gpu/ganesh/GrDrawingManager.cpp

/*
 * Copyright 2015 Google Inc.
 *
 * Use of this source code is governed by a BSD-style license that can be
 * found in the LICENSE file.
 */
#include "src/gpu/ganesh/GrDrawingManager.h"

#include "include/core/SkData.h"
#include "include/core/SkRect.h"
#include "include/core/SkSize.h"
#include "include/core/SkSurface.h"
#include "include/gpu/GpuTypes.h"
#include "include/gpu/ganesh/GrDirectContext.h"
#include "include/gpu/ganesh/GrRecordingContext.h"
#include "include/gpu/ganesh/GrTypes.h"
#include "include/private/base/SkAssert.h"
#include "include/private/base/SkTo.h"
#include "include/private/chromium/GrDeferredDisplayList.h"
#include "include/private/chromium/GrSurfaceCharacterization.h"
#include "include/private/gpu/ganesh/GrTypesPriv.h"
#include "src/base/SkTInternalLList.h"
#include "src/core/SkTraceEvent.h"
#include "src/gpu/GpuTypesPriv.h"
#include "src/gpu/ganesh/GrAuditTrail.h"
#include "src/gpu/ganesh/GrBufferTransferRenderTask.h"
#include "src/gpu/ganesh/GrBufferUpdateRenderTask.h"
#include "src/gpu/ganesh/GrClientMappedBufferManager.h"
#include "src/gpu/ganesh/GrCopyRenderTask.h"
#include "src/gpu/ganesh/GrDDLTask.h"
#include "src/gpu/ganesh/GrDeferredDisplayListPriv.h"
#include "src/gpu/ganesh/GrDirectContextPriv.h"
#include "src/gpu/ganesh/GrGpu.h"
#include "src/gpu/ganesh/GrGpuBuffer.h"
#include "src/gpu/ganesh/GrNativeRect.h"
#include "src/gpu/ganesh/GrOnFlushResourceProvider.h"
#include "src/gpu/ganesh/GrOpFlushState.h"
#include "src/gpu/ganesh/GrRecordingContextPriv.h"
#include "src/gpu/ganesh/GrRenderTargetProxy.h"
#include "src/gpu/ganesh/GrRenderTask.h"
#include "src/gpu/ganesh/GrRenderTaskCluster.h"
#include "src/gpu/ganesh/GrResourceAllocator.h"
#include "src/gpu/ganesh/GrResourceCache.h"
#include "src/gpu/ganesh/GrSurfaceProxyView.h"
#include "src/gpu/ganesh/GrTTopoSort.h"
#include "src/gpu/ganesh/GrTextureProxy.h"
#include "src/gpu/ganesh/GrTextureResolveManager.h"
#include "src/gpu/ganesh/GrTextureResolveRenderTask.h"
#include "src/gpu/ganesh/GrTracing.h"
#include "src/gpu/ganesh/GrTransferFromRenderTask.h"
#include "src/gpu/ganesh/GrWaitRenderTask.h"
#include "src/gpu/ganesh/GrWritePixelsRenderTask.h"
#include "src/gpu/ganesh/ops/GrOp.h"
#include "src/gpu/ganesh/ops/OpsTask.h"
#include "src/gpu/ganesh/ops/SoftwarePathRenderer.h"

#include <algorithm>
#include <memory>
#include <optional>
#include <utility>

using namespace skia_private;

///////////////////////////////////////////////////////////////////////////////////////////////////
GrDrawingManager::GrDrawingManager(GrRecordingContext* rContext,
                                   const PathRendererChain::Options& optionsForPathRendererChain,
                                   bool reduceOpsTaskSplitting)
        : fContext(rContext)
        , fOptionsForPathRendererChain(optionsForPathRendererChain)
        , fPathRendererChain(nullptr)
        , fSoftwarePathRenderer(nullptr)
        , fReduceOpsTaskSplitting(reduceOpsTaskSplitting) {
}

GrDrawingManager::~GrDrawingManager() {
    this->closeAllTasks();
    this->removeRenderTasks();
}

bool GrDrawingManager::wasAbandoned() const {
    return fContext->abandoned();
}

void GrDrawingManager::freeGpuResources() {
    for (int i = fOnFlushCBObjects.size() - 1; i >= 0; --i) {
        if (!fOnFlushCBObjects[i]->retainOnFreeGpuResources()) {
            // it's safe to just do this because we're iterating in reverse
            fOnFlushCBObjects.removeShuffle(i);
        }
    }

    // a path renderer may be holding onto resources
    fPathRendererChain = nullptr;
    fSoftwarePathRenderer = nullptr;
}

// MDB TODO: make use of the 'proxies' parameter.
bool GrDrawingManager::flush(SkSpan<GrSurfaceProxy*> proxies,
                             SkSurfaces::BackendSurfaceAccess access,
                             const GrFlushInfo& info,
                             const skgpu::MutableTextureState* newState) {
    GR_CREATE_TRACE_MARKER_CONTEXT("GrDrawingManager", "flush", fContext);

    if (fFlushing || this->wasAbandoned()) {
        if (info.fSubmittedProc) {
            info.fSubmittedProc(info.fSubmittedContext, false);
        }
        if (info.fFinishedProc) {
            info.fFinishedProc(info.fFinishedContext);
        }
        return false;
    }

    SkDEBUGCODE(this->validate());

