1 // Copyright (c) 2017-2025, 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/gen-tools.h> 13 #include <ceed/jit-tools.h> 14 15 #include <iostream> 16 #include <sstream> 17 #include <string> 18 19 #include "../hip-ref/ceed-hip-ref.h" 20 #include "../hip-shared/ceed-hip-shared.h" 21 #include "../hip/ceed-hip-common.h" 22 #include "../hip/ceed-hip-compile.h" 23 #include "ceed-hip-gen.h" 24 25 struct FieldReuse_Hip { 26 CeedInt index; 27 bool is_input; 28 CeedEvalMode eval_mode; 29 }; 30 31 //------------------------------------------------------------------------------ 32 // Calculate the block size used for launching the operator kernel 33 //------------------------------------------------------------------------------ 34 extern "C" int BlockGridCalculate_Hip_gen(const CeedInt dim, const CeedInt num_elem, const CeedInt P_1d, const CeedInt Q_1d, CeedInt *block_sizes) { 35 const CeedInt thread_1d = CeedIntMax(Q_1d, P_1d); 36 if (dim == 1) { 37 CeedInt elems_per_block = 64 * thread_1d > 256 ? 256 / thread_1d : 64; 38 39 elems_per_block = elems_per_block > 0 ? elems_per_block : 1; 40 block_sizes[0] = thread_1d; 41 block_sizes[1] = 1; 42 block_sizes[2] = elems_per_block; 43 } else if (dim == 2) { 44 const CeedInt elems_per_block = thread_1d < 4 ? 16 : 2; 45 46 block_sizes[0] = thread_1d; 47 block_sizes[1] = thread_1d; 48 block_sizes[2] = elems_per_block; 49 } else if (dim == 3) { 50 const CeedInt elems_per_block = thread_1d < 6 ? 4 : (thread_1d < 8 ? 2 : 1); 51 52 block_sizes[0] = thread_1d; 53 block_sizes[1] = thread_1d; 54 block_sizes[2] = elems_per_block; 55 } 56 return CEED_ERROR_SUCCESS; 57 } 58 59 //------------------------------------------------------------------------------ 60 // Determine type of operator 61 //------------------------------------------------------------------------------ 62 static int CeedOperatorBuildKernelData_Hip_gen(Ceed ceed, CeedInt num_input_fields, CeedOperatorField *op_input_fields, 63 CeedQFunctionField *qf_input_fields, CeedInt num_output_fields, CeedOperatorField *op_output_fields, 64 CeedQFunctionField *qf_output_fields, CeedInt *max_P, CeedInt *max_P_1d, CeedInt *Q, CeedInt *Q_1d, 65 CeedInt *max_dim, bool *is_all_tensor, bool *use_3d_slices) { 66 // Check if all are tensor 67 *is_all_tensor = true; 68 for (CeedInt i = 0; i < num_input_fields; i++) { 69 CeedBasis basis; 70 71 CeedCallBackend(CeedOperatorFieldGetBasis(op_input_fields[i], &basis)); 72 if (basis != CEED_BASIS_NONE) { 73 bool is_field_tensor; 74 75 CeedCallBackend(CeedBasisIsTensor(basis, &is_field_tensor)); 76 *is_all_tensor = *is_all_tensor && is_field_tensor; 77 } 78 CeedCallBackend(CeedBasisDestroy(&basis)); 79 } 80 for (CeedInt i = 0; i < num_output_fields; i++) { 81 CeedBasis basis; 82 83 CeedCallBackend(CeedOperatorFieldGetBasis(op_output_fields[i], &basis)); 84 if (basis != CEED_BASIS_NONE) { 85 bool is_field_tensor; 86 87 CeedCallBackend(CeedBasisIsTensor(basis, &is_field_tensor)); 88 *is_all_tensor = *is_all_tensor && is_field_tensor; 89 } 90 CeedCallBackend(CeedBasisDestroy(&basis)); 91 } 92 93 // Find max_P, max_P_1d, Q, and Q_1d 94 bool is_all_3d = true; 95 96 *max_P = 0; 97 *max_P_1d = 0; 98 *Q = 0; 99 *Q_1d = 0; 100 for (CeedInt i = 0; i < num_input_fields; i++) { 101 CeedBasis basis; 102 103 CeedCallBackend(CeedOperatorFieldGetBasis(op_input_fields[i], &basis)); 104 if (basis != CEED_BASIS_NONE) { 105 bool is_field_tensor; 106 CeedInt field_dim = 0, field_P = 0, field_P_1d = 0, field_Q = 0, field_Q_1d = 0; 107 108 // Check if 3D 109 CeedCallBackend(CeedBasisGetDimension(basis, &field_dim)); 110 is_all_3d = is_all_3d && (field_dim == 3); 111 *max_dim = CeedIntMax(*max_dim, field_dim); 112 113 // Collect P, P_1d, Q, and Q_1d 114 CeedCallBackend(CeedBasisGetNumNodes(basis, &field_P)); 115 *max_P = CeedIntMax(*max_P, field_P); 116 CeedCallBackend(CeedBasisIsTensor(basis, &is_field_tensor)); 117 if (is_field_tensor) { 118 CeedCallBackend(CeedBasisGetNumNodes1D(basis, &field_P_1d)); 119 *max_P_1d = CeedIntMax(*max_P_1d, field_P_1d); 120 } 121 CeedCallBackend(CeedBasisGetNumQuadraturePoints(basis, &field_Q)); 122 CeedCheck(*Q == 0 || field_Q == *Q, ceed, CEED_ERROR_BACKEND, "Quadrature spaces must be compatible"); 123 *Q = field_Q; 124 if (is_field_tensor) { 125 CeedCallBackend(CeedBasisGetNumQuadraturePoints1D(basis, &field_Q_1d)); 126 CeedCheck(*Q_1d == 0 || field_Q_1d == *Q_1d, ceed, CEED_ERROR_BACKEND, "Quadrature spaces must be compatible"); 127 *Q_1d = field_Q_1d; 128 } 129 } 130 CeedCallBackend(CeedBasisDestroy(&basis)); 131 } 132 for (CeedInt i = 0; i < num_output_fields; i++) { 133 CeedBasis basis; 134 135 CeedCallBackend(CeedOperatorFieldGetBasis(op_output_fields[i], &basis)); 136 if (basis != CEED_BASIS_NONE) { 137 bool is_field_tensor; 138 CeedInt field_dim = 0, field_P = 0, field_P_1d = 0, field_Q = 0, field_Q_1d = 0; 139 140 // Check if 3D 141 CeedCallBackend(CeedBasisGetDimension(basis, &field_dim)); 142 is_all_3d = is_all_3d && (field_dim == 3); 143 *max_dim = CeedIntMax(*max_dim, field_dim); 144 145 // Collect P, P_1d, Q, and Q_1d 146 CeedCallBackend(CeedBasisGetNumNodes(basis, &field_P)); 147 *max_P = CeedIntMax(*max_P, field_P); 148 CeedCallBackend(CeedBasisIsTensor(basis, &is_field_tensor)); 149 if (is_field_tensor) { 150 CeedCallBackend(CeedBasisGetNumNodes1D(basis, &field_P_1d)); 151 *max_P_1d = CeedIntMax(*max_P_1d, field_P_1d); 152 } 153 CeedCallBackend(CeedBasisGetNumQuadraturePoints(basis, &field_Q)); 154 CeedCheck(*Q == 0 || field_Q == *Q, ceed, CEED_ERROR_BACKEND, "Quadrature spaces must be compatible"); 155 *Q = field_Q; 156 if (is_field_tensor) { 157 CeedCallBackend(CeedBasisGetNumQuadraturePoints1D(basis, &field_Q_1d)); 158 CeedCheck(*Q_1d == 0 || field_Q_1d == *Q_1d, ceed, CEED_ERROR_BACKEND, "Quadrature spaces must be compatible"); 159 *Q_1d = field_Q_1d; 160 } 161 } 162 CeedCallBackend(CeedBasisDestroy(&basis)); 163 } 164 165 // Only use 3D collocated gradient parallelization strategy when gradient is computed 166 *use_3d_slices = false; 167 if (is_all_3d && *is_all_tensor) { 168 bool was_grad_found = false; 169 170 for (CeedInt i = 0; i < num_input_fields; i++) { 171 CeedEvalMode eval_mode; 172 173 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_input_fields[i], &eval_mode)); 174 if (eval_mode == CEED_EVAL_GRAD) { 175 CeedBasis_Hip_shared *basis_data; 176 CeedBasis basis; 177 178 CeedCallBackend(CeedOperatorFieldGetBasis(op_input_fields[i], &basis)); 179 CeedCallBackend(CeedBasisGetData(basis, &basis_data)); 180 *use_3d_slices = basis_data->d_collo_grad_1d && (was_grad_found ? *use_3d_slices : true); 181 was_grad_found = true; 182 CeedCallBackend(CeedBasisDestroy(&basis)); 183 } 184 } 185 for (CeedInt i = 0; i < num_output_fields; i++) { 186 CeedEvalMode eval_mode; 187 188 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_output_fields[i], &eval_mode)); 189 if (eval_mode == CEED_EVAL_GRAD) { 190 CeedBasis_Hip_shared *basis_data; 191 CeedBasis basis; 192 193 CeedCallBackend(CeedOperatorFieldGetBasis(op_output_fields[i], &basis)); 194 CeedCallBackend(CeedBasisGetData(basis, &basis_data)); 195 *use_3d_slices = basis_data->d_collo_grad_1d && (was_grad_found ? *use_3d_slices : true); 196 was_grad_found = true; 197 CeedCallBackend(CeedBasisDestroy(&basis)); 198 } 199 } 200 } 201 return CEED_ERROR_SUCCESS; 202 } 203 204 //------------------------------------------------------------------------------ 205 // Setup fields 206 //------------------------------------------------------------------------------ 207 static int CeedOperatorBuildKernelFieldData_Hip_gen(std::ostringstream &code, CeedOperator_Hip_gen *data, Tab &tab, CeedInt i, 208 CeedOperatorField op_field, CeedQFunctionField qf_field, FieldReuse_Hip field_reuse, 209 CeedInt max_dim, CeedInt Q, CeedInt Q_1d, bool is_input, bool is_all_tensor, bool is_at_points, 210 bool use_3d_slices) { 211 bool is_tensor = true; 212 CeedBasis basis; 213 CeedCallBackend(CeedOperatorFieldGetBasis(op_field, &basis)); 214 if (basis != CEED_BASIS_NONE) CeedCallBackend(CeedBasisIsTensor(basis, &is_tensor)); 215 216 const char *field_name; 217 std::string var_suffix = (is_input ? "_in_" : "_out_") + std::to_string(i); 218 std::string P_name = (is_tensor ? "P_1d" : "P") + var_suffix, Q_name = is_tensor ? "Q_1d" : "Q"; 219 std::string option_name = (is_input ? "inputs" : "outputs"); 220 CeedEvalMode eval_mode = CEED_EVAL_NONE; 221 CeedInt elem_size = 0, num_comp = 0, dim = max_dim, P_1d = 0; 222 CeedElemRestriction elem_rstr; 223 CeedBasis_Hip_shared *basis_data; 224 225 // Field reuse info 226 bool use_previous_field = field_reuse.index != -1; 227 228 CeedCallBackend(CeedOperatorFieldGetName(op_field, &field_name)); 229 code << tab << "// -- " << (is_input ? "Input" : "Output") << " field " << i << ": " << field_name << "\n"; 230 231 // Get field data 232 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_field, &elem_rstr)); 233 if (elem_rstr != CEED_ELEMRESTRICTION_NONE) { 234 CeedCallBackend(CeedElemRestrictionGetElementSize(elem_rstr, &elem_size)); 235 CeedCallBackend(CeedElemRestrictionGetNumComponents(elem_rstr, &num_comp)); 236 } 237 CeedCallBackend(CeedElemRestrictionDestroy(&elem_rstr)); 238 if (basis != CEED_BASIS_NONE) { 239 CeedCallBackend(CeedBasisGetData(basis, &basis_data)); 240 CeedCallBackend(CeedBasisGetDimension(basis, &dim)); 241 if (is_tensor) CeedCallBackend(CeedBasisGetNumNodes1D(basis, &P_1d)); 242 else CeedCallBackend(CeedBasisGetNumNodes(basis, &P_1d)); 243 } 244 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_field, &eval_mode)); 245 246 // Set field constants 247 code << tab << "const CeedInt dim" << var_suffix << " = " << dim << ";\n"; 248 if (is_tensor && !is_all_tensor) { 249 CeedInt P = 0; 250 251 CeedCallBackend(CeedBasisGetNumNodes(basis, &P)); 252 code << tab << "const CeedInt P" << var_suffix << " = " << (basis == CEED_BASIS_NONE ? Q : P) << ";\n"; 253 } 254 code << tab << "const CeedInt " << P_name << " = " << (basis == CEED_BASIS_NONE ? Q_1d : P_1d) << ";\n"; 255 if (eval_mode != CEED_EVAL_WEIGHT) { 256 code << tab << "const CeedInt num_comp" << var_suffix << " = " << num_comp << ";\n"; 257 } 258 259 // Load basis data 260 code << tab << "// EvalMode: " << CeedEvalModes[eval_mode] << "\n"; 261 switch (eval_mode) { 262 case CEED_EVAL_NONE: 263 break; 264 case CEED_EVAL_INTERP: 265 if (is_at_points) { 266 // AtPoints 267 if (!basis_data->d_chebyshev_interp_1d) { 268 CeedSize interp_bytes; 269 CeedScalar *chebyshev_interp_1d; 270 271 interp_bytes = P_1d * Q_1d * sizeof(CeedScalar); 272 CeedCallBackend(CeedCalloc(P_1d * Q_1d, &chebyshev_interp_1d)); 273 CeedCallBackend(CeedBasisGetChebyshevInterp1D(basis, chebyshev_interp_1d)); 274 CeedCallHip(CeedBasisReturnCeed(basis), hipMalloc((void **)&basis_data->d_chebyshev_interp_1d, interp_bytes)); 275 CeedCallHip(CeedBasisReturnCeed(basis), 276 hipMemcpy(basis_data->d_chebyshev_interp_1d, chebyshev_interp_1d, interp_bytes, hipMemcpyHostToDevice)); 277 CeedCallBackend(CeedFree(&chebyshev_interp_1d)); 278 } 279 if (is_input) data->B.inputs[i] = basis_data->d_chebyshev_interp_1d; 280 else data->B.outputs[i] = basis_data->d_chebyshev_interp_1d; 281 } else { 282 // Standard quadrature 283 if (is_input) data->B.inputs[i] = basis_data->d_interp_1d; 284 else data->B.outputs[i] = basis_data->d_interp_1d; 285 } 286 if (use_previous_field) { 287 std::string reuse_var = "s_B" + ((field_reuse.is_input ? "_in_" : "_out_") + std::to_string(field_reuse.index)); 288 289 code << tab << "CeedScalar *s_B" << var_suffix << " = " << reuse_var << ";\n"; 290 } else { 291 code << tab << "__shared__ CeedScalar s_B" << var_suffix << "[" << P_name << "*" << Q_name << "];\n"; 292 code << tab << "LoadMatrix<" << P_name << ", " << Q_name << ">(data, B." << option_name << "[" << i << "], s_B" << var_suffix << ");\n"; 293 } 294 break; 295 case CEED_EVAL_GRAD: 296 if (is_at_points) { 297 // AtPoints 298 if (!basis_data->d_chebyshev_interp_1d) { 299 CeedSize interp_bytes; 300 CeedScalar *chebyshev_interp_1d; 301 302 interp_bytes = P_1d * Q_1d * sizeof(CeedScalar); 303 CeedCallBackend(CeedCalloc(P_1d * Q_1d, &chebyshev_interp_1d)); 304 CeedCallBackend(CeedBasisGetChebyshevInterp1D(basis, chebyshev_interp_1d)); 305 CeedCallHip(CeedBasisReturnCeed(basis), hipMalloc((void **)&basis_data->d_chebyshev_interp_1d, interp_bytes)); 306 CeedCallHip(CeedBasisReturnCeed(basis), 307 hipMemcpy(basis_data->d_chebyshev_interp_1d, chebyshev_interp_1d, interp_bytes, hipMemcpyHostToDevice)); 308 CeedCallBackend(CeedFree(&chebyshev_interp_1d)); 309 } 310 if (is_input) data->B.inputs[i] = basis_data->d_chebyshev_interp_1d; 311 else data->B.outputs[i] = basis_data->d_chebyshev_interp_1d; 312 } else { 313 // Standard quadrature 314 if (is_input) data->B.inputs[i] = basis_data->d_interp_1d; 315 else data->B.outputs[i] = basis_data->d_interp_1d; 316 } 317 if (is_tensor) { 318 if (use_previous_field) { 319 std::string reuse_var = "s_B" + ((field_reuse.is_input ? "_in_" : "_out_") + std::to_string(field_reuse.index)); 320 321 code << tab << "CeedScalar *s_B" << var_suffix << " = " << reuse_var << ";\n"; 322 } else { 323 code << tab << "__shared__ CeedScalar s_B" << var_suffix << "[" << P_name << "*" << Q_name << "];\n"; 324 code << tab << "LoadMatrix<" << P_name << ", " << Q_name << ">(data, B." << option_name << "[" << i << "], s_B" << var_suffix << ");\n"; 325 } 326 } 327 if (is_at_points) break; // No G mat for AtPoints 328 if (use_3d_slices) { 329 if (is_input) data->G.inputs[i] = basis_data->d_collo_grad_1d; 330 else data->G.outputs[i] = basis_data->d_collo_grad_1d; 331 if (use_previous_field && field_reuse.eval_mode == CEED_EVAL_GRAD) { 332 std::string reuse_var = "s_G" + ((field_reuse.is_input ? "_in_" : "_out_") + std::to_string(field_reuse.index)); 333 334 code << tab << "CeedScalar *s_G" << var_suffix << " = " << reuse_var << ";\n"; 335 } else { 336 code << tab << "__shared__ CeedScalar s_G" << var_suffix << "[" << Q_name << "*" << Q_name << "];\n"; 337 code << tab << "LoadMatrix<" << Q_name << ", " << Q_name << ">(data, G." << option_name << "[" << i << "], s_G" << var_suffix << ");\n"; 338 } 339 } else { 340 bool has_collo_grad = basis_data->d_collo_grad_1d; 341 342 if (is_input) data->G.inputs[i] = has_collo_grad ? basis_data->d_collo_grad_1d : basis_data->d_grad_1d; 343 else data->G.outputs[i] = has_collo_grad ? basis_data->d_collo_grad_1d : basis_data->d_grad_1d; 344 if (has_collo_grad) { 345 if (use_previous_field && field_reuse.eval_mode == CEED_EVAL_GRAD) { 346 std::string reuse_var = "s_G" + ((field_reuse.is_input ? "_in_" : "_out_") + std::to_string(field_reuse.index)); 347 348 code << tab << "CeedScalar *s_G" << var_suffix << " = " << reuse_var << ";\n"; 349 } else { 350 code << tab << "__shared__ CeedScalar s_G" << var_suffix << "[" << Q_name << "*" << Q_name << "];\n"; 351 code << tab << "LoadMatrix<" << Q_name << ", " << Q_name << ">(data, G." << option_name << "[" << i << "], s_G" << var_suffix << ");\n"; 352 } 353 } else { 354 if (use_previous_field && field_reuse.eval_mode == CEED_EVAL_GRAD) { 355 std::string reuse_var = "s_G" + ((field_reuse.is_input ? "_in_" : "_out_") + std::to_string(field_reuse.index)); 356 357 code << tab << "CeedScalar *s_G" << var_suffix << " = " << reuse_var << ";\n"; 358 } else { 359 code << tab << "__shared__ CeedScalar s_G" << var_suffix << "[" << P_name << "*" << Q_name << (is_tensor ? "" : "*dim") 360 << (is_tensor ? "" : var_suffix) << "];\n"; 361 code << tab << "LoadMatrix<" << P_name << ", " << Q_name << (is_tensor ? "" : "*dim") << (is_tensor ? "" : var_suffix) << ">(data, G." 362 << option_name << "[" << i << "], s_G" << var_suffix << ");\n"; 363 } 364 } 365 } 366 break; 367 case CEED_EVAL_WEIGHT: 368 break; // No action 369 // LCOV_EXCL_START 370 case CEED_EVAL_DIV: 371 case CEED_EVAL_CURL: 372 break; // TODO: Not implemented 373 // LCOV_EXCL_STOP 374 } 375 CeedCallBackend(CeedBasisDestroy(&basis)); 376 return CEED_ERROR_SUCCESS; 377 } 378 379 //------------------------------------------------------------------------------ 380 // Restriction 381 //------------------------------------------------------------------------------ 382 static int CeedOperatorBuildKernelRestriction_Hip_gen(std::ostringstream &code, CeedOperator_Hip_gen *data, Tab &tab, CeedInt i, 383 CeedInt field_input_buffer[], CeedOperatorField op_field, CeedQFunctionField qf_field, 384 CeedInt max_dim, CeedInt Q_1d, bool is_input, bool is_all_tensor, bool is_at_points, 385 bool use_3d_slices) { 386 std::string var_suffix = (is_input ? "_in_" : "_out_") + std::to_string(i); 387 std::string P_name = (is_all_tensor ? "P_1d" : "P") + var_suffix; 388 CeedEvalMode eval_mode = CEED_EVAL_NONE; 389 CeedInt elem_size = 0, num_comp = 0; 390 CeedSize l_size; 391 CeedRestrictionType rstr_type = CEED_RESTRICTION_STANDARD; 392 CeedElemRestriction_Hip *rstr_data; 393 CeedElemRestriction elem_rstr; 394 395 // Get field data 396 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_field, &elem_rstr)); 397 if (elem_rstr != CEED_ELEMRESTRICTION_NONE) { 398 CeedCallBackend(CeedElemRestrictionGetType(elem_rstr, &rstr_type)); 399 CeedCallBackend(CeedElemRestrictionGetElementSize(elem_rstr, &elem_size)); 400 CeedCallBackend(CeedElemRestrictionGetNumComponents(elem_rstr, &num_comp)); 401 CeedCallBackend(CeedElemRestrictionGetData(elem_rstr, &rstr_data)); 402 } 403 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_field, &eval_mode)); 404 405 // Restriction 406 if (is_input) { 407 // Input 408 if (field_input_buffer[i] != i) { 409 std::string buffer_name = "r_e_in_" + std::to_string(field_input_buffer[i]); 410 411 // Restriction was already done for previous input 412 code << tab << "CeedScalar *r_e" << var_suffix << " = " << buffer_name << ";\n"; 413 } else if (eval_mode != CEED_EVAL_WEIGHT && !((eval_mode == CEED_EVAL_NONE) && use_3d_slices && is_at_points)) { 414 if (eval_mode == CEED_EVAL_NONE && rstr_type != CEED_RESTRICTION_POINTS) { 415 // No basis action, so r_e_in_* in also r_q_in_* and needs to be allocated 416 code << tab << "CeedScalar r_e" << var_suffix << "[num_comp" << var_suffix << "*" << P_name << "];\n"; 417 } else if (rstr_type != CEED_RESTRICTION_POINTS) { 418 // Otherwise we're using the scratch space 419 code << tab << "CeedScalar *r_e" << var_suffix << " = r_e_scratch;\n"; 420 } 421 switch (rstr_type) { 422 case CEED_RESTRICTION_STANDARD: { 423 CeedInt comp_stride; 424 425 CeedCallBackend(CeedElemRestrictionGetLVectorSize(elem_rstr, &l_size)); 426 code << tab << "const CeedInt l_size" << var_suffix << " = " << l_size << ";\n"; 427 CeedCallBackend(CeedElemRestrictionGetCompStride(elem_rstr, &comp_stride)); 428 code << tab << "const CeedInt comp_stride" << var_suffix << " = " << comp_stride << ";\n"; 429 data->indices.inputs[i] = (CeedInt *)rstr_data->d_offsets; 430 code << tab << "ReadLVecStandard" << (is_all_tensor ? max_dim : 1) << "d<num_comp" << var_suffix << ", comp_stride" << var_suffix << ", " 431 << P_name << ">(data, l_size" << var_suffix << ", elem, indices.inputs[" << i << "], d" << var_suffix << ", r_e" << var_suffix 432 << ");\n"; 433 break; 434 } 435 case CEED_RESTRICTION_STRIDED: { 436 bool has_backend_strides; 437 CeedInt num_elem; 438 439 CeedCallBackend(CeedElemRestrictionHasBackendStrides(elem_rstr, &has_backend_strides)); 440 CeedCallBackend(CeedElemRestrictionGetNumElements(elem_rstr, &num_elem)); 441 CeedInt strides[3] = {1, elem_size * num_elem, elem_size}; 442 443 if (!has_backend_strides) { 444 CeedCallBackend(CeedElemRestrictionGetStrides(elem_rstr, strides)); 445 } 446 code << tab << "const CeedInt strides" << var_suffix << "_0 = " << strides[0] << ", strides" << var_suffix << "_1 = " << strides[1] 447 << ", strides" << var_suffix << "_2 = " << strides[2] << ";\n"; 448 code << tab << "ReadLVecStrided" << (is_all_tensor ? max_dim : 1) << "d<num_comp" << var_suffix << ", " << P_name << ", strides" 449 << var_suffix << "_0, strides" << var_suffix << "_1, strides" << var_suffix << "_2>(data, elem, d" << var_suffix << ", r_e" 450 << var_suffix << ");\n"; 451 break; 452 } 453 case CEED_RESTRICTION_POINTS: { 454 CeedInt comp_stride; 455 456 CeedCallBackend(CeedElemRestrictionGetCompStride(elem_rstr, &comp_stride)); 457 code << tab << "const CeedInt comp_stride" << var_suffix << " = " << comp_stride << ";\n"; 458 data->indices.inputs[i] = (CeedInt *)rstr_data->d_offsets; 459 break; 460 } 461 // LCOV_EXCL_START 462 case CEED_RESTRICTION_ORIENTED: 463 case CEED_RESTRICTION_CURL_ORIENTED: 464 break; // TODO: Not implemented 465 // LCOV_EXCL_STOP 466 } 467 } 468 } else { 469 // Output 470 switch (rstr_type) { 471 case CEED_RESTRICTION_STANDARD: { 472 CeedInt comp_stride; 473 474 CeedCallBackend(CeedElemRestrictionGetLVectorSize(elem_rstr, &l_size)); 475 code << tab << "const CeedInt l_size" << var_suffix << " = " << l_size << ";\n"; 476 CeedCallBackend(CeedElemRestrictionGetCompStride(elem_rstr, &comp_stride)); 477 code << tab << "const CeedInt comp_stride" << var_suffix << " = " << comp_stride << ";\n"; 478 data->indices.outputs[i] = (CeedInt *)rstr_data->d_offsets; 479 code << tab << "WriteLVecStandard" << (is_all_tensor ? max_dim : 1) << "d<num_comp" << var_suffix << ", comp_stride" << var_suffix << ", " 480 << P_name << ">(data, l_size" << var_suffix << ", elem, indices.outputs[" << i << "], r_e" << var_suffix << ", d" << var_suffix 481 << ");\n"; 482 break; 483 } 484 case CEED_RESTRICTION_STRIDED: { 485 bool has_backend_strides; 486 CeedInt num_elem; 487 488 CeedCallBackend(CeedElemRestrictionHasBackendStrides(elem_rstr, &has_backend_strides)); 489 CeedCallBackend(CeedElemRestrictionGetNumElements(elem_rstr, &num_elem)); 490 CeedInt strides[3] = {1, elem_size * num_elem, elem_size}; 491 492 if (!has_backend_strides) { 493 CeedCallBackend(CeedElemRestrictionGetStrides(elem_rstr, strides)); 494 } 495 code << tab << "const CeedInt strides" << var_suffix << "_0 = " << strides[0] << ", strides" << var_suffix << "_1 = " << strides[1] 496 << ", strides" << var_suffix << "_2 = " << strides[2] << ";\n"; 497 code << tab << "WriteLVecStrided" << (is_all_tensor ? max_dim : 1) << "d<num_comp" << var_suffix << ", " << P_name << ", strides" 498 << var_suffix << "_0, strides" << var_suffix << "_1, strides" << var_suffix << "_2>(data, elem, r_e" << var_suffix << ", d" << var_suffix 499 << ");\n"; 500 break; 501 } 502 case CEED_RESTRICTION_POINTS: 503 data->indices.outputs[i] = (CeedInt *)rstr_data->d_offsets; 504 break; 505 // LCOV_EXCL_START 506 case CEED_RESTRICTION_ORIENTED: 507 case CEED_RESTRICTION_CURL_ORIENTED: 508 break; // TODO: Not implemented 509 // LCOV_EXCL_STOP 510 } 511 } 512 CeedCallBackend(CeedElemRestrictionDestroy(&elem_rstr)); 513 return CEED_ERROR_SUCCESS; 514 } 515 516 //------------------------------------------------------------------------------ 517 // Basis 518 //------------------------------------------------------------------------------ 519 static int CeedOperatorBuildKernelBasis_Hip_gen(std::ostringstream &code, CeedOperator_Hip_gen *data, Tab &tab, CeedInt i, CeedOperatorField op_field, 520 CeedQFunctionField qf_field, CeedInt max_dim, CeedInt Q_1d, bool is_input, bool is_all_tensor, 521 bool is_at_points, bool use_3d_slices) { 522 bool is_tensor = true; 523 CeedBasis basis; 524 CeedCallBackend(CeedOperatorFieldGetBasis(op_field, &basis)); 525 CeedCallBackend(CeedBasisIsTensor(basis, &is_tensor)); 