ladybird/Userland/Libraries/LibAudio/FlacLoader.cpp

1050 lines
47 KiB
C++
Raw Normal View History

/*
* Copyright (c) 2021, kleines Filmröllchen <filmroellchen@serenityos.org>
*
* SPDX-License-Identifier: BSD-2-Clause
*/
#include <AK/ByteString.h>
#include <AK/Debug.h>
#include <AK/DeprecatedFlyString.h>
#include <AK/FixedArray.h>
#include <AK/Format.h>
#include <AK/IntegralMath.h>
#include <AK/Math.h>
2023-01-25 20:19:05 +01:00
#include <AK/MemoryStream.h>
#include <AK/NonnullOwnPtr.h>
#include <AK/ScopeGuard.h>
#include <AK/StdLibExtras.h>
2021-11-27 17:00:19 +01:00
#include <AK/Try.h>
#include <AK/TypedTransfer.h>
2021-11-27 17:00:19 +01:00
#include <AK/UFixedBigInt.h>
#include <LibAudio/FlacLoader.h>
#include <LibAudio/FlacTypes.h>
#include <LibAudio/GenericTypes.h>
2021-11-27 17:00:19 +01:00
#include <LibAudio/LoaderError.h>
#include <LibAudio/MultiChannel.h>
#include <LibAudio/Resampler.h>
#include <LibAudio/VorbisComment.h>
#include <LibCore/File.h>
#include <LibCrypto/Checksum/ChecksumFunction.h>
#include <LibCrypto/Checksum/ChecksummingStream.h>
namespace Audio {
FlacLoaderPlugin::FlacLoaderPlugin(NonnullOwnPtr<SeekableStream> stream)
: LoaderPlugin(move(stream))
{
}
ErrorOr<NonnullOwnPtr<LoaderPlugin>, LoaderError> FlacLoaderPlugin::create(NonnullOwnPtr<SeekableStream> stream)
{
auto loader = make<FlacLoaderPlugin>(move(stream));
TRY(loader->initialize());
return loader;
}
MaybeLoaderError FlacLoaderPlugin::initialize()
{
2021-11-27 17:00:19 +01:00
TRY(parse_header());
TRY(reset());
return {};
}
bool FlacLoaderPlugin::sniff(SeekableStream& stream)
{
BigEndianInputBitStream bit_input { MaybeOwned<Stream>(stream) };
auto maybe_flac = bit_input.read_bits<u32>(32);
return !maybe_flac.is_error() && maybe_flac.value() == 0x664C6143; // "flaC"
}
// 11.5 STREAM
2021-11-27 17:00:19 +01:00
MaybeLoaderError FlacLoaderPlugin::parse_header()
{
BigEndianInputBitStream bit_input { MaybeOwned<Stream>(*m_stream) };
2021-11-27 17:00:19 +01:00
// A mixture of VERIFY and the non-crashing TRY().
#define FLAC_VERIFY(check, category, msg) \
do { \
if (!(check)) { \
return LoaderError { category, TRY(m_stream->tell()), ByteString::formatted("FLAC header: {}", msg) }; \
} \
} while (0)
// Magic number
u32 flac = TRY(bit_input.read_bits<u32>(32));
m_data_start_location += 4;
2021-11-27 17:00:19 +01:00
FLAC_VERIFY(flac == 0x664C6143, LoaderError::Category::Format, "Magic number must be 'flaC'"); // "flaC"
// Receive the streaminfo block
auto streaminfo = TRY(next_meta_block(bit_input));
2021-11-27 17:00:19 +01:00
FLAC_VERIFY(streaminfo.type == FlacMetadataBlockType::STREAMINFO, LoaderError::Category::Format, "First block must be STREAMINFO");
FixedMemoryStream streaminfo_data_memory { streaminfo.data.bytes() };
BigEndianInputBitStream streaminfo_data { MaybeOwned<Stream>(streaminfo_data_memory) };
// 11.10 METADATA_BLOCK_STREAMINFO
m_min_block_size = TRY(streaminfo_data.read_bits<u16>(16));
2021-11-27 17:00:19 +01:00
FLAC_VERIFY(m_min_block_size >= 16, LoaderError::Category::Format, "Minimum block size must be 16");
m_max_block_size = TRY(streaminfo_data.read_bits<u16>(16));
2021-11-27 17:00:19 +01:00
FLAC_VERIFY(m_max_block_size >= 16, LoaderError::Category::Format, "Maximum block size");
m_min_frame_size = TRY(streaminfo_data.read_bits<u32>(24));
m_max_frame_size = TRY(streaminfo_data.read_bits<u32>(24));
m_sample_rate = TRY(streaminfo_data.read_bits<u32>(20));
2021-11-27 17:00:19 +01:00
FLAC_VERIFY(m_sample_rate <= 655350, LoaderError::Category::Format, "Sample rate");
m_num_channels = TRY(streaminfo_data.read_bits<u8>(3)) + 1; // 0 = one channel
m_bits_per_sample = TRY(streaminfo_data.read_bits<u8>(5)) + 1;
if (m_bits_per_sample <= 8) {
// FIXME: Signed/Unsigned issues?
