1 // Copyright (c) 2017-2022, Lawrence Livermore National Security, LLC and other CEED contributors. 2 // All Rights Reserved. See the top-level LICENSE and NOTICE files for details. 3 // 4 // SPDX-License-Identifier: BSD-2-Clause 5 // 6 // This file is part of CEED: http://github.com/ceed 7 8 #include <ceed/backend.h> 9 #include <ceed/ceed.h> 10 #include <ceed/jit-tools.h> 11 12 #include <sycl/sycl.hpp> 13 #include <vector> 14 15 #include "../sycl/ceed-sycl-compile.hpp" 16 #include "ceed-sycl-ref.hpp" 17 18 template <int> 19 class CeedBasisSyclInterp; 20 template <int> 21 class CeedBasisSyclGrad; 22 class CeedBasisSyclWeight; 23 24 class CeedBasisSyclInterpNT; 25 class CeedBasisSyclGradNT; 26 class CeedBasisSyclWeightNT; 27 28 using SpecID = sycl::specialization_id<CeedInt>; 29 30 static constexpr SpecID BASIS_DIM_ID; 31 static constexpr SpecID BASIS_NUM_COMP_ID; 32 static constexpr SpecID BASIS_P_1D_ID; 33 static constexpr SpecID BASIS_Q_1D_ID; 34 35 //------------------------------------------------------------------------------ 36 // Interpolation kernel - tensor 37 //------------------------------------------------------------------------------ 38 template <int transpose> 39 static int CeedBasisApplyInterp_Sycl(sycl::queue &sycl_queue, const SyclModule_t &sycl_module, CeedInt num_elem, const CeedBasis_Sycl *impl, 40 const CeedScalar *u, CeedScalar *v) { 41 const CeedInt buf_len = impl->buf_len; 42 const CeedInt op_len = impl->op_len; 43 const CeedScalar *interp_1d = impl->d_interp_1d; 44 45 const sycl::device &sycl_device = sycl_queue.get_device(); 46 const CeedInt max_work_group_size = 32; 47 const CeedInt work_group_size = CeedIntMin(impl->num_qpts, max_work_group_size); 48 sycl::range<1> local_range(work_group_size); 49 sycl::range<1> global_range(num_elem * work_group_size); 50 sycl::nd_range<1> kernel_range(global_range, local_range); 51 52 // Order queue 53 sycl::event e = sycl_queue.ext_oneapi_submit_barrier(); 54 sycl_queue.submit([&](sycl::handler &cgh) { 55 cgh.depends_on({e}); 56 cgh.use_kernel_bundle(sycl_module); 57 58 sycl::local_accessor<CeedScalar> s_mem(op_len + 2 * buf_len, cgh); 59 60 cgh.parallel_for<CeedBasisSyclInterp<transpose>>(kernel_range, [=](sycl::nd_item<1> work_item, sycl::kernel_handler kh) { 61 //--------------------------------------------------------------> 62 // Retrieve spec constant values 63 const CeedInt dim = kh.get_specialization_constant<BASIS_DIM_ID>(); 64 const CeedInt num_comp = kh.get_specialization_constant<BASIS_NUM_COMP_ID>(); 65 const CeedInt P_1d = kh.get_specialization_constant<BASIS_P_1D_ID>(); 66 const CeedInt Q_1d = kh.get_specialization_constant<BASIS_Q_1D_ID>(); 67 //--------------------------------------------------------------> 68 const CeedInt num_nodes = CeedIntPow(P_1d, dim); 69 const CeedInt num_qpts = CeedIntPow(Q_1d, dim); 70 const CeedInt P = transpose ? Q_1d : P_1d; 71 const CeedInt Q = transpose ? P_1d : Q_1d; 72 const CeedInt stride_0 = transpose ? 1 : P_1d; 73 const CeedInt stride_1 = transpose ? P_1d : 1; 74 const CeedInt u_stride = transpose ? num_qpts : num_nodes; 75 const CeedInt v_stride = transpose ? num_nodes : num_qpts; 76 const CeedInt u_comp_stride = num_elem * u_stride; 77 const CeedInt v_comp_stride = num_elem * v_stride; 78 const CeedInt u_size = u_stride; 79 80 sycl::group work_group = work_item.get_group(); 81 const CeedInt i = work_item.get_local_linear_id(); 82 const CeedInt group_size = work_group.get_local_linear_range(); 83 const CeedInt elem = work_group.get_group_linear_id(); 84 85 