526 527 std::string var_suffix = (is_input ? "_in_" : "_out_") + std::to_string(i); 528 std::string P_name = (is_tensor ? "P_1d" : "P") + var_suffix, Q_name = is_tensor ? "Q_1d" : "Q"; 529 CeedEvalMode eval_mode = CEED_EVAL_NONE; 530 CeedInt dim = max_dim, elem_size = 0, num_comp = 0, P_1d = 0; 531 CeedElemRestriction elem_rstr; 532 533 // Get field data 534 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_field, &elem_rstr)); 535 if (elem_rstr != CEED_ELEMRESTRICTION_NONE) { 536 CeedCallBackend(CeedElemRestrictionGetElementSize(elem_rstr, &elem_size)); 537 CeedCallBackend(CeedElemRestrictionGetNumComponents(elem_rstr, &num_comp)); 538 } 539 CeedCallBackend(CeedElemRestrictionDestroy(&elem_rstr)); 540 if (basis != CEED_BASIS_NONE) { 541 CeedCallBackend(CeedBasisGetDimension(basis, &dim)); 542 if (is_tensor) CeedCallBackend(CeedBasisGetNumNodes1D(basis, &P_1d)); 543 else CeedCallBackend(CeedBasisGetNumNodes(basis, &P_1d)); 544 } 545 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_field, &eval_mode)); 546 547 // Basis 548 code << tab << "// EvalMode: " << CeedEvalModes[eval_mode] << "\n"; 549 if (is_input) { 550 switch (eval_mode) { 551 case CEED_EVAL_NONE: 552 if (!use_3d_slices && !is_at_points) { 553 code << tab << "CeedScalar *r_q" << var_suffix << " = r_e" << var_suffix << ";\n"; 554 } 555 break; 556 case CEED_EVAL_INTERP: 557 if (is_at_points) { 558 std::string function_name = (dim == 1 ? "Interp" : "InterpTensor") + std::to_string(dim) + "d"; 559 560 code << tab << "CeedScalar r_c" << var_suffix << "[num_comp" << var_suffix << "*" << (dim >= 3 ? Q_name : "1") << "];\n"; 561 code << tab << function_name << "<num_comp" << var_suffix << ", " << P_name << ", " << Q_name << ", OP_T_1D>(data, r_e" << var_suffix 562 << ", s_B" << var_suffix << ", r_c" << var_suffix << ");\n"; 563 } else { 564 std::string function_name = is_tensor 565 ? ((dim == 1 ? "Interp" : "InterpTensor") + std::to_string(dim) + "d" + (is_all_tensor ? "" : "Flattened")) 566 : "InterpNonTensor"; 567 std::string op_t_1d_name = (is_all_tensor || !is_tensor) ? "OP_T_1D" : (P_1d > Q_1d ? P_name : Q_name); 568 569 code << tab << "CeedScalar r_q" << var_suffix << "[num_comp" << var_suffix << "*" << (is_all_tensor && (dim >= 3) ? Q_name : "1") << "];\n"; 570 code << tab << function_name << "<num_comp" << var_suffix << ", " << P_name << ", " << Q_name << ", " << op_t_1d_name << ">(data, r_e" 571 << var_suffix << ", s_B" << var_suffix << ", r_q" << var_suffix << ");\n"; 572 } 573 break; 574 case CEED_EVAL_GRAD: 575 if (is_at_points) { 576 std::string function_name = (dim == 1 ? "Interp" : "InterpTensor") + std::to_string(dim) + "d"; 577 578 code << tab << "CeedScalar r_c" << var_suffix << "[num_comp" << var_suffix << "*" << (dim >= 3 ? Q_name : "1") << "];\n"; 579 code << tab << function_name << "<num_comp" << var_suffix << ", " << P_name << ", " << Q_name << ", OP_T_1D>(data, r_e" << var_suffix 580 << ", s_B" << var_suffix << ", r_c" << var_suffix << ");\n"; 581 } else if (use_3d_slices) { 582 std::string function_name = (dim > 1 ? "InterpTensor" : "Interp") + std::to_string(dim) + "d"; 583 584 code << tab << "CeedScalar r_q" << var_suffix << "[num_comp" << var_suffix << "*" << Q_name << "];\n"; 585 code << tab << function_name << "<num_comp" << var_suffix << ", " << P_name << ", " << Q_name << ", OP_T_1D>(data, r_e" << var_suffix 586 << ", s_B" << var_suffix << ", r_q" << var_suffix << ");\n"; 587 } else if (is_tensor) { 588 bool is_collocated = dim == 3 && Q_1d >= P_1d; 589 std::string function_name = (dim == 1 ? "Grad" : (is_collocated ? "GradTensorCollocated" : "GradTensor")) + std::to_string(dim) + "d" + 590 (is_all_tensor ? "" : "Flattened"); 591 std::string op_t_1d_name = is_all_tensor ? "OP_T_1D" : (P_1d > Q_1d ? P_name : Q_name); 592 593 code << tab << "CeedScalar r_q" << var_suffix << "[num_comp" << var_suffix << "*dim" << var_suffix << "*" 594 << (is_all_tensor && dim >= 3 ? Q_name : "1") << "];\n"; 595 code << tab << function_name << "<num_comp" << var_suffix << ", " << P_name << ", " << Q_name << ", " << op_t_1d_name << ">(data, r_e" 596 << var_suffix << ", s_B" << var_suffix << ", s_G" << var_suffix << ", r_q" << var_suffix << ");\n"; 597 } else { 598 std::string function_name = "GradNonTensor"; 599 600 code << tab << "CeedScalar r_q" << var_suffix << "[num_comp" << var_suffix << "*dim" << var_suffix << "];\n"; 601 code << tab << function_name << "<num_comp" << var_suffix << ", dim" << var_suffix << ", " << P_name << ", " << Q_name 602 << ", OP_T_1D>(data, r_e" << var_suffix << ", s_G" << var_suffix << ", r_q" << var_suffix << ");\n"; 603 } 604 break; 605 case CEED_EVAL_WEIGHT: { 606 if (is_at_points) { 607 code << tab << "// Nothing to do AtPoints\n"; 608 } else { 609 CeedBasis_Hip_shared *basis_data; 610 std::string function_name = is_tensor 611 ? ((dim == 1 ? "Weight" : "WeightTensor") + std::to_string(dim) + "d" + (is_all_tensor ? "" : "Flattened")) 612 : "WeightNonTensor"; 613 614 code << tab << "CeedScalar r_q" << var_suffix << "[" << (is_all_tensor && (dim >= 3) ? Q_name : "1") << "];\n"; 615 CeedCallBackend(CeedBasisGetData(basis, &basis_data)); 616 data->W = basis_data->d_q_weight_1d; 617 code << tab << function_name << "<" << P_name << ", " << Q_name << ">(data, W, r_q" << var_suffix << ");\n"; 618 } 619 break; 620 } 621 // LCOV_EXCL_START 622 case CEED_EVAL_DIV: 623 case CEED_EVAL_CURL: 624 break; // TODO: Not implemented 625 // LCOV_EXCL_STOP 626 } 627 } else { 628 switch (eval_mode) { 629 case CEED_EVAL_NONE: 630 code << tab << "CeedScalar *r_e" << var_suffix << " = r_q" << var_suffix << ";\n"; 631 break; // No action 632 case CEED_EVAL_INTERP: 633 code << tab << "CeedScalar *r_e" << var_suffix << " = r_e_scratch;\n"; 634 if (is_at_points) { 635 std::string function_name = (dim == 1 ? "InterpTranspose" : "InterpTransposeTensor") + std::to_string(dim) + "d"; 636 637 code << tab << function_name << "<num_comp" << var_suffix << ", " << P_name << ", " << Q_name << ", OP_T_1D>(data, r_c" << var_suffix 638 << ", s_B" << var_suffix << ", r_e" << var_suffix << ");\n"; 639 } else { 640 std::string function_name = 641 is_tensor ? ((dim == 1 ? "InterpTranspose" : "InterpTransposeTensor") + std::to_string(dim) + "d" + (is_all_tensor ? "" : "Flattened")) 642 : "InterpTransposeNonTensor"; 643 std::string op_t_1d_name = (is_all_tensor || !is_tensor) ? "OP_T_1D" : (P_1d > Q_1d ? P_name : Q_name); 644 645 code << tab << function_name << "<num_comp" << var_suffix << ", " << P_name << ", " << Q_name << ", " << op_t_1d_name << ">(data, r_q" 646 << var_suffix << ", s_B" << var_suffix << ", r_e" << var_suffix << ");\n"; 647 } 648 break; 649 case CEED_EVAL_GRAD: 650 code << tab << "CeedScalar *r_e" << var_suffix << " = r_e_scratch;\n"; 651 if (is_at_points) { 652 std::string function_name = (dim == 1 ? "InterpTranspose" : "InterpTransposeTensor") + std::to_string(dim) + "d"; 653 654 code << tab << function_name << "<num_comp" << var_suffix << ", " << P_name << ", " << Q_name << ", OP_T_1D>(data, r_c" << var_suffix 655 << ", s_B" << var_suffix << ", r_e" << var_suffix << ");\n"; 656 } else if (use_3d_slices) { 657 std::string function_name = (dim == 1 ? "InterpTranspose" : "InterpTransposeTensor") + std::to_string(dim) + "d"; 658 659 code << tab << function_name << "<num_comp" << var_suffix << ", " << P_name << ", " << Q_name << ", OP_T_1D>(data, r_q" << var_suffix 660 << ", s_B" << var_suffix << ", r_e" << var_suffix << ");\n"; 661 } else if (is_tensor) { 662 bool is_collocated = dim == 3 && Q_1d >= P_1d; 663 std::string function_name = (dim == 1 ? "GradTranspose" : (is_collocated ? "GradTransposeTensorCollocated" : "GradTransposeTensor")) + 664 std::to_string(dim) + "d" + (is_all_tensor ? "" : "Flattened"); 665 std::string op_t_1d_name = is_all_tensor ? "OP_T_1D" : (P_1d > Q_1d ? P_name : Q_name); 666 667 code << tab << function_name << "<num_comp" << var_suffix << ", " << P_name << ", " << Q_name << ", " << op_t_1d_name << ">(data, r_q" 668 << var_suffix << ", s_B" << var_suffix << ", s_G" << var_suffix << ", r_e" << var_suffix << ");\n"; 669 } else { 670 std::string function_name = "GradTransposeNonTensor"; 671 672 code << tab << function_name << "<num_comp" << var_suffix << ", dim" << var_suffix << ", " << P_name << ", " << Q_name 673 << ", OP_T_1D>(data, r_q" << var_suffix << ", s_G" << var_suffix << ", r_e" << var_suffix << ");\n"; 674 } 675 break; 676 // LCOV_EXCL_START 677 case CEED_EVAL_WEIGHT: 678 break; // Should not occur 679 case CEED_EVAL_DIV: 680 case CEED_EVAL_CURL: 681 break; // TODO: Not implemented 682 // LCOV_EXCL_STOP 683 } 684 } 685 CeedCallBackend(CeedBasisDestroy(&basis)); 686 return CEED_ERROR_SUCCESS; 687 } 688 689 //------------------------------------------------------------------------------ 690 // QFunction 691 //------------------------------------------------------------------------------ 692 static int CeedOperatorBuildKernelQFunction_Hip_gen(std::ostringstream &code, CeedOperator_Hip_gen *data, Tab &tab, CeedInt max_dim, 693 CeedInt max_num_points, CeedInt num_input_fields, CeedOperatorField *op_input_fields, 694 CeedQFunctionField *qf_input_fields, CeedInt num_output_fields, 695 CeedOperatorField *op_output_fields, CeedQFunctionField *qf_output_fields, 696 std::string qfunction_name, CeedInt Q_1d, bool is_all_tensor, bool is_at_points, 697 bool use_3d_slices) { 698 std::string Q_name = is_all_tensor ? "Q_1d" : "Q"; 699 CeedEvalMode eval_mode = CEED_EVAL_NONE; 700 CeedElemRestriction elem_rstr; 701 702 // Setup output arrays 703 code << "\n"; 704 code << tab << "// -- Output field setup\n"; 705 for (CeedInt i = 0; i < num_output_fields; i++) { 706 const char *field_name; 707 std::string var_suffix = "_out_" + std::to_string(i); 708 709 CeedCallBackend(CeedOperatorFieldGetName(op_output_fields[i], &field_name)); 710 code << tab << "// ---- Output field " << i << ": " << field_name << "\n"; 711 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_output_fields[i], &eval_mode)); 712 switch (eval_mode) { 713 case CEED_EVAL_NONE: 714 if (is_at_points) { 715 code << tab << "CeedScalar r_q" << var_suffix << "[num_comp" << var_suffix << "];\n"; 716 } else { 717 code << tab << "CeedScalar r_q" << var_suffix << "[num_comp" << var_suffix << "*" << (is_all_tensor && (max_dim >= 3) ? Q_name : "1") 718 << "];\n"; 719 } 720 break; 721 case CEED_EVAL_INTERP: 722 if (is_at_points) { 723 // Accumulator for point data 724 code << tab << "CeedScalar r_c" << var_suffix << "[num_comp" << var_suffix << "*" << (max_dim >= 3 ? Q_name : "1") << "];\n"; 725 code << tab << "for (CeedInt i = 0; i < num_comp" << var_suffix << "*" << (max_dim >= 3 ? Q_name : "1") << "; i++) r_c" << var_suffix 726 << "[i] = 0.0;\n"; 727 } else { 728 code << tab << "CeedScalar r_q" << var_suffix << "[num_comp" << var_suffix << "*" << (is_all_tensor && (max_dim >= 3) ? Q_name : "1") 729 << "];\n"; 730 } 731 break; 732 case CEED_EVAL_GRAD: 733 if (is_at_points) { 734 // Accumulator for point data 735 code << tab << "CeedScalar r_c" << var_suffix << "[num_comp" << var_suffix << "*" << (max_dim >= 3 ? Q_name : "1") << "];\n"; 736 code << tab << "for (CeedInt i = 0; i < num_comp" << var_suffix << "*" << (max_dim >= 3 ? Q_name : "1") << "; i++) r_c" << var_suffix 737 << "[i] = 0.0;\n"; 738 } else if (use_3d_slices) { 739 // Accumulator for gradient slices 740 code << tab << "CeedScalar r_q" << var_suffix << "[num_comp" << var_suffix << "*" << Q_name << "];\n"; 741 code << tab << "for (CeedInt i = 0; i < num_comp" << var_suffix << "*" << Q_name << "; i++) r_q" << var_suffix << "[i] = 0.0;\n"; 742 } else { 743 code << tab << "CeedScalar r_q" << var_suffix << "[num_comp" << var_suffix << "*dim" << var_suffix << "*" 744 << (is_all_tensor && (max_dim >= 3) ? Q_name : "1") << "];\n"; 745 } 746 break; 747 case CEED_EVAL_WEIGHT: 748 break; 749 // LCOV_EXCL_START 750 case CEED_EVAL_DIV: 751 case CEED_EVAL_CURL: 752 break; // TODO: Not implemented 753 // LCOV_EXCL_STOP 754 } 755 } 756 757 if (is_at_points) { 758 // We need to handle batches of points 759 code << "\n"; 760 code << tab << "// Note: Using batches of points\n"; 761 code << tab << "const CeedInt point_loop_bound = (blockDim.x*blockDim.y) * ceil((1.0*max_num_points) / (blockDim.x*blockDim.y));\n\n"; 762 code << tab << "#pragma unroll\n"; 763 code << tab << "for (CeedInt i = threadIdx.x + threadIdx.y*blockDim.x; i < point_loop_bound; i += blockDim.x*blockDim.y) {\n"; 764 tab.push(); 765 code << tab << "const CeedInt p = i % max_num_points;\n\n"; 766 767 code << tab << "// -- Coordinates\n"; 768 code << tab << "CeedScalar r_x[max_dim];\n"; 769 code << tab << "ReadPoint<max_dim, coords_comp_stride, max_num_points>(data, elem, p, max_num_points, points.indices, points.coords, r_x);\n\n"; 770 771 code << tab << "// -- Input fields\n"; 772 for (CeedInt i = 0; i < num_input_fields; i++) { 773 const char *field_name; 774 std::string var_suffix = "_in_" + std::to_string(i); 775 std::string P_name = "P_1d" + var_suffix; 776 777 CeedCallBackend(CeedOperatorFieldGetName(op_input_fields[i], &field_name)); 778 code << tab << "// ---- Input field " << i << ": " << field_name << "\n"; 779 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_input_fields[i], &eval_mode)); 780 // Basis action 781 code << tab << "// EvalMode: " << CeedEvalModes[eval_mode] << "\n"; 782 switch (eval_mode) { 783 case CEED_EVAL_NONE: 784 code << tab << "CeedScalar r_s" << var_suffix << "[num_comp" << var_suffix << "];\n"; 785 code << tab << "ReadPoint<num_comp" << var_suffix << ", comp_stride" << var_suffix 786 << ", max_num_points>(data, elem, p, max_num_points, indices.inputs[" << i << "], d" << var_suffix << ", r_s" << var_suffix << ");\n"; 787 break; 788 case CEED_EVAL_INTERP: 789 code << tab << "CeedScalar r_s" << var_suffix << "[num_comp" << var_suffix << "];\n"; 790 code << tab << "InterpAtPoints" << max_dim << "d<num_comp" << var_suffix << ", max_num_points, " << P_name << ", " << Q_name 791 << ">(data, i, r_c" << var_suffix << ", r_x, r_s" << var_suffix << ");\n"; 792 break; 793 case CEED_EVAL_GRAD: 794 code << tab << "CeedScalar r_s" << var_suffix << "[num_comp" << var_suffix << "*dim" << var_suffix << "];\n"; 795 code << tab << "GradAtPoints" << max_dim << "d<num_comp" << var_suffix << ", max_num_points, " << P_name << ", " << Q_name 796 << ">(data, i, r_c" << var_suffix << ", r_x, r_s" << var_suffix << ");\n"; 797 break; 798 case CEED_EVAL_WEIGHT: 799 code << tab << "CeedScalar r_s" << var_suffix << "[1];\n"; 800 code << tab << "r_s" << var_suffix << "[0] = 1.0;\n"; 801 break; 802 // LCOV_EXCL_START 803 case CEED_EVAL_DIV: 804 case CEED_EVAL_CURL: 805 break; // TODO: Not implemented 806 // LCOV_EXCL_STOP 807 } 808 } 809 code << "\n"; 810 code << tab << "// -- Output fields\n"; 811 for (CeedInt i = 0; i < num_output_fields; i++) { 812 const char *field_name; 813 std::string var_suffix = "_out_" + std::to_string(i); 814 815 CeedCallBackend(CeedOperatorFieldGetName(op_output_fields[i], &field_name)); 816 code << tab << "// ---- Output field " << i << ": " << field_name << "\n"; 817 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_output_fields[i], &eval_mode)); 818 // Basis action 819 switch (eval_mode) { 820 case CEED_EVAL_NONE: 821 code << tab << "CeedScalar r_s" << var_suffix << "[num_comp" << var_suffix << "];\n"; 822 break; 823 case CEED_EVAL_INTERP: 824 code << tab << "CeedScalar r_s" << var_suffix << "[num_comp" << var_suffix << "];\n"; 825 break; 826 case CEED_EVAL_GRAD: 827 code << tab << "CeedScalar r_s" << var_suffix << "[num_comp" << var_suffix << "*dim" << var_suffix << "];\n"; 828 break; 829 // LCOV_EXCL_START 830 case CEED_EVAL_WEIGHT: 831 break; // Should not occur 832 case CEED_EVAL_DIV: 833 case CEED_EVAL_CURL: 834 break; // TODO: Not implemented 835 // LCOV_EXCL_STOP 836 } 837 } 838 839 } else if (use_3d_slices) { 840 // We treat quadrature points per slice in 3d to save registers 841 code << "\n"; 842 code << tab << "// Note: Using planes of 3D elements\n"; 843 code << tab << "#pragma unroll\n"; 844 code << tab << "for (CeedInt q = 0; q < " << Q_name << "; q++) {\n"; 845 tab.push(); 846 code << tab << "// -- Input fields\n"; 847 for (CeedInt i = 0; i < num_input_fields; i++) { 848 const char *field_name; 849 std::string var_suffix = "_in_" + std::to_string(i); 850 851 CeedCallBackend(CeedOperatorFieldGetName(op_input_fields[i], &field_name)); 852 code << tab << "// ---- Input field " << i << ": " << field_name << "\n"; 853 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_input_fields[i], &eval_mode)); 854 // Basis action 855 code << tab << "// EvalMode: " << CeedEvalModes[eval_mode] << "\n"; 856 switch (eval_mode) { 857 case CEED_EVAL_NONE: 858 bool is_strided; 859 860 code << tab << "CeedScalar r_s" << var_suffix << "[num_comp" << var_suffix << "];\n"; 861 862 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_input_fields[i], &elem_rstr)); 863 CeedCallBackend(CeedElemRestrictionIsStrided(elem_rstr, &is_strided)); 864 if (is_strided) { 865 bool has_backend_strides; 866 CeedInt num_elem, elem_size; 867 868 CeedCallBackend(CeedElemRestrictionGetElementSize(elem_rstr, &elem_size)); 869 CeedCallBackend(CeedElemRestrictionHasBackendStrides(elem_rstr, &has_backend_strides)); 870 CeedCallBackend(CeedElemRestrictionGetNumElements(elem_rstr, &num_elem)); 871 CeedInt strides[3] = {1, elem_size * num_elem, elem_size}; 872 873 if (!has_backend_strides) { 874 CeedCallBackend(CeedElemRestrictionGetStrides(elem_rstr, strides)); 875 } 876 code << tab << "const CeedInt strides" << var_suffix << "_0 = " << strides[0] << ", strides" << var_suffix << "_1 = " << strides[1] 877 << ", strides" << var_suffix << "_2 = " << strides[2] << ";\n"; 878 code << tab << "ReadEVecSliceStrided3d<num_comp" << var_suffix << ", " << Q_name << ", strides" << var_suffix << "_0, strides" 879 << var_suffix << "_1, strides" << var_suffix << "_2>(data, elem, q, d" << var_suffix << ", r_s" << var_suffix << ");\n"; 880 } else { 881 CeedSize l_size = 0; 882 CeedInt comp_stride; 883 CeedElemRestriction_Hip *rstr_data; 884 885 CeedCallBackend(CeedElemRestrictionGetLVectorSize(elem_rstr, &l_size)); 886 code << tab << "const CeedInt l_size" << var_suffix << " = " << l_size << ";\n"; 887 CeedCallBackend(CeedElemRestrictionGetCompStride(elem_rstr, &comp_stride)); 888 code << tab << "const CeedInt comp_stride" << var_suffix << " = " << comp_stride << ";\n"; 889 CeedCallBackend(CeedElemRestrictionGetData(elem_rstr, &rstr_data)); 890 data->indices.inputs[i] = (CeedInt *)rstr_data->d_offsets; 891 code << tab << "ReadEVecSliceStandard3d<num_comp" << var_suffix << ", comp_stride" << var_suffix << ", " << Q_name << ">(data, l_size" 892 << var_suffix << ", elem, q, indices.inputs[" << i << "], d" << var_suffix << ", r_s" << var_suffix << ");\n"; 893 } 894 CeedCallBackend(CeedElemRestrictionDestroy(&elem_rstr)); 895 break; 896 case CEED_EVAL_INTERP: 897 code << tab << "CeedScalar r_s" << var_suffix << "[num_comp" << var_suffix << "];\n"; 898 code << tab << "for (CeedInt j = 0; j < num_comp" << var_suffix << "; j++) {\n"; 899 tab.push(); 900 code << tab << "r_s" << var_suffix << "[j] = r_q" << var_suffix << "[q + j*" << Q_name << "];\n"; 901 tab.pop(); 902 code << tab << "}\n"; 903 break; 904 case CEED_EVAL_GRAD: 905 code << tab << "CeedScalar r_s" << var_suffix << "[num_comp" << var_suffix << "*dim" << var_suffix << "];\n"; 906 code << tab << "GradColloSlice3d<num_comp" << var_suffix << ", " << Q_name << ", OP_T_1D>(data, q, r_q" << var_suffix << ", s_G" 907 << var_suffix << ", r_s" << var_suffix << ");\n"; 908 break; 909 case CEED_EVAL_WEIGHT: 910 code << tab << "CeedScalar r_s" << var_suffix << "[1];\n"; 911 code << tab << "r_s" << var_suffix << "[0] = r_q" << var_suffix << "[q];\n"; 912 break; 913 // LCOV_EXCL_START 914 case CEED_EVAL_DIV: 915 case CEED_EVAL_CURL: 916 break; // TODO: Not implemented 917 // LCOV_EXCL_STOP 918 } 919 } 920 code << "\n"; 921 code << tab << "// -- Output fields\n"; 922 for (CeedInt i = 0; i < num_output_fields; i++) { 923 const char *field_name; 924 std::string var_suffix = "_out_" + std::to_string(i); 925 926 CeedCallBackend(CeedOperatorFieldGetName(op_output_fields[i], &field_name)); 927 code << tab << "// ---- Output field " << i << ": " << field_name << "\n"; 928 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_output_fields[i], &eval_mode)); 929 // Basis action 930 switch (eval_mode) { 931 case CEED_EVAL_NONE: 932 code << tab << "CeedScalar r_s" << var_suffix << "[num_comp" << var_suffix << "];\n"; 933 break; 934 case CEED_EVAL_INTERP: 935 code << tab << "CeedScalar r_s" << var_suffix << "[num_comp" << var_suffix << "];\n"; 936 break; 937 case CEED_EVAL_GRAD: 938 code << tab << "CeedScalar r_s" << var_suffix << "[num_comp" << var_suffix << "*dim" << var_suffix << "];\n"; 939 break; 940 // LCOV_EXCL_START 941 case CEED_EVAL_WEIGHT: 942 break; // Should not occur 943 case CEED_EVAL_DIV: 944 case CEED_EVAL_CURL: 945 break; // TODO: Not implemented 946 // LCOV_EXCL_STOP 947 } 948 } 949 } else { 950 code << "\n"; 951 code << tab << "// Note: Using full elements\n"; 952 code << tab << "{\n"; 953 tab.push(); 954 code << tab << "// -- Input fields\n"; 955 for (CeedInt i = 0; i < num_input_fields; i++) { 956 const char *field_name; 957 958 CeedCallBackend(CeedOperatorFieldGetName(op_input_fields[i], &field_name)); 959 code << tab << "// ---- Input field " << i << ": " << field_name << "\n"; 960 code << tab << "CeedScalar *r_s_in_" << i << " = r_q_in_" << i << ";\n"; 961 } 962 code << tab << "// -- Output fields\n"; 963 for (CeedInt i = 0; i < num_output_fields; i++) { 964 const char *field_name; 965 966 CeedCallBackend(CeedOperatorFieldGetName(op_output_fields[i], &field_name)); 967 code << tab << "// ---- Output field " << i << ": " << field_name << "\n"; 968 code << tab << "CeedScalar *r_s_out_" << i << " = r_q_out_" << i << ";\n"; 969 } 970 } 971 972 // Input and output buffers 973 code << "\n"; 974 code << tab << "// -- QFunction inputs and outputs\n"; 975 code << tab << "// ---- Inputs\n"; 976 code << tab << "CeedScalar *inputs[" << CeedIntMax(num_input_fields, 1) << "];\n"; 977 for (CeedInt i = 0; i < num_input_fields; i++) { 978 const char *field_name; 979 980 CeedCallBackend(CeedOperatorFieldGetName(op_input_fields[i], &field_name)); 981 code << tab << "// ------ Input field " << i << ": " << field_name << "\n"; 982 code << tab << "inputs[" << i << "] = r_s_in_" << i << ";\n"; 983 } 984 code << tab << "// ---- Outputs\n"; 985 code << tab << "CeedScalar *outputs[" << CeedIntMax(num_output_fields, 1) << "];\n"; 986 for (CeedInt i = 0; i < num_output_fields; i++) { 987 const char *field_name; 988 989 CeedCallBackend(CeedOperatorFieldGetName(op_output_fields[i], &field_name)); 990 code << tab << "// ------ Output field " << i << ": " << field_name << "\n"; 991 code << tab << "outputs[" << i << "] = r_s_out_" << i << ";\n"; 992 } 993 994 // Apply QFunction 995 code << "\n"; 996 code << tab << "// -- Apply QFunction\n"; 997 code << tab << "" << qfunction_name << "(ctx, "; 998 if (max_dim != 3 || is_at_points || use_3d_slices || !is_all_tensor) { 999 code << "1"; 1000 } else { 1001 code << Q_name; 1002 } 1003 code << ", inputs, outputs);\n"; 1004 1005 if (is_at_points) { 1006 // Map back to coefficients 1007 code << "\n"; 1008 code << tab << "// -- Output fields\n"; 1009 for (CeedInt i = 0; i < num_output_fields; i++) { 1010 const char *field_name; 1011 std::string var_suffix = "_out_" + std::to_string(i); 1012 std::string P_name = "P_1d" + var_suffix; 1013 1014 CeedCallBackend(CeedOperatorFieldGetName(op_output_fields[i], &field_name)); 1015 code << tab << "// ---- Output field " << i << ": " << field_name << "\n"; 1016 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_output_fields[i], &eval_mode)); 1017 // Basis action 1018 code << tab << "// EvalMode: " << CeedEvalModes[eval_mode] << "\n"; 1019 switch (eval_mode) { 1020 case CEED_EVAL_NONE: { 1021 CeedInt comp_stride; 1022 CeedElemRestriction elem_rstr; 1023 1024 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_output_fields[i], &elem_rstr)); 1025 CeedCallBackend(CeedElemRestrictionGetCompStride(elem_rstr, &comp_stride)); 1026 CeedCallBackend(CeedElemRestrictionDestroy(&elem_rstr)); 1027 code << tab << "const CeedInt comp_stride" << var_suffix << " = " << comp_stride << ";\n"; 1028 code << tab << "WritePoint<num_comp" << var_suffix << ", comp_stride" << var_suffix 1029 << ", max_num_points>(data, elem, i, points.num_per_elem[elem], indices.outputs[" << i << "]" 1030 << ", r_s" << var_suffix << ", d" << var_suffix << ");\n"; 1031 break; 1032 } 1033 case CEED_EVAL_INTERP: 1034 code << tab << "if (i >= points.num_per_elem[elem]) {\n"; 1035 tab.push(); 1036 code << tab << "for (CeedInt j = 0; j < num_comp" << var_suffix << "; j++) r_s" << var_suffix << "[j] = 0.0;\n"; 1037 tab.pop(); 1038 code << tab << "}\n"; 1039 code << tab << "InterpTransposeAtPoints" << max_dim << "d<num_comp" << var_suffix << ", max_num_points, " << P_name << ", " << Q_name 1040 << ">(data, i, r_s" << var_suffix << ", r_x, r_c" << var_suffix << ");\n"; 1041 break; 1042 case CEED_EVAL_GRAD: 1043 code << tab << "if (i >= points.num_per_elem[elem]) {\n"; 1044 tab.push(); 1045 code << tab << "for (CeedInt j = 0; j < num_comp" << var_suffix << "*dim" << var_suffix << "; j++) r_s" << var_suffix << "[j] = 0.0;\n"; 1046 tab.pop(); 1047 code << tab << "}\n"; 1048 code << tab << "GradTransposeAtPoints" << max_dim << "d<num_comp" << var_suffix << ", max_num_points, " << P_name << ", " << Q_name 1049 << ">(data, i, r_s" << var_suffix << ", r_x, r_c" << var_suffix << ");\n"; 1050 break; 1051 // LCOV_EXCL_START 1052 case CEED_EVAL_WEIGHT: 1053 break; // Should not occur 1054 case CEED_EVAL_DIV: 1055 case CEED_EVAL_CURL: 1056 break; // TODO: Not implemented 1057 // LCOV_EXCL_STOP 1058 } 1059 } 1060 } else if (use_3d_slices) { 1061 // Copy or apply transpose grad, if needed 1062 code << "\n"; 1063 code << tab << "// -- Output fields\n"; 1064 for (CeedInt i = 0; i < num_output_fields; i++) { 1065 const char *field_name; 1066 std::string var_suffix = "_out_" + std::to_string(i); 1067 std::string P_name = "P_1d" + var_suffix; 1068 1069 CeedCallBackend(CeedOperatorFieldGetName(op_output_fields[i], &field_name)); 1070 code << tab << "// ---- Output field " << i << ": " << field_name << "\n"; 1071 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_output_fields[i], &eval_mode)); 1072 // Basis action 1073 code << tab << "// EvalMode: " << CeedEvalModes[eval_mode] << "\n"; 1074 switch (eval_mode) { 1075 case CEED_EVAL_NONE: 1076 code << tab << "for (CeedInt j = 0; j < num_comp" << var_suffix << " ; j++) {\n"; 1077 tab.push(); 1078 code << tab << "r_q" << var_suffix << "[q + j*" << Q_name << "] = r_s" << var_suffix << "[j];\n"; 1079 tab.pop(); 1080 code << tab << "}\n"; 1081 break; 1082 case CEED_EVAL_INTERP: 1083 code << tab << "for (CeedInt j = 0; j < num_comp" << var_suffix << " ; j++) {\n"; 1084 tab.push(); 1085 code << tab << "r_q" << var_suffix << "[q + j*" << Q_name << "] = r_s" << var_suffix << "[j];\n"; 1086 tab.pop(); 1087 code << tab << "}\n"; 1088 break; 1089 case CEED_EVAL_GRAD: 1090 code << tab << "GradColloSliceTranspose3d<num_comp" << var_suffix << ", " << Q_name << ", OP_T_1D>(data, q, r_s" << var_suffix << ", s_G" 1091 << var_suffix << ", r_q" << var_suffix << ");\n"; 1092 break; 1093 // LCOV_EXCL_START 1094 case CEED_EVAL_WEIGHT: 1095 break; // Should not occur 1096 case CEED_EVAL_DIV: 1097 case CEED_EVAL_CURL: 1098 break; // TODO: Not implemented 1099 // LCOV_EXCL_STOP 1100 } 1101 } 1102 } 1103 tab.pop(); 1104 code << tab << "}\n"; 1105 return CEED_ERROR_SUCCESS; 1106 } 1107 1108 //------------------------------------------------------------------------------ 1109 // Build single operator kernel 1110 //------------------------------------------------------------------------------ 1111 extern "C" int CeedOperatorBuildKernel_Hip_gen(CeedOperator op, bool *is_good_build) { 1112 bool is_all_tensor = true, is_all_nontensor = true, is_at_points = false, use_3d_slices = false; 1113 Ceed ceed; 1114 CeedInt Q = 0, Q_1d = 0, num_input_fields, num_output_fields, max_dim = 1, max_num_points = 0, coords_comp_stride = 0; 1115 CeedQFunctionField *qf_input_fields, *qf_output_fields; 1116 CeedQFunction_Hip_gen *qf_data; 1117 CeedQFunction qf; 1118 CeedOperatorField *op_input_fields, *op_output_fields; 1119 CeedOperator_Hip_gen *data; 1120 std::ostringstream code; 1121 Tab tab; 1122 1123 CeedCallBackend(CeedOperatorGetData(op, &data)); 1124 { 1125 bool is_setup_done; 1126 1127 CeedCallBackend(CeedOperatorIsSetupDone(op, &is_setup_done)); 1128 if (is_setup_done) { 1129 *is_good_build = !data->use_fallback; 1130 return CEED_ERROR_SUCCESS; 1131 } 1132 } 1133 1134 // Check field compatibility 1135 CeedCallBackend(CeedOperatorGetFields(op, &num_input_fields, &op_input_fields, &num_output_fields, &op_output_fields)); 1136 { 1137 bool has_shared_bases = true; 1138 1139 for (CeedInt i = 0; i < num_input_fields; i++) { 1140 CeedBasis basis; 1141 1142 CeedCallBackend(CeedOperatorFieldGetBasis(op_input_fields[i], &basis)); 1143 if (basis != CEED_BASIS_NONE) { 1144 bool is_tensor = true; 1145 const char *resource; 1146 char *resource_root; 1147 Ceed basis_ceed; 1148 1149 CeedCallBackend(CeedBasisIsTensor(basis, &is_tensor)); 1150 is_all_tensor = is_all_tensor && is_tensor; 1151 is_all_nontensor = is_all_nontensor && !is_tensor; 1152 CeedCallBackend(CeedBasisGetCeed(basis, &basis_ceed)); 1153 CeedCallBackend(CeedGetResource(basis_ceed, &resource)); 1154 CeedCallBackend(CeedGetResourceRoot(basis_ceed, resource, ":", &resource_root)); 1155 has_shared_bases = has_shared_bases && !strcmp(resource_root, "/gpu/hip/shared"); 1156 CeedCallBackend(CeedFree(&resource_root)); 1157 CeedCallBackend(CeedDestroy(&basis_ceed)); 1158 } 1159 CeedCallBackend(CeedBasisDestroy(&basis)); 1160 } 1161 1162 for (CeedInt i = 0; i < num_output_fields; i++) { 1163 CeedBasis basis; 1164 1165 CeedCallBackend(CeedOperatorFieldGetBasis(op_output_fields[i], &basis)); 1166 if (basis != CEED_BASIS_NONE) { 1167 bool is_tensor = true; 1168 const char *resource; 1169 char *resource_root; 1170 Ceed basis_ceed; 1171 1172 CeedCallBackend(CeedBasisIsTensor(basis, &is_tensor)); 1173 is_all_tensor = is_all_tensor && is_tensor; 1174 is_all_nontensor = is_all_nontensor && !is_tensor; 1175 1176 CeedCallBackend(CeedBasisGetCeed(basis, &basis_ceed)); 1177 CeedCallBackend(CeedGetResource(basis_ceed, &resource)); 1178 CeedCallBackend(CeedGetResourceRoot(basis_ceed, resource, ":", &resource_root)); 1179 has_shared_bases = has_shared_bases && !strcmp(resource_root, "/gpu/hip/shared"); 1180 CeedCallBackend(CeedFree(&resource_root)); 1181 CeedCallBackend(CeedDestroy(&basis_ceed)); 1182 } 1183 CeedCallBackend(CeedBasisDestroy(&basis)); 1184 } 1185 // -- Fallback to ref if not all bases are shared 1186 if (!has_shared_bases) { 1187 *is_good_build = false; 1188 return CEED_ERROR_SUCCESS; 1189 } 1190 } 1191 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 1192 CeedCallBackend(CeedOperatorGetQFunction(op, &qf)); 1193 CeedCallBackend(CeedQFunctionGetData(qf, &qf_data)); 1194 CeedCallBackend(CeedQFunctionGetFields(qf, NULL, &qf_input_fields, NULL, &qf_output_fields)); 1195 1196 // Get operator data 1197 CeedCallBackend(CeedOperatorIsAtPoints(op, &is_at_points)); 1198 { 1199 CeedInt max_P = 0, max_P_1d = 0; 1200 1201 CeedCallBackend(CeedOperatorBuildKernelData_Hip_gen(ceed, num_input_fields, op_input_fields, qf_input_fields, num_output_fields, op_output_fields, 1202 qf_output_fields, &max_P, &max_P_1d, &Q, &Q_1d, &max_dim, &is_all_tensor, &use_3d_slices)); 1203 data->max_P_1d = is_all_tensor ? max_P_1d : max_P; 1204 } 1205 if (max_dim == 0) max_dim = 1; 1206 data->dim = max_dim; 1207 if (is_at_points) { 1208 CeedElemRestriction_Hip *rstr_data; 1209 CeedElemRestriction rstr_points = NULL; 1210 1211 CeedCallBackend(CeedOperatorAtPointsGetPoints(op, &rstr_points, NULL)); 1212 CeedCallBackend(CeedElemRestrictionGetMaxPointsInElement(rstr_points, &max_num_points)); 1213 CeedCallBackend(CeedElemRestrictionGetCompStride(rstr_points, &coords_comp_stride)); 1214 CeedCallBackend(CeedElemRestrictionGetData(rstr_points, &rstr_data)); 1215 data->points.indices = (CeedInt *)rstr_data->d_offsets; 1216 CeedCallBackend(CeedElemRestrictionDestroy(&rstr_points)); 1217 } 1218 if (is_at_points) use_3d_slices = false; 1219 if (Q_1d == 0) { 1220 if (is_at_points) Q_1d = max_num_points; 1221 else CeedCallBackend(CeedOperatorGetNumQuadraturePoints(op, &Q_1d)); 1222 } 1223 if (Q == 0) Q = Q_1d; 1224 data->Q = Q; 1225 data->Q_1d = Q_1d; 1226 1227 // Check for restriction only identity operator 1228 { 1229 bool is_identity_qf; 1230 1231 CeedCallBackend(CeedQFunctionIsIdentity(qf, &is_identity_qf)); 1232 if (is_identity_qf) { 1233 CeedEvalMode eval_mode_in, eval_mode_out; 1234 1235 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_input_fields[0], &eval_mode_in)); 1236 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_output_fields[0], &eval_mode_out)); 1237 CeedCheck(eval_mode_in != CEED_EVAL_NONE || eval_mode_out != CEED_EVAL_NONE, ceed, CEED_ERROR_BACKEND, 1238 "Backend does not implement restriction only identity operators"); 1239 } 1240 } 1241 1242 // Load basis source files 1243 if (!is_all_nontensor) { 1244 code << tab << "// Tensor basis source\n"; 1245 code << tab << "#include <ceed/jit-source/hip/hip-shared-basis-tensor-templates.h>\n\n"; 1246 } 1247 if (!is_all_tensor) { 1248 code << tab << "// Non-tensor basis source\n"; 1249 code << tab << "#include <ceed/jit-source/hip/hip-shared-basis-nontensor-templates.h>\n\n"; 1250 } 1251 if (is_at_points) { 1252 code << tab << "// AtPoints basis source\n"; 1253 code << tab << "#include <ceed/jit-source/hip/hip-shared-basis-tensor-at-points-templates.h>\n\n"; 1254 } 1255 if (!is_all_tensor && !is_all_nontensor) { 1256 code << tab << "// Tensor basis source\n"; 1257 code << tab << "#include <ceed/jit-source/hip/hip-shared-basis-tensor-flattened-templates.h>\n\n"; 1258 } 1259 code << tab << "// CodeGen operator source\n"; 1260 code << tab << "#include <ceed/jit-source/hip/hip-gen-templates.h>\n\n"; 1261 1262 // Get QFunction name 1263 std::string qfunction_name(qf_data->qfunction_name); 1264 std::string operator_name; 1265 1266 operator_name = "CeedKernelHipGenOperator_" + qfunction_name; 1267 1268 // Define CEED_Q_VLA 1269 code << "\n" << tab << "#undef CEED_Q_VLA\n"; 1270 if (max_dim != 3 || is_at_points || use_3d_slices || !is_all_tensor) { 1271 code << tab << "#define CEED_Q_VLA 1\n\n"; 1272 } else { 1273 code << tab << "#define CEED_Q_VLA " << Q_1d << "\n\n"; 1274 } 1275 1276 // Add user QFunction source 1277 { 1278 const char *source_path; 1279 1280 CeedCallBackend(CeedQFunctionGetSourcePath(qf, &source_path)); 1281 CeedCheck(source_path, ceed, CEED_ERROR_UNSUPPORTED, "/gpu/hip/gen backend requires QFunction source code file"); 1282 1283 code << tab << "// User QFunction source\n"; 1284 