m_sample_format = PcmSampleFormat::Uint8;
} else if (m_bits_per_sample <= 16) {
m_sample_format = PcmSampleFormat::Int16;
} else if (m_bits_per_sample <= 24) {
m_sample_format = PcmSampleFormat::Int24;
} else if (m_bits_per_sample <= 32) {
m_sample_format = PcmSampleFormat::Int32;
} else {
FLAC_VERIFY(false, LoaderError::Category::Format, "Sample bit depth too large");
}
m_total_samples = TRY(streaminfo_data.read_bits<u64>(36));
if (m_total_samples == 0) {
// "A value of zero here means the number of total samples is unknown."
dbgln("FLAC Warning: File has unknown amount of samples, the loader will not stop before EOF");
m_total_samples = NumericLimits<decltype(m_total_samples)>::max();
}
VERIFY(streaminfo_data.is_aligned_to_byte_boundary());
TRY(streaminfo_data.read_until_filled({ m_md5_checksum, sizeof(m_md5_checksum) }));
// Parse other blocks
[[maybe_unused]] u16 meta_blocks_parsed = 1;
[[maybe_unused]] u16 total_meta_blocks = meta_blocks_parsed;
FlacRawMetadataBlock block = streaminfo;
while (!block.is_last_block) {
block = TRY(next_meta_block(bit_input));
switch (block.type) {
case (FlacMetadataBlockType::SEEKTABLE):
TRY(load_seektable(block));
break;
case FlacMetadataBlockType::PICTURE:
TRY(load_picture(block));
break;
case FlacMetadataBlockType::APPLICATION:
// Note: Third-party library can encode specific data in this.
dbgln("FLAC Warning: Unknown 'Application' metadata block encountered.");
[[fallthrough]];
case FlacMetadataBlockType::PADDING:
// Note: A padding block is empty and does not need any treatment.
break;
case FlacMetadataBlockType::VORBIS_COMMENT:
load_vorbis_comment(block);
break;
default:
// TODO: Parse the remaining metadata block types.
break;
}
++total_meta_blocks;
}
dbgln_if(AFLACLOADER_DEBUG, "Parsed FLAC header: blocksize {}-{}{}, framesize {}-{}, {}Hz, {}bit, {} channels, {} samples total ({:.2f}s), MD5 {}, data start at {:x} bytes, {} headers total (skipped {})", m_min_block_size, m_max_block_size, is_fixed_blocksize_stream() ? " (constant)" : "", m_min_frame_size, m_max_frame_size, m_sample_rate, pcm_bits_per_sample(m_sample_format), m_num_channels, m_total_samples, static_cast<float>(m_total_samples) / static_cast<float>(m_sample_rate), m_md5_checksum, m_data_start_location, total_meta_blocks, total_meta_blocks - meta_blocks_parsed);
TRY(m_seektable.insert_seek_point({ 0, 0 }));
2021-11-27 17:00:19 +01:00
return {};
}
// 11.19. METADATA_BLOCK_PICTURE
MaybeLoaderError FlacLoaderPlugin::load_picture(FlacRawMetadataBlock& block)
{
FixedMemoryStream memory_stream { block.data.bytes() };
BigEndianInputBitStream picture_block_bytes { MaybeOwned<Stream>(memory_stream) };
PictureData picture;
picture.type = static_cast<ID3PictureType>(TRY(picture_block_bytes.read_bits(32)));
auto const mime_string_length = TRY(picture_block_bytes.read_bits(32));
auto offset_before_seeking = memory_stream.offset();
if (offset_before_seeking + mime_string_length >= block.data.size())
return LoaderError { LoaderError::Category::Format, TRY(m_stream->tell()), "Picture MIME type exceeds available data" };
// "The MIME type string, in printable ASCII characters 0x20-0x7E."
picture.mime_string = TRY(String::from_stream(memory_stream, mime_string_length));
for (auto code_point : picture.mime_string.code_points()) {
if (code_point < 0x20 || code_point > 0x7E)
return LoaderError { LoaderError::Category::Format, TRY(m_stream->tell()), "Picture MIME type is not ASCII in range 0x20 - 0x7E" };
}
auto const description_string_length = TRY(picture_block_bytes.read_bits(32));
offset_before_seeking = memory_stream.offset();
if (offset_before_seeking + description_string_length >= block.data.size())
return LoaderError { LoaderError::Category::Format, TRY(m_stream->tell()), "Picture description exceeds available data" };
picture.description_string = TRY(String::from_stream(memory_stream, description_string_length));
picture.width = TRY(picture_block_bytes.read_bits(32));
picture.height = TRY(picture_block_bytes.read_bits(32));
picture.color_depth = TRY(picture_block_bytes.read_bits(32));
picture.colors = TRY(picture_block_bytes.read_bits(32));
auto const picture_size = TRY(picture_block_bytes.read_bits(32));
offset_before_seeking = memory_stream.offset();
if (offset_before_seeking + picture_size > block.data.size())
return LoaderError { LoaderError::Category::Format, static_cast<size_t>(TRY(m_stream->tell())), "Picture size exceeds available data" };
TRY(memory_stream.seek(picture_size, SeekMode::FromCurrentPosition));
picture.data = Vector<u8> { block.data.bytes().slice(offset_before_seeking, picture_size) };
m_pictures.append(move(picture));
return {};
}
// 11.15. METADATA_BLOCK_VORBIS_COMMENT
void FlacLoaderPlugin::load_vorbis_comment(FlacRawMetadataBlock& block)
{
auto metadata_or_error = Audio::load_vorbis_comment(block.data);
if (metadata_or_error.is_error()) {
dbgln("FLAC Warning: Vorbis comment invalid, error: {}", metadata_or_error.release_error());
return;
}
m_metadata = metadata_or_error.release_value();
}
// 11.13. METADATA_BLOCK_SEEKTABLE
MaybeLoaderError FlacLoaderPlugin::load_seektable(FlacRawMetadataBlock& block)
{
FixedMemoryStream memory_stream { block.data.bytes() };
BigEndianInputBitStream seektable_bytes { MaybeOwned<Stream>(memory_stream) };
for (size_t i = 0; i < block.length / 18; ++i) {
// 11.14. SEEKPOINT
u64 sample_index = TRY(seektable_bytes.read_bits<u64>(64));
u64 byte_offset = TRY(seektable_bytes.read_bits<u64>(64));
// The sample count of a seek point is not relevant to us.
