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 <hip/hip_runtime.h> 12 13 #include "../hip/ceed-hip-common.h" 14 #include "../hip/ceed-hip-compile.h" 15 #include "ceed-hip-ref.h" 16 17 //------------------------------------------------------------------------------ 18 // Basis apply - tensor 19 //------------------------------------------------------------------------------ 20 int CeedBasisApply_Hip(CeedBasis basis, const CeedInt num_elem, CeedTransposeMode t_mode, CeedEvalMode eval_mode, CeedVector u, CeedVector v) { 21 Ceed ceed; 22 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 23 Ceed_Hip *ceed_Hip; 24 CeedCallBackend(CeedGetData(ceed, &ceed_Hip)); 25 CeedBasis_Hip *data; 26 CeedCallBackend(CeedBasisGetData(basis, &data)); 27 const CeedInt transpose = t_mode == CEED_TRANSPOSE; 28 const int max_block_size = 64; 29 30 // Read vectors 31 const CeedScalar *d_u; 32 CeedScalar *d_v; 33 if (u != CEED_VECTOR_NONE) CeedCallBackend(CeedVectorGetArrayRead(u, CEED_MEM_DEVICE, &d_u)); 34 else CeedCheck(eval_mode == CEED_EVAL_WEIGHT, ceed, CEED_ERROR_BACKEND, "An input vector is required for this CeedEvalMode"); 35 CeedCallBackend(CeedVectorGetArrayWrite(v, CEED_MEM_DEVICE, &d_v)); 36 37 // Clear v for transpose operation 38 if (t_mode == CEED_TRANSPOSE) { 39 CeedSize length; 40 CeedCallBackend(CeedVectorGetLength(v, &length)); 41 CeedCallHip(ceed, hipMemset(d_v, 0, length * sizeof(CeedScalar))); 42 } 43 44 // Basis action 45 switch (eval_mode) { 46 case CEED_EVAL_INTERP: { 47 void *interp_args[] = {(void *)&num_elem, (void *)&transpose, &data->d_interp_1d, &d_u, &d_v}; 48 CeedInt Q_1d, dim; 49 CeedCallBackend(CeedBasisGetNumQuadraturePoints1D(basis, &Q_1d)); 50 CeedCallBackend(CeedBasisGetDimension(basis, &dim)); 51 CeedInt block_size = CeedIntMin(CeedIntPow(Q_1d, dim), max_block_size); 52 53 CeedCallBackend(CeedRunKernelHip(ceed, data->Interp, num_elem, block_size, interp_args)); 54 } break; 55 case CEED_EVAL_GRAD: { 56 void *grad_args[] = {(void *)&num_elem, (void *)&transpose, &data->d_interp_1d, &data->d_grad_1d, &d_u, &d_v}; 57 CeedInt block_size = max_block_size; 58 59 CeedCallBackend(CeedRunKernelHip(ceed, data->Grad, num_elem, block_size, grad_args)); 60 } break; 61 case CEED_EVAL_WEIGHT: { 62 void *weight_args[] = {(void *)&num_elem, (void *)&data->d_q_weight_1d, &d_v}; 63 const int block_size = 64; 64 int grid_size = num_elem / block_size; 65 if (block_size * grid_size < num_elem) grid_size += 1; 66 67 CeedCallBackend(CeedRunKernelHip(ceed, data->Weight, grid_size, block_size, weight_args)); 68 } break; 69 // LCOV_EXCL_START 70 // Evaluate the divergence to/from the quadrature points 71 case CEED_EVAL_DIV: 72 return CeedError(ceed, CEED_ERROR_BACKEND, "CEED_EVAL_DIV not supported"); 73 // Evaluate the curl to/from the quadrature points 74 case CEED_EVAL_CURL: 75 return CeedError(ceed, CEED_ERROR_BACKEND, "CEED_EVAL_CURL not supported"); 76 // Take no action, BasisApply should not have been called 77 case CEED_EVAL_NONE: 78 return CeedError(ceed, CEED_ERROR_BACKEND, "CEED_EVAL_NONE does not make sense in this context"); 79 // LCOV_EXCL_STOP 80 } 81 82 // Restore vectors 83 if (eval_mode != CEED_EVAL_WEIGHT) { 84 CeedCallBackend(CeedVectorRestoreArrayRead(u, &d_u)); 85 } 86 CeedCallBackend(CeedVectorRestoreArray(v, &d_v)); 87 return CEED_ERROR_SUCCESS; 88 } 89 90 //------------------------------------------------------------------------------ 91 // Basis apply - non-tensor 92 //------------------------------------------------------------------------------ 93 int CeedBasisApplyNonTensor_Hip(CeedBasis basis, const CeedInt num_elem, CeedTransposeMode t_mode, CeedEvalMode eval_mode, CeedVector u, 94 CeedVector v) { 95 Ceed ceed; 96 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 97 Ceed_Hip *ceed_Hip; 98 CeedCallBackend(CeedGetData(ceed, &ceed_Hip)); 99 CeedBasisNonTensor_Hip *data; 100 CeedCallBackend(CeedBasisGetData(basis, &data)); 101 CeedInt num_nodes, num_qpts; 102 CeedCallBackend(CeedBasisGetNumQuadraturePoints(basis, &num_qpts)); 103 CeedCallBackend(CeedBasisGetNumNodes(basis, &num_nodes)); 104 const CeedInt transpose = t_mode == CEED_TRANSPOSE; 105 int elemsPerBlock = 1; 106 int grid = num_elem / elemsPerBlock + ((num_elem / elemsPerBlock * elemsPerBlock < num_elem) ? 1 : 0); 107 108 // Read vectors 109 const CeedScalar *d_u; 110 CeedScalar *d_v; 111 if (eval_mode != CEED_EVAL_WEIGHT) { 112 CeedCallBackend(CeedVectorGetArrayRead(u, CEED_MEM_DEVICE, &d_u)); 113 } 114 CeedCallBackend(CeedVectorGetArrayWrite(v, CEED_MEM_DEVICE, &d_v)); 115 116 // Clear v for transpose operation 117 if (t_mode == CEED_TRANSPOSE) { 118 CeedSize length; 119 CeedCallBackend(CeedVectorGetLength(v, &length)); 120 CeedCallHip(ceed, hipMemset(d_v, 0, length * sizeof(CeedScalar))); 121 } 122 123 // Apply basis operation 124 switch (eval_mode) { 125 case CEED_EVAL_INTERP: { 126 void *interp_args[] = {(void *)&num_elem, (void *)&transpose, &data->d_interp, &d_u, &d_v}; 127 const int block_size_x = transpose ? num_nodes : num_qpts; 128 CeedCallBackend(CeedRunKernelDimHip(ceed, data->Interp, grid, block_size_x, 1, elemsPerBlock, interp_args)); 129 } break; 130 case CEED_EVAL_GRAD: { 131 void *grad_args[] = {(void *)&num_elem, (void *)&transpose, &data->d_grad, &d_u, &d_v}; 132 const int block_size_x = transpose ? num_nodes : num_qpts; 133 CeedCallBackend(CeedRunKernelDimHip(ceed, data->Grad, grid, block_size_x, 1, elemsPerBlock, grad_args)); 134 } break; 135 case CEED_EVAL_WEIGHT: { 136 void *weight_args[] = {(void *)&num_elem, (void *)&data->d_q_weight, &d_v}; 137 CeedCallBackend(CeedRunKernelDimHip(ceed, data->Weight, grid, num_qpts, 1, elemsPerBlock, weight_args)); 138 } break; 139 // LCOV_EXCL_START 140 // Evaluate the divergence to/from the quadrature points 141 case CEED_EVAL_DIV: 142 return CeedError(ceed, CEED_ERROR_BACKEND, "CEED_EVAL_DIV not supported"); 143 // Evaluate the curl to/from the quadrature points 144 case CEED_EVAL_CURL: 145 return CeedError(ceed, CEED_ERROR_BACKEND, "CEED_EVAL_CURL not supported"); 146 // Take no action, BasisApply should not have been called 147 case CEED_EVAL_NONE: 148 return CeedError(ceed, CEED_ERROR_BACKEND, "CEED_EVAL_NONE does not make sense in this context"); 149 // LCOV_EXCL_STOP 150 } 151 152 // Restore vectors 153 if (eval_mode != CEED_EVAL_WEIGHT) { 154 CeedCallBackend(CeedVectorRestoreArrayRead(u, &d_u)); 155 } 156 CeedCallBackend(CeedVectorRestoreArray(v, &d_v)); 157 return CEED_ERROR_SUCCESS; 158 } 159 160 //------------------------------------------------------------------------------ 161 // Destroy tensor basis 162 //------------------------------------------------------------------------------ 163 static int