1 // Copyright (c) 2017-2024, Lawrence Livermore National Security, LLC and other CEED contributors. 2 // All Rights Reserved. See the top-level LICENSE and NOTICE files for details. 3 // 4 // SPDX-License-Identifier: BSD-2-Clause 5 // 6 // This file is part of CEED: http://github.com/ceed 7 8 #include <ceed/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 is_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 std::vector<sycl::event> e; 53 54 if (!sycl_queue.is_in_order()) e = {sycl_queue.ext_oneapi_submit_barrier()}; 55 56 sycl_queue.submit([&](sycl::handler &cgh) { 57 cgh.depends_on(e); 58 cgh.use_kernel_bundle(sycl_module); 59 60 sycl::local_accessor<CeedScalar> s_mem(op_len + 2 * buf_len, cgh); 61 62 cgh.parallel_for<CeedBasisSyclInterp<is_transpose>>(kernel_range, [=](sycl::nd_item<1> work_item, sycl::kernel_handler kh) { 63 //--------------------------------------------------------------> 64 // Retrieve spec constant values 65 const CeedInt dim = kh.get_specialization_constant<BASIS_DIM_ID>(); 66 const CeedInt num_comp = kh.get_specialization_constant<BASIS_NUM_COMP_ID>(); 67 const CeedInt P_1d = kh.get_specialization_constant<BASIS_P_1D_ID>(); 68 const CeedInt Q_1d = kh.get_specialization_constant<BASIS_Q_1D_ID>(); 69 //--------------------------------------------------------------> 70 const CeedInt num_nodes = CeedIntPow(P_1d, dim); 71 const CeedInt num_qpts = CeedIntPow(Q_1d, dim); 72 const CeedInt P = is_transpose ? Q_1d : P_1d; 73 const CeedInt Q = is_transpose ? P_1d : Q_1d; 74 const CeedInt stride_0 = is_transpose ? 1 : P_1d; 75 const CeedInt stride_1 = is_transpose ? P_1d : 1; 76 const CeedInt u_stride = is_transpose ? num_qpts : num_nodes; 77 const CeedInt v_stride = is_transpose ? num_nodes : num_qpts; 78 const CeedInt u_comp_stride = num_elem * u_stride; 79 const CeedInt v_comp_stride = num_elem * v_stride; 80 const CeedInt u_size = u_stride; 81 82 sycl::group work_group = work_item.get_group(); 83 const CeedInt i = work_item.get_local_linear_id(); 84 const CeedInt group_size = work_group.get_local_linear_range(); 85 const CeedInt elem = work_group.get_group_linear_id(); 86 87 CeedScalar *s_interp_1d = s_mem.get_multi_ptr<sycl::access::decorated::yes>().get(); 88 CeedScalar *s_buffer_1 = s_interp_1d + Q * P; 89 CeedScalar *s_buffer_2 = s_buffer_1 + buf_len; 90 91 for (CeedInt k = i; k < P * Q; k += group_size) { 92 s_interp_1d[k] = interp_1d[k]; 93 } 94 95 // Apply basis element by element 96 for (CeedInt comp = 0; comp < num_comp; comp++) { 97 const CeedScalar *cur_u = u + elem * u_stride + comp * u_comp_stride; 98 CeedScalar *cur_v = v + elem * v_stride + comp * v_comp_stride; 99 100 for (CeedInt k = i; k < u_size; k += group_size) { 101 s_buffer_1[k] = cur_u[k]; 102 } 103 104 CeedInt pre = u_size; 105 CeedInt post = 1; 106 107 for (CeedInt d = 0; d < dim; d++) { 108 // Use older version of sycl workgroup barrier for performance reasons 109 // Can be updated in future to align with SYCL2020 spec if performance bottleneck is removed 110 // sycl::group_barrier(work_group); 111 work_item.barrier(sycl::access::fence_space::local_space); 112 113 pre /= P; 114 const CeedScalar *in = d % 2 ? s_buffer_2 : s_buffer_1; 115 CeedScalar *out = d == dim - 1 ? cur_v : (d % 2 ? s_buffer_1 : s_buffer_2); 116 117 // Contract along middle index 118 const CeedInt writeLen = pre * post * Q; 119 for (CeedInt k = i; k < writeLen; k += group_size) { 120 const CeedInt c = k % post; 121 const CeedInt j = (k / post) % Q; 122 const CeedInt a = k / (post * Q); 123 124 CeedScalar vk = 0; 125 for (CeedInt b = 0; b < P; b++) { 126 vk += s_interp_1d[j * stride_0 + b * stride_1] * in[(a * P + b) * post + c]; 127 } 128 out[k] = vk; 129 } 130 post *= Q; 131 } 132 } 133 }); 134 }); 135 return CEED_ERROR_SUCCESS; 136 } 137 138 //------------------------------------------------------------------------------ 139 // Gradient kernel - tensor 140 //------------------------------------------------------------------------------ 141 template <int is_transpose> 142 static int CeedBasisApplyGrad_Sycl(sycl::queue &sycl_queue, const SyclModule_t &sycl_module, CeedInt num_elem, const CeedBasis_Sycl *impl, 143 const CeedScalar *u, CeedScalar *v) { 144 const CeedInt buf_len = impl->buf_len; 145 const CeedInt op_len = impl->op_len; 146 const CeedScalar *interp_1d = impl->d_interp_1d; 147 const CeedScalar *grad_1d = impl->d_grad_1d; 148 149 const sycl::device &sycl_device = sycl_queue.get_device(); 150 const CeedInt work_group_size = 32; 151 sycl::range<1> local_range(work_group_size); 152 sycl::range<1> global_range(num_elem * work_group_size); 153 sycl::nd_range<1> kernel_range(global_range, local_range); 154 155 std::vector<sycl::event> e; 156 157 if (!sycl_queue.is_in_order()) e = {sycl_queue.ext_oneapi_submit_barrier()}; 158 159 sycl_queue.submit([&](sycl::handler &cgh) { 160 cgh.depends_on(e); 161 cgh.use_kernel_bundle(sycl_module); 162 163 sycl::local_accessor<CeedScalar> s_mem(2 * (op_len + buf_len), cgh); 164 165 cgh.parallel_for<CeedBasisSyclGrad<is_transpose>>(kernel_range, [=](sycl::nd_item<1> work_item, sycl::kernel_handler kh) { 166 //--------------------------------------------------------------> 167 // Retrieve spec constant values 168 const CeedInt dim = kh.get_specialization_constant<BASIS_DIM_ID>(); 169 const CeedInt num_comp = kh.get_specialization_constant<BASIS_NUM_COMP_ID>(); 170 const CeedInt P_1d = kh.get_specialization_constant<BASIS_P_1D_ID>(); 171 const CeedInt Q_1d = kh.get_specialization_constant<BASIS_Q_1D_ID>(); 172 //--------------------------------------------------------------> 173 const CeedInt num_nodes = CeedIntPow(P_1d, dim); 174 const CeedInt num_qpts = CeedIntPow(Q_1d, dim); 175 const CeedInt P = is_transpose ? Q_1d : P_1d; 176 const CeedInt Q = is_transpose ? P_1d : Q_1d; 177 const CeedInt stride_0 = is_transpose ? 1 : P_1d; 178 const CeedInt stride_1 = is_transpose ? P_1d : 1; 179 const CeedInt u_stride = is_transpose ? num_qpts : num_nodes; 180 const CeedInt v_stride = is_transpose ? num_nodes : num_qpts; 181 const CeedInt u_comp_stride = num_elem * u_stride; 182 const CeedInt v_comp_stride = num_elem * v_stride; 183 const CeedInt u_dim_stride = is_transpose ? num_elem * num_qpts * num_comp : 0; 184 const CeedInt v_dim_stride = is_transpose ? 