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 <stdbool.h> 12 #include <stddef.h> 13 #include <hip/hip_runtime.h> 14 15 #include "../hip/ceed-hip-common.h" 16 #include "../hip/ceed-hip-compile.h" 17 #include "ceed-hip-shared.h" 18 19 //------------------------------------------------------------------------------ 20 // Compute a block size based on required minimum threads 21 //------------------------------------------------------------------------------ 22 static CeedInt ComputeBlockSizeFromRequirement(const CeedInt required) { 23 CeedInt maxSize = 1024; // Max total threads per block 24 CeedInt currentSize = 64; // Start with one group 25 26 while (currentSize < maxSize) { 27 if (currentSize > required) break; 28 else currentSize = currentSize * 2; 29 } 30 return currentSize; 31 } 32 33 //------------------------------------------------------------------------------ 34 // Compute required thread block sizes for basis kernels given P, Q, dim, and 35 // num_comp (num_comp not currently used, but may be again in other basis 36 // parallelization options) 37 //------------------------------------------------------------------------------ 38 static int ComputeBasisThreadBlockSizes(const CeedInt dim, const CeedInt P_1d, const CeedInt Q_1d, const CeedInt num_comp, CeedInt *block_sizes) { 39 // Note that this will use the same block sizes for all dimensions when compiling, 40 // but as each basis object is defined for a particular dimension, we will never 41 // call any kernels except the ones for the dimension for which we have computed the 42 // block sizes. 43 const CeedInt thread_1d = CeedIntMax(P_1d, Q_1d); 44 45 switch (dim) { 46 case 1: { 47 // Interp kernels: 48 block_sizes[0] = 256; 49 50 // Grad kernels: 51 block_sizes[1] = 256; 52 53 // Weight kernels: 54 block_sizes[2] = 256; 55 } break; 56 case 2: { 57 // Interp kernels: 58 CeedInt required = thread_1d * thread_1d; 59 60 block_sizes[0] = CeedIntMax(256, ComputeBlockSizeFromRequirement(required)); 61 62 // Grad kernels: currently use same required minimum threads 63 block_sizes[1] = CeedIntMax(256, ComputeBlockSizeFromRequirement(required)); 64 65 // Weight kernels: 66 required = CeedIntMax(64, Q_1d * Q_1d); 67 block_sizes[2] = CeedIntMax(256, ComputeBlockSizeFromRequirement(required)); 68 69 } break; 70 case 3: { 71 // Interp kernels: 72 CeedInt required = thread_1d * thread_1d; 73 74 block_sizes[0] = CeedIntMax(256, ComputeBlockSizeFromRequirement(required)); 75 76 // Grad kernels: currently use same required minimum threads 77 block_sizes[1] = CeedIntMax(256, ComputeBlockSizeFromRequirement(required)); 78 79 // Weight kernels: 80 required = Q_1d * Q_1d * Q_1d; 81 block_sizes[2] = CeedIntMax(256, ComputeBlockSizeFromRequirement(required)); 82 } 83 } 84 85 return CEED_ERROR_SUCCESS; 86 } 87 88 //------------------------------------------------------------------------------ 89 // Apply basis 90 //------------------------------------------------------------------------------ 91 int CeedBasisApplyTensor_Hip_shared(CeedBasis basis, const CeedInt num_elem, CeedTransposeMode t_mode, CeedEvalMode eval_mode, CeedVector u, 92 CeedVector v) { 93 Ceed ceed; 94 Ceed_Hip *ceed_Hip; 95 CeedInt dim, num_comp; 96 