Make VEA test support videos with different coded size and visible size

content/common/gpu/media/video_encode_accelerator_unittest.cc

 // Copyright 2013 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/at_exit.h"
 #include "base/bind.h"
 #include "base/command_line.h"
 #include "base/file_util.h"
 #include "base/files/memory_mapped_file.h"
 #include "base/memory/scoped_vector.h"
@@ skipping 15 matching lines @@
 #endif
 
 #if defined(OS_CHROMEOS) && defined(ARCH_CPU_ARMEL)
 #include "content/common/gpu/media/v4l2_video_encode_accelerator.h"
 #elif defined(OS_CHROMEOS) && defined(ARCH_CPU_X86_FAMILY) && defined(USE_X11)
 #include "content/common/gpu/media/vaapi_video_encode_accelerator.h"
 #else
 #error The VideoEncodeAcceleratorUnittest is not supported on this platform.
 #endif
 
+#define ALIGN_64_BYTES(x) (((x) + 63) & ~63)
+
 using media::VideoEncodeAccelerator;
 
 namespace content {
 namespace {
 
 const media::VideoFrame::Format kInputFormat = media::VideoFrame::I420;
 
 // Arbitrarily chosen to add some depth to the pipeline.
 const unsigned int kNumOutputBuffers = 4;
 const unsigned int kNumExtraInputFrames = 4;
@@ skipping 26 matching lines @@
 
 // The syntax of multiple test streams is:
 //  test-stream1;test-stream2;test-stream3
 // The syntax of each test stream is:
 // "in_filename:width:height:out_filename:requested_bitrate:requested_framerate
 //  :requested_subsequent_bitrate:requested_subsequent_framerate"
 // - |in_filename| must be an I420 (YUV planar) raw stream
 //   (see http://www.fourcc.org/yuv.php#IYUV).
 // - |width| and |height| are in pixels.
 // - |profile| to encode into (values of media::VideoCodecProfile).
+// - |temp_file| is used to prepare input buffers.

[wuchengli, 2014/08/29 10:03:11]

This should go to struct TestStream. Right?

[henryhsu, 2014/09/01 05:53:45]

Done.

 // - |out_filename| filename to save the encoded stream to (optional).
 //   Output stream is saved for the simple encode test only.
 // Further parameters are optional (need to provide preceding positional
 // parameters if a specific subsequent parameter is required):
 // - |requested_bitrate| requested bitrate in bits per second.
 // - |requested_framerate| requested initial framerate.
 // - |requested_subsequent_bitrate| bitrate to switch to in the middle of the
 //                                  stream.
 // - |requested_subsequent_framerate| framerate to switch to in the middle
 //                                    of the stream.
 //   Bitrate is only forced for tests that test bitrate.
 const char* g_default_in_filename = "bear_320x192_40frames.yuv";
 const char* g_default_in_parameters = ":320:192:1:out.h264:200000";
 base::FilePath::StringType* g_test_stream_data;
 
 struct TestStream {
   TestStream()
       : requested_bitrate(0),
         requested_framerate(0),
         requested_subsequent_bitrate(0),
         requested_subsequent_framerate(0) {}
   ~TestStream() {}
 
   gfx::Size size;
+  std::string in_filename;
   base::MemoryMappedFile input_file;

[wuchengli, 2014/08/29 10:03:11]

in_filename, input_file, and temp_file are similar. Add comments to each
variable. Explain |input_file| is the memory mapped of |temp_file|.

[henryhsu, 2014/09/01 05:53:45]

Done.

   media::VideoCodecProfile requested_profile;
+  base::FilePath temp_file;

[wuchengli, 2014/08/29 10:03:11]

Move this after |input_file|.

[henryhsu, 2014/09/01 05:53:45]

Done.

   std::string out_filename;

[wuchengli, 2014/08/29 10:03:11]

Move this before requested_profile. Better to keep all file-related variables
together.

