one_copy_raster_buffer_provider.cc

// Copyright 2014 The Chromium Authors. All rights reserved.
// Use of this source code is governed by a BSD-style license that can be
// found in the LICENSE file.

#include "cc/raster/one_copy_raster_buffer_provider.h"

#include <stdint.h>

#include <algorithm>
#include <limits>
#include <utility>

#include "base/debug/alias.h"
#include "base/metrics/histogram_macros.h"
#include "base/strings/stringprintf.h"
#include "base/trace_event/process_memory_dump.h"
#include "base/trace_event/trace_event.h"
#include "cc/base/histograms.h"
#include "cc/base/math_util.h"
#include "components/viz/common/gpu/context_provider.h"
#include "components/viz/common/gpu/raster_context_provider.h"
#include "components/viz/common/resources/platform_color.h"
#include "components/viz/common/resources/resource_format.h"
#include "components/viz/common/resources/resource_sizes.h"
#include "gpu/GLES2/gl2extchromium.h"
#include "gpu/command_buffer/client/context_support.h"
#include "gpu/command_buffer/client/gpu_memory_buffer_manager.h"
#include "gpu/command_buffer/client/raster_interface.h"
#include "gpu/command_buffer/client/shared_image_interface.h"
#include "gpu/command_buffer/common/shared_image_trace_utils.h"
#include "gpu/command_buffer/common/shared_image_usage.h"
#include "ui/gfx/buffer_format_util.h"
#include "ui/gl/trace_util.h"

namespace cc {
namespace {

// 4MiB is the size of 4 512x512 tiles, which has proven to be a good
// default batch size for copy operations.
const int kMaxBytesPerCopyOperation = 1024 * 1024 * 4;

}  // namespace

// Subclass for InUsePoolResource that holds ownership of a one-copy backing
// and does cleanup of the backing when destroyed.
class OneCopyRasterBufferProvider::OneCopyGpuBacking
    : public ResourcePool::GpuBacking {
 public:
  ~OneCopyGpuBacking() override {
    if (mailbox.IsZero())
      return;
    auto* sii = worker_context_provider->SharedImageInterface();
    if (returned_sync_token.HasData())
      sii->DestroySharedImage(returned_sync_token, mailbox);
    else if (mailbox_sync_token.HasData())
      sii->DestroySharedImage(mailbox_sync_token, mailbox);
  }

  void OnMemoryDump(
      base::trace_event::ProcessMemoryDump* pmd,
      const base::trace_event::MemoryAllocatorDumpGuid& buffer_dump_guid,
      uint64_t tracing_process_id,
      int importance) const override {
    if (mailbox.IsZero())
      return;

    auto tracing_guid = gpu::GetSharedImageGUIDForTracing(mailbox);
    pmd->CreateSharedGlobalAllocatorDump(tracing_guid);
    pmd->AddOwnershipEdge(buffer_dump_guid, tracing_guid, importance);
  }

  // The ContextProvider used to clean up the mailbox
  viz::RasterContextProvider* worker_context_provider = nullptr;
};

OneCopyRasterBufferProvider::RasterBufferImpl::RasterBufferImpl(
    OneCopyRasterBufferProvider* client,
    gpu::GpuMemoryBufferManager* gpu_memory_buffer_manager,
    const ResourcePool::InUsePoolResource& in_use_resource,
    OneCopyGpuBacking* backing,
    uint64_t previous_content_id)
    : client_(client),
      backing_(backing),
      resource_size_(in_use_resource.size()),
      resource_format_(in_use_resource.format()),
      color_space_(in_use_resource.color_space()),
      previous_content_id_(previous_content_id),
      before_raster_sync_token_(backing->returned_sync_token),
      mailbox_(backing->mailbox),
      mailbox_texture_target_(backing->texture_target),
      mailbox_texture_is_overlay_candidate_(backing->overlay_candidate) {}