0 : num_elem * num_qpts * num_comp; 185 sycl::group work_group = work_item.get_group(); 186 const CeedInt i = work_item.get_local_linear_id(); 187 const CeedInt group_size = work_group.get_local_linear_range(); 188 const CeedInt elem = work_group.get_group_linear_id(); 189 190 CeedScalar *s_interp_1d = s_mem.get_multi_ptr<sycl::access::decorated::yes>().get(); 191 CeedScalar *s_grad_1d = s_interp_1d + P * Q; 192 CeedScalar *s_buffer_1 = s_grad_1d + P * Q; 193 CeedScalar *s_buffer_2 = s_buffer_1 + buf_len; 194 195 for (CeedInt k = i; k < P * Q; k += group_size) { 196 s_interp_1d[k] = interp_1d[k]; 197 s_grad_1d[k] = grad_1d[k]; 198 } 199 200 // Apply basis element by element 201 for (CeedInt comp = 0; comp < num_comp; comp++) { 202 for (CeedInt dim_1 = 0; dim_1 < dim; dim_1++) { 203 CeedInt pre = is_transpose ? num_qpts : num_nodes; 204 CeedInt post = 1; 205 const CeedScalar *cur_u = u + elem * u_stride + dim_1 * u_dim_stride + comp * u_comp_stride; 206 CeedScalar *cur_v = v + elem * v_stride + dim_1 * v_dim_stride + comp * v_comp_stride; 207 208 for (CeedInt dim_2 = 0; dim_2 < dim; dim_2++) { 209 // Use older version of sycl workgroup barrier for performance reasons 210 // Can be updated in future to align with SYCL2020 spec if performance bottleneck is removed 211 // sycl::group_barrier(work_group); 212 work_item.barrier(sycl::access::fence_space::local_space); 213 214 pre /= P; 215 const CeedScalar *op = dim_1 == dim_2 ? s_grad_1d : s_interp_1d; 216 const CeedScalar *in = (dim_2 == 0 ? cur_u : (dim_2 % 2 ? s_buffer_2 : s_buffer_1)); 217 CeedScalar *out = dim_2 == dim - 1 ? cur_v : (dim_2 % 2 ? s_buffer_1 : s_buffer_2); 218 219 // Contract along middle index 220 const CeedInt writeLen = pre * post * Q; 221 for (CeedInt k = i; k < writeLen; k += group_size) { 222 const CeedInt c = k % post; 223 const CeedInt j = (k / post) % Q; 224 const CeedInt a = k / (post * Q); 225 226 CeedScalar v_k = 0; 227 for (CeedInt b = 0; b < P; b++) v_k += op[j * stride_0 + b * stride_1] * in[(a * P + b) * post + c]; 228 229 if (is_transpose && dim_2 == dim - 1) out[k] += v_k; 230 else out[k] = v_k; 231 } 232 233 post *= Q; 234 } 235 } 236 } 237 }); 238 }); 239 return CEED_ERROR_SUCCESS; 240 } 241 242 //------------------------------------------------------------------------------ 243 // Weight kernel - tensor 244 //------------------------------------------------------------------------------ 245 static int CeedBasisApplyWeight_Sycl(sycl::queue &sycl_queue, CeedInt num_elem, const CeedBasis_Sycl *impl, CeedScalar *w) { 246 const CeedInt dim = impl->dim; 247 const CeedInt Q_1d = impl->Q_1d; 248 const CeedScalar *q_weight_1d = impl->d_q_weight_1d; 249 250 const CeedInt num_quad_x = Q_1d; 251 const CeedInt num_quad_y = (dim > 1) ? Q_1d : 1; 252 const CeedInt num_quad_z = (dim > 2) ? Q_1d : 1; 253 