 CeedCheck(data->d_interp_1d, ceed, CEED_ERROR_BACKEND, "%s not supported; interp_1d not set", CeedEvalModes[eval_mode]); 54 CeedCallBackend(CeedBasisGetNumNodes1D(basis, &P_1d)); 55 CeedCallBackend(CeedBasisGetNumQuadraturePoints1D(basis, &Q_1d)); 56 CeedInt thread_1d = CeedIntMax(Q_1d, P_1d); 57 58 void *interp_args[] = {(void *)&num_elem, &data->d_interp_1d, &d_u, &d_v}; 59 60 if (dim == 1) { 61 // avoid >512 total threads 62 CeedInt elems_per_block = CeedIntMin(ceed_Cuda->device_prop.maxThreadsDim[2], CeedIntMax(512 / thread_1d, 1)); 63 CeedInt grid = num_elem / elems_per_block + (num_elem % elems_per_block > 0); 64 CeedInt shared_mem = elems_per_block * thread_1d * sizeof(CeedScalar); 65 66 if (t_mode == CEED_TRANSPOSE) { 67 CeedCallBackend(CeedRunKernelDimShared_Cuda(ceed, apply_add ? data->InterpTransposeAdd : data->InterpTranspose, grid, thread_1d, 1, 68 elems_per_block, shared_mem, interp_args)); 69 } else { 70 CeedCallBackend(CeedRunKernelDimShared_Cuda(ceed, data->Interp, grid, thread_1d, 1, elems_per_block, shared_mem, interp_args)); 71 } 72 } else if (dim == 2) { 73 const CeedInt opt_elems[7] = {0, 32, 8, 6, 4, 2, 8}; 74 // elems_per_block must be at least 1 75 CeedInt elems_per_block = CeedIntMax(thread_1d < 7 ? opt_elems[thread_1d] / num_comp : 1, 1); 76 CeedInt grid = num_elem / elems_per_block + (num_elem % elems_per_block > 0); 77 CeedInt shared_mem = elems_per_block * thread_1d * thread_1d * sizeof(CeedScalar); 78 79 if (t_mode == CEED_TRANSPOSE) { 80 CeedCallBackend(CeedRunKernelDimShared_Cuda(ceed, apply_add ? data->InterpTransposeAdd : data->InterpTranspose, grid, thread_1d, thread_1d, 81 elems_per_block, shared_mem, interp_args)); 82 } else { 83 CeedCallBackend(CeedRunKernelDimShared_Cuda(ceed, data->Interp, grid, thread_1d, thread_1d, elems_per_block, shared_mem, interp_args)); 84 } 85 } else if (dim == 3) { 86 CeedInt elems_per_block = 1; 87 CeedInt grid = num_elem / elems_per_block + (num_elem % elems_per_block > 0); 88 CeedInt shared_mem = elems_per_block * thread_1d * thread_1d * sizeof(CeedScalar); 89 90 if (t_mode == CEED_TRANSPOSE) { 91 CeedCallBackend(CeedRunKernelDimShared_Cuda(ceed, apply_add ? data->InterpTransposeAdd : data->InterpTranspose, grid, thread_1d, thread_1d, 92 elems_per_block, shared_mem, interp_args)); 93 } else { 94 CeedCallBackend(CeedRunKernelDimShared_Cuda(ceed, data->Interp, grid, thread_1d, thread_1d, elems_per_block, shared_mem, interp_args)); 95 } 96 } 97 } break; 98 case CEED_EVAL_GRAD: { 99 CeedInt P_1d, Q_1d; 100 101 CeedCheck(data->d_grad_1d, ceed, CEED_ERROR_BACKEND, "%s not supported; grad_1d not set", CeedEvalModes[eval_mode]); 102 CeedCallBackend(CeedBasisGetNumNodes1D(basis, &P_1d)); 103 CeedCallBackend(CeedBasisGetNumQuadraturePoints1D(basis, &Q_1d)); 104 CeedInt thread_1d = CeedIntMax(Q_1d, P_1d); 105 CeedScalar *d_grad_1d = data->d_grad_1d; 106 107 if (data->d_collo_grad_1d) { 108 d_grad_1d = data->d_collo_grad_1d; 109 } 110 void *grad_args[] = {(void *)&num_elem, &data->d_interp_1d, &d_grad_1d, &d_u, &d_v}; 111 112 if (dim == 1) { 113 // avoid >512 total threads 114 CeedInt elems_per_block = CeedIntMin(ceed_Cuda->device_prop.maxThreadsDim[2], CeedIntMax(512 / thread_1d, 1)); 115 CeedInt grid = num_elem / elems_per_block + (num_elem % elems_per_block > 0); 116 CeedInt shared_mem = elems_per_block * thread_1d * sizeof(CeedScalar); 117 118 if (t_mode == CEED_TRANSPOSE) { 119 CeedCallBackend(CeedRunKernelDimShared_Cuda(ceed, apply_add ? data->GradTransposeAdd : data->GradTranspose, grid, thread_1d, 1, 120 elems_per_block, shared_mem, grad_args)); 121 } else { 122 CeedCallBackend(CeedRunKernelDimShared_Cuda(ceed, data->Grad, grid, thread_1d, 1, elems_per_block, shared_mem, grad_args)); 123 } 124 } else if (dim == 2) { 125 const CeedInt opt_elems[7] = {0, 32, 8, 6, 4, 2, 8}; 126 // elems_per_block must be at least 1 127 CeedInt elems_per_block = CeedIntMax(thread_1d < 7 ? opt_elems[thread_1d] / num_comp : 1, 1); 128 CeedInt grid = num_elem / elems_per_block + (num_elem % elems_per_block > 0); 129 CeedInt shared_mem = elems_per_block * thread_1d * thread_1d * sizeof(CeedScalar); 130 