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 sycl::group work_group = work_item.get_group(); 182 const CeedInt i = work_item.get_local_linear_id(); 183 const CeedInt group_size = work_group.get_local_linear_range(); 184 const CeedInt elem = work_group.get_group_linear_id(); 185 186 CeedScalar *s_interp_1d = s_mem.get_pointer(); 187 CeedScalar *s_grad_1d = s_interp_1d + P * Q; 188 CeedScalar *s_buffer_1 = s_grad_1d + P * Q; 189 CeedScalar *s_buffer_2 = s_buffer_1 + buf_len; 190 191 for (CeedInt k = i; k < P * Q; k += group_size) { 192 s_interp_1d[k] = interp_1d[k]; 193 s_grad_1d[k] = grad_1d[k]; 194 } 195 196 // Apply basis element by element 197 for (CeedInt comp = 0; comp < num_comp; comp++) { 198 for (CeedInt dim_1 = 0; dim_1 < dim; dim_1++) { 199 CeedInt pre = transpose ? num_qpts : num_nodes; 200 CeedInt post = 1; 201 const CeedScalar *cur_u = u + elem * u_stride + dim_1 * u_dim_stride + comp * u_comp_stride; 202 CeedScalar *cur_v = v + elem * v_stride + dim_1 * v_dim_stride + comp * v_comp_stride; 203 204 for (CeedInt dim_2 = 0; dim_2 < dim; dim_2++) { 205 // Use older version of sycl workgroup barrier for performance reasons 206 // Can be updated in future to align with SYCL2020 spec if performance bottleneck is removed 207 // sycl::group_barrier(work_group); 208 work_item.barrier(sycl::access::fence_space::local_space); 209 210 pre /= P; 211 const CeedScalar *op = dim_1 == dim_2 ? s_grad_1d : s_interp_1d; 212 const CeedScalar *in = (dim_2 == 0 ? cur_u : (dim_2 % 2 ? s_buffer_2 : s_buffer_1)); 213 CeedScalar *out = dim_2 == dim - 1 ? cur_v : (dim_2 % 2 ? s_buffer_1 : s_buffer_2); 214 215 // Contract along middle index 216 const CeedInt writeLen = pre * post * Q; 217 for (CeedInt k = i; k < writeLen; k += group_size) { 218 const CeedInt c = k % post; 219 const CeedInt j = (k / post) % Q; 220 const CeedInt a = k / (post * Q); 221 222 CeedScalar v_k = 0; 223 for (CeedInt b = 0; b < P; b++) v_k += op[j * stride_0 + b * stride_1] * in[(a * P + b) * post + c]; 224 225 if (transpose && dim_2 == dim - 1) out[k] += v_k; 226 else out[k] = v_k; 227 } 228 229 post *= Q; 230 } 231 } 232 } 233 }); 234 }); 235 return CEED_ERROR_SUCCESS; 236 } 237 238 //------------------------------------------------------------------------------ 239 // Weight kernel - tensor 240 //------------------------------------------------------------------------------ 241 static int CeedBasisApplyWeight_Sycl(sycl::queue &sycl_queue, CeedInt num_elem, const CeedBasis_Sycl *impl, CeedScalar *w) { 242 const CeedInt dim = impl->dim; 243 const CeedInt Q_1d = impl->Q_1d; 244 const CeedScalar *q_weight_1d = impl->d_q_weight_1d; 245 246 const CeedInt num_quad_x = Q_1d; 247 const CeedInt num_quad_y = (dim > 1) ? Q_1d : 1; 248 const CeedInt num_quad_z = (dim > 2) ? Q_1d : 1; 249 sycl::range<3> kernel_range(num_elem * num_quad_z, num_quad_y, num_quad_x); 250 251 // Order queue 252 sycl::event e = sycl_queue.ext_oneapi_submit_barrier(); 253 sycl_queue.parallel_for<CeedBasisSyclWeight>(kernel_range, {e}, [=](sycl::item<3> work_item) { 254 if (dim == 1) w[work_item.get_linear_id()] = q_weight_1d[work_item[2]]; 255 if (dim == 2) w[work_item.get_linear_id()] = q_weight_1d[work_item[2]] * q_weight_1d[work_item[1]]; 256 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]; 257 }); 258 return CEED_ERROR_SUCCESS; 259 } 260 261 //------------------------------------------------------------------------------ 262 // Basis apply - tensor 263 //------------------------------------------------------------------------------ 264 static int CeedBasisApply_Sycl(CeedBasis basis, const CeedInt num_elem, CeedTransposeMode t_mode, CeedEvalMode eval_mode, CeedVector u, 265 CeedVector v) { 266 Ceed ceed; 267 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 268 Ceed_Sycl *data; 269 CeedCallBackend(CeedGetData(ceed, &data)); 270 CeedBasis_Sycl *impl; 271 CeedCallBackend(CeedBasisGetData(basis, &impl)); 272 273 const CeedInt transpose = t_mode == CEED_TRANSPOSE; 274 275 // Read vectors 276 const CeedScalar *d_u; 277 CeedScalar *d_v; 278 if (eval_mode != CEED_EVAL_WEIGHT) { 279 CeedCallBackend(CeedVectorGetArrayRead(u, CEED_MEM_DEVICE, &d_u)); 280 } 281 CeedCallBackend(CeedVectorGetArrayWrite(v, CEED_MEM_DEVICE, &d_v)); 282 283 // Clear v for transpose operation 284 if (t_mode == CEED_TRANSPOSE) { 285 CeedSize length; 286 CeedCallBackend(CeedVectorGetLength(v, &length)); 287 // Order queue 288 sycl::event e = data->sycl_queue.ext_oneapi_submit_barrier(); 289 data->sycl_queue.fill<CeedScalar>(d_v, 0, length, {e}); 290 } 291 292 // Basis action 293 switch (eval_mode) { 294 case CEED_EVAL_INTERP: { 295 if (transpose) { 296 CeedCallBackend(CeedBasisApplyInterp_Sycl<CEED_TRANSPOSE>(data->sycl_queue, *impl->sycl_module, num_elem, impl, d_u, d_v)); 297 } else { 298 CeedCallBackend(CeedBasisApplyInterp_Sycl<CEED_NOTRANSPOSE>(data->sycl_queue, *impl->sycl_module, num_elem, impl, d_u, d_v)); 299 } 300 } break; 301 case CEED_EVAL_GRAD: { 302 if (transpose) { 303 CeedCallBackend(CeedBasisApplyGrad_Sycl<1>(data->sycl_queue, *impl->sycl_module, num_elem, impl, d_u, d_v)); 304 } else { 305 CeedCallBackend(CeedBasisApplyGrad_Sycl<0>(data->sycl_queue, *impl->sycl_module, num_elem, impl, d_u, d_v)); 306 } 307 } break; 308 case CEED_EVAL_WEIGHT: { 309 CeedCallBackend(CeedBasisApplyWeight_Sycl(data->sycl_queue, num_elem, impl, d_v)); 310 } break; 311 // LCOV_EXCL_START 312 // Evaluate the divergence to/from the quadrature points 313 case CEED_EVAL_DIV: 314 return CeedError(ceed, CEED_ERROR_BACKEND, "CEED_EVAL_DIV not supported"); 315 // Evaluate the curl to/from the quadrature points 316 case CEED_EVAL_CURL: 317 return CeedError(ceed, CEED_ERROR_BACKEND, "CEED_EVAL_CURL not supported"); 318 // Take no action, BasisApply should not have been called 319 case CEED_EVAL_NONE: 320 return CeedError(ceed, CEED_ERROR_BACKEND, "CEED_EVAL_NONE does not make sense in this context"); 321 // LCOV_EXCL_STOP 322 } 323 324 // Restore vectors 325 if (eval_mode != CEED_EVAL_WEIGHT) { 326 CeedCallBackend(CeedVectorRestoreArrayRead(u, &d_u)); 327 } 328 CeedCallBackend(CeedVectorRestoreArray(v, &d_v)); 329 return CEED_ERROR_SUCCESS; 330 } 331 332 //------------------------------------------------------------------------------ 333 // Interpolation kernel - non-tensor 334 //------------------------------------------------------------------------------ 335 static int CeedBasisApplyNonTensorInterp_Sycl(sycl::queue &sycl_queue, CeedInt num_elem, CeedInt transpose, const CeedBasisNonTensor_Sycl *impl, 336 const CeedScalar *d_U, CeedScalar *d_V) { 337 const CeedInt num_comp = impl->num_comp; 338 const CeedInt P = transpose ? impl->num_qpts : impl->num_nodes; 339 const CeedInt Q = transpose ? impl->num_nodes : impl->num_qpts; 340 const CeedInt stride_0 = transpose ? 