-
Grégoire Kubler authoredGrégoire Kubler authored
Code owners
Assign users and groups as approvers for specific file changes. Learn more.
Scheduler.cpp 27.06 KiB
/********************************************************************************
* Copyright (c) 2023 CEA-List
*
* This program and the accompanying materials are made available under the
* terms of the Eclipse Public License 2.0 which is available at
* http://www.eclipse.org/legal/epl-2.0.
*
* SPDX-License-Identifier: EPL-2.0
*
********************************************************************************/
#include "aidge/scheduler/Scheduler.hpp"
#include <chrono>
#include <memory>
#include <set>
#include <string>
#include <fmt/ranges.h>
#include <fmt/color.h>
#include "aidge/graph/GraphView.hpp"
#include "aidge/graph/Node.hpp"
#include "aidge/utils/Types.h"
#include "aidge/operator/OperatorTensor.hpp"
#include "aidge/operator/Producer.hpp"
#include "aidge/operator/Memorize.hpp"
#include "aidge/operator/MetaOperator.hpp"
void drawProgressBar(double progress, int barWidth, const std::string& additionalInfo = "") {
putchar('[');
int pos = static_cast<int>(barWidth * progress);
for (int i = 0; i < barWidth; ++i) {
if (i <= pos)
putchar('#');
else
putchar(' ');
}
fmt::print("] {}% | {}\r", static_cast<int>(progress * 100), additionalInfo);
fflush(stdout);
}
void Aidge::SequentialScheduler::generateScheduling(bool verbose) {
// TODO: For loop on the list of node to run
// run sequencially every runnable consumers once
// TODO: handle memory allocation in scheduler
// TODO: optimize memory usage
// 1) Setup initial consumers list:
// It is the list of input nodes
std::set<std::shared_ptr<Node>> consumers = mGraphView->inputNodes();
// Plus the list of nodes inside the graph connected to an inner producer
std::set<std::shared_ptr<Node>> producers;
for (const std::shared_ptr<Node>& nodePtr : mGraphView->getNodes()) {
if (nodePtr->type() == Producer_Op::Type) {
producers.insert(nodePtr);
}
}
const auto producersConsumers = getConsumers(producers);
consumers.insert(producersConsumers.begin(), producersConsumers.end());
std::map<std::shared_ptr<Node>, std::string> namePtrTable;
if (verbose) namePtrTable = mGraphView->getRankedNodesName("{0} ({1}#{3})");
// Still consumers are consumers that were run by can still consume data.
// They must be run AFTER the remaining consumer to ensure a non-greedy
// producers-consumers model!
std::set<std::shared_ptr<Node>> stillConsumers;
mStaticSchedule.push_back(std::vector<std::shared_ptr<Node>>());
do {
// 2) From the current consumers list, check if any prior consumer node
// is needed. A prior will generally be required for any node consuming
// parameters (weights and bias) that is not an input node.
// If for a given node, only parent producers (at any depth) are needed
// to satisfy its required data, it becomes a prior.
// If the prior node is a producer, it is added to the list of required
// producers.
// If the prior node is of another type, it replaces the initial consumer
// in the new priorConsumers list. The initial consumer will become
// again a consumer later, by construction.
if (verbose) fmt::print("List of consumers with their priors:\n");
std::set<std::shared_ptr<Node>> requiredProducers;
std::set<std::shared_ptr<Node>> priorConsumers;
mPriorCache.clear();
for (const auto& consumer : consumers) {
if (verbose) {
fmt::print("\t- consumer: ");
fmt::print(fg(fmt::color::orange), namePtrTable[consumer]);
fmt::print("\n");
}
const auto& prior = getPriorProducersConsumers(consumer);
if (prior.isPrior) {
if (verbose) {
std::vector<std::string> requiredProducersName;
std::transform(prior.requiredProducers.begin(), prior.requiredProducers.end(),
std::back_inserter(requiredProducersName),
[&namePtrTable](auto val){ return namePtrTable[val]; });
fmt::print("\t\trequired producers: {}\n", requiredProducersName);
std::vector<std::string> priorConsumersName;
std::transform(prior.priorConsumers.begin(), prior.priorConsumers.end(),
std::back_inserter(priorConsumersName),
[&namePtrTable](auto val){ return namePtrTable[val]; });
fmt::print("\t\tprior consumers: {}\n", priorConsumersName);
}
requiredProducers.insert(prior.requiredProducers.cbegin(), prior.requiredProducers.cend());
priorConsumers.insert(prior.priorConsumers.cbegin(), prior.priorConsumers.cend());
}
else {
priorConsumers.insert(consumer);
}
}
// 3) Prior consumers replace the initial consumers list.
