Add ArcGpuVideoDecodeAccelerator.

chrome/gpu/arc_gpu_video_decode_accelerator.cc

+// Copyright 2016 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 "base/callback_helpers.h"
+#include "base/logging.h"
+#include "base/run_loop.h"
+#include "chrome/gpu/arc_gpu_video_decode_accelerator.h"
+
+// TODO use Pawel's factory instead
+#include "content/public/common/create_vda.h"
+
+#undef DVLOG
+#define DVLOG VLOG
+
+namespace chromeos {
+namespace arc {
+
+ArcGpuVideoDecodeAccelerator::InputRecord::InputRecord(
+    int32_t bitstream_buffer_id,
+    uint32_t buffer_index,
+    int64_t timestamp)
+    : bitstream_buffer_id(bitstream_buffer_id),
+      buffer_index(buffer_index),
+      timestamp(timestamp) {}
+
+ArcGpuVideoDecodeAccelerator::InputBufferInfo::InputBufferInfo()
+    : offset(0), length(0) {}
+
+ArcGpuVideoDecodeAccelerator::InputBufferInfo::InputBufferInfo(
+    InputBufferInfo&& a) {
+  handle = std::move(a.handle);
+  offset = a.offset;
+  length = a.length;
+}
+
+ArcGpuVideoDecodeAccelerator::InputBufferInfo::~InputBufferInfo() {}
+
+ArcGpuVideoDecodeAccelerator::PortInfo::PortInfo()
+    : memory_type(MEMORY_DMABUF), num_buffers(0) {}
+
+ArcGpuVideoDecodeAccelerator::ArcGpuVideoDecodeAccelerator(
+    const scoped_refptr<base::SingleThreadTaskRunner>& io_task_runner)
+    : pending_eos_output_buffer_(false),
+      arc_client_(nullptr),
+      next_bitstream_buffer_id_(0),
+      io_task_runner_(io_task_runner),
+      output_buffer_size_(0) {}
+
+ArcGpuVideoDecodeAccelerator::~ArcGpuVideoDecodeAccelerator() {}
+
+bool ArcGpuVideoDecodeAccelerator::Initialize(
+    DeviceType device_type,
+    ArcVideoAccelerator::Client* client) {
+  DVLOG(5) << "Initialize()";
+  DCHECK(thread_checker_.CalledOnValidThread());
+  if (device_type != DEVICE_DECODER)
+    return false;
+  DCHECK(client);
+  DCHECK(!arc_client_);
+  arc_client_ = client;
+
+  vda_.reset(content::CreateVDA(AsWeakPtr(), io_task_runner_));
+  if (!vda_) {
+    DVLOG(1) << "Failed to create VDA.";
+    return false;
+  }
+  return true;
+}
+
+bool ArcGpuVideoDecodeAccelerator::SetBufferCount(PortType port,
+                                                  size_t* count) {

[Pawel Osciak, 2016/02/29 08:25:20]

We are not assigning to count, do we need it to be an in/out value? I believe
this is only used as a final count and not negotiable anymore? We get min.
number of buffers in ProvidePictureBuffers, send it via OnOutpufFormatChanged,
and then receive final count via SetBufferCount?

[Owen Lin, 2016/03/03 06:30:42]

That's because VDA doesn't returns false even on failure on
vda->AssignPictureBuffers. I don't think we can guarantee to support any number
of buffers? mmm.... maybe we can. since the buffer is not allocated from device.
In that case, sure let's make count a input argument only.

+  DVLOG(5) << "SetBufferCount(port=" << port << ", count=" << *count << ")";
+  DCHECK(thread_checker_.CalledOnValidThread());
+  if (!ValidatePort(port))
+    return false;
+  if (vda_ == nullptr) {
+    DVLOG(1) << "Must call setBufferFormat before setBufferCount";

[Pawel Osciak, 2016/02/29 08:25:19]

VDA is now created in Initialize()... But ideally we could possibly remove the
need for SetBufferFormat() (please see other comments).

[Owen Lin, 2016/03/03 06:30:42]

Done.

+    return false;
+  }
+  if (port == PORT_OUTPUT) {

[Pawel Osciak, 2016/02/29 08:25:19]

We should probably check we don't have any existing buffers already and refuse
this otherwise.

[Owen Lin, 2016/03/03 06:30:43]

I think VDA should handle this case?

