graphite/dawn/DawnBuffer.cpp

/*
 * Copyright 2022 Google LLC
 *
 * Use of this source code is governed by a BSD-style license that can be
 * found in the LICENSE file.
 */

#include "src/gpu/graphite/dawn/DawnBuffer.h"

#include "include/core/SkTraceMemoryDump.h"
#include "include/private/base/SkAlign.h"
#include "src/gpu/graphite/Log.h"
#include "src/gpu/graphite/dawn/DawnAsyncWait.h"
#include "src/gpu/graphite/dawn/DawnSharedContext.h"

namespace skgpu::graphite {
namespace {
#if defined(__EMSCRIPTEN__)
bool is_map_succeeded(WGPUBufferMapAsyncStatus status) {
    return status == WGPUBufferMapAsyncStatus_Success;
}

[[maybe_unused]]
void log_map_error(WGPUBufferMapAsyncStatus status, const char*) {
    const char* statusStr;
    LogPriority priority = LogPriority::kError;
    switch (status) {
        case WGPUBufferMapAsyncStatus_ValidationError:
            statusStr = "ValidationError";
            break;
        case WGPUBufferMapAsyncStatus_Unknown:
            statusStr = "Unknown";
            break;
        case WGPUBufferMapAsyncStatus_DeviceLost:
            statusStr = "DeviceLost";
            break;
        case WGPUBufferMapAsyncStatus_DestroyedBeforeCallback:
            statusStr = "DestroyedBeforeCallback";
            priority = LogPriority::kDebug;
            break;
        case WGPUBufferMapAsyncStatus_UnmappedBeforeCallback:
            statusStr = "UnmappedBeforeCallback";
            priority = LogPriority::kDebug;
            break;
        case WGPUBufferMapAsyncStatus_MappingAlreadyPending:
            statusStr = "MappingAlreadyPending";
            break;
        case WGPUBufferMapAsyncStatus_OffsetOutOfRange:
            statusStr = "OffsetOutOfRange";
            break;
        case WGPUBufferMapAsyncStatus_SizeOutOfRange:
            statusStr = "SizeOutOfRange";
            break;
        default:
            statusStr = "<other>";
            break;
    }
    SKGPU_LOG(priority, "Buffer async map failed with status %s.", statusStr);
}

#else

bool is_map_succeeded(wgpu::MapAsyncStatus status) {
    return status == wgpu::MapAsyncStatus::Success;
}

void log_map_error(wgpu::MapAsyncStatus status, wgpu::StringView message) {
    const char* statusStr;
    switch (status) {
        case wgpu::MapAsyncStatus::InstanceDropped:
            statusStr = "InstanceDropped";
            break;
        case wgpu::MapAsyncStatus::Error:
            statusStr = "Error";
            break;
        case wgpu::MapAsyncStatus::Aborted:
            statusStr = "Aborted";
            break;
        case wgpu::MapAsyncStatus::Unknown:
            statusStr = "Unknown";
            break;
        case wgpu::MapAsyncStatus::Success:
            SK_ABORT("This status is not an error");
            break;
    }
    SKGPU_LOG(LogPriority::kError,
              "Buffer async map failed with status %s, message '%.*s'.",
              statusStr,
              static_cast<int>(message.length),
              message.data);
}
#endif  // defined(__EMSCRIPTEN__)
}  // namespace

sk_sp<DawnBuffer> DawnBuffer::Make(const DawnSharedContext* sharedContext,
                                   size_t size,
                                   BufferType type,
                                   AccessPattern accessPattern) {
    if (size <= 0) {
        return nullptr;
    }

    wgpu::BufferUsage usage = wgpu::BufferUsage::None;

    switch (type) {
        case BufferType::kVertex:
            usage = wgpu::BufferUsage::Vertex | wgpu::BufferUsage::CopyDst;
            break;
        case BufferType::kIndex:
            usage = wgpu::BufferUsage::Index | wgpu::BufferUsage::CopyDst;
            break;
        case BufferType::kXferCpuToGpu:
            usage = wgpu::BufferUsage::CopySrc | wgpu::BufferUsage::MapWrite;
            break;
        case BufferType::kXferGpuToCpu:
            usage = wgpu::BufferUsage::CopyDst | wgpu::BufferUsage::MapRead;
            break;
        case BufferType::kUniform:
            usage = wgpu::BufferUsage::Uniform | wgpu::BufferUsage::CopyDst;
            break;
        case BufferType::kStorage:
            usage = wgpu::BufferUsage::Storage | wgpu::BufferUsage::CopyDst |
                    wgpu::BufferUsage::CopySrc;
            break;
        case BufferType::kQuery:
            usage = wgpu::BufferUsage::QueryResolve | wgpu::BufferUsage::CopySrc;
            break;
        case BufferType::kIndirect:
            usage = wgpu::BufferUsage::Indirect | wgpu::BufferUsage::Storage |
                    wgpu::BufferUsage::CopyDst;
            break;
        case BufferType::kVertexStorage:
            usage = wgpu::BufferUsage::Vertex | wgpu::BufferUsage::Storage;
            break;
        case BufferType::kIndexStorage:
            usage = wgpu::BufferUsage::Index | wgpu::BufferUsage::Storage;
            break;
    }