const CeedScalar *d_u; 97 CeedScalar *d_v; 98 CeedBasis_Hip_shared *data; 99 100 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 101 CeedCallBackend(CeedGetData(ceed, &ceed_Hip)); 102 CeedCallBackend(CeedBasisGetData(basis, &data)); 103 CeedCallBackend(CeedBasisGetDimension(basis, &dim)); 104 CeedCallBackend(CeedBasisGetNumComponents(basis, &num_comp)); 105 106 // Read vectors 107 if (u != CEED_VECTOR_NONE) CeedCallBackend(CeedVectorGetArrayRead(u, CEED_MEM_DEVICE, &d_u)); 108 else CeedCheck(eval_mode == CEED_EVAL_WEIGHT, ceed, CEED_ERROR_BACKEND, "An input vector is required for this CeedEvalMode"); 109 CeedCallBackend(CeedVectorGetArrayWrite(v, CEED_MEM_DEVICE, &d_v)); 110 111 // Apply basis operation 112 switch (eval_mode) { 113 case CEED_EVAL_INTERP: { 114 CeedInt P_1d, Q_1d; 115 CeedInt block_size = data->block_sizes[0]; 116 117 CeedCallBackend(CeedBasisGetNumNodes1D(basis, &P_1d)); 118 CeedCallBackend(CeedBasisGetNumQuadraturePoints1D(basis, &Q_1d)); 119 CeedInt thread_1d = CeedIntMax(Q_1d, P_1d); 120 void *interp_args[] = {(void *)&num_elem, &data->d_interp_1d, &d_u, &d_v}; 121 122 if (dim == 1) { 123 CeedInt elems_per_block = 64 * thread_1d > 256 ? 256 / thread_1d : 64; 124 elems_per_block = elems_per_block > 0 ? elems_per_block : 1; 125 CeedInt grid = num_elem / elems_per_block + ((num_elem / elems_per_block * elems_per_block < num_elem) ? 1 : 0); 126 CeedInt shared_mem = elems_per_block * thread_1d * sizeof(CeedScalar); 127 128 if (t_mode == CEED_TRANSPOSE) { 129 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, data->InterpTranspose, grid, thread_1d, 1, elems_per_block, shared_mem, interp_args)); 130 } else { 131 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, data->Interp, grid, thread_1d, 1, elems_per_block, shared_mem, interp_args)); 132 } 133 } else if (dim == 2) { 134 // Check if required threads is small enough to do multiple elems 135 const CeedInt elems_per_block = CeedIntMax(block_size / (thread_1d * thread_1d), 1); 136 CeedInt grid = num_elem / elems_per_block + ((num_elem / elems_per_block * elems_per_block < num_elem) ? 1 : 0); 137 CeedInt shared_mem = elems_per_block * thread_1d * thread_1d * sizeof(CeedScalar); 138 139 if (t_mode == CEED_TRANSPOSE) { 140 CeedCallBackend( 141 CeedRunKernelDimShared_Hip(ceed, data->InterpTranspose, grid, thread_1d, thread_1d, elems_per_block, shared_mem, interp_args)); 142 } else { 143 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, data->Interp, grid, thread_1d, thread_1d, elems_per_block, shared_mem, interp_args)); 144 } 145 } else if (dim == 3) { 146 const CeedInt elems_per_block = CeedIntMax(block_size / (thread_1d * thread_1d), 1); 147 CeedInt grid = num_elem / elems_per_block + ((num_elem / elems_per_block * elems_per_block < num_elem) ? 