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 <map> 13 #include <string_view> 14 #include <sycl/sycl.hpp> 15 16 #include "../sycl/ceed-sycl-compile.hpp" 17 #include "ceed-sycl-shared.hpp" 18 19 //------------------------------------------------------------------------------ 20 // Compute the local range of for basis kernels 21 //------------------------------------------------------------------------------ 22 static int ComputeLocalRange(Ceed ceed, CeedInt dim, CeedInt thread_1d, CeedInt *local_range, CeedInt max_group_size = 256) { 23 local_range[0] = thread_1d; 24 local_range[1] = (dim > 1) ? thread_1d : 1; 25 26 const CeedInt min_group_size = local_range[0] * local_range[1]; 27 CeedCheck(min_group_size <= max_group_size, ceed, CEED_ERROR_BACKEND, "Requested group size is smaller than the required minimum."); 28 29 local_range[2] = max_group_size / min_group_size; // elements per group 30 return CEED_ERROR_SUCCESS; 31 } 32 33 //------------------------------------------------------------------------------ 34 // Apply basis 35 //------------------------------------------------------------------------------ 36 int CeedBasisApplyTensor_Sycl_shared(CeedBasis basis, const CeedInt num_elem, CeedTransposeMode t_mode, CeedEvalMode eval_mode, CeedVector u, 37 CeedVector v) { 38 Ceed ceed; 39 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 40 Ceed_Sycl *ceed_Sycl; 41 CeedCallBackend(CeedGetData(ceed, &ceed_Sycl)); 42 CeedBasis_Sycl_shared *impl; 43 CeedCallBackend(CeedBasisGetData(basis, &impl)); 44 45 // Read vectors 46 const CeedScalar *d_u; 47 CeedScalar *d_v; 48 if (eval_mode != CEED_EVAL_WEIGHT) { 49 CeedCallBackend(CeedVectorGetArrayRead(u, CEED_MEM_DEVICE, &d_u)); 50 } 51 CeedCallBackend(CeedVectorGetArrayWrite(v, CEED_MEM_DEVICE, &d_v)); 52 53 // Apply basis operation 54 switch (eval_mode) { 55 case CEED_EVAL_INTERP: { 56 CeedInt *lrange = impl->interp_local_range; 57 const CeedInt &elem_per_group = lrange[2]; 58 const CeedInt group_count = (num_elem / elem_per_group) + !!(num_elem % elem_per_group); 59 //----------- 60 sycl::range<3> local_range(lrange[2], lrange[1], lrange[0]); 61 sycl::range<3> global_range(group_count * lrange[2], lrange[1], lrange[0]); 62 sycl::nd_range<3> kernel_range(global_range, local_range); 63 //----------- 64 sycl::kernel *interp_kernel = (t_mode == CEED_TRANSPOSE) ? impl->interp_transpose_kernel : impl->interp_kernel; 65 66 // Order queue 67 sycl::event e = ceed_Sycl->sycl_queue.ext_oneapi_submit_barrier(); 68 69 ceed_Sycl->sycl_queue.submit([&](sycl::handler &cgh) { 70 cgh.depends_on(e); 71 cgh.set_args(num_elem, impl->d_interp_1d, d_u, d_v); 72 cgh.parallel_for(kernel_range, *interp_kernel); 73 }); 74 75 } break; 76 case CEED_EVAL_GRAD: { 77 CeedInt *lrange = impl->grad_local_range; 78 const CeedInt &elem_per_group = lrange[2]; 79 const CeedInt group_count = (num_elem / elem_per_group) + !!(num_elem % elem_per_group); 80 //----------- 81 sycl::range<3> local_range(lrange[2], lrange[1], lrange[0]); 82 sycl::range<3> global_range(group_count * lrange[2], lrange[1], lrange[0]); 83 sycl::nd_range<3> kernel_range(global_range, local_range); 84 //----------- 85 sycl::kernel *grad_kernel = (t_mode == CEED_TRANSPOSE) ? impl->grad_transpose_kernel : impl->grad_kernel; 86 const CeedScalar *d_grad_1d = (impl->d_collo_grad_1d) ? impl->d_collo_grad_1d : impl->d_grad_1d; 87 // Order queue 88 sycl::event e = ceed_Sycl->sycl_queue.ext_oneapi_submit_barrier(); 89 90 ceed_Sycl->sycl_queue.submit([&](sycl::handler &cgh) { 91 cgh.depends_on(e); 92 cgh.set_args(num_elem, impl->d_interp_1d, d_grad_1d, d_u, d_v); 93 cgh.parallel_for(kernel_range, *grad_kernel); 94 }); 95 } break; 96 case CEED_EVAL_WEIGHT: { 97 CeedInt *lrange = impl->weight_local_range; 98 const CeedInt &elem_per_group = lrange[2]; 99 const CeedInt group_count = (num_elem / elem_per_group) + !!