[[maybe_unused]] u16 sample_count = TRY(seektable_bytes.read_bits<u16>(16));
// Placeholder, to be ignored.
if (sample_index == 0xFFFFFFFFFFFFFFFF)
continue;
SeekPoint seekpoint {
.sample_index = sample_index,
.byte_offset = byte_offset,
};
TRY(m_seektable.insert_seek_point(seekpoint));
}
dbgln_if(AFLACLOADER_DEBUG, "Loaded seektable of size {}", m_seektable.size());
return {};
}
// 11.6 METADATA_BLOCK
ErrorOr<FlacRawMetadataBlock, LoaderError> FlacLoaderPlugin::next_meta_block(BigEndianInputBitStream& bit_input)
{
// 11.7 METADATA_BLOCK_HEADER
bool is_last_block = TRY(bit_input.read_bit());
// The block type enum constants agree with the specification
FlacMetadataBlockType type = (FlacMetadataBlockType)TRY(bit_input.read_bits<u8>(7));
m_data_start_location += 1;
2021-11-27 17:00:19 +01:00
FLAC_VERIFY(type != FlacMetadataBlockType::INVALID, LoaderError::Category::Format, "Invalid metadata block");
u32 block_length = TRY(bit_input.read_bits<u32>(24));
m_data_start_location += 3;
// Blocks can be zero-sized, which would trip up the raw data reader below.
if (block_length == 0)
return FlacRawMetadataBlock {
.is_last_block = is_last_block,
.type = type,
.length = 0,
.data = TRY(ByteBuffer::create_uninitialized(0))
};
auto block_data_result = ByteBuffer::create_uninitialized(block_length);
FLAC_VERIFY(!block_data_result.is_error(), LoaderError::Category::IO, "Out of memory");
auto block_data = block_data_result.release_value();
TRY(bit_input.read_until_filled(block_data));
m_data_start_location += block_length;
return FlacRawMetadataBlock {
is_last_block,
type,
block_length,
block_data,
};
}
2021-11-27 17:00:19 +01:00
#undef FLAC_VERIFY
2021-11-27 17:00:19 +01:00
MaybeLoaderError FlacLoaderPlugin::reset()
{
TRY(seek(0));
m_current_frame.clear();
2021-11-27 17:00:19 +01:00
return {};
}
MaybeLoaderError FlacLoaderPlugin::seek(int int_sample_index)
{
auto sample_index = static_cast<size_t>(int_sample_index);
if (sample_index == m_loaded_samples)
return {};
auto maybe_target_seekpoint = m_seektable.seek_point_before(sample_index);
// No seektable or no fitting entry: Perform normal forward read
if (!maybe_target_seekpoint.has_value()) {
if (sample_index < m_loaded_samples) {
TRY(m_stream->seek(m_data_start_location, SeekMode::SetPosition));
m_loaded_samples = 0;
}
if (sample_index - m_loaded_samples == 0)
return {};
dbgln_if(AFLACLOADER_DEBUG, "Seeking {} samples manually", sample_index - m_loaded_samples);
} else {
auto target_seekpoint = maybe_target_seekpoint.release_value();
// When a small seek happens, we may already be closer to the target than the seekpoint.
if (sample_index - target_seekpoint.sample_index > sample_index - m_loaded_samples) {
dbgln_if(AFLACLOADER_DEBUG, "Close enough to target ({} samples): ignoring seek point", sample_index - m_loaded_samples);
} else {
dbgln_if(AFLACLOADER_DEBUG, "Seeking to seektable: sample index {}, byte offset {}", target_seekpoint.sample_index, target_seekpoint.byte_offset);
auto position = target_seekpoint.byte_offset + m_data_start_location;
if (m_stream->seek(static_cast<i64>(position), SeekMode::SetPosition).is_error())
return LoaderError { LoaderError::Category::IO, m_loaded_samples, ByteString::formatted("Invalid seek position {}", position) };
m_loaded_samples = target_seekpoint.sample_index;
}
}
// Skip frames until we're just before the target sample.