[henryhsu, 2014/09/01 05:53:45]

Done.

   unsigned int requested_bitrate;
   unsigned int requested_framerate;
   unsigned int requested_subsequent_bitrate;
   unsigned int requested_subsequent_framerate;
 };
 
+// ARM performs CPU cache management with CPU cache line granularity. We thus
+// need to ensure our buffers are CPU cache line-aligned (64 byte-aligned).
+// Otherwise newer kernels will refuse to accept them, and on older kernels
+// we'll be treating ourselves to random corruption.
+// Since we are just mmapping and passing chunks of the input file, to ensure
+// alignment, if the starting virtual addresses of YUV planes of the frames
+// in it were not 64 byte-aligned, we'd have to prepare a memory with 64
+// byte-aligned starting address and make sure the addresses of YUV planes of
+// each frame are 64 byte-aligned before sending to the encoder.
+// Now we test resolutions different from coded size and prepare chunks before
+// encoding to avoid performance impact.
+// Use |visible_size| and |coded_size| to copy YUV data into memory from
+// |in_filename|. The copied result will be saved in |input_file|. Also
+// calculate the byte size of an input frame and set it to |coded_buffer_size|.
+// |temp_file| is used to prepare input buffers and will be deleted after test
+// finished.
+static void PrepareInputBuffers(const gfx::Size& visible_size,
+                                const gfx::Size& coded_size,
+                                const std::string in_filename,
+                                base::MemoryMappedFile* input_file,
+                                base::FilePath* temp_file,
+                                size_t* coded_buffer_size) {
+  size_t input_num_planes = media::VideoFrame::NumPlanes(kInputFormat);
+  std::vector<size_t> padding_sizes(input_num_planes);
+  std::vector<size_t> coded_bpl(input_num_planes);
+  std::vector<size_t> visible_bpl(input_num_planes);
+  std::vector<size_t> visible_plane_rows(input_num_planes);
+
+  // YUV plane starting address should be 64 bytes alignment. Calculate padding
+  // size for each plane, and frame allocation size for coded size. Also store
+  // bytes per line information of coded size and visible size.
+  *coded_buffer_size = 0;
+  for (off_t i = 0; i < input_num_planes; i++) {
+    size_t size =
+        media::VideoFrame::PlaneAllocationSize(kInputFormat, i, coded_size);
+    size_t padding_bytes = ALIGN_64_BYTES(size) - size;
+    *coded_buffer_size += ALIGN_64_BYTES(size);
+
+    coded_bpl[i] =
+        media::VideoFrame::RowBytes(i, kInputFormat, coded_size.width());
+    visible_bpl[i] =
+        media::VideoFrame::RowBytes(i, kInputFormat, visible_size.width());
+    visible_plane_rows[i] =
+        media::VideoFrame::Rows(i, kInputFormat, visible_size.height());
+    size_t padding_rows =
+        media::VideoFrame::Rows(i, kInputFormat, coded_size.height()) -
+        visible_plane_rows[i];
+    padding_sizes[i] = padding_rows * coded_bpl[i] + padding_bytes;
+  }
+
+  // Test case may have many encoders and memory should be prepared once.
+  if (input_file->IsValid())
+    return;
+
+  base::MemoryMappedFile src_file;
+  CHECK(base::CreateTemporaryFile(temp_file));
+  CHECK(src_file.Initialize(base::FilePath(in_filename)));
+
+  size_t visible_buffer_size =
+      media::VideoFrame::AllocationSize(kInputFormat, visible_size);
+  size_t num_frames = src_file.length() / visible_buffer_size;
+  uint32 flags = base::File::FLAG_CREATE_ALWAYS | base::File::FLAG_WRITE |
+                 base::File::FLAG_READ;
+
+  // Create a temporary file with coded_size length.
+  base::File file(base::FilePath(*temp_file), flags);

[wuchengli, 2014/08/29 10:03:11]

|temp_file| is a FilePath. Do we need base::FilePath constructor here?

[henryhsu, 2014/09/01 05:53:45]

Done.

