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 #define CEED_DEBUG_COLOR 12 9 10 #include <ceed.h> 11 #include <ceed/backend.h> 12 #include <ceed/jit-tools.h> 13 #include <cuda_runtime.h> 14 15 #include <iostream> 16 #include <sstream> 17 #include <string> 18 19 #include "../cuda-ref/ceed-cuda-ref.h" 20 #include "../cuda-shared/ceed-cuda-shared.h" 21 #include "../cuda/ceed-cuda-common.h" 22 #include "../cuda/ceed-cuda-compile.h" 23 #include "ceed-cuda-gen.h" 24 25 struct FieldReuse_Cuda { 26 CeedInt index; 27 bool is_input; 28 CeedEvalMode eval_mode; 29 }; 30 31 //------------------------------------------------------------------------------ 32 // Determine type of operator 33 //------------------------------------------------------------------------------ 34 static int CeedOperatorBuildKernelData_Cuda_gen(Ceed ceed, CeedInt num_input_fields, CeedOperatorField *op_input_fields, 35 CeedQFunctionField *qf_input_fields, CeedInt num_output_fields, CeedOperatorField *op_output_fields, 36 CeedQFunctionField *qf_output_fields, CeedInt *max_P, CeedInt *max_P_1d, CeedInt *Q, CeedInt *Q_1d, 37 CeedInt *max_dim, bool *is_all_tensor, bool *use_3d_slices) { 38 // Check if all are tensor 39 *is_all_tensor = true; 40 for (CeedInt i = 0; i < num_input_fields; i++) { 41 CeedBasis basis; 42 43 CeedCallBackend(CeedOperatorFieldGetBasis(op_input_fields[i], &basis)); 44 if (basis != CEED_BASIS_NONE) { 45 bool is_field_tensor; 46 47 CeedCallBackend(CeedBasisIsTensor(basis, &is_field_tensor)); 48 *is_all_tensor = *is_all_tensor && is_field_tensor; 49 } 50 CeedCallBackend(CeedBasisDestroy(&basis)); 51 } 52 for (CeedInt i = 0; i < num_output_fields; i++) { 53 CeedBasis basis; 54 55 CeedCallBackend(CeedOperatorFieldGetBasis(op_output_fields[i], &basis)); 56 if (basis != CEED_BASIS_NONE) { 57 bool is_field_tensor; 58 59 CeedCallBackend(CeedBasisIsTensor(basis, &is_field_tensor)); 60 *is_all_tensor = *is_all_tensor && is_field_tensor; 61 } 62 CeedCallBackend(CeedBasisDestroy(&basis)); 63 } 64 65 // Find max_P, max_P_1d, Q, and Q_1d 66 bool is_all_3d = true; 67 68 *max_P = 0; 69 *max_P_1d = 0; 70 *Q = 0; 71 *Q_1d = 0; 72 for (CeedInt i = 0; i < num_input_fields; i++) { 73 CeedBasis basis; 74 75 CeedCallBackend(CeedOperatorFieldGetBasis(op_input_fields[i], &basis)); 76 if (basis != CEED_BASIS_NONE) { 77 bool is_field_tensor; 78 CeedInt field_dim = 0, field_P = 0, field_P_1d = 0, field_Q = 0, field_Q_1d = 0; 79 80 // Check if 3D 81 CeedCallBackend(CeedBasisGetDimension(basis, &field_dim)); 82 is_all_3d = is_all_3d && (field_dim == 3); 83 *max_dim = CeedIntMax(*max_dim, field_dim); 84 85 // Collect P, P_1d, Q, and Q_1d 86 CeedCallBackend(CeedBasisGetNumNodes(basis, &field_P)); 87 *max_P = CeedIntMax(*max_P, field_P); 88 CeedCallBackend(CeedBasisIsTensor(basis, &is_field_tensor)); 89 if (is_field_tensor) { 90 CeedCallBackend(CeedBasisGetNumNodes1D(basis, &field_P_1d)); 91 *max_P_1d = CeedIntMax(*max_P_1d, field_P_1d); 92 } 93 CeedCallBackend(CeedBasisGetNumQuadraturePoints(basis, &field_Q)); 94 CeedCheck(*Q == 0 || field_Q == *Q, ceed, CEED_ERROR_BACKEND, "Quadrature spaces must be compatible"); 95 *Q = field_Q; 96 if (is_field_tensor) { 97 CeedCallBackend(CeedBasisGetNumQuadraturePoints1D(basis, &field_Q_1d)); 98 CeedCheck(*Q_1d == 0 || field_Q_1d == *Q_1d, ceed, CEED_ERROR_BACKEND, "Quadrature spaces must be compatible"); 99 *Q_1d = field_Q_1d; 100 } 101 } 102 CeedCallBackend(CeedBasisDestroy(&basis)); 103 } 104 for (CeedInt i = 0; i < num_output_fields; i++) { 105 CeedBasis basis; 106 107 CeedCallBackend(CeedOperatorFieldGetBasis(op_output_fields[i], &basis)); 108 if (basis != CEED_BASIS_NONE) { 109 bool is_field_tensor; 110 CeedInt field_dim = 0, field_P = 0, field_P_1d = 0, field_Q = 0, field_Q_1d = 0; 111 112 // Check if 3D 113 CeedCallBackend(CeedBasisGetDimension(basis, &field_dim)); 114 is_all_3d = is_all_3d && (field_dim == 3); 115 *max_dim = CeedIntMax(*max_dim, field_dim); 116 117 // Collect P, P_1d, Q, and Q_1d 118 CeedCallBackend(CeedBasisGetNumNodes(basis, &field_P)); 119 *max_P = CeedIntMax(*max_P, field_P); 120 CeedCallBackend(CeedBasisIsTensor(basis, &is_field_tensor)); 121 if (is_field_tensor) { 122 CeedCallBackend(CeedBasisGetNumNodes1D(basis, &field_P_1d)); 123 *max_P_1d = CeedIntMax(*max_P_1d, field_P_1d); 124 } 125 CeedCallBackend(CeedBasisGetNumQuadraturePoints(basis, &field_Q)); 126 CeedCheck(*Q == 0 || field_Q == *Q, ceed, CEED_ERROR_BACKEND, "Quadrature spaces must be compatible"); 127 *Q = field_Q; 128 if (is_field_tensor) { 129 CeedCallBackend(CeedBasisGetNumQuadraturePoints1D(basis, &field_Q_1d)); 130 CeedCheck(*Q_1d == 0 || field_Q_1d == *Q_1d, ceed, CEED_ERROR_BACKEND, "Quadrature spaces must be compatible"); 131 *Q_1d = field_Q_1d; 132 } 133 } 134 CeedCallBackend(CeedBasisDestroy(&basis)); 135 } 136 137 // Only use 3D collocated gradient parallelization strategy when gradient is computed 138 *use_3d_slices = false; 139 if (is_all_3d && *is_all_tensor) { 140 bool was_grad_found = false; 141 142 for (CeedInt i = 0; i < num_input_fields; i++) { 143 CeedEvalMode eval_mode; 144 145 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_input_fields[i], &eval_mode)); 146 if (eval_mode == CEED_EVAL_GRAD) { 147 CeedBasis_Cuda_shared *basis_data; 148 CeedBasis basis; 149 150 CeedCallBackend(CeedOperatorFieldGetBasis(op_input_fields[i], &basis)); 151 CeedCallBackend(CeedBasisGetData(basis, &basis_data)); 152 *use_3d_slices = basis_data->d_collo_grad_1d && (was_grad_found ? *use_3d_slices : true); 153 was_grad_found = true; 154 CeedCallBackend(CeedBasisDestroy(&basis)); 155 } 156 } 157 for (CeedInt i = 0; i < num_output_fields; i++) { 158 CeedEvalMode eval_mode; 159 160 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_output_fields[i], &eval_mode)); 161 if (eval_mode == CEED_EVAL_GRAD) { 162 CeedBasis_Cuda_shared *basis_data; 163 CeedBasis basis; 164 165 CeedCallBackend(CeedOperatorFieldGetBasis(op_output_fields[i], &basis)); 166 CeedCallBackend(CeedBasisGetData(basis, &basis_data)); 167 *use_3d_slices = basis_data->d_collo_grad_1d && (was_grad_found ? *use_3d_slices : true); 168 was_grad_found = true; 169 CeedCallBackend(CeedBasisDestroy(&basis)); 170 } 171 } 172 } 173 return CEED_ERROR_SUCCESS; 174 } 175 176 //------------------------------------------------------------------------------ 177 // Setup fields 178 //------------------------------------------------------------------------------ 179 static int CeedOperatorBuildKernelFieldData_Cuda_gen(std::ostringstream &code, CeedOperator_Cuda_gen *data, CeedInt i, CeedOperatorField op_field, 180 CeedQFunctionField qf_field, FieldReuse_Cuda field_reuse, CeedInt max_dim, CeedInt Q, 181 CeedInt Q_1d, bool is_input, bool is_all_tensor, bool is_at_points, bool use_3d_slices) { 182 bool is_tensor = true; 183 CeedBasis basis; 184 CeedCallBackend(CeedOperatorFieldGetBasis(op_field, &basis)); 185 if (basis != CEED_BASIS_NONE) CeedCallBackend(CeedBasisIsTensor(basis, &is_tensor)); 186 187 const char *field_name; 188 std::string var_suffix = (is_input ? "_in_" : "_out_") + std::to_string(i); 189 std::string P_name = (is_tensor ? "P_1d" : "P") + var_suffix, Q_name = is_tensor ? "Q_1d" : "Q"; 190 std::string option_name = (is_input ? "inputs" : "outputs"); 191 CeedEvalMode eval_mode = CEED_EVAL_NONE; 192 CeedInt elem_size = 0, num_comp = 0, dim = max_dim, P_1d = 0; 193 CeedElemRestriction elem_rstr; 194 CeedBasis_Cuda_shared *basis_data; 195 196 // Field reuse info 197 bool use_previous_field = field_reuse.index != -1; 198 199 CeedCallBackend(CeedOperatorFieldGetName(op_field, &field_name)); 200 code << " // -- " << (is_input ? "Input" : "Output") << " field " << i << ": " << field_name << "\n"; 201 202 // Get field data 203 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_field, &elem_rstr)); 204 if (elem_rstr != CEED_ELEMRESTRICTION_NONE) { 205 CeedCallBackend(CeedElemRestrictionGetElementSize(elem_rstr, &elem_size)); 206 CeedCallBackend(CeedElemRestrictionGetNumComponents(elem_rstr, &num_comp)); 207 } 208 CeedCallBackend(CeedElemRestrictionDestroy(&elem_rstr)); 209 if (basis != CEED_BASIS_NONE) { 210 CeedCallBackend(CeedBasisGetData(basis, &basis_data)); 211 CeedCallBackend(CeedBasisGetDimension(basis, &dim)); 212 if (is_tensor) CeedCallBackend(CeedBasisGetNumNodes1D(basis, &P_1d)); 213 else CeedCallBackend(CeedBasisGetNumNodes(basis, &P_1d)); 214 } 215 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_field, &eval_mode)); 216 217 // Set field constants 218 code << " const CeedInt dim" << var_suffix << " = " << dim << ";\n"; 219 if (is_tensor && !is_all_tensor) { 220 CeedInt P = 0; 221 222 CeedCallBackend(CeedBasisGetNumNodes(basis, &P)); 