131 if (t_mode == CEED_TRANSPOSE) { 132 CeedCallBackend(CeedRunKernelDimShared_Cuda(ceed, apply_add ? data->GradTransposeAdd : data->GradTranspose, grid, thread_1d, thread_1d, 133 elems_per_block, shared_mem, grad_args)); 134 } else { 135 CeedCallBackend(CeedRunKernelDimShared_Cuda(ceed, data->Grad, grid, thread_1d, thread_1d, elems_per_block, shared_mem, grad_args)); 136 } 137 } else if (dim == 3) { 138 CeedInt elems_per_block = 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_Cuda(ceed, apply_add ? data->GradTransposeAdd : data->GradTranspose, grid, thread_1d, thread_1d, 144 elems_per_block, shared_mem, grad_args)); 145 } else { 146 CeedCallBackend(CeedRunKernelDimShared_Cuda(ceed, data->Grad, grid, thread_1d, thread_1d, elems_per_block, shared_mem, grad_args)); 147 } 148 } 149 } break; 150 case CEED_EVAL_WEIGHT: { 151 CeedInt Q_1d; 152 CeedInt block_size = 32; 153 154 CeedCheck(data->d_q_weight_1d, ceed, CEED_ERROR_BACKEND, "%s not supported; q_weights_1d not set", CeedEvalModes[eval_mode]); 155 CeedCallBackend(CeedBasisGetNumQuadraturePoints1D(basis, &Q_1d)); 156 void *weight_args[] = {(void *)&num_elem, (void *)&data->d_q_weight_1d, &d_v}; 157 if (dim == 1) { 158 const CeedInt elems_per_block = block_size / Q_1d; 159 const CeedInt grid_size = num_elem / elems_per_block + (num_elem % elems_per_block > 0); 160 161 CeedCallBackend(CeedRunKernelDim_Cuda(ceed, data->Weight, grid_size, Q_1d, elems_per_block, 1, weight_args)); 162 } else if (dim == 2) { 163 const CeedInt opt_elems = block_size / (Q_1d * Q_1d); 164 const CeedInt elems_per_block = opt_elems > 0 ? opt_elems : 1; 165 const CeedInt grid_size = num_elem / elems_per_block + (num_elem % elems_per_block > 0); 166 167 CeedCallBackend(CeedRunKernelDim_Cuda(ceed, data->Weight, grid_size, Q_1d, Q_1d, elems_per_block, weight_args)); 168 } else if (dim == 3) { 169 const CeedInt opt_elems = block_size / (Q_1d * Q_1d); 170 const CeedInt elems_per_block = opt_elems > 0 ? opt_elems : 1; 171 const CeedInt grid_size = num_elem / elems_per_block + (num_elem % elems_per_block > 0); 172 173 CeedCallBackend(CeedRunKernelDim_Cuda(ceed, data->Weight, grid_size, Q_1d, Q_1d, elems_per_block, weight_args)); 174 } 175 } break; 176 case CEED_EVAL_NONE: /* handled separately below */ 177 break; 178 // LCOV_EXCL_START 179 case CEED_EVAL_DIV: 180 case CEED_EVAL_CURL: 181 return CeedError(ceed, CEED_ERROR_BACKEND, "%s not supported", CeedEvalModes[eval_mode]); 182 // LCOV_EXCL_STOP 183 } 184 185 // Restore vectors, cover CEED_EVAL_NONE 186 CeedCallBackend(CeedVectorRestoreArray(v, &d_v)); 187 if (eval_mode == CEED_EVAL_NONE) CeedCallBackend(CeedVectorSetArray(v, CEED_MEM_DEVICE, CEED_COPY_VALUES, (CeedScalar *)d_u)); 188 if (eval_mode != CEED_EVAL_WEIGHT) CeedCallBackend(CeedVectorRestoreArrayRead(u, &d_u)); 189 CeedCallBackend(CeedDestroy(&ceed)); 190 return CEED_ERROR_SUCCESS; 191 } 192 193 static int CeedBasisApplyTensor_Cuda_shared(CeedBasis basis, const CeedInt num_elem, CeedTransposeMode t_mode, CeedEvalMode eval_mode, CeedVector u, 194 CeedVector v) { 195 CeedCallBackend(CeedBasisApplyTensorCore_Cuda_shared(basis, false, num_elem, t_mode, eval_mode, u, v)); 196 return CEED_ERROR_SUCCESS; 197 } 198 199 static int CeedBasisApplyAddTensor_Cuda_shared(CeedBasis basis, const CeedInt num_elem, CeedTransposeMode t_mode, CeedEvalMode eval_mode, 200 CeedVector u, CeedVector v) { 201 CeedCallBackend(CeedBasisApplyTensorCore_Cuda_shared(basis, true, num_elem, t_mode, eval_mode, u, v)); 202 return CEED_ERROR_SUCCESS; 203 } 204 205 //------------------------------------------------------------------------------ 206 // Basis apply - tensor AtPoints 207 //------------------------------------------------------------------------------ 208 static int CeedBasisApplyAtPointsCore_Cuda_shared(CeedBasis basis, bool apply_add, const CeedInt num_elem, const CeedInt *num_points, 209 CeedTransposeMode t_mode, CeedEvalMode eval_mode, CeedVector x_ref, CeedVector u, CeedVector v) { 210 Ceed ceed; 211 Ceed_Cuda *ceed_Cuda; 212 CeedInt Q_1d, dim, num_comp, max_num_points = num_points[0]; 213 const CeedInt is_transpose = t_mode == CEED_TRANSPOSE; 