1 : impl->num_nodes; 341 const CeedInt stride_1 = transpose ? impl->num_nodes : 1; 342 const CeedInt u_stride = P; 343 const CeedInt v_stride = Q; 344 const CeedInt u_comp_stride = u_stride * num_elem; 345 const CeedInt v_comp_stride = v_stride * num_elem; 346 const CeedInt u_size = P; 347 const CeedInt v_size = Q; 348 const CeedScalar *d_B = impl->d_interp; 349 350 sycl::range<2> kernel_range(num_elem, v_size); 351 352 // Order queue 353 sycl::event e = sycl_queue.ext_oneapi_submit_barrier(); 354 sycl_queue.parallel_for<CeedBasisSyclInterpNT>(kernel_range, {e}, [=](sycl::id<2> indx) { 355 const CeedInt i = indx[1]; 356 const CeedInt elem = indx[0]; 357 358 for (CeedInt comp = 0; comp < num_comp; comp++) { 359 const CeedScalar *U = d_U + elem * u_stride + comp * u_comp_stride; 360 CeedScalar V = 0.0; 361 362 for (CeedInt j = 0; j < u_size; ++j) { 363 V += d_B[i * stride_0 + j * stride_1] * U[j]; 364 } 365 d_V[i + elem * v_stride + comp * v_comp_stride] = V; 366 } 367 }); 368 return CEED_ERROR_SUCCESS; 369 } 370 371 //------------------------------------------------------------------------------ 372 // Gradient kernel - non-tensor 373 //------------------------------------------------------------------------------ 374 static int CeedBasisApplyNonTensorGrad_Sycl(sycl::queue &sycl_queue, CeedInt num_elem, CeedInt transpose, const CeedBasisNonTensor_Sycl *impl, 375 const CeedScalar *d_U, CeedScalar *d_V) { 376 const CeedInt num_comp = impl->num_comp; 377 const CeedInt P = transpose ? impl->num_qpts : impl->num_nodes; 378 const CeedInt Q = transpose ? impl->num_nodes : impl->num_qpts; 379 const CeedInt stride_0 = transpose ? 1 : impl->num_nodes; 380 const CeedInt stride_1 = transpose ? impl->num_nodes : 1; 381 const CeedInt g_dim_stride = P * Q; 382 const CeedInt u_stride = P; 383 const CeedInt v_stride = Q; 384 const CeedInt u_comp_stride = u_stride * num_elem; 385 const CeedInt v_comp_stride = v_stride * num_elem; 386 const CeedInt u_dim_stride = u_comp_stride * num_comp; 387 const CeedInt v_dim_stride = v_comp_stride * num_comp; 388 const CeedInt u_size = P; 389 const CeedInt v_size = Q; 390 const CeedInt in_dim = transpose ? impl->dim : 1; 391 const CeedInt out_dim = transpose ? 1 : impl->dim; 392 const CeedScalar *d_G = impl->d_grad; 393 394 sycl::range<2> kernel_range(num_elem, v_size); 395 396 // Order queue 397 sycl::event e = sycl_queue.ext_oneapi_submit_barrier(); 398 sycl_queue.parallel_for<CeedBasisSyclGradNT>(kernel_range, {e}, [=](sycl::id<2> indx) { 399 const CeedInt i = indx[1]; 400 const CeedInt elem = indx[0]; 401 402 for (CeedInt comp = 0; comp < num_comp; comp++) { 403 CeedScalar V[3] = {0.0, 0.0, 0.0}; 404 405 for (CeedInt d1 = 0; d1 < in_dim; ++d1) { 406 const CeedScalar *U = d_U + elem * u_stride + comp * u_comp_stride + d1 * u_dim_stride; 407 const CeedScalar *G = d_G + i * stride_0 + d1 * g_dim_stride; 408 409 for (CeedInt j = 0; j < u_size; ++j) { 410 const CeedScalar Uj = U[j]; 411 412 for (CeedInt d0 = 0; d0 < out_dim; ++d0) { 413 V[d0] += G[j * stride_1 + d0 * g_dim_stride] * Uj; 414 } 415 } 416 } 417 for (CeedInt d0 = 0; d0 < out_dim; ++d0) { 418 d_V[i + elem * v_stride + comp * v_comp_stride + d0 * v_dim_stride] = V[d0]; 419 } 420 } 421 }); 422 return CEED_ERROR_SUCCESS; 423 } 424 425 //------------------------------------------------------------------------------ 426 // Weight kernel - non-tensor 427 //------------------------------------------------------------------------------ 428 static int CeedBasisApplyNonTensorWeight_Sycl(sycl::queue &sycl_queue, CeedInt