VERIFY(m_loaded_samples <= sample_index);
size_t frame_start_location;
while (m_loaded_samples <= sample_index) {
frame_start_location = TRY(m_stream->tell());
(void)TRY(next_frame());
m_loaded_samples += m_current_frame->sample_count;
}
TRY(m_stream->seek(frame_start_location, SeekMode::SetPosition));
2021-11-27 17:00:19 +01:00
return {};
}
bool FlacLoaderPlugin::should_insert_seekpoint_at(u64 sample_index) const
{
auto const max_seekpoint_distance = (maximum_seekpoint_distance_ms * m_sample_rate) / 1000;
auto const seek_tolerance = (seek_tolerance_ms * m_sample_rate) / 1000;
auto const current_seekpoint_distance = m_seektable.seek_point_sample_distance_around(sample_index).value_or(NumericLimits<u64>::max());
auto const previous_seekpoint = m_seektable.seek_point_before(sample_index);
auto const distance_to_previous_seekpoint = previous_seekpoint.has_value() ? sample_index - previous_seekpoint->sample_index : NumericLimits<u64>::max();
// We insert a seekpoint only under two conditions:
// - The seek points around us are spaced too far for what the loader recommends.
// Prevents inserting too many seek points between pre-loaded seek points.
// - We are so far away from the previous seek point that seeking will become too imprecise if we don't insert a seek point at least here.
// Prevents inserting too many seek points at the end of files without pre-loaded seek points.
return current_seekpoint_distance >= max_seekpoint_distance && distance_to_previous_seekpoint >= seek_tolerance;
}
LibAudio: Move audio stream buffering into the loader Before, some loader plugins implemented their own buffering (FLAC&MP3), some didn't require any (WAV), and some didn't buffer at all (QOA). This meant that in practice, while you could load arbitrary amounts of samples from some loader plugins, you couldn't do that with some others. Also, it was ill-defined how many samples you would actually get back from a get_more_samples call. This commit fixes that by introducing a layer of abstraction between the loader and its plugins (because that's the whole point of having the extra class!). The plugins now only implement a load_chunks() function, which is much simpler to implement and allows plugins to play fast and loose with what they actually return. Basically, they can return many chunks of samples, where one chunk is simply a convenient block of samples to load. In fact, some loaders such as FLAC and QOA have separate internal functions for loading exactly one chunk. The loaders *should* load as many chunks as necessary for the sample count to be reached or surpassed (the latter simplifies loading loops in the implementations, since you don't need to know how large your next chunk is going to be; a problem for e.g. FLAC). If a plugin has no problems returning data of arbitrary size (currently WAV), it can return a single chunk that exactly (or roughly) matches the requested sample count. If a plugin is at the stream end, it can also return less samples than was requested! The loader can handle all of these cases and may call into load_chunk multiple times. If the plugin returns an empty chunk list (or only empty chunks; again, they can play fast and loose), the loader takes that as a stream end signal. Otherwise, the loader will always return exactly as many samples as the user requested. Buffering is handled by the loader, allowing any underlying plugin to deal with any weird sample count requirement the user throws at it (looking at you, SoundPlayer!). This (not accidentally!) makes QOA work in SoundPlayer.
2023-02-27 00:05:14 +01:00
ErrorOr<Vector<FixedArray<Sample>>, LoaderError> FlacLoaderPlugin::load_chunks(size_t samples_to_read_from_input)
{
ssize_t remaining_samples = static_cast<ssize_t>(m_total_samples - m_loaded_samples);
// The first condition is relevant for unknown-size streams (total samples = 0 in the header)
if (m_stream->is_eof() || (m_total_samples < NumericLimits<u64>::max() && remaining_samples <= 0))
LibAudio: Move audio stream buffering into the loader Before, some loader plugins implemented their own buffering (FLAC&MP3), some didn't require any (WAV), and some didn't buffer at all (QOA). This meant that in practice, while you could load arbitrary amounts of samples from some loader plugins, you couldn't do that with some others. Also, it was ill-defined how many samples you would actually get back from a get_more_samples call. This commit fixes that by introducing a layer of abstraction between the loader and its plugins (because that's the whole point of having the extra class!). The plugins now only implement a load_chunks() function, which is much simpler to implement and allows plugins to play fast and loose with what they actually return. Basically, they can return many chunks of samples, where one chunk is simply a convenient block of samples to load. In fact, some loaders such as FLAC and QOA have separate internal functions for loading exactly one chunk. The loaders *should* load as many chunks as necessary for the sample count to be reached or surpassed (the latter simplifies loading loops in the implementations, since you don't need to know how large your next chunk is going to be; a problem for e.g. FLAC). If a plugin has no problems returning data of arbitrary size (currently WAV), it can return a single chunk that exactly (or roughly) matches the requested sample count. If a plugin is at the stream end, it can also return less samples than was requested! The loader can handle all of these cases and may call into load_chunk multiple times. If the plugin returns an empty chunk list (or only empty chunks; again, they can play fast and loose), the loader takes that as a stream end signal. Otherwise, the loader will always return exactly as many samples as the user requested. Buffering is handled by the loader, allowing any underlying plugin to deal with any weird sample count requirement the user throws at it (looking at you, SoundPlayer!). This (not accidentally!) makes QOA work in SoundPlayer.