// By construction, initial consumers will necessarily become consumers
// again later.
consumers.swap(priorConsumers);
// 4) Make producers generate the required data.
// Producers are special nodes that generate data on demand.
for (const auto& requiredProducer : requiredProducers) {
requiredProducer->getOperator()->updateConsummerProducer();
mStaticSchedule.back().push_back(requiredProducer);
}
// 5) Find runnable consumers.
// A consumer is runnable if the required data is available for all of
// its inputs. At this point, not all consumers are necessarily
// runnable because some may depend on the execution of others (when
// there is multiple successive priors for example).
std::set<std::shared_ptr<Node>> runnableConsumers;
if (verbose) fmt::print("Updated list of consumers:\n");
for (const auto& consumer : consumers) {
if (verbose) { fmt::print("\t- consumer: ");
fmt::print(fg(fmt::color::orange), namePtrTable[consumer]);
fmt::print("\n\t\tC/R:\t");
for (IOIndex_t inId = 0; inId < consumer->nbInputs() - 1; ++inId) {
fmt::print("{}/{}\n\t\t\t", consumer->getOperator()->getNbConsumedData(inId),
consumer->getOperator()->getNbRequiredData(inId));
}
fmt::print("{}/{}", consumer->getOperator()->getNbConsumedData(static_cast<IOIndex_t>(consumer->nbInputs()) - 1),
consumer->getOperator()->getNbRequiredData(static_cast<IOIndex_t>(consumer->nbInputs()) - 1));
fmt::print("\n\t\tP:\t");
for (IOIndex_t outId = 0; outId < consumer->nbOutputs() - 1; ++outId) {
fmt::print("{}\n\t\t\t", consumer->getOperator()->getNbProducedData(outId));
}
fmt::print("{}", consumer->getOperator()->getNbProducedData(static_cast<IOIndex_t>(consumer->nbOutputs()) - 1));
fmt::print("\n");
}
bool isRunnable = true;
for (IOIndex_t inputIdx = 0; inputIdx < consumer->nbInputs(); ++inputIdx) {
if (/*consumer->getOperator()->getNbRequiredData(inputIdx) > 0
&& */(consumer->getOperator()->getNbConsumedData(inputIdx) + consumer->getOperator()->getNbRequiredData(inputIdx)) >
getNbAvailableData(consumer, inputIdx)) {
if (verbose) fmt::print(" not runnable: C{} + R{} > P{} for input #{}\n",
consumer->getOperator()->getNbConsumedData(inputIdx),
consumer->getOperator()->getNbRequiredData(inputIdx),
getNbAvailableData(consumer, inputIdx), inputIdx);
// not enough data to run
isRunnable = false;
break;
}
}
if (isRunnable) {
runnableConsumers.insert(consumer);
}
}
// 5) If not consumer is runnable, it is a stop condition!
if (runnableConsumers.empty()) {
if (verbose) fmt::print("********************\n");
// No consumer is runnable: some required data is missing for all of
// them. There is two possibilities:
// - At least one required data source is exhausted, which may be
// an expected stop condition.
// - There is a deadlock between consumers, if some one is waiting
// for data from the other and reciprocally.
break;
}
// 6) Push runnable consumers in the list of nodes to run and update the
// consumer producer system.