1 : 0); 148 CeedInt shared_mem = elems_per_block * thread_1d * thread_1d * sizeof(CeedScalar); 149 150 if (t_mode == CEED_TRANSPOSE) { 151 CeedCallBackend( 152 CeedRunKernelDimShared_Hip(ceed, data->InterpTranspose, grid, thread_1d, thread_1d, elems_per_block, shared_mem, interp_args)); 153 } else { 154 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, data->Interp, grid, thread_1d, thread_1d, elems_per_block, shared_mem, interp_args)); 155 } 156 } 157 } break; 158 case CEED_EVAL_GRAD: { 159 CeedInt P_1d, Q_1d; 160 CeedInt block_size = data->block_sizes[1]; 161 162 CeedCallBackend(CeedBasisGetNumNodes1D(basis, &P_1d)); 163 CeedCallBackend(CeedBasisGetNumQuadraturePoints1D(basis, &Q_1d)); 164 CeedInt thread_1d = CeedIntMax(Q_1d, P_1d); 165 CeedScalar *d_grad_1d = data->d_grad_1d; 166 167 if (data->d_collo_grad_1d) { 168 d_grad_1d = data->d_collo_grad_1d; 169 } 170 void *grad_args[] = {(void *)&num_elem, &data->d_interp_1d, &d_grad_1d, &d_u, &d_v}; 171 if (dim == 1) { 172 CeedInt elems_per_block = 64 * thread_1d > 256 ? 256 / thread_1d : 64; 173 elems_per_block = elems_per_block > 0 ? elems_per_block : 1; 174 CeedInt grid = num_elem / elems_per_block + ((num_elem / elems_per_block * elems_per_block < num_elem) ? 1 : 0); 175 CeedInt shared_mem = elems_per_block * thread_1d * sizeof(CeedScalar); 176 177 if (t_mode == CEED_TRANSPOSE) { 178 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, data->GradTranspose, grid, thread_1d, 1, elems_per_block, shared_mem, grad_args)); 179 } else { 180 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, data->Grad, grid, thread_1d, 1, elems_per_block, shared_mem, grad_args)); 181 } 182 } else if (dim == 2) { 183 // Check if required threads is small enough to do multiple elems 184 const CeedInt elems_per_block = CeedIntMax(block_size / (thread_1d * thread_1d), 1); 185 CeedInt grid = num_elem / elems_per_block + ((num_elem / elems_per_block * elems_per_block < num_elem) ? 1 : 0); 186 CeedInt shared_mem = elems_per_block * thread_1d * thread_1d * sizeof(CeedScalar); 187 188 if (t_mode == CEED_TRANSPOSE) { 189 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, data->GradTranspose, grid, thread_1d, thread_1d, elems_per_block, shared_mem, grad_args)); 190 } else { 191 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, data->Grad, grid, thread_1d, thread_1d, elems_per_block, shared_mem, grad_args)); 192 } 193 } else if (dim == 3) { 194 const CeedInt elems_per_block = CeedIntMax(block_size / (thread_1d * thread_1d), 1); 195 CeedInt grid = num_elem / elems_per_block + ((num_elem / elems_per_block * elems_per_block < num_elem) ? 1 : 0); 196 CeedInt shared_mem = elems_per_block * thread_1d * thread_1d * sizeof(CeedScalar); 197 198 if (t_mode == CEED_TRANSPOSE) { 199 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, data->GradTranspose, grid, thread_1d, thread_1d, elems_per_block, shared_mem, grad_args)); 200 } else { 201 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, data->Grad, grid, thread_1d, thread_1d, elems_per_block, shared_mem, grad_args)); 202 } 203 } 204 } break; 205 case CEED_EVAL_WEIGHT: { 206 CeedInt Q_1d; 207 CeedInt block_size = data->block_sizes[2]; 208 209 CeedCallBackend(CeedBasisGetNumQuadraturePoints1D(basis, &Q_1d)); 210 void *weight_args[] = {(void *)&num_elem, (void *)&data->d_q_weight_1d, &d_v}; 211 212 if (dim == 1) { 213 const CeedInt opt_elems = block_size / Q_1d; 214 const CeedInt elems_per_block = opt_elems > 0 ? opt_elems : 1; 215 const CeedInt grid_size = num_elem / elems_per_block + ((num_elem / elems_per_block * elems_per_block < num_elem) ? 