2023-02-27 00:05:14 +01:00
return Vector<FixedArray<Sample>> {};
LibAudio: Move audio stream buffering into the loader Before, some loader plugins implemented their own buffering (FLAC&MP3), some didn't require any (WAV), and some didn't buffer at all (QOA). This meant that in practice, while you could load arbitrary amounts of samples from some loader plugins, you couldn't do that with some others. Also, it was ill-defined how many samples you would actually get back from a get_more_samples call. This commit fixes that by introducing a layer of abstraction between the loader and its plugins (because that's the whole point of having the extra class!). The plugins now only implement a load_chunks() function, which is much simpler to implement and allows plugins to play fast and loose with what they actually return. Basically, they can return many chunks of samples, where one chunk is simply a convenient block of samples to load. In fact, some loaders such as FLAC and QOA have separate internal functions for loading exactly one chunk. The loaders *should* load as many chunks as necessary for the sample count to be reached or surpassed (the latter simplifies loading loops in the implementations, since you don't need to know how large your next chunk is going to be; a problem for e.g. FLAC). If a plugin has no problems returning data of arbitrary size (currently WAV), it can return a single chunk that exactly (or roughly) matches the requested sample count. If a plugin is at the stream end, it can also return less samples than was requested! The loader can handle all of these cases and may call into load_chunk multiple times. If the plugin returns an empty chunk list (or only empty chunks; again, they can play fast and loose), the loader takes that as a stream end signal. Otherwise, the loader will always return exactly as many samples as the user requested. Buffering is handled by the loader, allowing any underlying plugin to deal with any weird sample count requirement the user throws at it (looking at you, SoundPlayer!). This (not accidentally!) makes QOA work in SoundPlayer.
2023-02-27 00:05:14 +01:00
size_t samples_to_read = min(samples_to_read_from_input, remaining_samples);
Vector<FixedArray<Sample>> frames;
// In this case we can know exactly how many frames we're going to read.
if (is_fixed_blocksize_stream() && m_current_frame.has_value())
TRY(frames.try_ensure_capacity(samples_to_read / m_current_frame->sample_count + 1));
size_t sample_index = 0;
while (!m_stream->is_eof() && sample_index < samples_to_read) {
LibAudio: Move audio stream buffering into the loader Before, some loader plugins implemented their own buffering (FLAC&MP3), some didn't require any (WAV), and some didn't buffer at all (QOA). This meant that in practice, while you could load arbitrary amounts of samples from some loader plugins, you couldn't do that with some others. Also, it was ill-defined how many samples you would actually get back from a get_more_samples call. This commit fixes that by introducing a layer of abstraction between the loader and its plugins (because that's the whole point of having the extra class!). The plugins now only implement a load_chunks() function, which is much simpler to implement and allows plugins to play fast and loose with what they actually return. Basically, they can return many chunks of samples, where one chunk is simply a convenient block of samples to load. In fact, some loaders such as FLAC and QOA have separate internal functions for loading exactly one chunk. The loaders *should* load as many chunks as necessary for the sample count to be reached or surpassed (the latter simplifies loading loops in the implementations, since you don't need to know how large your next chunk is going to be; a problem for e.g. FLAC). If a plugin has no problems returning data of arbitrary size (currently WAV), it can return a single chunk that exactly (or roughly) matches the requested sample count. If a plugin is at the stream end, it can also return less samples than was requested! The loader can handle all of these cases and may call into load_chunk multiple times. If the plugin returns an empty chunk list (or only empty chunks; again, they can play fast and loose), the loader takes that as a stream end signal. Otherwise, the loader will always return exactly as many samples as the user requested. Buffering is handled by the loader, allowing any underlying plugin to deal with any weird sample count requirement the user throws at it (looking at you, SoundPlayer!). This (not accidentally!) makes QOA work in SoundPlayer.