// At this point, simultaneously runnable consumers have no data
// dependency and could be run in parallel!
for (const auto& runnable : runnableConsumers) {
if (verbose) fmt::print("Runnable: {}\n", namePtrTable[runnable]);
runnable->getOperator()->updateConsummerProducer();
mStaticSchedule.back().push_back(runnable);
}
// 7) Update consumers list
if (verbose) fmt::print("Updating producer and consumer lists...\n");
for (const auto& consumer : runnableConsumers) {
if (verbose) {
fmt::print("\t- consumer: {}\n\t\tC/R:\t",
namePtrTable[consumer]);
for (IOIndex_t inId = 0; inId < consumer->nbInputs() - 1; ++inId) {
fmt::print("{}/{}\n\t\t\t", consumer->getOperator()->getNbConsumedData(inId),
consumer->getOperator()->getNbRequiredData(inId));
} fmt::print("{}/{}", consumer->getOperator()->getNbConsumedData(static_cast<IOIndex_t>(consumer->nbInputs()) - 1),
consumer->getOperator()->getNbRequiredData(static_cast<IOIndex_t>(consumer->nbInputs()) - 1));
fmt::print("\n\t\tP:\t");
for (IOIndex_t outId = 0; outId < consumer->nbOutputs() - 1; ++outId) {
fmt::print("{}\n\t\t\t", consumer->getOperator()->getNbProducedData(outId));
}
fmt::print("{}", consumer->getOperator()->getNbProducedData(static_cast<IOIndex_t>(consumer->nbOutputs()) - 1));
fmt::print("\n");
}
// 7.1) If the current consumer has still data to consume, it will
// be put back in the consumers list once the remaining consumers
// have been exhausted.
bool isStillConsumer = false;
for (IOIndex_t inputIdx = 0; inputIdx < consumer->nbInputs(); ++inputIdx) {
if (consumer->getOperator()->getNbConsumedData(inputIdx) <
getNbAvailableData(consumer, inputIdx)) {
if (verbose) fmt::print(" still consumer: C{} < P{} for input #{}\n",
consumer->getOperator()->getNbConsumedData(inputIdx),
getNbAvailableData(consumer, inputIdx), inputIdx);
// there is still data to consume
isStillConsumer = true;
break;
}
}
// 7.2) If the current consumer becomes a producer for other nodes,
// its childs become consumers.
bool isProducer = false;
for (IOIndex_t outId = 0; outId < consumer->nbOutputs(); ++outId) {
for (const auto& child : consumer->getChildren(outId)) {
if (child) {
IOIndex_t inputIdx = 0;
for (const auto& childParent : child->getParents()) {
if (childParent == consumer) {
if (consumer->getOperator()->getNbProducedData(outId) > child->getOperator()->getNbConsumedData(inputIdx)) {
isProducer = true;
}
}
++inputIdx;
}
}
}
/*
if (consumer->getOperator()->getNbProducedData(outId) > 0) {
if (verbose) fmt::print(" also producer\n");
// make sure consumer is also a producer
producers.insert(consumer);
const auto& newConsumers = getConsumers({consumer});
consumers.insert(newConsumers.cbegin(), newConsumers.cend());
break;
}
*/
}
consumers.erase(consumer);
if (isProducer) {
if (verbose) fmt::print(" also producer\n");
// make sure consumer is also a producer
producers.insert(consumer);
const auto& newConsumers = getConsumers({consumer});
consumers.insert(newConsumers.cbegin(), newConsumers.cend());
}
if (isStillConsumer) {
// If there is still data to consume, the consumer will be // run AFTER the other remaining consumers
// (= non-greedy consumers)
stillConsumers.insert(consumer);
}
}
// 8) If there is no more consumers, swap with possible "still consumers"
// This ensures that the "non-greedy" consumer behavior
if (consumers.empty()) {
consumers.swap(stillConsumers);
stillConsumers.clear();
}
if (verbose) fmt::print("********************\n");
} while (!consumers.empty());
if (verbose) {
if (!consumers.empty()) {
fmt::print("/!\\ Remaining consumers: possible dead-lock\n");
fmt::print("********************\n");
}
}
}
void Aidge::SequentialScheduler::resetScheduling() {
for (auto node : mGraphView->getNodes()) {
node->getOperator()->resetConsummerProducer();
}
mStaticSchedule.clear();
mStaticScheduleStep = 0;
mScheduling.clear();
}
/**
* This version is a simplified version without special handling of concatenation.