+    std::vector<media::PictureBuffer> buffers;
+    for (int32_t id = 0, n = *count; id < n; ++id) {
+      // TODO: Make sure the size is what we want.
+      buffers.push_back(media::PictureBuffer(id, coded_size_, 0));
+    }
+    vda_->AssignPictureBuffers(buffers);
+  }
+  PortInfo* port_info = &port_info_[port];
+  port_info->num_buffers = *count;
+  if (port == PORT_INPUT) {

[Pawel Osciak, 2016/02/29 08:25:19]

Is it ok to allow SetBufferCount at any time for inputs? For example, if we are
still decoding, but client calls SetBufferCount for INPUT, we drop the inputs,
but the VDA still has them and may return them to us. So we could signal we are
done for buffers that were dropped. Perhaps better not to allow SetBufferCount
at all if we already have buffers?

[Owen Lin, 2016/03/03 06:30:43]

Will move this to be part of initialization.

+    input_buffer_info_.clear();
+    input_buffer_info_.resize(port_info->num_buffers);
+  }
+  return true;
+}
+
+bool ArcGpuVideoDecodeAccelerator::SetBufferFormat(PortType port,

[Pawel Osciak, 2016/02/29 08:25:20]

IIRC, we could not change format and memory type after initial setup. If so,
perhaps we should pass formats to AVA::Initialize() and remove this method?

[Owen Lin, 2016/03/03 06:30:43]

I agree. But this will involve changes of different projects in both arc and
chromium side. I would prefer to do that in a separate CL.

+                                                   const BufferFormat& format) {
+  DVLOG(5) << "SetBufferFormat(port=" << port
+           << ", format=(memory_type=" << format.memory_type
+           << ", pixel_format=" << format.pixel_format << "))";
+  DCHECK(thread_checker_.CalledOnValidThread());
+  if (!ValidatePort(port))
+    return false;
+  switch (port) {
+    case PORT_INPUT: {
+      PortInfo* port_info = &port_info_[port];
+      if (format.memory_type != MEMORY_SHARED_MEMORY) {
+        DVLOG(1) << "Only SharedMemory is supported for input buffers";
+        return false;
+      }
+      port_info->memory_type = format.memory_type;
+      media::VideoDecodeAccelerator::Config config;
+      switch (format.pixel_format) {
+        case HAL_PIXEL_FORMAT_H264:
+          config.profile = media::H264PROFILE_MAIN;

[Pawel Osciak, 2016/02/29 08:25:20]

On second thought, could we pass the exact profile and use it here? A VDA may
support a subset of profiles for a given codec and the VDA interface already
handles it.

[Owen Lin, 2016/03/03 06:30:42]

I don't think we know the exact profile. MediaCodecs just tell us the mime type.

+          break;
+        case HAL_PIXEL_FORMAT_VP8:
+          config.profile = media::VP8PROFILE_ANY;
+          break;
+        default:
+          DVLOG(1) << "Unsupported input format: " << format.pixel_format;
+          return false;
+      }
+      config.output_mode =
+          media::VideoDecodeAccelerator::Config::OutputMode::IMPORT;
+      if (!vda_->Initialize(config, this)) {

[Pawel Osciak, 2016/02/29 08:25:20]

We should probably make sure vda_ is not null, otherwise the client may crash
us.

[Owen Lin, 2016/03/03 06:30:42]

Done.

+        DVLOG(1) << "VDA::Initialize() failed.";
+        return false;
+      }
+      break;
+    }
+    case PORT_OUTPUT: {
+      PortInfo* port_info = &port_info_[port];
+      if (format.memory_type != MEMORY_DMABUF) {
+        DVLOG(1) << "Only DMA buffer is supported for output buffers";

[Pawel Osciak, 2016/02/29 08:25:19]

s/DMA buffer/dmabuf/

[Owen Lin, 2016/03/03 06:30:42]

Done.

+        return false;
+      }
+      port_info->memory_type = format.memory_type;
+      break;
+    }
+    default:
+      NOTREACHED() << "Invalid port: " << port;
+      return false;
+  }
+  return true;
+}
+
+bool ArcGpuVideoDecodeAccelerator::ValidatePort(PortType port) {
+  if (port != PORT_INPUT && port != PORT_OUTPUT) {
+    DVLOG(1) << "Invalid port: " << port;
+    return false;
+  }
+  return true;
+}
+
+bool ArcGpuVideoDecodeAccelerator::ValidatePortAndIndex(PortType port,
+                                                        uint32_t index) {
+  if (!ValidatePort(port))
+    return false;
+  if (index >= port_info_[port].num_buffers) {
+    DVLOG(1) << "Invalid buffer - port: " << port << ", index: " << index;
+    return false;
+  }
+  return true;
+}
+
+bool ArcGpuVideoDecodeAccelerator::BindSharedMemory(PortType port,
+                                                    uint32_t index,
+                                                    int ashmem_fd,
+                                                    size_t offset,

[Pawel Osciak, 2016/02/29 08:25:20]

off_t ?

