1 // Copyright (c) 2017-2024, 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 <string.h> 14 #include <hip/hip_runtime.h> 15 16 #include "../hip/ceed-hip-common.h" 17 #include "../hip/ceed-hip-compile.h" 18 #include "ceed-hip-shared.h" 19 20 //------------------------------------------------------------------------------ 21 // Compute a block size based on required minimum threads 22 //------------------------------------------------------------------------------ 23 static CeedInt ComputeBlockSizeFromRequirement(const CeedInt required) { 24 CeedInt maxSize = 1024; // Max total threads per block 25 CeedInt currentSize = 64; // Start with one group 26 27 while (currentSize < maxSize) { 28 if (currentSize > required) break; 29 else currentSize = currentSize * 2; 30 } 31 return currentSize; 32 } 33 34 //------------------------------------------------------------------------------ 35 // Compute required thread block sizes for basis kernels given P, Q, dim, and 36 // num_comp (num_comp not currently used, but may be again in other basis 37 // parallelization options) 38 //------------------------------------------------------------------------------ 39 static int ComputeBasisThreadBlockSizes(const CeedInt dim, const CeedInt P_1d, const CeedInt Q_1d, const CeedInt num_comp, CeedInt *block_sizes) { 40 // Note that this will use the same block sizes for all dimensions when compiling, 41 // but as each basis object is defined for a particular dimension, we will never 42 // call any kernels except the ones for the dimension for which we have computed the 43 // block sizes. 44 const CeedInt thread_1d = CeedIntMax(P_1d, Q_1d); 45 46 switch (dim) { 47 case 1: { 48 // Interp kernels: 49 block_sizes[0] = 256; 50 51 // Grad kernels: 52 block_sizes[1] = 256; 53 54 // Weight kernels: 55 block_sizes[2] = 256; 56 } break; 57 case 2: { 58 // Interp kernels: 59 CeedInt required = thread_1d * thread_1d; 60 61 block_sizes[0] = CeedIntMax(256, ComputeBlockSizeFromRequirement(required)); 62 63 // Grad kernels: currently use same required minimum threads 64 block_sizes[1] = CeedIntMax(256, ComputeBlockSizeFromRequirement(required)); 65 66 // Weight kernels: 67 required = CeedIntMax(64, Q_1d * Q_1d); 68 block_sizes[2] = CeedIntMax(256, ComputeBlockSizeFromRequirement(required)); 69 70 } break; 71 case 3: { 72 // Interp kernels: 73 CeedInt required = thread_1d * thread_1d; 74 75 block_sizes[0] = CeedIntMax(256, ComputeBlockSizeFromRequirement(required)); 76 77 // Grad kernels: currently use same required minimum threads 78 block_sizes[1] = CeedIntMax(256, ComputeBlockSizeFromRequirement(required)); 79 80 // Weight kernels: 81 required = Q_1d * Q_1d * Q_1d; 82 block_sizes[2] = CeedIntMax(256, ComputeBlockSizeFromRequirement(required)); 83 } 84 } 85 return CEED_ERROR_SUCCESS; 86 } 87 88 //------------------------------------------------------------------------------ 89 // Apply basis 90 //------------------------------------------------------------------------------ 91 static int CeedBasisApplyTensorCore_Hip_shared(CeedBasis basis, bool apply_add, const CeedInt num_elem, CeedTransposeMode t_mode, 92 CeedEvalMode eval_mode, CeedVector u, 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 // Get read/write access to u, v 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 if (apply_add) CeedCallBackend(CeedVectorGetArray(v, CEED_MEM_DEVICE, &d_v)); 110 else CeedCallBackend(CeedVectorGetArrayWrite(v, CEED_MEM_DEVICE, &d_v)); 111 112 // Apply basis operation 113 switch (eval_mode) { 114 case CEED_EVAL_INTERP: { 115 CeedInt P_1d, Q_1d; 116 CeedInt block_size = data->block_sizes[0]; 117 118 CeedCallBackend(CeedBasisGetNumNodes1D(basis, &P_1d)); 119 CeedCallBackend(CeedBasisGetNumQuadraturePoints1D(basis, &Q_1d)); 120 CeedInt thread_1d = CeedIntMax(Q_1d, P_1d); 121 void *interp_args[] = {(void *)&num_elem, &data->d_interp_1d, &d_u, &d_v}; 122 123 if (dim == 1) { 124 CeedInt elems_per_block = 64 * thread_1d > 256 ? 