*/
Aidge::MemoryManager Aidge::SequentialScheduler::generateMemory(bool incProducers, bool wrapAroundBuffer) const {
MemoryManager memManager;
for (const auto& shedule : mStaticSchedule) {
for (const auto& node : shedule) {
if (!incProducers && node->type() == Producer_Op::Type) {
memManager.releaseDependencies(node);
continue;
}
const auto childs = node->getChildren();
AIDGE_ASSERT(node->getOperator()->operatorType() == OperatorType::Tensor, "Operator must be of Tensor type.");
const auto op = std::static_pointer_cast<OperatorTensor>(node->getOperator());
std::vector<const MemoryManager::MemoryPlane*> wrapAroundMemPlane;
// Allocate a memory plane for each node's output
for (IOIndex_t outputIdx = 0; outputIdx < node->nbOutputs(); ++outputIdx) {
const size_t requiredSize = op->getRequiredMemory(outputIdx, {});
// By default, specifies a fully monolithic memory block
size_t size = requiredSize;
size_t stride = 0;
size_t length = 1;
size_t count = 1;
if (op->getOutput(outputIdx) && op->getOutput(outputIdx)->dims().size() > 3) {
// If it is possible, assume a NCHW layout
size = op->getOutput(outputIdx)->dims().end()[-3];
stride = size;
length = op->getOutput(outputIdx)->dims().end()[-1];
count = op->getOutput(outputIdx)->dims().end()[-2];
}
// Check if wrap around buffer is possible for this node
// (re-using previous node outputs memory for this node outputs).
// => only if this node is the only child of its parent(s)
size_t wrapAroundSize = 0;
size_t wrapAroundExtra = 0;
wrapAroundMemPlane.push_back(nullptr);
// Select the best parent among all allocable nodes for
// reallocation, which is the one with most memory (in order
// to minimize the reallocation size).
IOIndex_t inputIdx = 0;
for (const auto& parent : node->dataInputs()) {
if (parent.first && parent.first->getChildren(parent.second).size() == 1
// there might be no existing plane if the parent was
// not yet scheduled (because it may be a recurrent connection)
&& memManager.getNbPlanes(parent.first) >= parent.first->nbOutputs()
// memSpace should not be already released
&& memManager.getPlanes(parent.first).end()[-parent.first->nbOutputs()+parent.second].memSpace->released == -1)
{
const bool isWrappable = (op->getNbRequiredProtected(inputIdx) < op->getNbRequiredData(inputIdx));
const MemoryManager::MemoryPlane& memPlane = memManager.getPlanes(parent.first).end()[-parent.first->nbOutputs()+parent.second];
if (isWrappable || !memManager.isWrapAround(
memPlane.memSpace,
memPlane.getFinalOffset()
- memPlane.memSpace->offset,
requiredSize))
{
if (memPlane.getSize() > wrapAroundSize + op->getNbRequiredProtected(inputIdx)
&& std::find(wrapAroundMemPlane.begin(), wrapAroundMemPlane.end(), &memPlane) == wrapAroundMemPlane.end())
{
wrapAroundSize = memPlane.getSize() - op->getNbRequiredProtected(inputIdx);
if (requiredSize > wrapAroundSize) {
wrapAroundExtra = requiredSize - wrapAroundSize;
}
wrapAroundMemPlane[outputIdx] = &memPlane;
}
if (wrapAroundExtra == 0) {
break;
}
}
}
++inputIdx;
}
// MemoryPlane to (re)use
const MemoryManager::MemoryPlane& memPlane
= (wrapAroundBuffer && wrapAroundSize > 0)
? (*wrapAroundMemPlane[outputIdx]) :
memManager.allocate(requiredSize, childs, stride, length, count);
if (wrapAroundBuffer && wrapAroundSize > 0) {
memManager.reallocate(memPlane,
node, 0,
requiredSize, true, wrapAroundExtra, childs, stride, length, count);
}
else {
memManager.reallocate(memPlane.memSpace,
node, memPlane.offset,
requiredSize, false, 0, childs, stride, length, count);
}
}
memManager.releaseDependencies(node);
memManager.tick();
}
}
return memManager;
}
void Aidge::SequentialScheduler::connectInputs(std::vector<std::shared_ptr<Aidge::Tensor>> data){
// This version of connect inputs only connects tensor inputs in input data producers.