num_elem, const CeedBasisNonTensor_Sycl *impl, CeedScalar *d_V) { 429 const CeedInt num_qpts = impl->num_qpts; 430 const CeedScalar *q_weight = impl->d_q_weight; 431 432 sycl::range<2> kernel_range(num_elem, num_qpts); 433 434 // Order queue 435 sycl::event e = sycl_queue.ext_oneapi_submit_barrier(); 436 sycl_queue.parallel_for<CeedBasisSyclWeightNT>(kernel_range, {e}, [=](sycl::id<2> indx) { 437 const CeedInt i = indx[1]; 438 const CeedInt elem = indx[0]; 439 d_V[i + elem * num_qpts] = q_weight[i]; 440 }); 441 return CEED_ERROR_SUCCESS; 442 } 443 444 //------------------------------------------------------------------------------ 445 // Basis apply - non-tensor 446 //------------------------------------------------------------------------------ 447 static int CeedBasisApplyNonTensor_Sycl(CeedBasis basis, const CeedInt num_elem, CeedTransposeMode t_mode, CeedEvalMode eval_mode, CeedVector u, 448 CeedVector v) { 449 Ceed ceed; 450 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 451 CeedBasisNonTensor_Sycl *impl; 452 CeedCallBackend(CeedBasisGetData(basis, &impl)); 453 Ceed_Sycl *data; 454 CeedCallBackend(CeedGetData(ceed, &data)); 455 456 const CeedInt transpose = t_mode == CEED_TRANSPOSE; 457 458 // Read vectors 459 const CeedScalar *d_u; 460 CeedScalar *d_v; 461 if (eval_mode != CEED_EVAL_WEIGHT) { 462 CeedCallBackend(CeedVectorGetArrayRead(u, CEED_MEM_DEVICE, &d_u)); 463 } 464 CeedCallBackend(CeedVectorGetArrayWrite(v, CEED_MEM_DEVICE, &d_v)); 465 466 // Clear v for transpose operation 467 if (transpose) { 468 CeedSize length; 469 CeedCallBackend(CeedVectorGetLength(v, &length)); 470 // Order queue 471 sycl::event e = data->sycl_queue.ext_oneapi_submit_barrier(); 472 data->sycl_queue.fill<CeedScalar>(d_v, 0, length, {e}); 473 } 474 475 // Apply basis operation 476 switch (eval_mode) { 477 case CEED_EVAL_INTERP: { 478 CeedCallBackend(CeedBasisApplyNonTensorInterp_Sycl(data->sycl_queue, num_elem, transpose, impl, d_u, d_v)); 479 } break; 480 case CEED_EVAL_GRAD: { 481 CeedCallBackend(CeedBasisApplyNonTensorGrad_Sycl(data->sycl_queue, num_elem, transpose, impl, d_u, d_v)); 482 } break; 483 case CEED_EVAL_WEIGHT: { 484 CeedCallBackend(CeedBasisApplyNonTensorWeight_Sycl(data->sycl_queue, num_elem, impl, d_v)); 485 } break; 486 // LCOV_EXCL_START 487 // Evaluate the divergence to/from the quadrature points 488 case CEED_EVAL_DIV: 489 return CeedError(ceed, CEED_ERROR_BACKEND, "CEED_EVAL_DIV not supported"); 490 // Evaluate the curl to/from the quadrature points 491 case CEED_EVAL_CURL: 492 return CeedError(ceed, CEED_ERROR_BACKEND, "CEED_EVAL_CURL not supported"); 493 // Take no action, BasisApply should not have been called 494 case CEED_EVAL_NONE: 495 return CeedError(ceed, CEED_ERROR_BACKEND, "CEED_EVAL_NONE does not make sense in this context"); 496 // LCOV_EXCL_STOP 497 } 498 499 // Restore vectors 500 if (eval_mode != CEED_EVAL_WEIGHT) { 501 CeedCallBackend(CeedVectorRestoreArrayRead(u, &d_u)); 502 } 503 504 CeedCallBackend(CeedVectorRestoreArray(v, &d_v)); 505 return CEED_ERROR_SUCCESS; 506 } 507 508 //------------------------------------------------------------------------------ 509 // Destroy tensor basis 510 //------------------------------------------------------------------------------ 511 static int CeedBasisDestroy_Sycl(CeedBasis basis) { 512 Ceed ceed; 513 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 514 CeedBasis_Sycl *impl; 515 CeedCallBackend(CeedBasisGetData(basis, &impl)); 