223 code << " const CeedInt P" << var_suffix << " = " << (basis == CEED_BASIS_NONE ? Q : P) << ";\n"; 224 } 225 code << " const CeedInt " << P_name << " = " << (basis == CEED_BASIS_NONE ? Q_1d : P_1d) << ";\n"; 226 if (eval_mode != CEED_EVAL_WEIGHT) { 227 code << " const CeedInt num_comp" << var_suffix << " = " << num_comp << ";\n"; 228 } 229 230 // Load basis data 231 code << " // EvalMode: " << CeedEvalModes[eval_mode] << "\n"; 232 switch (eval_mode) { 233 case CEED_EVAL_NONE: 234 break; 235 case CEED_EVAL_INTERP: 236 if (is_at_points) { 237 // AtPoints 238 if (!basis_data->d_chebyshev_interp_1d) { 239 CeedSize interp_bytes; 240 CeedScalar *chebyshev_interp_1d; 241 242 interp_bytes = P_1d * Q_1d * sizeof(CeedScalar); 243 CeedCallBackend(CeedCalloc(P_1d * Q_1d, &chebyshev_interp_1d)); 244 CeedCallBackend(CeedBasisGetChebyshevInterp1D(basis, chebyshev_interp_1d)); 245 CeedCallCuda(CeedBasisReturnCeed(basis), cudaMalloc((void **)&basis_data->d_chebyshev_interp_1d, interp_bytes)); 246 CeedCallCuda(CeedBasisReturnCeed(basis), 247 cudaMemcpy(basis_data->d_chebyshev_interp_1d, chebyshev_interp_1d, interp_bytes, cudaMemcpyHostToDevice)); 248 CeedCallBackend(CeedFree(&chebyshev_interp_1d)); 249 } 250 if (is_input) data->B.inputs[i] = basis_data->d_chebyshev_interp_1d; 251 else data->B.outputs[i] = basis_data->d_chebyshev_interp_1d; 252 } else { 253 // Standard quadrature 254 if (is_input) data->B.inputs[i] = basis_data->d_interp_1d; 255 else data->B.outputs[i] = basis_data->d_interp_1d; 256 } 257 if (use_previous_field) { 258 std::string reuse_var = "s_B" + ((field_reuse.is_input ? "_in_" : "_out_") + std::to_string(field_reuse.index)); 259 260 code << " CeedScalar *s_B" << var_suffix << " = " << reuse_var << ";\n"; 261 } else { 262 code << " __shared__ CeedScalar s_B" << var_suffix << "[" << P_name << "*" << Q_name << "];\n"; 263 code << " LoadMatrix<" << P_name << ", " << Q_name << ">(data, B." << option_name << "[" << i << "], s_B" << var_suffix << ");\n"; 264 } 265 break; 266 case CEED_EVAL_GRAD: 267 if (is_at_points) { 268 // AtPoints 269 if (!basis_data->d_chebyshev_interp_1d) { 270 CeedSize interp_bytes; 271 CeedScalar *chebyshev_interp_1d; 272 273 interp_bytes = P_1d * Q_1d * sizeof(CeedScalar); 274 CeedCallBackend(CeedCalloc(P_1d * Q_1d, &chebyshev_interp_1d)); 275 CeedCallBackend(CeedBasisGetChebyshevInterp1D(basis, chebyshev_interp_1d)); 276 CeedCallCuda(CeedBasisReturnCeed(basis), cudaMalloc((void **)&basis_data->d_chebyshev_interp_1d, interp_bytes)); 277 CeedCallCuda(CeedBasisReturnCeed(basis), 278 cudaMemcpy(basis_data->d_chebyshev_interp_1d, chebyshev_interp_1d, interp_bytes, cudaMemcpyHostToDevice)); 279 CeedCallBackend(CeedFree(&chebyshev_interp_1d)); 280 } 281 if (is_input) data->B.inputs[i] = basis_data->d_chebyshev_interp_1d; 282 else data->B.outputs[i] = basis_data->d_chebyshev_interp_1d; 283 } else { 284 // Standard quadrature 285 if (is_input) data->B.inputs[i] = basis_data->d_interp_1d; 286 else data->B.outputs[i] = basis_data->d_interp_1d; 287 } 288 if (is_tensor) { 289 if (use_previous_field) { 290 std::string reuse_var = "s_B" + ((field_reuse.is_input ? "_in_" : "_out_") + std::to_string(field_reuse.index)); 291 292 code << " CeedScalar *s_B" << var_suffix << " = " << reuse_var << ";\n"; 293 } else { 294 code << " __shared__ CeedScalar s_B" << var_suffix << "[" << P_name << "*" << Q_name << "];\n"; 295 code << " LoadMatrix<" << P_name << ", " << Q_name << ">(data, B." << option_name << "[" << i << "], s_B" << var_suffix << ");\n"; 296 } 297 } 298 if (is_at_points) break; // No G mat for AtPoints 299 if (use_3d_slices) { 300 if (is_input) data->G.inputs[i] = basis_data->d_collo_grad_1d; 301 else data->G.outputs[i] = basis_data->d_collo_grad_1d; 302 if (use_previous_field && field_reuse.eval_mode == CEED_EVAL_GRAD) { 303 std::string reuse_var = "s_G" + ((field_reuse.is_input ? "_in_" : "_out_") + std::to_string(field_reuse.index)); 304 305 code << " CeedScalar *s_G" << var_suffix << " = " << reuse_var << ";\n"; 306 } else { 307 code << " __shared__ CeedScalar s_G" << var_suffix << "[" << Q_name << "*" << Q_name << "];\n"; 308 code << " LoadMatrix<" << Q_name << ", " << Q_name << ">(data, G." << option_name << "[" << i << "], s_G" << var_suffix << ");\n"; 309 } 310 } else { 311 bool has_collo_grad = basis_data->d_collo_grad_1d; 312 313 if (is_input) data->G.inputs[i] = has_collo_grad ? basis_data->d_collo_grad_1d : basis_data->d_grad_1d; 314 else data->G.outputs[i] = has_collo_grad ? basis_data->d_collo_grad_1d : basis_data->d_grad_1d; 315 if (has_collo_grad) { 316 if (use_previous_field && field_reuse.eval_mode == CEED_EVAL_GRAD) { 317 std::string reuse_var = "s_G" + ((field_reuse.is_input ? "_in_" : "_out_") + std::to_string(field_reuse.index)); 318 319 code << " CeedScalar *s_G" << var_suffix << " = " << reuse_var << ";\n"; 320 } else { 321 code << " __shared__ CeedScalar s_G" << var_suffix << "[" << Q_name << "*" << Q_name << "];\n"; 322 code << " LoadMatrix<" << Q_name << ", " << Q_name << ">(data, G." << option_name << "[" << i << "], s_G" << var_suffix << ");\n"; 323 } 324 } else { 325 if (use_previous_field && field_reuse.eval_mode == CEED_EVAL_GRAD) { 326 std::string reuse_var = "s_G" + ((field_reuse.is_input ? "_in_" : "_out_") + std::to_string(field_reuse.index)); 327 328 code << " CeedScalar *s_G" << var_suffix << " = " << reuse_var << ";\n"; 329 } else { 330 code << " __shared__ CeedScalar s_G" << var_suffix << "[" << P_name << "*" << Q_name << (is_tensor ? "" : "*dim") 331 << (is_tensor ? "" : var_suffix) << "];\n"; 332 code << " LoadMatrix<" << P_name << ", " << Q_name << (is_tensor ? "" : "*dim") << (is_tensor ? "" : var_suffix) << ">(data, G." 333 << option_name << "[" << i << "], s_G" << var_suffix << ");\n"; 334 } 335 } 336 } 337 break; 338 case CEED_EVAL_WEIGHT: 339 break; // No action 340 // LCOV_EXCL_START 341 case CEED_EVAL_DIV: 342 case CEED_EVAL_CURL: 343 break; // TODO: Not implemented 344 // LCOV_EXCL_STOP 345 } 346 CeedCallBackend(CeedBasisDestroy(&basis)); 347 return CEED_ERROR_SUCCESS; 348 } 349 350 //------------------------------------------------------------------------------ 351 // Restriction 352 //------------------------------------------------------------------------------ 353 static int CeedOperatorBuildKernelRestriction_Cuda_gen(std::ostringstream &code, CeedOperator_Cuda_gen *data, CeedInt i, CeedInt field_input_buffer[], 354 CeedOperatorField op_field, CeedQFunctionField qf_field, CeedInt max_dim, CeedInt Q_1d, 355 bool is_input, bool is_all_tensor, bool is_at_points, bool use_3d_slices) { 356 std::string var_suffix = (is_input ? "_in_" : "_out_") + std::to_string(i); 357 std::string P_name = (is_all_tensor ? "P_1d" : "P") + var_suffix; 358 CeedEvalMode eval_mode = CEED_EVAL_NONE; 359 CeedInt elem_size = 0, num_comp = 0; 360 CeedSize l_size; 361 CeedRestrictionType rstr_type = CEED_RESTRICTION_STANDARD; 362 CeedElemRestriction_Cuda *rstr_data; 363 CeedElemRestriction elem_rstr; 364 365 // Get field data 366 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_field, &elem_rstr)); 367 if (elem_rstr != CEED_ELEMRESTRICTION_NONE) { 368 CeedCallBackend(CeedElemRestrictionGetType(elem_rstr, &rstr_type)); 369 CeedCallBackend(CeedElemRestrictionGetElementSize(elem_rstr, &elem_size)); 370 CeedCallBackend(CeedElemRestrictionGetNumComponents(elem_rstr, &num_comp)); 371 CeedCallBackend(CeedElemRestrictionGetData(elem_rstr, &rstr_data)); 372 } 373 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_field, &eval_mode)); 374 375 // Restriction 376 if (is_input) { 377 // Input 378 if (field_input_buffer[i] != i) { 379 std::string buffer_name = "r_e_in_" + std::to_string(field_input_buffer[i]); 380 381 // Restriction was already done for previous input 382 code << " CeedScalar *r_e" << var_suffix << " = " << buffer_name << ";\n"; 383 } else if (eval_mode != CEED_EVAL_WEIGHT && !