    // As of now we only short-circuit if we got an explicit list of surfaces to flush.
    if (!proxies.empty() && !info.fNumSemaphores && !info.fFinishedProc &&
        access == SkSurfaces::BackendSurfaceAccess::kNoAccess && !newState) {
        bool allUnused = std::all_of(proxies.begin(), proxies.end(), [&](GrSurfaceProxy* proxy) {
            bool used = std::any_of(fDAG.begin(), fDAG.end(), [&](auto& task) {
                return task && task->isUsed(proxy);
            });
            return !used;
        });
        if (allUnused) {
            if (info.fSubmittedProc) {
                info.fSubmittedProc(info.fSubmittedContext, true);
            }
            return false;
        }
    }

    auto dContext = fContext->asDirectContext();
    SkASSERT(dContext);
    dContext->priv().clientMappedBufferManager()->process();

    GrGpu* gpu = dContext->priv().getGpu();
    // We have a non abandoned and direct GrContext. It must have a GrGpu.
    SkASSERT(gpu);

    fFlushing = true;

    auto resourceProvider = dContext->priv().resourceProvider();
    auto resourceCache = dContext->priv().getResourceCache();

    // Semi-usually the GrRenderTasks are already closed at this point, but sometimes Ganesh needs
    // to flush mid-draw. In that case, the SkGpuDevice's opsTasks won't be closed but need to be
    // flushed anyway. Closing such opsTasks here will mean new ones will be created to replace them
    // if the SkGpuDevice(s) write to them again.
    this->closeAllTasks();
    fActiveOpsTask = nullptr;

    this->sortTasks();

    if (!fCpuBufferCache) {
        // We cache more buffers when the backend is using client side arrays. Otherwise, we
        // expect each pool will use a CPU buffer as a staging buffer before uploading to a GPU
        // buffer object. Each pool only requires one staging buffer at a time.
        int maxCachedBuffers = fContext->priv().caps()->preferClientSideDynamicBuffers() ? 2 : 6;
        fCpuBufferCache = GrBufferAllocPool::CpuBufferCache::Make(maxCachedBuffers);
    }

    GrOpFlushState flushState(gpu, resourceProvider, &fTokenTracker, fCpuBufferCache);

    std::optional<GrTimerQuery> timerQuery;
    if (info.fFinishedWithStatsProc && (info.fGpuStatsFlags & skgpu::GpuStatsFlags::kElapsedTime)) {
        timerQuery = gpu->startTimerQuery();
    }
    GrOnFlushResourceProvider onFlushProvider(this);

    // Prepare any onFlush op lists (e.g. atlases).
    bool preFlushSuccessful = true;
    for (GrOnFlushCallbackObject* onFlushCBObject : fOnFlushCBObjects) {
        preFlushSuccessful &= onFlushCBObject->preFlush(&onFlushProvider);
    }

    bool cachePurgeNeeded = false;

    if (preFlushSuccessful) {
        bool usingReorderedDAG = false;
        GrResourceAllocator resourceAllocator(dContext);
        if (fReduceOpsTaskSplitting) {
            usingReorderedDAG = this->reorderTasks(&resourceAllocator);
            if (!usingReorderedDAG) {
                resourceAllocator.reset();
            }
        }

#if 0
        // Enable this to print out verbose GrOp information
        SkDEBUGCODE(SkDebugf("RenderTasks (%d):\n", fDAG.count()));
        for (const auto& task : fDAG) {
            SkDEBUGCODE(task->dump(/* printDependencies */ true);)
        }
#endif

        if (!resourceAllocator.failedInstantiation()) {
            if (!usingReorderedDAG) {
                for (const auto& task : fDAG) {
                    SkASSERT(task);
                    task->gatherProxyIntervals(&resourceAllocator);
                }
                resourceAllocator.planAssignment();
            }
            resourceAllocator.assign();
        }

        cachePurgeNeeded = !resourceAllocator.failedInstantiation() &&
                           this->executeRenderTasks(&flushState);
    }
    this->removeRenderTasks();

    gpu->executeFlushInfo(proxies, access, info, std::move(timerQuery), newState);

    // Give the cache a chance to purge resources that become purgeable due to flushing.
    if (cachePurgeNeeded) {
        resourceCache->purgeAsNeeded();
        cachePurgeNeeded = false;
    }
    for (GrOnFlushCallbackObject* onFlushCBObject : fOnFlushCBObjects) {
        onFlushCBObject->postFlush(fTokenTracker.nextFlushToken());
        cachePurgeNeeded = true;
    }
    if (cachePurgeNeeded) {
        resourceCache->purgeAsNeeded();
    }
    fFlushing = false;

    return true;
}

bool GrDrawingManager::submitToGpu() {
    if (fFlushing || this->wasAbandoned()) {
        return false;
    }

    auto direct = fContext->asDirectContext();
    if (!direct) {
        return false; // Can't submit while DDL recording
    }
    GrGpu* gpu = direct->priv().getGpu();
    return gpu->submitToGpu();
}

bool GrDrawingManager::executeRenderTasks(GrOpFlushState* flushState) {
#if GR_FLUSH_TIME_OP_SPEW
    SkDebugf("Flushing %d opsTasks\n", fDAG.size());
    for (int i = 0; i < fDAG.size(); ++i) {
        if (fDAG[i]) {
            SkString label;
            label.printf("task %d/%d", i, fDAG.size());
            fDAG[i]->dump(label, {}, true, true);
        }
    }
#endif

    bool anyRenderTasksExecuted = false;

    for (const auto& renderTask : fDAG) {
        if (!renderTask || !renderTask->isInstantiated()) {
             continue;
        }

