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 #include <hip/hip_runtime.h> 12 #include <stddef.h> 13 14 #include "../hip/ceed-hip-common.h" 15 #include "../hip/ceed-hip-compile.h" 16 #include "ceed-hip-shared.h" 17 18 //------------------------------------------------------------------------------ 19 // Compute a block size based on required minimum threads 20 //------------------------------------------------------------------------------ 21 static CeedInt ComputeBlockSizeFromRequirement(const CeedInt required) { 22 CeedInt maxSize = 1024; // Max total threads per block 23 CeedInt currentSize = 64; // Start with one group 24 25 while (currentSize < maxSize) { 26 if (currentSize > required) break; 27 else currentSize = currentSize * 2; 28 } 29 return currentSize; 30 } 31 32 //------------------------------------------------------------------------------ 33 // Compute required thread block sizes for basis kernels given P, Q, dim, and 34 // num_comp (num_comp not currently used, but may be again in other basis 35 // parallelization options) 36 //------------------------------------------------------------------------------ 37 static int ComputeBasisThreadBlockSizes(const CeedInt dim, const CeedInt P_1d, const CeedInt Q_1d, const CeedInt num_comp, CeedInt *block_sizes) { 38 // Note that this will use the same block sizes for all dimensions when compiling, 39 // but as each basis object is defined for a particular dimension, we will never 40 // call any kernels except the ones for the dimension for which we have computed the 41 // block sizes. 42 const CeedInt thread_1d = CeedIntMax(P_1d, Q_1d); 43 switch (dim) { 44 case 1: { 45 // Interp kernels: 46 block_sizes[0] = 256; 47 48 // Grad kernels: 49 block_sizes[1] = 256; 50 51 // Weight kernels: 52 block_sizes[2] = 256; 53 } break; 54 case 2: { 55 // Interp kernels: 56 CeedInt required = thread_1d * thread_1d; 57 block_sizes[0] = CeedIntMax(256, ComputeBlockSizeFromRequirement(required)); 58 59 // Grad kernels: currently use same required minimum threads 60 block_sizes[1] = CeedIntMax(256, ComputeBlockSizeFromRequirement(required)); 61 62 // Weight kernels: 63 required = CeedIntMax(64, Q_1d * Q_1d); 64 block_sizes[2] = CeedIntMax(256, ComputeBlockSizeFromRequirement(required)); 65 66 } break; 67 case 3: { 68 // Interp kernels: 69 CeedInt required = thread_1d * thread_1d; 70 block_sizes[0] = CeedIntMax(256, ComputeBlockSizeFromRequirement(required)); 71 72 // Grad kernels: currently use same required minimum threads 73 block_sizes[1] = CeedIntMax(256, ComputeBlockSizeFromRequirement(required)); 74 75 // Weight kernels: 76 required = Q_1d * Q_1d * Q_1d; 77 block_sizes[2] = CeedIntMax(256, ComputeBlockSizeFromRequirement(required)); 78 } 79 } 80 81 return CEED_ERROR_SUCCESS; 82 } 83 84 //------------------------------------------------------------------------------ 85 // Apply basis 86 //------------------------------------------------------------------------------ 87 int CeedBasisApplyTensor_Hip_shared(CeedBasis basis, const CeedInt num_elem, CeedTransposeMode t_mode, CeedEvalMode eval_mode, CeedVector u, 88 CeedVector v) { 89 Ceed ceed; 90 