+  file.SetLength(*coded_buffer_size * num_frames);
+  CHECK(input_file->Initialize(file.Pass(), true));
+
+  off_t src_offset = 0, dest_offset = 0;
+  while (src_offset < static_cast<off_t>(src_file.length())) {
+    for (off_t i = 0; i < input_num_planes; i++) {
+#if defined(ARCH_CPU_ARMEL)
+      // Assert that each plane of frame starts at 64-byte boundary.
+      const uint8* ptr = input_file->data() + dest_offset;
+      ASSERT_EQ(reinterpret_cast<off_t>(ptr) & 63, 0)
+          << "Planes of frame should be mapped at a 64 byte boundary";
+#endif
+      for (off_t j = 0; j < visible_plane_rows[i]; j++) {
+        const uint8* src = src_file.data() + src_offset;
+        uint8* dest = input_file->data() + dest_offset;
+        memcpy(dest, src, visible_bpl[i]);
+        src_offset += visible_bpl[i];
+        dest_offset += coded_bpl[i];
+      }
+      dest_offset += padding_sizes[i];
+    }
+  }
+}
+
 // Parse |data| into its constituent parts, set the various output fields
 // accordingly, read in video stream, and store them to |test_streams|.
 static void ParseAndReadTestStreamData(const base::FilePath::StringType& data,
                                        ScopedVector<TestStream>* test_streams) {
   // Split the string to individual test stream data.
   std::vector<base::FilePath::StringType> test_streams_data;
   base::SplitString(data, ';', &test_streams_data);
   CHECK_GE(test_streams_data.size(), 1U) << data;
 
   // Parse each test stream data and read the input file.
   for (size_t index = 0; index < test_streams_data.size(); ++index) {
     std::vector<base::FilePath::StringType> fields;
     base::SplitString(test_streams_data[index], ':', &fields);
     CHECK_GE(fields.size(), 4U) << data;
     CHECK_LE(fields.size(), 9U) << data;
     TestStream* test_stream = new TestStream();
 
-    base::FilePath::StringType filename = fields[0];
+    test_stream->in_filename = fields[0];
     int width, height;
     CHECK(base::StringToInt(fields[1], &width));
     CHECK(base::StringToInt(fields[2], &height));
     test_stream->size = gfx::Size(width, height);
     CHECK(!test_stream->size.IsEmpty());
     int profile;
     CHECK(base::StringToInt(fields[3], &profile));
     CHECK_GT(profile, media::VIDEO_CODEC_PROFILE_UNKNOWN);
     CHECK_LE(profile, media::VIDEO_CODEC_PROFILE_MAX);
     test_stream->requested_profile =
@@ skipping 11 matching lines @@
     if (fields.size() >= 8 && !fields[7].empty()) {
       CHECK(base::StringToUint(fields[7],
                                &test_stream->requested_subsequent_bitrate));
     }
 
     if (fields.size() >= 9 && !fields[8].empty()) {
       CHECK(base::StringToUint(fields[8],
                                &test_stream->requested_subsequent_framerate));
     }
 
-    CHECK(test_stream->input_file.Initialize(base::FilePath(filename)));
     test_streams->push_back(test_stream);
   }
 }
 
 // Set default parameters of |test_streams| and update the parameters according
 // to |mid_stream_bitrate_switch| and |mid_stream_framerate_switch|.
 static void UpdateTestStreamData(bool mid_stream_bitrate_switch,
                                  bool mid_stream_framerate_switch,
                                  ScopedVector<TestStream>* test_streams) {
   for (size_t i = 0; i < test_streams->size(); i++) {
@@ skipping 186 matching lines @@
     validator.reset(new VP8Validator(frame_cb));
   } else {
     LOG(FATAL) << "Unsupported profile: " << profile;
   }
 
   return validator.Pass();
 }
 
 class VEAClient : public VideoEncodeAccelerator::Client {
  public:
-  VEAClient(const TestStream& test_stream,
+  VEAClient(TestStream& test_stream,
             ClientStateNotification<ClientState>* note,
             bool save_to_file,
             unsigned int keyframe_period,
             bool force_bitrate,
             bool test_perf);
   virtual ~VEAClient();
   void CreateEncoder();
   void DestroyEncoder();
 