        SkASSERT(renderTask->deferredProxiesAreInstantiated());

        renderTask->prepare(flushState);
    }

    // Upload all data to the GPU
    flushState->preExecuteDraws();

    // For Vulkan, if we have too many oplists to be flushed we end up allocating a lot of resources
    // for each command buffer associated with the oplists. If this gets too large we can cause the
    // devices to go OOM. In practice we usually only hit this case in our tests, but to be safe we
    // put a cap on the number of oplists we will execute before flushing to the GPU to relieve some
    // memory pressure.
    static constexpr int kMaxRenderTasksBeforeFlush = 100;
    int numRenderTasksExecuted = 0;

    // Execute the normal op lists.
    for (const auto& renderTask : fDAG) {
        SkASSERT(renderTask);
        if (!renderTask->isInstantiated()) {
            continue;
        }

        if (renderTask->execute(flushState)) {
            anyRenderTasksExecuted = true;
        }
        if (++numRenderTasksExecuted >= kMaxRenderTasksBeforeFlush) {
            flushState->gpu()->submitToGpu();
            numRenderTasksExecuted = 0;
        }
    }

    SkASSERT(!flushState->opsRenderPass());
    SkASSERT(fTokenTracker.nextDrawToken() == fTokenTracker.nextFlushToken());

    // We reset the flush state before the RenderTasks so that the last resources to be freed are
    // those that are written to in the RenderTasks. This helps to make sure the most recently used
    // resources are the last to be purged by the resource cache.
    flushState->reset();

    return anyRenderTasksExecuted;
}

void GrDrawingManager::removeRenderTasks() {
    for (const auto& task : fDAG) {
        SkASSERT(task);
        if (!task->unique() || task->requiresExplicitCleanup()) {
            // TODO: Eventually uniqueness should be guaranteed: http://skbug.com/7111.
            // DDLs, however, will always require an explicit notification for when they
            // can clean up resources.
            task->endFlush(this);
        }
        task->disown(this);
    }
    fDAG.clear();
    fReorderBlockerTaskIndices.clear();
    fLastRenderTasks.reset();
}

void GrDrawingManager::sortTasks() {
    // We separately sort the ranges around non-reorderable tasks.
    for (size_t i = 0, start = 0, end; start < SkToSizeT(fDAG.size()); ++i, start = end + 1) {
        end = i == fReorderBlockerTaskIndices.size() ? fDAG.size() : fReorderBlockerTaskIndices[i];
        SkSpan span(fDAG.begin() + start, end - start);

        SkASSERT(std::none_of(span.begin(), span.end(), [](const auto& t) {
            return t->blocksReordering();
        }));
        SkASSERT(span.end() == fDAG.end() || fDAG[end]->blocksReordering());

#if defined(SK_DEBUG)
        // In order to partition the dag array like this it must be the case that each partition
        // only depends on nodes in the partition or earlier partitions.
        auto check = [&](const GrRenderTask* task, auto&& check) -> void {
            SkASSERT(GrRenderTask::TopoSortTraits::WasOutput(task) ||
                     std::find_if(span.begin(), span.end(), [task](const auto& n) {
                         return n.get() == task; }));
            for (int i = 0; i < task->fDependencies.size(); ++i) {
                check(task->fDependencies[i], check);
            }
        };
        for (const auto& node : span) {
            check(node.get(), check);
        }
#endif

        bool sorted = GrTTopoSort<GrRenderTask, GrRenderTask::TopoSortTraits>(span, start);
        if (!sorted) {
            SkDEBUGFAIL("Render task topo sort failed.");
        }

#ifdef SK_DEBUG
        if (sorted && !span.empty()) {
            // This block checks for any unnecessary splits in the opsTasks. If two sequential
            // opsTasks could have merged it means the opsTask was artificially split.
            auto prevOpsTask = span[0]->asOpsTask();
            for (size_t j = 1; j < span.size(); ++j) {
                auto curOpsTask = span[j]->asOpsTask();

                if (prevOpsTask && curOpsTask) {
                    SkASSERT(!prevOpsTask->canMerge(curOpsTask));
                }

                prevOpsTask = curOpsTask;
            }
        }
#endif
    }
}

// Reorder the array to match the llist without reffing & unreffing sk_sp's.
// Both args must contain the same objects.
// This is basically a shim because clustering uses LList but the rest of drawmgr uses array.
template <typename T>
static void reorder_array_by_llist(const SkTInternalLList<T>& llist, TArray<sk_sp<T>>* array) {
    int i = 0;
    for (T* t : llist) {
        // Release the pointer that used to live here so it doesn't get unreffed.
        [[maybe_unused]] T* old = array->at(i).release();
        array->at(i++).reset(t);
    }
    SkASSERT(i == array->size());
}

bool GrDrawingManager::reorderTasks(GrResourceAllocator* resourceAllocator) {
    SkASSERT(fReduceOpsTaskSplitting);
    // We separately sort the ranges around non-reorderable tasks.
    bool clustered = false;
    SkTInternalLList<GrRenderTask> llist;
    for (size_t i = 0, start = 0, end; start < SkToSizeT(fDAG.size()); ++i, start = end + 1) {
        end = i == fReorderBlockerTaskIndices.size() ? fDAG.size() : fReorderBlockerTaskIndices[i];
        SkSpan span(fDAG.begin() + start, end - start);
        SkASSERT(std::none_of(span.begin(), span.end(), [](const auto& t) {
            return t->blocksReordering();
        }));

