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

content/common/gpu/media/video_encode_accelerator_unittest.cc

Index: content/common/gpu/media/video_encode_accelerator_unittest.cc
diff --git a/content/common/gpu/media/video_encode_accelerator_unittest.cc b/content/common/gpu/media/video_encode_accelerator_unittest.cc
index 9d1e3b69dc7c9edccf56b2fab6021bc0fb20f23b..7830c84efee89e6dceb03dfa8b4e8887548f43ac 100644
--- a/content/common/gpu/media/video_encode_accelerator_unittest.cc
+++ b/content/common/gpu/media/video_encode_accelerator_unittest.cc
@@ -33,6 +33,8 @@
 #error The VideoEncodeAcceleratorUnittest is not supported on this platform.
 #endif
 
+#define ALIGN_64_BYTES(x) (((x) + 63) & ~63)
+
 using media::VideoEncodeAccelerator;
 
 namespace content {
@@ -103,6 +105,7 @@ struct TestStream {
   ~TestStream() {}
 
   gfx::Size size;
+  std::string in_filename;
   base::MemoryMappedFile input_file;
   media::VideoCodecProfile requested_profile;
   std::string out_filename;
@@ -112,6 +115,105 @@ struct TestStream {
   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 use a separate set of input

[wuchengli, 2014/08/28 08:33:00]

Please remove "use a separate set of input buffers and copy the frames". Just
explain how you achieve the alignment now.

[henryhsu, 2014/08/28 09:36:45]

Done.

+// buffers and copy the frames into them before sending to the encoder.
+// Now we test resolutions different from coded size and prepare chunks before
+// testing to avoid performance impact.

[wuchengli, 2014/08/28 08:33:01]

s/testing/encoding/

[henryhsu, 2014/08/28 09:36:44]

Done.

+// Use |coded_size| to copy YUV data into memory from |test_stream| input file.
+// The copyied result will be saved in test_stream->input_file.

[wuchengli, 2014/08/28 08:33:01]

s/copyied/copied/

[henryhsu, 2014/08/28 09:36:45]

Done.

+// Also calculate the byte size of an input frame and set it to
+// |input_buffer_size|.
+static void PrepareInputMemory(const gfx::Size& coded_size,
+                               TestStream* test_stream,

[wuchengli, 2014/08/28 08:33:01]

It is clearer to pass test_stream->size, test_stream->input_file, and
test_stream->in_filename to the function parameters.

[henryhsu, 2014/08/28 09:36:44]

Done.

+                               size_t* input_buffer_size) {
+  base::FilePath temp_file;
+  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);
+  size_t visible_frame_size =

[wuchengli, 2014/08/28 08:33:01]

I think it's easier to understand to s/visible_frame_size/visible_buffer_size/
and s/input_buffer_size/coded_buffer_size/.

[henryhsu, 2014/08/28 09:36:45]

Done.

+      media::VideoFrame::AllocationSize(kInputFormat, test_stream->size);
+
+  // YUV plane starting address should be 64 bytes alignment.
+  // Calculate padding size for each plane, and frame size for visible size

[wuchengli, 2014/08/28 08:33:00]

s/frame size/frame allocation size/. Remove visible size because it's not
calculated here. Update the comment. This loop calculates other stuff, too.

[henryhsu, 2014/08/28 09:36:44]

Done.