((eval_mode == CEED_EVAL_NONE) && use_3d_slices && is_at_points)) { 384 if (eval_mode == CEED_EVAL_NONE && rstr_type != CEED_RESTRICTION_POINTS) { 385 // No basis action, so r_e_in_* in also r_q_in_* and needs to be allocated 386 code << " CeedScalar r_e" << var_suffix << "[num_comp" << var_suffix << "*" << P_name << "];\n"; 387 } else if (rstr_type != CEED_RESTRICTION_POINTS) { 388 // Otherwise we're using the scratch space 389 code << " CeedScalar *r_e" << var_suffix << " = r_e_scratch;\n"; 390 } 391 switch (rstr_type) { 392 case CEED_RESTRICTION_STANDARD: { 393 CeedInt comp_stride; 394 395 CeedCallBackend(CeedElemRestrictionGetLVectorSize(elem_rstr, &l_size)); 396 code << " const CeedInt l_size" << var_suffix << " = " << l_size << ";\n"; 397 CeedCallBackend(CeedElemRestrictionGetCompStride(elem_rstr, &comp_stride)); 398 code << " // CompStride: " << comp_stride << "\n"; 399 data->indices.inputs[i] = (CeedInt *)rstr_data->d_offsets; 400 code << " ReadLVecStandard" << (is_all_tensor ? max_dim : 1) << "d<num_comp" << var_suffix << ", " << comp_stride << ", " << P_name 401 << ">(data, l_size" << var_suffix << ", elem, indices.inputs[" << i << "], d" << var_suffix << ", r_e" << var_suffix << ");\n"; 402 break; 403 } 404 case CEED_RESTRICTION_STRIDED: { 405 bool has_backend_strides; 406 CeedInt num_elem; 407 408 CeedCallBackend(CeedElemRestrictionHasBackendStrides(elem_rstr, &has_backend_strides)); 409 CeedCallBackend(CeedElemRestrictionGetNumElements(elem_rstr, &num_elem)); 410 CeedInt strides[3] = {1, elem_size * num_elem, elem_size}; 411 412 if (!has_backend_strides) { 413 CeedCallBackend(CeedElemRestrictionGetStrides(elem_rstr, strides)); 414 } 415 code << " // Strides: {" << strides[0] << ", " << strides[1] << ", " << strides[2] << "}\n"; 416 code << " ReadLVecStrided" << (is_all_tensor ? max_dim : 1) << "d<num_comp" << var_suffix << ", " << P_name << ", " << strides[0] << ", " 417 << strides[1] << ", " << strides[2] << ">(data, elem, d" << var_suffix << ", r_e" << var_suffix << ");\n"; 418 break; 419 } 420 case CEED_RESTRICTION_POINTS: { 421 CeedInt comp_stride; 422 423 CeedCallBackend(CeedElemRestrictionGetCompStride(elem_rstr, &comp_stride)); 424 code << " const CeedInt comp_stride" << var_suffix << " = " << comp_stride << ";\n"; 425 data->indices.inputs[i] = (CeedInt *)rstr_data->d_offsets; 426 break; 427 } 428 // LCOV_EXCL_START 429 case CEED_RESTRICTION_ORIENTED: 430 case CEED_RESTRICTION_CURL_ORIENTED: 431 break; // TODO: Not implemented 432 // LCOV_EXCL_STOP 433 } 434 } 435 } else { 436 // Output 437 switch (rstr_type) { 438 case CEED_RESTRICTION_STANDARD: { 439 CeedInt comp_stride; 440 441 CeedCallBackend(CeedElemRestrictionGetLVectorSize(elem_rstr, &l_size)); 442 code << " const CeedInt l_size" << var_suffix << " = " << l_size << ";\n"; 443 CeedCallBackend(CeedElemRestrictionGetCompStride(elem_rstr, &comp_stride)); 444 code << " // CompStride: " << comp_stride << "\n"; 445 data->indices.outputs[i] = (CeedInt *)rstr_data->d_offsets; 446 code << " WriteLVecStandard" << (is_all_tensor ? max_dim : 1) << "d<num_comp" << var_suffix << ", " << comp_stride << ", " << P_name 447 << ">(data, l_size" << var_suffix << ", elem, indices.outputs[" << i << "], r_e" << var_suffix << ", d" << var_suffix << ");\n"; 448 break; 449 } 450 case CEED_RESTRICTION_STRIDED: { 451 bool has_backend_strides; 452 CeedInt num_elem; 453 454 CeedCallBackend(CeedElemRestrictionHasBackendStrides(elem_rstr, &has_backend_strides)); 455 CeedCallBackend(CeedElemRestrictionGetNumElements(elem_rstr, &num_elem)); 456 CeedInt strides[3] = {1, elem_size * num_elem, elem_size}; 457 458 if (!has_backend_strides) { 459 CeedCallBackend(CeedElemRestrictionGetStrides(elem_rstr, strides)); 460 } 461 code << " // Strides: {" << strides[0] << ", " << strides[1] << ", " << strides[2] << "}\n"; 462 code << " WriteLVecStrided" << (is_all_tensor ? max_dim : 1) << "d<num_comp" << var_suffix << ", " << P_name << ", " << strides[0] << ", " 463 << strides[1] << ", " << strides[2] << ">(data, elem, r_e" << var_suffix << ", d" << var_suffix << ");\n"; 464 break; 465 } 466 case CEED_RESTRICTION_POINTS: 467 data->indices.outputs[i] = (CeedInt *)rstr_data->d_offsets; 468 break; 469 // LCOV_EXCL_START 470 case CEED_RESTRICTION_ORIENTED: 471 case CEED_RESTRICTION_CURL_ORIENTED: 472 break; // TODO: Not implemented 473 // LCOV_EXCL_STOP 474 } 475 } 476 CeedCallBackend(CeedElemRestrictionDestroy(&elem_rstr)); 477 return CEED_ERROR_SUCCESS; 478 } 479 480 //------------------------------------------------------------------------------ 481 // Basis 482 //------------------------------------------------------------------------------ 483 static int CeedOperatorBuildKernelBasis_Cuda_gen(std::ostringstream &code, CeedOperator_Cuda_gen *data, CeedInt i, CeedOperatorField op_field, 484 CeedQFunctionField qf_field, CeedInt max_dim, CeedInt Q_1d, bool is_input, bool is_all_tensor, 485 bool is_at_points, bool use_3d_slices) { 486 bool is_tensor = true; 487 CeedBasis basis; 488 CeedCallBackend(CeedOperatorFieldGetBasis(op_field, &basis)); 489 CeedCallBackend(CeedBasisIsTensor(basis, &is_tensor)); 490 491 std::string var_suffix = (is_input ? "_in_" : "_out_") + std::to_string(i); 492 std::string P_name = (is_tensor ? "P_1d" : "P") + var_suffix, Q_name = is_tensor ? "Q_1d" : "Q"; 493 CeedEvalMode eval_mode = CEED_EVAL_NONE; 494 CeedInt dim = max_dim, elem_size = 0, num_comp = 0, P_1d = 0; 495 CeedElemRestriction elem_rstr; 496 497 // Get field data 498 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_field, &elem_rstr)); 499 if (elem_rstr != CEED_ELEMRESTRICTION_NONE) { 500 CeedCallBackend(CeedElemRestrictionGetElementSize(elem_rstr, &elem_size)); 501 CeedCallBackend(CeedElemRestrictionGetNumComponents(elem_rstr, &num_comp)); 502 } 503 CeedCallBackend(CeedElemRestrictionDestroy(&elem_rstr)); 504 if (basis != CEED_BASIS_NONE) { 505 CeedCallBackend(CeedBasisGetDimension(basis, &dim)); 506 if (is_tensor) CeedCallBackend(CeedBasisGetNumNodes1D(basis, &P_1d)); 507 else CeedCallBackend(CeedBasisGetNumNodes(basis, &P_1d)); 508 } 509 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_field, &eval_mode)); 510 511 // Basis 512 code << " // EvalMode: " << CeedEvalModes[eval_mode] << "\n"; 513 if (is_input) { 514 switch (eval_mode) { 515 case CEED_EVAL_NONE: 516 if (!use_3d_slices && !is_at_points) { 517 code << " CeedScalar *r_q" << var_suffix << " = r_e" << var_suffix << ";\n"; 518 } 519 break; 520 case CEED_EVAL_INTERP: 521 if (is_at_points) { 522 std::string function_name = (dim == 1 ? "Interp" : "InterpTensor") + std::to_string(dim) + "d"; 523 524 code << " CeedScalar r_c" << var_suffix << "[num_comp" << var_suffix << "*" << (dim >= 3 ? Q_name : "1") << "];\n"; 525 code << " " << function_name << "<num_comp" << var_suffix << ", " << P_name << ", " << Q_name << ", OP_T_1D>(data, r_e" << var_suffix 526 << ", s_B" << var_suffix << ", r_c" << var_suffix << ");\n"; 527 } else { 528 std::string function_name = is_tensor 529 ? ((dim == 1 ? "Interp" : "InterpTensor") + std::to_string(dim) + "d" + (is_all_tensor ? "" : "Flattened")) 530 : "InterpNonTensor"; 531 std::string op_t_1d_name = (is_all_tensor || !is_tensor) ? "OP_T_1D" : (P_1d > Q_1d ? P_name : Q_name); 532 533 code << " CeedScalar r_q" << var_suffix << "[num_comp" << var_suffix << "*" << (is_all_tensor && (dim >= 3) ? Q_name : "1") << "];\n"; 534 code << " " << function_name << "<num_comp" << var_suffix << ", " << P_name << ", " << Q_name << ", " << op_t_1d_name << ">(data, r_e" 535 << var_suffix << ", s_B" << var_suffix << ", r_q" << var_suffix << ");\n"; 536 } 537 break; 538 case CEED_EVAL_GRAD: 539 if (is_at_points) { 540 std::string function_name = (dim == 1 ? "Interp" : "InterpTensor") + std::to_string(dim) + "d"; 541 542 code << " CeedScalar r_c" << var_suffix << "[num_comp" << var_suffix << "*" << (dim >= 3 ? Q_name : "1") << "];\n"; 543 code << " " << function_name << "<num_comp" << var_suffix << ", " << P_name << ", " << Q_name << ", OP_T_1D>(data, r_e" << var_suffix 544 << ", s_B" << var_suffix << ", r_c" << var_suffix << ");\n"; 545 } else if (use_3d_slices) { 546 std::string function_name = (dim > 1 ? "InterpTensor" : "Interp") + std::to_string(dim) + "d"; 547 548 code << " CeedScalar r_q" << var_suffix << "[num_comp" << var_suffix << "*" << Q_name << "];\n"; 549 code << " " << function_name << "<num_comp" << var_suffix << ", " << P_name << ", " << Q_name << ", OP_T_1D>(data, r_e" << var_suffix 550 << ", s_B" << var_suffix << ", r_q" << var_suffix << ");\n"; 551 } else if (is_tensor) { 552 bool is_collocated = dim == 3 && Q_1d >= P_1d; 553 std::string function_name = (dim == 1 ? "Grad" : (is_collocated ? "GradTensorCollocated" : "GradTensor")) + std::to_string(dim) + "d" + 554 (is_all_tensor ? "" : "Flattened"); 555 std::string op_t_1d_name = is_all_tensor ? "OP_T_1D" : (P_1d > Q_1d ? P_name : Q_name); 556 557 code << " CeedScalar r_q" << var_suffix << "[num_comp" << var_suffix << "*dim" << var_suffix << "*" 558 << (is_all_tensor && dim >= 3 ? Q_name : "1") << "];\n"; 559 code << " " << function_name << "<num_comp" << var_suffix << ", " << P_name << ", " << Q_name << ", " << op_t_1d_name << ">(data, r_e" 560 << var_suffix << ", s_B" << var_suffix << ", s_G" << var_suffix << ", r_q" << var_suffix << ");\n"; 561 } else { 562 std::string function_name = "GradNonTensor"; 563 564 code << " CeedScalar