[Owen Lin, 2016/03/03 06:30:42]

Done.

+                                                    size_t length) {
+  DVLOG(5) << "ArcGVDA::BindSharedMemory, offset: " << offset
+           << ", length: " << length;
+  DCHECK(thread_checker_.CalledOnValidThread());
+
+  // Make sure we will close the file descriptor.
+  base::ScopedFD handle(ashmem_fd);
+  if (!ValidatePortAndIndex(port, index))
+    return false;
+  if (port != PORT_INPUT) {

[Pawel Osciak, 2016/02/29 08:25:20]

Should we instead check against port_info->memory_type for given port?

[Owen Lin, 2016/03/03 06:30:43]

port_info is just removed.

+    DVLOG(1) << "SharedBuffer is only supported for input";
+    return false;
+  }
+  InputBufferInfo* input_info = &input_buffer_info_[index];
+  input_info->handle = std::move(handle);
+  input_info->offset = offset;
+  input_info->length = length;
+  return true;
+}
+
+bool ArcGpuVideoDecodeAccelerator::BindDmabuf(PortType port,
+                                              uint32_t index,
+                                              int dmabuf_fd) {
+  DCHECK(thread_checker_.CalledOnValidThread());
+  base::ScopedFD handle(dmabuf_fd);
+
+  // Make sure we will close the file descriptor.
+  if (!ValidatePortAndIndex(port, index))
+    return false;
+  if (port != PORT_OUTPUT) {

[Pawel Osciak, 2016/02/29 08:25:20]

Here as well, should we instead check against port_info->memory_type for given
port?

[Owen Lin, 2016/03/03 06:30:42]

same as above.

+    DVLOG(1) << "GraphicBuffer is only supported for input";
+    return false;
+  }
+  std::vector<base::ScopedFD> dmabuf_fds;
+  dmabuf_fds.push_back(std::move(handle));
+  vda_->ImportBufferForPicture(index, std::move(dmabuf_fds));
+  return true;
+}
+
+void ArcGpuVideoDecodeAccelerator::UseBuffer(PortType port,
+                                             uint32_t index,
+                                             const BufferMetadata& metadata) {
+  DVLOG(5) << "UseBuffer(port=" << port << ", index=" << index
+           << ", metadata=(bytes_used=" << metadata.bytes_used
+           << ", timestamp=" << metadata.timestamp << "))";
+  DCHECK(thread_checker_.CalledOnValidThread());
+  if (!ValidatePortAndIndex(port, index)) {
+    arc_client_->OnError(INVALID_ARGUMENT);
+    return;
+  }
+  switch (port) {
+    case PORT_INPUT: {
+      InputBufferInfo* input_info = &input_buffer_info_[index];
+      if (metadata.flags & BUFFER_FLAG_EOS) {
+        // TODO(owenlin): Ask VDA to return all output pictures so that we
+        // can output an EOS picture when Flush() is done.
+        vda_->Flush();
+      }
+      if (metadata.bytes_used == 0) {

[Pawel Osciak, 2016/02/29 08:25:19]

Could we handle this case in VDAs? We already do I believe.

[Owen Lin, 2016/03/03 06:30:42]

Done.

+        arc_client_->OnBufferDone(PORT_INPUT, index, BufferMetadata());
+        return;
+      }
+      if (metadata.bytes_used > input_info->length) {

[Pawel Osciak, 2016/02/29 08:25:19]

I think this also should be validated in the VDA, as it has to do it anyway?

[Owen Lin, 2016/03/03 06:30:42]

Yes, you're right.

