2 files changed, 23 insertions, 7 deletions

diff --git a/src/audio/audio_task.cpp b/src/audio/audio_task.cpp
index d4c1d27a..24bc7be7 100644
--- a/src/audio/audio_task.cpp
+++ b/src/audio/audio_task.cpp
@@ -173,11 +173,11 @@ void AudioTaskMain(std::unique_ptr<Pipeline> pipeline, IAudioSink* sink) {
       float samples_sunk = bytes_sunk;
       samples_sunk /= pcm.channels;
 
-      int8_t bps = pcm.bits_per_sample;
-      if (bps == 24) {
-        bps = 32;
-      }
-      samples_sunk /= (bps / 8);
+      // Samples must be aligned to 16 bits. The number of actual bytes per
+      // sample is therefore the bps divided by 16, rounded up (align to word),
+      // times two (convert to bytes).
+      uint8_t bytes_per_sample = ((pcm.bits_per_sample + 16 - 1) / 16) * 2;
+      samples_sunk /= bytes_per_sample;
 
       current_sample_in_second += samples_sunk;
       while (current_sample_in_second >= pcm.sample_rate) {
diff --git a/src/audio/fatfs_audio_input.cpp b/src/audio/fatfs_audio_input.cpp
index 86b455f0..ca5b02a1 100644
--- a/src/audio/fatfs_audio_input.cpp
+++ b/src/audio/fatfs_audio_input.cpp
@@ -9,6 +9,7 @@
 
 #include <algorithm>
 #include <chrono>
+#include <cstddef>
 #include <cstdint>
 #include <future>
 #include <memory>
@@ -71,8 +72,7 @@ auto FatfsAudioInput::OpenFile(const std::string& path) -> bool {
   database::TrackTags tags;
   if (!tag_parser.ReadAndParseTags(path, &tags)) {
     ESP_LOGE(kTag, "failed to read tags");
-    tags.encoding = database::Encoding::kFlac;
-    // return false;
+    return false;
   }
 
   auto stream_type = ContainerToStreamType(tags.encoding);
@@ -115,6 +115,8 @@ auto FatfsAudioInput::NeedsToProcess() const -> bool {
 
 auto FatfsAudioInput::Process(const std::vector<InputStream>& inputs,
                               OutputStream* output) -> void {
+  // If the next path is being given to us asynchronously, then we need to check
+  // in regularly to see if it's available yet.
   if (pending_path_) {
     if (!pending_path_->valid()) {
       pending_path_ = {};
@@ -133,11 +135,15 @@ auto FatfsAudioInput::Process(const std::vector<InputStream>& inputs,
     return;
   }
 
+  // If the output buffer isn't ready for a new stream, then we need to wait.
   if (!has_prepared_output_ && !output->prepare(*current_format_)) {
     return;
   }
   has_prepared_output_ = true;
 
+  // Performing many small reads is inefficient; it's better to do fewer, larger
+  // reads. Try to achieve this by only reading in new bytes if the output
+  // buffer has been mostly drained.
   std::size_t max_size = output->data().size_bytes();
   if (max_size < output->data().size_bytes() / 2) {
     return;
@@ -148,6 +154,7 @@ auto FatfsAudioInput::Process(const std::vector<InputStream>& inputs,
       f_read(&current_file_, output->data().data(), max_size, &size);
   if (result != FR_OK) {
     ESP_LOGE(kTag, "file I/O error %d", result);
+    output->mark_producer_finished();
     // TODO(jacqueline): Handle errors.
     return;
   }
@@ -155,6 +162,15 @@ auto FatfsAudioInput::Process(const std::vector<InputStream>& inputs,
   output->add(size);
 
   if (size < max_size || f_eof(&current_file_)) {
+    // HACK: In order to decode the last frame of a file, libmad requires 8
+    // 0-bytes ( == MAD_GUARD_BYTES) to be appended to the end of the stream.
+    // It would be better to do this within mad.cpp, but so far it's the only
+    // decoder that has such a requirement.
+    if (current_container_ == database::Encoding::kMp3) {
+      std::fill_n(output->data().begin(), 8, std::byte(0));
+      output->add(8);
+    }
+
     f_close(&current_file_);
     is_file_open_ = false;
     has_prepared_output_ = false;