+  // and coded size.
+  *input_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_bpl[i] =
+        media::VideoFrame::RowBytes(i, kInputFormat, coded_size.width());
+    visible_bpl[i] =
+        media::VideoFrame::RowBytes(i, kInputFormat, test_stream->size.width());
+    visible_plane_rows[i] =
+        media::VideoFrame::Rows(i, kInputFormat, test_stream->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;
+    *input_buffer_size += ALIGN_64_BYTES(size);
+  }
+
+  // Test case may have many encoders and memory should be prepared once.
+  if (test_stream->input_file.IsValid())
+    return;
+
+  base::MemoryMappedFile input_file;
+  CHECK(base::CreateTemporaryFile(&temp_file));
+  CHECK(input_file.Initialize(base::FilePath(test_stream->in_filename)));
+
+  size_t num_frames = input_file.length() / visible_frame_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);
+  file.Write(*input_buffer_size * num_frames - 1, ".", 1);
+  CHECK(test_stream->input_file.Initialize(file.Pass()));
+
+#if defined(ARCH_CPU_ARMEL)
+  // Assert that the frame size is a multiple of 64 bytes. This ensures all
+  // frames start at 64-byte boundary, because MemoryMappedFile should be
+  // mmap()ed at virtual page start as well.
+  ASSERT_EQ(*input_buffer_size & 63, 0u)

[wuchengli, 2014/08/28 08:33:01]

This can be removed. The above for loop ensured it.

[henryhsu, 2014/08/28 09:36:44]

Done.

+      << "Frame size has to be a multiple of 64 bytes";
+  ASSERT_EQ(reinterpret_cast<off_t>(test_stream->input_file.data()) & 63, 0)

[wuchengli, 2014/08/28 08:33:01]

This can be removed. The first time of ASSERT_EQ below ensures it.

[henryhsu, 2014/08/28 09:36:44]

Done.

+      << "Mapped file should be mapped at a 64 byte boundary";
+#endif
+
+  off_t src_offset = 0, dest_offset = 0;
+  while (src_offset < static_cast<off_t>(input_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.
+      uint8* ptr =
+          const_cast<uint8*>(test_stream->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++) {
+        uint8* src = const_cast<uint8*>(input_file.data() + src_offset);

[wuchengli, 2014/08/28 08:33:01]

nit: |src| can be const and the casting can be removed.

[henryhsu, 2014/08/28 09:36:45]

Done.

+        uint8* dest =
+            const_cast<uint8*>(test_stream->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];
+    }
+  }
+  base::DeleteFile(temp_file, false);

[wuchengli, 2014/08/28 08:33:00]

Can we delete the file and still access it from test_stream->input_file?

[henryhsu, 2014/08/28 09:36:44]

Yes, I tried this and encoded results are correct.

[wuchengli, 2014/08/28 10:17:48]

This doesn't make sense. Why we can access the mapped file after it is removed?

[henryhsu, 2014/08/29 06:36:40]

Not sure. But for safety, we can rollback to original way to delete file after
test finished.

+}
+
 // 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,
@@ -129,7 +231,7 @@ static void ParseAndReadTestStreamData(const base::FilePath::StringType& 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));
@@ -161,7 +263,6 @@ static void ParseAndReadTestStreamData(const base::FilePath::StringType& data,
                                &test_stream->requested_subsequent_framerate));
     }
 
-    CHECK(test_stream->input_file.Initialize(base::FilePath(filename)));
     test_streams->push_back(test_stream);
   }
 }
@@ -552,29 +653,6 @@ VEAClient::VEAClient(const TestStream& test_stream,
     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();
 }
 
@@ -629,35 +707,31 @@ void VEAClient::RequireBitstreamBuffers(unsigned int input_count,
   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.
-  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
+  PrepareInputMemory(input_coded_size,

[wuchengli, 2014/08/28 08:33:01]

s/PrepareInputMemory/PrepareInputBuffers/ to be consistent with
|input_buffer_size_|.

[henryhsu, 2014/08/28 09:36:44]

Done.

+                     const_cast<TestStream*>(&test_stream_),
+                     &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;
   num_required_input_buffers_ = input_count;
   ASSERT_GT(num_required_input_buffers_, 0UL);
 
@@ -743,8 +817,14 @@ void VEAClient::InputNoLongerNeededCallback(int32 input_id) {
 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 =
@@ -756,9 +836,9 @@ scoped_refptr<media::VideoFrame> VEAClient::PrepareInputFrame(off_t position) {
           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_),