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 91 Ceed_Hip *ceed_Hip; 92 CeedCallBackend(CeedGetData(ceed, &ceed_Hip)); 93 CeedBasis_Hip_shared *data; 94 CeedCallBackend(CeedBasisGetData(basis, &data)); 95 CeedInt dim, num_comp; 96 CeedCallBackend(CeedBasisGetDimension(basis, &dim)); 97 CeedCallBackend(CeedBasisGetNumComponents(basis, &num_comp)); 98 99 // Read vectors 100 const CeedScalar *d_u; 101 CeedScalar *d_v; 102 if (eval_mode != CEED_EVAL_WEIGHT) { 103 CeedCallBackend(CeedVectorGetArrayRead(u, CEED_MEM_DEVICE, &d_u)); 104 } 105 CeedCallBackend(CeedVectorGetArrayWrite(v, CEED_MEM_DEVICE, &d_v)); 106 107 // Apply basis operation 108 switch (eval_mode) { 109 case CEED_EVAL_INTERP: { 110 CeedInt P_1d, Q_1d; 111 CeedInt block_size = data->block_sizes[0]; 112 CeedCallBackend(CeedBasisGetNumNodes1D(basis, &P_1d)); 113 CeedCallBackend(CeedBasisGetNumQuadraturePoints1D(basis, &Q_1d)); 114 CeedInt thread_1d = CeedIntMax(Q_1d, P_1d); 115 void *interp_args[] = {(void *)&num_elem, &data->d_interp_1d, &d_u, &d_v}; 116 if (dim == 1) { 117 CeedInt elems_per_block = 64 * thread_1d > 256 ? 256 / thread_1d : 64; 118 elems_per_block = elems_per_block > 0 ? elems_per_block : 1; 119 CeedInt grid = num_elem / elems_per_block + ((num_elem / elems_per_block * elems_per_block < num_elem) ? 1 : 0); 120 CeedInt shared_mem = elems_per_block * thread_1d * sizeof(CeedScalar); 121 if (t_mode == CEED_TRANSPOSE) { 122 CeedCallBackend(CeedRunKernelDimSharedHip(ceed, data->InterpTranspose, grid, thread_1d, 1, elems_per_block, shared_mem, interp_args)); 123 } else { 124 CeedCallBackend(CeedRunKernelDimSharedHip(ceed, data->Interp, grid, thread_1d, 1, elems_per_block, shared_mem, interp_args)); 125 } 126 } else if (dim == 2) { 127 // Check if required threads is small enough to do multiple elems 128 const CeedInt elems_per_block = CeedIntMax(block_size / (thread_1d * thread_1d), 1); 129 CeedInt grid = num_elem / elems_per_block + ((num_elem / elems_per_block * elems_per_block < num_elem) ? 1 : 0); 130 CeedInt shared_mem = elems_per_block * thread_1d * thread_1d * sizeof(CeedScalar); 131 if (t_mode == CEED_TRANSPOSE) { 132 CeedCallBackend( 133 CeedRunKernelDimSharedHip(ceed, data->InterpTranspose, grid, thread_1d, thread_1d, elems_per_block, shared_mem, interp_args)); 134 } else { 135 CeedCallBackend(CeedRunKernelDimSharedHip(ceed, data->Interp, grid, thread_1d, thread_1d, elems_per_block, shared_mem, interp_args)); 136 } 137 } else if (dim == 3) { 138 const CeedInt elems_per_block = CeedIntMax(block_size / (thread_1d * thread_1d), 1); 139 CeedInt grid = num_elem / elems_per_block + ((num_elem / elems_per_block * elems_per_block < num_elem) ? 1 : 0); 140 CeedInt shared_mem = elems_per_block * thread_1d * thread_1d * sizeof(CeedScalar); 141 if (t_mode == CEED_TRANSPOSE) { 142 CeedCallBackend( 143 CeedRunKernelDimSharedHip(ceed, data->InterpTranspose, grid, thread_1d, thread_1d, elems_per_block, shared_mem, interp_args)); 144 } else { 145 CeedCallBackend(CeedRunKernelDimSharedHip(ceed, data->Interp, grid, thread_1d, thread_1d, elems_per_block, shared_mem, interp_args)); 146 } 147 } 148 } break; 149 case CEED_EVAL_GRAD: { 150 CeedInt