code << tab << "#include \"" << source_path << "\"\n\n"; 1285 } 1286 1287 // Setup 1288 code << "\n" << tab << "// -----------------------------------------------------------------------------\n"; 1289 code << tab << "// Operator Kernel\n"; 1290 code << tab << "// \n"; 1291 code << tab << "// d_[in,out]_i: CeedVector device array\n"; 1292 code << tab << "// r_[in,out]_e_i: Element vector register\n"; 1293 code << tab << "// r_[in,out]_q_i: Quadrature space vector register\n"; 1294 code << tab << "// r_[in,out]_c_i: AtPoints Chebyshev coefficients register\n"; 1295 code << tab << "// r_[in,out]_s_i: Quadrature space slice vector register\n"; 1296 code << tab << "// \n"; 1297 code << tab << "// s_B_[in,out]_i: Interpolation matrix, shared memory\n"; 1298 code << tab << "// s_G_[in,out]_i: Gradient matrix, shared memory\n"; 1299 code << tab << "// -----------------------------------------------------------------------------\n"; 1300 code << tab << "extern \"C\" __launch_bounds__(BLOCK_SIZE)\n"; 1301 code << "__global__ void " << operator_name 1302 << "(CeedInt num_elem, void* ctx, FieldsInt_Hip indices, Fields_Hip fields, Fields_Hip B, Fields_Hip G, CeedScalar* W, Points_Hip points) {\n"; 1303 tab.push(); 1304 1305 // Scratch buffers 1306 for (CeedInt i = 0; i < num_input_fields; i++) { 1307 CeedEvalMode eval_mode; 1308 1309 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_input_fields[i], &eval_mode)); 1310 if (eval_mode != CEED_EVAL_WEIGHT) { // Skip CEED_EVAL_WEIGHT 1311 code << tab << "const CeedScalar *__restrict__ d_in_" << i << " = fields.inputs[" << i << "];\n"; 1312 } 1313 } 1314 for (CeedInt i = 0; i < num_output_fields; i++) { 1315 code << tab << "CeedScalar *__restrict__ d_out_" << i << " = fields.outputs[" << i << "];\n"; 1316 } 1317 1318 code << tab << "const CeedInt max_dim = " << max_dim << ";\n"; 1319 if (!is_all_tensor) { 1320 code << tab << "const CeedInt Q = " << Q << ";\n"; 1321 } 1322 if (!is_all_nontensor) { 1323 code << tab << "const CeedInt Q_1d = " << Q_1d << ";\n"; 1324 } 1325 if (is_at_points) { 1326 code << tab << "const CeedInt max_num_points = " << max_num_points << ";\n"; 1327 code << tab << "const CeedInt coords_comp_stride = " << coords_comp_stride << ";\n"; 1328 } 1329 1330 // Shared data 1331 code << tab << "extern __shared__ CeedScalar slice[];\n"; 1332 code << tab << "SharedData_Hip data;\n"; 1333 code << tab << "data.t_id_x = threadIdx.x;\n"; 1334 code << tab << "data.t_id_y = threadIdx.y;\n"; 1335 code << tab << "data.t_id_z = threadIdx.z;\n"; 1336 code << tab << "data.t_id = threadIdx.x + threadIdx.y*blockDim.x + threadIdx.z*blockDim.y*blockDim.x;\n"; 1337 code << tab << "data.slice = slice + data.t_id_z*OP_T_1D" << ((!is_all_tensor || max_dim == 1) ? "" : "*OP_T_1D") << ";\n"; 1338 1339 // -- Determine input mat reuse 1340 FieldReuse_Hip input_matrix_reuse[CEED_FIELD_MAX]; 1341 1342 for (CeedInt i = 0; i < num_input_fields; i++) { 1343 input_matrix_reuse[i].index = -1; 1344 } 1345 for (CeedInt i = 0; i < num_input_fields; i++) { 1346 bool is_tensor = true; 1347 CeedEvalMode eval_mode_i; 1348 CeedBasis basis_i; 1349 1350 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_input_fields[i], &eval_mode_i)); 1351 if (eval_mode_i == CEED_EVAL_WEIGHT) continue; 1352 CeedCallBackend(CeedOperatorFieldGetBasis(op_input_fields[i], &basis_i)); 1353 CeedCallBackend(CeedBasisIsTensor(basis_i, &is_tensor)); 1354 for (CeedInt j = 0; (input_matrix_reuse[i].index == -1) && (j < i); j++) { 1355 CeedEvalMode eval_mode_j; 1356 CeedBasis basis_j; 1357 1358 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_input_fields[j], &eval_mode_j)); 1359 if (eval_mode_j == CEED_EVAL_WEIGHT) continue; 1360 CeedCallBackend(CeedOperatorFieldGetBasis(op_input_fields[j], &basis_j)); 1361 if (basis_i == basis_j) { 1362 if (is_tensor) { 1363 input_matrix_reuse[i].index = j; 1364 input_matrix_reuse[i].is_input = true; 1365 input_matrix_reuse[i].eval_mode = eval_mode_j; 1366 } else { 1367 // For non-tensor can only re-use with the same eval mode 1368 if (eval_mode_i == eval_mode_j) { 1369 input_matrix_reuse[i].index = j; 1370 input_matrix_reuse[i].is_input = true; 1371 input_matrix_reuse[i].eval_mode = eval_mode_j; 1372 } 1373 } 1374 } 1375 CeedCallBackend(CeedBasisDestroy(&basis_j)); 1376 } 1377 CeedCallBackend(CeedBasisDestroy(&basis_i)); 1378 } 1379 1380 // -- Determine output mat reuse 1381 FieldReuse_Hip output_matrix_reuse[CEED_FIELD_MAX]; 1382 1383 for (CeedInt i = 0; i < num_output_fields; i++) { 1384 output_matrix_reuse[i].index = -1; 1385 } 1386 for (CeedInt i = 0; i < num_output_fields; i++) { 1387 bool is_tensor = true; 1388 CeedEvalMode eval_mode_i; 1389 CeedBasis basis_i; 1390 1391 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_output_fields[i], &eval_mode_i)); 1392 CeedCallBackend(CeedOperatorFieldGetBasis(op_output_fields[i], &basis_i)); 1393 for (CeedInt j = 0; (output_matrix_reuse[i].index == -1) && (j < num_input_fields); j++) { 1394 CeedEvalMode eval_mode_j; 1395 CeedBasis basis_j; 1396 1397 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_input_fields[j], &eval_mode_j)); 1398 if (eval_mode_j == CEED_EVAL_WEIGHT) continue; 1399 CeedCallBackend(CeedOperatorFieldGetBasis(op_input_fields[j], &basis_j)); 1400 if (basis_i == basis_j) { 1401 if (is_tensor) { 1402 output_matrix_reuse[i].index = j; 1403 output_matrix_reuse[i].is_input = true; 1404 output_matrix_reuse[i].eval_mode = eval_mode_j; 1405 } else { 1406 // For non-tensor can only re-use with the same eval mode 1407 if (eval_mode_i == eval_mode_j) { 1408 output_matrix_reuse[i].index = j; 1409 output_matrix_reuse[i].is_input = true; 1410 output_matrix_reuse[i].eval_mode = eval_mode_j; 1411 } 1412 } 1413 } 1414 CeedCallBackend(CeedBasisDestroy(&basis_j)); 1415 } 1416 for (CeedInt j = 0; (output_matrix_reuse[i].index == -1) && (j < i); j++) { 1417 CeedEvalMode eval_mode_j; 1418 CeedBasis basis_j; 1419 1420 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_output_fields[j], &eval_mode_j)); 1421 if (eval_mode_j == CEED_EVAL_WEIGHT) continue; 1422 CeedCallBackend(CeedOperatorFieldGetBasis(op_output_fields[j], &basis_j)); 1423 CeedCallBackend(CeedBasisIsTensor(basis_i, &is_tensor)); 1424 if (basis_i == basis_j) { 1425 if (is_tensor) { 1426 output_matrix_reuse[i].index = j; 1427 output_matrix_reuse[i].is_input = false; 1428 output_matrix_reuse[i].eval_mode = eval_mode_j; 1429 } else { 1430 // For non-tensor can only re-use with the same eval mode 1431 if (eval_mode_i == eval_mode_j) { 1432 output_matrix_reuse[i].index = j; 1433 output_matrix_reuse[i].is_input = false; 1434 output_matrix_reuse[i].eval_mode = eval_mode_j; 1435 } 1436 } 1437 } 1438 CeedCallBackend(CeedBasisDestroy(&basis_j)); 1439 } 1440 CeedCallBackend(CeedBasisDestroy(&basis_i)); 1441 } 1442 1443 // Initialize constants, and matrices B and G 1444 code << "\n" << tab << "// Input field constants and basis data\n"; 1445 for (CeedInt i = 0; i < num_input_fields; i++) { 1446 CeedCallBackend(CeedOperatorBuildKernelFieldData_Hip_gen(code, data, tab, i, op_input_fields[i], qf_input_fields[i], input_matrix_reuse[i], 1447 max_dim, Q, Q_1d, true, is_all_tensor, is_at_points, use_3d_slices)); 1448 } 1449 code << "\n" << tab << "// Output field constants and basis data\n"; 1450 for (CeedInt i = 0; i < num_output_fields; i++) { 1451 CeedCallBackend(CeedOperatorBuildKernelFieldData_Hip_gen(code, data, tab, i, op_output_fields[i], qf_output_fields[i], output_matrix_reuse[i], 1452 max_dim, Q, Q_1d, false, is_all_tensor, is_at_points, use_3d_slices)); 1453 } 1454 1455 // Loop over all elements 1456 code << "\n" << tab << "// Element loop\n"; 1457 code << tab << "__syncthreads();\n"; 1458 code << tab << "for (CeedInt elem = blockIdx.x*blockDim.z + threadIdx.z; elem < num_elem; elem += gridDim.x*blockDim.z) {\n"; 1459 tab.push(); 1460 1461 // -- Compute minimum buffer space needed 1462 CeedInt max_rstr_buffer_size = 1; 1463 1464 for (CeedInt i = 0; i < num_input_fields; i++) { 1465 CeedEvalMode eval_mode; 1466 1467 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_input_fields[i], &eval_mode)); 1468 if (eval_mode != CEED_EVAL_NONE && eval_mode != CEED_EVAL_WEIGHT) { 1469 CeedInt num_comp; 1470 CeedElemRestriction elem_rstr; 1471 1472 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_input_fields[i], &elem_rstr)); 1473 CeedCallBackend(CeedElemRestrictionGetNumComponents(elem_rstr, &num_comp)); 1474 max_rstr_buffer_size = CeedIntMax(max_rstr_buffer_size, num_comp * (is_all_tensor && (max_dim >= 3) ? Q_1d : 1)); 1475 CeedCallBackend(CeedElemRestrictionDestroy(&elem_rstr)); 1476 } 1477 } 1478 for (CeedInt i = 0; i < num_output_fields; i++) { 1479 CeedEvalMode eval_mode; 1480 1481 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_output_fields[i], &eval_mode)); 1482 if (eval_mode != CEED_EVAL_NONE) { 1483 CeedInt num_comp; 1484 CeedElemRestriction elem_rstr; 1485 1486 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_output_fields[i], &elem_rstr)); 1487 CeedCallBackend(CeedElemRestrictionGetNumComponents(elem_rstr, &num_comp)); 1488 max_rstr_buffer_size = CeedIntMax(max_rstr_buffer_size, num_comp * (is_all_tensor && (max_dim >= 3) ? Q_1d : 1)); 1489 CeedCallBackend(CeedElemRestrictionDestroy(&elem_rstr)); 1490 } 1491 } 1492 code << tab << "// Scratch restriction buffer space\n"; 1493 code << tab << "CeedScalar r_e_scratch[" << max_rstr_buffer_size << "];\n"; 1494 1495 // -- Determine best input field processing order 1496 CeedInt field_rstr_in_buffer[CEED_FIELD_MAX], input_field_order[CEED_FIELD_MAX]; 1497 1498 for (CeedInt i = 0; i < num_input_fields; i++) { 1499 field_rstr_in_buffer[i] = -1; 1500 input_field_order[i] = -1; 1501 } 1502 { 1503 bool is_ordered[CEED_FIELD_MAX]; 1504 CeedInt curr_index = 0; 1505 1506 for (CeedInt i = 0; i < num_input_fields; i++) is_ordered[i] = false; 1507 for (CeedInt i = 0; i < num_input_fields; i++) { 1508 CeedVector vec_i; 1509 CeedElemRestriction rstr_i; 1510 1511 if (is_ordered[i]) continue; 1512 field_rstr_in_buffer[i] = i; 1513 is_ordered[i] = true; 1514 input_field_order[curr_index] = i; 1515 curr_index++; 1516 CeedCallBackend(CeedOperatorFieldGetVector(op_input_fields[i], &vec_i)); 1517 if (vec_i == CEED_VECTOR_NONE) continue; // CEED_EVAL_WEIGHT 1518 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_input_fields[i], &rstr_i)); 1519 for (CeedInt j = i + 1; j < num_input_fields; j++) { 1520 CeedVector vec_j; 1521 CeedElemRestriction rstr_j; 1522 1523 CeedCallBackend(CeedOperatorFieldGetVector(op_input_fields[j], &vec_j)); 1524 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_input_fields[j], &rstr_j)); 1525 if (rstr_i == rstr_j && vec_i == vec_j) { 1526 field_rstr_in_buffer[j] = i; 1527 is_ordered[j] = true; 1528 input_field_order[curr_index] = j; 1529 curr_index++; 1530 } 1531 CeedCallBackend(CeedVectorDestroy(&vec_j)); 1532 CeedCallBackend(CeedElemRestrictionDestroy(&rstr_j)); 1533 } 1534 CeedCallBackend(CeedVectorDestroy(&vec_i)); 1535 CeedCallBackend(CeedElemRestrictionDestroy(&rstr_i)); 1536 } 1537 } 1538 1539 // -- Input restriction and basis 1540 code << "\n" << tab << "// -- Input field restrictions and basis actions\n"; 1541 for (CeedInt i = 0; i < num_input_fields; i++) { 1542 const char *field_name; 1543 const CeedInt f = input_field_order[i]; 1544 1545 CeedCallBackend(CeedOperatorFieldGetName(op_input_fields[f], &field_name)); 1546 code << tab << "// ---- Input field " << f << ": " << field_name << "\n"; 1547 1548 // ---- Restriction 1549 CeedCallBackend(CeedOperatorBuildKernelRestriction_Hip_gen(code, data, tab, f, field_rstr_in_buffer, op_input_fields[f], qf_input_fields[f], 1550 max_dim, Q_1d, true, is_all_tensor, is_at_points, use_3d_slices)); 