ArcBridge: Add VideoEncodeAccelerator implementation.

chrome/gpu/gpu_arc_video_encode_accelerator.cc

+// Copyright 2017 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 "chrome/gpu/gpu_arc_video_encode_accelerator.h"
+
+#include <utility>
+
+#include "base/logging.h"
+#include "base/sys_info.h"
+#include "media/base/video_types.h"
+#include "media/gpu/gpu_video_encode_accelerator_factory.h"
+#include "mojo/public/cpp/bindings/strong_binding.h"
+#include "mojo/public/cpp/bindings/type_converter.h"
+#include "mojo/public/cpp/system/platform_handle.h"
+
+#define DVLOGF(x) DVLOG(x) << __func__ << "(): "
+
+namespace chromeos {
+namespace arc {
+
+GpuArcVideoEncodeAccelerator::GpuArcVideoEncodeAccelerator(
+    const gpu::GpuPreferences& gpu_preferences)
+    : gpu_preferences_(gpu_preferences) {}
+
+GpuArcVideoEncodeAccelerator::~GpuArcVideoEncodeAccelerator() = default;
+
+// VideoEncodeAccelerator::Client implementation.
+void GpuArcVideoEncodeAccelerator::RequireBitstreamBuffers(
+    unsigned int input_count,
+    const gfx::Size& coded_size,
+    size_t output_buffer_size) {
+  DVLOGF(2) << "input_count=" << input_count
+            << ", coded_size=" << coded_size.ToString()
+            << ", output_buffer_size=" << output_buffer_size;
+  DCHECK(client_);
+  coded_size_ = coded_size;
+  client_->RequireBitstreamBuffers(input_count, coded_size, output_buffer_size);
+}
+
+void GpuArcVideoEncodeAccelerator::BitstreamBufferReady(
+    int32_t bitstream_buffer_id,
+    size_t payload_size,
+    bool key_frame,
+    base::TimeDelta timestamp) {
+  DVLOGF(2) << "id=" << bitstream_buffer_id;
+  DCHECK(client_);
+  client_->BitstreamBufferReady(bitstream_buffer_id, payload_size, key_frame,
+                                timestamp.InMicroseconds());
+}
+
+void GpuArcVideoEncodeAccelerator::NotifyError(Error error) {
+  DVLOGF(2) << "error=" << error;
+  DCHECK(client_);
+  client_->NotifyError(error);
+}
+
+// ::arc::mojom::VideoEncodeAccelerator implementation.
+void GpuArcVideoEncodeAccelerator::GetSupportedProfiles(
+    const GetSupportedProfilesCallback& callback) {
+  callback.Run(media::GpuVideoEncodeAcceleratorFactory::GetSupportedProfiles(
+      gpu_preferences_));
+}
+
+void GpuArcVideoEncodeAccelerator::Initialize(
+    VideoPixelFormat input_format,
+    const gfx::Size& visible_size,
+    VideoEncodeAccelerator::StorageType input_storage,
+    VideoCodecProfile output_profile,
+    uint32_t initial_bitrate,
+    VideoEncodeClientPtr client,
+    const InitializeCallback& callback) {
+  DVLOGF(2) << "visible_size=" << visible_size.ToString()
+            << ", profile=" << output_profile;
+  if (!client) {

[dcheng, 2017/06/21 09:42:29]

Does Mojo actually allow a null interface pointer to passed if the parameter
wasn't marked optional? 

[Owen Lin, 2017/06/22 07:43:14]

You're right. I tested it, it failed in the IPC message validation phase. Code
removed.

+    LOG(ERROR) << "Invalid client";
+    callback.Run(false);
+    return;
+  }
+
+  input_pixel_format_ = static_cast<media::VideoPixelFormat>(input_format);

[dcheng, 2017/06/21 09:42:28]

Do we need these casts? Seems like they're the same time

Nit: also spell VideoPixelFormat consistently, sometimes it's spelled without
media::, but other times, we include the media:: qualifier.