sycl::range<3> kernel_range(num_elem * num_quad_z, num_quad_y, num_quad_x); 254 255 std::vector<sycl::event> e; 256 257 if (!sycl_queue.is_in_order()) e = {sycl_queue.ext_oneapi_submit_barrier()}; 258 259 sycl_queue.parallel_for<CeedBasisSyclWeight>(kernel_range, e, [=](sycl::item<3> work_item) { 260 if (dim == 1) w[work_item.get_linear_id()] = q_weight_1d[work_item[2]]; 261 if (dim == 2) w[work_item.get_linear_id()] = q_weight_1d[work_item[2]] * q_weight_1d[work_item[1]]; 262 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]; 263 }); 264 return CEED_ERROR_SUCCESS; 265 } 266 267 //------------------------------------------------------------------------------ 268 // Basis apply - tensor 269 //------------------------------------------------------------------------------ 270 static int CeedBasisApply_Sycl(CeedBasis basis, const CeedInt num_elem, CeedTransposeMode t_mode, CeedEvalMode eval_mode, CeedVector u, 271 CeedVector v) { 272 Ceed ceed; 273 const CeedInt is_transpose = t_mode == CEED_TRANSPOSE; 274 const CeedScalar *d_u; 275 CeedScalar *d_v; 276 Ceed_Sycl *data; 277 CeedBasis_Sycl *impl; 278 279 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 280 CeedCallBackend(CeedGetData(ceed, &data)); 281 CeedCallBackend(CeedBasisGetData(basis, &impl)); 282 283 // Get read/write access to u, v 284 if (u != CEED_VECTOR_NONE) CeedCallBackend(CeedVectorGetArrayRead(u, CEED_MEM_DEVICE, &d_u)); 285 else CeedCheck(eval_mode == CEED_EVAL_WEIGHT, ceed, CEED_ERROR_BACKEND, "An input vector is required for this CeedEvalMode"); 286 CeedCallBackend(CeedVectorGetArrayWrite(v, CEED_MEM_DEVICE, &d_v)); 287 288 // Clear v for transpose operation 289 if (is_transpose) { 290 CeedSize length; 291 CeedCallBackend(CeedVectorGetLength(v, &length)); 292 std::vector<sycl::event> e; 293 294 if (!data->sycl_queue.is_in_order()) e = {data->sycl_queue.ext_oneapi_submit_barrier()}; 295 data->sycl_queue.fill<CeedScalar>(d_v, 0, length, e); 296 } 297 298 // Basis action 299 switch (eval_mode) { 300 case CEED_EVAL_INTERP: 301 if (is_transpose) { 302 CeedCallBackend(CeedBasisApplyInterp_Sycl<true>(data->sycl_queue, *impl->sycl_module, num_elem, impl, d_u, d_v)); 303 } else { 304 CeedCallBackend(CeedBasisApplyInterp_Sycl<false>(data->sycl_queue, *impl->sycl_module, num_elem, impl, d_u, d_v)); 305 } 306 break; 307 case CEED_EVAL_GRAD: 308 if (is_transpose) { 309 CeedCallBackend(CeedBasisApplyGrad_Sycl<true>(data->sycl_queue, *impl->sycl_module, num_elem, impl, d_u, d_v)); 310 } else { 311 CeedCallBackend(CeedBasisApplyGrad_Sycl<false>(data->sycl_queue, *impl->sycl_module, num_elem, impl, d_u, d_v)); 312 } 313 break; 314 case CEED_EVAL_WEIGHT: 315 CeedCallBackend(CeedBasisApplyWeight_Sycl(data->sycl_queue, num_elem, impl, d_v)); 316 break; 317 case CEED_EVAL_NONE: /* handled separately below */ 318 break; 319 // LCOV_EXCL_START 320 case CEED_EVAL_DIV: 321 case CEED_EVAL_CURL: 322 return CeedError(ceed, CEED_ERROR_BACKEND, "%s not supported", CeedEvalModes[eval_mode]); 323 // LCOV_EXCL_STOP 324 } 325 326 // Restore