CeedScalar *s_interp_1d = s_mem.get_pointer(); 86 CeedScalar *s_buffer_1 = s_interp_1d + Q * P; 87 CeedScalar *s_buffer_2 = s_buffer_1 + buf_len; 88 89 for (CeedInt k = i; k < P * Q; k += group_size) { 90 s_interp_1d[k] = interp_1d[k]; 91 } 92 93 // Apply basis element by element 94 for (CeedInt comp = 0; comp < num_comp; comp++) { 95 const CeedScalar *cur_u = u + elem * u_stride + comp * u_comp_stride; 96 CeedScalar *cur_v = v + elem * v_stride + comp * v_comp_stride; 97 98 for (CeedInt k = i; k < u_size; k += group_size) { 99 s_buffer_1[k] = cur_u[k]; 100 } 101 102 CeedInt pre = u_size; 103 CeedInt post = 1; 104 105 for (CeedInt d = 0; d < dim; d++) { 106 // Use older version of sycl workgroup barrier for performance reasons 107 // Can be updated in future to align with SYCL2020 spec if performance bottleneck is removed 108 // sycl::group_barrier(work_group); 109 work_item.barrier(sycl::access::fence_space::local_space); 110 111 pre /= P; 112 const CeedScalar *in = d % 2 ? s_buffer_2 : s_buffer_1; 113 CeedScalar *out = d == dim - 1 ? cur_v : (d % 2 ? s_buffer_1 : s_buffer_2); 114 115 // Contract along middle index 116 const CeedInt writeLen = pre * post * Q; 117 for (CeedInt k = i; k < writeLen; k += group_size) { 118 const CeedInt c = k % post; 119 const CeedInt j = (k / post) % Q; 120 const CeedInt a = k / (post * Q); 121 122 CeedScalar vk = 0; 123 for (CeedInt b = 0; b < P; b++) { 124 vk += s_interp_1d[j * stride_0 + b * stride_1] * in[(a * P + b) * post + c]; 125 } 126 out[k] = vk; 127 } 128 post *= Q; 129 } 130 } 131 }); 132 }); 133 return CEED_ERROR_SUCCESS; 134 } 135 136 //------------------------------------------------------------------------------ 137 // Gradient kernel - tensor 138 //------------------------------------------------------------------------------ 139 template <int transpose> 140 static int CeedBasisApplyGrad_Sycl(sycl::queue &sycl_queue, const SyclModule_t &sycl_module, CeedInt num_elem, const CeedBasis_Sycl *impl, 141 const CeedScalar *u, CeedScalar *v) { 142 const CeedInt buf_len = impl->buf_len; 143 const CeedInt op_len = impl->op_len; 144 const CeedScalar *interp_1d = impl->d_interp_1d; 145 const CeedScalar *grad_1d = impl->d_grad_1d; 146 147 const sycl::device &sycl_device = sycl_queue.get_device(); 148 const CeedInt work_group_size = 32; 149 sycl::range<1> local_range(work_group_size); 150 sycl::range<1> global_range(num_elem * work_group_size); 151 sycl::nd_range<1> kernel_range(global_range, local_range); 152 153 // Order queue 154 sycl::event e = sycl_queue.ext_oneapi_submit_barrier(); 155 sycl_queue.submit([&](sycl::handler &cgh) { 156 cgh.depends_on({e}); 157 cgh.use_kernel_bundle(sycl_module); 158 159 sycl::local_accessor<CeedScalar> s_mem(2 * (op_len + buf_len), cgh); 160 161 cgh.parallel_for<CeedBasisSyclGrad<transpose>>(kernel_range, [=](sycl::nd_item<1> work_item, sycl::kernel_handler kh) { 162 //--------------------------------------------------------------> 163 // Retrieve spec constant values 164 const CeedInt dim = kh.get_specialization_constant<BASIS_DIM_ID>(); 165 const CeedInt num_comp = kh.get_specialization_constant<BASIS_NUM_COMP_ID>(); 166 const CeedInt P_1d = kh.get_specialization_constant<BASIS_P_1D_ID>(); 167 const CeedInt Q_1d = kh.get_specialization_constant<BASIS_Q_1D_ID>(); 168 //--------------------------------------------------------------> 169 const CeedInt num_nodes = CeedIntPow(P_1d, dim); 170 const CeedInt num_qpts = CeedIntPow(Q_1d, dim); 171 const CeedInt P = transpose ? Q_1d : P_1d; 172 const CeedInt Q = transpose ? P_1d : Q_1d; 173 const CeedInt stride_0 = transpose ? 