P_1d, Q_1d; 151 CeedInt block_size = data->block_sizes[1]; 152 CeedCallBackend(CeedBasisGetNumNodes1D(basis, &P_1d)); 153 CeedCallBackend(CeedBasisGetNumQuadraturePoints1D(basis, &Q_1d)); 154 CeedInt thread_1d = CeedIntMax(Q_1d, P_1d); 155 CeedScalar *d_grad_1d = data->d_grad_1d; 156 if (data->d_collo_grad_1d) { 157 d_grad_1d = data->d_collo_grad_1d; 158 } 159 void *grad_args[] = {(void *)&num_elem, &data->d_interp_1d, &d_grad_1d, &d_u, &d_v}; 160 if (dim == 1) { 161 CeedInt elems_per_block = 64 * thread_1d > 256 ? 256 / thread_1d : 64; 162 elems_per_block = elems_per_block > 0 ? elems_per_block : 1; 163 CeedInt grid = num_elem / elems_per_block + ((num_elem / elems_per_block * elems_per_block < num_elem) ? 1 : 0); 164 CeedInt shared_mem = elems_per_block * thread_1d * sizeof(CeedScalar); 165 if (t_mode == CEED_TRANSPOSE) { 166 CeedCallBackend(CeedRunKernelDimSharedHip(ceed, data->GradTranspose, grid, thread_1d, 1, elems_per_block, shared_mem, grad_args)); 167 } else { 168 CeedCallBackend(CeedRunKernelDimSharedHip(ceed, data->Grad, grid, thread_1d, 1, elems_per_block, shared_mem, grad_args)); 169 } 170 } else if (dim == 2) { 171 // Check if required threads is small enough to do multiple elems 172 const CeedInt elems_per_block = CeedIntMax(block_size / (thread_1d * thread_1d), 1); 173 CeedInt grid = num_elem / elems_per_block + ((num_elem / elems_per_block * elems_per_block < num_elem) ? 1 : 0); 174 CeedInt shared_mem = elems_per_block * thread_1d * thread_1d * sizeof(CeedScalar); 175 if (t_mode == CEED_TRANSPOSE) { 176 CeedCallBackend(CeedRunKernelDimSharedHip(ceed, data->GradTranspose, grid, thread_1d, thread_1d, elems_per_block, shared_mem, grad_args)); 177 } else { 178 CeedCallBackend(CeedRunKernelDimSharedHip(ceed, data->Grad, grid, thread_1d, thread_1d, elems_per_block, shared_mem, grad_args)); 179 } 180 } else if (dim == 3) { 181 const CeedInt elems_per_block = CeedIntMax(block_size / (thread_1d * thread_1d), 1); 182 CeedInt grid = num_elem / elems_per_block + ((num_elem / elems_per_block * elems_per_block < num_elem) ? 1 : 0); 183 CeedInt shared_mem = elems_per_block * thread_1d * thread_1d * sizeof(CeedScalar); 184 if (t_mode == CEED_TRANSPOSE) { 185 CeedCallBackend(CeedRunKernelDimSharedHip(ceed, data->GradTranspose, grid, thread_1d, thread_1d, elems_per_block, shared_mem, grad_args)); 186 } else { 187 CeedCallBackend(CeedRunKernelDimSharedHip(ceed, data->Grad, grid, thread_1d, thread_1d, elems_per_block, shared_mem, grad_args)); 188 } 189 } 190 } break; 191 case CEED_EVAL_WEIGHT: { 192 CeedInt Q_1d; 193 CeedInt block_size = data->block_sizes[2]; 194 CeedCallBackend(CeedBasisGetNumQuadraturePoints1D(basis, &Q_1d)); 195 void *weight_args[] = {(void *)&num_elem, (void *)&data->d_q_weight_1d, &d_v}; 196 if (dim == 1) { 197 const CeedInt opt_elems = block_size / Q_1d; 198 const CeedInt elems_per_block = opt_elems > 0 ? opt_elems : 1; 199 const CeedInt grid_size = num_elem / elems_per_block + ((num_elem / elems_per_block * elems_per_block < num_elem) ? 