[Owen Lin, 2017/06/22 07:43:14]

Removed. It should be removed after we type mapping the enums. 

+  visible_size_ = visible_size;
+  accelerator_ = media::GpuVideoEncodeAcceleratorFactory::CreateVEA(
+      input_pixel_format_, visible_size_,
+      static_cast<media::VideoCodecProfile>(output_profile), initial_bitrate,

[dcheng, 2017/06/21 09:42:29]

Similarly, I'm not sure why this requires a cast.

[Owen Lin, 2017/06/22 07:43:14]

Done.

+      this, gpu_preferences_);
+  if (accelerator_ == nullptr) {
+    LOG(ERROR) << "Failed to create a VideoEncodeAccelerator.";

[dcheng, 2017/06/21 09:42:29]

Would personally suggest DLOG for everything that uses LOG here.

[Owen Lin, 2017/06/22 07:43:14]

Those messages should only be logged when errors happen, when it happens, the
info would be helpful.
Why would you suggest not to use LOG?

[dcheng, 2017/06/22 20:05:13]

Right, but is it useful to non-developers? Using LOG instead of DLOG means these
strings are shipped to all CrOS users.

[Owen Lin, 2017/06/23 01:37:54]

OK. I got your point. We would like to use LOG simply because we don't run debug
build in the lab.

[dcheng, 2017/06/23 07:38:44]

I don't think we should be bloating official builds with logging strings that
are only used for internal debugging. Among other things, this contradicts the
general advice about logging. See
https://chromium.googlesource.com/chromium/src/+/master/styleguide/c++/c++.md...

If this is being used in labs, it shouldn't be that hard to make a build which
enables these logs and test on that machine/config.

+    callback.Run(false);
+    return;
+  }
+  client_ = std::move(client);
+  callback.Run(true);
+}
+
+static void DropSharedMemory(std::unique_ptr<base::SharedMemory> shm) {
+  // Just let |shm| fall out of scope.
+}
+
+void GpuArcVideoEncodeAccelerator::Encode(
+    mojo::ScopedHandle handle,
+    std::vector<::arc::VideoFramePlane> planes,
+    int64_t timestamp,
+    bool force_keyframe) {
+  DVLOGF(2) << "timestamp=" << timestamp;
+  if (!accelerator_) {
+    LOG(ERROR) << "Accelerator is not initialized.";
+    return;
+  }
+
+  if (planes.empty()) {  // EOS
+    accelerator_->Encode(media::VideoFrame::CreateEOSFrame(), force_keyframe);
+    return;
+  }
+
+  base::ScopedFD fd = UnwrapFdFromMojoHandle(std::move(handle));
+  if (!fd.is_valid())
+    return;
+
+  size_t allocation_size =
+      media::VideoFrame::AllocationSize(input_pixel_format_, coded_size_);
+
+  // TODO(rockot): Pass GUIDs through Mojo. https://crbug.com/713763.
+  // TODO(rockot): This fd comes from a mojo::ScopedHandle in
+  // GpuArcVideoService::BindSharedMemory. That should be passed through,
+  // rather than pulling out the fd. https://crbug.com/713763.
+  // TODO(rockot): Pass through a real size rather than |0|.
+  base::UnguessableToken guid = base::UnguessableToken::Create();
+  base::SharedMemoryHandle shm_handle(base::FileDescriptor(fd.release(), true),
+                                      0u, guid);
+  auto shm = base::MakeUnique<base::SharedMemory>(shm_handle, true);
+
+  base::CheckedNumeric<off_t> map_offset = planes[0].offset;
+  base::CheckedNumeric<size_t> map_size = allocation_size;
+  const uint32_t aligned_offset =
+      planes[0].offset % base::SysInfo::VMAllocationGranularity();
+  map_offset -= aligned_offset;
+  map_size += aligned_offset;
+
+  if (!map_offset.IsValid() || !map_size.IsValid()) {
+    LOG(ERROR) << "Invalid map_offset or map_size";
+    client_->NotifyError(Error::kInvalidArgumentError);
+    return;
+  }
+  if (!shm->MapAt(map_offset.ValueOrDie(), map_size.ValueOrDie())) {
+    LOG(ERROR) << "Failed to map memory.";
+    client_->NotifyError(Error::kPlatformFailureError);
+    return;
+  }
+
+  uint8_t* shm_memory = reinterpret_cast<uint8_t*>(shm->memory());
+  auto frame = media::VideoFrame::WrapExternalSharedMemory(
+      input_pixel_format_, coded_size_, gfx::Rect(visible_size_), visible_size_,
+      shm_memory + aligned_offset, allocation_size, shm_handle,
+      planes[0].offset, base::TimeDelta::FromMicroseconds(timestamp));
+
+  // Wrap |shm| in a callback and add it as a destruction observer, so it
+  // stays alive and mapped until |frame| goes out of scope.
+  frame->AddDestructionObserver(
+      base::Bind(&DropSharedMemory, base::Passed(&shm)));
+  accelerator_->Encode(frame, force_keyframe);
+}
+
+void GpuArcVideoEncodeAccelerator::UseOutputBitstreamBuffer(
+    int32_t bitstream_buffer_id,
+    mojo::ScopedHandle shmem_fd,
+    uint32_t offset,
+    uint32_t size) {
+  DVLOGF(2) << "id=" << bitstream_buffer_id;
+  if (!accelerator_) {
+    LOG(ERROR) << "Accelerator is not initialized.";
+    return;
+  }
+
+  base::ScopedFD fd = UnwrapFdFromMojoHandle(std::move(shmem_fd));
+  if (!fd.is_valid())
+    return;
+
+  // TODO(rockot): Pass GUIDs through Mojo. https://crbug.com/713763.
+  // TODO(rockot): This fd comes from a mojo::ScopedHandle in
+  // GpuArcVideoService::BindSharedMemory. That should be passed through,
+  // rather than pulling out the fd. https://crbug.com/713763.
+  // TODO(rockot): Pass through a real size rather than |0|.
+  base::UnguessableToken guid = base::UnguessableToken::Create();
+  base::SharedMemoryHandle shm_handle(base::FileDescriptor(fd.release(), true),
+                                      0u, guid);