        SkTInternalLList<GrRenderTask> subllist;
        if (GrClusterRenderTasks(span, &subllist)) {
            clustered = true;
        }

        if (i < fReorderBlockerTaskIndices.size()) {
            SkASSERT(fDAG[fReorderBlockerTaskIndices[i]]->blocksReordering());
            subllist.addToTail(fDAG[fReorderBlockerTaskIndices[i]].get());
        }
        llist.concat(std::move(subllist));
    }
    if (!clustered) {
        return false;
    }

    for (GrRenderTask* task : llist) {
        task->gatherProxyIntervals(resourceAllocator);
    }
    if (!resourceAllocator->planAssignment()) {
        return false;
    }
    if (!resourceAllocator->makeBudgetHeadroom()) {
        auto dContext = fContext->asDirectContext();
        SkASSERT(dContext);
        dContext->priv().getGpu()->stats()->incNumReorderedDAGsOverBudget();
        return false;
    }
    reorder_array_by_llist(llist, &fDAG);

    int newCount = 0;
    for (int i = 0; i < fDAG.size(); i++) {
        sk_sp<GrRenderTask>& task = fDAG[i];
        if (auto opsTask = task->asOpsTask()) {
            size_t remaining = fDAG.size() - i - 1;
            SkSpan<sk_sp<GrRenderTask>> nextTasks{fDAG.end() - remaining, remaining};
            int removeCount = opsTask->mergeFrom(nextTasks);
            for (const auto& removed : nextTasks.first(removeCount)) {
                removed->disown(this);
            }
            i += removeCount;
        }
        fDAG[newCount++] = std::move(task);
    }
    fDAG.resize_back(newCount);
    return true;
}

void GrDrawingManager::closeAllTasks() {
    for (auto& task : fDAG) {
        if (task) {
            task->makeClosed(fContext);
        }
    }
}

GrRenderTask* GrDrawingManager::insertTaskBeforeLast(sk_sp<GrRenderTask> task) {
    if (!task) {
        return nullptr;
    }
    if (fDAG.empty()) {
        return fDAG.push_back(std::move(task)).get();
    }
    if (!fReorderBlockerTaskIndices.empty() && fReorderBlockerTaskIndices.back() == fDAG.size()) {
        fReorderBlockerTaskIndices.back()++;
    }
    fDAG.push_back(std::move(task));
    auto& penultimate = fDAG.fromBack(1);
    fDAG.back().swap(penultimate);
    return penultimate.get();
}

GrRenderTask* GrDrawingManager::appendTask(sk_sp<GrRenderTask> task) {
    if (!task) {
        return nullptr;
    }
    if (task->blocksReordering()) {
        fReorderBlockerTaskIndices.push_back(fDAG.size());
    }
    return fDAG.push_back(std::move(task)).get();
}

static void resolve_and_mipmap(GrGpu* gpu, GrSurfaceProxy* proxy) {
    if (!proxy->isInstantiated()) {
        return;
    }

    // In the flushSurfaces case, we need to resolve MSAA immediately after flush. This is
    // because clients expect the flushed surface's backing texture to be fully resolved
    // upon return.
    if (proxy->requiresManualMSAAResolve()) {
        auto* rtProxy = proxy->asRenderTargetProxy();
        SkASSERT(rtProxy);
        if (rtProxy->isMSAADirty()) {
            SkASSERT(rtProxy->peekRenderTarget());
            gpu->resolveRenderTarget(rtProxy->peekRenderTarget(), rtProxy->msaaDirtyRect());
            gpu->submitToGpu();
            rtProxy->markMSAAResolved();
        }
    }
    // If, after a flush, any of the proxies of interest have dirty mipmaps, regenerate them in
    // case their backend textures are being stolen.
    // (This special case is exercised by the ReimportImageTextureWithMipLevels test.)
    // FIXME: It may be more ideal to plumb down a "we're going to steal the backends" flag.
    if (auto* textureProxy = proxy->asTextureProxy()) {
        if (textureProxy->mipmapsAreDirty()) {
            SkASSERT(textureProxy->peekTexture());
            gpu->regenerateMipMapLevels(textureProxy->peekTexture());
            textureProxy->markMipmapsClean();
        }
    }
}

GrSemaphoresSubmitted GrDrawingManager::flushSurfaces(SkSpan<GrSurfaceProxy*> proxies,
                                                      SkSurfaces::BackendSurfaceAccess access,
                                                      const GrFlushInfo& info,
                                                      const skgpu::MutableTextureState* newState) {
    if (this->wasAbandoned()) {
        if (info.fSubmittedProc) {
            info.fSubmittedProc(info.fSubmittedContext, false);
        }
        if (info.fFinishedProc) {
            info.fFinishedProc(info.fFinishedContext);
        }
        return GrSemaphoresSubmitted::kNo;
    }
    SkDEBUGCODE(this->validate());

    auto direct = fContext->asDirectContext();
    SkASSERT(direct);
    GrGpu* gpu = direct->priv().getGpu();
    // We have a non abandoned and direct GrContext. It must have a GrGpu.
    SkASSERT(gpu);

    // TODO: It is important to upgrade the drawingmanager to just flushing the
    // portion of the DAG required by 'proxies' in order to restore some of the
    // semantics of this method.
    bool didFlush = this->flush(proxies, access, info, newState);
    for (GrSurfaceProxy* proxy : proxies) {
        resolve_and_mipmap(gpu, proxy);
    }

    SkDEBUGCODE(this->validate());

    if (!didFlush || (!direct->priv().caps()->backendSemaphoreSupport() && info.fNumSemaphores)) {
        return GrSemaphoresSubmitted::kNo;
    }
    return GrSemaphoresSubmitted::kYes;
}

void GrDrawingManager::addOnFlushCallbackObject(GrOnFlushCallbackObject* onFlushCBObject) {
    fOnFlushCBObjects.push_back(onFlushCBObject);
}

