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 <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 return CEED_ERROR_SUCCESS; 85 } 86 87 //------------------------------------------------------------------------------ 88 // Apply basis 89 //------------------------------------------------------------------------------ 90 int CeedBasisApplyTensor_Hip_shared(CeedBasis basis, const CeedInt num_elem, CeedTransposeMode t_mode, CeedEvalMode eval_mode, CeedVector u, 91 CeedVector v) { 92 Ceed ceed; 93 Ceed_Hip *ceed_Hip; 94 CeedInt dim, num_comp; 95 const CeedScalar *d_u; 96 CeedScalar *d_v; 97 CeedBasis_Hip_shared *data; 98 99 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 100 CeedCallBackend(CeedGetData(ceed, &ceed_Hip)); 101 CeedCallBackend(CeedBasisGetData(basis, &data)); 102 CeedCallBackend(CeedBasisGetDimension(basis, &dim)); 103 CeedCallBackend(CeedBasisGetNumComponents(basis, &num_comp)); 104 105 // Get read/write access to u, v 106 if (u != CEED_VECTOR_NONE) CeedCallBackend(CeedVectorGetArrayRead(u, CEED_MEM_DEVICE, &d_u)); 107 else CeedCheck(eval_mode == CEED_EVAL_WEIGHT, ceed, CEED_ERROR_BACKEND, "An input vector is required for this CeedEvalMode"); 108 CeedCallBackend(CeedVectorGetArrayWrite(v, CEED_MEM_DEVICE, &d_v)); 109 110 // Apply basis operation 111 switch (eval_mode) { 112 case CEED_EVAL_INTERP: { 113 CeedInt P_1d, Q_1d; 114 CeedInt block_size = data->block_sizes[0]; 115 116 CeedCallBackend(CeedBasisGetNumNodes1D(basis, &P_1d)); 117 CeedCallBackend(CeedBasisGetNumQuadraturePoints1D(basis, &Q_1d)); 118 CeedInt thread_1d = CeedIntMax(Q_1d, P_1d); 119 void *interp_args[] = {(void *)&num_elem, &data->d_interp_1d, &d_u, &d_v}; 120 121 if (dim == 1) { 122 CeedInt elems_per_block = 64 * thread_1d > 256 ? 256 / thread_1d : 64; 123 elems_per_block = elems_per_block > 0 ? elems_per_block : 1; 124 CeedInt grid = num_elem / elems_per_block + ((num_elem / elems_per_block * elems_per_block < num_elem) ? 1 : 0); 125 CeedInt shared_mem = elems_per_block * thread_1d * sizeof(CeedScalar); 126 127 if (t_mode == CEED_TRANSPOSE) { 128 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, data->InterpTranspose, grid, thread_1d, 1, elems_per_block, shared_mem, interp_args)); 129 } else { 130 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, data->Interp, grid, thread_1d, 1, elems_per_block, shared_mem, interp_args)); 131 } 132 } else if (dim == 2) { 133 // Check if required threads is small enough to do multiple elems 134 const CeedInt elems_per_block = CeedIntMax(block_size / (thread_1d * thread_1d), 1); 135 CeedInt grid = num_elem / elems_per_block + ((num_elem / elems_per_block * elems_per_block < num_elem) ? 1 : 0); 136 CeedInt shared_mem = elems_per_block * thread_1d * thread_1d * sizeof(CeedScalar); 137 138 if (t_mode == CEED_TRANSPOSE) { 139 CeedCallBackend( 140 CeedRunKernelDimShared_Hip(ceed, data->InterpTranspose, grid, thread_1d, thread_1d, elems_per_block, shared_mem, interp_args)); 141 } else { 142 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, data->Interp, grid, thread_1d, thread_1d, elems_per_block, shared_mem, interp_args)); 143 } 144 } else if (dim == 3) { 145 const CeedInt elems_per_block = CeedIntMax(block_size / (thread_1d * thread_1d), 1); 146 CeedInt grid = num_elem / elems_per_block + ((num_elem / elems_per_block * elems_per_block < num_elem) ? 