[dcheng, 2017/06/21 09:42:29]

What is the behavior if we just pass 0 here? Does the Map() later just fail?

[Owen Lin, 2017/06/22 07:43:14]

AFAIK, this argument is not used yet. rockot@ should know the details. (also see
crbug.com/713763)

+  accelerator_->UseOutputBitstreamBuffer(
+      media::BitstreamBuffer(bitstream_buffer_id, shm_handle, size, offset));
+}
+
+void GpuArcVideoEncodeAccelerator::RequestEncodingParametersChange(
+    uint32_t bitrate,
+    uint32_t framerate) {
+  DVLOGF(2) << "bitrate=" << bitrate << ", framerate=" << framerate;
+  if (!accelerator_) {
+    LOG(ERROR) << "Accelerator is not initialized.";
+    return;
+  }
+  accelerator_->RequestEncodingParametersChange(bitrate, framerate);
+}
+
+base::ScopedFD GpuArcVideoEncodeAccelerator::UnwrapFdFromMojoHandle(
+    mojo::ScopedHandle handle) {
+  DCHECK(client_);
+  if (!handle.is_valid()) {
+    LOG(ERROR) << "handle is invalid.";
+    client_->NotifyError(Error::kInvalidArgumentError);
+    return base::ScopedFD();
+  }
+
+  base::PlatformFile platform_file;
+  MojoResult mojo_result =
+      mojo::UnwrapPlatformFile(std::move(handle), &platform_file);
+  if (mojo_result != MOJO_RESULT_OK) {
+    LOG(ERROR) << "UnwrapPlatformFile failed: " << mojo_result;
+    client_->NotifyError(Error::kPlatformFailureError);
+    return base::ScopedFD();
+  }
+
+  return base::ScopedFD(platform_file);
+}
+
+}  // namespace arc
+}  // namespace chromeos