1 : P_1d; 174 const CeedInt stride_1 = transpose ? P_1d : 1; 175 const CeedInt u_stride = transpose ? num_qpts : num_nodes; 176 const CeedInt v_stride = transpose ? num_nodes : num_qpts; 177 const CeedInt u_comp_stride = num_elem * u_stride; 178 const CeedInt v_comp_stride = num_elem * v_stride; 179 const CeedInt u_dim_stride = transpose ? num_elem * num_qpts * num_comp : 0; 180 const CeedInt v_dim_stride = transpose ? 0 : num_elem * num_qpts * num_comp; 181 182 sycl::group work_group = work_item.get_group(); 183 const CeedInt i = work_item.get_local_linear_id(); 184 const CeedInt group_size = work_group.get_local_linear_range(); 185 const CeedInt elem = work_group.get_group_linear_id(); 186 187 CeedScalar *s_interp_1d = s_mem.get_pointer(); 188 CeedScalar *s_grad_1d = s_interp_1d + P * Q; 189 CeedScalar *s_buffer_1 = s_grad_1d + P * Q; 190 CeedScalar *s_buffer_2 = s_buffer_1 + buf_len; 191 192 for (CeedInt k = i; k < P * Q; k += group_size) { 193 s_interp_1d[k] = interp_1d[k]; 194 s_grad_1d[k] = grad_1d[k]; 195 } 196 197 // Apply basis element by element 198 for (CeedInt comp = 0; comp < num_comp; comp++) { 199 for (CeedInt dim_1 = 0; dim_1 < dim; dim_1++) { 200 CeedInt pre = transpose ? num_qpts : num_nodes; 201 CeedInt post = 1; 202 const CeedScalar *cur_u = u + elem * u_stride + dim_1 * u_dim_stride + comp * u_comp_stride; 203 CeedScalar *cur_v = v + elem * v_stride + dim_1 * v_dim_stride + comp * v_comp_stride; 204 205 for (CeedInt dim_2 = 0; dim_2 < dim; dim_2++) { 206 // Use older version of sycl workgroup barrier for performance reasons 207 // Can be updated in future to align with SYCL2020 spec if performance bottleneck is removed 208 // sycl::group_barrier(work_group); 209 work_item.barrier(sycl::access::fence_space::local_space); 210 211 pre /= P; 212 const CeedScalar *op = dim_1 == dim_2 ? s_grad_1d : s_interp_1d; 213 const CeedScalar *in = (dim_2 == 0 ? cur_u : (dim_2 % 2 ? s_buffer_2 : s_buffer_1)); 214 CeedScalar *out = dim_2 == dim - 1 ? cur_v : (dim_2 % 2 ? s_buffer_1 : s_buffer_2); 215 216 // Contract along middle index 217 const CeedInt writeLen = pre * post * Q; 218 for (CeedInt k = i; k < writeLen; k += group_size) { 219 const CeedInt c = k % post; 220 const CeedInt j = (k / post) % Q; 221 const CeedInt a = k / (post * Q); 222 223 CeedScalar v_k = 0; 224 for (CeedInt b = 0; b < P; b++) v_k += op[j * stride_0 + b * stride_1] * in[(a * P + b) * post + c]; 225 226 if (transpose && dim_2 == dim - 1) out[k] += v_k; 227 else out[k] = v_k; 228 } 229 230 post *= Q; 231 } 232 } 233 } 234 }); 235 }); 236 return CEED_ERROR_SUCCESS; 237 } 238 239 //------------------------------------------------------------------------------ 240 // Weight kernel - tensor 241 //------------------------------------------------------------------------------ 242 static int CeedBasisApplyWeight_Sycl(sycl::queue &sycl_queue, CeedInt num_elem, const CeedBasis_Sycl *impl, CeedScalar *w) { 243 const CeedInt dim = impl->dim; 244 const CeedInt Q_1d = impl->Q_1d; 245 const CeedScalar *q_weight_1d = impl->d_q_weight_1d; 246 247 const CeedInt num_quad_x = Q_1d; 248 const CeedInt num_quad_y = (dim > 1) ? Q_1d : 1; 249 const CeedInt num_quad_z = (dim > 2) ? Q_1d : 1; 250 sycl::range<3> kernel_range(num_elem * num_quad_z, num_quad_y, num_quad_x); 251 252 // Order queue 253 sycl::event e = sycl_queue.ext_oneapi_submit_barrier(); 254 sycl_queue.parallel_for<CeedBasisSyclWeight>(kernel_range, {e}, [=](sycl::item<3> work_item) { 255 if (dim == 1) w[work_item.get_linear_id()] = q_weight_1d[work_item[2]]; 256 if (dim == 2) w[work_item.get_linear_id()] = q_weight_1d[work_item[2]] * q_weight_1d[work_item[1]]; 257 if (dim == 3) w[work_item.get_linear_id()] = q_weight_1d[work_item[2]] * q_weight_1d[work_item[1]] * q_weight_1d[work_item[0] % Q_1d]; 258 }); 259 return CEED_ERROR_SUCCESS; 260 } 261 262 //------------------------------------------------------------------------------ 263 // Basis apply - tensor 264 //------------------------------------------------------------------------------ 265 static int CeedBasisApply_Sycl(CeedBasis basis, const CeedInt num_elem, CeedTransposeMode t_mode, CeedEvalMode eval_mode, CeedVector u, 266 CeedVector v) { 267 Ceed ceed; 268 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 269 Ceed_Sycl *data; 270 CeedCallBackend(CeedGetData(ceed, &data)); 271 CeedBasis_Sycl *impl; 272 CeedCallBackend(CeedBasisGetData(basis, &impl)); 273 274 const CeedInt transpose = t_mode == CEED_TRANSPOSE; 275 276 // Read vectors 277 const CeedScalar *d_u; 278 CeedScalar *d_v; 279 if (eval_mode != CEED_EVAL_WEIGHT) { 280 CeedCallBackend(CeedVectorGetArrayRead(u, CEED_MEM_DEVICE, &d_u)); 281 } 282 CeedCallBackend(CeedVectorGetArrayWrite(v, CEED_MEM_DEVICE, &d_v)); 283 284 // Clear v for transpose operation 285 if (t_mode == CEED_TRANSPOSE) { 286 CeedSize length; 287 CeedCallBackend(CeedVectorGetLength(v, &length)); 288 // Order queue 289 sycl::event e = data->sycl_queue.ext_oneapi_submit_barrier(); 290 data->sycl_queue.fill<CeedScalar>(d_v, 0, length, {e}); 291 } 292 293 // Basis action 294 switch (eval_mode) { 295 case CEED_EVAL_INTERP: { 296 if (transpose) { 297 CeedCallBackend(CeedBasisApplyInterp_Sycl<CEED_TRANSPOSE>(data->sycl_queue, *impl->sycl_module, num_elem, impl, d_u, d_v)); 298 } else { 299 CeedCallBackend(CeedBasisApplyInterp_Sycl<CEED_NOTRANSPOSE>(data->sycl_queue, *impl->sycl_module, num_elem, impl, d_u, d_v)); 300 } 301 } break; 302 case CEED_EVAL_GRAD: { 303 if (transpose) { 304 CeedCallBackend(CeedBasisApplyGrad_Sycl<1>(data->sycl_queue, *impl->sycl_module, num_elem, impl, d_u, d_v)); 305 } else { 306 CeedCallBackend(CeedBasisApplyGrad_Sycl<0>(data->sycl_queue, *impl->sycl_module, num_elem, impl, d_u, d_v)); 307 } 308 } break; 309 case CEED_EVAL_WEIGHT: { 310 CeedCallBackend(CeedBasisApplyWeight_Sycl(data->sycl_queue, num_elem, impl, d_v)); 311 } break; 312 // LCOV_EXCL_START 313 // Evaluate the divergence to/from the quadrature points 314 case CEED_EVAL_DIV: 315 return CeedError(ceed, CEED_ERROR_BACKEND, "CEED_EVAL_DIV not supported"); 316 // Evaluate the curl to/from the quadrature points 317 case CEED_EVAL_CURL: 318 return CeedError(ceed, CEED_ERROR_BACKEND, "CEED_EVAL_CURL not supported"); 319 // Take no action, BasisApply should not have been called 320 case CEED_EVAL_NONE: 321 return CeedError(ceed, CEED_ERROR_BACKEND, "CEED_EVAL_NONE does not make sense in this context"); 322 // LCOV_EXCL_STOP 323 } 324 325 // Restore vectors 326 if (eval_mode != CEED_EVAL_WEIGHT) { 327 CeedCallBackend(CeedVectorRestoreArrayRead(u, &d_u)); 328 } 329 CeedCallBackend(CeedVectorRestoreArray(v, &d_v)); 330 return CEED_ERROR_SUCCESS; 331 } 332 333 //------------------------------------------------------------------------------ 334 // Interpolation kernel - non-tensor 335 //------------------------------------------------------------------------------ 336 static int CeedBasisApplyNonTensorInterp_Sycl(sycl::queue &sycl_queue, CeedInt num_elem, CeedInt transpose, const CeedBasisNonTensor_Sycl *impl, 337 const CeedScalar *d_U, CeedScalar *d_V) { 338 const CeedInt num_comp = impl->num_comp; 339 const CeedInt P = transpose ? impl->num_qpts : impl->num_nodes; 340 const CeedInt Q = transpose ? impl->num_nodes : impl->num_qpts; 341 const CeedInt stride_0 = transpose ? 