2023-02-27 00:05:14 +01:00
TRY(frames.try_append(TRY(next_frame())));
sample_index += m_current_frame->sample_count;
}
m_loaded_samples += sample_index;
LibAudio+Userland: Use new audio queue in client-server communication Previously, we were sending Buffers to the server whenever we had new audio data for it. This meant that for every audio enqueue action, we needed to create a new shared memory anonymous buffer, send that buffer's file descriptor over IPC (+recfd on the other side) and then map the buffer into the audio server's memory to be able to play it. This was fine for sending large chunks of audio data, like when playing existing audio files. However, in the future we want to move to real-time audio in some applications like Piano. This means that the size of buffers that are sent need to be very small, as just the size of a buffer itself is part of the audio latency. If we were to try real-time audio with the existing system, we would run into problems really quickly. Dealing with a continuous stream of new anonymous files like the current audio system is rather expensive, as we need Kernel help in multiple places. Additionally, every enqueue incurs an IPC call, which are not optimized for >1000 calls/second (which would be needed for real-time audio with buffer sizes of ~40 samples). So a fundamental change in how we handle audio sending in userspace is necessary. This commit moves the audio sending system onto a shared single producer circular queue (SSPCQ) (introduced with one of the previous commits). This queue is intended to live in shared memory and be accessed by multiple processes at the same time. It was specifically written to support the audio sending case, so e.g. it only supports a single producer (the audio client). Now, audio sending follows these general steps: - The audio client connects to the audio server. - The audio client creates a SSPCQ in shared memory. - The audio client sends the SSPCQ's file descriptor to the audio server with the set_buffer() IPC call. - The audio server receives the SSPCQ and maps it. - The audio client signals start of playback with start_playback(). - At the same time: - The audio client writes its audio data into the shared-memory queue. - The audio server reads audio data from the shared-memory queue(s). Both sides have additional before-queue/after-queue buffers, depending on the exact application. - Pausing playback is just an IPC call, nothing happens to the buffer except that the server stops reading from it until playback is resumed. - Muting has nothing to do with whether audio data is read or not. - When the connection closes, the queues are unmapped on both sides. This should already improve audio playback performance in a bunch of places. Implementation & commit notes: - Audio loaders don't create LegacyBuffers anymore. LegacyBuffer is kept for WavLoader, see previous commit message. - Most intra-process audio data passing is done with FixedArray<Sample> or Vector<Sample>. - Improvements to most audio-enqueuing applications. (If necessary I can try to extract some of the aplay improvements.) - New APIs on LibAudio/ClientConnection which allows non-realtime applications to enqueue audio in big chunks like before. - Removal of status APIs from the audio server connection for information that can be directly obtained from the shared queue. - Split the pause playback API into two APIs with more intuitive names. I know this is a large commit, and you can kinda tell from the commit message. It's basically impossible to break this up without hacks, so please forgive me. These are some of the best changes to the audio subsystem and I hope that that makes up for this :yaktangle: commit. :yakring:
2022-02-20 13:01:22 +01:00
LibAudio: Move audio stream buffering into the loader Before, some loader plugins implemented their own buffering (FLAC&MP3), some didn't require any (WAV), and some didn't buffer at all (QOA). This meant that in practice, while you could load arbitrary amounts of samples from some loader plugins, you couldn't do that with some others. Also, it was ill-defined how many samples you would actually get back from a get_more_samples call. This commit fixes that by introducing a layer of abstraction between the loader and its plugins (because that's the whole point of having the extra class!). The plugins now only implement a load_chunks() function, which is much simpler to implement and allows plugins to play fast and loose with what they actually return. Basically, they can return many chunks of samples, where one chunk is simply a convenient block of samples to load. In fact, some loaders such as FLAC and QOA have separate internal functions for loading exactly one chunk. The loaders *should* load as many chunks as necessary for the sample count to be reached or surpassed (the latter simplifies loading loops in the implementations, since you don't need to know how large your next chunk is going to be; a problem for e.g. FLAC). If a plugin has no problems returning data of arbitrary size (currently WAV), it can return a single chunk that exactly (or roughly) matches the requested sample count. If a plugin is at the stream end, it can also return less samples than was requested! The loader can handle all of these cases and may call into load_chunk multiple times. If the plugin returns an empty chunk list (or only empty chunks; again, they can play fast and loose), the loader takes that as a stream end signal. Otherwise, the loader will always return exactly as many samples as the user requested. Buffering is handled by the loader, allowing any underlying plugin to deal with any weird sample count requirement the user throws at it (looking at you, SoundPlayer!). This (not accidentally!) makes QOA work in SoundPlayer.
2023-02-27 00:05:14 +01:00
return frames;
}
// 11.21. FRAME
LibAudio: Move audio stream buffering into the loader Before, some loader plugins implemented their own buffering (FLAC&MP3), some didn't require any (WAV), and some didn't buffer at all (QOA). This meant that in practice, while you could load arbitrary amounts of samples from some loader plugins, you couldn't do that with some others. Also, it was ill-defined how many samples you would actually get back from a get_more_samples call. This commit fixes that by introducing a layer of abstraction between the loader and its plugins (because that's the whole point of having the extra class!). The plugins now only implement a load_chunks() function, which is much simpler to implement and allows plugins to play fast and loose with what they actually return. Basically, they can return many chunks of samples, where one chunk is simply a convenient block of samples to load. In fact, some loaders such as FLAC and QOA have separate internal functions for loading exactly one chunk. The loaders *should* load as many chunks as necessary for the sample count to be reached or surpassed (the latter simplifies loading loops in the implementations, since you don't need to know how large your next chunk is going to be; a problem for e.g. FLAC). If a plugin has no problems returning data of arbitrary size (currently WAV), it can return a single chunk that exactly (or roughly) matches the requested sample count. If a plugin is at the stream end, it can also return less samples than was requested! The loader can handle all of these cases and may call into load_chunk multiple times. If the plugin returns an empty chunk list (or only empty chunks; again, they can play fast and loose), the loader takes that as a stream end signal. Otherwise, the loader will always return exactly as many samples as the user requested. Buffering is handled by the loader, allowing any underlying plugin to deal with any weird sample count requirement the user throws at it (looking at you, SoundPlayer!). This (not accidentally!) makes QOA work in SoundPlayer.