1 : 0); 147 CeedInt shared_mem = elems_per_block * thread_1d * thread_1d * sizeof(CeedScalar); 148 149 if (t_mode == CEED_TRANSPOSE) { 150 CeedCallBackend( 151 CeedRunKernelDimShared_Hip(ceed, data->InterpTranspose, grid, thread_1d, thread_1d, elems_per_block, shared_mem, interp_args)); 152 } else { 153 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, data->Interp, grid, thread_1d, thread_1d, elems_per_block, shared_mem, interp_args)); 154 } 155 } 156 } break; 157 case CEED_EVAL_GRAD: { 158 CeedInt P_1d, Q_1d; 159 CeedInt block_size = data->block_sizes[1]; 160 161 CeedCallBackend(CeedBasisGetNumNodes1D(basis, &P_1d)); 162 CeedCallBackend(CeedBasisGetNumQuadraturePoints1D(basis, &Q_1d)); 163 CeedInt thread_1d = CeedIntMax(Q_1d, P_1d); 164 CeedScalar *d_grad_1d = data->d_grad_1d; 165 166 if (data->d_collo_grad_1d) { 167 d_grad_1d = data->d_collo_grad_1d; 168 } 169 void *grad_args[] = {(void *)&num_elem, &data->d_interp_1d, &d_grad_1d, &d_u, &d_v}; 170 if (dim == 1) { 171 CeedInt elems_per_block = 64 * thread_1d > 256 ? 256 / thread_1d : 64; 172 elems_per_block = elems_per_block > 0 ? elems_per_block : 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 * sizeof(CeedScalar); 175 176 if (t_mode == CEED_TRANSPOSE) { 177 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, data->GradTranspose, grid, thread_1d, 1, elems_per_block, shared_mem, grad_args)); 178 } else { 179 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, data->Grad, grid, thread_1d, 1, elems_per_block, shared_mem, grad_args)); 180 } 181 } else if (dim == 2) { 182 // Check if required threads is small enough to do multiple elems 183 const CeedInt elems_per_block = CeedIntMax(block_size / (thread_1d * thread_1d), 1); 184 CeedInt grid = num_elem / elems_per_block + ((num_elem / elems_per_block * elems_per_block < num_elem) ? 1 : 0); 185 CeedInt shared_mem = elems_per_block * thread_1d * thread_1d * sizeof(CeedScalar); 186 187 if (t_mode == CEED_TRANSPOSE) { 188 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, data->GradTranspose, grid, thread_1d, thread_1d, elems_per_block, shared_mem, grad_args)); 189 } else { 190 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, data->Grad, grid, thread_1d, thread_1d, elems_per_block, shared_mem, grad_args)); 191 } 192 } else if (dim == 3) { 193 const CeedInt elems_per_block = CeedIntMax(block_size / (thread_1d * thread_1d), 1); 194 CeedInt grid = num_elem / elems_per_block + ((num_elem / elems_per_block * elems_per_block < num_elem) ? 