   // Return the number of encoded frames per second.
@@ skipping 41 matching lines @@
   // the performance test.
   void VerifyPerf();
 
   // Prepare and return a frame wrapping the data at |position| bytes in
   // the input stream, ready to be sent to encoder.
   scoped_refptr<media::VideoFrame> PrepareInputFrame(off_t position);
 
   ClientState state_;
   scoped_ptr<VideoEncodeAccelerator> encoder_;
 
-  const TestStream& test_stream_;
+  TestStream& test_stream_;
   // Used to notify another thread about the state. VEAClient does not own this.
   ClientStateNotification<ClientState>* note_;
 
   // Ids assigned to VideoFrames (start at 1 for easy comparison with
   // num_encoded_frames_).
   std::set<int32> inputs_at_client_;
   int32 next_input_id_;
 
   // Ids for output BitstreamBuffers.
   typedef std::map<int32, base::SharedMemory*> IdToSHM;
@@ skipping 56 matching lines @@
   // The time when the encoding started.
   base::TimeTicks encode_start_time_;
 
   // The time when the last encoded frame is ready.
   base::TimeTicks last_frame_ready_time_;
 
   // All methods of this class should be run on the same thread.
   base::ThreadChecker thread_checker_;
 };
 
-VEAClient::VEAClient(const TestStream& test_stream,
+VEAClient::VEAClient(TestStream& test_stream,
                      ClientStateNotification<ClientState>* note,
                      bool save_to_file,
                      unsigned int keyframe_period,
                      bool force_bitrate,
                      bool test_perf)
     : state_(CS_CREATED),
       test_stream_(test_stream),
       note_(note),
       next_input_id_(1),
       next_output_buffer_id_(0),
@@ skipping 24 matching lines @@
   CHECK(validator_.get());
 
   if (save_to_file_) {
     CHECK(!test_stream_.out_filename.empty());
     base::FilePath out_filename(test_stream_.out_filename);
     // This creates or truncates out_filename.
     // Without it, AppendToFile() will not work.
     EXPECT_EQ(0, base::WriteFile(out_filename, NULL, 0));
   }
 
-  input_buffer_size_ =
-      media::VideoFrame::AllocationSize(kInputFormat, test_stream.size);
-  CHECK_GT(input_buffer_size_, 0UL);
-
-  // Calculate the number of frames in the input stream by dividing its length
-  // in bytes by frame size in bytes.
-  CHECK_EQ(test_stream_.input_file.length() % input_buffer_size_, 0U)
-      << "Stream byte size is not a product of calculated frame byte size";
-  num_frames_in_stream_ = test_stream_.input_file.length() / input_buffer_size_;
-  CHECK_GT(num_frames_in_stream_, 0UL);
-  CHECK_LE(num_frames_in_stream_, kMaxFrameNum);
-
-  // We may need to loop over the stream more than once if more frames than
-  // provided is required for bitrate tests.
-  if (force_bitrate_ && num_frames_in_stream_ < kMinFramesForBitrateTests) {
-    DVLOG(1) << "Stream too short for bitrate test (" << num_frames_in_stream_
-             << " frames), will loop it to reach " << kMinFramesForBitrateTests
-             << " frames";
-    num_frames_to_encode_ = kMinFramesForBitrateTests;
-  } else {
-    num_frames_to_encode_ = num_frames_in_stream_;
-  }
-
   thread_checker_.DetachFromThread();
 }
 
 VEAClient::~VEAClient() { CHECK(!has_encoder()); }
 
 void VEAClient::CreateEncoder() {
   DCHECK(thread_checker_.CalledOnValidThread());
   CHECK(!has_encoder());
 
