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 const CeedInt min_group_size = local_range[0] * local_range[1]; 26 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 Ceed_Sycl *ceed_Sycl; 40 const CeedScalar *d_u; 41 CeedScalar *d_v; 42 CeedBasis_Sycl_shared *impl; 43 44 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 45 CeedCallBackend(CeedGetData(ceed, &ceed_Sycl)); 46 CeedCallBackend(CeedBasisGetData(basis, &impl)); 47 48 // Read vectors 49 if (eval_mode != CEED_EVAL_WEIGHT) { 50 CeedCallBackend(CeedVectorGetArrayRead(u, CEED_MEM_DEVICE, &d_u)); 51 } 52 CeedCallBackend(CeedVectorGetArrayWrite(v, CEED_MEM_DEVICE, &d_v)); 53 54 // Apply basis operation 55 switch (eval_mode) { 56 case CEED_EVAL_INTERP: { 57 CeedInt *lrange = impl->interp_local_range; 58 const CeedInt &elem_per_group = lrange[2]; 59 const CeedInt group_count = (num_elem / elem_per_group) + !!(num_elem % elem_per_group); 60 //----------- 61 sycl::range<3> local_range(lrange[2], lrange[1], lrange[0]); 62 sycl::range<3> global_range(group_count * lrange[2], lrange[1], lrange[0]); 63 sycl::nd_range<3> kernel_range(global_range, local_range); 64 //----------- 65 sycl::kernel *interp_kernel = (t_mode == CEED_TRANSPOSE) ? impl->interp_transpose_kernel : impl->interp_kernel; 66 67 // Order queue 68 sycl::event e = ceed_Sycl->sycl_queue.ext_oneapi_submit_barrier(); 69 70 ceed_Sycl->sycl_queue.submit([&](sycl::handler &cgh) { 71 cgh.depends_on(e); 72 cgh.set_args(num_elem, impl->d_interp_1d, d_u, d_v); 73 cgh.parallel_for(kernel_range, *interp_kernel); 74 }); 75 76 } break; 77 case CEED_EVAL_GRAD: { 78 CeedInt *lrange = impl->grad_local_range; 79 const CeedInt &elem_per_group = lrange[2]; 80 const CeedInt group_count = (num_elem / elem_per_group) + !!(num_elem % elem_per_group); 81 //----------- 82 sycl::range<3> local_range(lrange[2], lrange[1], lrange[0]); 83 sycl::range<3> global_range(group_count * lrange[2], lrange[1], lrange[0]); 84 sycl::nd_range<3> kernel_range(global_range, local_range); 85 //----------- 86 sycl::kernel *grad_kernel = (t_mode == CEED_TRANSPOSE) ? impl->grad_transpose_kernel : impl->grad_kernel; 87 const CeedScalar *d_grad_1d = (impl->d_collo_grad_1d) ? impl->d_collo_grad_1d : impl->d_grad_1d; 88 // Order queue 89 sycl::event e = ceed_Sycl->sycl_queue.ext_oneapi_submit_barrier(); 90 91 ceed_Sycl->sycl_queue.submit([&](sycl::handler &cgh) { 92 cgh.depends_on(e); 93 cgh.set_args(num_elem, impl->d_interp_1d, d_grad_1d, d_u, d_v); 94 cgh.parallel_for(kernel_range, *grad_kernel); 95 }); 96 } break; 97 case CEED_EVAL_WEIGHT: { 98 CeedInt *lrange = impl->weight_local_range; 99 const CeedInt &elem_per_group = lrange[2]; 100 const CeedInt group_count = (num_elem / elem_per_group) + !!(num_elem % elem_per_group); 101 //----------- 102 sycl::range<3> local_range(lrange[2], lrange[1], lrange[0]); 103 sycl::range<3> global_range(group_count * lrange[2], lrange[1], lrange[0]); 104 sycl::nd_range<3> kernel_range(global_range, local_range); 105 //----------- 106 // Order queue 107 sycl::event e = ceed_Sycl->sycl_queue.ext_oneapi_submit_barrier(); 108 109 ceed_Sycl->sycl_queue.submit([&](sycl::handler &cgh) { 110 cgh.depends_on(e); 111 cgh.set_args(num_elem, impl->d_q_weight_1d, d_v); 112 cgh.parallel_for(kernel_range, *(impl->weight_kernel)); 113 }); 114 } break; 115 // LCOV_EXCL_START 116 // Evaluate the divergence to/from the quadrature points 117 case CEED_EVAL_DIV: 118 return