256 / thread_1d : 64; 125 elems_per_block = elems_per_block > 0 ? elems_per_block : 1; 126 CeedInt grid = num_elem / elems_per_block + (num_elem % elems_per_block > 0); 127 CeedInt shared_mem = elems_per_block * thread_1d * sizeof(CeedScalar); 128 129 if (t_mode == CEED_TRANSPOSE) { 130 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, apply_add ? data->InterpTransposeAdd : data->InterpTranspose, grid, thread_1d, 1, 131 elems_per_block, shared_mem, interp_args)); 132 } else { 133 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, data->Interp, grid, thread_1d, 1, elems_per_block, shared_mem, interp_args)); 134 } 135 } else if (dim == 2) { 136 // Check if required threads is small enough to do multiple elems 137 const CeedInt elems_per_block = CeedIntMax(block_size / (thread_1d * thread_1d), 1); 138 CeedInt grid = num_elem / elems_per_block + (num_elem % elems_per_block > 0); 139 CeedInt shared_mem = elems_per_block * thread_1d * thread_1d * sizeof(CeedScalar); 140 141 if (t_mode == CEED_TRANSPOSE) { 142 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, apply_add ? data->InterpTransposeAdd : data->InterpTranspose, grid, thread_1d, thread_1d, 143 elems_per_block, shared_mem, interp_args)); 144 } else { 145 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, data->Interp, grid, thread_1d, thread_1d, elems_per_block, shared_mem, interp_args)); 146 } 147 } else if (dim == 3) { 148 const CeedInt elems_per_block = CeedIntMax(block_size / (thread_1d * thread_1d), 1); 149 CeedInt grid = num_elem / elems_per_block + (num_elem % elems_per_block > 0); 150 CeedInt shared_mem = elems_per_block * thread_1d * thread_1d * sizeof(CeedScalar); 151 152 if (t_mode == CEED_TRANSPOSE) { 153 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, apply_add ? data->InterpTransposeAdd : data->InterpTranspose, grid, thread_1d, thread_1d, 154 elems_per_block, shared_mem, interp_args)); 155 } else { 156 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, data->Interp, grid, thread_1d, thread_1d, elems_per_block, shared_mem, interp_args)); 157 } 158 } 159 } break; 160 case CEED_EVAL_GRAD: { 161 CeedInt P_1d, Q_1d; 162 CeedInt block_size = data->block_sizes[1]; 163 164 CeedCallBackend(CeedBasisGetNumNodes1D(basis, &P_1d)); 165 CeedCallBackend(CeedBasisGetNumQuadraturePoints1D(basis, &Q_1d)); 166 CeedInt thread_1d = CeedIntMax(Q_1d, P_1d); 167 CeedScalar *d_grad_1d = data->d_grad_1d; 168 169 if (data->d_collo_grad_1d) { 170 d_grad_1d = data->d_collo_grad_1d; 171 } 172 void *grad_args[] = {(void *)&num_elem, &data->d_interp_1d, &d_grad_1d, &d_u, &d_v}; 173 if (dim == 1) { 174 CeedInt elems_per_block = 64 * thread_1d > 256 ? 