r_q" << var_suffix << "[num_comp" << var_suffix << "*dim" << var_suffix << "];\n"; 565 code << " " << function_name << "<num_comp" << var_suffix << ", dim" << var_suffix << ", " << P_name << ", " << Q_name 566 << ", OP_T_1D>(data, r_e" << var_suffix << ", s_G" << var_suffix << ", r_q" << var_suffix << ");\n"; 567 } 568 break; 569 case CEED_EVAL_WEIGHT: { 570 if (is_at_points) { 571 code << " // Nothing to do AtPoints\n"; 572 } else { 573 CeedBasis_Cuda_shared *basis_data; 574 std::string function_name = is_tensor 575 ? ((dim == 1 ? "Weight" : "WeightTensor") + std::to_string(dim) + "d" + (is_all_tensor ? "" : "Flattened")) 576 : "WeightNonTensor"; 577 578 code << " CeedScalar r_q" << var_suffix << "[" << (is_all_tensor && (dim >= 3) ? Q_name : "1") << "];\n"; 579 CeedCallBackend(CeedBasisGetData(basis, &basis_data)); 580 data->W = basis_data->d_q_weight_1d; 581 code << " " << function_name << "<" << P_name << ", " << Q_name << ">(data, W, r_q" << var_suffix << ");\n"; 582 } 583 break; 584 } 585 // LCOV_EXCL_START 586 case CEED_EVAL_DIV: 587 case CEED_EVAL_CURL: 588 break; // TODO: Not implemented 589 // LCOV_EXCL_STOP 590 } 591 } else { 592 switch (eval_mode) { 593 case CEED_EVAL_NONE: 594 code << " CeedScalar *r_e" << var_suffix << " = r_q" << var_suffix << ";\n"; 595 break; // No action 596 case CEED_EVAL_INTERP: 597 code << " CeedScalar *r_e" << var_suffix << " = r_e_scratch;\n"; 598 if (is_at_points) { 599 std::string function_name = (dim == 1 ? "InterpTranspose" : "InterpTransposeTensor") + std::to_string(dim) + "d"; 600 601 code << " " << function_name << "<num_comp" << var_suffix << ", " << P_name << ", " << Q_name << ", OP_T_1D>(data, r_c" << var_suffix 602 << ", s_B" << var_suffix << ", r_e" << var_suffix << ");\n"; 603 } else { 604 std::string function_name = 605 is_tensor ? ((dim == 1 ? "InterpTranspose" : "InterpTransposeTensor") + std::to_string(dim) + "d" + (is_all_tensor ? "" : "Flattened")) 606 : "InterpTransposeNonTensor"; 607 std::string op_t_1d_name = (is_all_tensor || !is_tensor) ? "OP_T_1D" : (P_1d > Q_1d ? P_name : Q_name); 608 609 code << " " << function_name << "<num_comp" << var_suffix << ", " << P_name << ", " << Q_name << ", " << op_t_1d_name << ">(data, r_q" 610 << var_suffix << ", s_B" << var_suffix << ", r_e" << var_suffix << ");\n"; 611 } 612 break; 613 case CEED_EVAL_GRAD: 614 code << " CeedScalar *r_e" << var_suffix << " = r_e_scratch;\n"; 615 if (is_at_points) { 616 std::string function_name = (dim == 1 ? "InterpTranspose" : "InterpTransposeTensor") + std::to_string(dim) + "d"; 617 618 code << " " << function_name << "<num_comp" << var_suffix << ", " << P_name << ", " << Q_name << ", OP_T_1D>(data, r_c" << var_suffix 619 << ", s_B" << var_suffix << ", r_e" << var_suffix << ");\n"; 620 } else if (use_3d_slices) { 621 std::string function_name = (dim == 1 ? "InterpTranspose" : "InterpTransposeTensor") + std::to_string(dim) + "d"; 622 623 code << " " << function_name << "<num_comp" << var_suffix << ", " << P_name << ", " << Q_name << ", OP_T_1D>(data, r_q" << var_suffix 624 << ", s_B" << var_suffix << ", r_e" << var_suffix << ");\n"; 625 } else if (is_tensor) { 626 bool is_collocated = dim == 3 && Q_1d >= P_1d; 627 std::string function_name = (dim == 1 ? "GradTranspose" : (is_collocated ? "GradTransposeTensorCollocated" : "GradTransposeTensor")) + 628 std::to_string(dim) + "d" + (is_all_tensor ? "" : "Flattened"); 629 std::string op_t_1d_name = is_all_tensor ? "OP_T_1D" : (P_1d > Q_1d ? P_name : Q_name); 630 631 code << " " << function_name << "<num_comp" << var_suffix << ", " << P_name << ", " << Q_name << ", " << op_t_1d_name << ">(data, r_q" 632 << var_suffix << ", s_B" << var_suffix << ", s_G" << var_suffix << ", r_e" << var_suffix << ");\n"; 633 } else { 634 std::string function_name = "GradTransposeNonTensor"; 635 636 code << " " << function_name << "<num_comp" << var_suffix << ", dim" << var_suffix << ", " << P_name << ", " << Q_name 637 << ", OP_T_1D>(data, r_q" << var_suffix << ", s_G" << var_suffix << ", r_e" << var_suffix << ");\n"; 638 } 639 break; 640 // LCOV_EXCL_START 641 case CEED_EVAL_WEIGHT: 642 break; // Should not occur 643 case CEED_EVAL_DIV: 644 case CEED_EVAL_CURL: 645 break; // TODO: Not implemented 646 // LCOV_EXCL_STOP 647 } 648 } 649 CeedCallBackend(CeedBasisDestroy(&basis)); 650 return CEED_ERROR_SUCCESS; 651 } 652 653 //------------------------------------------------------------------------------ 654 // QFunction 655 //------------------------------------------------------------------------------ 656 static int CeedOperatorBuildKernelQFunction_Cuda_gen(std::ostringstream &code, CeedOperator_Cuda_gen *data, CeedInt max_dim, CeedInt max_num_points, 657 CeedInt num_input_fields, CeedOperatorField *op_input_fields, 658 CeedQFunctionField *qf_input_fields, CeedInt num_output_fields, 659 CeedOperatorField *op_output_fields, CeedQFunctionField *qf_output_fields, 660 std::string qfunction_name, CeedInt Q_1d, bool is_all_tensor, bool is_at_points, 661 bool use_3d_slices) { 662 std::string Q_name = is_all_tensor ? "Q_1d" : "Q"; 663 CeedEvalMode eval_mode = CEED_EVAL_NONE; 664 CeedElemRestriction elem_rstr; 665 666 // Setup output arrays 667 code << "\n // -- Output field setup\n"; 668 for (CeedInt i = 0; i < num_output_fields; i++) { 669 const char *field_name; 670 std::string var_suffix = "_out_" + std::to_string(i); 671 672 CeedCallBackend(CeedOperatorFieldGetName(op_output_fields[i], &field_name)); 673 code << " // ---- Output field " << i << ": " << field_name << "\n"; 674 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_output_fields[i], &eval_mode)); 675 switch (eval_mode) { 676 case CEED_EVAL_NONE: 677 if (is_at_points) { 678 code << " CeedScalar r_q" << var_suffix << "[num_comp" << var_suffix << "];\n"; 679 } else { 680 code << " CeedScalar r_q" << var_suffix << "[num_comp" << var_suffix << "*" << (is_all_tensor && (max_dim >= 3) ? Q_name : "1") 681 << "];\n"; 682 } 683 break; 684 case CEED_EVAL_INTERP: 685 if (is_at_points) { 686 // Accumulator for point data 687 code << " CeedScalar r_c" << var_suffix << "[num_comp" << var_suffix << "*" << (max_dim >= 3 ? Q_name : "1") << "];\n"; 688 code << " for (CeedInt i = 0; i < num_comp" << var_suffix << "*" << (max_dim >= 3 ? Q_name : "1") << "; i++) {\n"; 689 code << " r_c" << var_suffix << "[i] = 0.0;\n"; 690 code << " }\n"; 691 } else { 692 code << " CeedScalar r_q" << var_suffix << "[num_comp" << var_suffix << "*" << (is_all_tensor && (max_dim >= 3) ? Q_name : "1") 693 << "];\n"; 694 } 695 break; 696 case CEED_EVAL_GRAD: 697 if (is_at_points) { 698 // Accumulator for point data 699 code << " CeedScalar r_c" << var_suffix << "[num_comp" << var_suffix << "*" << (max_dim >= 3 ? Q_name : "1") << "];\n"; 700 code << " for (CeedInt i = 0; i < num_comp" << var_suffix << "*" << (max_dim >= 3 ? Q_name : "1") << "; i++) {\n"; 701 code << " r_c" << var_suffix << "[i] = 0.0;\n"; 702 code << " }\n"; 703 } else if (use_3d_slices) { 704 // Accumulator for gradient slices 705 code << " CeedScalar r_q" << var_suffix << "[num_comp" << var_suffix << "*" << Q_name << "];\n"; 706 code << " for (CeedInt i = 0; i < num_comp" << var_suffix << "*" << Q_name << "; i++) {\n"; 707 code << " r_q" << var_suffix << "[i] = 0.0;\n"; 708 code << " }\n"; 709 } else { 710 code << " CeedScalar r_q" << var_suffix << "[num_comp" << var_suffix << "*dim" << var_suffix << "*" 711 << (is_all_tensor && (max_dim >= 3) ? Q_name : "1") << "];\n"; 712 } 713 break; 714 case CEED_EVAL_WEIGHT: 715 break; 716 // LCOV_EXCL_START 717 case CEED_EVAL_DIV: 718 case CEED_EVAL_CURL: 719 break; // TODO: Not implemented 720 // LCOV_EXCL_STOP 721 } 722 } 723 724 if (is_at_points) { 725 // We need to handle batches of points 726 code << "\n // Note: Using batches of points\n"; 727 code << " const CeedInt point_loop_bound = (blockDim.x*blockDim.y) * ceil((1.0*max_num_points) / (blockDim.x*blockDim.y));\n\n"; 728 code << " #pragma unroll\n"; 729 code << " for (CeedInt i = threadIdx.x + threadIdx.y*blockDim.x; i < point_loop_bound; i += blockDim.x*blockDim.y) {\n"; 730 code << " const CeedInt p = i % max_num_points;\n\n"; 731 732 code << " // -- Coordinates\n"; 733 code << " CeedScalar r_x[max_dim];\n"; 734 code << " ReadPoint<max_dim, coords_comp_stride, max_num_points>(data, elem, p, max_num_points, points.indices, points.coords, r_x);\n\n"; 735 736 code << " // -- Input fields\n"; 737 for (CeedInt i = 0; i < num_input_fields; i++) { 738 const char *field_name; 739 std::string var_suffix = "_in_" + std::to_string(i); 740 std::string P_name = "P_1d" + var_suffix; 741 742 CeedCallBackend(CeedOperatorFieldGetName(op_input_fields[i], &field_name)); 