    if (sharedContext->caps()->drawBufferCanBeMapped() &&
        accessPattern == AccessPattern::kHostVisible && type != BufferType::kXferGpuToCpu) {
        if (type == BufferType::kQuery) {
            // We can map the query buffer to get the results directly rather than having to copy to
            // a transfer buffer.
            usage |= wgpu::BufferUsage::MapRead;
        } else {
            // If the buffer is intended to be mappable, add MapWrite usage and remove
            // CopyDst.
            // We don't want to allow both CPU and GPU to write to the same buffer.
            usage |= wgpu::BufferUsage::MapWrite;
            usage &= ~wgpu::BufferUsage::CopyDst;
        }
    }

    wgpu::BufferDescriptor desc;
    desc.usage = usage;
    desc.size  = size;
    // Specifying mappedAtCreation avoids clearing the buffer on the GPU which can cause MapAsync to
    // be very slow as it waits for GPU execution to complete.
    desc.mappedAtCreation = SkToBool(usage & wgpu::BufferUsage::MapWrite);

    auto buffer = sharedContext->device().CreateBuffer(&desc);
    if (!buffer) {
        return {};
    }

    void* mappedAtCreationPtr = nullptr;
    if (desc.mappedAtCreation) {
        mappedAtCreationPtr = buffer.GetMappedRange();
        SkASSERT(mappedAtCreationPtr);
    }

    return sk_sp<DawnBuffer>(
            new DawnBuffer(sharedContext, size, std::move(buffer), mappedAtCreationPtr));
}

DawnBuffer::DawnBuffer(const DawnSharedContext* sharedContext,
                       size_t size,
                       wgpu::Buffer buffer,
                       void* mappedAtCreationPtr)
        : Buffer(sharedContext,
                 size,
                 Protected::kNo, // Dawn doesn't support protected memory
                 /*commandBufferRefsAsUsageRefs=*/buffer.GetUsage() & wgpu::BufferUsage::MapWrite)
        , fBuffer(std::move(buffer)) {
    fMapPtr = mappedAtCreationPtr;
}

#if defined(__EMSCRIPTEN__)
void DawnBuffer::prepareForReturnToCache(const std::function<void()>& takeRef) {
    // This function is only useful for Emscripten where we have to pre-map the buffer
    // once it is returned to the cache.
    SkASSERT(this->sharedContext()->caps()->bufferMapsAreAsync());

    // This implementation is almost Dawn-agnostic. However, Buffer base class doesn't have any
    // way of distinguishing a buffer that is mappable for writing from one mappable for reading.
    // We only need to re-map the former.
    if (!(fBuffer.GetUsage() & wgpu::BufferUsage::MapWrite)) {
        return;
    }
    // We cannot start an async map while the GPU is still using the buffer. We asked that
    // our Resource convert command buffer refs to usage refs. So we should never have any
    // command buffer refs.
    SkASSERT(!this->debugHasCommandBufferRef());
    // Note that the map state cannot change on another thread when we are here. We got here
    // because there were no UsageRefs on the buffer but async mapping holds a UsageRef until it
    // completes.
    if (this->isMapped()) {
        return;
    }
    takeRef();
    this->asyncMap([](void* ctx, skgpu::CallbackResult result) {
                       sk_sp<DawnBuffer> buffer(static_cast<DawnBuffer*>(ctx));
                       if (result != skgpu::CallbackResult::kSuccess) {
                           buffer->setDeleteASAP();
                       }
                   },
                   this);
}

void DawnBuffer::onAsyncMap(GpuFinishedProc proc, GpuFinishedContext ctx) {
    // This function is only useful for Emscripten where we have to use asyncMap().
    SkASSERT(this->sharedContext()->caps()->bufferMapsAreAsync());

    if (proc) {
        SkAutoMutexExclusive ex(fAsyncMutex);
        if (this->isMapped()) {
            proc(ctx, CallbackResult::kSuccess);
            return;
        }
        fAsyncMapCallbacks.push_back(RefCntedCallback::Make(proc, ctx));
    }
    if (this->isUnmappable()) {
        return;
    }
    SkASSERT(fBuffer);
    SkASSERT((fBuffer.GetUsage() & wgpu::BufferUsage::MapRead) ||
             (fBuffer.GetUsage() & wgpu::BufferUsage::MapWrite));
    SkASSERT(fBuffer.GetMapState() == wgpu::BufferMapState::Unmapped);
    bool isWrite = fBuffer.GetUsage() & wgpu::BufferUsage::MapWrite;
    auto buffer = sk_ref_sp(this);

    fBuffer.MapAsync(
            isWrite ? wgpu::MapMode::Write : wgpu::MapMode::Read,
            0,
            fBuffer.GetSize(),
            [](WGPUBufferMapAsyncStatus s, void* userData) {
                sk_sp<DawnBuffer> buffer(static_cast<DawnBuffer*>(userData));
                buffer->mapCallback(s, /*message=*/nullptr);
            },
            buffer.release());
}

void DawnBuffer::onMap() {
    SKGPU_LOG_W("Synchronous buffer mapping not supported in Dawn. Failing map request.");
}