214 const CeedScalar *d_x, *d_u; 215 CeedScalar *d_v; 216 CeedBasis_Cuda_shared *data; 217 218 CeedCallBackend(CeedBasisGetData(basis, &data)); 219 CeedCallBackend(CeedBasisGetNumQuadraturePoints1D(basis, &Q_1d)); 220 CeedCallBackend(CeedBasisGetDimension(basis, &dim)); 221 CeedCallBackend(CeedBasisGetNumComponents(basis, &num_comp)); 222 223 // Weight handled separately 224 if (eval_mode == CEED_EVAL_WEIGHT) { 225 CeedCallBackend(CeedVectorSetValue(v, 1.0)); 226 return CEED_ERROR_SUCCESS; 227 } 228 229 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 230 CeedCallBackend(CeedGetData(ceed, &ceed_Cuda)); 231 232 // Check padded to uniform number of points per elem 233 for (CeedInt i = 1; i < num_elem; i++) max_num_points = CeedIntMax(max_num_points, num_points[i]); 234 { 235 CeedInt q_comp; 236 CeedSize len, len_required; 237 CeedCallBackend(CeedBasisGetNumQuadratureComponents(basis, eval_mode, &q_comp)); 238 CeedCallBackend(CeedVectorGetLength(is_transpose ? u : v, &len)); 239 len_required = (CeedSize)num_comp * (CeedSize)q_comp * (CeedSize)num_elem * (CeedSize)max_num_points; 240 CeedCheck(len >= len_required, ceed, CEED_ERROR_BACKEND, 241 "Vector at points must be padded to the same number of points in each element for BasisApplyAtPoints on GPU backends." 242 " Found %" CeedSize_FMT ", Required %" CeedSize_FMT, 243 len, len_required); 244 } 245 246 // Move num_points array to device 247 if (is_transpose) { 248 const CeedInt num_bytes = num_elem * sizeof(CeedInt); 249 250 if (num_elem != data->num_elem_at_points) { 251 data->num_elem_at_points = num_elem; 252 253 if (data->d_points_per_elem) CeedCallCuda(ceed, cudaFree(data->d_points_per_elem)); 254 CeedCallCuda(ceed, cudaMalloc((void **)&data->d_points_per_elem, num_bytes)); 255 CeedCallBackend(CeedFree(&data->h_points_per_elem)); 256 CeedCallBackend(CeedCalloc(num_elem, &data->h_points_per_elem)); 257 } 258 if (memcmp(data->h_points_per_elem, num_points, num_bytes)) { 259 memcpy(data->h_points_per_elem, num_points, num_bytes); 260 CeedCallCuda(ceed, cudaMemcpy(data->d_points_per_elem, num_points, num_bytes, cudaMemcpyHostToDevice)); 261 } 262 } 263 264 // Build kernels if needed 265 if (data->num_points != max_num_points) { 266 CeedInt P_1d; 267 268 CeedCallBackend(CeedBasisGetNumNodes1D(basis, &P_1d)); 269 data->num_points = max_num_points; 270 271 // -- Create interp matrix to Chebyshev coefficients 272 if (!data->d_chebyshev_interp_1d) { 273 CeedSize interp_bytes; 274 CeedScalar *chebyshev_interp_1d; 275 276 interp_bytes = P_1d * Q_1d * sizeof(CeedScalar); 277 CeedCallBackend(CeedCalloc(P_1d * Q_1d, &chebyshev_interp_1d)); 278 CeedCallBackend(CeedBasisGetChebyshevInterp1D(basis, chebyshev_interp_1d)); 279 CeedCallCuda(ceed, cudaMalloc((void **)&data->d_chebyshev_interp_1d, interp_bytes)); 280 CeedCallCuda(ceed, cudaMemcpy(data->d_chebyshev_interp_1d, chebyshev_interp_1d, interp_bytes, cudaMemcpyHostToDevice)); 281 CeedCallBackend(CeedFree(&chebyshev_interp_1d)); 282 } 283 284 // -- Compile kernels 285 const char basis_kernel_source[] = "// AtPoints basis source\n#include <ceed/jit-source/cuda/cuda-shared-basis-tensor-at-points.h>\n"; 286 CeedInt num_comp; 287 288 if (data->moduleAtPoints) CeedCallCuda(ceed, cuModuleUnload(data->moduleAtPoints)); 289 CeedCallBackend(CeedBasisGetNumComponents(basis, &num_comp)); 290 CeedCallBackend(CeedCompile_Cuda(ceed, basis_kernel_source, &data->moduleAtPoints, 8, "BASIS_Q_1D", Q_1d, "BASIS_P_1D", P_1d, "T_1D", 291 CeedIntMax(Q_1d, P_1d), "BASIS_DIM", dim, "BASIS_NUM_COMP", num_comp, "BASIS_NUM_NODES", CeedIntPow(P_1d, dim), 292 "BASIS_NUM_QPTS", CeedIntPow(Q_1d, dim), "BASIS_NUM_PTS", max_num_points)); 293 CeedCallBackend(CeedGetKernel_Cuda(ceed, data->moduleAtPoints, "InterpAtPoints", &data->InterpAtPoints)); 294 CeedCallBackend(CeedGetKernel_Cuda(ceed, data->moduleAtPoints, "InterpTransposeAtPoints", &data->InterpTransposeAtPoints)); 295 CeedCallBackend(CeedGetKernel_Cuda(ceed, data->moduleAtPoints, "GradAtPoints", &data->GradAtPoints)); 296 CeedCallBackend(CeedGetKernel_Cuda(ceed, data->moduleAtPoints, "GradTransposeAtPoints", &data->GradTransposeAtPoints)); 