743 code << " // ---- Input field " << i << ": " << field_name << "\n"; 744 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_input_fields[i], &eval_mode)); 745 // Basis action 746 code << " // EvalMode: " << CeedEvalModes[eval_mode] << "\n"; 747 switch (eval_mode) { 748 case CEED_EVAL_NONE: 749 code << " CeedScalar r_s" << var_suffix << "[num_comp" << var_suffix << "];\n"; 750 code << " ReadPoint<num_comp" << var_suffix << ", comp_stride" << var_suffix 751 << ", max_num_points>(data, elem, p, max_num_points, indices.inputs[" << i << "], d" << var_suffix << ", r_s" << var_suffix << ");\n"; 752 break; 753 case CEED_EVAL_INTERP: 754 code << " CeedScalar r_s" << var_suffix << "[num_comp" << var_suffix << "];\n"; 755 code << " InterpAtPoints" << max_dim << "d<num_comp" << var_suffix << ", max_num_points, " << P_name << ", " << Q_name 756 << ">(data, i, r_c" << var_suffix << ", r_x, r_s" << var_suffix << ");\n"; 757 break; 758 case CEED_EVAL_GRAD: 759 code << " CeedScalar r_s" << var_suffix << "[num_comp" << var_suffix << "*dim" << var_suffix << "];\n"; 760 code << " GradAtPoints" << max_dim << "d<num_comp" << var_suffix << ", max_num_points, " << P_name << ", " << Q_name 761 << ">(data, i, r_c" << var_suffix << ", r_x, r_s" << var_suffix << ");\n"; 762 break; 763 case CEED_EVAL_WEIGHT: 764 code << " CeedScalar r_s" << var_suffix << "[1];\n"; 765 code << " r_s" << var_suffix << "[0] = 1.0;\n"; 766 break; 767 // LCOV_EXCL_START 768 case CEED_EVAL_DIV: 769 case CEED_EVAL_CURL: 770 break; // TODO: Not implemented 771 // LCOV_EXCL_STOP 772 } 773 } 774 code << "\n // -- Output fields\n"; 775 for (CeedInt i = 0; i < num_output_fields; i++) { 776 const char *field_name; 777 std::string var_suffix = "_out_" + std::to_string(i); 778 779 CeedCallBackend(CeedOperatorFieldGetName(op_output_fields[i], &field_name)); 780 code << " // ---- Output field " << i << ": " << field_name << "\n"; 781 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_output_fields[i], &eval_mode)); 782 // Basis action 783 switch (eval_mode) { 784 case CEED_EVAL_NONE: 785 code << " CeedScalar r_s" << var_suffix << "[num_comp" << var_suffix << "];\n"; 786 break; 787 case CEED_EVAL_INTERP: 788 code << " CeedScalar r_s" << var_suffix << "[num_comp" << var_suffix << "];\n"; 789 break; 790 case CEED_EVAL_GRAD: 791 code << " CeedScalar r_s" << var_suffix << "[num_comp" << var_suffix << "*dim" << var_suffix << "];\n"; 792 break; 793 // LCOV_EXCL_START 794 case CEED_EVAL_WEIGHT: 795 break; // Should not occur 796 case CEED_EVAL_DIV: 797 case CEED_EVAL_CURL: 798 break; // TODO: Not implemented 799 // LCOV_EXCL_STOP 800 } 801 } 802 803 } else if (use_3d_slices) { 804 // We treat quadrature points per slice in 3d to save registers 805 code << "\n // Note: Using planes of 3D elements\n"; 806 code << " #pragma unroll\n"; 807 code << " for (CeedInt q = 0; q < " << Q_name << "; q++) {\n"; 808 code << " // -- Input fields\n"; 809 for (CeedInt i = 0; i < num_input_fields; i++) { 810 const char *field_name; 811 std::string var_suffix = "_in_" + std::to_string(i); 812 813 CeedCallBackend(CeedOperatorFieldGetName(op_input_fields[i], &field_name)); 814 code << " // ---- Input field " << i << ": " << field_name << "\n"; 815 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_input_fields[i], &eval_mode)); 816 // Basis action 817 code << " // EvalMode: " << CeedEvalModes[eval_mode] << "\n"; 818 switch (eval_mode) { 819 case CEED_EVAL_NONE: 820 bool is_strided; 821 822 code << " CeedScalar r_s" << var_suffix << "[num_comp" << var_suffix << "];\n"; 823 824 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_input_fields[i], &elem_rstr)); 825 CeedCallBackend(CeedElemRestrictionIsStrided(elem_rstr, &is_strided)); 826 if (is_strided) { 827 bool has_backend_strides; 828 CeedInt num_elem, elem_size; 829 830 CeedCallBackend(CeedElemRestrictionGetElementSize(elem_rstr, &elem_size)); 831 CeedCallBackend(CeedElemRestrictionHasBackendStrides(elem_rstr, &has_backend_strides)); 832 CeedCallBackend(CeedElemRestrictionGetNumElements(elem_rstr, &num_elem)); 833 CeedInt strides[3] = {1, elem_size * num_elem, elem_size}; 834 835 if (!has_backend_strides) { 836 CeedCallBackend(CeedElemRestrictionGetStrides(elem_rstr, strides)); 837 } 838 code << " // Strides: {" << strides[0] << ", " << strides[1] << ", " << strides[2] << "}\n"; 839 code << " ReadEVecSliceStrided3d<num_comp" << var_suffix << ", " << Q_name << ", " << strides[0] << ", " << strides[1] << ", " 840 << strides[2] << ">(data, elem, q, d" << var_suffix << ", r_s" << var_suffix << ");\n"; 841 } else { 842 CeedSize l_size = 0; 843 CeedInt comp_stride; 844 CeedElemRestriction_Cuda *rstr_data; 845 846 CeedCallBackend(CeedElemRestrictionGetLVectorSize(elem_rstr, &l_size)); 847 code << " const CeedInt l_size" << var_suffix << " = " << l_size << ";\n"; 848 CeedCallBackend(CeedElemRestrictionGetCompStride(elem_rstr, &comp_stride)); 849 code << " // CompStride: " << comp_stride << "\n"; 850 CeedCallBackend(CeedElemRestrictionGetData(elem_rstr, &rstr_data)); 851 data->indices.inputs[i] = (CeedInt *)rstr_data->d_offsets; 852 code << " ReadEVecSliceStandard3d<num_comp" << var_suffix << ", " << comp_stride << ", " << Q_name << ">(data, l_size" << var_suffix 853 << ", elem, q, indices.inputs[" << i << "], d" << var_suffix << ", r_s" << var_suffix << ");\n"; 854 } 855 CeedCallBackend(CeedElemRestrictionDestroy(&elem_rstr)); 856 break; 857 case CEED_EVAL_INTERP: 858 code << " CeedScalar r_s" << var_suffix << "[num_comp" << var_suffix << "];\n"; 859 code << " for (CeedInt j = 0; j < num_comp" << var_suffix << "; j++) {\n"; 860 code << " r_s" << var_suffix << "[j] = r_q" << var_suffix << "[q + j*" << Q_name << "];\n"; 861 code << " }\n"; 862 break; 863 case CEED_EVAL_GRAD: 864 code << " CeedScalar r_s" << var_suffix << "[num_comp" << var_suffix << "*dim" << var_suffix << "];\n"; 865 code << " GradColloSlice3d<num_comp" << var_suffix << ", " << Q_name << ", OP_T_1D>(data, q, r_q" << var_suffix << ", s_G" 866 << var_suffix << ", r_s" << var_suffix << ");\n"; 867 break; 868 case CEED_EVAL_WEIGHT: 869 code << " CeedScalar r_s" << var_suffix << "[1];\n"; 870 code << " r_s" << var_suffix << "[0] = r_q" << var_suffix << "[q];\n"; 871 break; 872 // LCOV_EXCL_START 873 case CEED_EVAL_DIV: 874 case CEED_EVAL_CURL: 875 break; // TODO: Not implemented 876 // LCOV_EXCL_STOP 877 } 878 } 879 code << "\n // -- Output fields\n"; 880 for (CeedInt i = 0; i < num_output_fields; i++) { 881 const char *field_name; 882 std::string var_suffix = "_out_" + std::to_string(i); 883 884 CeedCallBackend(CeedOperatorFieldGetName(op_output_fields[i], &field_name)); 885 code << " // ---- Output field " << i << ": " << field_name << "\n"; 886 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_output_fields[i], &eval_mode)); 887 // Basis action 888 switch (eval_mode) { 889 case CEED_EVAL_NONE: 890 code << " CeedScalar r_s" << var_suffix << "[num_comp" << var_suffix << "];\n"; 891 break; 892 case CEED_EVAL_INTERP: 893 code << " CeedScalar r_s" << var_suffix << "[num_comp" << var_suffix << "];\n"; 894 break; 895 case CEED_EVAL_GRAD: 896 code << " CeedScalar r_s" << var_suffix << "[num_comp" << var_suffix << "*dim" << var_suffix << "];\n"; 897 break; 898 // LCOV_EXCL_START 899 case CEED_EVAL_WEIGHT: 900 break; // Should not occur 901 case CEED_EVAL_DIV: 902 case CEED_EVAL_CURL: 903 break; // TODO: Not implemented 904 // LCOV_EXCL_STOP 905 } 906 } 907 } else { 908 code << "\n // Note: Using full elements\n"; 909 code << " {\n"; 910 code << " // -- Input fields\n"; 911 for (CeedInt i = 0; i < num_input_fields; i++) { 912 const char *field_name; 913 914 CeedCallBackend(CeedOperatorFieldGetName(op_input_fields[i], &field_name)); 915 code << " // ---- Input field " << i << ": " << field_name << "\n"; 916 code << " CeedScalar *r_s_in_" << i << " = r_q_in_" << i << ";\n"; 917 } 918 code << " // -- Output fields\n"; 919 for (CeedInt i = 0; i < num_output_fields; i++) { 920 const char *field_name; 921 922 CeedCallBackend(CeedOperatorFieldGetName(op_output_fields[i], &field_name)); 923 code << " // ---- Output field " << i << ": " << field_name << "\n"; 924 code << " CeedScalar *r_s_out_" << i << " = r_q_out_" << i << ";\n"; 925 } 926 } 927 928 // Input and output buffers 929 code << "\n // -- QFunction inputs and outputs\n"; 930 code << " // ---- Inputs\n"; 931 code << " CeedScalar *inputs[" << CeedIntMax(num_input_fields, 1) << "];\n"; 932 for (CeedInt i = 0; i < num_input_fields; i++) { 933 const char *field_name; 934 935 CeedCallBackend(CeedOperatorFieldGetName(op_input_fields[i], &field_name)); 936 code << " // ------ Input