// Copyright Advanced Micro Devices, Inc., or its affiliates. // SPDX-License-Identifier: MIT #pragma once #include #include #include #include #include #include namespace rocRoller { template concept CInvokableTo = std::invocable && std::convertible_to>; template Generator generateNodes(Scheduling::SchedulerPtr scheduler, std::set nodes, std::set& completedNodes, CInvokableTo, Node> auto generateNode, CInvokableTo auto nodeIsReady, CInvokableTo auto nodeCategory, CInvokableTo auto categoryLimit, CInvokableTo auto comparePriorities) { std::set allCategories; for(auto node : nodes) allCategories.insert(nodeCategory(node)); std::map> generating; using PQ = std::priority_queue, decltype(comparePriorities)>; std::map ready; for(auto category : allCategories) ready.emplace(category, PQ{comparePriorities}); auto fillReady = [&]() { for(auto iter = nodes.begin(); iter != nodes.end();) { auto idx = *iter; if(nodeIsReady(idx)) { ready.at(nodeCategory(idx)).push(idx); iter = nodes.erase(iter); } else { ++iter; } } }; auto generate = [&](Node idx) -> Generator { auto category = nodeCategory(idx); co_yield generateNode(idx); completedNodes.insert(idx); generating[category].erase(idx); }; auto anyRemainingNodes = [&]() -> bool { if(!nodes.empty()) return true; for(auto const& [category, readyQueue] : ready) { if(!readyQueue.empty()) return true; } for(auto const& [category, generatingSet] : generating) { AssertFatal(generatingSet.empty()); } return false; }; while(anyRemainingNodes()) { std::vector> schedulable; auto fillSchedulable = [&]() { for(auto& [category, readyQueue] : ready) { auto limit = categoryLimit(category); AssertFatal(limit > 0); while(!readyQueue.empty() && generating[category].size() < limit) { auto idx = readyQueue.top(); readyQueue.pop(); generating[category].insert(idx); schedulable.push_back(generate(idx)); } } }; fillReady(); fillSchedulable(); if(!scheduler->supportsAddingStreams()) { co_yield (*scheduler)(schedulable); } else { auto generator = (*scheduler)(schedulable); auto prevDoneSize = completedNodes.size(); for(auto iter = generator.begin(); iter != generator.end(); ++iter) { co_yield *iter; if(prevDoneSize != completedNodes.size()) { fillReady(); prevDoneSize = completedNodes.size(); } fillSchedulable(); } } } } }