516 Ceed_Sycl *data; 517 CeedCallBackend(CeedGetData(ceed, &data)); 518 519 // Wait for all work to finish before freeing memory 520 CeedCallSycl(ceed, data->sycl_queue.wait_and_throw()); 521 522 CeedCallSycl(ceed, sycl::free(impl->d_q_weight_1d, data->sycl_context)); 523 CeedCallSycl(ceed, sycl::free(impl->d_interp_1d, data->sycl_context)); 524 CeedCallSycl(ceed, sycl::free(impl->d_grad_1d, data->sycl_context)); 525 526 CeedCallBackend(CeedFree(&impl)); 527 return CEED_ERROR_SUCCESS; 528 } 529 530 //------------------------------------------------------------------------------ 531 // Destroy non-tensor basis 532 //------------------------------------------------------------------------------ 533 static int CeedBasisDestroyNonTensor_Sycl(CeedBasis basis) { 534 Ceed ceed; 535 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 536 CeedBasisNonTensor_Sycl *impl; 537 CeedCallBackend(CeedBasisGetData(basis, &impl)); 538 Ceed_Sycl *data; 539 CeedCallBackend(CeedGetData(ceed, &data)); 540 541 // Wait for all work to finish before freeing memory 542 CeedCallSycl(ceed, data->sycl_queue.wait_and_throw()); 543 544 CeedCallSycl(ceed, sycl::free(impl->d_q_weight, data->sycl_context)); 545 CeedCallSycl(ceed, sycl::free(impl->d_interp, data->sycl_context)); 546 CeedCallSycl(ceed, sycl::free(impl->d_grad, data->sycl_context)); 547 548 CeedCallBackend(CeedFree(&impl)); 549 return CEED_ERROR_SUCCESS; 550 } 551 552 //------------------------------------------------------------------------------ 553 // Create tensor 554 //------------------------------------------------------------------------------ 555 int CeedBasisCreateTensorH1_Sycl(CeedInt dim, CeedInt P_1d, CeedInt Q_1d, const CeedScalar *interp_1d, const CeedScalar *grad_1d, 556 const CeedScalar *q_ref_1d, const CeedScalar *q_weight_1d, CeedBasis basis) { 557 Ceed ceed; 558 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 559 CeedBasis_Sycl *impl; 560 CeedCallBackend(CeedCalloc(1, &impl)); 561 Ceed_Sycl *data; 562 CeedCallBackend(CeedGetData(ceed, &data)); 563 564 CeedInt num_comp; 565 CeedCallBackend(CeedBasisGetNumComponents(basis, &num_comp)); 566 567 const CeedInt num_nodes = CeedIntPow(P_1d, dim); 568 const CeedInt num_qpts = CeedIntPow(Q_1d, dim); 569 570 impl->dim = dim; 571 impl->P_1d = P_1d; 572 impl->Q_1d = Q_1d; 573 impl->num_comp = num_comp; 574 impl->num_nodes = num_nodes; 575 impl->num_qpts = num_qpts; 576 impl->buf_len = num_comp * CeedIntMax(num_nodes, num_qpts); 577 impl->op_len = Q_1d * P_1d; 578 579 // Order queue 580 sycl::event e = data->sycl_queue.ext_oneapi_submit_barrier(); 581 582 CeedCallSycl(ceed, impl->d_q_weight_1d = sycl::malloc_device<CeedScalar>(Q_1d, data->sycl_device, data->sycl_context)); 583 sycl::event copy_weight = data->sycl_queue.copy<CeedScalar>(q_weight_1d, impl->d_q_weight_1d, Q_1d, {e}); 584 585 const CeedInt interp_length = Q_1d * P_1d; 586 CeedCallSycl(ceed, impl->d_interp_1d = sycl::malloc_device<CeedScalar>(interp_length, data->sycl_device, data->sycl_context)); 587 sycl::event copy_interp = data->sycl_queue.copy<CeedScalar>(interp_1d, impl->d_interp_1d, interp_length, {e}); 588 589 CeedCallSycl(ceed, impl->d_grad_1d = sycl::malloc_device<CeedScalar>(interp_length, data->sycl_device, data->sycl_context)); 590 sycl::event copy_grad = data->sycl_queue.copy<CeedScalar>(grad_1d, impl->d_grad_1d, interp_length, {e}); 591 592 CeedCallSycl(ceed, sycl::event::wait_and_throw({copy_weight, copy_interp, copy_grad})); 