vectors, cover CEED_EVAL_NONE 327 CeedCallBackend(CeedVectorRestoreArray(v, &d_v)); 328 if (eval_mode == CEED_EVAL_NONE) CeedCallBackend(CeedVectorSetArray(v, CEED_MEM_DEVICE, CEED_COPY_VALUES, (CeedScalar *)d_u)); 329 if (eval_mode != CEED_EVAL_WEIGHT) CeedCallBackend(CeedVectorRestoreArrayRead(u, &d_u)); 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 is_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 = is_transpose ? impl->num_qpts : impl->num_nodes; 340 const CeedInt Q = is_transpose ? impl->num_nodes : impl->num_qpts; 341 const CeedInt stride_0 = is_transpose ? 1 : impl->num_nodes; 342 const CeedInt stride_1 = is_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 std::vector<sycl::event> e; 354 355 if (!sycl_queue.is_in_order()) e = {sycl_queue.ext_oneapi_submit_barrier()}; 356 357 sycl_queue.parallel_for<CeedBasisSyclInterpNT>(kernel_range, e, [=](sycl::id<2> indx) { 358 const CeedInt i = indx[1]; 359 const CeedInt elem = indx[0]; 360 361 for (CeedInt comp = 0; comp < num_comp; comp++) { 362 const CeedScalar *U = d_U + elem * u_stride + comp * u_comp_stride; 363 CeedScalar V = 0.0; 364 365 for (CeedInt j = 0; j < u_size; ++j) { 366 V += d_B[i * stride_0 + j * stride_1] * U[j]; 367 } 368 d_V[i + elem * v_stride + comp * v_comp_stride] = V; 369 } 370 }); 371 return CEED_ERROR_SUCCESS; 372 } 373 374 //------------------------------------------------------------------------------ 375 // Gradient kernel - non-tensor 376 //------------------------------------------------------------------------------ 377 static int CeedBasisApplyNonTensorGrad_Sycl(sycl::queue &sycl_queue, CeedInt num_elem, CeedInt is_transpose, const CeedBasisNonTensor_Sycl *impl, 378 const CeedScalar *d_U, CeedScalar *d_V) { 379 const CeedInt num_comp = impl->num_comp; 380 const CeedInt P = is_transpose ? impl->num_qpts : impl->num_nodes; 381 const CeedInt Q = is_transpose ? impl->num_nodes : impl->num_qpts; 382 const CeedInt stride_0 = is_transpose ? 1 : impl->num_nodes; 383 const CeedInt stride_1 = is_transpose ? impl->num_nodes : 1; 384 const CeedInt g_dim_stride = P * Q; 385 const CeedInt u_stride = P; 386 const CeedInt v_stride = Q; 387 const CeedInt u_comp_stride = u_stride * num_elem; 388 const CeedInt v_comp_stride = v_stride * num_elem; 389 const CeedInt u_dim_stride = u_comp_stride * num_comp; 390 const CeedInt v_dim_stride = v_comp_stride * num_comp; 391 const CeedInt u_size = P; 392 const CeedInt v_size = Q; 393 const CeedInt in_dim = is_transpose ? impl->dim : 1; 394 const CeedInt out_dim = is_transpose ? 1 : impl->dim; 395 const CeedScalar *d_G = impl->d_grad; 396 397 sycl::range<2> kernel_range(num_elem, v_size); 398 399 std::vector<sycl::event> e; 400 401 if (!sycl_queue.is_in_order()) e = {sycl_queue.ext_oneapi_submit_barrier()}; 402 403 sycl_queue.parallel_for<CeedBasisSyclGradNT>(kernel_range, e, [=](sycl::id<2> indx) { 404 const CeedInt i = indx[1]; 405 const CeedInt elem = indx[0]; 406 407 for (CeedInt comp = 0; comp < num_comp; comp++) { 408 CeedScalar V[3] = {0.0, 0.0, 