 #if defined(OS_CHROMEOS) && defined(ARCH_CPU_ARMEL)
@@ skipping 34 matching lines @@
   return num_encoded_frames_ / duration.InSecondsF();
 }
 
 void VEAClient::RequireBitstreamBuffers(unsigned int input_count,
                                         const gfx::Size& input_coded_size,
                                         size_t output_size) {
   DCHECK(thread_checker_.CalledOnValidThread());
   ASSERT_EQ(state_, CS_INITIALIZED);
   SetState(CS_ENCODING);
 
-  // TODO(posciak): For now we only support input streams that meet encoder
-  // size requirements exactly (i.e. coded size == visible size), so that we
-  // can simply mmap the stream file and feed the encoder directly with chunks
-  // of that, instead of memcpying from mmapped file into a separate set of
-  // input buffers that would meet the coded size and alignment requirements.
-  // If/when this is changed, the ARM-specific alignment check below should be
-  // redone as well.
+  PrepareInputBuffers(
+      test_stream_.size,
+      input_coded_size,
+      test_stream_.in_filename,
+      &test_stream_.input_file,
+      &test_stream_.temp_file,
+      &input_buffer_size_);
+  CHECK_GT(input_buffer_size_, 0UL);
+
+  // Calculate the number of frames in the input stream by dividing its length
+  // in bytes by frame size in bytes.
+  CHECK_EQ(test_stream_.input_file.length() % input_buffer_size_, 0U)
+      << "Stream byte size is not a product of calculated frame byte size";
+  num_frames_in_stream_ = test_stream_.input_file.length() / input_buffer_size_;
+  CHECK_GT(num_frames_in_stream_, 0UL);
+  CHECK_LE(num_frames_in_stream_, kMaxFrameNum);
+
+  // We may need to loop over the stream more than once if more frames than
+  // provided is required for bitrate tests.
+  if (force_bitrate_ && num_frames_in_stream_ < kMinFramesForBitrateTests) {
+    DVLOG(1) << "Stream too short for bitrate test (" << num_frames_in_stream_
+             << " frames), will loop it to reach " << kMinFramesForBitrateTests
+             << " frames";
+    num_frames_to_encode_ = kMinFramesForBitrateTests;
+  } else {
+    num_frames_to_encode_ = num_frames_in_stream_;
+  }
+
   input_coded_size_ = input_coded_size;
-  ASSERT_EQ(input_coded_size_, test_stream_.size);
-#if defined(ARCH_CPU_ARMEL)
-  // ARM performs CPU cache management with CPU cache line granularity. We thus
-  // need to ensure our buffers are CPU cache line-aligned (64 byte-aligned).
-  // Otherwise newer kernels will refuse to accept them, and on older kernels
-  // we'll be treating ourselves to random corruption.
-  // Since we are just mmapping and passing chunks of the input file, to ensure
-  // alignment, if the starting virtual addresses of the frames in it were not
-  // 64 byte-aligned, we'd have to use a separate set of input buffers and copy
-  // the frames into them before sending to the encoder. It would have been an
-  // overkill here though, because, for now at least, we only test resolutions
-  // that result in proper alignment, and it would have also interfered with
-  // performance testing. So just assert that the frame size is a multiple of
-  // 64 bytes. This ensures all frames start at 64-byte boundary, because
-  // MemoryMappedFile should be mmapp()ed at virtual page start as well.
-  ASSERT_EQ(input_buffer_size_ & 63, 0u)
-      << "Frame size has to be a multiple of 64 bytes";
-  ASSERT_EQ(reinterpret_cast<off_t>(test_stream_.input_file.data()) & 63, 0)
-      << "Mapped file should be mapped at a 64 byte boundary";
-#endif
-
   num_required_input_buffers_ = input_count;
   ASSERT_GT(num_required_input_buffers_, 0UL);
 
   output_buffer_size_ = output_size;
   ASSERT_GT(output_buffer_size_, 0UL);
 