1 : 0); 195 CeedInt shared_mem = elems_per_block * thread_1d * thread_1d * sizeof(CeedScalar); 196 197 if (t_mode == CEED_TRANSPOSE) { 198 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, data->GradTranspose, grid, thread_1d, thread_1d, elems_per_block, shared_mem, grad_args)); 199 } else { 200 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, data->Grad, grid, thread_1d, thread_1d, elems_per_block, shared_mem, grad_args)); 201 } 202 } 203 } break; 204 case CEED_EVAL_WEIGHT: { 205 CeedInt Q_1d; 206 CeedInt block_size = data->block_sizes[2]; 207 208 CeedCheck(data->d_q_weight_1d, ceed, CEED_ERROR_BACKEND, "%s not supported; q_weights_1d not set", CeedEvalModes[eval_mode]); 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 case CEED_EVAL_NONE: /* handled separately below */ 233 break; 234 // LCOV_EXCL_START 235 case CEED_EVAL_DIV: 236 case CEED_EVAL_CURL: 237 return CeedError(ceed, CEED_ERROR_BACKEND, "%s not supported", CeedEvalModes[eval_mode]); 238 // LCOV_EXCL_STOP 239 } 240 241 // Restore vectors, cover CEED_EVAL_NONE 242 CeedCallBackend(CeedVectorRestoreArray(v, &d_v)); 243 if (eval_mode == CEED_EVAL_NONE) CeedCallBackend(CeedVectorSetArray(v, CEED_MEM_DEVICE, CEED_COPY_VALUES, (CeedScalar *)d_u)); 244 if (eval_mode != CEED_EVAL_WEIGHT) CeedCallBackend(CeedVectorRestoreArrayRead(u, &d_u)); 245 return CEED_ERROR_SUCCESS; 246 } 247 248 //------------------------------------------------------------------------------ 249 // Basis apply - tensor AtPoints 250 //------------------------------------------------------------------------------ 251 int CeedBasisApplyAtPoints_Hip_shared(CeedBasis basis, const CeedInt num_elem, const CeedInt *num_points, CeedTransposeMode t_mode, 252 CeedEvalMode eval_mode, CeedVector x_ref, CeedVector u, CeedVector v) { 253 Ceed ceed; 254 CeedInt Q_1d, dim, max_num_points = num_points[0]; 255 const CeedInt is_transpose = t_mode == CEED_TRANSPOSE; 256 const int max_block_size = 32; 257 const CeedScalar *d_x, *d_u; 258 CeedScalar *d_v; 259 CeedBasis_Hip_shared *data; 260 261 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 262 CeedCallBackend(CeedBasisGetData(basis, &data)); 263 CeedCallBackend(CeedBasisGetNumQuadraturePoints1D(basis, &Q_1d)); 264 CeedCallBackend(CeedBasisGetDimension(basis, &dim)); 265 266 // Check uniform number of points per elem 267 for (CeedInt i = 1; i < num_elem; i++) { 268 CeedCheck(max_num_points == num_points[i], ceed, CEED_ERROR_BACKEND, 269 "BasisApplyAtPoints only supported for the same number of points in each element"); 270 } 271 272 // Weight handled separately 273 if (eval_mode == CEED_EVAL_WEIGHT) { 274 CeedCall(CeedVectorSetValue(v, 1.0)); 275 return CEED_ERROR_SUCCESS; 276 } 277 278 // Build kernels if needed 279 if (data->num_points != max_num_points) { 280 CeedInt P_1d; 281 282 CeedCallBackend(CeedBasisGetNumNodes1D(basis, &P_1d)); 283 data->num_points = max_num_points; 284 285 // -- Create interp matrix to Chebyshev coefficients 286 if (!data->d_chebyshev_interp_1d) { 287 CeedSize interp_bytes; 288 CeedScalar *chebyshev_interp_1d; 289 290 interp_bytes = P_1d * Q_1d * sizeof(CeedScalar); 291 CeedCallBackend(CeedCalloc(P_1d * Q_1d, &chebyshev_interp_1d)); 