#if defined(GPU_TEST_UTILS)
void GrDrawingManager::testingOnly_removeOnFlushCallbackObject(GrOnFlushCallbackObject* cb) {
    int n = std::find(fOnFlushCBObjects.begin(), fOnFlushCBObjects.end(), cb) -
            fOnFlushCBObjects.begin();
    SkASSERT(n < fOnFlushCBObjects.size());
    fOnFlushCBObjects.removeShuffle(n);
}
#endif

void GrDrawingManager::setLastRenderTask(const GrSurfaceProxy* proxy, GrRenderTask* task) {
#ifdef SK_DEBUG
    if (auto prior = this->getLastRenderTask(proxy)) {
        SkASSERT(prior->isClosed() || prior == task);
    }
#endif
    uint32_t key = proxy->uniqueID().asUInt();
    if (task) {
        fLastRenderTasks.set(key, task);
    } else if (fLastRenderTasks.find(key)) {
        fLastRenderTasks.remove(key);
    }
}

GrRenderTask* GrDrawingManager::getLastRenderTask(const GrSurfaceProxy* proxy) const {
    auto entry = fLastRenderTasks.find(proxy->uniqueID().asUInt());
    return entry ? *entry : nullptr;
}

skgpu::ganesh::OpsTask* GrDrawingManager::getLastOpsTask(const GrSurfaceProxy* proxy) const {
    GrRenderTask* task = this->getLastRenderTask(proxy);
    return task ? task->asOpsTask() : nullptr;
}

void GrDrawingManager::moveRenderTasksToDDL(GrDeferredDisplayList* ddl) {
    SkDEBUGCODE(this->validate());

    // no renderTask should receive a new command after this
    this->closeAllTasks();
    fActiveOpsTask = nullptr;

    this->sortTasks();

    fDAG.swap(ddl->fRenderTasks);
    SkASSERT(fDAG.empty());
    fReorderBlockerTaskIndices.clear();

    for (auto& renderTask : ddl->fRenderTasks) {
        renderTask->disown(this);
        renderTask->prePrepare(fContext);
    }

    ddl->fArenas = std::move(fContext->priv().detachArenas());

    fContext->priv().detachProgramData(&ddl->fProgramData);

    SkDEBUGCODE(this->validate());
}

void GrDrawingManager::createDDLTask(sk_sp<const GrDeferredDisplayList> ddl,
                                     sk_sp<GrRenderTargetProxy> newDest) {
    SkDEBUGCODE(this->validate());

    if (fActiveOpsTask) {
        // This is a temporary fix for the partial-MDB world. In that world we're not
        // reordering so ops that (in the single opsTask world) would've just glommed onto the
        // end of the single opsTask but referred to a far earlier RT need to appear in their
        // own opsTask.
        fActiveOpsTask->makeClosed(fContext);
        fActiveOpsTask = nullptr;
    }

    // Propagate the DDL proxy's state information to the replay target.
    if (ddl->priv().targetProxy()->isMSAADirty()) {
        auto nativeRect = GrNativeRect::MakeIRectRelativeTo(
                ddl->characterization().origin(),
                ddl->priv().targetProxy()->backingStoreDimensions().height(),
                ddl->priv().targetProxy()->msaaDirtyRect());
        newDest->markMSAADirty(nativeRect);
    }
    GrTextureProxy* newTextureProxy = newDest->asTextureProxy();
    if (newTextureProxy && skgpu::Mipmapped::kYes == newTextureProxy->mipmapped()) {
        newTextureProxy->markMipmapsDirty();
    }

    // Here we jam the proxy that backs the current replay SkSurface into the LazyProxyData.
    // The lazy proxy that references it (in the DDL opsTasks) will then steal its GrTexture.
    ddl->fLazyProxyData->fReplayDest = newDest.get();

    // Add a task to handle drawing and lifetime management of the DDL.
    SkDEBUGCODE(auto ddlTask =) this->appendTask(sk_make_sp<GrDDLTask>(this,
                                                                       std::move(newDest),
                                                                       std::move(ddl)));
    SkASSERT(ddlTask->isClosed());

    SkDEBUGCODE(this->validate());
}

#ifdef SK_DEBUG
void GrDrawingManager::validate() const {
    if (fActiveOpsTask) {
        SkASSERT(!fDAG.empty());
        SkASSERT(!fActiveOpsTask->isClosed());
        SkASSERT(fActiveOpsTask == fDAG.back().get());
    }

    for (int i = 0; i < fDAG.size(); ++i) {
        if (fActiveOpsTask != fDAG[i].get()) {
            // The resolveTask associated with the activeTask remains open for as long as the
            // activeTask does.
            bool isActiveResolveTask =
                fActiveOpsTask && fActiveOpsTask->fTextureResolveTask == fDAG[i].get();
            bool isAtlas = fDAG[i]->isSetFlag(GrRenderTask::kAtlas_Flag);
            SkASSERT(isActiveResolveTask || isAtlas || fDAG[i]->isClosed());
        }
    }