1 : impl->num_nodes; 342 const CeedInt stride_1 = transpose ? impl->num_nodes : 1; 343 const CeedInt u_stride = P; 344 const CeedInt v_stride = Q; 345 const CeedInt u_comp_stride = u_stride * num_elem; 346 const CeedInt v_comp_stride = v_stride * num_elem; 347 const CeedInt u_size = P; 348 const CeedInt v_size = Q; 349 const CeedScalar *d_B = impl->d_interp; 350 351 sycl::range<2> kernel_range(num_elem, v_size); 352 353 // Order queue 354 sycl::event e = sycl_queue.ext_oneapi_submit_barrier(); 355 sycl_queue.parallel_for<CeedBasisSyclInterpNT>(kernel_range, {e}, [=](sycl::id<2> indx) { 356 const CeedInt i = indx[1]; 357 const CeedInt elem = indx[0]; 358 359 for (CeedInt comp = 0; comp < num_comp; comp++) { 360 const CeedScalar *U = d_U + elem * u_stride + comp * u_comp_stride; 361 CeedScalar V = 0.0; 362 363 for (CeedInt j = 0; j < u_size; ++j) { 364 V += d_B[i * stride_0 + j * stride_1] * U[j]; 365 } 366 d_V[i + elem * v_stride + comp * v_comp_stride] = V; 367 } 368 }); 369 return CEED_ERROR_SUCCESS; 370 } 371 372 //------------------------------------------------------------------------------ 373 // Gradient kernel - non-tensor 374 //------------------------------------------------------------------------------ 375 static int CeedBasisApplyNonTensorGrad_Sycl(sycl::queue &sycl_queue, CeedInt num_elem, CeedInt transpose, const CeedBasisNonTensor_Sycl *impl, 376 const CeedScalar *d_U, CeedScalar *d_V) { 377 const CeedInt num_comp = impl->num_comp; 378 const CeedInt P = transpose ? impl->num_qpts : impl->num_nodes; 379 const CeedInt Q = transpose ? impl->num_nodes : impl->num_qpts; 380 const CeedInt stride_0 = transpose ? 1 : impl->num_nodes; 381 const CeedInt stride_1 = transpose ? impl->num_nodes : 1; 382 const CeedInt g_dim_stride = P * Q; 383 const CeedInt u_stride = P; 384 const CeedInt v_stride = Q; 385 const CeedInt u_comp_stride = u_stride * num_elem; 386 const CeedInt v_comp_stride = v_stride * num_elem; 387 const CeedInt u_dim_stride = u_comp_stride * num_comp; 388 const CeedInt v_dim_stride = v_comp_stride * num_comp; 389 const CeedInt u_size = P; 390 const CeedInt v_size = Q; 391 const CeedInt in_dim = transpose ? impl->dim : 1; 392 const CeedInt out_dim = transpose ? 1 : impl->dim; 393 const CeedScalar *d_G = impl->d_grad; 394 395 sycl::range<2> kernel_range(num_elem, v_size); 396 397 // Order queue 398 sycl::event e = sycl_queue.ext_oneapi_submit_barrier(); 399 sycl_queue.parallel_for<CeedBasisSyclGradNT>(kernel_range, {e}, [=](sycl::id<2> indx) { 400 const CeedInt i = indx[1]; 401 const CeedInt elem = indx[0]; 402 403 for (CeedInt comp = 0; comp < num_comp; comp++) { 404 CeedScalar V[3] = {0.0, 0.0, 0.0}; 405 406 for (CeedInt d1 = 0; d1 < in_dim; ++d1) { 407 const CeedScalar *U = d_U + elem * u_stride + comp * u_comp_stride + d1 * u_dim_stride; 408 const CeedScalar *G = d_G + i * stride_0 + d1 * g_dim_stride; 409 410 for (CeedInt j = 0; j < u_size; ++j) { 411 const CeedScalar Uj = U[j]; 412 413 for (CeedInt d0 = 0; d0 < out_dim; ++d0) { 414 V[d0] += G[j * stride_1 + d0 * g_dim_stride] * Uj; 415 } 416 } 417 } 418 for (CeedInt d0 = 0; d0 < out_dim; ++d0) { 419 d_V[i + elem * v_stride + comp * v_comp_stride + d0 * v_dim_stride] = V[d0]; 420 } 421 } 422 }); 423 return CEED_ERROR_SUCCESS; 424 } 425 426 //------------------------------------------------------------------------------ 427 // Weight kernel - non-tensor 428 //------------------------------------------------------------------------------ 429 static int CeedBasisApplyNonTensorWeight_Sycl(sycl::queue &sycl_queue, CeedInt num_elem, const CeedBasisNonTensor_Sycl *impl, CeedScalar *d_V) { 430 const CeedInt num_qpts = impl->num_qpts; 431 const CeedScalar *q_weight = impl->d_q_weight; 432 433 sycl::range<2> kernel_range(num_elem, num_qpts); 434 435 // Order queue 436 sycl::event e = sycl_queue.ext_oneapi_submit_barrier(); 437 sycl_queue.parallel_for<CeedBasisSyclWeightNT>(kernel_range, {e}, [=](sycl::id<2> indx) { 438 const CeedInt i = indx[1]; 439 const CeedInt elem = indx[0]; 440 d_V[i + elem * num_qpts] = q_weight[i]; 441 }); 442 return CEED_ERROR_SUCCESS; 443 } 444 445 //------------------------------------------------------------------------------ 446 // Basis apply - non-tensor 447 //------------------------------------------------------------------------------ 448 static int CeedBasisApplyNonTensor_Sycl(CeedBasis basis, const CeedInt num_elem, CeedTransposeMode t_mode, CeedEvalMode eval_mode, CeedVector u, 449 CeedVector v) { 450 Ceed ceed; 451 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 452 CeedBasisNonTensor_Sycl *impl; 453 CeedCallBackend(CeedBasisGetData(basis, &impl)); 454 Ceed_Sycl *data; 455 CeedCallBackend(CeedGetData(ceed, &data)); 456 457 const CeedInt transpose = t_mode == CEED_TRANSPOSE; 458 459 // Read vectors 460 const CeedScalar *d_u; 461 CeedScalar *d_v; 462 if (eval_mode != CEED_EVAL_WEIGHT) { 463 CeedCallBackend(CeedVectorGetArrayRead(u, CEED_MEM_DEVICE, &d_u)); 464 } 465 CeedCallBackend(CeedVectorGetArrayWrite(v, CEED_MEM_DEVICE, &d_v)); 466 467 // Clear v for transpose operation 468 if (transpose) { 469 CeedSize length; 470 CeedCallBackend(CeedVectorGetLength(v, &length)); 471 // Order queue 472 sycl::event e = data->sycl_queue.ext_oneapi_submit_barrier(); 473 data->sycl_queue.fill<CeedScalar>(d_v, 0, length, {e}); 474 } 475 476 // Apply basis operation 477 switch (eval_mode) { 478 case CEED_EVAL_INTERP: { 479 CeedCallBackend(CeedBasisApplyNonTensorInterp_Sycl(data->sycl_queue, num_elem, transpose, impl, d_u, d_v)); 480 } break; 481 case CEED_EVAL_GRAD: { 482 CeedCallBackend(CeedBasisApplyNonTensorGrad_Sycl(data->sycl_queue, num_elem, transpose, impl, d_u, d_v)); 483 } break; 484 case CEED_EVAL_WEIGHT: { 485 CeedCallBackend(CeedBasisApplyNonTensorWeight_Sycl(data->sycl_queue, num_elem, impl, d_v)); 486 } break; 487 // LCOV_EXCL_START 488 // Evaluate the divergence to/from the quadrature points 489 case CEED_EVAL_DIV: 490 return CeedError(ceed, CEED_ERROR_BACKEND, "CEED_EVAL_DIV not supported"); 491 // Evaluate the curl to/from the quadrature points 492 case CEED_EVAL_CURL: 493 return CeedError(ceed, CEED_ERROR_BACKEND, "CEED_EVAL_CURL not supported"); 494 // Take no action, BasisApply should not have been called 495 case CEED_EVAL_NONE: 496 return CeedError(ceed, CEED_ERROR_BACKEND, "CEED_EVAL_NONE does not make sense in this context"); 497 // LCOV_EXCL_STOP 498 } 499 500 // Restore vectors 501 if (eval_mode != CEED_EVAL_WEIGHT) { 502 CeedCallBackend(CeedVectorRestoreArrayRead(u, &d_u)); 503 } 504 505 CeedCallBackend(CeedVectorRestoreArray(v, &d_v)); 506 return CEED_ERROR_SUCCESS; 507 } 508 509 //------------------------------------------------------------------------------ 510 // Destroy tensor basis 511 //------------------------------------------------------------------------------ 512 