OneCopyRasterBufferProvider::RasterBufferImpl::~RasterBufferImpl() {
  // This SyncToken was created on the worker context after uploading the
  // texture content.
  backing_->mailbox_sync_token = after_raster_sync_token_;
  if (after_raster_sync_token_.HasData()) {
    // The returned SyncToken was waited on in Playback. We know Playback
    // happened if the |after_raster_sync_token_| was set.
    backing_->returned_sync_token = gpu::SyncToken();
  }
  backing_->mailbox = mailbox_;
}

void OneCopyRasterBufferProvider::RasterBufferImpl::Playback(
    const RasterSource* raster_source,
    const gfx::Rect& raster_full_rect,
    const gfx::Rect& raster_dirty_rect,
    uint64_t new_content_id,
    const gfx::AxisTransform2d& transform,
    const RasterSource::PlaybackSettings& playback_settings,
    const GURL& url) {
  TRACE_EVENT0("cc", "OneCopyRasterBuffer::Playback");
  // The |before_raster_sync_token_| passed in here was created on the
  // compositor thread, or given back with the texture for reuse. This call
  // returns another SyncToken generated on the worker thread to synchronize
  // with after the raster is complete.
  after_raster_sync_token_ = client_->PlaybackAndCopyOnWorkerThread(
      &mailbox_, mailbox_texture_target_, mailbox_texture_is_overlay_candidate_,
      before_raster_sync_token_, raster_source, raster_full_rect,
      raster_dirty_rect, transform, resource_size_, resource_format_,
      color_space_, playback_settings, previous_content_id_, new_content_id);
}

OneCopyRasterBufferProvider::OneCopyRasterBufferProvider(
    scoped_refptr<base::SequencedTaskRunner> task_runner,
    viz::ContextProvider* compositor_context_provider,
    viz::RasterContextProvider* worker_context_provider,
    gpu::GpuMemoryBufferManager* gpu_memory_buffer_manager,
    int max_copy_texture_chromium_size,
    bool use_partial_raster,
    bool use_gpu_memory_buffer_resources,
    int max_staging_buffer_usage_in_bytes,
    viz::ResourceFormat tile_format)
    : compositor_context_provider_(compositor_context_provider),
      worker_context_provider_(worker_context_provider),
      gpu_memory_buffer_manager_(gpu_memory_buffer_manager),
      max_bytes_per_copy_operation_(
          max_copy_texture_chromium_size
              ? std::min(kMaxBytesPerCopyOperation,
                         max_copy_texture_chromium_size)
              : kMaxBytesPerCopyOperation),
      use_partial_raster_(use_partial_raster),
      use_gpu_memory_buffer_resources_(use_gpu_memory_buffer_resources),
      bytes_scheduled_since_last_flush_(0),
      tile_format_(tile_format),
      staging_pool_(std::move(task_runner),
                    worker_context_provider,
                    use_partial_raster,
                    max_staging_buffer_usage_in_bytes) {
  DCHECK(compositor_context_provider);
  DCHECK(worker_context_provider);
  DCHECK(!IsResourceFormatCompressed(tile_format));
}

OneCopyRasterBufferProvider::~OneCopyRasterBufferProvider() {}

std::unique_ptr<RasterBuffer>
OneCopyRasterBufferProvider::AcquireBufferForRaster(
    const ResourcePool::InUsePoolResource& resource,
    uint64_t resource_content_id,
    uint64_t previous_content_id) {
  if (!resource.gpu_backing()) {
    auto backing = std::make_unique<OneCopyGpuBacking>();
    backing->worker_context_provider = worker_context_provider_;
    backing->InitOverlayCandidateAndTextureTarget(
        resource.format(), compositor_context_provider_->ContextCapabilities(),
        use_gpu_memory_buffer_resources_);
    resource.set_gpu_backing(std::move(backing));
  }
  OneCopyGpuBacking* backing =
      static_cast<OneCopyGpuBacking*>(resource.gpu_backing());
  // TODO(danakj): If resource_content_id != 0, we only need to copy/upload
  // the dirty rect.
  return std::make_unique<RasterBufferImpl>(
      this, gpu_memory_buffer_manager_, resource, backing, previous_content_id);
}

void OneCopyRasterBufferProvider::Flush() {
  // This flush on the compositor context flushes queued work on all contexts,
  // including the raster worker. Tile raster inserted a SyncToken which is
  // waited for in order to tell if a tile is ready for draw, but a flush
  // is needed to ensure the work is sent for those queries to get the right
  // answer.
  compositor_context_provider_->ContextSupport()->FlushPendingWork();
}

viz::ResourceFormat OneCopyRasterBufferProvider::GetResourceFormat() const {
  return tile_format_;
}

bool OneCopyRasterBufferProvider::IsResourcePremultiplied() const {
  // TODO(ericrk): Handle unpremultiply/dither in one-copy case as well.
  // https://crbug.com/789153
  return true;
}

bool OneCopyRasterBufferProvider::CanPartialRasterIntoProvidedResource() const {
  // While OneCopyRasterBufferProvider has an internal partial raster
  // implementation, it cannot directly partial raster into the externally
  // owned resource provided in AcquireBufferForRaster.
  return false;
}

bool OneCopyRasterBufferProvider::IsResourceReadyToDraw(
    const ResourcePool::InUsePoolResource& resource) const {
  const gpu::SyncToken& sync_token = resource.gpu_backing()->mailbox_sync_token;
  // This SyncToken() should have been set by calling OrderingBarrier() before
  // calling this.
  DCHECK(sync_token.HasData());

