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.h> 9 #include <ceed/backend.h> 10 #include <ceed/jit-tools.h> 11 #include <cuda.h> 12 #include <cuda_runtime.h> 13 14 #include "../cuda/ceed-cuda-common.h" 15 #include "../cuda/ceed-cuda-compile.h" 16 #include "ceed-cuda-ref.h" 17 18 //------------------------------------------------------------------------------ 19 // Basis apply - tensor 20 //------------------------------------------------------------------------------ 21 int CeedBasisApply_Cuda(CeedBasis basis, const CeedInt num_elem, CeedTransposeMode t_mode, CeedEvalMode eval_mode, CeedVector u, CeedVector v) { 22 Ceed ceed; 23 CeedInt Q_1d, dim; 24 const CeedInt is_transpose = t_mode == CEED_TRANSPOSE; 25 const int max_block_size = 32; 26 const CeedScalar *d_u; 27 CeedScalar *d_v; 28 CeedBasis_Cuda *data; 29 30 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 31 CeedCallBackend(CeedBasisGetData(basis, &data)); 32 33 // Get read/write access to u, v 34 if (u != CEED_VECTOR_NONE) CeedCallBackend(CeedVectorGetArrayRead(u, CEED_MEM_DEVICE, &d_u)); 35 else CeedCheck(eval_mode == CEED_EVAL_WEIGHT, ceed, CEED_ERROR_BACKEND, "An input vector is required for this CeedEvalMode"); 36 CeedCallBackend(CeedVectorGetArrayWrite(v, CEED_MEM_DEVICE, &d_v)); 37 38 // Clear v for transpose operation 39 if (is_transpose) { 40 CeedSize length; 41 42 CeedCallBackend(CeedVectorGetLength(v, &length)); 43 CeedCallCuda(ceed, cudaMemset(d_v, 0, length * sizeof(CeedScalar))); 44 } 45 CeedCallBackend(CeedBasisGetNumQuadraturePoints1D(basis, &Q_1d)); 46 CeedCallBackend(CeedBasisGetDimension(basis, &dim)); 47 48 // Basis action 49 switch (eval_mode) { 50 case CEED_EVAL_INTERP: { 51 void *interp_args[] = {(void *)&num_elem, (void *)&is_transpose, &data->d_interp_1d, &d_u, &d_v}; 52 const CeedInt block_size = CeedIntMin(CeedIntPow(Q_1d, dim), max_block_size); 53 54 CeedCallBackend(CeedRunKernel_Cuda(ceed, data->Interp, num_elem, block_size, interp_args)); 55 } break; 56 case CEED_EVAL_GRAD: { 57 void *grad_args[] = {(void *)&num_elem, (void *)&is_transpose, &data->d_interp_1d, &data->d_grad_1d, &d_u, &d_v}; 58 const CeedInt block_size = max_block_size; 59 60 CeedCallBackend(CeedRunKernel_Cuda(ceed, data->Grad, num_elem, block_size, grad_args)); 61 } break; 62 case CEED_EVAL_WEIGHT: { 63 void *weight_args[] = {(void *)&num_elem, (void *)&data->d_q_weight_1d, &d_v}; 64 const int block_size_x = Q_1d; 65 const int block_size_y = dim >= 2 ? Q_1d : 1; 66 67 CeedCallBackend(CeedRunKernelDim_Cuda(ceed, data->Weight, num_elem, block_size_x, block_size_y, 1, weight_args)); 68 } break; 69 case CEED_EVAL_NONE: /* handled separately below */ 70 break; 71 // LCOV_EXCL_START 72 case CEED_EVAL_DIV: 73 case CEED_EVAL_CURL: 74 return CeedError(ceed, CEED_ERROR_BACKEND, "%s not supported", CeedEvalModes[eval_mode]); 75 // LCOV_EXCL_STOP 76 } 77 78 // Restore vectors, cover CEED_EVAL_NONE 79 CeedCallBackend(CeedVectorRestoreArray(v, &d_v)); 80 if (eval_mode == CEED_EVAL_NONE) CeedCallBackend(CeedVectorSetArray(v, CEED_MEM_DEVICE, CEED_COPY_VALUES, (CeedScalar *)d_u)); 81 if (eval_mode != CEED_EVAL_WEIGHT) CeedCallBackend(CeedVectorRestoreArrayRead(u, &d_u)); 82 return CEED_ERROR_SUCCESS; 83 } 84 85 //------------------------------------------------------------------------------ 86 // Basis apply - non-tensor 87 //------------------------------------------------------------------------------ 88 int CeedBasisApplyNonTensor_Cuda(CeedBasis basis, const CeedInt num_elem, CeedTransposeMode t_mode, CeedEvalMode eval_mode, CeedVector u, 89 CeedVector v) { 90 Ceed ceed; 91 CeedInt num_nodes, num_qpts; 92 const CeedInt is_transpose = t_mode == CEED_TRANSPOSE; 93 const int elems_per_block = 1; 94 const int grid = CeedDivUpInt(num_elem, elems_per_block); 95 const CeedScalar *d_u; 96 CeedScalar *d_v; 97 CeedBasisNonTensor_Cuda *data; 98 99 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 100 CeedCallBackend(CeedBasisGetData(basis, &data)); 101 CeedCallBackend(CeedBasisGetNumQuadraturePoints(basis, &num_qpts)); 102 CeedCallBackend(CeedBasisGetNumNodes(basis, &num_nodes)); 103 104 // Get read/write access to u, v 105 if (u != CEED_VECTOR_NONE) CeedCallBackend(CeedVectorGetArrayRead(u, CEED_MEM_DEVICE, &d_u)); 106 else CeedCheck(eval_mode == CEED_EVAL_WEIGHT, ceed, CEED_ERROR_BACKEND, "An input vector is required for this CeedEvalMode"); 107 CeedCallBackend(CeedVectorGetArrayWrite(v, CEED_MEM_DEVICE, &d_v)); 108 109 // Clear v for transpose operation 110 if (is_transpose) { 111 CeedSize length; 112 113 CeedCallBackend(CeedVectorGetLength(v, &length)); 114 CeedCallCuda(ceed, cudaMemset(d_v, 0, length * sizeof(CeedScalar))); 115 } 116 117 // Apply basis operation 118 switch (eval_mode) { 119 case CEED_EVAL_INTERP: { 120 void *interp_args[] = {(void *)&num_elem, &data->d_interp, &d_u, &d_v}; 121 const int block_size_x = is_transpose ? num_nodes : num_qpts; 122 123 if (is_transpose) { 124 CeedCallBackend(CeedRunKernelDim_Cuda(ceed, data->InterpTranspose, grid, block_size_x, 1, elems_per_block, interp_args)); 125 } else { 126 CeedCallBackend(CeedRunKernelDim_Cuda(ceed, data->Interp, grid, block_size_x, 1, elems_per_block, interp_args)); 127 } 128 } break; 129 case CEED_EVAL_GRAD: { 130 void *grad_args[] = {(void *)&num_elem, &data->d_grad, &d_u, &d_v}; 131 const int block_size_x = is_transpose ? num_nodes : num_qpts; 132 133 if (is_transpose) { 134 CeedCallBackend(CeedRunKernelDim_Cuda(ceed, data->DerivTranspose, grid, block_size_x, 1, elems_per_block, grad_args)); 135 } else { 136 CeedCallBackend(CeedRunKernelDim_Cuda(ceed, data->Deriv, grid, block_size_x, 1, elems_per_block, grad_args)); 137 } 138 } break; 139 case CEED_EVAL_DIV: { 140 void *div_args[] = {(void *)&num_elem, &data->d_div, &d_u, &d_v}; 141 const int block_size_x = is_transpose ? num_nodes : num_qpts; 142 143 if (is_transpose) { 144 CeedCallBackend(CeedRunKernelDim_Cuda(ceed, data->DerivTranspose, grid, block_size_x, 1, elems_per_block, div_args)); 145 } else { 146 CeedCallBackend(CeedRunKernelDim_Cuda(ceed, data->Deriv, grid, block_size_x, 1, elems_per_block, div_args)); 147 } 148 } break; 149 case CEED_EVAL_CURL: { 150 void *curl_args[] = {(void *)&num_elem, &data->d_curl, &d_u, &d_v}; 151 const int block_size_x = is_transpose ? num_nodes : num_qpts; 152 153 if (is_transpose) { 154 CeedCallBackend(CeedRunKernelDim_Cuda(ceed, data->DerivTranspose, grid, block_size_x, 1, elems_per_block, curl_args)); 155 } else { 156 CeedCallBackend(CeedRunKernelDim_Cuda(ceed, data->Deriv, grid, block_size_x, 1, elems_per_block, curl_args)); 157 } 158 } break; 159 case CEED_EVAL_WEIGHT: { 160 void *weight_args[] = {(void *)&num_elem, (void *)&data->d_q_weight, &d_v}; 161 162 CeedCallBackend(CeedRunKernelDim_Cuda(ceed, data->Weight, grid, num_qpts, 1, elems_per_block, weight_args)); 163 } break; 164 case CEED_EVAL_NONE: /* handled separately below */ 165 break; 166 } 167 168 // Restore vectors, cover CEED_EVAL_NONE 169 CeedCallBackend(CeedVectorRestoreArray(v, &d_v)); 170 if (eval_mode == CEED_EVAL_NONE) CeedCallBackend(CeedVectorSetArray(v, CEED_MEM_DEVICE, CEED_COPY_VALUES, (CeedScalar *)d_u)); 171 if (eval_mode != CEED_EVAL_WEIGHT) CeedCallBackend(CeedVectorRestoreArrayRead(u, &d_u)); 172 return CEED_ERROR_SUCCESS; 173 } 174 175 //------------------------------------------------------------------------------ 176 // Destroy tensor basis 177 //------------------------------------------------------------------------------ 178 static int CeedBasisDestroy_Cuda(CeedBasis basis) { 179 Ceed ceed; 180 CeedBasis_Cuda *data; 181 182 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 183 CeedCallBackend(CeedBasisGetData(basis, &data)); 184 CeedCallCuda(ceed, cuModuleUnload(data->module)); 185 CeedCallCuda(ceed, cudaFree(data->d_q_weight_1d)); 186 CeedCallCuda(ceed, cudaFree(data->d_interp_1d)); 187 CeedCallCuda(ceed, cudaFree(data->d_grad_1d)); 188 CeedCallBackend(CeedFree(&data)); 189 return CEED_ERROR_SUCCESS; 190 } 191 192 //------------------------------------------------------------------------------ 193 // Destroy non-tensor basis 194 //------------------------------------------------------------------------------ 195 static int CeedBasisDestroyNonTensor_Cuda(CeedBasis basis) { 196 Ceed ceed; 197 CeedBasisNonTensor_Cuda *data; 198 199 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 200 CeedCallBackend(CeedBasisGetData(basis, &data)); 201 CeedCallCuda(ceed, cuModuleUnload(data->module)); 202 CeedCallCuda(ceed, cudaFree(data->d_q_weight)); 203 CeedCallCuda(ceed, cudaFree(data->d_interp)); 204 CeedCallCuda(ceed, cudaFree(data->d_grad)); 205 CeedCallCuda(ceed, cudaFree(data->d_div)); 206 CeedCallCuda(ceed, cudaFree(data->d_curl)); 207 CeedCallBackend(CeedFree(&data)); 208 return CEED_ERROR_SUCCESS; 209 } 210 211 //------------------------------------------------------------------------------ 212 // Create tensor 213 //------------------------------------------------------------------------------ 214 int CeedBasisCreateTensorH1_Cuda(CeedInt dim, CeedInt P_1d, CeedInt Q_1d, const CeedScalar *interp_1d, const CeedScalar *grad_1d, 215 const CeedScalar *q_ref_1d, const CeedScalar *q_weight_1d, CeedBasis basis) { 216 Ceed ceed; 217 char *basis_kernel_path, *basis_kernel_source; 218 CeedInt num_comp; 219 const CeedInt q_bytes = Q_1d * sizeof(CeedScalar); 220 const CeedInt interp_bytes = q_bytes * P_1d; 221 CeedBasis_Cuda *data; 222 223 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 224 CeedCallBackend(CeedCalloc(1, &data)); 225 226 // Copy data to GPU 227 CeedCallCuda(ceed, cudaMalloc((void **)&data->d_q_weight_1d, q_bytes)); 228 CeedCallCuda(ceed, cudaMemcpy(data->d_q_weight_1d, q_weight_1d, q_bytes, cudaMemcpyHostToDevice)); 229 CeedCallCuda(ceed, cudaMalloc((void **)&data->d_interp_1d, interp_bytes)); 230 CeedCallCuda(ceed, cudaMemcpy(data->d_interp_1d, interp_1d, interp_bytes, cudaMemcpyHostToDevice)); 231 CeedCallCuda(ceed, cudaMalloc((void **)&data->d_grad_1d, interp_bytes)); 232 CeedCallCuda(ceed, cudaMemcpy(data->d_grad_1d, grad_1d, interp_bytes, cudaMemcpyHostToDevice)); 233 234 // Compile basis kernels 235 CeedCallBackend(CeedBasisGetNumComponents(basis, &num_comp)); 236 CeedCallBackend(CeedGetJitAbsolutePath(ceed, "ceed/jit-source/cuda/cuda-ref-basis-tensor.h", &basis_kernel_path)); 237 CeedDebug256(ceed, CEED_DEBUG_COLOR_SUCCESS, "----- Loading Basis Kernel Source -----\n"); 238 CeedCallBackend(CeedLoadSourceToBuffer(ceed, basis_kernel_path, &basis_kernel_source)); 239 CeedDebug256(ceed, CEED_DEBUG_COLOR_SUCCESS, "----- Loading Basis Kernel Source Complete! -----\n"); 240 CeedCallBackend(CeedCompile_Cuda(ceed, basis_kernel_source, &data->module, 7, "BASIS_Q_1D", Q_1d, "BASIS_P_1D", P_1d, "BASIS_BUF_LEN", 241 num_comp * CeedIntPow(Q_1d > P_1d ? Q_1d : P_1d, dim), "BASIS_DIM", dim, "BASIS_NUM_COMP", num_comp, 242 "BASIS_NUM_NODES", CeedIntPow(P_1d, dim), "BASIS_NUM_QPTS", CeedIntPow(Q_1d, dim))); 243 CeedCallBackend(CeedGetKernel_Cuda(ceed, data->module, "Interp", &data->Interp)); 244 CeedCallBackend(CeedGetKernel_Cuda(ceed, data->module, "Grad", &data->Grad)); 245 CeedCallBackend(CeedGetKernel_Cuda(ceed, data->module, "Weight", &data->Weight)); 246 CeedCallBackend(CeedFree(&basis_kernel_path)); 247 CeedCallBackend(CeedFree(&basis_kernel_source)); 248 249 CeedCallBackend(CeedBasisSetData(basis, data)); 250 251 // Register backend functions 252 CeedCallBackend(CeedSetBackendFunction(ceed, "Basis", basis, "Apply", CeedBasisApply_Cuda)); 253 CeedCallBackend(CeedSetBackendFunction(ceed, "Basis", basis, "Destroy", CeedBasisDestroy_Cuda)); 254 return CEED_ERROR_SUCCESS; 255 } 256 257 //------------------------------------------------------------------------------ 258 // Create non-tensor H^1 259 //------------------------------------------------------------------------------ 260 int CeedBasisCreateH1_Cuda(CeedElemTopology topo, CeedInt dim, CeedInt num_nodes, CeedInt num_qpts, const CeedScalar *interp, const CeedScalar *grad, 261 const CeedScalar *q_ref, const CeedScalar *q_weight, CeedBasis basis) { 262 Ceed ceed; 263 char *basis_kernel_path, *basis_kernel_source; 264 CeedInt num_comp, q_comp_interp, q_comp_grad; 265 const CeedInt q_bytes = num_qpts * sizeof(CeedScalar); 266 CeedBasisNonTensor_Cuda *data; 267 268 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 269 CeedCallBackend(CeedCalloc(1, &data)); 270 271 // Copy basis data to GPU 272 