    // The active opsTask, if any, should always be at the back of the DAG.
    if (!fDAG.empty()) {
        if (fDAG.back()->isSetFlag(GrRenderTask::kAtlas_Flag)) {
            SkASSERT(fActiveOpsTask == nullptr);
            SkASSERT(!fDAG.back()->isClosed());
        } else if (fDAG.back()->isClosed()) {
            SkASSERT(fActiveOpsTask == nullptr);
        } else {
            SkASSERT(fActiveOpsTask == fDAG.back().get());
        }
    } else {
        SkASSERT(fActiveOpsTask == nullptr);
    }
}
#endif // SK_DEBUG

void GrDrawingManager::closeActiveOpsTask() {
    if (fActiveOpsTask) {
        // This is a temporary fix for the partial-MDB world. In that world we're not
        // reordering so ops that (in the single opsTask world) would've just glommed onto the
        // end of the single opsTask but referred to a far earlier RT need to appear in their
        // own opsTask.
        fActiveOpsTask->makeClosed(fContext);
        fActiveOpsTask = nullptr;
    }
}

sk_sp<skgpu::ganesh::OpsTask> GrDrawingManager::newOpsTask(GrSurfaceProxyView surfaceView,
                                                           sk_sp<GrArenas> arenas) {
    SkDEBUGCODE(this->validate());
    SkASSERT(fContext);

    this->closeActiveOpsTask();

    sk_sp<skgpu::ganesh::OpsTask> opsTask(new skgpu::ganesh::OpsTask(
            this, std::move(surfaceView), fContext->priv().auditTrail(), std::move(arenas)));

    SkASSERT(this->getLastRenderTask(opsTask->target(0)) == opsTask.get());

    this->appendTask(opsTask);

    fActiveOpsTask = opsTask.get();

    SkDEBUGCODE(this->validate());
    return opsTask;
}

void GrDrawingManager::addAtlasTask(sk_sp<GrRenderTask> atlasTask,
                                    GrRenderTask* previousAtlasTask) {
    SkDEBUGCODE(this->validate());
    SkASSERT(fContext);

    if (previousAtlasTask) {
        previousAtlasTask->makeClosed(fContext);
        for (GrRenderTask* previousAtlasUser : previousAtlasTask->dependents()) {
            // Make the new atlas depend on everybody who used the old atlas, and close their tasks.
            // This guarantees that the previous atlas is totally out of service before we render
            // the next one, meaning there is only ever one atlas active at a time and that they can
            // all share the same texture.
            atlasTask->addDependency(previousAtlasUser);
            previousAtlasUser->makeClosed(fContext);
            if (previousAtlasUser == fActiveOpsTask) {
                fActiveOpsTask = nullptr;
            }
        }
    }

    atlasTask->setFlag(GrRenderTask::kAtlas_Flag);
    this->insertTaskBeforeLast(std::move(atlasTask));

    SkDEBUGCODE(this->validate());
}

GrTextureResolveRenderTask* GrDrawingManager::newTextureResolveRenderTaskBefore(
        const GrCaps& caps) {
    // Unlike in the "new opsTask" case, we do not want to close the active opsTask, nor (if we are
    // in sorting and opsTask reduction mode) the render tasks that depend on any proxy's current
    // state. This is because those opsTasks can still receive new ops and because if they refer to
    // the mipmapped version of 'proxy', they will then come to depend on the render task being
    // created here.
    //
    // Add the new textureResolveTask before the fActiveOpsTask (if not in
    // sorting/opsTask-splitting-reduction mode) because it will depend upon this resolve task.
    // NOTE: Putting it here will also reduce the amount of work required by the topological sort.
    GrRenderTask* task = this->insertTaskBeforeLast(sk_make_sp<GrTextureResolveRenderTask>());
    return static_cast<GrTextureResolveRenderTask*>(task);
}

void GrDrawingManager::newTextureResolveRenderTask(sk_sp<GrSurfaceProxy> proxy,
                                                   GrSurfaceProxy::ResolveFlags flags,
                                                   const GrCaps& caps) {
    SkDEBUGCODE(this->validate());
    SkASSERT(fContext);

    if (!proxy->requiresManualMSAAResolve()) {
        SkDEBUGCODE(this->validate());
        return;
    }

    GrRenderTask* lastTask = this->getLastRenderTask(proxy.get());
    if (!proxy->asRenderTargetProxy()->isMSAADirty() && (!lastTask || lastTask->isClosed())) {
        SkDEBUGCODE(this->validate());
        return;
    }

    this->closeActiveOpsTask();

    auto resolveTask = sk_make_sp<GrTextureResolveRenderTask>();
    // Add proxy also adds all the needed dependencies we need
    resolveTask->addProxy(this, std::move(proxy), flags, caps);

    auto task = this->appendTask(std::move(resolveTask));
    task->makeClosed(fContext);

    // We have closed the previous active oplist but since a new oplist isn't being added there
    // shouldn't be an active one.
    SkASSERT(!fActiveOpsTask);
    SkDEBUGCODE(this->validate());
}

void GrDrawingManager::newWaitRenderTask(const sk_sp<GrSurfaceProxy>& proxy,
                                         std::unique_ptr<std::unique_ptr<GrSemaphore>[]> semaphores,
                                         int numSemaphores) {
    SkDEBUGCODE(this->validate());
    SkASSERT(fContext);

    sk_sp<GrWaitRenderTask> waitTask = sk_make_sp<GrWaitRenderTask>(GrSurfaceProxyView(proxy),
                                                                    std::move(semaphores),
                                                                    numSemaphores);

    if (fActiveOpsTask && (fActiveOpsTask->target(0) == proxy.get())) {
        SkASSERT(this->getLastRenderTask(proxy.get()) == fActiveOpsTask);
        this->insertTaskBeforeLast(waitTask);
        // In this case we keep the current renderTask open but just insert the new waitTask
        // before it in the list. The waitTask will never need to trigger any resolves or mip
        // map generation which is the main advantage of going through the proxy version.
        // Additionally we would've had to temporarily set the wait task as the lastRenderTask
        // on the proxy, add the dependency, and then reset the lastRenderTask to
        // fActiveOpsTask. Additionally we make the waitTask depend on all of fActiveOpsTask
        // dependencies so that we don't unnecessarily reorder the waitTask before them.
        // Note: Any previous Ops already in fActiveOpsTask will get blocked by the wait
        // semaphore even though they don't need to be for correctness.