297 } 298 299 // Get read/write access to u, v 300 CeedCallBackend(CeedVectorGetArrayRead(x_ref, CEED_MEM_DEVICE, &d_x)); 301 if (u != CEED_VECTOR_NONE) CeedCallBackend(CeedVectorGetArrayRead(u, CEED_MEM_DEVICE, &d_u)); 302 else CeedCheck(eval_mode == CEED_EVAL_WEIGHT, ceed, CEED_ERROR_BACKEND, "An input vector is required for this CeedEvalMode"); 303 if (apply_add) { 304 CeedCallBackend(CeedVectorGetArray(v, CEED_MEM_DEVICE, &d_v)); 305 } else { 306 // Clear v for transpose operation 307 if (is_transpose) CeedCallBackend(CeedVectorSetValue(v, 0.0)); 308 CeedCallBackend(CeedVectorGetArrayWrite(v, CEED_MEM_DEVICE, &d_v)); 309 } 310 311 // Basis action 312 switch (eval_mode) { 313 case CEED_EVAL_INTERP: { 314 CeedInt P_1d, Q_1d; 315 316 CeedCallBackend(CeedBasisGetNumNodes1D(basis, &P_1d)); 317 CeedCallBackend(CeedBasisGetNumQuadraturePoints1D(basis, &Q_1d)); 318 CeedInt thread_1d = CeedIntMax(Q_1d, P_1d); 319 320 void *interp_args[] = {(void *)&num_elem, &data->d_chebyshev_interp_1d, &data->d_points_per_elem, &d_x, &d_u, &d_v}; 321 322 if (dim == 1) { 323 // avoid >512 total threads 324 CeedInt elems_per_block = CeedIntMin(ceed_Cuda->device_prop.maxThreadsDim[2], CeedIntMax(512 / thread_1d, 1)); 325 CeedInt grid = num_elem / elems_per_block + (num_elem % elems_per_block > 0); 326 CeedInt shared_mem = elems_per_block * thread_1d * sizeof(CeedScalar); 327 328 CeedCallBackend(CeedRunKernelDimShared_Cuda(ceed, is_transpose ? data->InterpTransposeAtPoints : data->InterpAtPoints, grid, thread_1d, 1, 329 elems_per_block, shared_mem, interp_args)); 330 } else if (dim == 2) { 331 const CeedInt opt_elems[7] = {0, 32, 8, 6, 4, 2, 8}; 332 // elems_per_block must be at least 1 333 CeedInt elems_per_block = CeedIntMax(thread_1d < 7 ? opt_elems[thread_1d] / num_comp : 1, 1); 334 CeedInt grid = num_elem / elems_per_block + (num_elem % elems_per_block > 0); 335 CeedInt shared_mem = elems_per_block * thread_1d * thread_1d * sizeof(CeedScalar); 336 337 CeedCallBackend(CeedRunKernelDimShared_Cuda(ceed, is_transpose ? data->InterpTransposeAtPoints : data->InterpAtPoints, grid, thread_1d, 338 thread_1d, elems_per_block, shared_mem, interp_args)); 339 } else if (dim == 3) { 340 CeedInt elems_per_block = 1; 341 CeedInt grid = num_elem / elems_per_block + (num_elem % elems_per_block > 0); 342 CeedInt shared_mem = elems_per_block * thread_1d * thread_1d * sizeof(CeedScalar); 343 344 CeedCallBackend(CeedRunKernelDimShared_Cuda(ceed, is_transpose ? data->InterpTransposeAtPoints : data->InterpAtPoints, grid, thread_1d, 345 thread_1d, elems_per_block, shared_mem, interp_args)); 346 } 347 } break; 348 case CEED_EVAL_GRAD: { 349 CeedInt P_1d, Q_1d; 350 351 CeedCallBackend(CeedBasisGetNumNodes1D(basis, &P_1d)); 352 CeedCallBackend(CeedBasisGetNumQuadraturePoints1D(basis, &Q_1d)); 353 CeedInt thread_1d = CeedIntMax(Q_1d, P_1d); 354 355 void *grad_args[] = {(void *)&num_elem, &data->d_chebyshev_interp_1d, &data->d_points_per_elem, &d_x, &d_u, &d_v}; 356 357 if (dim == 1) { 358 // avoid >512 total threads 359 CeedInt elems_per_block = CeedIntMin(ceed_Cuda->device_prop.maxThreadsDim[2], CeedIntMax(512 / thread_1d, 1)); 360 CeedInt grid = num_elem / elems_per_block + (num_elem % elems_per_block > 0); 361 CeedInt shared_mem = elems_per_block * thread_1d * sizeof(CeedScalar); 362 363 CeedCallBackend(CeedRunKernelDimShared_Cuda(ceed, is_transpose ? data->GradTransposeAtPoints : data->GradAtPoints, grid, thread_1d, 1, 364 elems_per_block, shared_mem, grad_args)); 365 } else if (dim == 2) { 366 const CeedInt opt_elems[7] = {0, 32, 8, 6, 4, 2, 8}; 367 // elems_per_block must be at least 1 368 CeedInt elems_per_block = CeedIntMax(thread_1d < 7 ? opt_elems[thread_1d] / num_comp : 1, 1); 369 CeedInt grid = num_elem / elems_per_block + (num_elem % elems_per_block > 0); 370 CeedInt shared_mem = elems_per_block * thread_1d * thread_1d * sizeof(CeedScalar); 371 372 CeedCallBackend(CeedRunKernelDimShared_Cuda(ceed, is_transpose ? data->GradTransposeAtPoints : data->GradAtPoints, grid, thread_1d, thread_1d, 373 elems_per_block, shared_mem, grad_args)); 374 } else if (dim == 3) { 