1 : 0); 216 217 CeedCallBackend(CeedRunKernelDim_Hip(ceed, data->Weight, grid_size, Q_1d, elems_per_block, 1, weight_args)); 218 } else if (dim == 2) { 219 const CeedInt opt_elems = block_size / (Q_1d * Q_1d); 220 const CeedInt elems_per_block = opt_elems > 0 ? opt_elems : 1; 221 const CeedInt grid_size = num_elem / elems_per_block + ((num_elem / elems_per_block * elems_per_block < num_elem) ? 1 : 0); 222 223 CeedCallBackend(CeedRunKernelDim_Hip(ceed, data->Weight, grid_size, Q_1d, Q_1d, elems_per_block, weight_args)); 224 } else if (dim == 3) { 225 const CeedInt opt_elems = block_size / (Q_1d * Q_1d); 226 const CeedInt elems_per_block = opt_elems > 0 ? opt_elems : 1; 227 const CeedInt grid_size = num_elem / elems_per_block + ((num_elem / elems_per_block * elems_per_block < num_elem) ? 1 : 0); 228 229 CeedCallBackend(CeedRunKernelDim_Hip(ceed, data->Weight, grid_size, Q_1d, Q_1d, elems_per_block, weight_args)); 230 } 231 } break; 232 // LCOV_EXCL_START 233 // Evaluate the divergence to/from the quadrature points 234 case CEED_EVAL_DIV: 235 return CeedError(ceed, CEED_ERROR_BACKEND, "CEED_EVAL_DIV not supported"); 236 // Evaluate the curl to/from the quadrature points 237 case CEED_EVAL_CURL: 238 return CeedError(ceed, CEED_ERROR_BACKEND, "CEED_EVAL_CURL not supported"); 239 // Take no action, BasisApply should not have been called 240 case CEED_EVAL_NONE: 241 return CeedError(ceed, CEED_ERROR_BACKEND, "CEED_EVAL_NONE does not make sense in this context"); 242 // LCOV_EXCL_STOP 243 } 244 245 // Restore vectors 246 if (eval_mode != CEED_EVAL_WEIGHT) { 247 CeedCallBackend(CeedVectorRestoreArrayRead(u, &d_u)); 248 } 249 CeedCallBackend(CeedVectorRestoreArray(v, &d_v)); 250 return CEED_ERROR_SUCCESS; 251 } 252 253 //------------------------------------------------------------------------------ 254 // Destroy basis 255 //------------------------------------------------------------------------------ 256 static int CeedBasisDestroy_Hip_shared(CeedBasis basis) { 257 Ceed ceed; 258 CeedBasis_Hip_shared *data; 259 260 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 261 CeedCallBackend(CeedBasisGetData(basis, &data)); 262 CeedCallHip(ceed, hipModuleUnload(data->module)); 263 CeedCallHip(ceed, hipFree(data->d_q_weight_1d)); 264 CeedCallHip(ceed, hipFree(data->d_interp_1d)); 265 CeedCallHip(ceed, hipFree(data->d_grad_1d)); 266 CeedCallHip(ceed, hipFree(data->d_collo_grad_1d)); 267 CeedCallBackend(CeedFree(&data)); 268 return CEED_ERROR_SUCCESS; 269 } 270 271 //------------------------------------------------------------------------------ 272 // Create tensor basis 273 //------------------------------------------------------------------------------ 274 int CeedBasisCreateTensorH1_Hip_shared(CeedInt dim, CeedInt P_1d, CeedInt Q_1d, const CeedScalar *interp_1d, const CeedScalar *grad_1d, 275 const CeedScalar *q_ref_1d, const CeedScalar *q_weight_1d, CeedBasis basis) { 276 Ceed ceed; 277 char *basis_kernel_path, *basis_kernel_source; 278 CeedInt num_comp; 279 const CeedInt q_bytes = Q_1d * sizeof(CeedScalar); 280 const CeedInt i_bytes = q_bytes * P_1d; 281 CeedBasis_Hip_shared *data; 282 283 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 284 CeedCallBackend(CeedCalloc(1, &data)); 285 286 // Copy basis data to GPU 287 CeedCallHip(ceed, hipMalloc((void **)&data->d_q_weight_1d, q_bytes)); 288 CeedCallHip(ceed, hipMemcpy(data->d_q_weight_1d, q_weight_1d, q_bytes, hipMemcpyHostToDevice)); 289 CeedCallHip(ceed, hipMalloc((void **)&data->d_interp_1d, i_bytes)); 290 CeedCallHip(ceed, hipMemcpy(data->d_interp_1d, interp_1d, i_bytes, hipMemcpyHostToDevice)); 291 CeedCallHip(ceed, hipMalloc((void **)&data->d_grad_1d, i_bytes)); 292 CeedCallHip(ceed, hipMemcpy(data->d_grad_1d, grad_1d, i_bytes, hipMemcpyHostToDevice)); 293 294 // Compute collocated gradient and copy to GPU 295 data->d_collo_grad_1d = NULL; 296 bool has_collocated_grad = dim == 3 && Q_1d >= P_1d; 297 298 if (has_collocated_grad) { 299 CeedScalar *collo_grad_1d; 300 301 CeedCallBackend(CeedMalloc(Q_1d * Q_1d, &collo_grad_1d)); 302 CeedCallBackend(CeedBasisGetCollocatedGrad(basis, collo_grad_1d)); 303 CeedCallHip(ceed, hipMalloc((void **)&data->d_collo_grad_1d, q_bytes * Q_1d)); 304 CeedCallHip(ceed, hipMemcpy(data->d_collo_grad_1d, collo_grad_1d, q_bytes * Q_1d, hipMemcpyHostToDevice)); 305 CeedCallBackend(CeedFree(&collo_grad_1d)); 306 } 307 308 // Set number of threads per block for basis kernels 309 CeedCallBackend(CeedBasisGetNumComponents(basis, &num_comp)); 310 CeedCallBackend(ComputeBasisThreadBlockSizes(dim, P_1d, Q_1d, num_comp, data->block_sizes)); 311 312 // Compile basis kernels 313 CeedCallBackend(CeedGetJitAbsolutePath(ceed, "ceed/jit-source/hip/hip-shared-basis-tensor.h", &basis_kernel_path)); 314 CeedDebug256(ceed, CEED_DEBUG_COLOR_SUCCESS, "----- Loading Basis Kernel Source -----\n"); 315 CeedCallBackend(CeedLoadSourceToBuffer(ceed, basis_kernel_path, &basis_kernel_source)); 316 CeedDebug256(ceed, CEED_DEBUG_COLOR_SUCCESS, "----- Loading Basis Kernel Source Complete! -----\n"); 317 CeedCallBackend(CeedCompile_Hip(ceed, basis_kernel_source, &data->module, 11, "BASIS_Q_1D", Q_1d, "BASIS_P_1D", P_1d, "T_1D", 318 CeedIntMax(Q_1d, P_1d), "BASIS_DIM", dim, "BASIS_NUM_COMP", num_comp, "BASIS_NUM_NODES", CeedIntPow(P_1d, dim), 319 "BASIS_NUM_QPTS", CeedIntPow(Q_1d, dim), "BASIS_INTERP_BLOCK_SIZE", data->block_sizes[0], "BASIS_GRAD_BLOCK_SIZE", 320 data->block_sizes[1], "BASIS_WEIGHT_BLOCK_SIZE", data->block_sizes[2], "BASIS_HAS_COLLOCATED_GRAD", 321 has_collocated_grad)); 322 CeedCallBackend(CeedGetKernel_Hip(ceed, data->module, "Interp", &data->Interp)); 323 CeedCallBackend(CeedGetKernel_Hip(ceed, data->module, "InterpTranspose", &data->InterpTranspose)); 324 CeedCallBackend(CeedGetKernel_Hip(ceed, data->module, "Grad", &data->Grad)); 325 CeedCallBackend(CeedGetKernel_Hip(ceed, data->module, "GradTranspose", &data->GradTranspose)); 326 CeedCallBackend(CeedGetKernel_Hip(ceed, data->module, "Weight", &data->Weight)); 327 CeedCallBackend(CeedFree(&basis_kernel_path)); 328 CeedCallBackend(CeedFree(&basis_kernel_source)); 329 330 CeedCallBackend(CeedBasisSetData(basis, data)); 331 332 // Register backend functions 333 CeedCallBackend(CeedSetBackendFunction(ceed, "Basis", basis, "Apply", CeedBasisApplyTensor_Hip_shared)); 334 CeedCallBackend(CeedSetBackendFunction(ceed, "Basis", basis, "Destroy", CeedBasisDestroy_Hip_shared)); 335 return CEED_ERROR_SUCCESS; 336 } 337 338 //------------------------------------------------------------------------------ 339