593 594 std::vector<sycl::kernel_id> kernel_ids = {sycl::get_kernel_id<CeedBasisSyclInterp<1>>(), sycl::get_kernel_id<CeedBasisSyclInterp<0>>(), 595 sycl::get_kernel_id<CeedBasisSyclGrad<1>>(), sycl::get_kernel_id<CeedBasisSyclGrad<0>>()}; 596 597 sycl::kernel_bundle<sycl::bundle_state::input> input_bundle = sycl::get_kernel_bundle<sycl::bundle_state::input>(data->sycl_context, kernel_ids); 598 input_bundle.set_specialization_constant<BASIS_DIM_ID>(dim); 599 input_bundle.set_specialization_constant<BASIS_NUM_COMP_ID>(num_comp); 600 input_bundle.set_specialization_constant<BASIS_Q_1D_ID>(Q_1d); 601 input_bundle.set_specialization_constant<BASIS_P_1D_ID>(P_1d); 602 603 CeedCallSycl(ceed, impl->sycl_module = new SyclModule_t(sycl::build(input_bundle))); 604 605 CeedCallBackend(CeedBasisSetData(basis, impl)); 606 607 // Register backend functions 608 CeedCallBackend(CeedSetBackendFunctionCpp(ceed, "Basis", basis, "Apply", CeedBasisApply_Sycl)); 609 CeedCallBackend(CeedSetBackendFunctionCpp(ceed, "Basis", basis, "Destroy", CeedBasisDestroy_Sycl)); 610 return CEED_ERROR_SUCCESS; 611 } 612 613 //------------------------------------------------------------------------------ 614 // Create non-tensor 615 //------------------------------------------------------------------------------ 616 int CeedBasisCreateH1_Sycl(CeedElemTopology topo, CeedInt dim, CeedInt num_nodes, CeedInt num_qpts, const CeedScalar *interp, const CeedScalar *grad, 617 const CeedScalar *q_ref, const CeedScalar *q_weight, CeedBasis basis) { 618 Ceed ceed; 619 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 620 CeedBasisNonTensor_Sycl *impl; 621 CeedCallBackend(CeedCalloc(1, &impl)); 622 Ceed_Sycl *data; 623 CeedCallBackend(CeedGetData(ceed, &data)); 624 625 CeedInt num_comp; 626 CeedCallBackend(CeedBasisGetNumComponents(basis, &num_comp)); 627 628 impl->dim = dim; 629 impl->num_comp = num_comp; 630 impl->num_nodes = num_nodes; 631 impl->num_qpts = num_qpts; 632 633 // Order queue 634 sycl::event e = data->sycl_queue.ext_oneapi_submit_barrier(); 635 636 CeedCallSycl(ceed, impl->d_q_weight = sycl::malloc_device<CeedScalar>(num_qpts, data->sycl_device, data->sycl_context)); 637 sycl::event copy_weight = data->sycl_queue.copy<CeedScalar>(q_weight, impl->d_q_weight, num_qpts, {e}); 638 639 const CeedInt interp_length = num_qpts * num_nodes; 640 CeedCallSycl(ceed, impl->d_interp = sycl::malloc_device<CeedScalar>(interp_length, data->sycl_device, data->sycl_context)); 641 sycl::event copy_interp = data->sycl_queue.copy<CeedScalar>(interp, impl->d_interp, interp_length, {e}); 642 643 const CeedInt grad_length = num_qpts * num_nodes * dim; 644 CeedCallSycl(ceed, impl->d_grad = sycl::malloc_device<CeedScalar>(grad_length, data->sycl_device, data->sycl_context)); 645 sycl::event copy_grad = data->sycl_queue.copy<CeedScalar>(grad, impl->d_grad, grad_length, {e}); 646 647 CeedCallSycl(ceed, sycl::event::wait_and_throw({copy_weight, copy_interp, copy_grad})); 648 649 CeedCallBackend(CeedBasisSetData(basis, impl)); 650 651 // Register backend functions 652 CeedCallBackend(CeedSetBackendFunctionCpp(ceed, "Basis", basis, "Apply", CeedBasisApplyNonTensor_Sycl)); 653 CeedCallBackend(CeedSetBackendFunctionCpp(ceed, "Basis", basis, "Destroy", CeedBasisDestroyNonTensor_Sycl)); 654 return CEED_ERROR_SUCCESS; 655 } 656 657 //------------------------------------------------------------------------------ 658