  // IsSyncTokenSignaled is thread-safe, no need for worker context lock.
  return worker_context_provider_->ContextSupport()->IsSyncTokenSignaled(
      sync_token);
}

uint64_t OneCopyRasterBufferProvider::SetReadyToDrawCallback(
    const std::vector<const ResourcePool::InUsePoolResource*>& resources,
    base::OnceClosure callback,
    uint64_t pending_callback_id) const {
  gpu::SyncToken latest_sync_token;
  for (const auto* in_use : resources) {
    const gpu::SyncToken& sync_token =
        in_use->gpu_backing()->mailbox_sync_token;
    if (sync_token.release_count() > latest_sync_token.release_count())
      latest_sync_token = sync_token;
  }
  uint64_t callback_id = latest_sync_token.release_count();
  DCHECK_NE(callback_id, 0u);

  // If the callback is different from the one the caller is already waiting on,
  // pass the callback through to SignalSyncToken. Otherwise the request is
  // redundant.
  if (callback_id != pending_callback_id) {
    // Use the compositor context because we want this callback on the
    // compositor thread.
    compositor_context_provider_->ContextSupport()->SignalSyncToken(
        latest_sync_token, std::move(callback));
  }

  return callback_id;
}

void OneCopyRasterBufferProvider::Shutdown() {
  staging_pool_.Shutdown();
}

gpu::SyncToken OneCopyRasterBufferProvider::PlaybackAndCopyOnWorkerThread(
    gpu::Mailbox* mailbox,
    GLenum mailbox_texture_target,
    bool mailbox_texture_is_overlay_candidate,
    const gpu::SyncToken& sync_token,
    const RasterSource* raster_source,
    const gfx::Rect& raster_full_rect,
    const gfx::Rect& raster_dirty_rect,
    const gfx::AxisTransform2d& transform,
    const gfx::Size& resource_size,
    viz::ResourceFormat resource_format,
    const gfx::ColorSpace& color_space,
    const RasterSource::PlaybackSettings& playback_settings,
    uint64_t previous_content_id,
    uint64_t new_content_id) {
  std::unique_ptr<StagingBuffer> staging_buffer =
      staging_pool_.AcquireStagingBuffer(resource_size, resource_format,
                                         previous_content_id);

  PlaybackToStagingBuffer(staging_buffer.get(), raster_source, raster_full_rect,
                          raster_dirty_rect, transform, resource_format,
                          color_space, playback_settings, previous_content_id,
                          new_content_id);

  gpu::SyncToken sync_token_after_upload = CopyOnWorkerThread(
      staging_buffer.get(), raster_source, raster_full_rect, resource_format,
      resource_size, mailbox, mailbox_texture_target,
      mailbox_texture_is_overlay_candidate, sync_token, color_space);
  staging_pool_.ReleaseStagingBuffer(std::move(staging_buffer));
  return sync_token_after_upload;
}

void OneCopyRasterBufferProvider::PlaybackToStagingBuffer(
    StagingBuffer* staging_buffer,
    const RasterSource* raster_source,
    const gfx::Rect& raster_full_rect,
    const gfx::Rect& raster_dirty_rect,
    const gfx::AxisTransform2d& transform,
    viz::ResourceFormat format,
    const gfx::ColorSpace& dst_color_space,
    const RasterSource::PlaybackSettings& playback_settings,
    uint64_t previous_content_id,
    uint64_t new_content_id) {
  // Allocate GpuMemoryBuffer if necessary.
  if (!staging_buffer->gpu_memory_buffer) {
    staging_buffer->gpu_memory_buffer =
        gpu_memory_buffer_manager_->CreateGpuMemoryBuffer(
            staging_buffer->size, BufferFormat(format),
            gfx::BufferUsage::GPU_READ_CPU_READ_WRITE, gpu::kNullSurfaceHandle);
  }

  gfx::Rect playback_rect = raster_full_rect;
  if (use_partial_raster_ && previous_content_id) {
    // Reduce playback rect to dirty region if the content id of the staging
    // buffer matches the prevous content id.
    if (previous_content_id == staging_buffer->content_id)
      playback_rect.Intersect(raster_dirty_rect);
  }