        // Make sure we add the dependencies of fActiveOpsTask to waitTask first or else we'll
        // get a circular self dependency of waitTask on waitTask.
        waitTask->addDependenciesFromOtherTask(fActiveOpsTask);
        fActiveOpsTask->addDependency(waitTask.get());
    } else {
        // In this case we just close the previous RenderTask and start and append the waitTask
        // to the DAG. Since it is the last task now we call setLastRenderTask on the proxy. If
        // there is a lastTask on the proxy we make waitTask depend on that task. This
        // dependency isn't strictly needed but it does keep the DAG from reordering the
        // waitTask earlier and blocking more tasks.
        if (GrRenderTask* lastTask = this->getLastRenderTask(proxy.get())) {
            waitTask->addDependency(lastTask);
        }
        this->setLastRenderTask(proxy.get(), waitTask.get());
        this->closeActiveOpsTask();
        this->appendTask(waitTask);
    }
    waitTask->makeClosed(fContext);

    SkDEBUGCODE(this->validate());
}

void GrDrawingManager::newTransferFromRenderTask(const sk_sp<GrSurfaceProxy>& srcProxy,
                                                 const SkIRect& srcRect,
                                                 GrColorType surfaceColorType,
                                                 GrColorType dstColorType,
                                                 sk_sp<GrGpuBuffer> dstBuffer,
                                                 size_t dstOffset) {
    SkDEBUGCODE(this->validate());
    SkASSERT(fContext);
    this->closeActiveOpsTask();

    GrRenderTask* task = this->appendTask(sk_make_sp<GrTransferFromRenderTask>(
            srcProxy, srcRect, surfaceColorType, dstColorType,
            std::move(dstBuffer), dstOffset));

    const GrCaps& caps = *fContext->priv().caps();

    // We always say skgpu::Mipmapped::kNo here since we are always just copying from the base
    // layer. We don't need to make sure the whole mip map chain is valid.
    task->addDependency(
            this, srcProxy.get(), skgpu::Mipmapped::kNo, GrTextureResolveManager(this), caps);
    task->makeClosed(fContext);

    // We have closed the previous active oplist but since a new oplist isn't being added there
    // shouldn't be an active one.
    SkASSERT(!fActiveOpsTask);
    SkDEBUGCODE(this->validate());
}

void GrDrawingManager::newBufferTransferTask(sk_sp<GrGpuBuffer> src,
                                             size_t srcOffset,
                                             sk_sp<GrGpuBuffer> dst,
                                             size_t dstOffset,
                                             size_t size) {
    SkASSERT(src);
    SkASSERT(dst);
    SkASSERT(srcOffset + size <= src->size());
    SkASSERT(dstOffset + size <= dst->size());
    SkASSERT(src->intendedType() == GrGpuBufferType::kXferCpuToGpu);
    SkASSERT(dst->intendedType() != GrGpuBufferType::kXferCpuToGpu);

    SkDEBUGCODE(this->validate());
    SkASSERT(fContext);

    this->closeActiveOpsTask();

    sk_sp<GrRenderTask> task = GrBufferTransferRenderTask::Make(std::move(src),
                                                                srcOffset,
                                                                std::move(dst),
                                                                dstOffset,
                                                                size);
    SkASSERT(task);

    this->appendTask(task);
    task->makeClosed(fContext);

    // We have closed the previous active oplist but since a new oplist isn't being added there
    // shouldn't be an active one.
    SkASSERT(!fActiveOpsTask);
    SkDEBUGCODE(this->validate());
}

void GrDrawingManager::newBufferUpdateTask(sk_sp<SkData> src,
                                           sk_sp<GrGpuBuffer> dst,
                                           size_t dstOffset) {
    SkASSERT(src);
    SkASSERT(dst);
    SkASSERT(dstOffset + src->size() <= dst->size());
    SkASSERT(dst->intendedType() != GrGpuBufferType::kXferCpuToGpu);
    SkASSERT(!dst->isMapped());

    SkDEBUGCODE(this->validate());
    SkASSERT(fContext);

    this->closeActiveOpsTask();

    sk_sp<GrRenderTask> task = GrBufferUpdateRenderTask::Make(std::move(src),
                                                              std::move(dst),
                                                              dstOffset);
    SkASSERT(task);

    this->appendTask(task);
    task->makeClosed(fContext);

    // We have closed the previous active oplist but since a new oplist isn't being added there
    // shouldn't be an active one.
    SkASSERT(!fActiveOpsTask);
    SkDEBUGCODE(this->validate());
}

sk_sp<GrRenderTask> GrDrawingManager::newCopyRenderTask(sk_sp<GrSurfaceProxy> dst,
                                                        SkIRect dstRect,
                                                        const sk_sp<GrSurfaceProxy>& src,
                                                        SkIRect srcRect,
                                                        GrSamplerState::Filter filter,
                                                        GrSurfaceOrigin origin) {
    SkDEBUGCODE(this->validate());
    SkASSERT(fContext);