CeedCallBackend(CeedBasisGetNumQuadratureComponents(basis, CEED_EVAL_INTERP, &q_comp_interp)); 273 CeedCallBackend(CeedBasisGetNumQuadratureComponents(basis, CEED_EVAL_GRAD, &q_comp_grad)); 274 CeedCallCuda(ceed, cudaMalloc((void **)&data->d_q_weight, q_bytes)); 275 CeedCallCuda(ceed, cudaMemcpy(data->d_q_weight, q_weight, q_bytes, cudaMemcpyHostToDevice)); 276 if (interp) { 277 const CeedInt interp_bytes = q_bytes * num_nodes * q_comp_interp; 278 279 CeedCallCuda(ceed, cudaMalloc((void **)&data->d_interp, interp_bytes)); 280 CeedCallCuda(ceed, cudaMemcpy(data->d_interp, interp, interp_bytes, cudaMemcpyHostToDevice)); 281 } 282 if (grad) { 283 const CeedInt grad_bytes = q_bytes * num_nodes * q_comp_grad; 284 285 CeedCallCuda(ceed, cudaMalloc((void **)&data->d_grad, grad_bytes)); 286 CeedCallCuda(ceed, cudaMemcpy(data->d_grad, grad, grad_bytes, cudaMemcpyHostToDevice)); 287 } 288 289 // Compile basis kernels 290 CeedCallBackend(CeedBasisGetNumComponents(basis, &num_comp)); 291 CeedCallBackend(CeedGetJitAbsolutePath(ceed, "ceed/jit-source/cuda/cuda-ref-basis-nontensor.h", &basis_kernel_path)); 292 CeedDebug256(ceed, CEED_DEBUG_COLOR_SUCCESS, "----- Loading Basis Kernel Source -----\n"); 293 CeedCallBackend(CeedLoadSourceToBuffer(ceed, basis_kernel_path, &basis_kernel_source)); 294 CeedDebug256(ceed, CEED_DEBUG_COLOR_SUCCESS, "----- Loading Basis Kernel Source Complete! -----\n"); 295 CeedCallBackend(CeedCompile_Cuda(ceed, basis_kernel_source, &data->module, 5, "BASIS_Q", num_qpts, "BASIS_P", num_nodes, "BASIS_Q_COMP_INTERP", 296 q_comp_interp, "BASIS_Q_COMP_DERIV", q_comp_grad, "BASIS_NUM_COMP", num_comp)); 297 CeedCallBackend(CeedGetKernel_Cuda(ceed, data->module, "Interp", &data->Interp)); 298 CeedCallBackend(CeedGetKernel_Cuda(ceed, data->module, "InterpTranspose", &data->InterpTranspose)); 299 CeedCallBackend(CeedGetKernel_Cuda(ceed, data->module, "Deriv", &data->Deriv)); 300 CeedCallBackend(CeedGetKernel_Cuda(ceed, data->module, "DerivTranspose", &data->DerivTranspose)); 301 CeedCallBackend(CeedGetKernel_Cuda(ceed, data->module, "Weight", &data->Weight)); 302 CeedCallBackend(CeedFree(&basis_kernel_path)); 303 CeedCallBackend(CeedFree(&basis_kernel_source)); 304 305 CeedCallBackend(CeedBasisSetData(basis, data)); 306 307 // Register backend functions 308 CeedCallBackend(CeedSetBackendFunction(ceed, "Basis", basis, "Apply", CeedBasisApplyNonTensor_Cuda)); 309 CeedCallBackend(CeedSetBackendFunction(ceed, "Basis", basis, "Destroy", CeedBasisDestroyNonTensor_Cuda)); 310 return CEED_ERROR_SUCCESS; 311 } 312 313 //------------------------------------------------------------------------------ 314 // Create non-tensor H(div) 315 //------------------------------------------------------------------------------ 316 int CeedBasisCreateHdiv_Cuda(CeedElemTopology topo, CeedInt dim, CeedInt num_nodes, CeedInt num_qpts, const CeedScalar *interp, const CeedScalar *div, 317 const CeedScalar *q_ref, const CeedScalar *q_weight, CeedBasis basis) { 318 Ceed ceed; 319 char *basis_kernel_path, *basis_kernel_source; 320 CeedInt num_comp, q_comp_interp, q_comp_div; 321 