CeedBasisDestroy_Hip(CeedBasis basis) { 164 Ceed ceed; 165 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 166 167 CeedBasis_Hip *data; 168 CeedCallBackend(CeedBasisGetData(basis, &data)); 169 170 CeedCallHip(ceed, hipModuleUnload(data->module)); 171 172 CeedCallHip(ceed, hipFree(data->d_q_weight_1d)); 173 CeedCallHip(ceed, hipFree(data->d_interp_1d)); 174 CeedCallHip(ceed, hipFree(data->d_grad_1d)); 175 CeedCallBackend(CeedFree(&data)); 176 177 return CEED_ERROR_SUCCESS; 178 } 179 180 //------------------------------------------------------------------------------ 181 // Destroy non-tensor basis 182 //------------------------------------------------------------------------------ 183 static int CeedBasisDestroyNonTensor_Hip(CeedBasis basis) { 184 Ceed ceed; 185 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 186 187 CeedBasisNonTensor_Hip *data; 188 CeedCallBackend(CeedBasisGetData(basis, &data)); 189 190 CeedCallHip(ceed, hipModuleUnload(data->module)); 191 192 CeedCallHip(ceed, hipFree(data->d_q_weight)); 193 CeedCallHip(ceed, hipFree(data->d_interp)); 194 CeedCallHip(ceed, hipFree(data->d_grad)); 195 CeedCallBackend(CeedFree(&data)); 196 197 return CEED_ERROR_SUCCESS; 198 } 199 200 //------------------------------------------------------------------------------ 201 // Create tensor 202 //------------------------------------------------------------------------------ 203 int CeedBasisCreateTensorH1_Hip(CeedInt dim, CeedInt P_1d, CeedInt Q_1d, const CeedScalar *interp_1d, const CeedScalar *grad_1d, 204 const CeedScalar *q_ref_1d, const CeedScalar *q_weight_1d, CeedBasis basis) { 205 Ceed ceed; 206 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 207 CeedBasis_Hip *data; 208 CeedCallBackend(CeedCalloc(1, &data)); 209 210 // Copy data to GPU 211 const CeedInt q_bytes = Q_1d * sizeof(CeedScalar); 212 CeedCallHip(ceed, hipMalloc((void **)&data->d_q_weight_1d, q_bytes)); 213 CeedCallHip(ceed, hipMemcpy(data->d_q_weight_1d, q_weight_1d, q_bytes, hipMemcpyHostToDevice)); 214 215 const CeedInt interp_bytes = q_bytes * P_1d; 216 CeedCallHip(ceed, hipMalloc((void **)&data->d_interp_1d, interp_bytes)); 217 CeedCallHip(ceed, hipMemcpy(data->d_interp_1d, interp_1d, interp_bytes, hipMemcpyHostToDevice)); 218 219 CeedCallHip(ceed, hipMalloc((void **)&data->d_grad_1d, interp_bytes)); 220 CeedCallHip(ceed, hipMemcpy(data->d_grad_1d, grad_1d, interp_bytes, hipMemcpyHostToDevice)); 221 222 // Compile basis kernels 223 CeedInt ncomp; 224 CeedCallBackend(CeedBasisGetNumComponents(basis, &ncomp)); 225 char *basis_kernel_path, *basis_kernel_source; 226 CeedCallBackend(CeedGetJitAbsolutePath(ceed, "ceed/jit-source/hip/hip-ref-basis-tensor.h", &basis_kernel_path)); 227 CeedDebug256(ceed, 2, "----- Loading Basis Kernel Source -----\n"); 228 CeedCallBackend(CeedLoadSourceToBuffer(ceed, basis_kernel_path, &basis_kernel_source)); 229 CeedDebug256(ceed, 2, "----- Loading Basis Kernel Source Complete! -----\n"); 230 CeedCallBackend(CeedCompileHip(ceed, basis_kernel_source, &data->module, 7, "BASIS_Q_1D", Q_1d, "BASIS_P_1D", P_1d, "BASIS_BUF_LEN", 231 ncomp * CeedIntPow(Q_1d > P_1d ? Q_1d : P_1d, dim), "BASIS_DIM", dim, "BASIS_NUM_COMP", ncomp, "BASIS_NUM_NODES", 232 CeedIntPow(P_1d, dim), "BASIS_NUM_QPTS", CeedIntPow(Q_1d, dim))); 233 CeedCallBackend(CeedGetKernelHip(ceed, data->module, "Interp", &data->Interp)); 