field " << i << ": " << field_name << "\n"; 937 code << " inputs[" << i << "] = r_s_in_" << i << ";\n"; 938 } 939 code << " // ---- Outputs\n"; 940 code << " CeedScalar *outputs[" << CeedIntMax(num_output_fields, 1) << "];\n"; 941 for (CeedInt i = 0; i < num_output_fields; i++) { 942 const char *field_name; 943 944 CeedCallBackend(CeedOperatorFieldGetName(op_output_fields[i], &field_name)); 945 code << " // ------ Output field " << i << ": " << field_name << "\n"; 946 code << " outputs[" << i << "] = r_s_out_" << i << ";\n"; 947 } 948 949 // Apply QFunction 950 code << "\n // -- Apply QFunction\n"; 951 code << " " << qfunction_name << "(ctx, "; 952 if (max_dim != 3 || is_at_points || use_3d_slices || !is_all_tensor) { 953 code << "1"; 954 } else { 955 code << Q_name; 956 } 957 code << ", inputs, outputs);\n"; 958 959 if (is_at_points) { 960 // Map back to coefficients 961 code << "\n // -- Output fields\n"; 962 for (CeedInt i = 0; i < num_output_fields; i++) { 963 const char *field_name; 964 std::string var_suffix = "_out_" + std::to_string(i); 965 std::string P_name = "P_1d" + var_suffix; 966 967 CeedCallBackend(CeedOperatorFieldGetName(op_output_fields[i], &field_name)); 968 code << " // ---- Output field " << i << ": " << field_name << "\n"; 969 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_output_fields[i], &eval_mode)); 970 // Basis action 971 code << " // EvalMode: " << CeedEvalModes[eval_mode] << "\n"; 972 switch (eval_mode) { 973 case CEED_EVAL_NONE: { 974 CeedInt comp_stride; 975 CeedElemRestriction elem_rstr; 976 977 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_output_fields[i], &elem_rstr)); 978 CeedCallBackend(CeedElemRestrictionGetCompStride(elem_rstr, &comp_stride)); 979 CeedCallBackend(CeedElemRestrictionDestroy(&elem_rstr)); 980 code << " const CeedInt comp_stride" << var_suffix << " = " << comp_stride << ";\n"; 981 code << " WritePoint<num_comp" << var_suffix << ", comp_stride" << var_suffix 982 << ", max_num_points>(data, elem, i, points.num_per_elem[elem], indices.outputs[" << i << "]" 983 << ", r_s" << var_suffix << ", d" << var_suffix << ");\n"; 984 break; 985 } 986 case CEED_EVAL_INTERP: 987 code << " if (i >= points.num_per_elem[elem]) {\n"; 988 code << " for (CeedInt j = 0; j < num_comp" << var_suffix << "; j++) r_s" << var_suffix << "[j] = 0.0;\n"; 989 code << " }\n"; 990 code << " InterpTransposeAtPoints" << max_dim << "d<num_comp" << var_suffix << ", max_num_points, " << P_name << ", " << Q_name 991 << ">(data, i, r_s" << var_suffix << ", r_x, r_c" << var_suffix << ");\n"; 992 break; 993 case CEED_EVAL_GRAD: 994 code << " if (i >= points.num_per_elem[elem]) {\n"; 995 code << " for (CeedInt j = 0; j < num_comp" << var_suffix << "*dim" << var_suffix << "; j++) r_s" << var_suffix << "[j] = 0.0;\n"; 996 code << " }\n"; 997 code << " GradTransposeAtPoints" << max_dim << "d<num_comp" << var_suffix << ", max_num_points, " << P_name << ", " << Q_name 998 << ">(data, i, r_s" << var_suffix << ", r_x, r_c" << var_suffix << ");\n"; 999 break; 1000 // LCOV_EXCL_START 1001 case CEED_EVAL_WEIGHT: 1002 break; // Should not occur 1003 case CEED_EVAL_DIV: 1004 case CEED_EVAL_CURL: 1005 break; // TODO: Not implemented 1006 // LCOV_EXCL_STOP 1007 } 1008 } 1009 } else if (use_3d_slices) { 1010 // Copy or apply transpose grad, if needed 1011 code << "\n // -- Output fields\n"; 1012 for (CeedInt i = 0; i < num_output_fields; i++) { 1013 const char *field_name; 1014 std::string var_suffix = "_out_" + std::to_string(i); 1015 std::string P_name = "P_1d" + var_suffix; 1016 1017 CeedCallBackend(CeedOperatorFieldGetName(op_output_fields[i], &field_name)); 1018 code << " // ---- Output field " << i << ": " << field_name << "\n"; 1019 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_output_fields[i], &eval_mode)); 1020 // Basis action 1021 code << " // EvalMode: " << CeedEvalModes[eval_mode] << "\n"; 1022 switch (eval_mode) { 1023 case CEED_EVAL_NONE: 1024 code << " for (CeedInt j = 0; j < num_comp" << var_suffix << " ; j++) {\n"; 1025 code << " r_q" << var_suffix << "[q + j*" << Q_name << "] = r_s" << var_suffix << "[j];\n"; 1026 code << " }\n"; 1027 break; 1028 case CEED_EVAL_INTERP: 1029 code << " for (CeedInt j = 0; j < num_comp" << var_suffix << " ; j++) {\n"; 1030 code << " r_q" << var_suffix << "[q + j*" << Q_name << "] = r_s" << var_suffix << "[j];\n"; 1031 code << " }\n"; 1032 break; 1033 case CEED_EVAL_GRAD: 1034 code << " GradColloSliceTranspose3d<num_comp" << var_suffix << ", " << Q_name << ", OP_T_1D>(data, q, r_s" << var_suffix << ", s_G" 1035 << var_suffix << ", r_q" << var_suffix << ");\n"; 1036 break; 1037 // LCOV_EXCL_START 1038 case CEED_EVAL_WEIGHT: 1039 break; // Should not occur 1040 case CEED_EVAL_DIV: 1041 case CEED_EVAL_CURL: 1042 break; // TODO: Not implemented 1043 // LCOV_EXCL_STOP 1044 } 1045 } 1046 } 1047 code << " }\n"; 1048 return CEED_ERROR_SUCCESS; 1049 } 1050 1051 //------------------------------------------------------------------------------ 1052 // Build single operator kernel 1053 //------------------------------------------------------------------------------ 1054 extern "C" int CeedOperatorBuildKernel_Cuda_gen(CeedOperator op, bool *is_good_build) { 1055 bool is_all_tensor = true, is_all_nontensor = true, is_at_points = false, use_3d_slices = false; 1056 Ceed ceed; 1057 CeedInt Q = 0, Q_1d = 0, num_input_fields, num_output_fields, max_dim = 1, max_num_points = 0, coords_comp_stride = 0; 1058 CeedQFunctionField *qf_input_fields, *qf_output_fields; 1059 CeedQFunction_Cuda_gen *qf_data; 1060 CeedQFunction qf; 1061 CeedOperatorField *op_input_fields, *op_output_fields; 1062 CeedOperator_Cuda_gen *data; 1063 std::ostringstream code; 1064 1065 CeedCallBackend(CeedOperatorGetData(op, &data)); 1066 { 1067 bool is_setup_done; 1068 1069 CeedCallBackend(CeedOperatorIsSetupDone(op, &is_setup_done)); 1070 if (is_setup_done) { 1071 *is_good_build = !data->use_fallback; 1072 return CEED_ERROR_SUCCESS; 1073 } 1074 } 1075 1076 // Check field compatibility 1077 CeedCallBackend(CeedOperatorGetFields(op, &num_input_fields, &op_input_fields, &num_output_fields, &op_output_fields)); 1078 { 1079 bool has_shared_bases = true; 1080 1081 for (CeedInt i = 0; i < num_input_fields; i++) { 1082 CeedBasis basis; 1083 1084 CeedCallBackend(CeedOperatorFieldGetBasis(op_input_fields[i], &basis)); 1085 if (basis != CEED_BASIS_NONE) { 1086 bool is_tensor = true; 1087 const char *resource; 1088 char *resource_root; 1089 Ceed basis_ceed; 1090 1091 CeedCallBackend(CeedBasisIsTensor(basis, &is_tensor)); 1092 is_all_tensor = is_all_tensor && is_tensor; 1093 is_all_nontensor = is_all_nontensor && !is_tensor; 1094 CeedCallBackend(CeedBasisGetCeed(basis, &basis_ceed)); 1095 CeedCallBackend(CeedGetResource(basis_ceed, &resource)); 1096 CeedCallBackend(CeedGetResourceRoot(basis_ceed, resource, ":", &resource_root)); 1097 has_shared_bases = has_shared_bases && !strcmp(resource_root, "/gpu/cuda/shared"); 1098 CeedCallBackend(CeedFree(&resource_root)); 1099 CeedCallBackend(CeedDestroy(&basis_ceed)); 1100 } 1101 CeedCallBackend(CeedBasisDestroy(&basis)); 1102 } 1103 1104 for (CeedInt i = 0; i < num_output_fields; i++) { 1105 CeedBasis basis; 1106 1107 CeedCallBackend(CeedOperatorFieldGetBasis(op_output_fields[i], &basis)); 1108 if (basis != CEED_BASIS_NONE) { 1109 bool is_tensor = true; 1110 const char *resource; 1111 char *resource_root; 1112 Ceed basis_ceed; 1113 1114 CeedCallBackend(CeedBasisIsTensor(basis, &is_tensor)); 1115 is_all_tensor = is_all_tensor && is_tensor; 1116 is_all_nontensor = is_all_nontensor && !is_tensor; 1117 1118 CeedCallBackend(CeedBasisGetCeed(basis, &basis_ceed)); 1119 CeedCallBackend(CeedGetResource(basis_ceed, &resource)); 1120 CeedCallBackend(CeedGetResourceRoot(basis_ceed, resource, ":", &resource_root)); 1121 has_shared_bases = has_shared_bases && !strcmp(resource_root, "/gpu/cuda/shared"); 1122 CeedCallBackend(CeedFree(&resource_root)); 1123 CeedCallBackend(CeedDestroy(&basis_ceed)); 1124 } 1125 CeedCallBackend(CeedBasisDestroy(&basis)); 1126 } 1127 // -- Fallback to ref if not all bases are shared 1128 if (!has_shared_bases) { 1129 *is_good_build = false; 1130 return CEED_ERROR_SUCCESS; 1131 } 1132 } 1133 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 1134 CeedCallBackend(CeedOperatorGetQFunction(op, &qf)); 1135 CeedCallBackend(CeedQFunctionGetData(qf, &qf_data)); 1136 CeedCallBackend(CeedQFunctionGetFields(qf, NULL, &qf_input_fields, NULL, &qf_output_fields)); 1137 1138 // Get operator data 1139 CeedCallBackend(CeedOperatorIsAtPoints(op, &is_at_points)); 1140 { 1141 CeedInt max_P = 0, max_P_1d = 0; 1142 1143 