Mac AEC: Fix two errors in upmixing.

The code was skipping the first sample and calculated output sample
positions slightly wrong. These two issues lead to discontinuities in
the audio.

Bug: chromium:776327
Cq-Include-Trybots: luci.chromium.try:android_optional_gpu_tests_rel;luci.chromium.try:linux_optional_gpu_tests_rel;luci.chromium.try:mac_optional_gpu_tests_rel;master.tryserver.chromium.win:win_optional_gpu_tests_rel
Change-Id: Ia39071a1f5f14082f5fcd1cb6032cf45ea5a244a
Reviewed-on: https://chromium-review.googlesource.com/983492
Commit-Queue: Oskar Sundbom <ossu@chromium.org>
Reviewed-by: Henrik Grunell <grunell@chromium.org>
Reviewed-by: Max Morin <maxmorin@chromium.org>
Cr-Commit-Position: refs/heads/master@{#546476}

media/audio/mac/audio_low_latency_input_mac.cc

@@ -36,22 +36,6 @@ OSStatus AudioDeviceDuck(AudioDeviceID inDevice,
                          Float32 inRampDuration) __attribute__((weak_import));
 }
 
-void UpmixMonoToStereoInPlace(AudioBuffer* audio_buffer, int bytes_per_sample) {
-  constexpr int channels = 2;
-  const int total_bytes = audio_buffer->mDataByteSize;
-  const int frames = total_bytes / bytes_per_sample / channels;
-  char* byte_ptr = reinterpret_cast<char*>(audio_buffer->mData);
-  for (int i = frames - 1; i > 0; --i) {
-    int in_offset = (bytes_per_sample * i);
-    int out_offset = (channels * bytes_per_sample * i);
-    for (int b = 0; b != bytes_per_sample; ++b) {
-      const char byte = byte_ptr[in_offset + b];
-      byte_ptr[out_offset + bytes_per_sample + b] = byte;
-      byte_ptr[out_offset + bytes_per_sample * 2 + b] = byte;
-    }
-  }
-}
-
 }  // namespace
 
 namespace media {
@@ -1368,4 +1352,41 @@ void AUAudioInputStream::ReportAndResetStats() {
   largest_glitch_frames_ = 0;
 }
 
+// TODO(ossu): Ideally, we'd just use the mono stream directly. However, since
+// mono or stereo (may) depend on if we want to run the echo canceller, and
+// since we can't provide two sets of AudioParameters for a device, this is the
+// best we can do right now.
+//
+// The algorithm works by copying a sample at offset N to 2*N and 2*N + 1, e.g.:
+//  ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ----
+// | a1 | a2 | a3 | b1 | b2 | b3 | c1 | c2 | c3 | -- | -- | -- | -- | -- | ...
+//  ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ----
+//  into
+//  ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ----
+// | a1 | a2 | a3 | a1 | a2 | a3 | b1 | b2 | b3 | b1 | b2 | b3 | c1 | c2 | ...
+//  ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ----
+//
+// To support various different sample sizes, this is done byte-by-byte. Only
+// the first half of the buffer will be used as input. It is expected to contain
+// mono audio. The second half is output only. Since the data is expanding, the
+// algorithm starts copying from the last sample. Otherwise it would overwrite
+// data not already copied.
+void AUAudioInputStream::UpmixMonoToStereoInPlace(AudioBuffer* audio_buffer,
+                                                  int bytes_per_sample) {
+  constexpr int channels = 2;
+  DCHECK_EQ(audio_buffer->mNumberChannels, static_cast<UInt32>(channels));
+  const int total_bytes = audio_buffer->mDataByteSize;
+  const int frames = total_bytes / bytes_per_sample / channels;
+  char* byte_ptr = reinterpret_cast<char*>(audio_buffer->mData);
+  for (int i = frames - 1; i >= 0; --i) {
+    int in_offset = (bytes_per_sample * i);
+    int out_offset = (channels * bytes_per_sample * i);
+    for (int b = 0; b < bytes_per_sample; ++b) {
+      const char byte = byte_ptr[in_offset + b];
+      byte_ptr[out_offset + b] = byte;
+      byte_ptr[out_offset + bytes_per_sample + b] = byte;
+    }
+  }
+}
+
 }  // namespace media

media/audio/mac/audio_low_latency_input_mac.h

@@ -98,6 +98,11 @@ class MEDIA_EXPORT AUAudioInputStream
   }
   AudioUnit audio_unit() const { return audio_unit_; }
 
+  // Fan out the data from the first half of audio_buffer into interleaved
+  // stereo across the whole of audio_buffer. Public for testing only.
+  static void UpmixMonoToStereoInPlace(AudioBuffer* audio_buffer,
+                                       int bytes_per_sample);
+
  private:
   bool OpenAUHAL();
   bool OpenVoiceProcessingAU();

media/audio/mac/audio_low_latency_input_mac_unittest.cc

@@ -285,4 +285,54 @@ TEST_F(MacAudioInputTest, DISABLED_AUAudioInputStreamRecordToFile) {
   ais->Close();
 }
 
+TEST(MacAudioInputUpmixerTest, Upmix16bit) {
+  constexpr int kNumFrames = 512;
+  constexpr int kBytesPerSample = sizeof(int16_t);
+  int16_t mono[kNumFrames];
+  int16_t stereo[kNumFrames * 2];
+
+  // Fill the mono buffer and the first half of the stereo buffer with data
+  for (int i = 0; i != kNumFrames; ++i) {
+    mono[i] = i;
+    stereo[i] = i;
+  }
+
+  AudioBuffer audio_buffer;
+  audio_buffer.mNumberChannels = 2;
+  audio_buffer.mDataByteSize = kNumFrames * kBytesPerSample * 2;
+  audio_buffer.mData = stereo;
+  AUAudioInputStream::UpmixMonoToStereoInPlace(&audio_buffer, kBytesPerSample);
+
+  // Assert that the samples have been distributed properly
+  for (int i = 0; i != kNumFrames; ++i) {
+    ASSERT_EQ(mono[i], stereo[i * 2]);
+    ASSERT_EQ(mono[i], stereo[i * 2 + 1]);
+  }
+}
+
+TEST(MacAudioInputUpmixerTest, Upmix32bit) {
+  constexpr int kNumFrames = 512;
+  constexpr int kBytesPerSample = sizeof(int32_t);
+  int32_t mono[kNumFrames];
+  int32_t stereo[kNumFrames * 2];
+
+  // Fill the mono buffer and the first half of the stereo buffer with data
+  for (int i = 0; i != kNumFrames; ++i) {
+    mono[i] = i;
+    stereo[i] = i;
+  }
+
+  AudioBuffer audio_buffer;
+  audio_buffer.mNumberChannels = 2;
+  audio_buffer.mDataByteSize = kNumFrames * kBytesPerSample * 2;
+  audio_buffer.mData = stereo;
+  AUAudioInputStream::UpmixMonoToStereoInPlace(&audio_buffer, kBytesPerSample);
+
+  // Assert that the samples have been distributed properly
+  for (int i = 0; i != kNumFrames; ++i) {
+    ASSERT_EQ(mono[i], stereo[i * 2]);
+    ASSERT_EQ(mono[i], stereo[i * 2 + 1]);
+  }
+}
+
 }  // namespace media