CeedError(ceed, CEED_ERROR_BACKEND, "CEED_EVAL_DIV not supported"); 119 // Evaluate the curl to/from the quadrature points 120 case CEED_EVAL_CURL: 121 return CeedError(ceed, CEED_ERROR_BACKEND, "CEED_EVAL_CURL not supported"); 122 // Take no action, BasisApply should not have been called 123 case CEED_EVAL_NONE: 124 return CeedError(ceed, CEED_ERROR_BACKEND, "CEED_EVAL_NONE does not make sense in this context"); 125 // LCOV_EXCL_STOP 126 } 127 128 // Restore vectors 129 if (eval_mode != CEED_EVAL_WEIGHT) { 130 CeedCallBackend(CeedVectorRestoreArrayRead(u, &d_u)); 131 } 132 CeedCallBackend(CeedVectorRestoreArray(v, &d_v)); 133 return CEED_ERROR_SUCCESS; 134 } 135 136 //------------------------------------------------------------------------------ 137 // Destroy basis 138 //------------------------------------------------------------------------------ 139 static int CeedBasisDestroy_Sycl_shared(CeedBasis basis) { 140 Ceed ceed; 141 Ceed_Sycl *data; 142 CeedBasis_Sycl_shared *impl; 143 144 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 145 CeedCallBackend(CeedBasisGetData(basis, &impl)); 146 CeedCallBackend(CeedGetData(ceed, &data)); 147 CeedCallSycl(ceed, data->sycl_queue.wait_and_throw()); 148 CeedCallSycl(ceed, sycl::free(impl->d_q_weight_1d, data->sycl_context)); 149 CeedCallSycl(ceed, sycl::free(impl->d_interp_1d, data->sycl_context)); 150 CeedCallSycl(ceed, sycl::free(impl->d_grad_1d, data->sycl_context)); 151 CeedCallSycl(ceed, sycl::free(impl->d_collo_grad_1d, data->sycl_context)); 152 153 delete impl->interp_kernel; 154 delete impl->interp_transpose_kernel; 155 delete impl->grad_kernel; 156 delete impl->grad_transpose_kernel; 157 delete impl->weight_kernel; 158 delete impl->sycl_module; 159 160 CeedCallBackend(CeedFree(&impl)); 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 Ceed_Sycl *data; 172 char *basis_kernel_path, *basis_kernel_source; 173 CeedInt num_comp; 174 CeedBasis_Sycl_shared *impl; 175 176 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 177 CeedCallBackend(CeedCalloc(1, &impl)); 178 CeedCallBackend(CeedGetData(ceed, &data)); 179 CeedCallBackend(CeedBasisGetNumComponents(basis, &num_comp)); 180 181 const CeedInt thread_1d = CeedIntMax(Q_1d, P_1d); 182 const CeedInt num_nodes = CeedIntPow(P_1d, dim); 183 const CeedInt num_qpts = CeedIntPow(Q_1d, dim); 184 185 CeedInt *interp_lrange = impl->interp_local_range; 186 187 CeedCallBackend(ComputeLocalRange(ceed, dim, thread_1d, interp_lrange)); 188 const CeedInt interp_group_size = interp_lrange[0] * interp_lrange[1] * interp_lrange[2]; 189 190 CeedInt *grad_lrange = impl->grad_local_range; 191 192 CeedCallBackend(ComputeLocalRange(ceed, dim, thread_1d, grad_lrange)); 193 const CeedInt grad_group_size = grad_lrange[0] * grad_lrange[1] * grad_lrange[2]; 194 195 CeedCallBackend(ComputeLocalRange(ceed, dim, Q_1d, impl->weight_local_range)); 196 197 // Copy basis data to GPU 198 CeedCallSycl(ceed, impl->d_q_weight_1d = sycl::malloc_device<CeedScalar>(Q_1d, data->sycl_device, data->sycl_context)); 199 sycl::event copy_weight = data->sycl_queue.copy<CeedScalar>(q_weight_1d, impl->d_q_weight_1d, Q_1d); 200 201 const CeedInt interp_length = Q_1d * P_1d; 202 CeedCallSycl(ceed, impl->d_interp_1d = sycl::malloc_device<CeedScalar>(interp_length, data->sycl_device, data->sycl_context)); 203 sycl::event copy_interp = data->sycl_queue.copy<CeedScalar>(interp_1d, impl->d_interp_1d, interp_length); 204 205 CeedCallSycl(ceed, impl->d_grad_1d = sycl::malloc_device<CeedScalar>(interp_length, data->sycl_device, data->sycl_context)); 206 sycl::event copy_grad = data->sycl_queue.copy<CeedScalar>(grad_1d, impl->d_grad_1d, interp_length); 207 208 CeedCallSycl(ceed, sycl::event::wait_and_throw({copy_weight, copy_interp, copy_grad})); 209 210 // Compute collocated gradient and copy to GPU 211 impl->d_collo_grad_1d = NULL; 212 const bool has_collocated_grad = (dim == 3) && (Q_1d >= P_1d); 213 214 if (has_collocated_grad) { 215 CeedScalar *collo_grad_1d; 216 const CeedInt cgrad_length = Q_1d * Q_1d; 217 218 CeedCallBackend(CeedMalloc(Q_1d * Q_1d, &collo_grad_1d)); 219 CeedCallBackend(CeedBasisGetCollocatedGrad(basis, collo_grad_1d)); 220 CeedCallSycl(ceed, impl->d_collo_grad_1d = sycl::malloc_device<CeedScalar>(cgrad_length, data->sycl_device, data->sycl_context)); 221 CeedCallSycl(ceed, data->sycl_queue.copy<CeedScalar>(collo_grad_1d, impl->d_collo_grad_1d, cgrad_length).wait_and_throw()); 222 CeedCallBackend(CeedFree(&collo_grad_1d)); 223 } 224 225 // ---[Refactor into separate function]------> 226 // Define compile-time constants 227 std::map<std::string, CeedInt> jit_constants; 228 jit_constants["BASIS_DIM"] = dim; 229 jit_constants["BASIS_Q_1D"] = Q_1d; 230 jit_constants["BASIS_P_1D"] = P_1d; 231 jit_constants["T_1D"] = thread_1d; 232 jit_constants["BASIS_NUM_COMP"] = num_comp; 233 jit_constants["BASIS_NUM_NODES"] = num_nodes; 234 jit_constants["BASIS_NUM_QPTS"] = num_qpts; 235 jit_constants["BASIS_HAS_COLLOCATED_GRAD"] = has_collocated_grad; 236 jit_constants["BASIS_INTERP_SCRATCH_SIZE"] = interp_group_size; 237 jit_constants["BASIS_GRAD_SCRATCH_SIZE"] = grad_group_size; 238 239 // Load kernel source 240 CeedCallBackend(CeedGetJitAbsolutePath(ceed, "ceed/jit-source/sycl/sycl-shared-basis-tensor.h", &basis_kernel_path)); 241 CeedDebug256(ceed, CEED_DEBUG_COLOR_SUCCESS, "----- Loading Basis Kernel Source -----\n"); 242 CeedCallBackend(CeedLoadSourceToBuffer(ceed, basis_kernel_path, &basis_kernel_source)); 243 CeedDebug256(ceed, CEED_DEBUG_COLOR_SUCCESS, "----- Loading Basis Kernel Source Complete -----\n"); 244 245 // Compile kernels into a kernel bundle 246 CeedCallBackend(CeedBuildModule_Sycl(ceed, basis_kernel_source, &impl->sycl_module, jit_constants)); 247 248 // Load kernel functions 249 CeedCallBackend(CeedGetKernel_Sycl(ceed, impl->sycl_module, "Interp", &impl->interp_kernel)); 250 CeedCallBackend(CeedGetKernel_Sycl(ceed, impl->sycl_module, "InterpTranspose", &impl->interp_transpose_kernel)); 251 CeedCallBackend(CeedGetKernel_Sycl(ceed, impl->sycl_module, "Grad", &impl->grad_kernel)); 252 CeedCallBackend(CeedGetKernel_Sycl(ceed, impl->sycl_module, "GradTranspose", &impl->grad_transpose_kernel)); 253 CeedCallBackend(CeedGetKernel_Sycl(ceed, impl->sycl_module, "Weight", &impl->weight_kernel)); 254 255 // Clean-up 256 CeedCallBackend(CeedFree(&basis_kernel_path)); 257 CeedCallBackend(CeedFree(&basis_kernel_source)); 258 // <---[Refactor into separate function]------ 259 260 CeedCallBackend(CeedBasisSetData(basis, impl)); 261 262 // Register backend functions 263 CeedCallBackend(CeedSetBackendFunctionCpp(ceed, "Basis", basis, "Apply", CeedBasisApplyTensor_Sycl_shared)); 264 CeedCallBackend(CeedSetBackendFunctionCpp(ceed, "Basis", basis, "Destroy", CeedBasisDestroy_Sycl_shared)); 265 return CEED_ERROR_SUCCESS; 266 } 267 268 //------------------------------------------------------------------------------ 269