375 CeedInt elems_per_block = 1; 376 CeedInt grid = num_elem / elems_per_block + (num_elem % elems_per_block > 0); 377 CeedInt shared_mem = elems_per_block * thread_1d * thread_1d * sizeof(CeedScalar); 378 379 CeedCallBackend(CeedRunKernelDimShared_Cuda(ceed, is_transpose ? data->GradTransposeAtPoints : data->GradAtPoints, grid, thread_1d, thread_1d, 380 elems_per_block, shared_mem, grad_args)); 381 } 382 } break; 383 case CEED_EVAL_WEIGHT: 384 case CEED_EVAL_NONE: /* handled separately below */ 385 break; 386 // LCOV_EXCL_START 387 case CEED_EVAL_DIV: 388 case CEED_EVAL_CURL: 389 return CeedError(ceed, CEED_ERROR_BACKEND, "%s not supported", CeedEvalModes[eval_mode]); 390 // LCOV_EXCL_STOP 391 } 392 393 // Restore vectors, cover CEED_EVAL_NONE 394 CeedCallBackend(CeedVectorRestoreArrayRead(x_ref, &d_x)); 395 CeedCallBackend(CeedVectorRestoreArray(v, &d_v)); 396 if (eval_mode == CEED_EVAL_NONE) CeedCallBackend(CeedVectorSetArray(v, CEED_MEM_DEVICE, CEED_COPY_VALUES, (CeedScalar *)d_u)); 397 if (eval_mode != CEED_EVAL_WEIGHT) CeedCallBackend(CeedVectorRestoreArrayRead(u, &d_u)); 398 CeedCallBackend(CeedDestroy(&ceed)); 399 return CEED_ERROR_SUCCESS; 400 } 401 402 static int CeedBasisApplyAtPoints_Cuda_shared(CeedBasis basis, const CeedInt num_elem, const CeedInt *num_points, CeedTransposeMode t_mode, 403 CeedEvalMode eval_mode, CeedVector x_ref, CeedVector u, CeedVector v) { 404 CeedCallBackend(CeedBasisApplyAtPointsCore_Cuda_shared(basis, false, num_elem, num_points, t_mode, eval_mode, x_ref, u, v)); 405 return CEED_ERROR_SUCCESS; 406 } 407 408 static int CeedBasisApplyAddAtPoints_Cuda_shared(CeedBasis basis, const CeedInt num_elem, const CeedInt *num_points, CeedTransposeMode t_mode, 409 CeedEvalMode eval_mode, CeedVector x_ref, CeedVector u, CeedVector v) { 410 CeedCallBackend(CeedBasisApplyAtPointsCore_Cuda_shared(basis, true, num_elem, num_points, t_mode, eval_mode, x_ref, u, v)); 411 return CEED_ERROR_SUCCESS; 412 } 413 414 //------------------------------------------------------------------------------ 415 // Apply non-tensor basis 416 //------------------------------------------------------------------------------ 417 static int CeedBasisApplyNonTensorCore_Cuda_shared(CeedBasis basis, bool apply_add, const CeedInt num_elem, CeedTransposeMode t_mode, 418 CeedEvalMode eval_mode, CeedVector u, CeedVector v) { 419 Ceed ceed; 420 Ceed_Cuda *ceed_Cuda; 421 CeedInt dim; 422 const CeedScalar *d_u; 423 CeedScalar *d_v; 424 CeedBasis_Cuda_shared *data; 425 426 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 427 CeedCallBackend(CeedGetData(ceed, &ceed_Cuda)); 428 CeedCallBackend(CeedBasisGetData(basis, &data)); 429 CeedCallBackend(CeedBasisGetDimension(basis, &dim)); 430 431 // Get read/write access to u, v 432 if (u != CEED_VECTOR_NONE) CeedCallBackend(CeedVectorGetArrayRead(u, CEED_MEM_DEVICE, &d_u)); 433 else CeedCheck(eval_mode == CEED_EVAL_WEIGHT, ceed, CEED_ERROR_BACKEND, "An input vector is required for this CeedEvalMode"); 434 if (apply_add) { 435 CeedCallBackend(CeedVectorGetArray(v, CEED_MEM_DEVICE, &d_v)); 436 } else { 437 CeedCallBackend(CeedVectorGetArrayWrite(v, CEED_MEM_DEVICE, &d_v)); 438 } 439 440 // Apply basis operation 441 switch (eval_mode) { 442 case CEED_EVAL_INTERP: { 443 CeedInt P, Q; 444 445 CeedCheck(data->d_interp_1d, ceed, CEED_ERROR_BACKEND, "%s not supported; interp not set", CeedEvalModes[eval_mode]); 446 CeedCallBackend(CeedBasisGetNumNodes(basis, &P)); 447 CeedCallBackend(CeedBasisGetNumQuadraturePoints(basis, &Q)); 448 CeedInt thread = CeedIntMax(Q, P); 449 450 void *interp_args[] = {(void *)&num_elem, &data->d_interp_1d, &d_u, &d_v}; 451 452 { 453 // avoid >512 total threads 454 CeedInt elems_per_block = CeedIntMin(ceed_Cuda->device_prop.maxThreadsDim[2], CeedIntMax(512 / thread, 1)); 455 CeedInt grid = num_elem / elems_per_block + (num_elem % elems_per_block > 0); 456 CeedInt shared_mem = elems_per_block * thread * sizeof(CeedScalar); 457 458 if (t_mode == CEED_TRANSPOSE) { 459 CeedCallBackend(CeedRunKernelDimShared_Cuda(ceed, apply_add ? data->InterpTransposeAdd : data->InterpTranspose, grid, thread, 1, 460 