  // Log a histogram of the percentage of pixels that were saved due to
  // partial raster.
  const char* client_name = GetClientNameForMetrics();
  float full_rect_size = raster_full_rect.size().GetArea();
  if (full_rect_size > 0 && client_name) {
    float fraction_partial_rastered =
        static_cast<float>(playback_rect.size().GetArea()) / full_rect_size;
    float fraction_saved = 1.0f - fraction_partial_rastered;
    UMA_HISTOGRAM_PERCENTAGE(
        base::StringPrintf("Renderer4.%s.PartialRasterPercentageSaved.OneCopy",
                           client_name),
        100.0f * fraction_saved);
  }

  if (staging_buffer->gpu_memory_buffer) {
    gfx::GpuMemoryBuffer* buffer = staging_buffer->gpu_memory_buffer.get();
    DCHECK_EQ(1u,
              gfx::NumberOfPlanesForLinearBufferFormat(buffer->GetFormat()));
    bool rv = buffer->Map();
    DCHECK(rv);
    DCHECK(buffer->memory(0));
    // RasterBufferProvider::PlaybackToMemory only supports unsigned strides.
    DCHECK_GE(buffer->stride(0), 0);

    // TODO(https://crbug.com/870663): Temporary diagnostics.
    base::debug::Alias(&playback_rect);
    base::debug::Alias(&full_rect_size);
    base::debug::Alias(&rv);
    void* buffer_memory = buffer->memory(0);
    base::debug::Alias(&buffer_memory);
    gfx::Size staging_buffer_size = staging_buffer->size;
    base::debug::Alias(&staging_buffer_size);
    gfx::Size buffer_size = buffer->GetSize();
    base::debug::Alias(&buffer_size);

    DCHECK(!playback_rect.IsEmpty())
        << "Why are we rastering a tile that's not dirty?";
    RasterBufferProvider::PlaybackToMemory(
        buffer->memory(0), format, staging_buffer->size, buffer->stride(0),
        raster_source, raster_full_rect, playback_rect, transform,
        dst_color_space, /*gpu_compositing=*/true, playback_settings);
    buffer->Unmap();
    staging_buffer->content_id = new_content_id;
  }
}

gpu::SyncToken OneCopyRasterBufferProvider::CopyOnWorkerThread(
    StagingBuffer* staging_buffer,
    const RasterSource* raster_source,
    const gfx::Rect& rect_to_copy,
    viz::ResourceFormat resource_format,
    const gfx::Size& resource_size,
    gpu::Mailbox* mailbox,
    GLenum mailbox_texture_target,
    bool mailbox_texture_is_overlay_candidate,
    const gpu::SyncToken& sync_token,
    const gfx::ColorSpace& color_space) {
  auto* sii = worker_context_provider_->SharedImageInterface();
  DCHECK(sii);

  if (!staging_buffer->gpu_memory_buffer.get()) {
    // If GpuMemoryBuffer allocation failed (https://crbug.com/554541), then
    // we don't have anything to give to copy into the resource. We report a
    // zero mailbox that will result in checkerboarding, and be treated as OOM
    // which should retry.
    if (!mailbox->IsZero()) {
      sii->DestroySharedImage(sync_token, *mailbox);
      mailbox->SetZero();
    }
    return gpu::SyncToken();
  }

  if (mailbox->IsZero()) {
    uint32_t flags =
        gpu::SHARED_IMAGE_USAGE_DISPLAY | gpu::SHARED_IMAGE_USAGE_RASTER;
    if (mailbox_texture_is_overlay_candidate)
      flags |= gpu::SHARED_IMAGE_USAGE_SCANOUT;
    *mailbox = sii->CreateSharedImage(resource_format, resource_size,
                                      color_space, flags);
  }

