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-source/hip/hip-types.h> 11 #include <stddef.h> 12 13 #include "../hip/ceed-hip-common.h" 14 #include "../hip/ceed-hip-compile.h" 15 #include "ceed-hip-gen-operator-build.h" 16 #include "ceed-hip-gen.h" 17 18 //------------------------------------------------------------------------------ 19 // Destroy operator 20 //------------------------------------------------------------------------------ 21 static int CeedOperatorDestroy_Hip_gen(CeedOperator op) { 22 CeedOperator_Hip_gen *impl; 23 24 CeedCallBackend(CeedOperatorGetData(op, &impl)); 25 CeedCallBackend(CeedFree(&impl)); 26 return CEED_ERROR_SUCCESS; 27 } 28 29 //------------------------------------------------------------------------------ 30 // Apply and add to output 31 //------------------------------------------------------------------------------ 32 static int CeedOperatorApplyAdd_Hip_gen(CeedOperator op, CeedVector input_vec, CeedVector output_vec, CeedRequest *request) { 33 Ceed ceed; 34 CeedInt num_elem, num_input_fields, num_output_fields; 35 CeedEvalMode eval_mode; 36 CeedVector output_vecs[CEED_FIELD_MAX] = {NULL}; 37 CeedQFunctionField *qf_input_fields, *qf_output_fields; 38 CeedQFunction_Hip_gen *qf_data; 39 CeedQFunction qf; 40 CeedOperatorField *op_input_fields, *op_output_fields; 41 CeedOperator_Hip_gen *data; 42 43 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 44 CeedCallBackend(CeedOperatorGetData(op, &data)); 45 CeedCallBackend(CeedOperatorGetQFunction(op, &qf)); 46 CeedCallBackend(CeedQFunctionGetData(qf, &qf_data)); 47 CeedCallBackend(CeedOperatorGetNumElements(op, &num_elem)); 48 CeedCallBackend(CeedOperatorGetFields(op, &num_input_fields, &op_input_fields, &num_output_fields, &op_output_fields)); 49 CeedCallBackend(CeedQFunctionGetFields(qf, NULL, &qf_input_fields, NULL, &qf_output_fields)); 50 51 // Check for tensor-product bases 52 { 53 bool has_tensor_bases; 54 55 CeedCallBackend(CeedOperatorHasTensorBases(op, &has_tensor_bases)); 56 // -- Fallback to shared if not all bases are tensor-product 57 if (!has_tensor_bases) { 58 CeedOperator op_fallback; 59 60 CeedCallBackend(CeedOperatorGetFallback(op, &op_fallback)); 61 CeedCallBackend(CeedOperatorApplyAdd(op_fallback, input_vec, output_vec, request)); 62 return CEED_ERROR_SUCCESS; 63 } 64 } 65 66 // Creation of the operator 67 CeedCallBackend(CeedOperatorBuildKernel_Hip_gen(op)); 68 69 // Input vectors 70 for (CeedInt i = 0; i < num_input_fields; i++) { 71 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_input_fields[i], &eval_mode)); 72 if (eval_mode == CEED_EVAL_WEIGHT) { // Skip 73 data->fields.inputs[i] = NULL; 74 } else { 75 CeedVector vec; 76 77 // Get input vector 78 CeedCallBackend(CeedOperatorFieldGetVector(op_input_fields[i], &vec)); 79 if (vec == CEED_VECTOR_ACTIVE) vec = input_vec; 80 CeedCallBackend(CeedVectorGetArrayRead(vec, CEED_MEM_DEVICE, &data->fields.inputs[i])); 81 } 82 } 83 84 // Output vectors 85 for (CeedInt i = 0; i < num_output_fields; i++) { 86 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_output_fields[i], &eval_mode)); 87 if (eval_mode == CEED_EVAL_WEIGHT) { // Skip 88 data->fields.outputs[i] = NULL; 89 } else { 90 CeedVector vec; 91 92 // Get output vector 93 CeedCallBackend(CeedOperatorFieldGetVector(op_output_fields[i], &vec)); 94 if (vec == CEED_VECTOR_ACTIVE) vec = output_vec; 95 output_vecs[i] = vec; 96 // Check for multiple output modes 97 CeedInt index = -1; 98 for (CeedInt j = 0; j < i; j++) { 99 if (vec == output_vecs[j]) { 100 index = j; 101 break; 102 } 103 } 104 if (index == -1) { 105 CeedCallBackend(CeedVectorGetArray(vec, CEED_MEM_DEVICE, &data->fields.outputs[i])); 106 } else { 107 data->fields.outputs[i] = data->fields.outputs[index]; 108 } 109 } 