2023-02-27 00:05:14 +01:00
LoaderSamples FlacLoaderPlugin::next_frame()
{
#define FLAC_VERIFY(check, category, msg) \
do { \
if (!(check)) { \
return LoaderError { category, static_cast<size_t>(m_current_sample_or_frame), ByteString::formatted("FLAC header: {}", msg) }; \
} \
} while (0)
auto frame_byte_index = TRY(m_stream->tell());
auto sample_index = m_loaded_samples;
// Insert a new seek point if we don't have enough here.
if (should_insert_seekpoint_at(sample_index)) {
dbgln_if(AFLACLOADER_DEBUG, "Inserting ad-hoc seek point for sample {} at byte {:x} (seekpoint spacing {} samples)", sample_index, frame_byte_index, m_seektable.seek_point_sample_distance_around(sample_index).value_or(NumericLimits<u64>::max()));
auto maybe_error = m_seektable.insert_seek_point({ .sample_index = sample_index, .byte_offset = frame_byte_index - m_data_start_location });
if (maybe_error.is_error())
dbgln("FLAC Warning: Inserting seek point for sample {} failed: {}", sample_index, maybe_error.release_error());
}
auto frame_checksum_stream = TRY(try_make<Crypto::Checksum::ChecksummingStream<IBMCRC>>(MaybeOwned<Stream>(*m_stream)));
auto header_checksum_stream = TRY(try_make<Crypto::Checksum::ChecksummingStream<FlacFrameHeaderCRC>>(MaybeOwned<Stream>(*frame_checksum_stream)));
BigEndianInputBitStream bit_stream { MaybeOwned<Stream> { *header_checksum_stream } };
// 11.22. FRAME_HEADER
u16 sync_code = TRY(bit_stream.read_bits<u16>(14));
2021-11-27 17:00:19 +01:00
FLAC_VERIFY(sync_code == 0b11111111111110, LoaderError::Category::Format, "Sync code");
bool reserved_bit = TRY(bit_stream.read_bit());
2021-11-27 17:00:19 +01:00
FLAC_VERIFY(reserved_bit == 0, LoaderError::Category::Format, "Reserved frame header bit");
// 11.22.2. BLOCKING STRATEGY
[[maybe_unused]] bool blocking_strategy = TRY(bit_stream.read_bit());
u32 sample_count = TRY(convert_sample_count_code(TRY(bit_stream.read_bits<u8>(4))));
u32 frame_sample_rate = TRY(convert_sample_rate_code(TRY(bit_stream.read_bits<u8>(4))));
u8 channel_type_num = TRY(bit_stream.read_bits<u8>(4));
FLAC_VERIFY(channel_type_num < 0b1011, LoaderError::Category::Format, "Channel assignment");
FlacFrameChannelType channel_type = (FlacFrameChannelType)channel_type_num;
u8 bit_depth = TRY(convert_bit_depth_code(TRY(bit_stream.read_bits<u8>(3))));
reserved_bit = TRY(bit_stream.read_bit());
2021-11-27 17:00:19 +01:00
FLAC_VERIFY(reserved_bit == 0, LoaderError::Category::Format, "Reserved frame header end bit");
// 11.22.8. CODED NUMBER
m_current_sample_or_frame = TRY(read_utf8_char(bit_stream));
// Conditional header variables
// 11.22.9. BLOCK SIZE INT
if (sample_count == FLAC_BLOCKSIZE_AT_END_OF_HEADER_8) {
sample_count = TRY(bit_stream.read_bits<u32>(8)) + 1;
} else if (sample_count == FLAC_BLOCKSIZE_AT_END_OF_HEADER_16) {
sample_count = TRY(bit_stream.read_bits<u32>(16)) + 1;
}
// 11.22.10. SAMPLE RATE INT
if (frame_sample_rate == FLAC_SAMPLERATE_AT_END_OF_HEADER_8) {
frame_sample_rate = TRY(bit_stream.read_bits<u32>(8)) * 1000;
} else if (frame_sample_rate == FLAC_SAMPLERATE_AT_END_OF_HEADER_16) {
frame_sample_rate = TRY(bit_stream.read_bits<u32>(16));
} else if (frame_sample_rate == FLAC_SAMPLERATE_AT_END_OF_HEADER_16X10) {
frame_sample_rate = TRY(bit_stream.read_bits<u32>(16)) * 10;
}
// It does not matter whether we extract the checksum from the digest here, or extract the digest 0x00 after processing the checksum.