256 / thread_1d : 64; 175 elems_per_block = elems_per_block > 0 ? elems_per_block : 1; 176 CeedInt grid = num_elem / elems_per_block + (num_elem % elems_per_block > 0); 177 CeedInt shared_mem = elems_per_block * thread_1d * sizeof(CeedScalar); 178 179 if (t_mode == CEED_TRANSPOSE) { 180 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, apply_add ? data->GradTransposeAdd : data->GradTranspose, grid, thread_1d, 1, 181 elems_per_block, shared_mem, grad_args)); 182 } else { 183 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, data->Grad, grid, thread_1d, 1, elems_per_block, shared_mem, grad_args)); 184 } 185 } else if (dim == 2) { 186 // Check if required threads is small enough to do multiple elems 187 const CeedInt elems_per_block = CeedIntMax(block_size / (thread_1d * thread_1d), 1); 188 CeedInt grid = num_elem / elems_per_block + (num_elem % elems_per_block > 0); 189 CeedInt shared_mem = elems_per_block * thread_1d * thread_1d * sizeof(CeedScalar); 190 191 if (t_mode == CEED_TRANSPOSE) { 192 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, apply_add ? data->GradTransposeAdd : data->GradTranspose, grid, thread_1d, thread_1d, 193 elems_per_block, shared_mem, grad_args)); 194 } else { 195 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, data->Grad, grid, thread_1d, thread_1d, elems_per_block, shared_mem, grad_args)); 196 } 197 } else if (dim == 3) { 198 const CeedInt elems_per_block = CeedIntMax(block_size / (thread_1d * thread_1d), 1); 199 CeedInt grid = num_elem / elems_per_block + (num_elem % elems_per_block > 0); 200 CeedInt shared_mem = elems_per_block * thread_1d * thread_1d * sizeof(CeedScalar); 201 202 if (t_mode == CEED_TRANSPOSE) { 203 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, apply_add ? data->GradTransposeAdd : data->GradTranspose, grid, thread_1d, thread_1d, 204 elems_per_block, shared_mem, grad_args)); 205 } else { 206 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, data->Grad, grid, thread_1d, thread_1d, elems_per_block, shared_mem, grad_args)); 207 } 208 } 209 } break; 210 case CEED_EVAL_WEIGHT: { 211 CeedInt Q_1d; 212 CeedInt block_size = data->block_sizes[2]; 213 214 CeedCheck(data->d_q_weight_1d, ceed, CEED_ERROR_BACKEND, "%s not supported; q_weights_1d not set", CeedEvalModes[eval_mode]); 215 CeedCallBackend(CeedBasisGetNumQuadraturePoints1D(basis, &Q_1d)); 216 void *weight_args[] = {(void *)&num_elem, (void *)&data->d_q_weight_1d, &d_v}; 217 218 if (dim == 1) { 219 const CeedInt opt_elems = block_size / 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 > 0); 222 223 CeedCallBackend(CeedRunKernelDim_Hip(ceed, data->Weight, grid_size, Q_1d, elems_per_block, 1, weight_args)); 224 } else if (dim == 2) { 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 > 0); 228 229 CeedCallBackend(CeedRunKernelDim_Hip(ceed, data->Weight, grid_size, Q_1d, Q_1d, elems_per_block, weight_args)); 230 } else if (dim == 3) { 231 const CeedInt opt_elems = block_size / (Q_1d * Q_1d); 232 const CeedInt elems_per_block = opt_elems > 0 ? opt_elems : 1; 233 const CeedInt grid_size = num_elem / elems_per_block + (num_elem % elems_per_block > 0); 234 235 CeedCallBackend(CeedRunKernelDim_Hip(ceed, data->Weight, grid_size, Q_1d, Q_1d, elems_per_block, weight_args)); 236 } 237 } break; 238 case CEED_EVAL_NONE: /* handled separately below */ 239 break; 240 // LCOV_EXCL_START 241 case CEED_EVAL_DIV: 242 case CEED_EVAL_CURL: 243 return CeedError(ceed, CEED_ERROR_BACKEND, "%s not supported", CeedEvalModes[eval_mode]); 244 // LCOV_EXCL_STOP 245 } 246 247 // Restore vectors, cover CEED_EVAL_NONE 248 CeedCallBackend(CeedVectorRestoreArray(v, &d_v)); 249 if (eval_mode == CEED_EVAL_NONE) CeedCallBackend(CeedVectorSetArray(v, CEED_MEM_DEVICE, CEED_COPY_VALUES, (CeedScalar *)d_u)); 250 if (eval_mode != CEED_EVAL_WEIGHT) CeedCallBackend(CeedVectorRestoreArrayRead(u, &d_u)); 