static int CeedBasisDestroy_Sycl(CeedBasis basis) { 513 Ceed ceed; 514 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 515 CeedBasis_Sycl *impl; 516 CeedCallBackend(CeedBasisGetData(basis, &impl)); 517 Ceed_Sycl *data; 518 CeedCallBackend(CeedGetData(ceed, &data)); 519 520 // Wait for all work to finish before freeing memory 521 CeedCallSycl(ceed, data->sycl_queue.wait_and_throw()); 522 523 CeedCallSycl(ceed, sycl::free(impl->d_q_weight_1d, data->sycl_context)); 524 CeedCallSycl(ceed, sycl::free(impl->d_interp_1d, data->sycl_context)); 525 CeedCallSycl(ceed, sycl::free(impl->d_grad_1d, data->sycl_context)); 526 527 CeedCallBackend(CeedFree(&impl)); 528 529 return CEED_ERROR_SUCCESS; 530 } 531 532 //------------------------------------------------------------------------------ 533 // Destroy non-tensor basis 534 //------------------------------------------------------------------------------ 535 static int CeedBasisDestroyNonTensor_Sycl(CeedBasis basis) { 536 Ceed ceed; 537 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 538 CeedBasisNonTensor_Sycl *impl; 539 CeedCallBackend(CeedBasisGetData(basis, &impl)); 540 Ceed_Sycl *data; 541 CeedCallBackend(CeedGetData(ceed, &data)); 542 543 // Wait for all work to finish before freeing memory 544 CeedCallSycl(ceed, data->sycl_queue.wait_and_throw()); 545 546 CeedCallSycl(ceed, sycl::free(impl->d_q_weight, data->sycl_context)); 547 CeedCallSycl(ceed, sycl::free(impl->d_interp, data->sycl_context)); 548 CeedCallSycl(ceed, sycl::free(impl->d_grad, data->sycl_context)); 549 550 CeedCallBackend(CeedFree(&impl)); 551 552 return CEED_ERROR_SUCCESS; 553 } 554 555 //------------------------------------------------------------------------------ 556 // Create tensor 557 //------------------------------------------------------------------------------ 558 int CeedBasisCreateTensorH1_Sycl(CeedInt dim, CeedInt P_1d, CeedInt Q_1d, const CeedScalar *interp_1d, const CeedScalar *grad_1d, 559 const CeedScalar *q_ref_1d, const CeedScalar *q_weight_1d, CeedBasis basis) { 560 Ceed ceed; 561 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 562 CeedBasis_Sycl *impl; 563 CeedCallBackend(CeedCalloc(1, &impl)); 564 Ceed_Sycl *data; 565 CeedCallBackend(CeedGetData(ceed, &data)); 566 567 CeedInt num_comp; 568 CeedCallBackend(CeedBasisGetNumComponents(basis, &num_comp)); 569 570 const CeedInt num_nodes = CeedIntPow(P_1d, dim); 571 const CeedInt num_qpts = CeedIntPow(Q_1d, dim); 572 573 impl->dim = dim; 574 impl->P_1d = P_1d; 575 impl->Q_1d = Q_1d; 576 impl->num_comp = num_comp; 577 impl->num_nodes = num_nodes; 578 impl->num_qpts = num_qpts; 579 impl->buf_len = num_comp * CeedIntMax(num_nodes, num_qpts); 580 impl->op_len = Q_1d * P_1d; 581 582 // Order queue 583 sycl::event e = data->sycl_queue.ext_oneapi_submit_barrier(); 584 585 CeedCallSycl(ceed, impl->d_q_weight_1d = sycl::malloc_device<CeedScalar>(Q_1d, data->sycl_device, data->sycl_context)); 586 sycl::event copy_weight = data->sycl_queue.copy<CeedScalar>(q_weight_1d, impl->d_q_weight_1d, Q_1d, {e}); 587 588 const CeedInt interp_length = Q_1d * P_1d; 589 CeedCallSycl(ceed, impl->d_interp_1d = sycl::malloc_device<CeedScalar>(interp_length, data->sycl_device, data->sycl_context)); 590 sycl::event copy_interp = data->sycl_queue.copy<CeedScalar>(interp_1d, impl->d_interp_1d, interp_length, {e}); 591 592 CeedCallSycl(ceed, impl->d_grad_1d = sycl::malloc_device<CeedScalar>(interp_length, data->sycl_device, data->sycl_context)); 593 sycl::event copy_grad = data->sycl_queue.copy<CeedScalar>(grad_1d, impl->d_grad_1d, interp_length, {e}); 