elems_per_block, shared_mem, interp_args)); 461 } else { 462 CeedCallBackend(CeedRunKernelDimShared_Cuda(ceed, data->Interp, grid, thread, 1, elems_per_block, shared_mem, interp_args)); 463 } 464 } 465 } break; 466 case CEED_EVAL_GRAD: { 467 CeedInt P, Q; 468 469 CeedCheck(data->d_grad_1d, ceed, CEED_ERROR_BACKEND, "%s not supported; grad not set", CeedEvalModes[eval_mode]); 470 CeedCallBackend(CeedBasisGetNumNodes(basis, &P)); 471 CeedCallBackend(CeedBasisGetNumQuadraturePoints(basis, &Q)); 472 CeedInt thread = CeedIntMax(Q, P); 473 474 void *grad_args[] = {(void *)&num_elem, &data->d_grad_1d, &d_u, &d_v}; 475 476 { 477 // avoid >512 total threads 478 CeedInt elems_per_block = CeedIntMin(ceed_Cuda->device_prop.maxThreadsDim[2], CeedIntMax(512 / thread, 1)); 479 CeedInt grid = num_elem / elems_per_block + (num_elem % elems_per_block > 0); 480 CeedInt shared_mem = elems_per_block * thread * sizeof(CeedScalar); 481 482 if (t_mode == CEED_TRANSPOSE) { 483 CeedCallBackend(CeedRunKernelDimShared_Cuda(ceed, apply_add ? data->GradTransposeAdd : data->GradTranspose, grid, thread, 1, 484 elems_per_block, shared_mem, grad_args)); 485 } else { 486 CeedCallBackend(CeedRunKernelDimShared_Cuda(ceed, data->Grad, grid, thread, 1, elems_per_block, shared_mem, grad_args)); 487 } 488 } 489 } break; 490 case CEED_EVAL_WEIGHT: { 491 CeedInt P, Q; 492 493 CeedCheck(data->d_q_weight_1d, ceed, CEED_ERROR_BACKEND, "%s not supported; q_weights not set", CeedEvalModes[eval_mode]); 494 CeedCallBackend(CeedBasisGetNumNodes(basis, &P)); 495 CeedCallBackend(CeedBasisGetNumQuadraturePoints(basis, &Q)); 496 CeedInt thread = CeedIntMax(Q, P); 497 498 void *weight_args[] = {(void *)&num_elem, (void *)&data->d_q_weight_1d, &d_v}; 499 500 { 501 // avoid >512 total threads 502 CeedInt elems_per_block = CeedIntMin(ceed_Cuda->device_prop.maxThreadsDim[2], CeedIntMax(512 / thread, 1)); 503 CeedInt grid = num_elem / elems_per_block + (num_elem % elems_per_block > 0); 504 505 CeedCallBackend(CeedRunKernelDim_Cuda(ceed, data->Weight, grid, thread, elems_per_block, 1, weight_args)); 506 } 507 } break; 508 case CEED_EVAL_NONE: /* handled separately below */ 509 break; 510 // LCOV_EXCL_START 511 case CEED_EVAL_DIV: 512 case CEED_EVAL_CURL: 513 return CeedError(ceed, CEED_ERROR_BACKEND, "%s not supported", CeedEvalModes[eval_mode]); 514 // LCOV_EXCL_STOP 515 } 516 517 // Restore vectors, cover CEED_EVAL_NONE 518 CeedCallBackend(CeedVectorRestoreArray(v, &d_v)); 519 if (eval_mode == CEED_EVAL_NONE) CeedCallBackend(CeedVectorSetArray(v, CEED_MEM_DEVICE, CEED_COPY_VALUES, (CeedScalar *)d_u)); 520 if (eval_mode != CEED_EVAL_WEIGHT) CeedCallBackend(CeedVectorRestoreArrayRead(u, &d_u)); 521 CeedCallBackend(CeedDestroy(&ceed)); 522 return CEED_ERROR_SUCCESS; 523 } 524 525 static int CeedBasisApplyNonTensor_Cuda_shared(CeedBasis basis, const CeedInt num_elem, CeedTransposeMode t_mode, CeedEvalMode eval_mode, 526 CeedVector u, CeedVector v) { 527 CeedCallBackend(CeedBasisApplyNonTensorCore_Cuda_shared(basis, false, num_elem, t_mode, eval_mode, u, v)); 528 return CEED_ERROR_SUCCESS; 529 } 530 531 static int CeedBasisApplyAddNonTensor_Cuda_shared(CeedBasis basis, const CeedInt num_elem, CeedTransposeMode t_mode, CeedEvalMode eval_mode, 532 CeedVector u, CeedVector v) { 533 CeedCallBackend(CeedBasisApplyNonTensorCore_Cuda_shared(basis, true, num_elem, t_mode, eval_mode, u, v)); 534 return CEED_ERROR_SUCCESS; 535 } 536 537 //------------------------------------------------------------------------------ 538 // Destroy basis 539 //------------------------------------------------------------------------------ 540 static int CeedBasisDestroy_Cuda_shared(CeedBasis basis) { 541 Ceed ceed; 542 CeedBasis_Cuda_shared *data; 543 544 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 545 CeedCallBackend(CeedBasisGetData(basis, &data)); 546 CeedCallCuda(ceed, cuModuleUnload(data->module)); 547 if (data->moduleAtPoints) CeedCallCuda(ceed, cuModuleUnload(data->moduleAtPoints)); 548 if (data->d_q_weight_1d) CeedCallCuda(ceed, cudaFree(data->d_q_weight_1d)); 549 CeedCallBackend(CeedFree(&data->h_points_per_elem)); 550 