110 } 111 112 // Get context data 113 CeedCallBackend(CeedQFunctionGetInnerContextData(qf, CEED_MEM_DEVICE, &qf_data->d_c)); 114 115 // Apply operator 116 void *opargs[] = {(void *)&num_elem, &qf_data->d_c, &data->indices, &data->fields, &data->B, &data->G, &data->W}; 117 const CeedInt dim = data->dim; 118 const CeedInt Q_1d = data->Q_1d; 119 const CeedInt P_1d = data->max_P_1d; 120 const CeedInt thread_1d = CeedIntMax(Q_1d, P_1d); 121 CeedInt block_sizes[3]; 122 123 CeedCallBackend(BlockGridCalculate_Hip_gen(dim, num_elem, P_1d, Q_1d, block_sizes)); 124 if (dim == 1) { 125 CeedInt grid = num_elem / block_sizes[2] + ((num_elem / block_sizes[2] * block_sizes[2] < num_elem) ? 1 : 0); 126 CeedInt sharedMem = block_sizes[2] * thread_1d * sizeof(CeedScalar); 127 128 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, data->op, grid, block_sizes[0], block_sizes[1], block_sizes[2], sharedMem, opargs)); 129 } else if (dim == 2) { 130 CeedInt grid = num_elem / block_sizes[2] + ((num_elem / block_sizes[2] * block_sizes[2] < num_elem) ? 1 : 0); 131 CeedInt sharedMem = block_sizes[2] * thread_1d * thread_1d * sizeof(CeedScalar); 132 133 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, data->op, grid, block_sizes[0], block_sizes[1], block_sizes[2], sharedMem, opargs)); 134 } else if (dim == 3) { 135 CeedInt grid = num_elem / block_sizes[2] + ((num_elem / block_sizes[2] * block_sizes[2] < num_elem) ? 1 : 0); 136 CeedInt sharedMem = block_sizes[2] * thread_1d * thread_1d * sizeof(CeedScalar); 137 138 CeedCallBackend(CeedRunKernelDimShared_Hip(ceed, data->op, grid, block_sizes[0], block_sizes[1], block_sizes[2], sharedMem, opargs)); 139 } 140 141 // Restore input arrays 142 for (CeedInt i = 0; i < num_input_fields; i++) { 143 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_input_fields[i], &eval_mode)); 144 if (eval_mode == CEED_EVAL_WEIGHT) { // Skip 145 } else { 146 CeedVector vec; 147 148 CeedCallBackend(CeedOperatorFieldGetVector(op_input_fields[i], &vec)); 149 if (vec == CEED_VECTOR_ACTIVE) vec = input_vec; 150 CeedCallBackend(CeedVectorRestoreArrayRead(vec, &data->fields.inputs[i])); 151 } 152 } 153 154 // Restore output arrays 155 for (CeedInt i = 0; i < num_output_fields; i++) { 156 CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_output_fields[i], &eval_mode)); 157 if (eval_mode == CEED_EVAL_WEIGHT) { // Skip 158 } else { 159 CeedVector vec; 160 161 CeedCallBackend(CeedOperatorFieldGetVector(op_output_fields[i], &vec)); 162 if (vec == CEED_VECTOR_ACTIVE) vec = output_vec; 163 // Check for multiple output modes 164 CeedInt index = -1; 165 for (CeedInt j = 0; j < i; j++) { 166 if (vec == output_vecs[j]) { 167 index = j; 168 break; 169 } 170 } 171 if (index == -1) { 172 CeedCallBackend(CeedVectorRestoreArray(vec, &data->fields.outputs[i])); 173 } 174 } 175 } 176 177 // Restore context data 178 CeedCallBackend(CeedQFunctionRestoreInnerContextData(qf, &qf_data->d_c)); 179 return CEED_ERROR_SUCCESS; 180 } 181 182 //------------------------------------------------------------------------------ 183 // Create operator 184 //------------------------------------------------------------------------------ 185 int CeedOperatorCreate_Hip_gen(CeedOperator op) { 186 Ceed ceed; 187 CeedOperator_Hip_gen *impl; 188 189 CeedCallBackend(CeedOperatorGetCeed(op, &ceed)); 190 CeedCallBackend(CeedCalloc(1, &impl)); 191 CeedCallBackend(CeedOperatorSetData(op, impl)); 192 CeedCallBackend(CeedSetBackendFunction(ceed, "Operator", op, "ApplyAdd", CeedOperatorApplyAdd_Hip_gen)); 193 CeedCallBackend(CeedSetBackendFunction(ceed, "Operator", op, "Destroy", CeedOperatorDestroy_Hip_gen)); 194 return CEED_ERROR_SUCCESS; 195 } 196 197 //------------------------------------------------------------------------------ 198