1 : 0); 200 CeedCallBackend(CeedRunKernelDimHip(ceed, data->Weight, grid_size, Q_1d, elems_per_block, 1, weight_args)); 201 } else if (dim == 2) { 202 const CeedInt opt_elems = block_size / (Q_1d * Q_1d); 203 const CeedInt elems_per_block = opt_elems > 0 ? opt_elems : 1; 204 const CeedInt grid_size = num_elem / elems_per_block + ((num_elem / elems_per_block * elems_per_block < num_elem) ? 1 : 0); 205 CeedCallBackend(CeedRunKernelDimHip(ceed, data->Weight, grid_size, Q_1d, Q_1d, elems_per_block, weight_args)); 206 } else if (dim == 3) { 207 const CeedInt opt_elems = block_size / (Q_1d * Q_1d); 208 const CeedInt elems_per_block = opt_elems > 0 ? opt_elems : 1; 209 const CeedInt grid_size = num_elem / elems_per_block + ((num_elem / elems_per_block * elems_per_block < num_elem) ? 1 : 0); 210 CeedCallBackend(CeedRunKernelDimHip(ceed, data->Weight, grid_size, Q_1d, Q_1d, elems_per_block, weight_args)); 211 } 212 } break; 213 // LCOV_EXCL_START 214 // Evaluate the divergence to/from the quadrature points 215 case CEED_EVAL_DIV: 216 return CeedError(ceed, CEED_ERROR_BACKEND, "CEED_EVAL_DIV not supported"); 217 // Evaluate the curl to/from the quadrature points 218 case CEED_EVAL_CURL: 219 return CeedError(ceed, CEED_ERROR_BACKEND, "CEED_EVAL_CURL not supported"); 220 // Take no action, BasisApply should not have been called 221 case CEED_EVAL_NONE: 222 return CeedError(ceed, CEED_ERROR_BACKEND, "CEED_EVAL_NONE does not make sense in this context"); 223 // LCOV_EXCL_STOP 224 } 225 226 // Restore vectors 227 if (eval_mode != CEED_EVAL_WEIGHT) { 228 CeedCallBackend(CeedVectorRestoreArrayRead(u, &d_u)); 229 } 230 CeedCallBackend(CeedVectorRestoreArray(v, &d_v)); 231 return CEED_ERROR_SUCCESS; 232 } 233 234 //------------------------------------------------------------------------------ 235 // Destroy basis 236 //------------------------------------------------------------------------------ 237 static int CeedBasisDestroy_Hip_shared(CeedBasis basis) { 238 Ceed ceed; 239 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 240 241 CeedBasis_Hip_shared *data; 242 CeedCallBackend(CeedBasisGetData(basis, &data)); 243 244 CeedCallHip(ceed, hipModuleUnload(data->module)); 245 246 CeedCallHip(ceed, hipFree(data->d_q_weight_1d)); 247 CeedCallHip(ceed, hipFree(data->d_interp_1d)); 248 CeedCallHip(ceed, hipFree(data->d_grad_1d)); 249 CeedCallHip(ceed, hipFree(data->d_collo_grad_1d)); 250 CeedCallBackend(CeedFree(&data)); 251 252 return CEED_ERROR_SUCCESS; 253 } 254 255 //------------------------------------------------------------------------------ 256 // Create tensor basis 257 //------------------------------------------------------------------------------ 258 int CeedBasisCreateTensorH1_Hip_shared(CeedInt dim, CeedInt P_1d, CeedInt Q_1d, const CeedScalar *interp_1d, const CeedScalar *grad_1d, 259 const CeedScalar *q_ref1d, const CeedScalar *q_weight_1d, CeedBasis basis) { 260 Ceed ceed; 261 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 262 CeedBasis_Hip_shared *data; 263 CeedCallBackend(CeedCalloc(1, &data)); 264 265 // Copy basis data to GPU 266 const CeedInt qBytes = Q_1d * sizeof(CeedScalar); 267 CeedCallHip(ceed, hipMalloc((void **)&data->d_q_weight_1d, qBytes)); 268 CeedCallHip(ceed, hipMemcpy(data->d_q_weight_1d, q_weight_1d, qBytes, hipMemcpyHostToDevice)); 