auto inputNodes = mGraphView->getOrderedInputs();
// Assert that the number of input data producers corresponds to the number of data input
assert(data.size() == inputNodes.size() && "Scheduler connectInput error - Inconsistent number of graph inputs and inputs passed to the graph");
for (std::size_t i = 0; i < data.size(); ++i){
// TODO : maybe shallow copy instead of deepcopy
inputNodes[i].first->getOperator()->setInput(inputNodes[i].second, data[i]);
}
}
void Aidge::SequentialScheduler::forward(bool forwardDims, bool verbose, std::vector<std::shared_ptr<Aidge::Tensor>> data) {
// Collect all data input of the graph (that are producers)
if (!data.empty()){
connectInputs(data);
}
// Forward dims (if allowed)
if (forwardDims) {mGraphView->forwardDims(); }
// Generate scheduling *only if empty*
// If scheduling was already generated (in one or several steps, i.e. one or
// several successive call to generateScheduling()), do not generate it twice
if (mStaticSchedule.empty()) {
this->generateScheduling(verbose);
}
std::map<std::shared_ptr<Node>, std::string> namePtrTable;
if (verbose) namePtrTable = mGraphView->getRankedNodesName("{0} ({1}#{3})");
size_t cpt = 0;
for (const auto& runnable : mStaticSchedule.at(mStaticScheduleStep)) {
if (verbose)
fmt::print("run: {}\n", namePtrTable[runnable]);
else
drawProgressBar(static_cast<float>(cpt) / static_cast<float>(mStaticSchedule.size()), 50,
(std::string("running ") + namePtrTable[runnable]));
const auto tStart = std::chrono::high_resolution_clock::now();
runnable->forward();
const auto tEnd = std::chrono::high_resolution_clock::now();
mScheduling.push_back(SchedulingElement(runnable, tStart, tEnd));
cpt++;
}
if (!verbose) drawProgressBar(1.0, 50, " ");
fmt::print("\n");
++mStaticScheduleStep;
if (mStaticScheduleStep == mStaticSchedule.size()) {
mStaticScheduleStep = 0;
}
}
void Aidge::SequentialScheduler::saveSchedulingDiagram(const std::string& fileName) const {
auto fp = std::unique_ptr<FILE, decltype(&std::fclose)>(std::fopen((fileName + ".mmd").c_str(), "w"), &std::fclose);
if (!fp) {
AIDGE_THROW_OR_ABORT(std::runtime_error,
"Could not create scheduling diagram log file: {}", fileName + ".mmd");
}
fmt::print(fp.get(), "gantt\ndateFormat x\naxisFormat %Q µs\n\n");
if (!mScheduling.empty()) { const std::map<std::shared_ptr<Node>, std::string> namePtrTable
= mGraphView->getRankedNodesName("{0} ({1}#{3})");
const auto globalStart = mScheduling[0].start;
for (const auto& element : mScheduling) {
auto name = namePtrTable.at(element.node);
// Mermaid does not allow : character in task title
std::replace(name.begin(), name.end(), ':', '_');
fmt::print(fp.get(), "{} :{}, {}\n",
name,
std::chrono::duration_cast<std::chrono::microseconds>(element.start - globalStart).count(),
std::chrono::duration_cast<std::chrono::microseconds>(element.end - globalStart).count());
}
}
fmt::print(fp.get(), "\n");
}
std::set<std::shared_ptr<Aidge::Node>> Aidge::SequentialScheduler::getConsumers(
const std::set<std::shared_ptr<Node>>& producers) const {
std::set<std::shared_ptr<Node>> consumers;
for (const auto& producer : producers) {
const auto& childs = producer->getChildren();
for (const auto& child : childs) {
// Do not schedule childs outside current graph!