if (data->d_points_per_elem) CeedCallCuda(ceed, cudaFree(data->d_points_per_elem)); 551 CeedCallCuda(ceed, cudaFree(data->d_interp_1d)); 552 CeedCallCuda(ceed, cudaFree(data->d_grad_1d)); 553 CeedCallCuda(ceed, cudaFree(data->d_collo_grad_1d)); 554 CeedCallCuda(ceed, cudaFree(data->d_chebyshev_interp_1d)); 555 CeedCallBackend(CeedFree(&data)); 556 CeedCallBackend(CeedDestroy(&ceed)); 557 return CEED_ERROR_SUCCESS; 558 } 559 560 //------------------------------------------------------------------------------ 561 // Create tensor basis 562 //------------------------------------------------------------------------------ 563 int CeedBasisCreateTensorH1_Cuda_shared(CeedInt dim, CeedInt P_1d, CeedInt Q_1d, const CeedScalar *interp_1d, const CeedScalar *grad_1d, 564 const CeedScalar *q_ref_1d, const CeedScalar *q_weight_1d, CeedBasis basis) { 565 Ceed ceed; 566 CeedInt num_comp; 567 const CeedInt q_bytes = Q_1d * sizeof(CeedScalar); 568 const CeedInt interp_bytes = q_bytes * P_1d; 569 CeedBasis_Cuda_shared *data; 570 571 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 572 CeedCallBackend(CeedCalloc(1, &data)); 573 574 // Copy basis data to GPU 575 if (q_weight_1d) { 576 CeedCallCuda(ceed, cudaMalloc((void **)&data->d_q_weight_1d, q_bytes)); 577 CeedCallCuda(ceed, cudaMemcpy(data->d_q_weight_1d, q_weight_1d, q_bytes, cudaMemcpyHostToDevice)); 578 } 579 CeedCallCuda(ceed, cudaMalloc((void **)&data->d_interp_1d, interp_bytes)); 580 CeedCallCuda(ceed, cudaMemcpy(data->d_interp_1d, interp_1d, interp_bytes, cudaMemcpyHostToDevice)); 581 CeedCallCuda(ceed, cudaMalloc((void **)&data->d_grad_1d, interp_bytes)); 582 CeedCallCuda(ceed, cudaMemcpy(data->d_grad_1d, grad_1d, interp_bytes, cudaMemcpyHostToDevice)); 583 584 // Compute collocated gradient and copy to GPU 585 data->d_collo_grad_1d = NULL; 586 bool has_collocated_grad = dim == 3 && Q_1d >= P_1d; 587 588 if (has_collocated_grad) { 589 CeedScalar *collo_grad_1d; 590 591 CeedCallBackend(CeedMalloc(Q_1d * Q_1d, &collo_grad_1d)); 592 CeedCallBackend(CeedBasisGetCollocatedGrad(basis, collo_grad_1d)); 593 CeedCallCuda(ceed, cudaMalloc((void **)&data->d_collo_grad_1d, q_bytes * Q_1d)); 594 CeedCallCuda(ceed, cudaMemcpy(data->d_collo_grad_1d, collo_grad_1d, q_bytes * Q_1d, cudaMemcpyHostToDevice)); 595 CeedCallBackend(CeedFree(&collo_grad_1d)); 596 } 597 598 // Compile basis kernels 599 const char basis_kernel_source[] = "// Tensor basis source\n#include <ceed/jit-source/cuda/cuda-shared-basis-tensor.h>\n"; 600 601 CeedCallBackend(CeedBasisGetNumComponents(basis, &num_comp)); 602 CeedCallBackend(CeedCompile_Cuda(ceed, basis_kernel_source, &data->module, 8, "BASIS_Q_1D", Q_1d, "BASIS_P_1D", P_1d, "T_1D", 603 CeedIntMax(Q_1d, P_1d), "BASIS_DIM", dim, "BASIS_NUM_COMP", num_comp, "BASIS_NUM_NODES", CeedIntPow(P_1d, dim), 604 "BASIS_NUM_QPTS", CeedIntPow(Q_1d, dim), "BASIS_HAS_COLLOCATED_GRAD", has_collocated_grad)); 605 CeedCallBackend(CeedGetKernel_Cuda(ceed, data->module, "Interp", &data->Interp)); 606 CeedCallBackend(CeedGetKernel_Cuda(ceed, data->module, "InterpTranspose", &data->InterpTranspose)); 607 CeedCallBackend(CeedGetKernel_Cuda(ceed, data->module, "InterpTransposeAdd", &data->InterpTransposeAdd)); 608 CeedCallBackend(CeedGetKernel_Cuda(ceed, data->module, "Grad", &data->Grad)); 609 CeedCallBackend(CeedGetKernel_Cuda(ceed, data->module, "GradTranspose", &data->GradTranspose)); 610 CeedCallBackend(CeedGetKernel_Cuda(ceed, data->module, "GradTransposeAdd", &data->GradTransposeAdd)); 611 CeedCallBackend(CeedGetKernel_Cuda(ceed, data->module, "Weight", &data->Weight)); 612 613 CeedCallBackend(CeedBasisSetData(basis, data)); 614 615 // Register backend functions 616 CeedCallBackend(CeedSetBackendFunction(ceed, "Basis", basis, "Apply", CeedBasisApplyTensor_Cuda_shared)); 617 CeedCallBackend(CeedSetBackendFunction(ceed, "Basis", basis, "ApplyAdd", CeedBasisApplyAddTensor_Cuda_shared)); 618 CeedCallBackend(CeedSetBackendFunction(ceed, "Basis", basis, "ApplyAtPoints", CeedBasisApplyAtPoints_Cuda_shared)); 619 CeedCallBackend(CeedSetBackendFunction(ceed, "Basis", basis, "ApplyAddAtPoints", CeedBasisApplyAddAtPoints_Cuda_shared)); 