234 CeedCallBackend(CeedGetKernelHip(ceed, data->module, "Grad", &data->Grad)); 235 CeedCallBackend(CeedGetKernelHip(ceed, data->module, "Weight", &data->Weight)); 236 CeedCallBackend(CeedFree(&basis_kernel_path)); 237 CeedCallBackend(CeedFree(&basis_kernel_source)); 238 239 CeedCallBackend(CeedBasisSetData(basis, data)); 240 241 CeedCallBackend(CeedSetBackendFunction(ceed, "Basis", basis, "Apply", CeedBasisApply_Hip)); 242 CeedCallBackend(CeedSetBackendFunction(ceed, "Basis", basis, "Destroy", CeedBasisDestroy_Hip)); 243 return CEED_ERROR_SUCCESS; 244 } 245 246 //------------------------------------------------------------------------------ 247 // Create non-tensor 248 //------------------------------------------------------------------------------ 249 int CeedBasisCreateH1_Hip(CeedElemTopology topo, CeedInt dim, CeedInt num_nodes, CeedInt num_qpts, const CeedScalar *interp, const CeedScalar *grad, 250 const CeedScalar *q_ref, const CeedScalar *q_weight, CeedBasis basis) { 251 Ceed ceed; 252 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 253 CeedBasisNonTensor_Hip *data; 254 CeedCallBackend(CeedCalloc(1, &data)); 255 256 // Copy basis data to GPU 257 const CeedInt q_bytes = num_qpts * sizeof(CeedScalar); 258 CeedCallHip(ceed, hipMalloc((void **)&data->d_q_weight, q_bytes)); 259 CeedCallHip(ceed, hipMemcpy(data->d_q_weight, q_weight, q_bytes, hipMemcpyHostToDevice)); 260 261 const CeedInt interp_bytes = q_bytes * num_nodes; 262 CeedCallHip(ceed, hipMalloc((void **)&data->d_interp, interp_bytes)); 263 CeedCallHip(ceed, hipMemcpy(data->d_interp, interp, interp_bytes, hipMemcpyHostToDevice)); 264 265 const CeedInt grad_bytes = q_bytes * num_nodes * dim; 266 CeedCallHip(ceed, hipMalloc((void **)&data->d_grad, grad_bytes)); 267 CeedCallHip(ceed, hipMemcpy(data->d_grad, grad, grad_bytes, hipMemcpyHostToDevice)); 268 269 // Compile basis kernels 270 CeedInt ncomp; 271 CeedCallBackend(CeedBasisGetNumComponents(basis, &ncomp)); 272 char *basis_kernel_path, *basis_kernel_source; 273 CeedCallBackend(CeedGetJitAbsolutePath(ceed, "ceed/jit-source/hip/hip-ref-basis-nontensor.h", &basis_kernel_path)); 274 CeedDebug256(ceed, 2, "----- Loading Basis Kernel Source -----\n"); 275 CeedCallBackend(CeedLoadSourceToBuffer(ceed, basis_kernel_path, &basis_kernel_source)); 276 CeedDebug256(ceed, 2, "----- Loading Basis Kernel Source Complete! -----\n"); 277 CeedCallBackend(CeedCompileHip(ceed, basis_kernel_source, &data->module, 4, "BASIS_Q", num_qpts, "BASIS_P", num_nodes, "BASIS_DIM", dim, 278 "BASIS_NUM_COMP", ncomp)); 279 CeedCallBackend(CeedGetKernelHip(ceed, data->module, "Interp", &data->Interp)); 280 CeedCallBackend(CeedGetKernelHip(ceed, data->module, "Grad", &data->Grad)); 281 CeedCallBackend(CeedGetKernelHip(ceed, data->module, "Weight", &data->Weight)); 282 CeedCallBackend(CeedFree(&basis_kernel_path)); 283 CeedCallBackend(CeedFree(&basis_kernel_source)); 284 CeedCallBackend(CeedBasisSetData(basis, data)); 285 286 // Register backend functions 287 CeedCallBackend(CeedSetBackendFunction(ceed, "Basis", basis, "Apply", CeedBasisApplyNonTensor_Hip)); 288 CeedCallBackend(CeedSetBackendFunction(ceed, "Basis", basis, "Destroy", CeedBasisDestroyNonTensor_Hip)); 289 return CEED_ERROR_SUCCESS; 290 } 291 292 //------------------------------------------------------------------------------ 293