if (mGraphView->inView(child)) {
consumers.insert(child);
}
}
}
return consumers;
}
Aidge::NbElts_t Aidge::SequentialScheduler::getNbAvailableData(const std::shared_ptr<Node>& node, const IOIndex_t inputIdx) const {
const auto parent = node->inputs()[inputIdx];
if (parent.first) {
// Parent is connected, everything if fine!
return parent.first->getOperator()->getNbProducedData(parent.second);
}
else if (std::shared_ptr<Node> upperNode = mUpperNode.lock()) {
// We are inside an upper operator (for instance a MetaOperator)
// We need to connect the "local" producer-consumer model to the upper
// one, by mapping local node inputs to the upper node inputs.
IOIndex_t nodeInputIdx = 0;
for (const auto& input : mGraphView->getOrderedInputs()) {
if (input.first == node) {
// Current node is an input
const auto upperInput = upperNode->inputs()[nodeInputIdx];
if (upperInput.first) {
return upperInput.first->getOperator()->getNbProducedData(upperInput.second);
}
}
++nodeInputIdx;
}
}
// Otherwise, two cases:
if (node->getOperator()->getRawInput(inputIdx)) {
// Input is not connected but a valid tensor exists
// => This means data was fed manually to the input, without a Producer
// In this case, we assume a single-use data (unlike a Producer, which
// keep producing the data each time it is needed).
fmt::print("No producer node attached to input#{} for node {} ({})\n", inputIdx, node->name(), node->type());
return std::static_pointer_cast<Tensor>(node->getOperator()->getRawInput(inputIdx))->size();
}
else { // Input is not connected, this is an error
AIDGE_THROW_OR_ABORT(std::runtime_error, "Missing input#{} for node {} ({})\n", inputIdx, node->name(), node->type());
}
return 0;
}
Aidge::SequentialScheduler::PriorProducersConsumers Aidge::SequentialScheduler::getPriorProducersConsumers(
const std::shared_ptr<Node>& node) const
{
const auto priorCache = mPriorCache.find(node);
if (priorCache != mPriorCache.end()) {
return priorCache->second;
}
PriorProducersConsumers prior;
IOIndex_t inputIdx = 0;
for (const auto& parent : node->inputs()) {
if (parent.first &&
(node->getOperator()->getNbConsumedData(inputIdx) + node->getOperator()->getNbRequiredData(inputIdx)) >
parent.first->getOperator()->getNbProducedData(parent.second))
{
if (!mGraphView->inView(parent.first)) {
// Do not schedule prior outside the current graph!
return PriorProducersConsumers();
}
if (parent.first->type() == Producer_Op::Type) {
prior.requiredProducers.insert(parent.first);
prior.priorConsumers.insert(node);
}
else if (parent.first->type() == Memorize_Op::Type) {
// Break cycles
return PriorProducersConsumers();
}
else {
const auto& parentPrior = getPriorProducersConsumers(parent.first);
if (!parentPrior.isPrior) {
return PriorProducersConsumers();
}
else {
prior.requiredProducers.insert(parentPrior.requiredProducers.cbegin(), parentPrior.requiredProducers.cend());
prior.priorConsumers.insert(parentPrior.priorConsumers.cbegin(), parentPrior.priorConsumers.cend());
}
}
}
++inputIdx;
}
prior.isPrior = true;
if (prior.priorConsumers.empty()) {
prior.priorConsumers.insert(node);
}
mPriorCache.insert(std::make_pair(node, prior));
return prior;
}