269 270 const CeedInt iBytes = qBytes * P_1d; 271 CeedCallHip(ceed, hipMalloc((void **)&data->d_interp_1d, iBytes)); 272 CeedCallHip(ceed, hipMemcpy(data->d_interp_1d, interp_1d, iBytes, hipMemcpyHostToDevice)); 273 274 CeedCallHip(ceed, hipMalloc((void **)&data->d_grad_1d, iBytes)); 275 CeedCallHip(ceed, hipMemcpy(data->d_grad_1d, grad_1d, iBytes, hipMemcpyHostToDevice)); 276 277 // Compute collocated gradient and copy to GPU 278 data->d_collo_grad_1d = NULL; 279 bool has_collocated_grad = dim == 3 && Q_1d >= P_1d; 280 if (has_collocated_grad) { 281 CeedScalar *collo_grad_1d; 282 CeedCallBackend(CeedMalloc(Q_1d * Q_1d, &collo_grad_1d)); 283 CeedCallBackend(CeedBasisGetCollocatedGrad(basis, collo_grad_1d)); 284 CeedCallHip(ceed, hipMalloc((void **)&data->d_collo_grad_1d, qBytes * Q_1d)); 285 CeedCallHip(ceed, hipMemcpy(data->d_collo_grad_1d, collo_grad_1d, qBytes * Q_1d, hipMemcpyHostToDevice)); 286 CeedCallBackend(CeedFree(&collo_grad_1d)); 287 } 288 289 // Set number of threads per block for basis kernels 290 CeedInt num_comp; 291 CeedCallBackend(CeedBasisGetNumComponents(basis, &num_comp)); 292 CeedCallBackend(ComputeBasisThreadBlockSizes(dim, P_1d, Q_1d, num_comp, data->block_sizes)); 293 294 // Compile basis kernels 295 char *basis_kernel_path, *basis_kernel_source; 296 CeedCallBackend(CeedGetJitAbsolutePath(ceed, "ceed/jit-source/hip/hip-shared-basis-tensor.h", &basis_kernel_path)); 297 CeedDebug256(ceed, 2, "----- Loading Basis Kernel Source -----\n"); 298 CeedCallBackend(CeedLoadSourceToBuffer(ceed, basis_kernel_path, &basis_kernel_source)); 299 CeedDebug256(ceed, 2, "----- Loading Basis Kernel Source Complete! -----\n"); 300 CeedCallBackend(CeedCompileHip(ceed, basis_kernel_source, &data->module, 11, "BASIS_Q_1D", Q_1d, "BASIS_P_1D", P_1d, "T_1D", CeedIntMax(Q_1d, P_1d), 301 "BASIS_DIM", dim, "BASIS_NUM_COMP", num_comp, "BASIS_NUM_NODES", CeedIntPow(P_1d, dim), "BASIS_NUM_QPTS", 302 CeedIntPow(Q_1d, dim), "BASIS_INTERP_BLOCK_SIZE", data->block_sizes[0], "BASIS_GRAD_BLOCK_SIZE", 303 data->block_sizes[1], "BASIS_WEIGHT_BLOCK_SIZE", data->block_sizes[2], "BASIS_HAS_COLLOCATED_GRAD", 304 has_collocated_grad)); 305 CeedCallBackend(CeedGetKernelHip(ceed, data->module, "Interp", &data->Interp)); 306 CeedCallBackend(CeedGetKernelHip(ceed, data->module, "InterpTranspose", &data->InterpTranspose)); 307 CeedCallBackend(CeedGetKernelHip(ceed, data->module, "Grad", &data->Grad)); 308 CeedCallBackend(CeedGetKernelHip(ceed, data->module, "GradTranspose", &data->GradTranspose)); 309 CeedCallBackend(CeedGetKernelHip(ceed, data->module, "Weight", &data->Weight)); 310 CeedCallBackend(CeedFree(&basis_kernel_path)); 311 CeedCallBackend(CeedFree(&basis_kernel_source)); 312 313 CeedCallBackend(CeedBasisSetData(basis, data)); 314 315 // Register backend functions 316 CeedCallBackend(CeedSetBackendFunction(ceed, "Basis", basis, "Apply", CeedBasisApplyTensor_Hip_shared)); 317 CeedCallBackend(CeedSetBackendFunction(ceed, "Basis", basis, "Destroy", CeedBasisDestroy_Hip_shared)); 318 return CEED_ERROR_SUCCESS; 319 } 320 //------------------------------------------------------------------------------ 321