251 CeedCallBackend(CeedDestroy(&ceed)); 252 return CEED_ERROR_SUCCESS; 253 } 254 255 int CeedBasisApplyTensor_Hip_shared(CeedBasis basis, const CeedInt num_elem, CeedTransposeMode t_mode, CeedEvalMode eval_mode, CeedVector u, 256 CeedVector v) { 257 CeedCallBackend(CeedBasisApplyTensorCore_Hip_shared(basis, false, num_elem, t_mode, eval_mode, u, v)); 258 return CEED_ERROR_SUCCESS; 259 } 260 261 int CeedBasisApplyAddTensor_Hip_shared(CeedBasis basis, const CeedInt num_elem, CeedTransposeMode t_mode, CeedEvalMode eval_mode, CeedVector u, 262 CeedVector v) { 263 CeedCallBackend(CeedBasisApplyTensorCore_Hip_shared(basis, true, num_elem, t_mode, eval_mode, u, v)); 264 return CEED_ERROR_SUCCESS; 265 } 266 267 //------------------------------------------------------------------------------ 268 // Basis apply - tensor AtPoints 269 //------------------------------------------------------------------------------ 270 static int CeedBasisApplyAtPointsCore_Hip_shared(CeedBasis basis, bool apply_add, const CeedInt num_elem, const CeedInt *num_points, 271 CeedTransposeMode t_mode, CeedEvalMode eval_mode, CeedVector x_ref, CeedVector u, CeedVector v) { 272 Ceed ceed; 273 CeedInt Q_1d, dim, max_num_points = num_points[0]; 274 const CeedInt is_transpose = t_mode == CEED_TRANSPOSE; 275 const CeedScalar *d_x, *d_u; 276 CeedScalar *d_v; 277 CeedBasis_Hip_shared *data; 278 279 CeedCallBackend(CeedBasisGetData(basis, &data)); 280 CeedCallBackend(CeedBasisGetNumQuadraturePoints1D(basis, &Q_1d)); 281 CeedCallBackend(CeedBasisGetDimension(basis, &dim)); 282 283 // Weight handled separately 284 if (eval_mode == CEED_EVAL_WEIGHT) { 285 CeedCallBackend(CeedVectorSetValue(v, 1.0)); 286 return CEED_ERROR_SUCCESS; 287 } 288 289 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 290 291 // Check padded to uniform number of points per elem 292 for (CeedInt i = 1; i < num_elem; i++) max_num_points = CeedIntMax(max_num_points, num_points[i]); 293 { 294 CeedInt num_comp, q_comp; 295 CeedSize len, len_required; 296 297 CeedCallBackend(CeedBasisGetNumComponents(basis, &num_comp)); 298 CeedCallBackend(CeedBasisGetNumQuadratureComponents(basis, eval_mode, &q_comp)); 299 CeedCallBackend(CeedVectorGetLength(is_transpose ? u : v, &len)); 300 len_required = (CeedSize)num_comp * (CeedSize)q_comp * (CeedSize)num_elem * (CeedSize)max_num_points; 301 CeedCheck(len >= len_required, ceed, CEED_ERROR_BACKEND, 302 "Vector at points must be padded to the same number of points in each element for BasisApplyAtPoints on GPU backends." 303 " Found %" CeedSize_FMT ", Required %" CeedSize_FMT, 304 len, len_required); 305 } 306 307 // Move num_points array to device 308 if (is_transpose) { 309 const CeedInt num_bytes = num_elem * sizeof(CeedInt); 310 311 if (num_elem != data->num_elem_at_points) { 312 data->num_elem_at_points = num_elem; 313 314 if (data->d_points_per_elem) CeedCallHip(ceed, hipFree(data->d_points_per_elem)); 315 CeedCallHip(ceed, hipMalloc((void **)&data->d_points_per_elem, num_bytes)); 316 CeedCallBackend(CeedFree(&data->h_points_per_elem)); 317 CeedCallBackend(CeedCalloc(num_elem, &data->h_points_per_elem)); 318 } 319 if (memcmp(data->h_points_per_elem, num_points, num_bytes)) { 320 memcpy(data->h_points_per_elem, num_points, num_bytes); 321 CeedCallHip(ceed, hipMemcpy(data->d_points_per_elem, num_points, num_bytes, hipMemcpyHostToDevice)); 322 } 323 } 324 325 // Build kernels if needed 326 if (data->num_points != max_num_points) { 327 CeedInt P_1d; 328 329 CeedCallBackend(CeedBasisGetNumNodes1D(basis, &P_1d)); 330 data->num_points = max_num_points; 331 332 // -- Create interp matrix to Chebyshev coefficients 333 if (!data->d_chebyshev_interp_1d) { 334 CeedSize interp_bytes; 335 CeedScalar *chebyshev_interp_1d; 336 337 interp_bytes = P_1d * Q_1d * sizeof(CeedScalar); 338 CeedCallBackend(CeedCalloc(P_1d * Q_1d, &chebyshev_interp_1d)); 339 CeedCallBackend(CeedBasisGetChebyshevInterp1D(basis, chebyshev_interp_1d)); 340 CeedCallHip(ceed, hipMalloc((void **)&data->d_chebyshev_interp_1d, interp_bytes)); 341 CeedCallHip(ceed, hipMemcpy(data->d_chebyshev_interp_1d, chebyshev_interp_1d, interp_bytes, hipMemcpyHostToDevice)); 342 CeedCallBackend(CeedFree(&chebyshev_interp_1d)); 343 } 344 345 // -- Compile kernels 346 const char basis_kernel_source[] = "// AtPoints basis source\n#include <ceed/jit-source/hip/hip-shared-basis-tensor-at-points.h>\n"; 347 CeedInt num_comp; 348 349 if (data->moduleAtPoints) CeedCallHip(ceed, hipModuleUnload(data->moduleAtPoints)); 350 CeedCallBackend(CeedBasisGetNumComponents(basis, &num_comp)); 351 CeedCallBackend(CeedCompile_Hip(ceed, basis_kernel_source, &data->moduleAtPoints, 9, "BASIS_Q_1D", Q_1d, "BASIS_P_1D", P_1d, "T_1D", 352 CeedIntMax(Q_1d, P_1d), "BASIS_DIM", dim, "BASIS_NUM_COMP", num_comp, "BASIS_NUM_NODES", CeedIntPow(P_1d, dim), 353 "BASIS_NUM_QPTS", CeedIntPow(Q_1d, dim), "BASIS_NUM_PTS", max_num_points, "BASIS_INTERP_BLOCK_SIZE", 354 data->block_sizes[0])); 355 CeedCallBackend(CeedGetKernel_Hip(ceed, data->moduleAtPoints, "InterpAtPoints", &data->InterpAtPoints)); 356 CeedCallBackend(CeedGetKernel_Hip(ceed, data->moduleAtPoints, "InterpTransposeAtPoints", &data->InterpTransposeAtPoints)); 357 CeedCallBackend(CeedGetKernel_Hip(ceed, data->moduleAtPoints, "GradAtPoints", &data->GradAtPoints)); 358 CeedCallBackend(CeedGetKernel_Hip(ceed, data->moduleAtPoints, "GradTransposeAtPoints", &data->GradTransposeAtPoints)); 359 } 360 361 // Get read/write access to u, v 362 CeedCallBackend(CeedVectorGetArrayRead(x_ref, CEED_MEM_DEVICE, &d_x)); 363 if (u != CEED_VECTOR_NONE) CeedCallBackend(CeedVectorGetArrayRead(u, CEED_MEM_DEVICE, &d_u)); 364 else CeedCheck(eval_mode == CEED_EVAL_WEIGHT, ceed, CEED_ERROR_BACKEND, "An input vector is required for this CeedEvalMode"); 365 if (apply_add) CeedCallBackend(CeedVectorGetArray(v, CEED_MEM_DEVICE, &d_v)); 366 else CeedCallBackend(CeedVectorGetArrayWrite(v, CEED_MEM_DEVICE, &d_v)); 367 368 // Clear v for transpose operation 369 if (is_transpose && !apply_add) { 370 CeedInt num_comp, q_comp, num_nodes; 371 CeedSize length; 372 373 CeedCallBackend(CeedBasisGetNumComponents(basis, &num_comp)); 374 CeedCallBackend(CeedBasisGetNumQuadratureComponents(basis, eval_mode, &q_comp)); 375 CeedCallBackend(CeedBasisGetNumNodes(basis, &num_nodes)); 376 length = 377 (CeedSize)num_elem * (CeedSize)num_comp * (t_mode == CEED_TRANSPOSE ? (CeedSize)num_nodes : ((CeedSize)max_num_points * (CeedSize)q_comp)); 378 CeedCallHip(ceed, hipMemset(d_v, 0, length * sizeof(CeedScalar))); 379 } 380 381 // Basis action 382 switch (eval_mode) { 383 case CEED_EVAL_INTERP: { 384 CeedInt P_1d, Q_1d; 385 CeedInt block_size = data->block_sizes[0]; 386 387 CeedCallBackend(CeedBasisGetNumNodes1D(basis, &P_1d)); 388 CeedCallBackend(CeedBasisGetNumQuadraturePoints1D(basis, &Q_1d)); 389 CeedInt thread_1d = CeedIntMax(Q_1d, P_1d); 390 void *interp_args[] = {(void *)&num_elem, &data->d_chebyshev_interp_1d, &data->d_points_per_elem, &d_x, &d_u, &d_v}; 391 392 if (dim == 1) { 393 CeedInt elems_per_block = 64 * thread_1d > 256 ? 