+        DVLOG(1) << "Invalid bytes_used: " << metadata.bytes_used << " > "
+                 << input_info->length;
+        arc_client_->OnError(INVALID_ARGUMENT);
+        return;
+      }
+      int32_t bitstream_buffer_id = next_bitstream_buffer_id_;
+      // Mask against 30 bits, to avoid (undefined wraparound on signed integer)
+      next_bitstream_buffer_id_ = (next_bitstream_buffer_id_ + 1) & 0x3FFFFFFF;
+      SetInputRecord(bitstream_buffer_id, index, metadata.timestamp);
+      int dup_fd = HANDLE_EINTR(dup(input_info->handle.get()));
+      if (dup_fd < 0) {
+        DVLOG(1) << "dup() failed.";
+        arc_client_->OnError(PLATFORM_FAILURE);
+        return;
+      }
+      vda_->Decode(media::BitstreamBuffer(
+          bitstream_buffer_id, base::SharedMemoryHandle(dup_fd, true),
+          metadata.bytes_used, input_info->offset));
+      break;
+    }
+    case PORT_OUTPUT: {
+      SendEosIfNeededOrReusePicture(index);
+      break;
+    }
+    default:
+      NOTREACHED();
+  }
+}
+
+void ArcGpuVideoDecodeAccelerator::Reset() {
+  DCHECK(thread_checker_.CalledOnValidThread());
+  base::RunLoop loop;
+  reset_done_callback_ = loop.QuitClosure();
+  vda_->Reset();
+  base::MessageLoop::ScopedNestableTaskAllower allow(
+      base::MessageLoop::current());
+  // Wait for the ResetDone callback.
+  loop.Run();
+}
+
+void ArcGpuVideoDecodeAccelerator::ProvidePictureBuffers(
+    uint32_t requested_num_of_buffers,
+    const gfx::Size& dimensions,
+    uint32_t texture_target) {
+  DVLOG(5) << "ProvidePictureBuffers("
+           << "requested_num_of_buffers=" << requested_num_of_buffers
+           << ", dimensions=" << dimensions.ToString() << ")";
+  DCHECK(thread_checker_.CalledOnValidThread());
+  coded_size_ = dimensions;
+  // TODO(owenlin): use VDA::GetOutputFormat() here and calculate correct
+  //                |image_size|.
+  VideoFormat video_format;
+  video_format.image_size = dimensions.GetArea() * 3 / 2;
+  output_buffer_size_ = video_format.image_size;
+  video_format.min_num_buffers = requested_num_of_buffers;
+  video_format.coded_width = dimensions.width();
+  video_format.coded_height = dimensions.height();
+  // TODO(owenlin): How to get visible size?

[Pawel Osciak, 2016/02/29 08:25:20]

Do we need it at this time for the client?

[Owen Lin, 2016/03/03 06:30:42]

Maybe not, I think client needs the visible size before getting the first
decoded frame. However, I would prefer to address this issue later. 

+  video_format.crop_top = 0;
+  video_format.crop_left = 0;
+  video_format.crop_width = dimensions.width();
+  video_format.crop_height = dimensions.height();
+  arc_client_->OnOutputFormatChanged(video_format);
+}
+
+void ArcGpuVideoDecodeAccelerator::DismissPictureBuffer(
+    int32_t picture_buffer) {
+  // no-op

[Pawel Osciak, 2016/02/29 08:25:20]

I'm wondering if we should still clean up here, i.e. close the fd, mark the
buffer as dismissed and return it to the client. Next time we get it from the
client we could just ignore it.

[Owen Lin, 2016/03/03 06:30:42]

There should not be any fd owned by us and need to be closed here.

+}
+
+void ArcGpuVideoDecodeAccelerator::PictureReady(const media::Picture& picture) {
+  DVLOG(5) << "PictureReady(picture_buffer_id=" << picture.picture_buffer_id()
+           << ", bitstream_buffer_id=" << picture.bitstream_buffer_id();
+  DCHECK(thread_checker_.CalledOnValidThread());
+  if (!ValidatePortAndIndex(PORT_OUTPUT, picture.picture_buffer_id())) {

[Pawel Osciak, 2016/02/29 08:25:20]

Do we need to validate PR(), since it's coming from the VDA?

[Owen Lin, 2016/03/03 06:30:43]

Done.