1551 1552 // ---- Basis action 1553 CeedCallBackend(CeedOperatorBuildKernelBasis_Hip_gen(code, data, tab, f, op_input_fields[f], qf_input_fields[f], max_dim, Q_1d, true, 1554 is_all_tensor, is_at_points, use_3d_slices)); 1555 } 1556 1557 // -- Q function 1558 CeedCallBackend(CeedOperatorBuildKernelQFunction_Hip_gen(code, data, tab, max_dim, max_num_points, num_input_fields, op_input_fields, 1559 qf_input_fields, num_output_fields, op_output_fields, qf_output_fields, qfunction_name, 1560 Q_1d, is_all_tensor, is_at_points, use_3d_slices)); 1561 1562 // -- Output basis and restriction 1563 code << "\n" << tab << "// -- Output field basis action and restrictions\n"; 1564 for (CeedInt i = 0; i < num_output_fields; i++) { 1565 const char *field_name; 1566 1567 CeedCallBackend(CeedOperatorFieldGetName(op_output_fields[i], &field_name)); 1568 code << tab << "// ---- Output field " << i << ": " << field_name << "\n"; 1569 1570 // ---- Basis action 1571 CeedCallBackend(CeedOperatorBuildKernelBasis_Hip_gen(code, data, tab, i, op_output_fields[i], qf_output_fields[i], max_dim, Q_1d, false, 1572 is_all_tensor, is_at_points, use_3d_slices)); 1573 1574 // ---- Restriction 1575 CeedCallBackend(CeedOperatorBuildKernelRestriction_Hip_gen(code, data, tab, i, NULL, op_output_fields[i], qf_output_fields[i], max_dim, Q_1d, 1576 false, is_all_tensor, is_at_points, use_3d_slices)); 1577 } 1578 1579 // Close loop and function 1580 tab.pop(); 1581 code << tab << "}\n"; 1582 tab.pop(); 1583 code << tab << "}\n"; 1584 code << tab << "// -----------------------------------------------------------------------------\n\n"; 1585 1586 CeedInt block_sizes[3] = {0, 0, 0}; 1587 CeedInt num_elem; 1588 1589 // Compile 1590 CeedCallBackend(CeedOperatorGetNumElements(op, &num_elem)); 1591 CeedCallBackend(BlockGridCalculate_Hip_gen(is_all_tensor ? max_dim : 1, num_elem, data->max_P_1d, is_all_tensor ? Q_1d : Q, block_sizes)); 1592 if (is_at_points) block_sizes[2] = 1; 1593 { 1594 bool is_compile_good = false; 1595 1596 data->thread_1d = block_sizes[0]; 1597 CeedCallBackend(CeedTryCompile_Hip(ceed, code.str().c_str(), &is_compile_good, &data->module, 2, "OP_T_1D", block_sizes[0], "BLOCK_SIZE", 1598 block_sizes[0] * block_sizes[1] * block_sizes[2])); 1599 if (is_compile_good) { 1600 *is_good_build = true; 1601 CeedCallBackend(CeedGetKernel_Hip(ceed, data->module, operator_name.c_str(), &data->op)); 1602 } else { 1603 *is_good_build = false; 1604 data->use_fallback = true; 1605 } 1606 } 1607 CeedCallBackend(CeedOperatorSetSetupDone(op)); 1608 CeedCallBackend(CeedDestroy(&ceed)); 1609 CeedCallBackend(CeedQFunctionDestroy(&qf)); 1610 return CEED_ERROR_SUCCESS; 1611 } 1612 1613 //------------------------------------------------------------------------------ 1614 // Build AtPoints assembly operator kernel 1615 //------------------------------------------------------------------------------ 1616 static int CeedOperatorBuildKernelAssemblyAtPoints_Hip_gen(CeedOperator op, bool is_full, bool *is_good_build) { 1617 bool is_all_tensor = true, is_at_points = false, use_3d_slices = false; 1618 Ceed ceed; 1619 CeedInt Q, Q_1d, num_input_fields, num_output_fields, max_dim = 1, max_num_points = 0, coords_comp_stride = 0; 1620 CeedQFunctionField *qf_input_fields, *qf_output_fields; 1621 CeedQFunction_Hip_gen *qf_data; 1622 CeedQFunction qf; 1623 CeedOperatorField *op_input_fields, *op_output_fields; 1624 CeedOperator_Hip_gen *data; 1625 std::ostringstream code; 1626 Tab tab; 1627 1628 // Check compatibility 1629 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 1630 CeedCallBackend(CeedOperatorIsAtPoints(op, &is_at_points)); 1631 CeedCheck(is_at_points, ceed, CEED_ERROR_BACKEND, "Only AtPoints operator assembly supported"); 1632 1633 // Retrieve operator data 1634 CeedCallBackend(CeedOperatorGetData(op, &data)); 1635 Q = data->Q; 1636 Q_1d = data->Q_1d; 1637 max_dim = data->dim; 1638 { 1639 CeedElemRestriction rstr_points = NULL; 1640 1641 CeedCallBackend(CeedOperatorAtPointsGetPoints(op, &rstr_points, NULL)); 1642 CeedCallBackend(CeedElemRestrictionGetMaxPointsInElement(rstr_points, &max_num_points)); 1643 CeedCallBackend(CeedElemRestrictionGetCompStride(rstr_points, &coords_comp_stride)); 1644 CeedCallBackend(CeedElemRestrictionDestroy(&rstr_points)); 1645 } 1646 CeedCallBackend(CeedOperatorGetQFunction(op, &qf)); 1647 CeedCallBackend(CeedQFunctionGetData(qf, &qf_data)); 1648 CeedCallBackend(CeedQFunctionGetFields(qf, NULL, &qf_input_fields, NULL, &qf_output_fields)); 1649 CeedCallBackend(CeedOperatorGetFields(op, &num_input_fields, &op_input_fields, &num_output_fields, &op_output_fields)); 1650 1651 // Load basis source files 1652 code << tab << "// Tensor basis source\n"; 1653 code << tab << "#include <ceed/jit-source/hip/hip-shared-basis-tensor-templates.h>\n\n"; 1654 code << tab << "// AtPoints basis source\n"; 1655 code << tab << "#include <ceed/jit-source/hip/hip-shared-basis-tensor-at-points-templates.h>\n\n"; 1656 code << tab << "// CodeGen operator source\n"; 1657 code << tab << "#include <ceed/jit-source/hip/hip-gen-templates.h>\n\n"; 1658 1659 // Get QFunction name 1660 std::string qfunction_name(qf_data->qfunction_name); 1661 std::string operator_name; 1662 1663 if (is_full) { 1664 operator_name = "CeedKernelHipGenOperatorFullAssembly_" + qfunction_name; 1665 } else { 1666 operator_name = "CeedKernelHipGenOperatorDiagonalAssembly_" + qfunction_name; 1667 } 1668 1669 // Define CEED_Q_VLA 1670 code << "\n" << tab << "#undef CEED_Q_VLA\n"; 1671 code << tab << "#define CEED_Q_VLA 1\n\n"; 1672 1673 // Add user QFunction source 1674 { 1675 const char *source_path; 1676 1677 CeedCallBackend(CeedQFunctionGetSourcePath(qf, &source_path)); 1678 CeedCheck(source_path, ceed, CEED_ERROR_UNSUPPORTED, "/gpu/hip/gen backend requires QFunction source code file"); 1679 1680 code << tab << "// User QFunction source\n"; 1681 code << tab << "#include \"" << source_path << "\"\n\n"; 1682 } 1683 1684 // Setup 1685 code << "\n" << tab << "// -----------------------------------------------------------------------------\n"; 1686 code << tab << "// Operator Assembly Kernel\n"; 1687 code << tab << "// \n"; 1688 code << tab << "// d_[in,out]_i: CeedVector device array\n"; 1689 code << tab << "// r_[in,out]_e_i: Element vector register\n"; 1690 code << tab << "// r_[in,out]_q_i: Quadrature space vector register\n"; 1691 code << tab << "// r_[in,out]_c_i: AtPoints Chebyshev coefficients register\n"; 1692 code << tab << "// r_[in,out]_s_i: Quadrature space slice vector register\n"; 1693 code << tab << "// \n"; 1694 code << tab << "// s_B_[in,out]_i: Interpolation matrix, shared memory\n"; 1695 code << tab << "// s_G_[in,out]_i: Gradient matrix, shared memory\n"; 1696 code << tab << "// -----------------------------------------------------------------------------\n"; 1697 code << tab << "extern \"C\" __global__ void " << operator_name 1698 << "(CeedInt num_elem, void* ctx, FieldsInt_Hip indices, Fields_Hip fields, Fields_Hip B, Fields_Hip G, CeedScalar *W, Points_Hip " 1699 "points, CeedScalar *__restrict__ values_array) {\n"; 1700 tab.push(); 1701 1702 // Scratch buffers 1703 for (CeedInt i = 0; i < num_input_fields; i++) { 1704 CeedEvalMode eval_mode; 1705 1706 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_input_fields[i], &eval_mode)); 1707 if (eval_mode != CEED_EVAL_WEIGHT) { // Skip CEED_EVAL_WEIGHT 1708 code << tab << "const CeedScalar *__restrict__ d_in_" << i << " = fields.inputs[" << i << "];\n"; 1709 } 1710 } 1711 for (CeedInt i = 0; i < num_output_fields; i++) { 1712 code << tab << "CeedScalar *__restrict__ d_out_" << i << " = fields.outputs[" << i << "];\n"; 1713 } 1714 1715 code << tab << "const CeedInt max_dim = " << max_dim << ";\n"; 1716 code << tab << "const CeedInt Q_1d = " << Q_1d << ";\n"; 1717 code << tab << "const CeedInt max_num_points = " << max_num_points << ";\n"; 1718 code << tab << "const CeedInt coords_comp_stride = " << coords_comp_stride << ";\n"; 1719 1720 // Shared data 1721 code << tab << "extern __shared__ CeedScalar slice[];\n"; 1722 code << tab << "SharedData_Hip data;\n"; 1723 code << tab << "data.t_id_x = threadIdx.x;\n"; 1724 code << tab << "data.t_id_y = threadIdx.y;\n"; 1725 code << tab << "data.t_id_z = threadIdx.z;\n"; 1726 code << tab << "data.t_id = threadIdx.x + threadIdx.y*blockDim.x + threadIdx.z*blockDim.y*blockDim.x;\n"; 1727 code << tab << "data.slice = slice + data.t_id_z*OP_T_1D" << ((!is_all_tensor || max_dim == 1) ? "" : "*OP_T_1D") << ";\n"; 1728 1729 // -- Determine input mat reuse 1730 FieldReuse_Hip input_matrix_reuse[CEED_FIELD_MAX]; 1731 1732 for (CeedInt i = 0; i < num_input_fields; i++) { 1733 input_matrix_reuse[i].index = -1; 1734 } 1735 for (CeedInt i = 0; i < num_input_fields; i++) { 1736 CeedEvalMode eval_mode_i; 1737 CeedBasis basis_i; 1738 1739 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_input_fields[i], &eval_mode_i)); 1740 if (eval_mode_i == CEED_EVAL_WEIGHT) continue; 1741 CeedCallBackend(CeedOperatorFieldGetBasis(op_input_fields[i], &basis_i)); 1742 for (CeedInt j = 0; (input_matrix_reuse[i].index == -1) && (j < i); j++) { 1743 CeedEvalMode eval_mode_j; 1744 CeedBasis basis_j; 1745 1746 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_input_fields[j], &eval_mode_j)); 1747 if (eval_mode_j == CEED_EVAL_WEIGHT) continue; 1748 CeedCallBackend(CeedOperatorFieldGetBasis(op_input_fields[j], &basis_j)); 1749 if (basis_i == basis_j) { 1750 input_matrix_reuse[i].index = j; 1751 input_matrix_reuse[i].is_input = true; 1752 input_matrix_reuse[i].eval_mode = eval_mode_j; 1753 } 1754 CeedCallBackend(CeedBasisDestroy(&basis_j)); 1755 } 1756 CeedCallBackend(CeedBasisDestroy(&basis_i)); 1757 } 1758 1759 // -- Determine output mat reuse 1760 FieldReuse_Hip output_matrix_reuse[CEED_FIELD_MAX]; 1761 1762 for (CeedInt i = 0; i < num_output_fields; i++) { 1763 output_matrix_reuse[i].index = -1; 1764 } 1765 for (CeedInt i = 0; i < num_output_fields; i++) { 1766 CeedEvalMode eval_mode_i; 1767 CeedBasis basis_i; 1768 1769 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_output_fields[i], &eval_mode_i)); 1770 CeedCallBackend(CeedOperatorFieldGetBasis(op_output_fields[i], &basis_i)); 1771 for (CeedInt j = 0; (output_matrix_reuse[i].index == -1) && (j < num_input_fields); j++) { 1772 CeedEvalMode eval_mode_j; 1773 CeedBasis basis_j; 1774 1775 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_input_fields[j], &eval_mode_j)); 1776 if (eval_mode_j == CEED_EVAL_WEIGHT) continue; 1777 CeedCallBackend(CeedOperatorFieldGetBasis(op_input_fields[j], &basis_j)); 1778 if (basis_i == basis_j) { 1779 output_matrix_reuse[i].index = j; 1780 output_matrix_reuse[i].is_input = true; 1781 output_matrix_reuse[i].eval_mode = eval_mode_j; 1782 } 1783 CeedCallBackend(CeedBasisDestroy(&basis_j)); 1784 } 1785 for (CeedInt j = 0; (output_matrix_reuse[i].index == -1) && (j < i); j++) { 1786 CeedEvalMode eval_mode_j; 1787 CeedBasis basis_j; 1788 1789 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_output_fields[j], &eval_mode_j)); 1790 if (eval_mode_j == CEED_EVAL_WEIGHT) continue; 1791 CeedCallBackend(CeedOperatorFieldGetBasis(op_output_fields[j], &basis_j)); 1792 if (basis_i == basis_j) { 1793 output_matrix_reuse[i].index = j; 1794 output_matrix_reuse[i].is_input = false; 1795 output_matrix_reuse[i].eval_mode = eval_mode_j; 1796 } 1797 CeedCallBackend(CeedBasisDestroy(&basis_j)); 1798 } 1799 CeedCallBackend(CeedBasisDestroy(&basis_i)); 1800 } 1801 1802 // Initialize constants, and matrices B and G 1803 code << "\n" << tab << "// Input field constants and basis data\n"; 1804 for (CeedInt i = 0; i < num_input_fields; i++) { 