256 / thread_1d : 64; 394 elems_per_block = elems_per_block > 0 ? elems_per_block : 1; 395 CeedInt grid = num_elem / elems_per_block + (num_elem % elems_per_block > 0); 396 CeedInt shared_mem = elems_per_block * thread_1d * sizeof(CeedScalar); 397 398 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, is_transpose ? data->InterpTransposeAtPoints : data->InterpAtPoints, grid, thread_1d, 1, 399 elems_per_block, shared_mem, interp_args)); 400 } else if (dim == 2) { 401 // Check if required threads is small enough to do multiple elems 402 const CeedInt elems_per_block = CeedIntMax(block_size / (thread_1d * thread_1d), 1); 403 CeedInt grid = num_elem / elems_per_block + (num_elem % elems_per_block > 0); 404 CeedInt shared_mem = elems_per_block * thread_1d * thread_1d * sizeof(CeedScalar); 405 406 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, is_transpose ? data->InterpTransposeAtPoints : data->InterpAtPoints, grid, thread_1d, 407 thread_1d, elems_per_block, shared_mem, interp_args)); 408 } else if (dim == 3) { 409 const CeedInt elems_per_block = CeedIntMax(block_size / (thread_1d * thread_1d), 1); 410 CeedInt grid = num_elem / elems_per_block + (num_elem % elems_per_block > 0); 411 CeedInt shared_mem = elems_per_block * thread_1d * thread_1d * sizeof(CeedScalar); 412 413 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, is_transpose ? data->InterpTransposeAtPoints : data->InterpAtPoints, grid, thread_1d, 414 thread_1d, elems_per_block, shared_mem, interp_args)); 415 } 416 } break; 417 case CEED_EVAL_GRAD: { 418 CeedInt P_1d, Q_1d; 419 CeedInt block_size = data->block_sizes[0]; 420 421 CeedCallBackend(CeedBasisGetNumNodes1D(basis, &P_1d)); 422 CeedCallBackend(CeedBasisGetNumQuadraturePoints1D(basis, &Q_1d)); 423 CeedInt thread_1d = CeedIntMax(Q_1d, P_1d); 424 void *grad_args[] = {(void *)&num_elem, &data->d_chebyshev_interp_1d, &data->d_points_per_elem, &d_x, &d_u, &d_v}; 425 426 if (dim == 1) { 427 CeedInt elems_per_block = 64 * thread_1d > 256 ? 256 / thread_1d : 64; 428 elems_per_block = elems_per_block > 0 ? elems_per_block : 1; 429 CeedInt grid = num_elem / elems_per_block + (num_elem % elems_per_block > 0); 430 CeedInt shared_mem = elems_per_block * thread_1d * sizeof(CeedScalar); 431 432 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, is_transpose ? data->GradTransposeAtPoints : data->GradAtPoints, grid, thread_1d, 1, 433 elems_per_block, shared_mem, grad_args)); 434 } else if (dim == 2) { 435 // Check if required threads is small enough to do multiple elems 436 const CeedInt elems_per_block = CeedIntMax(block_size / (thread_1d * thread_1d), 1); 437 CeedInt grid = num_elem / elems_per_block + (num_elem % elems_per_block > 0); 438 CeedInt shared_mem = elems_per_block * thread_1d * thread_1d * sizeof(CeedScalar); 439 440 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, is_transpose ? data->GradTransposeAtPoints : data->GradAtPoints, grid, thread_1d, thread_1d, 441 elems_per_block, shared_mem, grad_args)); 442 } else if (dim == 3) { 443 const CeedInt elems_per_block = CeedIntMax(block_size / (thread_1d * thread_1d), 1); 444 CeedInt grid = num_elem / elems_per_block + (num_elem % elems_per_block > 0); 445 CeedInt shared_mem = elems_per_block * thread_1d * thread_1d * sizeof(CeedScalar); 446 447 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, is_transpose ? data->GradTransposeAtPoints : data->GradAtPoints, grid, thread_1d, thread_1d, 448 elems_per_block, shared_mem, grad_args)); 449 } 450 } break; 451 case CEED_EVAL_WEIGHT: 452 case CEED_EVAL_NONE: /* handled separately below */ 453 break; 454 // LCOV_EXCL_START 455 case CEED_EVAL_DIV: 456 case CEED_EVAL_CURL: 457 return CeedError(ceed, CEED_ERROR_BACKEND, "%s not supported", CeedEvalModes[eval_mode]); 458 // LCOV_EXCL_STOP 459 } 460 461 // Restore vectors, cover CEED_EVAL_NONE 462 CeedCallBackend(CeedVectorRestoreArrayRead(x_ref, &d_x)); 463 CeedCallBackend(CeedVectorRestoreArray(v, &d_v)); 464 if (eval_mode == CEED_EVAL_NONE) CeedCallBackend(CeedVectorSetArray(v, CEED_MEM_DEVICE, CEED_COPY_VALUES, (CeedScalar *)d_u)); 465 if (eval_mode != CEED_EVAL_WEIGHT) CeedCallBackend(CeedVectorRestoreArrayRead(u, &d_u)); 466 return CEED_ERROR_SUCCESS; 467 } 468 469 static int CeedBasisApplyAtPoints_Hip_shared(CeedBasis basis, const CeedInt num_elem, const CeedInt *num_points, CeedTransposeMode t_mode, 470 CeedEvalMode eval_mode, CeedVector x_ref, CeedVector u, CeedVector v) { 471 CeedCallBackend(CeedBasisApplyAtPointsCore_Hip_shared(basis, false, num_elem, num_points, t_mode, eval_mode, x_ref, u, v)); 472 return CEED_ERROR_SUCCESS; 473 } 474 475 static int CeedBasisApplyAddAtPoints_Hip_shared(CeedBasis basis, const CeedInt num_elem, const CeedInt *num_points, CeedTransposeMode t_mode, 476 CeedEvalMode eval_mode, CeedVector x_ref, CeedVector u, CeedVector v) { 477 CeedCallBackend(CeedBasisApplyAtPointsCore_Hip_shared(basis, true, num_elem, num_points, t_mode, eval_mode, x_ref, u, v)); 478 return CEED_ERROR_SUCCESS; 479 } 480 481 //------------------------------------------------------------------------------ 482 // Destroy basis 483 //------------------------------------------------------------------------------ 484 static int CeedBasisDestroy_Hip_shared(CeedBasis basis) { 485 Ceed ceed; 486 CeedBasis_Hip_shared *data; 487 488 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 489 CeedCallBackend(CeedBasisGetData(basis, &data)); 490 CeedCallHip(ceed, hipModuleUnload(data->module)); 491 if (data->moduleAtPoints) CeedCallHip(ceed, hipModuleUnload(data->moduleAtPoints)); 492 if (data->d_q_weight_1d) CeedCallHip(ceed, hipFree(data->d_q_weight_1d)); 493 CeedCallBackend(CeedFree(&data->h_points_per_elem)); 494 if (data->d_points_per_elem) CeedCallHip(ceed, hipFree(data->d_points_per_elem)); 495 CeedCallHip(ceed, hipFree(data->d_interp_1d)); 496 CeedCallHip(ceed, hipFree(data->d_grad_1d)); 497 CeedCallHip(ceed, hipFree(data->d_collo_grad_1d)); 498 CeedCallHip(ceed, hipFree(data->d_chebyshev_interp_1d)); 499 CeedCallBackend(CeedFree(&data)); 500 return CEED_ERROR_SUCCESS; 501 } 502 503 //------------------------------------------------------------------------------ 504 // Create tensor basis 505 //------------------------------------------------------------------------------ 506 int CeedBasisCreateTensorH1_Hip_shared(CeedInt dim, CeedInt P_1d, CeedInt Q_1d, const CeedScalar *interp_1d, const CeedScalar *grad_1d, 507 const CeedScalar *q_ref_1d, const CeedScalar *q_weight_1d, CeedBasis basis) { 508 Ceed ceed; 509 CeedInt num_comp; 510 const CeedInt q_bytes = Q_1d * sizeof(CeedScalar); 511 const CeedInt interp_bytes = q_bytes * P_1d; 512 CeedBasis_Hip_shared *data; 513 514 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 515 CeedCallBackend(CeedCalloc(1, &data)); 516 517 // Copy basis data to GPU 518 if (q_weight_1d) { 519 CeedCallHip(ceed, hipMalloc((void **)&data->d_q_weight_1d, q_bytes)); 520 CeedCallHip(ceed, hipMemcpy(data->d_q_weight_1d, q_weight_1d, q_bytes, hipMemcpyHostToDevice)); 521 } 522 CeedCallHip(ceed, hipMalloc((void **)&data->d_interp_1d, interp_bytes)); 523 CeedCallHip(ceed, hipMemcpy(data->d_interp_1d, interp_1d, interp_bytes, hipMemcpyHostToDevice)); 524 CeedCallHip(ceed, hipMalloc((void **)&data->d_grad_1d, interp_bytes)); 525 CeedCallHip(ceed, hipMemcpy(data->d_grad_1d, grad_1d, interp_bytes, hipMemcpyHostToDevice)); 526 527 // Compute collocated gradient and copy to GPU 528 data->d_collo_grad_1d = NULL; 529 bool has_collocated_grad = dim == 3 && Q_1d >= P_1d; 530 531 if (has_collocated_grad) { 532 CeedScalar *collo_grad_1d; 533 534 CeedCallBackend(CeedMalloc(Q_1d * Q_1d, &collo_grad_1d)); 535 CeedCallBackend(CeedBasisGetCollocatedGrad(basis, collo_grad_1d)); 536 CeedCallHip(ceed, hipMalloc((void **)&data->d_collo_grad_1d, q_bytes * Q_1d)); 537 CeedCallHip(ceed, hipMemcpy(data->d_collo_grad_1d, collo_grad_1d, q_bytes * Q_1d, hipMemcpyHostToDevice)); 538 CeedCallBackend(CeedFree(&collo_grad_1d)); 539 } 540 541 // Set number of threads per block for basis kernels 542 CeedCallBackend(CeedBasisGetNumComponents(basis, &num_comp)); 543 CeedCallBackend(ComputeBasisThreadBlockSizes(dim, P_1d, Q_1d, num_comp, data->block_sizes)); 544 545 // Compile basis kernels 546 const char basis_kernel_source[] = "// Tensor basis source\n#include <ceed/jit-source/hip/hip-shared-basis-tensor.h>\n"; 547 548 CeedCallBackend(CeedCompile_Hip(ceed, basis_kernel_source, &data->module, 11, "BASIS_Q_1D", Q_1d, "BASIS_P_1D", P_1d, "T_1D", 549 CeedIntMax(Q_1d, P_1d), "BASIS_DIM", dim, "BASIS_NUM_COMP", num_comp, "BASIS_NUM_NODES", CeedIntPow(P_1d, dim), 550 "BASIS_NUM_QPTS", CeedIntPow(Q_1d, dim), "BASIS_INTERP_BLOCK_SIZE", data->block_sizes[0], "BASIS_GRAD_BLOCK_SIZE", 551 data->block_sizes[1], "BASIS_WEIGHT_BLOCK_SIZE", data->block_sizes[2], "BASIS_HAS_COLLOCATED_GRAD", 552 has_collocated_grad)); 553 CeedCallBackend(CeedGetKernel_Hip(ceed, data->module, "Interp", &data->Interp)); 554 CeedCallBackend(CeedGetKernel_Hip(ceed, data->module, "InterpTranspose", &data->InterpTranspose)); 555 CeedCallBackend(CeedGetKernel_Hip(ceed, data->module, "InterpTransposeAdd", &data->InterpTransposeAdd)); 556 CeedCallBackend(CeedGetKernel_Hip(ceed, data->module, "Grad", &data->Grad)); 557 CeedCallBackend(CeedGetKernel_Hip(ceed, data->module, "GradTranspose", &data->GradTranspose)); 558 CeedCallBackend(CeedGetKernel_Hip(ceed, data->module, "GradTransposeAdd", &data->GradTransposeAdd)); 559 CeedCallBackend(CeedGetKernel_Hip(ceed, data->module, "Weight", &data->Weight)); 560 561 CeedCallBackend(CeedBasisSetData(basis, data)); 562 563 // Register backend functions 564 CeedCallBackend(CeedSetBackendFunction(ceed, "Basis", basis, "Apply", CeedBasisApplyTensor_Hip_shared)); 565 CeedCallBackend(CeedSetBackendFunction(ceed, "Basis", basis, "ApplyAdd", CeedBasisApplyAddTensor_Hip_shared)); 566 CeedCallBackend(CeedSetBackendFunction(ceed, "Basis", basis, "ApplyAtPoints", CeedBasisApplyAtPoints_Hip_shared)); 567 CeedCallBackend(CeedSetBackendFunction(ceed, "Basis", basis, "ApplyAddAtPoints", CeedBasisApplyAddAtPoints_Hip_shared)); 568 CeedCallBackend(CeedSetBackendFunction(ceed, "Basis", basis, "Destroy", CeedBasisDestroy_Hip_shared)); 569 CeedCallBackend(CeedDestroy(&ceed)); 570 return CEED_ERROR_SUCCESS; 571 } 572 573 //------------------------------------------------------------------------------ 574