CeedCallBackend(CeedOperatorBuildKernelData_Cuda_gen(ceed, num_input_fields, op_input_fields, qf_input_fields, num_output_fields, 1144 op_output_fields, qf_output_fields, &max_P, &max_P_1d, &Q, &Q_1d, &max_dim, &is_all_tensor, 1145 &use_3d_slices)); 1146 data->max_P_1d = is_all_tensor ? max_P_1d : max_P; 1147 } 1148 if (max_dim == 0) max_dim = 1; 1149 data->dim = max_dim; 1150 if (is_at_points) { 1151 CeedElemRestriction_Cuda *rstr_data; 1152 CeedElemRestriction rstr_points = NULL; 1153 1154 CeedCallBackend(CeedOperatorAtPointsGetPoints(op, &rstr_points, NULL)); 1155 CeedCallBackend(CeedElemRestrictionGetMaxPointsInElement(rstr_points, &max_num_points)); 1156 CeedCallBackend(CeedElemRestrictionGetCompStride(rstr_points, &coords_comp_stride)); 1157 CeedCallBackend(CeedElemRestrictionGetData(rstr_points, &rstr_data)); 1158 data->points.indices = (CeedInt *)rstr_data->d_offsets; 1159 CeedCallBackend(CeedElemRestrictionDestroy(&rstr_points)); 1160 } 1161 if (is_at_points) use_3d_slices = false; 1162 if (Q_1d == 0) { 1163 if (is_at_points) Q_1d = max_num_points; 1164 else CeedCallBackend(CeedOperatorGetNumQuadraturePoints(op, &Q_1d)); 1165 } 1166 if (Q == 0) Q = Q_1d; 1167 data->Q = Q; 1168 data->Q_1d = Q_1d; 1169 1170 // Check for restriction only identity operator 1171 { 1172 bool is_identity_qf; 1173 1174 CeedCallBackend(CeedQFunctionIsIdentity(qf, &is_identity_qf)); 1175 if (is_identity_qf) { 1176 CeedEvalMode eval_mode_in, eval_mode_out; 1177 1178 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_input_fields[0], &eval_mode_in)); 1179 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_output_fields[0], &eval_mode_out)); 1180 CeedCheck(eval_mode_in != CEED_EVAL_NONE || eval_mode_out != CEED_EVAL_NONE, ceed, CEED_ERROR_BACKEND, 1181 "Backend does not implement restriction only identity operators"); 1182 } 1183 } 1184 1185 // Add atomicAdd function for old NVidia architectures 1186 { 1187 Ceed_Cuda *ceed_data; 1188 struct cudaDeviceProp prop; 1189 1190 CeedCallBackend(CeedGetData(ceed, &ceed_data)); 1191 CeedCallBackend(cudaGetDeviceProperties(&prop, ceed_data->device_id)); 1192 if ((prop.major < 6) && (CEED_SCALAR_TYPE != CEED_SCALAR_FP32)) { 1193 code << "// AtomicAdd fallback source\n"; 1194 code << "#include <ceed/jit-source/cuda/cuda-atomic-add-fallback.h>\n\n"; 1195 } 1196 } 1197 1198 // Load basis source files 1199 if (!is_all_nontensor) { 1200 code << "// Tensor basis source\n"; 1201 code << "#include <ceed/jit-source/cuda/cuda-shared-basis-tensor-templates.h>\n\n"; 1202 } 1203 if (!is_all_tensor) { 1204 code << "// Non-tensor basis source\n"; 1205 code << "#include <ceed/jit-source/cuda/cuda-shared-basis-nontensor-templates.h>\n\n"; 1206 } 1207 if (!is_all_tensor && !is_all_nontensor) { 1208 code << "// Tensor basis source\n"; 1209 code << "#include <ceed/jit-source/cuda/cuda-shared-basis-tensor-flattened-templates.h>\n\n"; 1210 } 1211 if (is_at_points) { 1212 code << "// AtPoints basis source\n"; 1213 code << "#include <ceed/jit-source/cuda/cuda-shared-basis-tensor-at-points-templates.h>\n\n"; 1214 } 1215 code << "// CodeGen operator source\n"; 1216 code << "#include <ceed/jit-source/cuda/cuda-gen-templates.h>\n\n"; 1217 1218 // Get QFunction name 1219 std::string qfunction_name(qf_data->qfunction_name); 1220 std::string operator_name; 1221 1222 operator_name = "CeedKernelCudaGenOperator_" + qfunction_name; 1223 1224 // Define CEED_Q_VLA 1225 code << "\n#undef CEED_Q_VLA\n"; 1226 if (max_dim != 3 || is_at_points || use_3d_slices || !is_all_tensor) { 1227 code << "#define CEED_Q_VLA 1\n\n"; 1228 } else { 1229 code << "#define CEED_Q_VLA " << Q_1d << "\n\n"; 1230 } 1231 1232 // Add user QFunction source 1233 { 1234 const char *source_path; 1235 1236 CeedCallBackend(CeedQFunctionGetSourcePath(qf, &source_path)); 1237 CeedCheck(source_path, ceed, CEED_ERROR_UNSUPPORTED, "/gpu/cuda/gen backend requires QFunction source code file"); 1238 1239 code << "// User QFunction source\n"; 1240 code << "#include \"" << source_path << "\"\n\n"; 1241 } 1242 1243 // Setup 1244 code << "\n// -----------------------------------------------------------------------------\n"; 1245 code << "// Operator Kernel\n"; 1246 code << "// \n"; 1247 code << "// d_[in,out]_i: CeedVector device array\n"; 1248 code << "// r_[in,out]_e_i: Element vector register\n"; 1249 code << "// r_[in,out]_q_i: Quadrature space vector register\n"; 1250 code << "// r_[in,out]_c_i: AtPoints Chebyshev coefficients register\n"; 1251 code << "// r_[in,out]_s_i: Quadrature space slice vector register\n"; 1252 code << "// \n"; 1253 code << "// s_B_[in,out]_i: Interpolation matrix, shared memory\n"; 1254 code << "// s_G_[in,out]_i: Gradient matrix, shared memory\n"; 1255 code << "// -----------------------------------------------------------------------------\n"; 1256 code << "extern \"C\" __global__ void " << operator_name 1257 << "(CeedInt num_elem, void* ctx, FieldsInt_Cuda indices, Fields_Cuda fields, Fields_Cuda B, Fields_Cuda G, CeedScalar *W, Points_Cuda " 1258 "points) {\n"; 1259 1260 // Scratch buffers 1261 for (CeedInt i = 0; i < num_input_fields; i++) { 1262 CeedEvalMode eval_mode; 1263 1264 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_input_fields[i], &eval_mode)); 1265 if (eval_mode != CEED_EVAL_WEIGHT) { // Skip CEED_EVAL_WEIGHT 1266 code << " const CeedScalar *__restrict__ d_in_" << i << " = fields.inputs[" << i << "];\n"; 1267 } 1268 } 1269 for (CeedInt i = 0; i < num_output_fields; i++) { 1270 code << " CeedScalar *__restrict__ d_out_" << i << " = fields.outputs[" << i << "];\n"; 1271 } 1272 1273 code << " const CeedInt max_dim = " << max_dim << ";\n"; 1274 if (!is_all_tensor) { 1275 code << " const CeedInt Q = " << Q << ";\n"; 1276 } 1277 if (!is_all_nontensor) { 1278 code << " const CeedInt Q_1d = " << Q_1d << ";\n"; 1279 } 1280 if (is_at_points) { 1281 code << " const CeedInt max_num_points = " << max_num_points << ";\n"; 1282 code << " const CeedInt coords_comp_stride = " << coords_comp_stride << ";\n"; 1283 } 1284 1285 // Shared data 1286 code << " extern __shared__ CeedScalar slice[];\n"; 1287 code << " SharedData_Cuda data;\n"; 1288 code << " data.t_id_x = threadIdx.x;\n"; 1289 code << " data.t_id_y = threadIdx.y;\n"; 1290 code << " data.t_id_z = threadIdx.z;\n"; 1291 code << " data.t_id = threadIdx.x + threadIdx.y*blockDim.x + threadIdx.z*blockDim.y*blockDim.x;\n"; 1292 code << " data.slice = slice + data.t_id_z*OP_T_1D" << ((!is_all_tensor || max_dim == 1) ? "" : "*OP_T_1D") << ";\n"; 1293 1294 // -- Determine input mat reuse 1295 FieldReuse_Cuda input_matrix_reuse[CEED_FIELD_MAX]; 1296 1297 for (CeedInt i = 0; i < num_input_fields; i++) { 1298 input_matrix_reuse[i].index = -1; 1299 } 1300 for (CeedInt i = 0; i < num_input_fields; i++) { 1301 bool is_tensor = true; 1302 CeedEvalMode eval_mode_i; 1303 CeedBasis basis_i; 1304 1305 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_input_fields[i], &eval_mode_i)); 1306 if (eval_mode_i == CEED_EVAL_WEIGHT) continue; 1307 CeedCallBackend(CeedOperatorFieldGetBasis(op_input_fields[i], &basis_i)); 1308 CeedCallBackend(CeedBasisIsTensor(basis_i, &is_tensor)); 1309 for (CeedInt j = 0; (input_matrix_reuse[i].index == -1) && (j < i); j++) { 1310 CeedEvalMode eval_mode_j; 1311 CeedBasis basis_j; 1312 1313 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_input_fields[j], &eval_mode_j)); 1314 if (eval_mode_j == CEED_EVAL_WEIGHT) continue; 1315 CeedCallBackend(CeedOperatorFieldGetBasis(op_input_fields[j], &basis_j)); 1316 if (basis_i == basis_j) { 1317 if (is_tensor) { 1318 input_matrix_reuse[i].index = j; 1319 input_matrix_reuse[i].is_input = true; 1320 input_matrix_reuse[i].eval_mode = eval_mode_j; 1321 } else { 1322 // For non-tensor can only re-use with the same eval mode 1323 if (eval_mode_i == eval_mode_j) { 1324 input_matrix_reuse[i].index = j; 1325 input_matrix_reuse[i].is_input = true; 1326 input_matrix_reuse[i].eval_mode = eval_mode_j; 1327 } 1328 } 1329 } 1330 CeedCallBackend(CeedBasisDestroy(&basis_j)); 1331 } 1332 CeedCallBackend(CeedBasisDestroy(&basis_i)); 1333 } 1334 1335 // -- Determine output mat reuse 1336 FieldReuse_Cuda output_matrix_reuse[CEED_FIELD_MAX]; 1337 1338 for (CeedInt i = 0; i < num_output_fields; i++) { 1339 output_matrix_reuse[i].index = -1; 1340 } 1341 for (CeedInt i = 0; i < num_output_fields; i++) { 1342 bool is_tensor = true; 1343 CeedEvalMode eval_mode_i; 1344 CeedBasis basis_i; 1345 1346 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_output_fields[i], &eval_mode_i)); 1347 CeedCallBackend(CeedOperatorFieldGetBasis(op_output_fields[i], &basis_i)); 1348 CeedCallBackend(CeedBasisIsTensor(basis_i, &is_tensor)); 