1805 CeedCallBackend(CeedOperatorBuildKernelFieldData_Hip_gen(code, data, tab, i, op_input_fields[i], qf_input_fields[i], input_matrix_reuse[i], 1806 max_dim, Q, Q_1d, true, is_all_tensor, is_at_points, use_3d_slices)); 1807 } 1808 code << "\n" << tab << "// Output field constants and basis data\n"; 1809 for (CeedInt i = 0; i < num_output_fields; i++) { 1810 CeedCallBackend(CeedOperatorBuildKernelFieldData_Hip_gen(code, data, tab, i, op_output_fields[i], qf_output_fields[i], output_matrix_reuse[i], 1811 max_dim, Q, Q_1d, false, is_all_tensor, is_at_points, use_3d_slices)); 1812 } 1813 1814 // Loop over all elements 1815 code << "\n" << tab << "// Element loop\n"; 1816 code << tab << "__syncthreads();\n"; 1817 code << tab << "for (CeedInt elem = blockIdx.x*blockDim.z + threadIdx.z; elem < num_elem; elem += gridDim.x*blockDim.z) {\n"; 1818 tab.push(); 1819 1820 // -- Compute minimum buffer space needed 1821 CeedInt max_rstr_buffer_size = 1; 1822 1823 for (CeedInt i = 0; i < num_input_fields; i++) { 1824 CeedEvalMode eval_mode; 1825 1826 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_input_fields[i], &eval_mode)); 1827 if (eval_mode != CEED_EVAL_NONE && eval_mode != CEED_EVAL_WEIGHT) { 1828 CeedInt num_comp; 1829 CeedElemRestriction elem_rstr; 1830 1831 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_input_fields[i], &elem_rstr)); 1832 CeedCallBackend(CeedElemRestrictionGetNumComponents(elem_rstr, &num_comp)); 1833 max_rstr_buffer_size = CeedIntMax(max_rstr_buffer_size, num_comp * (is_all_tensor && (max_dim >= 3) ? Q_1d : 1)); 1834 CeedCallBackend(CeedElemRestrictionDestroy(&elem_rstr)); 1835 } 1836 } 1837 for (CeedInt i = 0; i < num_output_fields; i++) { 1838 CeedEvalMode eval_mode; 1839 1840 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_output_fields[i], &eval_mode)); 1841 if (eval_mode != CEED_EVAL_NONE) { 1842 CeedInt num_comp; 1843 CeedElemRestriction elem_rstr; 1844 1845 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_output_fields[i], &elem_rstr)); 1846 CeedCallBackend(CeedElemRestrictionGetNumComponents(elem_rstr, &num_comp)); 1847 max_rstr_buffer_size = CeedIntMax(max_rstr_buffer_size, num_comp * (is_all_tensor && (max_dim >= 3) ? Q_1d : 1)); 1848 CeedCallBackend(CeedElemRestrictionDestroy(&elem_rstr)); 1849 } 1850 } 1851 code << tab << "// Scratch restriction buffer space\n"; 1852 code << tab << "CeedScalar r_e_scratch[" << max_rstr_buffer_size << "];\n"; 1853 1854 // -- Determine best input field processing order 1855 CeedInt field_rstr_in_buffer[CEED_FIELD_MAX], input_field_order[CEED_FIELD_MAX]; 1856 1857 for (CeedInt i = 0; i < num_input_fields; i++) { 1858 field_rstr_in_buffer[i] = -1; 1859 input_field_order[i] = -1; 1860 } 1861 { 1862 bool is_ordered[CEED_FIELD_MAX]; 1863 CeedInt curr_index = 0; 1864 1865 for (CeedInt i = 0; i < num_input_fields; i++) is_ordered[i] = false; 1866 for (CeedInt i = 0; i < num_input_fields; i++) { 1867 CeedVector vec_i; 1868 CeedElemRestriction rstr_i; 1869 1870 if (is_ordered[i]) continue; 1871 field_rstr_in_buffer[i] = i; 1872 is_ordered[i] = true; 1873 input_field_order[curr_index] = i; 1874 curr_index++; 1875 CeedCallBackend(CeedOperatorFieldGetVector(op_input_fields[i], &vec_i)); 1876 if (vec_i == CEED_VECTOR_NONE) continue; // CEED_EVAL_WEIGHT 1877 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_input_fields[i], &rstr_i)); 1878 for (CeedInt j = i + 1; j < num_input_fields; j++) { 1879 CeedVector vec_j; 1880 CeedElemRestriction rstr_j; 1881 1882 CeedCallBackend(CeedOperatorFieldGetVector(op_input_fields[j], &vec_j)); 1883 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_input_fields[j], &rstr_j)); 1884 if (rstr_i == rstr_j && vec_i == vec_j) { 1885 field_rstr_in_buffer[j] = i; 1886 is_ordered[j] = true; 1887 input_field_order[curr_index] = j; 1888 curr_index++; 1889 } 1890 CeedCallBackend(CeedVectorDestroy(&vec_j)); 1891 CeedCallBackend(CeedElemRestrictionDestroy(&rstr_j)); 1892 } 1893 CeedCallBackend(CeedVectorDestroy(&vec_i)); 1894 CeedCallBackend(CeedElemRestrictionDestroy(&rstr_i)); 1895 } 1896 } 1897 1898 // -- Input restriction and basis 1899 code << "\n" << tab << "// -- Input field restrictions and basis actions\n"; 1900 CeedInt active_field_index = -1; 1901 1902 for (CeedInt i = 0; i < num_input_fields; i++) { 1903 bool is_active = false; 1904 const char *field_name; 1905 const CeedInt f = input_field_order[i]; 1906 1907 { 1908 CeedVector vec; 1909 1910 CeedCallBackend(CeedOperatorFieldGetVector(op_input_fields[f], &vec)); 1911 is_active = vec == CEED_VECTOR_ACTIVE; 1912 CeedCallBackend(CeedVectorDestroy(&vec)); 1913 } 1914 1915 CeedCallBackend(CeedOperatorFieldGetName(op_input_fields[f], &field_name)); 1916 code << tab << "// ---- Input field " << f << ": " << field_name << "\n"; 1917 1918 if (is_active) { 1919 std::string var_suffix = "_in_" + std::to_string(f); 1920 1921 code << tab << "// Active field - no restriction or basis action here\n"; 1922 if (active_field_index == -1) { 1923 active_field_index = f; 1924 code << tab << "CeedScalar r_e" << var_suffix << "[num_comp" << var_suffix << "*" << (max_dim >= 3 ? "P_1d" + var_suffix : "1") 1925 << "] = {0.0};\n"; 1926 } else { 1927 code << tab << "CeedScalar *r_e" << var_suffix << " = r_e_in_" << active_field_index << ";\n"; 1928 } 1929 } else { 1930 // ---- Restriction 1931 CeedCallBackend(CeedOperatorBuildKernelRestriction_Hip_gen(code, data, tab, f, field_rstr_in_buffer, op_input_fields[f], qf_input_fields[f], 1932 max_dim, Q_1d, true, is_all_tensor, is_at_points, use_3d_slices)); 1933 1934 // ---- Basis action 1935 CeedCallBackend(CeedOperatorBuildKernelBasis_Hip_gen(code, data, tab, f, op_input_fields[f], qf_input_fields[f], max_dim, Q_1d, true, 1936 is_all_tensor, is_at_points, use_3d_slices)); 1937 } 1938 } 1939 1940 // -- Loop over active field 1941 std::string active_var_suffix = "_in_" + std::to_string(active_field_index); 1942 1943 code << "\n" << tab << "// Loop over nodes in active field\n"; 1944 code << tab << "for (CeedInt n = 0; n < num_comp" << active_var_suffix << "*P_1d" << active_var_suffix 1945 << (max_dim > 1 ? "*P_1d" + active_var_suffix : "") << (max_dim > 2 ? "*P_1d" + active_var_suffix : "") << "; n++) {\n"; 1946 tab.push(); 1947 1948 // -- Set current active node and component to 1 1949 code << tab << "// Set current active node and component to 1.0\n"; 1950 code << tab << "SetEVecStandard" << max_dim << "d_Single<num_comp" << active_var_suffix << ", P_1d" << active_var_suffix << ">(data, n, 1.0, r_e" 1951 << active_var_suffix << ");\n\n"; 1952 1953 for (CeedInt i = 0; i < num_input_fields; i++) { 1954 bool is_active = false; 1955 const char *field_name; 1956 const CeedInt f = input_field_order[i]; 1957 1958 { 1959 CeedVector vec; 1960 1961 CeedCallBackend(CeedOperatorFieldGetVector(op_input_fields[f], &vec)); 1962 is_active = vec == CEED_VECTOR_ACTIVE; 1963 CeedCallBackend(CeedVectorDestroy(&vec)); 1964 } 1965 if (!is_active) continue; 1966 1967 CeedCallBackend(CeedOperatorFieldGetName(op_input_fields[f], &field_name)); 1968 code << tab << "// ---- Input field " << f << ": " << field_name << "\n"; 1969 1970 // ---- Basis action 1971 CeedCallBackend(CeedOperatorBuildKernelBasis_Hip_gen(code, data, tab, f, op_input_fields[f], qf_input_fields[f], max_dim, Q_1d, true, 1972 is_all_tensor, is_at_points, use_3d_slices)); 1973 } 1974 1975 // -- Q function 1976 CeedCallBackend(CeedOperatorBuildKernelQFunction_Hip_gen(code, data, tab, max_dim, max_num_points, num_input_fields, op_input_fields, 1977 qf_input_fields, num_output_fields, op_output_fields, qf_output_fields, qfunction_name, 1978 Q_1d, is_all_tensor, is_at_points, use_3d_slices)); 1979 1980 // -- Output basis and restriction 1981 code << "\n" << tab << "// -- Output field basis action and restrictions\n"; 1982 for (CeedInt i = 0; i < num_output_fields; i++) { 1983 bool is_active = false; 1984 const char *field_name; 1985 1986 { 1987 CeedVector vec; 1988 1989 CeedCallBackend(CeedOperatorFieldGetVector(op_output_fields[i], &vec)); 1990 is_active = vec == CEED_VECTOR_ACTIVE; 1991 CeedCallBackend(CeedVectorDestroy(&vec)); 1992 } 1993 if (!is_active) continue; 1994 1995 CeedCallBackend(CeedOperatorFieldGetName(op_output_fields[i], &field_name)); 1996 code << tab << "// ---- Output field " << i << ": " << field_name << "\n"; 1997 1998 // ---- Basis action 1999 CeedCallBackend(CeedOperatorBuildKernelBasis_Hip_gen(code, data, tab, i, op_output_fields[i], qf_output_fields[i], max_dim, Q_1d, false, 2000 is_all_tensor, is_at_points, use_3d_slices)); 2001 2002 // ---- Restriction 2003 if (is_full) { 2004 // TODO: UPDATE OUTPUTS FOR FULL ASSEMBLY 2005 } else { 2006 std::string var_suffix = "_out_" + std::to_string(i); 2007 CeedInt comp_stride; 2008 CeedSize l_size; 2009 CeedElemRestriction elem_rstr; 2010 2011 CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_output_fields[i], &elem_rstr)); 2012 CeedCallBackend(CeedElemRestrictionGetLVectorSize(elem_rstr, &l_size)); 2013 code << tab << "const CeedInt l_size" << var_suffix << " = " << l_size << ";\n"; 2014 CeedCallBackend(CeedElemRestrictionGetCompStride(elem_rstr, &comp_stride)); 2015 code << tab << "const CeedInt comp_stride" << var_suffix << " = " << comp_stride << ";\n"; 2016 code << tab << "WriteLVecStandard" << max_dim << "d_Single<num_comp" << var_suffix << ", comp_stride" << var_suffix << ", P_1d" + var_suffix 2017 << ">(data, l_size" << var_suffix << ", elem, n, indices.outputs[" << i << "], r_e" << var_suffix << ", values_array);\n"; 2018 CeedCallBackend(CeedElemRestrictionDestroy(&elem_rstr)); 2019 } 2020 } 2021 2022 // -- Reset current active node and component 2023 code << "\n" << tab << "// Reset current active node and component to 0.0\n"; 2024 code << tab << "SetEVecStandard" << max_dim << "d_Single<num_comp" << active_var_suffix << ", P_1d" << active_var_suffix << ">(data, n, 0.0, r_e" 2025 << active_var_suffix << ");\n"; 2026 2027 // -- End of loop over active field 2028 tab.pop(); 2029 code << tab << "}\n"; 2030 2031 // Close loop and function 2032 tab.pop(); 2033 code << tab << "}\n"; 2034 tab.pop(); 2035 code << tab << "}\n"; 2036 code << tab << "// -----------------------------------------------------------------------------\n\n"; 2037 2038 CeedInt block_sizes[3] = {0, 0, 0}; 2039 CeedInt num_elem; 2040 2041 // Compile 2042 CeedCallBackend(CeedOperatorGetNumElements(op, &num_elem)); 2043 CeedCallBackend(BlockGridCalculate_Hip_gen(max_dim, num_elem, data->max_P_1d, Q_1d, block_sizes)); 2044 block_sizes[2] = 1; 2045 { 2046 bool is_compile_good = false; 2047 2048 data->thread_1d = block_sizes[0]; 2049 CeedCallBackend(CeedTryCompile_Hip(ceed, code.str().c_str(), &is_compile_good, 2050 is_full ? &data->module_assemble_full : &data->module_assemble_diagonal, 2, "OP_T_1D", block_sizes[0], 2051 "BLOCK_SIZE", block_sizes[0] * block_sizes[1] * block_sizes[2])); 2052 if (is_compile_good) { 2053 *is_good_build = true; 2054 CeedCallBackend(CeedGetKernel_Hip(ceed, is_full ? data->module_assemble_full : data->module_assemble_diagonal, operator_name.c_str(), 2055 is_full ? &data->assemble_full : &data->assemble_diagonal)); 2056 } else { 2057 *is_good_build = false; 2058 data->use_assembly_fallback = true; 2059 } 2060 } 2061 CeedCallBackend(CeedDestroy(&ceed)); 2062 CeedCallBackend(CeedQFunctionDestroy(&qf)); 2063 return CEED_ERROR_SUCCESS; 2064 } 2065 2066 extern "C" int CeedOperatorBuildKernelDiagonalAssemblyAtPoints_Hip_gen(CeedOperator op, bool *is_good_build) { 2067 return CeedOperatorBuildKernelAssemblyAtPoints_Hip_gen(op, false, is_good_build); 2068 } 2069 2070 //------------------------------------------------------------------------------ 2071