292 CeedCall(CeedBasisGetChebyshevInterp1D(basis, chebyshev_interp_1d)); 293 CeedCallHip(ceed, hipMalloc((void **)&data->d_chebyshev_interp_1d, interp_bytes)); 294 CeedCallHip(ceed, hipMemcpy(data->d_chebyshev_interp_1d, chebyshev_interp_1d, interp_bytes, hipMemcpyHostToDevice)); 295 CeedCallBackend(CeedFree(&chebyshev_interp_1d)); 296 } 297 298 // -- Compile kernels 299 char *basis_kernel_source; 300 const char *basis_kernel_path; 301 CeedInt num_comp; 302 303 if (data->moduleAtPoints) CeedCallHip(ceed, hipModuleUnload(data->moduleAtPoints)); 304 CeedCallBackend(CeedBasisGetNumComponents(basis, &num_comp)); 305 CeedCallBackend(CeedGetJitAbsolutePath(ceed, "ceed/jit-source/hip/hip-ref-basis-tensor-at-points.h", &basis_kernel_path)); 306 CeedDebug256(ceed, CEED_DEBUG_COLOR_SUCCESS, "----- Loading Basis Kernel Source -----\n"); 307 CeedCallBackend(CeedLoadSourceToBuffer(ceed, basis_kernel_path, &basis_kernel_source)); 308 CeedDebug256(ceed, CEED_DEBUG_COLOR_SUCCESS, "----- Loading Basis Kernel Source Complete! -----\n"); 309 CeedCallBackend(CeedCompile_Hip(ceed, basis_kernel_source, &data->moduleAtPoints, 9, "BASIS_Q_1D", Q_1d, "BASIS_P_1D", P_1d, "BASIS_BUF_LEN", 310 Q_1d * CeedIntPow(Q_1d > P_1d ? Q_1d : P_1d, dim - 1), "BASIS_DIM", dim, "BASIS_NUM_COMP", num_comp, 311 "BASIS_NUM_NODES", CeedIntPow(P_1d, dim), "BASIS_NUM_QPTS", CeedIntPow(Q_1d, dim), "BASIS_NUM_PTS", 312 max_num_points, "POINTS_BUFF_LEN", CeedIntPow(Q_1d, dim - 1))); 313 CeedCallBackend(CeedGetKernel_Hip(ceed, data->moduleAtPoints, "InterpAtPoints", &data->InterpAtPoints)); 314 CeedCallBackend(CeedGetKernel_Hip(ceed, data->moduleAtPoints, "GradAtPoints", &data->GradAtPoints)); 315 CeedCallBackend(CeedFree(&basis_kernel_path)); 316 CeedCallBackend(CeedFree(&basis_kernel_source)); 317 } 318 319 // Get read/write access to u, v 320 CeedCallBackend(CeedVectorGetArrayRead(x_ref, CEED_MEM_DEVICE, &d_x)); 321 if (u != CEED_VECTOR_NONE) CeedCallBackend(CeedVectorGetArrayRead(u, CEED_MEM_DEVICE, &d_u)); 322 else CeedCheck(eval_mode == CEED_EVAL_WEIGHT, ceed, CEED_ERROR_BACKEND, "An input vector is required for this CeedEvalMode"); 323 CeedCallBackend(CeedVectorGetArrayWrite(v, CEED_MEM_DEVICE, &d_v)); 324 325 // Clear v for transpose operation 326 if (is_transpose) { 327 CeedSize length; 328 329 CeedCallBackend(CeedVectorGetLength(v, &length)); 330 CeedCallHip(ceed, hipMemset(d_v, 0, length * sizeof(CeedScalar))); 331 } 332 333 // Basis action 334 switch (eval_mode) { 335 case CEED_EVAL_INTERP: { 336 void *interp_args[] = {(void *)&num_elem, (void *)&is_transpose, &data->d_chebyshev_interp_1d, &d_x, &d_u, &d_v}; 337 const CeedInt block_size = CeedIntMin(CeedIntPow(Q_1d, dim), max_block_size); 338 339 CeedCallBackend(CeedRunKernel_Hip(ceed, data->InterpAtPoints, num_elem, block_size, interp_args)); 340 } break; 341 case CEED_EVAL_GRAD: { 342 void *grad_args[] = {(void *)&num_elem, (void *)&is_transpose, &data->d_chebyshev_interp_1d, &d_x, &d_u, &d_v}; 343 const CeedInt block_size = CeedIntMin(CeedIntPow(Q_1d, dim), max_block_size); 344 345 CeedCallBackend(CeedRunKernel_Hip(ceed, data->GradAtPoints, num_elem, block_size, grad_args)); 346 } break; 347 case CEED_EVAL_WEIGHT: 348 case CEED_EVAL_NONE: /* handled separately below */ 349 break; 350 // LCOV_EXCL_START 351 case CEED_EVAL_DIV: 352 case CEED_EVAL_CURL: 353 return CeedError(ceed, CEED_ERROR_BACKEND, "%s not supported", CeedEvalModes[eval_mode]); 354 // LCOV_EXCL_STOP 355 } 356 357 // Restore vectors, cover CEED_EVAL_NONE 358 CeedCallBackend(CeedVectorRestoreArrayRead(x_ref, &d_x)); 359 CeedCallBackend(CeedVectorRestoreArray(v, &d_v)); 360 if (eval_mode == CEED_EVAL_NONE) CeedCallBackend(CeedVectorSetArray(v, CEED_MEM_DEVICE, CEED_COPY_VALUES, (CeedScalar *)d_u)); 361 if (eval_mode != CEED_EVAL_WEIGHT) CeedCallBackend(CeedVectorRestoreArrayRead(u, &d_u)); 362 return CEED_ERROR_SUCCESS; 363 } 364 365 //------------------------------------------------------------------------------ 366 // Destroy basis 367 //------------------------------------------------------------------------------ 368 static int CeedBasisDestroy_Hip_shared(CeedBasis basis) { 369 Ceed ceed; 370 CeedBasis_Hip_shared *data; 371 372 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 373 CeedCallBackend(CeedBasisGetData(basis, &data)); 374 CeedCallHip(ceed, hipModuleUnload(data->module)); 375 if (data->moduleAtPoints) CeedCallHip(ceed, hipModuleUnload(data->moduleAtPoints)); 376 if (data->d_q_weight_1d) CeedCallHip(ceed, hipFree(data->d_q_weight_1d)); 377 CeedCallHip(ceed, hipFree(data->d_interp_1d)); 378 CeedCallHip(ceed, hipFree(data->d_grad_1d)); 379 CeedCallHip(ceed, hipFree(data->d_collo_grad_1d)); 380 CeedCallHip(ceed, hipFree(data->d_chebyshev_interp_1d)); 381 CeedCallBackend(CeedFree(&data)); 382 return CEED_ERROR_SUCCESS; 383 } 384 385 //------------------------------------------------------------------------------ 386 // Create tensor basis 387 //------------------------------------------------------------------------------ 388 int CeedBasisCreateTensorH1_Hip_shared(CeedInt dim, CeedInt P_1d, CeedInt Q_1d, const CeedScalar *interp_1d, const CeedScalar *grad_1d, 389 const CeedScalar *q_ref_1d, const CeedScalar *q_weight_1d, CeedBasis basis) { 390 Ceed ceed; 391 char *basis_kernel_source; 392 const char *basis_kernel_path; 393 CeedInt num_comp; 394 const CeedInt q_bytes = Q_1d * sizeof(CeedScalar); 395 const CeedInt interp_bytes = q_bytes * P_1d; 396 CeedBasis_Hip_shared *data; 397 398 CeedCallBackend(CeedBasisGetCeed(basis, &ceed)); 399 CeedCallBackend(CeedCalloc(1, &data)); 400 401 // Copy basis data to GPU 402 if (q_weight_1d) { 403 CeedCallHip(ceed, hipMalloc((void **)&data->d_q_weight_1d, q_bytes)); 404 CeedCallHip(ceed, hipMemcpy(data->d_q_weight_1d, q_weight_1d, q_bytes, hipMemcpyHostToDevice)); 405 } 406 CeedCallHip(ceed, hipMalloc((void **)&data->d_interp_1d, interp_bytes)); 407 CeedCallHip(ceed, hipMemcpy(data->d_interp_1d, interp_1d, interp_bytes, hipMemcpyHostToDevice)); 408 CeedCallHip(ceed, hipMalloc((void **)&data->d_grad_1d, interp_bytes)); 409 CeedCallHip(ceed, hipMemcpy(data->d_grad_1d, grad_1d, interp_bytes, hipMemcpyHostToDevice)); 410 411 // Compute collocated gradient and copy to GPU 412 data->d_collo_grad_1d = NULL; 413 bool has_collocated_grad = dim == 3 && Q_1d >= P_1d; 414 415 if (has_collocated_grad) { 416 CeedScalar *collo_grad_1d; 417 418 CeedCallBackend(CeedMalloc(Q_1d * Q_1d, &collo_grad_1d)); 419 CeedCallBackend(CeedBasisGetCollocatedGrad(basis, collo_grad_1d)); 420 CeedCallHip(ceed, hipMalloc((void **)&data->d_collo_grad_1d, q_bytes * Q_1d)); 421 CeedCallHip(ceed, hipMemcpy(data->d_collo_grad_1d, collo_grad_1d, q_bytes * Q_1d, hipMemcpyHostToDevice)); 422 CeedCallBackend(CeedFree(&collo_grad_1d)); 423 } 424 425 // Set number of threads per block for basis kernels 426 CeedCallBackend(CeedBasisGetNumComponents(basis, &num_comp)); 427 CeedCallBackend(ComputeBasisThreadBlockSizes(dim, P_1d, Q_1d, num_comp, data->block_sizes)); 428 429 // Compile basis kernels 430 CeedCallBackend(CeedGetJitAbsolutePath(ceed, "ceed/jit-source/hip/hip-shared-basis-tensor.h", &basis_kernel_path)); 431 CeedDebug256(ceed, CEED_DEBUG_COLOR_SUCCESS, "----- Loading Basis Kernel Source -----\n"); 432 CeedCallBackend(CeedLoadSourceToBuffer(ceed, basis_kernel_path, &basis_kernel_source)); 433 CeedDebug256(ceed, CEED_DEBUG_COLOR_SUCCESS, "----- Loading Basis Kernel Source Complete! -----\n"); 434 CeedCallBackend(CeedCompile_Hip(ceed, basis_kernel_source, &data->module, 11, "BASIS_Q_1D", Q_1d, "BASIS_P_1D", P_1d, "T_1D", 435 CeedIntMax(Q_1d, P_1d), "BASIS_DIM", dim, "BASIS_NUM_COMP", num_comp, "BASIS_NUM_NODES", CeedIntPow(P_1d, dim), 436 "BASIS_NUM_QPTS", CeedIntPow(Q_1d, dim), "BASIS_INTERP_BLOCK_SIZE", data->block_sizes[0], "BASIS_GRAD_BLOCK_SIZE", 437 data->block_sizes[1], "BASIS_WEIGHT_BLOCK_SIZE", data->block_sizes[2], "BASIS_HAS_COLLOCATED_GRAD", 438 has_collocated_grad)); 439 CeedCallBackend(CeedGetKernel_Hip(ceed, data->module, "Interp", &data->Interp)); 440 CeedCallBackend(CeedGetKernel_Hip(ceed, data->module, "InterpTranspose", &data->InterpTranspose)); 441 CeedCallBackend(CeedGetKernel_Hip(ceed, data->module, "Grad", &data->Grad)); 442 CeedCallBackend(CeedGetKernel_Hip(ceed, data->module, "GradTranspose", &data->GradTranspose)); 443 CeedCallBackend(CeedGetKernel_Hip(ceed, data->module, "Weight", &data->Weight)); 444 CeedCallBackend(CeedFree(&basis_kernel_path)); 445 CeedCallBackend(CeedFree(&basis_kernel_source)); 446 447 CeedCallBackend(CeedBasisSetData(basis, data)); 448 449 // Register backend functions 450 CeedCallBackend(CeedSetBackendFunction(ceed, "Basis", basis, "Apply", CeedBasisApplyTensor_Hip_shared)); 451 CeedCallBackend(CeedSetBackendFunction(ceed, "Basis", basis, "ApplyAtPoints", CeedBasisApplyAtPoints_Hip_shared)); 452 CeedCallBackend(CeedSetBackendFunction(ceed, "Basis", basis, "Destroy", CeedBasisDestroy_Hip_shared)); 453 return CEED_ERROR_SUCCESS; 454 } 455 456 //------------------------------------------------------------------------------ 457