    // It'd be nicer to check this in GrCopyRenderTask::Make. This gets complicated because of
    // "active ops task" tracking. dst will be the target of our copy task but it might also be the
    // target of the active ops task. We currently require the active ops task to be closed before
    // making a new task that targets the same proxy. However, if we first close the active ops
    // task, then fail to make a copy task, the next active ops task may target the same proxy. This
    // will trip an assert related to unnecessary ops task splitting.
    if (src->framebufferOnly()) {
        return nullptr;
    }

    this->closeActiveOpsTask();

    sk_sp<GrRenderTask> task = GrCopyRenderTask::Make(this,
                                                      std::move(dst),
                                                      dstRect,
                                                      src,
                                                      srcRect,
                                                      filter,
                                                      origin);
    if (!task) {
        return nullptr;
    }

    this->appendTask(task);

    const GrCaps& caps = *fContext->priv().caps();
    // We always say skgpu::Mipmapped::kNo here since we are always just copying from the base layer
    // to another base layer. We don't need to make sure the whole mip map chain is valid.
    task->addDependency(
            this, src.get(), skgpu::Mipmapped::kNo, GrTextureResolveManager(this), caps);
    task->makeClosed(fContext);

    // We have closed the previous active oplist but since a new oplist isn't being added there
    // shouldn't be an active one.
    SkASSERT(!fActiveOpsTask);
    SkDEBUGCODE(this->validate());
    return task;
}

bool GrDrawingManager::newWritePixelsTask(sk_sp<GrSurfaceProxy> dst,
                                          SkIRect rect,
                                          GrColorType srcColorType,
                                          GrColorType dstColorType,
                                          const GrMipLevel levels[],
                                          int levelCount) {
    SkDEBUGCODE(this->validate());
    SkASSERT(fContext);

    this->closeActiveOpsTask();
    const GrCaps& caps = *fContext->priv().caps();

    // On platforms that prefer flushes over VRAM use (i.e., ANGLE) we're better off forcing a
    // complete flush here.
    if (!caps.preferVRAMUseOverFlushes()) {
        this->flushSurfaces(SkSpan<GrSurfaceProxy*>{},
                            SkSurfaces::BackendSurfaceAccess::kNoAccess,
                            GrFlushInfo{},
                            nullptr);
    }

    GrRenderTask* task = this->appendTask(GrWritePixelsTask::Make(this,
                                                                  std::move(dst),
                                                                  rect,
                                                                  srcColorType,
                                                                  dstColorType,
                                                                  levels,
                                                                  levelCount));
    if (!task) {
        return false;
    }

    task->makeClosed(fContext);

    // We have closed the previous active oplist but since a new oplist isn't being added there
    // shouldn't be an active one.
    SkASSERT(!fActiveOpsTask);
    SkDEBUGCODE(this->validate());
    return true;
}

/*
 * This method finds a path renderer that can draw the specified path on
 * the provided target.
 * Due to its expense, the software path renderer has split out so it can
 * can be individually allowed/disallowed via the "allowSW" boolean.
 */
skgpu::ganesh::PathRenderer* GrDrawingManager::getPathRenderer(
        const PathRenderer::CanDrawPathArgs& args,
        bool allowSW,
        PathRendererChain::DrawType drawType,
        PathRenderer::StencilSupport* stencilSupport) {
    if (!fPathRendererChain) {
        fPathRendererChain =
                std::make_unique<PathRendererChain>(fContext, fOptionsForPathRendererChain);
    }

    auto pr = fPathRendererChain->getPathRenderer(args, drawType, stencilSupport);
    if (!pr && allowSW) {
        auto swPR = this->getSoftwarePathRenderer();
        if (PathRenderer::CanDrawPath::kNo != swPR->canDrawPath(args)) {
            pr = swPR;
        }
    }

#if GR_PATH_RENDERER_SPEW
    if (pr) {
        SkDebugf("getPathRenderer: %s\n", pr->name());
    }
#endif

    return pr;
}

skgpu::ganesh::PathRenderer* GrDrawingManager::getSoftwarePathRenderer() {
    if (!fSoftwarePathRenderer) {
        fSoftwarePathRenderer.reset(new skgpu::ganesh::SoftwarePathRenderer(
                fContext->priv().proxyProvider(),
                fOptionsForPathRendererChain.fAllowPathMaskCaching));
    }
    return fSoftwarePathRenderer.get();
}

skgpu::ganesh::AtlasPathRenderer* GrDrawingManager::getAtlasPathRenderer() {
    if (!fPathRendererChain) {
        fPathRendererChain = std::make_unique<PathRendererChain>(fContext,
                                                                 fOptionsForPathRendererChain);
    }
    return fPathRendererChain->getAtlasPathRenderer();
}

skgpu::ganesh::PathRenderer* GrDrawingManager::getTessellationPathRenderer() {
    if (!fPathRendererChain) {
        fPathRendererChain = std::make_unique<PathRendererChain>(fContext,
                                                                 fOptionsForPathRendererChain);
    }
    return fPathRendererChain->getTessellationPathRenderer();
}

void GrDrawingManager::flushIfNecessary() {
    auto direct = fContext->asDirectContext();
    if (!direct) {
        return;
    }

    auto resourceCache = direct->priv().getResourceCache();
    if (resourceCache && resourceCache->requestsFlush()) {
        if (this->flush({}, SkSurfaces::BackendSurfaceAccess::kNoAccess, GrFlushInfo(), nullptr)) {
            this->submitToGpu();
        }
        resourceCache->purgeAsNeeded();
    }
}