#else

void DawnBuffer::onMap() {
    SkASSERT(!this->sharedContext()->caps()->bufferMapsAreAsync());
    SkASSERT(fBuffer);
    SkASSERT((fBuffer.GetUsage() & wgpu::BufferUsage::MapRead) ||
             (fBuffer.GetUsage() & wgpu::BufferUsage::MapWrite));
    bool isWrite = fBuffer.GetUsage() & wgpu::BufferUsage::MapWrite;

    // Use wgpu::Future and WaitAny with timeout=0 to trigger callback immediately.
    // This should work because our resource tracking mechanism should make sure that
    // the buffer is free of any GPU use at this point.
    wgpu::FutureWaitInfo mapWaitInfo{};

    mapWaitInfo.future = fBuffer.MapAsync(
            isWrite ? wgpu::MapMode::Write : wgpu::MapMode::Read,
            0,
            fBuffer.GetSize(),
            wgpu::CallbackMode::WaitAnyOnly,
            [this](wgpu::MapAsyncStatus s, wgpu::StringView m) { this->mapCallback(s, m); });

    wgpu::Device device = static_cast<const DawnSharedContext*>(sharedContext())->device();
    wgpu::Instance instance = device.GetAdapter().GetInstance();
    [[maybe_unused]] auto status = instance.WaitAny(1, &mapWaitInfo, /*timeoutNS=*/0);

    if (status != wgpu::WaitStatus::Success) {
        // WaitAny(timeout=0) might fail in this scenario:
        // - Allocates a buffer.
        // - Encodes a command buffer to copy a texture to this buffer.
        // - Submits the command buffer. If OOM happens, this command buffer will fail to
        // be submitted.
        // - The buffer is *supposed* to be free of any GPU use since the command buffer that would
        // have used it wasn't submitted successfully.
        // - If we try to map this buffer at this point, internally Dawn will try to use GPU to
        // clear this buffer to zeros, since this is its 1st use. WaitAny(timeout=0) won't work
        // since the buffer now has a pending GPU clear operation.
        //
        // To work around this, we need to try again with a blocking WaitAny(), to wait for the
        // clear operation to finish.
        // Notes:
        // - This fallback should be rare since it is caused by an OOM error during buffer
        // readbacks.
        // - For buffer writing cases, since we use mappedAtCreation, the GPU clear won't happen.
        status = instance.WaitAny(
                1, &mapWaitInfo, /*timeoutNS=*/std::numeric_limits<uint64_t>::max());
    }

    SkASSERT(status == wgpu::WaitStatus::Success);
    SkASSERT(mapWaitInfo.completed);
}
#endif  // defined(__EMSCRIPTEN__)

void DawnBuffer::onUnmap() {
    SkASSERT(fBuffer);
    SkASSERT(this->isUnmappable());

    fMapPtr = nullptr;
    fBuffer.Unmap();
}

template <typename StatusT, typename MessageT>
void DawnBuffer::mapCallback(StatusT status, MessageT message) {
    SkAutoMutexExclusive em(this->fAsyncMutex);
    if (is_map_succeeded(status)) {
        if (this->fBuffer.GetUsage() & wgpu::BufferUsage::MapWrite) {
            this->fMapPtr = this->fBuffer.GetMappedRange();
        } else {
            // If buffer is only created with MapRead usage, Dawn only allows returning
            // constant pointer. We need to use const_cast as a workaround here.
            this->fMapPtr = const_cast<void*>(this->fBuffer.GetConstMappedRange());
        }
    } else {
        log_map_error(status, message);
        for (auto& cb : this->fAsyncMapCallbacks) {
            cb->setFailureResult();
        }
    }
    this->fAsyncMapCallbacks.clear();
}

bool DawnBuffer::isUnmappable() const {
    return fBuffer.GetMapState() != wgpu::BufferMapState::Unmapped;
}

void DawnBuffer::freeGpuData() {
    if (fBuffer) {
        // Explicitly destroy the buffer since it might be ref'd by cached bind groups which are
        // not immediately cleaned up. Graphite should already guarantee that all command buffers
        // using this buffer (indirectly via BindGroups) are already completed.
        fBuffer.Destroy();
        fBuffer = nullptr;
    }
}

void DawnBuffer::setBackendLabel(char const* label) {
    SkASSERT(label);
    if (sharedContext()->caps()->setBackendLabels()) {
        fBuffer.SetLabel(label);
    }
}

} // namespace skgpu::graphite