+    DVLOG(1) << "Invalid index: " << picture.picture_buffer_id();
+    arc_client_->OnError(PLATFORM_FAILURE);
+    return;
+  }
+
+  // Empty buffer, returned in Flushing.
+  if (picture.bitstream_buffer_id() == -1) {
+    buffers_pending_eos_.push(picture.picture_buffer_id());
+    return;
+  }
+  InputRecord* input_record = FindInputRecord(picture.bitstream_buffer_id());
+  if (input_record == nullptr) {
+    DVLOG(1) << "Cannot find for bitstream buffer id: "
+             << picture.bitstream_buffer_id();
+    arc_client_->OnError(PLATFORM_FAILURE);
+    return;
+  }
+
+  BufferMetadata metadata;
+  metadata.timestamp = input_record->timestamp;
+  metadata.bytes_used = output_buffer_size_;
+  arc_client_->OnBufferDone(PORT_OUTPUT, picture.picture_buffer_id(), metadata);
+}
+
+void ArcGpuVideoDecodeAccelerator::NotifyEndOfBitstreamBuffer(
+    int32_t bitstream_buffer_id) {
+  DVLOG(5) << "NotifyEndOfBitstreamBuffer(" << bitstream_buffer_id << ")";
+  DCHECK(thread_checker_.CalledOnValidThread());
+  InputRecord* input_record = FindInputRecord(bitstream_buffer_id);
+  if (input_record == nullptr) {
+    arc_client_->OnError(PLATFORM_FAILURE);
+    return;
+  }
+  arc_client_->OnBufferDone(PORT_INPUT, input_record->buffer_index,
+                            BufferMetadata());
+}
+
+void ArcGpuVideoDecodeAccelerator::NotifyFlushDone() {
+  DCHECK(thread_checker_.CalledOnValidThread());
+  pending_eos_output_buffer_ = true;
+  while (!buffers_pending_eos_.empty()) {
+    SendEosIfNeededOrReusePicture(buffers_pending_eos_.front());
+    buffers_pending_eos_.pop();
+  }
+}
+
+void ArcGpuVideoDecodeAccelerator::NotifyResetDone() {
+  DCHECK(thread_checker_.CalledOnValidThread());
+  base::ResetAndReturn(&reset_done_callback_).Run();
+}
+
+static ArcVideoAccelerator::Error ConvertErrorCode(
+    media::VideoDecodeAccelerator::Error error) {
+  switch (error) {
+    case media::VideoDecodeAccelerator::ILLEGAL_STATE:
+      return ArcVideoAccelerator::ILLEGAL_STATE;
+    case media::VideoDecodeAccelerator::INVALID_ARGUMENT:
+      return ArcVideoAccelerator::INVALID_ARGUMENT;
+    case media::VideoDecodeAccelerator::UNREADABLE_INPUT:
+      return ArcVideoAccelerator::UNREADABLE_INPUT;
+    case media::VideoDecodeAccelerator::PLATFORM_FAILURE:
+      return ArcVideoAccelerator::PLATFORM_FAILURE;
+    default:
+      DVLOG(1) << "Unknown error: " << error;
+      return ArcVideoAccelerator::PLATFORM_FAILURE;
+  }
+}
+
+void ArcGpuVideoDecodeAccelerator::NotifyError(
+    media::VideoDecodeAccelerator::Error error) {
+  DCHECK(thread_checker_.CalledOnValidThread());
+  arc_client_->OnError(ConvertErrorCode(error));
+}
+
+void ArcGpuVideoDecodeAccelerator::SendEosIfNeededOrReusePicture(
+    uint32_t index) {
+  if (pending_eos_output_buffer_) {
+    BufferMetadata metadata;
+    metadata.flags = BUFFER_FLAG_EOS;
+    arc_client_->OnBufferDone(PORT_OUTPUT, index, metadata);
+    pending_eos_output_buffer_ = false;
+  } else {
+    vda_->ReusePictureBuffer(index);
+  }
+}
+
+void ArcGpuVideoDecodeAccelerator::SetInputRecord(int32_t bitstream_buffer_id,

[Pawel Osciak, 2016/02/29 08:25:20]

Since we can't have more inputs than buffer indices, could we move the
InputRecord data into InputBufferInfo directly, and remove the InputRecord class
entirely?

[Owen Lin, 2016/03/03 06:30:42]

InputRecord could be more than InputBuffers. Assume there are only 2 input
buffers, in the case of h264, the frames will be reordered so will not returned
so quickly.

+                                                  uint32_t buffer_index,
+                                                  int64_t timestamp) {
+  input_records_.push_front(
+      InputRecord(bitstream_buffer_id, buffer_index, timestamp));
+  // The same value copied from media::GpuVideoDecoder.
+  const size_t kMaxNumberOfInputRecords = 128;
+  if (input_records_.size() > kMaxNumberOfInputRecords)
+    input_records_.pop_back();
+}
+
+ArcGpuVideoDecodeAccelerator::InputRecord*
+ArcGpuVideoDecodeAccelerator::FindInputRecord(int32_t bitstream_buffer_id) {
+  for (auto& record : input_records_) {
+    if (record.bitstream_buffer_id == bitstream_buffer_id)
+      return &record;
+  }
+  return nullptr;
+}
+
+}  // namespace arc
+}  // namespace chromeos