0.0}; 409 410 for (CeedInt d1 = 0; d1 < in_dim; ++d1) { 411 const CeedScalar *U = d_U + elem * u_stride + comp * u_comp_stride + d1 * u_dim_stride; 412 const CeedScalar *G = d_G + i * stride_0 + d1 * g_dim_stride; 413 414 for (CeedInt j = 0; j < u_size; ++j) { 415 const CeedScalar Uj = U[j]; 416 417 for (CeedInt d0 = 0; d0 < out_dim; ++d0) { 418 V[d0] += G[j * stride_1 + d0 * g_dim_stride] * Uj; 419 } 420 } 421 } 422 for (CeedInt d0 = 0; d0 < out_dim; ++d0) { 423 d_V[i + elem * v_stride + comp * v_comp_stride + d0 * v_dim_stride] = V[d0]; 424 } 425 } 426 }); 427 return CEED_ERROR_SUCCESS; 428 } 429 430 //------------------------------------------------------------------------------ 431 // Weight kernel - non-tensor 432 //------------------------------------------------------------------------------ 433 static int CeedBasisApplyNonTensorWeight_Sycl(sycl::queue &sycl_queue, CeedInt num_elem, const CeedBasisNonTensor_Sycl *impl, CeedScalar *d_V) { 434 const CeedInt num_qpts = impl->num_qpts; 435 const CeedScalar *q_weight = impl->d_q_weight; 436 437 sycl::range<2> kernel_range(num_elem, num_qpts); 438 439 std::vector<sycl::event> e; 440 441 if (!sycl_queue.is_in_order()) e = {sycl_queue.ext_oneapi_submit_barrier()}; 442 443 sycl_queue.parallel_for<CeedBasisSyclWeightNT>(kernel_range, e, [=](sycl::id<2> indx) { 444 const CeedInt i = indx[1]; 445 const CeedInt elem = indx[0]; 446 d_V[i + elem * num_qpts] = q_weight[i]; 447 }); 448 return CEED_ERROR_SUCCESS; 449 } 450 451 //------------------------------------------------------------------------------ 452 // Basis apply - non-tensor 453 //------------------------------------------------------------------------------ 454 static int CeedBasisApplyNonTensor_Sycl(CeedBasis basis, const CeedInt num_elem, CeedTransposeMode t_mode, CeedEvalMode eval_mode, CeedVector u, 455 CeedVector v) { 456 Ceed ceed; 457 const CeedInt is_transpose = t_mode == CEED_TRANSPOSE; 458 const CeedScalar *d_u; 459 CeedScalar *d_v; 460 CeedBasisNonTensor_Sycl *impl; 461 Ceed_Sycl *data; 462 463 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 464 CeedCallBackend(CeedBasisGetData(basis, &impl)); 465 CeedCallBackend(CeedGetData(ceed, &data)); 466 467 // Get read/write access to u, v 468 if (u != CEED_VECTOR_NONE) CeedCallBackend(CeedVectorGetArrayRead(u, CEED_MEM_DEVICE, &d_u)); 469 else CeedCheck(eval_mode == CEED_EVAL_WEIGHT, ceed, CEED_ERROR_BACKEND, "An input vector is required for this CeedEvalMode"); 470 CeedCallBackend(CeedVectorGetArrayWrite(v, CEED_MEM_DEVICE, &d_v)); 471 472 // Clear v for transpose operation 473 if (is_transpose) { 474 CeedSize length; 475 CeedCallBackend(CeedVectorGetLength(v, &length)); 476 // Order queue 477 sycl::event e = data->sycl_queue.ext_oneapi_submit_barrier(); 478 data->sycl_queue.fill<CeedScalar>(d_v, 0, length, {e}); 479 } 480 481 // Apply basis operation 482 switch (eval_mode) { 483 case CEED_EVAL_INTERP: 484 CeedCallBackend(CeedBasisApplyNonTensorInterp_Sycl(data->sycl_queue, num_elem, is_transpose, impl, d_u, d_v)); 485 