620 CeedCallBackend(CeedSetBackendFunction(ceed, "Basis", basis, "Destroy", CeedBasisDestroy_Cuda_shared)); 621 CeedCallBackend(CeedDestroy(&ceed)); 622 return CEED_ERROR_SUCCESS; 623 } 624 625 //------------------------------------------------------------------------------ 626 // Create non-tensor basis 627 //------------------------------------------------------------------------------ 628 int CeedBasisCreateH1_Cuda_shared(CeedElemTopology topo, CeedInt dim, CeedInt num_nodes, CeedInt num_qpts, const CeedScalar *interp, 629 const CeedScalar *grad, const CeedScalar *q_ref, const CeedScalar *q_weight, CeedBasis basis) { 630 Ceed ceed; 631 CeedInt num_comp, q_comp_interp, q_comp_grad; 632 const CeedInt q_bytes = num_qpts * sizeof(CeedScalar); 633 CeedBasis_Cuda_shared *data; 634 635 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 636 637 // Check shared memory size 638 { 639 Ceed_Cuda *cuda_data; 640 641 CeedCallBackend(CeedGetData(ceed, &cuda_data)); 642 if (((size_t)num_nodes * (size_t)num_qpts * (size_t)dim + (size_t)CeedIntMax(num_nodes, num_qpts)) * sizeof(CeedScalar) > 643 cuda_data->device_prop.sharedMemPerBlock) { 644 CeedCallBackend(CeedBasisCreateH1Fallback(ceed, topo, dim, num_nodes, num_qpts, interp, grad, q_ref, q_weight, basis)); 645 CeedCallBackend(CeedDestroy(&ceed)); 646 return CEED_ERROR_SUCCESS; 647 } 648 } 649 650 CeedCallBackend(CeedCalloc(1, &data)); 651 652 // Copy basis data to GPU 653 CeedCallBackend(CeedBasisGetNumQuadratureComponents(basis, CEED_EVAL_INTERP, &q_comp_interp)); 654 CeedCallBackend(CeedBasisGetNumQuadratureComponents(basis, CEED_EVAL_GRAD, &q_comp_grad)); 655 if (q_weight) { 656 CeedCallCuda(ceed, cudaMalloc((void **)&data->d_q_weight_1d, q_bytes)); 657 CeedCallCuda(ceed, cudaMemcpy(data->d_q_weight_1d, q_weight, q_bytes, cudaMemcpyHostToDevice)); 658 } 659 if (interp) { 660 const CeedInt interp_bytes = q_bytes * num_nodes * q_comp_interp; 661 662 CeedCallCuda(ceed, cudaMalloc((void **)&data->d_interp_1d, interp_bytes)); 663 CeedCallCuda(ceed, cudaMemcpy(data->d_interp_1d, interp, interp_bytes, cudaMemcpyHostToDevice)); 664 } 665 if (grad) { 666 const CeedInt grad_bytes = q_bytes * num_nodes * q_comp_grad; 667 668 CeedCallCuda(ceed, cudaMalloc((void **)&data->d_grad_1d, grad_bytes)); 669 CeedCallCuda(ceed, cudaMemcpy(data->d_grad_1d, grad, grad_bytes, cudaMemcpyHostToDevice)); 670 } 671 672 // Compile basis kernels 673 const char basis_kernel_source[] = "// Non-tensor basis source\n#include <ceed/jit-source/cuda/cuda-shared-basis-nontensor.h>\n"; 674 675 CeedCallBackend(CeedBasisGetNumComponents(basis, &num_comp)); 676 CeedCallBackend(CeedCompile_Cuda(ceed, basis_kernel_source, &data->module, 5, "BASIS_Q", num_qpts, "BASIS_P", num_nodes, "T_1D", 677 CeedIntMax(num_qpts, num_nodes), "BASIS_DIM", dim, "BASIS_NUM_COMP", num_comp)); 678 CeedCallBackend(CeedGetKernel_Cuda(ceed, data->module, "Interp", &data->Interp)); 679 CeedCallBackend(CeedGetKernel_Cuda(ceed, data->module, "InterpTranspose", &data->InterpTranspose)); 680 CeedCallBackend(CeedGetKernel_Cuda(ceed, data->module, "InterpTransposeAdd", &data->InterpTransposeAdd)); 681 CeedCallBackend(CeedGetKernel_Cuda(ceed, data->module, "Grad", &data->Grad)); 682 CeedCallBackend(CeedGetKernel_Cuda(ceed, data->module, "GradTranspose", &data->GradTranspose)); 683 CeedCallBackend(CeedGetKernel_Cuda(ceed, data->module, "GradTransposeAdd", &data->GradTransposeAdd)); 684 CeedCallBackend(CeedGetKernel_Cuda(ceed, data->module, "Weight", &data->Weight)); 685 686 CeedCallBackend(CeedBasisSetData(basis, data)); 687 688 // Register backend functions 689 CeedCallBackend(CeedSetBackendFunction(ceed, "Basis", basis, "Apply", CeedBasisApplyNonTensor_Cuda_shared)); 690 CeedCallBackend(CeedSetBackendFunction(ceed, "Basis", basis, "ApplyAdd", CeedBasisApplyAddNonTensor_Cuda_shared)); 691 CeedCallBackend(CeedSetBackendFunction(ceed, "Basis", basis, "Destroy", CeedBasisDestroy_Cuda_shared)); 692 CeedCallBackend(CeedDestroy(&ceed)); 693 return CEED_ERROR_SUCCESS; 694 } 695 696 //------------------------------------------------------------------------------ 697