   for (unsigned int i = 0; i < kNumOutputBuffers; ++i) {
     base::SharedMemory* shm = new base::SharedMemory();
     CHECK(shm->CreateAndMapAnonymous(output_buffer_size_));
     output_shms_.push_back(shm);
@@ skipping 65 matching lines @@
 void VEAClient::InputNoLongerNeededCallback(int32 input_id) {
   std::set<int32>::iterator it = inputs_at_client_.find(input_id);
   ASSERT_NE(it, inputs_at_client_.end());
   inputs_at_client_.erase(it);
   FeedEncoderWithInputs();
 }
 
 scoped_refptr<media::VideoFrame> VEAClient::PrepareInputFrame(off_t position) {
   CHECK_LE(position + input_buffer_size_, test_stream_.input_file.length());
 
-  uint8* frame_data =
+  uint8* frame_data_y =
       const_cast<uint8*>(test_stream_.input_file.data() + position);
+  uint8* frame_data_u =
+      frame_data_y + ALIGN_64_BYTES(media::VideoFrame::PlaneAllocationSize(
+                         kInputFormat, 0, input_coded_size_));
+  uint8* frame_data_v =
+      frame_data_u + ALIGN_64_BYTES(media::VideoFrame::PlaneAllocationSize(
+                         kInputFormat, 1, input_coded_size_));
 
   CHECK_GT(current_framerate_, 0U);
   scoped_refptr<media::VideoFrame> frame =
       media::VideoFrame::WrapExternalYuvData(
           kInputFormat,
           input_coded_size_,
           gfx::Rect(test_stream_.size),
           test_stream_.size,
           input_coded_size_.width(),
           input_coded_size_.width() / 2,
           input_coded_size_.width() / 2,
-          frame_data,
-          frame_data + input_coded_size_.GetArea(),
-          frame_data + (input_coded_size_.GetArea() * 5 / 4),
+          frame_data_y,
+          frame_data_u,
+          frame_data_v,
           base::TimeDelta().FromMilliseconds(
               next_input_id_ * base::Time::kMillisecondsPerSecond /
               current_framerate_),
           media::BindToCurrentLoop(
               base::Bind(&VEAClient::InputNoLongerNeededCallback,
                          base::Unretained(this),
                          next_input_id_)));
 
   CHECK(inputs_at_client_.insert(next_input_id_).second);
   ++next_input_id_;
@@ skipping 194 matching lines @@
   for (size_t state_no = 0; state_no < arraysize(state_transitions); ++state_no)
     for (size_t i = 0; i < num_concurrent_encoders; i++)
       ASSERT_EQ(notes[i]->Wait(), state_transitions[state_no]);
 
   for (size_t i = 0; i < num_concurrent_encoders; ++i) {
     encoder_thread.message_loop()->PostTask(
         FROM_HERE,
         base::Bind(&VEAClient::DestroyEncoder, base::Unretained(clients[i])));
   }
 
+  for (size_t i = 0; i < test_streams.size(); i++)

[wuchengli, 2014/08/29 10:03:11]

We should close the memory mapped files before deleting them.
MemoryMappedFile::CloseHandles.

[henryhsu, 2014/09/01 05:53:45]

Done.

+    base::DeleteFile(test_streams[i]->temp_file, false);
+
   // This ensures all tasks have finished.
   encoder_thread.Stop();
 }
 
 INSTANTIATE_TEST_CASE_P(
     SimpleEncode,
     VideoEncodeAcceleratorTest,
     ::testing::Values(MakeTuple(1, true, 0, false, false, false, false)));
 
 INSTANTIATE_TEST_CASE_P(
@@ skipping 68 matching lines @@
       test_stream_data->assign(it->second.c_str());
       continue;
     }
     if (it->first == "v" || it->first == "vmodule")
       continue;
     LOG(FATAL) << "Unexpected switch: " << it->first << ":" << it->second;
   }
 
   return RUN_ALL_TESTS();
 }