break; 486 case CEED_EVAL_GRAD: 487 CeedCallBackend(CeedBasisApplyNonTensorGrad_Sycl(data->sycl_queue, num_elem, is_transpose, impl, d_u, d_v)); 488 break; 489 case CEED_EVAL_WEIGHT: 490 CeedCallBackend(CeedBasisApplyNonTensorWeight_Sycl(data->sycl_queue, num_elem, impl, d_v)); 491 break; 492 case CEED_EVAL_NONE: /* handled separately below */ 493 break; 494 // LCOV_EXCL_START 495 case CEED_EVAL_DIV: 496 case CEED_EVAL_CURL: 497 return CeedError(ceed, CEED_ERROR_BACKEND, "%s not supported", CeedEvalModes[eval_mode]); 498 // LCOV_EXCL_STOP 499 } 500 501 // Restore vectors, cover CEED_EVAL_NONE 502 CeedCallBackend(CeedVectorRestoreArray(v, &d_v)); 503 if (eval_mode == CEED_EVAL_NONE) CeedCallBackend(CeedVectorSetArray(v, CEED_MEM_DEVICE, CEED_COPY_VALUES, (CeedScalar *)d_u)); 504 if (eval_mode != CEED_EVAL_WEIGHT) CeedCallBackend(CeedVectorRestoreArrayRead(u, &d_u)); 505 506 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 return CEED_ERROR_SUCCESS; 529 } 530 531 //------------------------------------------------------------------------------ 532 // Destroy non-tensor basis 533 //------------------------------------------------------------------------------ 534 static int CeedBasisDestroyNonTensor_Sycl(CeedBasis basis) { 535 Ceed ceed; 536 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 537 CeedBasisNonTensor_Sycl *impl; 538 CeedCallBackend(CeedBasisGetData(basis, &impl)); 539 Ceed_Sycl *data; 540 CeedCallBackend(CeedGetData(ceed, &data)); 541 542 // Wait for all work to finish before freeing memory 543 CeedCallSycl(ceed, data->sycl_queue.wait_and_throw()); 544 545 CeedCallSycl(ceed, sycl::free(impl->d_q_weight, data->sycl_context)); 546 CeedCallSycl(ceed, sycl::free(impl->d_interp, data->sycl_context)); 547 CeedCallSycl(ceed, sycl::free(impl->d_grad, data->sycl_context)); 548 549 CeedCallBackend(CeedFree(&impl)); 550 return CEED_ERROR_SUCCESS; 551 } 552 553 //------------------------------------------------------------------------------ 554 // Create tensor 555 //------------------------------------------------------------------------------ 556 int CeedBasisCreateTensorH1_Sycl(CeedInt dim, CeedInt P_1d, CeedInt Q_1d, const CeedScalar *interp_1d, const CeedScalar *grad_1d, 557 const CeedScalar *q_ref_1d, const CeedScalar *q_weight_1d, CeedBasis basis) { 558 Ceed ceed; 559 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 560 CeedBasis_Sycl *impl; 561 CeedCallBackend(CeedCalloc(1, &impl)); 562 Ceed_Sycl *data; 563 CeedCallBackend(CeedGetData(ceed, &data)); 564 565 CeedInt num_comp; 566 CeedCallBackend(CeedBasisGetNumComponents(basis, &num_comp)); 567 568 const CeedInt num_nodes = CeedIntPow(P_1d, dim); 569 const CeedInt num_qpts = CeedIntPow(Q_1d, dim); 570 571 impl->dim = dim; 572 impl->P_1d = P_1d; 573 impl->Q_1d = Q_1d; 574 impl->num_comp = num_comp; 575 impl->num_nodes = num_nodes; 576 impl->num_qpts = num_qpts; 577 impl->buf_len = num_comp * CeedIntMax(num_nodes, num_qpts); 578 impl->op_len = Q_1d * P_1d; 579 580 std::vector<sycl::event> e; 581 582 if (!data->sycl_queue.is_in_order()) e = {data->sycl_queue.ext_oneapi_submit_barrier()}; 583 584 CeedCallSycl(ceed, impl->d_q_weight_1d = sycl::malloc_device<CeedScalar>(Q_1d, data->sycl_device, data->sycl_context)); 585 sycl::event copy_weight = data->sycl_queue.copy<CeedScalar>(q_weight_1d, impl->d_q_weight_1d, Q_1d, e); 586 587 const CeedInt interp_length = Q_1d * P_1d; 588 CeedCallSycl(ceed, impl->d_interp_1d = sycl::malloc_device<CeedScalar>(interp_length, data->sycl_device, data->sycl_context)); 589 sycl::event copy_interp = data->sycl_queue.copy<CeedScalar>(interp_1d, impl->d_interp_1d, interp_length, e); 590 591 CeedCallSycl(ceed, impl->d_grad_1d = sycl::malloc_device<CeedScalar>(interp_length, data->sycl_device, data->sycl_context)); 592 sycl::event copy_grad = data->sycl_queue.copy<CeedScalar>(grad_1d, impl->d_grad_1d, interp_length, e); 593 594 CeedCallSycl(ceed, sycl::event::wait_and_throw({copy_weight, copy_interp, copy_grad})); 595 596 std::vector<sycl::kernel_id> kernel_ids = {sycl::get_kernel_id<CeedBasisSyclInterp<1>>(), sycl::get_kernel_id<CeedBasisSyclInterp<0>>(), 597 sycl::get_kernel_id<CeedBasisSyclGrad<1>>(), sycl::get_kernel_id<CeedBasisSyclGrad<0>>()}; 598 599 sycl::kernel_bundle<sycl::bundle_state::input> input_bundle = sycl::get_kernel_bundle<sycl::bundle_state::input>(data->sycl_context, kernel_ids); 600 input_bundle.set_specialization_constant<BASIS_DIM_ID>(dim); 601 input_bundle.set_specialization_constant<BASIS_NUM_COMP_ID>(num_comp); 602 input_bundle.set_specialization_constant<BASIS_Q_1D_ID>(Q_1d); 603 input_bundle.set_specialization_constant<BASIS_P_1D_ID>(P_1d); 604 605 CeedCallSycl(ceed, impl->sycl_module = new SyclModule_t(sycl::build(input_bundle))); 606 607 CeedCallBackend(CeedBasisSetData(basis, impl)); 608 609 // Register backend functions 610 CeedCallBackend(CeedSetBackendFunctionCpp(ceed, "Basis", basis, "Apply", CeedBasisApply_Sycl)); 611 CeedCallBackend(CeedSetBackendFunctionCpp(ceed, "Basis", basis, "Destroy", CeedBasisDestroy_Sycl)); 612 return CEED_ERROR_SUCCESS; 613 } 614 615 //------------------------------------------------------------------------------ 616 // Create non-tensor 617 //------------------------------------------------------------------------------ 618 int CeedBasisCreateH1_Sycl(CeedElemTopology topo, CeedInt dim, CeedInt num_nodes, CeedInt num_qpts, const CeedScalar *interp, const CeedScalar *grad, 619 const CeedScalar *q_ref, const CeedScalar *q_weight, CeedBasis basis) { 620 Ceed ceed; 621 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 622 CeedBasisNonTensor_Sycl *impl; 623 CeedCallBackend(CeedCalloc(1, &impl)); 624 Ceed_Sycl *data; 625 CeedCallBackend(CeedGetData(ceed, &data)); 626 627 CeedInt num_comp; 628 CeedCallBackend(CeedBasisGetNumComponents(basis, &num_comp)); 629 630 impl->dim = dim; 631 impl->num_comp = num_comp; 632 impl->num_nodes = num_nodes; 633 impl->num_qpts = num_qpts; 634 635 std::vector<sycl::event> e; 636 637 if (!data->sycl_queue.is_in_order()) 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