594 595 CeedCallSycl(ceed, sycl::event::wait_and_throw({copy_weight, copy_interp, copy_grad})); 596 597 std::vector<sycl::kernel_id> kernel_ids = {sycl::get_kernel_id<CeedBasisSyclInterp<1>>(), sycl::get_kernel_id<CeedBasisSyclInterp<0>>(), 598 sycl::get_kernel_id<CeedBasisSyclGrad<1>>(), sycl::get_kernel_id<CeedBasisSyclGrad<0>>()}; 599 600 sycl::kernel_bundle<sycl::bundle_state::input> input_bundle = sycl::get_kernel_bundle<sycl::bundle_state::input>(data->sycl_context, kernel_ids); 601 input_bundle.set_specialization_constant<BASIS_DIM_ID>(dim); 602 input_bundle.set_specialization_constant<BASIS_NUM_COMP_ID>(num_comp); 603 input_bundle.set_specialization_constant<BASIS_Q_1D_ID>(Q_1d); 604 input_bundle.set_specialization_constant<BASIS_P_1D_ID>(P_1d); 605 606 CeedCallSycl(ceed, impl->sycl_module = new SyclModule_t(sycl::build(input_bundle))); 607 608 CeedCallBackend(CeedBasisSetData(basis, impl)); 609 610 // Register backend functions 611 CeedCallBackend(CeedSetBackendFunctionCpp(ceed, "Basis", basis, "Apply", CeedBasisApply_Sycl)); 612 CeedCallBackend(CeedSetBackendFunctionCpp(ceed, "Basis", basis, "Destroy", CeedBasisDestroy_Sycl)); 613 return CEED_ERROR_SUCCESS; 614 } 615 616 //------------------------------------------------------------------------------ 617 // Create non-tensor 618 //------------------------------------------------------------------------------ 619 int CeedBasisCreateH1_Sycl(CeedElemTopology topo, CeedInt dim, CeedInt num_nodes, CeedInt num_qpts, const CeedScalar *interp, const CeedScalar *grad, 620 const CeedScalar *qref, const CeedScalar *q_weight, CeedBasis basis) { 621 Ceed ceed; 622 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 623 CeedBasisNonTensor_Sycl *impl; 624 CeedCallBackend(CeedCalloc(1, &impl)); 625 Ceed_Sycl *data; 626 CeedCallBackend(CeedGetData(ceed, &data)); 627 628 CeedInt num_comp; 629 CeedCallBackend(CeedBasisGetNumComponents(basis, &num_comp)); 630 631 impl->dim = dim; 632 impl->num_comp = num_comp; 633 impl->num_nodes = num_nodes; 634 impl->num_qpts = num_qpts; 635 636 // Order queue 637 sycl::event e = data->sycl_queue.ext_oneapi_submit_barrier(); 638 639 CeedCallSycl(ceed, impl->d_q_weight = sycl::malloc_device<CeedScalar>(num_qpts, data->sycl_device, data->sycl_context)); 640 sycl::event copy_weight = data->sycl_queue.copy<CeedScalar>(q_weight, impl->d_q_weight, num_qpts, {e}); 641 642 const CeedInt interp_length = num_qpts * num_nodes; 643 CeedCallSycl(ceed, impl->d_interp = sycl::malloc_device<CeedScalar>(interp_length, data->sycl_device, data->sycl_context)); 644 sycl::event copy_interp = data->sycl_queue.copy<CeedScalar>(interp, impl->d_interp, interp_length, {e}); 645 646 const CeedInt grad_length = num_qpts * num_nodes * dim; 647 CeedCallSycl(ceed, impl->d_grad = sycl::malloc_device<CeedScalar>(grad_length, data->sycl_device, data->sycl_context)); 648 sycl::event copy_grad = data->sycl_queue.copy<CeedScalar>(grad, impl->d_grad, grad_length, {e}); 649 650 CeedCallSycl(ceed, sycl::event::wait_and_throw({copy_weight, copy_interp, copy_grad})); 651 652 CeedCallBackend(CeedBasisSetData(basis, impl)); 653 654 // Register backend functions 655 CeedCallBackend(CeedSetBackendFunctionCpp(ceed, "Basis", basis, "Apply", CeedBasisApplyNonTensor_Sycl)); 656 CeedCallBackend(CeedSetBackendFunctionCpp(ceed, "Basis", basis, "Destroy", CeedBasisDestroyNonTensor_Sycl)); 657 return CEED_ERROR_SUCCESS; 658 } 659 660 //------------------------------------------------------------------------------ 661