1349 for (CeedInt j = 0; (output_matrix_reuse[i].index == -1) && (j < num_input_fields); j++) { 1350 CeedEvalMode eval_mode_j; 1351 CeedBasis basis_j; 1352 1353 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_input_fields[j], &eval_mode_j)); 1354 if (eval_mode_j == CEED_EVAL_WEIGHT) continue; 1355 CeedCallBackend(CeedOperatorFieldGetBasis(op_input_fields[j], &basis_j)); 1356 if (basis_i == basis_j) { 1357 if (is_tensor) { 1358 output_matrix_reuse[i].index = j; 1359 output_matrix_reuse[i].is_input = true; 1360 output_matrix_reuse[i].eval_mode = eval_mode_j; 1361 } else { 1362 // For non-tensor can only re-use with the same eval mode 1363 if (eval_mode_i == eval_mode_j) { 1364 output_matrix_reuse[i].index = j; 1365 output_matrix_reuse[i].is_input = true; 1366 output_matrix_reuse[i].eval_mode = eval_mode_j; 1367 } 1368 } 1369 } 1370 CeedCallBackend(CeedBasisDestroy(&basis_j)); 1371 } 1372 for (CeedInt j = 0; (output_matrix_reuse[i].index == -1) && (j < i); j++) { 1373 CeedEvalMode eval_mode_j; 1374 CeedBasis basis_j; 1375 1376 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_output_fields[j], &eval_mode_j)); 1377 if (eval_mode_j == CEED_EVAL_WEIGHT) continue; 1378 CeedCallBackend(CeedOperatorFieldGetBasis(op_output_fields[j], &basis_j)); 1379 if (basis_i == basis_j) { 1380 if (is_tensor) { 1381 output_matrix_reuse[i].index = j; 1382 output_matrix_reuse[i].is_input = false; 1383 output_matrix_reuse[i].eval_mode = eval_mode_j; 1384 } else { 1385 // For non-tensor can only re-use with the same eval mode 1386 if (eval_mode_i == eval_mode_j) { 1387 output_matrix_reuse[i].index = j; 1388 output_matrix_reuse[i].is_input = false; 1389 output_matrix_reuse[i].eval_mode = eval_mode_j; 1390 } 1391 } 1392 } 1393 CeedCallBackend(CeedBasisDestroy(&basis_j)); 1394 } 1395 CeedCallBackend(CeedBasisDestroy(&basis_i)); 1396 } 1397 1398 // Initialize constants, and matrices B and G 1399 code << "\n // Input field constants and basis data\n"; 1400 for (CeedInt i = 0; i < num_input_fields; i++) { 1401 CeedCallBackend(CeedOperatorBuildKernelFieldData_Cuda_gen(code, data, i, op_input_fields[i], qf_input_fields[i], input_matrix_reuse[i], max_dim, 1402 Q, Q_1d, true, is_all_tensor, is_at_points, use_3d_slices)); 1403 } 1404 code << "\n // Output field constants and basis data\n"; 1405 for (CeedInt i = 0; i < num_output_fields; i++) { 1406 CeedCallBackend(CeedOperatorBuildKernelFieldData_Cuda_gen(code, data, i, op_output_fields[i], qf_output_fields[i], output_matrix_reuse[i], 1407 max_dim, Q, Q_1d, false, is_all_tensor, is_at_points, use_3d_slices)); 1408 } 1409 1410 // Loop over all elements 1411 code << "\n // Element loop\n"; 1412 code << " __syncthreads();\n"; 1413 code << " for (CeedInt elem = blockIdx.x*blockDim.z + threadIdx.z; elem < num_elem; elem += gridDim.x*blockDim.z) {\n"; 1414 1415 // -- Compute minimum buffer space needed 1416 CeedInt max_rstr_buffer_size = 1; 1417 1418 for (CeedInt i = 0; i < num_input_fields; i++) { 1419 CeedInt num_comp; 1420 CeedElemRestriction elem_rstr; 1421 1422 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_input_fields[i], &elem_rstr)); 1423 CeedCallBackend(CeedElemRestrictionGetNumComponents(elem_rstr, &num_comp)); 1424 max_rstr_buffer_size = CeedIntMax(max_rstr_buffer_size, num_comp * (is_all_tensor && (max_dim >= 3) ? Q_1d : 1)); 1425 CeedCallBackend(CeedElemRestrictionDestroy(&elem_rstr)); 1426 } 1427 for (CeedInt i = 0; i < num_output_fields; i++) { 1428 CeedInt num_comp; 1429 CeedElemRestriction elem_rstr; 1430 1431 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_output_fields[i], &elem_rstr)); 1432 CeedCallBackend(CeedElemRestrictionGetNumComponents(elem_rstr, &num_comp)); 1433 max_rstr_buffer_size = CeedIntMax(max_rstr_buffer_size, num_comp * (is_all_tensor && (max_dim >= 3) ? Q_1d : 1)); 1434 CeedCallBackend(CeedElemRestrictionDestroy(&elem_rstr)); 1435 } 1436 code << " // Scratch restriction buffer space\n"; 1437 code << " CeedScalar r_e_scratch[" << max_rstr_buffer_size << "];\n"; 1438 1439 // -- Determine best input field processing order 1440 CeedInt field_rstr_in_buffer[CEED_FIELD_MAX], input_field_order[CEED_FIELD_MAX]; 1441 1442 for (CeedInt i = 0; i < num_input_fields; i++) { 1443 field_rstr_in_buffer[i] = -1; 1444 input_field_order[i] = -1; 1445 } 1446 { 1447 bool is_ordered[CEED_FIELD_MAX]; 1448 CeedInt curr_index = 0; 1449 1450 for (CeedInt i = 0; i < num_input_fields; i++) is_ordered[i] = false; 1451 for (CeedInt i = 0; i < num_input_fields; i++) { 1452 CeedVector vec_i; 1453 CeedElemRestriction rstr_i; 1454 1455 if (is_ordered[i]) continue; 1456 field_rstr_in_buffer[i] = i; 1457 is_ordered[i] = true; 1458 input_field_order[curr_index] = i; 1459 curr_index++; 1460 CeedCallBackend(CeedOperatorFieldGetVector(op_input_fields[i], &vec_i)); 1461 if (vec_i == CEED_VECTOR_NONE) continue; // CEED_EVAL_WEIGHT 1462 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_input_fields[i], &rstr_i)); 1463 for (CeedInt j = i + 1; j < num_input_fields; j++) { 1464 CeedVector vec_j; 1465 CeedElemRestriction rstr_j; 1466 1467 CeedCallBackend(CeedOperatorFieldGetVector(op_input_fields[j], &vec_j)); 1468 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_input_fields[j], &rstr_j)); 1469 if (rstr_i == rstr_j && vec_i == vec_j) { 1470 field_rstr_in_buffer[j] = i; 1471 is_ordered[j] = true; 1472 input_field_order[curr_index] = j; 1473 curr_index++; 1474 } 1475 CeedCallBackend(CeedVectorDestroy(&vec_j)); 1476 CeedCallBackend(CeedElemRestrictionDestroy(&rstr_j)); 1477 } 1478 CeedCallBackend(CeedVectorDestroy(&vec_i)); 1479 CeedCallBackend(CeedElemRestrictionDestroy(&rstr_i)); 1480 } 1481 } 1482 1483 // -- Input restriction and basis 1484 code << "\n // -- Input field restrictions and basis actions\n"; 1485 for (CeedInt i = 0; i < num_input_fields; i++) { 1486 const char *field_name; 1487 const CeedInt f = input_field_order[i]; 1488 1489 CeedCallBackend(CeedOperatorFieldGetName(op_input_fields[f], &field_name)); 1490 code << " // ---- Input field " << f << ": " << field_name << "\n"; 1491 1492 // ---- Restriction 1493 CeedCallBackend(CeedOperatorBuildKernelRestriction_Cuda_gen(code, data, f, field_rstr_in_buffer, op_input_fields[f], qf_input_fields[f], max_dim, 1494 Q_1d, true, is_all_tensor, is_at_points, use_3d_slices)); 1495 1496 // ---- Basis action 1497 CeedCallBackend(CeedOperatorBuildKernelBasis_Cuda_gen(code, data, f, op_input_fields[f], qf_input_fields[f], max_dim, Q_1d, true, is_all_tensor, 1498 is_at_points, use_3d_slices)); 1499 } 1500 1501 // -- Q function 1502 CeedCallBackend(CeedOperatorBuildKernelQFunction_Cuda_gen(code, data, max_dim, max_num_points, num_input_fields, op_input_fields, qf_input_fields, 1503 num_output_fields, op_output_fields, qf_output_fields, qfunction_name, Q_1d, 1504 is_all_tensor, is_at_points, use_3d_slices)); 1505 1506 // -- Output basis and restriction 1507 code << "\n // -- Output field basis action and restrictions\n"; 1508 for (CeedInt i = 0; i < num_output_fields; i++) { 1509 const char *field_name; 1510 1511 CeedCallBackend(CeedOperatorFieldGetName(op_output_fields[i], &field_name)); 1512 code << " // ---- Output field " << i << ": " << field_name << "\n"; 1513 1514 // ---- Basis action 1515 CeedCallBackend(CeedOperatorBuildKernelBasis_Cuda_gen(code, data, i, op_output_fields[i], qf_output_fields[i], max_dim, Q_1d, false, 1516 is_all_tensor, is_at_points, use_3d_slices)); 1517 1518 // ---- Restriction 1519 CeedCallBackend(CeedOperatorBuildKernelRestriction_Cuda_gen(code, data, i, NULL, op_output_fields[i], qf_output_fields[i], max_dim, Q_1d, false, 1520 is_all_tensor, is_at_points, use_3d_slices)); 1521 } 1522 1523 // Close loop and function 1524 code << " }\n"; 1525 code << "}\n"; 1526 code << "// -----------------------------------------------------------------------------\n\n"; 1527 1528 // Compile 1529 { 1530 bool is_compile_good = false; 1531 const CeedInt T_1d = CeedIntMax(is_all_tensor ? Q_1d : Q, data->max_P_1d); 1532 1533 data->thread_1d = T_1d; 1534 CeedCallBackend(CeedTryCompile_Cuda(ceed, code.str().c_str(), &is_compile_good, &data->module, 1, "OP_T_1D", T_1d)); 1535 if (is_compile_good) { 1536 *is_good_build = true; 1537 CeedCallBackend(CeedGetKernel_Cuda(ceed, data->module, operator_name.c_str(), &data->op)); 1538 } else { 1539 *is_good_build = false; 1540 data->use_fallback = true; 1541 } 1542 } 1543 CeedCallBackend(CeedOperatorSetSetupDone(op)); 1544 CeedCallBackend(CeedDestroy(&ceed)); 1545 CeedCallBackend(CeedQFunctionDestroy(&qf)); 1546 return CEED_ERROR_SUCCESS; 1547 } 1548 1549 //------------------------------------------------------------------------------ 1550