const CeedInt q_bytes = num_qpts * sizeof(CeedScalar); 322 CeedBasisNonTensor_Cuda *data; 323 324 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 325 CeedCallBackend(CeedCalloc(1, &data)); 326 327 // Copy basis data to GPU 328 CeedCallBackend(CeedBasisGetNumQuadratureComponents(basis, CEED_EVAL_INTERP, &q_comp_interp)); 329 CeedCallBackend(CeedBasisGetNumQuadratureComponents(basis, CEED_EVAL_DIV, &q_comp_div)); 330 CeedCallCuda(ceed, cudaMalloc((void **)&data->d_q_weight, q_bytes)); 331 CeedCallCuda(ceed, cudaMemcpy(data->d_q_weight, q_weight, q_bytes, cudaMemcpyHostToDevice)); 332 if (interp) { 333 const CeedInt interp_bytes = q_bytes * num_nodes * q_comp_interp; 334 335 CeedCallCuda(ceed, cudaMalloc((void **)&data->d_interp, interp_bytes)); 336 CeedCallCuda(ceed, cudaMemcpy(data->d_interp, interp, interp_bytes, cudaMemcpyHostToDevice)); 337 } 338 if (div) { 339 const CeedInt div_bytes = q_bytes * num_nodes * q_comp_div; 340 341 CeedCallCuda(ceed, cudaMalloc((void **)&data->d_div, div_bytes)); 342 CeedCallCuda(ceed, cudaMemcpy(data->d_div, div, div_bytes, cudaMemcpyHostToDevice)); 343 } 344 345 // Compile basis kernels 346 CeedCallBackend(CeedBasisGetNumComponents(basis, &num_comp)); 347 CeedCallBackend(CeedGetJitAbsolutePath(ceed, "ceed/jit-source/cuda/cuda-ref-basis-nontensor.h", &basis_kernel_path)); 348 CeedDebug256(ceed, CEED_DEBUG_COLOR_SUCCESS, "----- Loading Basis Kernel Source -----\n"); 349 CeedCallBackend(CeedLoadSourceToBuffer(ceed, basis_kernel_path, &basis_kernel_source)); 350 CeedDebug256(ceed, CEED_DEBUG_COLOR_SUCCESS, "----- Loading Basis Kernel Source Complete! -----\n"); 351 CeedCallBackend(CeedCompile_Cuda(ceed, basis_kernel_source, &data->module, 5, "BASIS_Q", num_qpts, "BASIS_P", num_nodes, "BASIS_Q_COMP_INTERP", 352 q_comp_interp, "BASIS_Q_COMP_DERIV", q_comp_div, "BASIS_NUM_COMP", num_comp)); 353 CeedCallBackend(CeedGetKernel_Cuda(ceed, data->module, "Interp", &data->Interp)); 354 CeedCallBackend(CeedGetKernel_Cuda(ceed, data->module, "InterpTranspose", &data->InterpTranspose)); 355 CeedCallBackend(CeedGetKernel_Cuda(ceed, data->module, "Deriv", &data->Deriv)); 356 CeedCallBackend(CeedGetKernel_Cuda(ceed, data->module, "DerivTranspose", &data->DerivTranspose)); 357 CeedCallBackend(CeedGetKernel_Cuda(ceed, data->module, "Weight", &data->Weight)); 358 CeedCallBackend(CeedFree(&basis_kernel_path)); 359 CeedCallBackend(CeedFree(&basis_kernel_source)); 360 361 CeedCallBackend(CeedBasisSetData(basis, data)); 362 363 // Register backend functions 364 CeedCallBackend(CeedSetBackendFunction(ceed, "Basis", basis, "Apply", CeedBasisApplyNonTensor_Cuda)); 365 CeedCallBackend(CeedSetBackendFunction(ceed, "Basis", basis, "Destroy", CeedBasisDestroyNonTensor_Cuda)); 366 return CEED_ERROR_SUCCESS; 367 } 368 369 //------------------------------------------------------------------------------ 370 // Create non-tensor H(curl) 371 //------------------------------------------------------------------------------ 372 int CeedBasisCreateHcurl_Cuda(CeedElemTopology topo, CeedInt dim, CeedInt num_nodes, CeedInt num_qpts, const CeedScalar *interp, 373 const CeedScalar *curl, const CeedScalar *q_ref, const CeedScalar *q_weight, CeedBasis basis) { 374 Ceed ceed; 375 char *basis_kernel_path, *basis_kernel_source; 376 CeedInt num_comp, q_comp_interp, q_comp_curl; 377 const CeedInt q_bytes = num_qpts * sizeof(CeedScalar); 378 CeedBasisNonTensor_Cuda *data; 379 380 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 381 CeedCallBackend(CeedCalloc(1, &data)); 382 383 // Copy basis data to GPU 384 CeedCallBackend(CeedBasisGetNumQuadratureComponents(basis, CEED_EVAL_INTERP, &q_comp_interp)); 385 CeedCallBackend(CeedBasisGetNumQuadratureComponents(basis, CEED_EVAL_CURL, &q_comp_curl)); 386 CeedCallCuda(ceed, cudaMalloc((void **)&data->d_q_weight, q_bytes)); 387 CeedCallCuda(ceed, cudaMemcpy(data->d_q_weight, q_weight, q_bytes, cudaMemcpyHostToDevice)); 388 if (interp) { 389 const CeedInt interp_bytes = q_bytes * num_nodes * q_comp_interp; 390 391 CeedCallCuda(ceed, cudaMalloc((void **)&data->d_interp, interp_bytes)); 392 CeedCallCuda(ceed, cudaMemcpy(data->d_interp, interp, interp_bytes, cudaMemcpyHostToDevice)); 393 } 394 if (curl) { 395 const CeedInt curl_bytes = q_bytes * num_nodes * q_comp_curl; 396 397 CeedCallCuda(ceed, cudaMalloc((void **)&data->d_curl, curl_bytes)); 398 CeedCallCuda(ceed, cudaMemcpy(data->d_curl, curl, curl_bytes, cudaMemcpyHostToDevice)); 399 } 400 401 // Compile basis kernels 402 CeedCallBackend(CeedBasisGetNumComponents(basis, &num_comp)); 403 CeedCallBackend(CeedGetJitAbsolutePath(ceed, "ceed/jit-source/cuda/cuda-ref-basis-nontensor.h", &basis_kernel_path)); 404 CeedDebug256(ceed, CEED_DEBUG_COLOR_SUCCESS, "----- Loading Basis Kernel Source -----\n"); 405 CeedCallBackend(CeedLoadSourceToBuffer(ceed, basis_kernel_path, &basis_kernel_source)); 406 CeedDebug256(ceed, CEED_DEBUG_COLOR_SUCCESS, "----- Loading Basis Kernel Source Complete! -----\n"); 407 CeedCallBackend(CeedCompile_Cuda(ceed, basis_kernel_source, &data->module, 5, "BASIS_Q", num_qpts, "BASIS_P", num_nodes, "BASIS_Q_COMP_INTERP", 408 q_comp_interp, "BASIS_Q_COMP_DERIV", q_comp_curl, "BASIS_NUM_COMP", num_comp)); 409 CeedCallBackend(CeedGetKernel_Cuda(ceed, data->module, "Interp", &data->Interp)); 410 CeedCallBackend(CeedGetKernel_Cuda(ceed, data->module, "InterpTranspose", &data->InterpTranspose)); 411 CeedCallBackend(CeedGetKernel_Cuda(ceed, data->module, "Deriv", &data->Deriv)); 412 CeedCallBackend(CeedGetKernel_Cuda(ceed, data->module, "DerivTranspose", &data->DerivTranspose)); 413 CeedCallBackend(CeedGetKernel_Cuda(ceed, data->module, "Weight", &data->Weight)); 414 CeedCallBackend(CeedFree(&basis_kernel_path)); 415 CeedCallBackend(CeedFree(&basis_kernel_source)); 416 417 CeedCallBackend(CeedBasisSetData(basis, data)); 418 419 // Register backend functions 420 CeedCallBackend(CeedSetBackendFunction(ceed, "Basis", basis, "Apply", CeedBasisApplyNonTensor_Cuda)); 421 CeedCallBackend(CeedSetBackendFunction(ceed, "Basis", basis, "Destroy", CeedBasisDestroyNonTensor_Cuda)); 422 return CEED_ERROR_SUCCESS; 423 } 424 425 //------------------------------------------------------------------------------ 426