(num_elem % elem_per_group); 100 //----------- 101 sycl::range<3> local_range(lrange[2], lrange[1], lrange[0]); 102 sycl::range<3> global_range(group_count * lrange[2], lrange[1], lrange[0]); 103 sycl::nd_range<3> kernel_range(global_range, local_range); 104 //----------- 105 // Order queue 106 sycl::event e = ceed_Sycl->sycl_queue.ext_oneapi_submit_barrier(); 107 108 ceed_Sycl->sycl_queue.submit([&](sycl::handler &cgh) { 109 cgh.depends_on(e); 110 cgh.set_args(num_elem, impl->d_q_weight_1d, d_v); 111 cgh.parallel_for(kernel_range, *(impl->weight_kernel)); 112 }); 113 } break; 114 // LCOV_EXCL_START 115 // Evaluate the divergence to/from the quadrature points 116 case CEED_EVAL_DIV: 117 return CeedError(ceed, CEED_ERROR_BACKEND, "CEED_EVAL_DIV not supported"); 118 // Evaluate the curl to/from the quadrature points 119 case CEED_EVAL_CURL: 120 return CeedError(ceed, CEED_ERROR_BACKEND, "CEED_EVAL_CURL not supported"); 121 // Take no action, BasisApply should not have been called 122 case CEED_EVAL_NONE: 123 return CeedError(ceed, CEED_ERROR_BACKEND, "CEED_EVAL_NONE does not make sense in this context"); 124 // LCOV_EXCL_STOP 125 } 126 127 // Restore vectors 128 if (eval_mode != CEED_EVAL_WEIGHT) { 129 CeedCallBackend(CeedVectorRestoreArrayRead(u, &d_u)); 130 } 131 CeedCallBackend(CeedVectorRestoreArray(v, &d_v)); 132 return CEED_ERROR_SUCCESS; 133 } 134 135 //------------------------------------------------------------------------------ 136 // Destroy basis 137 //------------------------------------------------------------------------------ 138 static int CeedBasisDestroy_Sycl_shared(CeedBasis basis) { 139 Ceed ceed; 140 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 141 CeedBasis_Sycl_shared *impl; 142 CeedCallBackend(CeedBasisGetData(basis, &impl)); 143 Ceed_Sycl *data; 144 CeedCallBackend(CeedGetData(ceed, &data)); 145 146 CeedCallSycl(ceed, data->sycl_queue.wait_and_throw()); 147 CeedCallSycl(ceed, sycl::free(impl->d_q_weight_1d, data->sycl_context)); 148 CeedCallSycl(ceed, sycl::free(impl->d_interp_1d, data->sycl_context)); 149 CeedCallSycl(ceed, sycl::free(impl->d_grad_1d, data->sycl_context)); 150 CeedCallSycl(ceed, sycl::free(impl->d_collo_grad_1d, data->sycl_context)); 151 152 delete impl->interp_kernel; 153 delete impl->interp_transpose_kernel; 154 delete impl->grad_kernel; 155 delete impl->grad_transpose_kernel; 156 delete impl->weight_kernel; 157 delete impl->sycl_module; 158 159 CeedCallBackend(CeedFree(&impl)); 160 161 return CEED_ERROR_SUCCESS; 162 } 163 164 //------------------------------------------------------------------------------ 165 // Create tensor basis 166 // TODO: Refactor 167 //------------------------------------------------------------------------------ 168 int CeedBasisCreateTensorH1_Sycl_shared(CeedInt dim, CeedInt P_1d, CeedInt Q_1d, const CeedScalar *interp_1d, const CeedScalar *grad_1d, 169 const CeedScalar *q_ref_1d, const CeedScalar *q_weight_1d, CeedBasis basis) { 170 Ceed ceed; 171 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 172 CeedBasis_Sycl_shared *impl; 173 CeedCallBackend(CeedCalloc(1, &impl)); 174 Ceed_Sycl *data; 175 CeedCallBackend(CeedGetData(ceed, &data)); 176 177 CeedInt num_comp; 178 CeedCallBackend(CeedBasisGetNumComponents(basis, &num_comp)); 179 180 const CeedInt thread_1d = CeedIntMax(Q_1d, P_1d); 181 const CeedInt num_nodes = CeedIntPow(P_1d, dim); 182 const CeedInt num_qpts = CeedIntPow(Q_1d, dim); 183 184 CeedInt *interp_lrange = impl->interp_local_range; 185 CeedCallBackend(ComputeLocalRange(ceed, dim, thread_1d, interp_lrange)); 186 const CeedInt interp_group_size = interp_lrange[0] * interp_lrange[1] * interp_lrange[2]; 187 188 CeedInt *grad_lrange = impl->grad_local_range; 189 CeedCallBackend(ComputeLocalRange(ceed, dim, thread_1d, grad_lrange)); 190 const CeedInt grad_group_size = grad_lrange[0] * grad_lrange[1] * grad_lrange[2]; 191 192 CeedCallBackend(ComputeLocalRange(ceed, dim, Q_1d, impl->weight_local_range)); 193 194 // Copy basis data to GPU 195 CeedCallSycl(ceed, impl->d_q_weight_1d = sycl::malloc_device<CeedScalar>(Q_1d, data->sycl_device, data->sycl_context)); 196 sycl::event copy_weight = data->sycl_queue.copy<CeedScalar>(q_weight_1d, impl->d_q_weight_1d, Q_1d); 197 198 const CeedInt interp_length = Q_1d * P_1d; 199 CeedCallSycl(ceed, impl->d_interp_1d = sycl::malloc_device<CeedScalar>(interp_length, data->sycl_device, data->sycl_context)); 200 sycl::event copy_interp = data->sycl_queue.copy<CeedScalar>(interp_1d, impl->d_interp_1d, interp_length); 201 202 CeedCallSycl(ceed, impl->d_grad_1d = sycl::malloc_device<CeedScalar>(interp_length, data->sycl_device, data->sycl_context)); 203 sycl::event copy_grad = data->sycl_queue.copy<CeedScalar>(grad_1d, impl->d_grad_1d, interp_length); 204 205 CeedCallSycl(ceed, sycl::event::wait_and_throw({copy_weight, copy_interp, copy_grad})); 206 207 // Compute collocated gradient and copy to GPU 208 impl->d_collo_grad_1d = NULL; 209 const bool has_collocated_grad = (dim == 3) && (Q_1d >= P_1d); 210 if (has_collocated_grad) { 211 CeedScalar *collo_grad_1d; 212 CeedCallBackend(CeedMalloc(Q_1d * Q_1d, &collo_grad_1d)); 213 CeedCallBackend(CeedBasisGetCollocatedGrad(basis, collo_grad_1d)); 214 const CeedInt cgrad_length = Q_1d * Q_1d; 215 CeedCallSycl(ceed, impl->d_collo_grad_1d = sycl::malloc_device<CeedScalar>(cgrad_length, data->sycl_device, data->sycl_context)); 216 CeedCallSycl(ceed, data->sycl_queue.copy<CeedScalar>(collo_grad_1d, impl->d_collo_grad_1d, cgrad_length).wait_and_throw()); 217 CeedCallBackend(CeedFree(&collo_grad_1d)); 218 } 219 220 // ---[Refactor into separate function]------> 221 // Define compile-time constants 222 std::map<std::string, CeedInt> jit_constants; 223 jit_constants["BASIS_DIM"] = dim; 224 jit_constants["BASIS_Q_1D"] = Q_1d; 225 jit_constants["BASIS_P_1D"] = P_1d; 226 jit_constants["T_1D"] = thread_1d; 227 jit_constants["BASIS_NUM_COMP"] = num_comp; 228 jit_constants["BASIS_NUM_NODES"] = num_nodes; 229 jit_constants["BASIS_NUM_QPTS"] = num_qpts; 230 jit_constants["BASIS_HAS_COLLOCATED_GRAD"] = has_collocated_grad; 231 jit_constants["BASIS_INTERP_SCRATCH_SIZE"] = interp_group_size; 232 jit_constants["BASIS_GRAD_SCRATCH_SIZE"] = grad_group_size; 233 234 // Load kernel source 235 char *basis_kernel_path, *basis_kernel_source; 236 CeedCallBackend(CeedGetJitAbsolutePath(ceed, "ceed/jit-source/sycl/sycl-shared-basis-tensor.h", &basis_kernel_path)); 237 CeedDebug256(ceed, 2, "----- Loading Basis Kernel Source -----\n"); 238 CeedCallBackend(CeedLoadSourceToBuffer(ceed, basis_kernel_path, &basis_kernel_source)); 239 CeedDebug256(ceed, 2, "----- Loading Basis Kernel Source Complete -----\n"); 240 241 // Compile kernels into a kernel bundle 242 CeedCallBackend(CeedJitBuildModule_Sycl(ceed, basis_kernel_source, &impl->sycl_module, jit_constants)); 243 244 // Load kernel functions 245 CeedCallBackend(CeedJitGetKernel_Sycl(ceed, impl->sycl_module, "Interp", &impl->interp_kernel)); 246 CeedCallBackend(CeedJitGetKernel_Sycl(ceed, impl->sycl_module, "InterpTranspose", &impl->interp_transpose_kernel)); 247 CeedCallBackend(CeedJitGetKernel_Sycl(ceed, impl->sycl_module, "Grad", &impl->grad_kernel)); 248 CeedCallBackend(CeedJitGetKernel_Sycl(ceed, impl->sycl_module, "GradTranspose", &impl->grad_transpose_kernel)); 249 CeedCallBackend(CeedJitGetKernel_Sycl(ceed, impl->sycl_module, "Weight", &impl->weight_kernel)); 250 251 // Clean-up 252 CeedCallBackend(CeedFree(&basis_kernel_path)); 253 CeedCallBackend(CeedFree(&basis_kernel_source)); 254 // <---[Refactor into separate function]------ 255 256 CeedCallBackend(CeedBasisSetData(basis, impl)); 257 258 // Register backend functions 259 CeedCallBackend(CeedSetBackendFunctionCpp(ceed, "Basis", basis, "Apply", CeedBasisApplyTensor_Sycl_shared)); 260 CeedCallBackend(CeedSetBackendFunctionCpp(ceed, "Basis", basis, "Destroy", CeedBasisDestroy_Sycl_shared)); 261 return CEED_ERROR_SUCCESS; 262 } 263 //------------------------------------------------------------------------------ 264