  // Create staging shared image.
  if (staging_buffer->mailbox.IsZero()) {
    uint32_t flags =
        gpu::SHARED_IMAGE_USAGE_DISPLAY | gpu::SHARED_IMAGE_USAGE_RASTER;
    staging_buffer->mailbox =
        sii->CreateSharedImage(staging_buffer->gpu_memory_buffer.get(),
                               gpu_memory_buffer_manager_, color_space, flags);
  } else {
    sii->UpdateSharedImage(staging_buffer->sync_token, staging_buffer->mailbox);
  }

  viz::RasterContextProvider::ScopedRasterContextLock scoped_context(
      worker_context_provider_);
  gpu::raster::RasterInterface* ri = scoped_context.RasterInterface();
  DCHECK(ri);
  ri->WaitSyncTokenCHROMIUM(sync_token.GetConstData());
  ri->WaitSyncTokenCHROMIUM(sii->GenUnverifiedSyncToken().GetConstData());

  // Do not use queries unless COMMANDS_COMPLETED queries are supported, or
  // COMMANDS_ISSUED queries are sufficient.
  GLenum query_target = GL_NONE;

  if (worker_context_provider_->ContextCapabilities().sync_query) {
    // Use GL_COMMANDS_COMPLETED_CHROMIUM when supported because native
    // GpuMemoryBuffers can be accessed by the GPU after commands are issued
    // until GPU reads are done.
    query_target = GL_COMMANDS_COMPLETED_CHROMIUM;
  }

#if defined(OS_CHROMEOS) && defined(ARCH_CPU_ARM_FAMILY)
  // TODO(reveman): This avoids a performance problem on ARM ChromeOS devices.
  // https://crbug.com/580166
  query_target = GL_COMMANDS_ISSUED_CHROMIUM;
#endif

  // COMMANDS_ISSUED is sufficient for shared memory GpuMemoryBuffers because
  // they're uploaded using glTexImage2D (see gl::GLImageMemory::CopyTexImage).
  const auto* buffer = staging_buffer->gpu_memory_buffer.get();
  if (buffer &&
      buffer->GetType() == gfx::GpuMemoryBufferType::SHARED_MEMORY_BUFFER) {
    query_target = GL_COMMANDS_ISSUED_CHROMIUM;
  }

  if (query_target != GL_NONE) {
    if (!staging_buffer->query_id)
      ri->GenQueriesEXT(1, &staging_buffer->query_id);

    ri->BeginQueryEXT(query_target, staging_buffer->query_id);
  }

  int bytes_per_row = viz::ResourceSizes::UncheckedWidthInBytes<int>(
      rect_to_copy.width(), staging_buffer->format);
  int chunk_size_in_rows =
      std::max(1, max_bytes_per_copy_operation_ / bytes_per_row);
  // Align chunk size to 4. Required to support compressed texture formats.
  chunk_size_in_rows = MathUtil::UncheckedRoundUp(chunk_size_in_rows, 4);
  int y = 0;
  int height = rect_to_copy.height();
  while (y < height) {
    // Copy at most |chunk_size_in_rows|.
    int rows_to_copy = std::min(chunk_size_in_rows, height - y);
    DCHECK_GT(rows_to_copy, 0);

    ri->CopySubTexture(staging_buffer->mailbox, *mailbox,
                       mailbox_texture_target, 0, y, 0, y, rect_to_copy.width(),
                       rows_to_copy);
    y += rows_to_copy;

    // Increment |bytes_scheduled_since_last_flush_| by the amount of memory
    // used for this copy operation.
    bytes_scheduled_since_last_flush_ += rows_to_copy * bytes_per_row;

    if (bytes_scheduled_since_last_flush_ >= max_bytes_per_copy_operation_) {
      ri->ShallowFlushCHROMIUM();
      bytes_scheduled_since_last_flush_ = 0;
    }
  }

  if (query_target != GL_NONE)
    ri->EndQueryEXT(query_target);

  // Generate sync token on the worker context that will be sent to and waited
  // for by the display compositor before using the content generated here.
  // The same sync token is used to synchronize operations on the staging
  // buffer. Note, the query completion is generally enough to guarantee
  // ordering, but there are some paths (e.g.
  // StagingBufferPool::ReduceMemoryUsage) that may destroy the staging buffer
  // without waiting for the query completion.
  gpu::SyncToken out_sync_token =
      viz::ClientResourceProvider::GenerateSyncTokenHelper(ri);
  staging_buffer->sync_token = out_sync_token;
  return out_sync_token;
}

bool OneCopyRasterBufferProvider::CheckRasterFinishedQueries() {
  return false;
}

}  // namespace cc