auto const calculated_header_checksum = header_checksum_stream->digest();
// 11.22.11. FRAME CRC
u8 specified_header_checksum = TRY(bit_stream.read_bits<u8>(8));
VERIFY(bit_stream.is_aligned_to_byte_boundary());
if (specified_header_checksum != calculated_header_checksum)
dbgln("FLAC frame {}: Calculated header checksum {:02x} is different from specified checksum {:02x}", m_current_sample_or_frame, calculated_header_checksum, specified_header_checksum);
dbgln_if(AFLACLOADER_DEBUG, "Frame: {} samples, {}bit {}Hz, channeltype {:x}, {} number {}, header checksum {:02x}{}", sample_count, bit_depth, frame_sample_rate, channel_type_num, blocking_strategy ? "sample" : "frame", m_current_sample_or_frame, specified_header_checksum, specified_header_checksum != calculated_header_checksum ? " (checksum error)"sv : ""sv);
m_current_frame = FlacFrameHeader {
.sample_rate = frame_sample_rate,
.sample_count = static_cast<u16>(sample_count),
.sample_or_frame_index = static_cast<u32>(m_current_sample_or_frame),
.blocking_strategy = static_cast<BlockingStrategy>(blocking_strategy),
.channels = channel_type,
.bit_depth = bit_depth,
.checksum = specified_header_checksum,
};
u8 subframe_count = frame_channel_type_to_channel_count(channel_type);
TRY(m_subframe_buffers.try_resize_and_keep_capacity(subframe_count));
float sample_rescale = 1 / static_cast<float>(1 << (m_current_frame->bit_depth - 1));
dbgln_if(AFLACLOADER_DEBUG, "Samples will be rescaled from {} bits: factor {:.8f}", m_current_frame->bit_depth, sample_rescale);
for (u8 i = 0; i < subframe_count; ++i) {
FlacSubframeHeader new_subframe = TRY(next_subframe_header(bit_stream, i));
auto& subframe_samples = m_subframe_buffers[i];
subframe_samples.clear_with_capacity();
TRY(parse_subframe(subframe_samples, new_subframe, bit_stream));
// We only verify the sample count for the common case of a constant sample rate.
if (m_sample_rate == m_current_frame->sample_rate)
VERIFY(subframe_samples.size() == m_current_frame->sample_count);
}
// 11.2. Overview ("The audio data is composed of...")
bit_stream.align_to_byte_boundary();
// 11.23. FRAME_FOOTER
auto const calculated_frame_checksum = frame_checksum_stream->digest();
auto const specified_frame_checksum = TRY(bit_stream.read_bits<u16>(16));
if (calculated_frame_checksum != specified_frame_checksum)
dbgln("FLAC frame {}: Calculated frame checksum {:04x} is different from specified checksum {:04x}", m_current_sample_or_frame, calculated_frame_checksum, specified_frame_checksum);
dbgln_if(AFLACLOADER_DEBUG, "Subframe footer checksum: {:04x}{}", specified_frame_checksum, specified_frame_checksum != calculated_frame_checksum ? " (checksum error)"sv : ""sv);
FixedArray<Sample> samples;
switch (channel_type) {
case FlacFrameChannelType::Mono:
case FlacFrameChannelType::Stereo:
case FlacFrameChannelType::StereoCenter:
case FlacFrameChannelType::Surround4p0:
case FlacFrameChannelType::Surround5p0:
case FlacFrameChannelType::Surround5p1:
case FlacFrameChannelType::Surround6p1:
case FlacFrameChannelType::Surround7p1: {
auto new_samples = TRY(downmix_surround_to_stereo<Vector<i64>>(m_subframe_buffers, sample_rescale));
samples.swap(new_samples);
break;
}
case FlacFrameChannelType::LeftSideStereo: {
auto new_samples = TRY(FixedArray<Sample>::create(m_current_frame->sample_count));
samples.swap(new_samples);
// channels are left (0) and side (1)
LibAudio: Move audio stream buffering into the loader Before, some loader plugins implemented their own buffering (FLAC&MP3), some didn't require any (WAV), and some didn't buffer at all (QOA). This meant that in practice, while you could load arbitrary amounts of samples from some loader plugins, you couldn't do that with some others. Also, it was ill-defined how many samples you would actually get back from a get_more_samples call. This commit fixes that by introducing a layer of abstraction between the loader and its plugins (because that's the whole point of having the extra class!). The plugins now only implement a load_chunks() function, which is much simpler to implement and allows plugins to play fast and loose with what they actually return. Basically, they can return many chunks of samples, where one chunk is simply a convenient block of samples to load. In fact, some loaders such as FLAC and QOA have separate internal functions for loading exactly one chunk. The loaders *should* load as many chunks as necessary for the sample count to be reached or surpassed (the latter simplifies loading loops in the implementations, since you don't need to know how large your next chunk is going to be; a problem for e.g. FLAC). If a plugin has no problems returning data of arbitrary size (currently WAV), it can return a single chunk that exactly (or roughly) matches the requested sample count. If a plugin is at the stream end, it can also return less samples than was requested! The loader can handle all of these cases and may call into load_chunk multiple times. If the plugin returns an empty chunk list (or only empty chunks; again, they can play fast and loose), the loader takes that as a stream end signal. Otherwise, the loader will always return exactly as many samples as the user requested. Buffering is handled by the loader, allowing any underlying plugin to deal with any weird sample count requirement the user throws at it (looking at you, SoundPlayer!). This (not accidentally!) makes QOA work in SoundPlayer.
2023-02-27 00:05:14 +01:00
for (size_t i = 0; i < m_current_frame->sample_count; ++i) {
// right = left - side
samples[i] = { static_cast<float>(m_subframe_buffers[0][i]) * sample_rescale,
static_cast<float>(m_subframe_buffers[0][i] - m_subframe_buffers[1][i]) * sample_rescale };
}
break;
}
case FlacFrameChannelType::RightSideStereo: {
auto new_samples = TRY(FixedArray<Sample>::create(m_current_frame->sample_count));
samples.swap(new_samples);
// channels are side (0) and right (1)
LibAudio: Move audio stream buffering into the loader Before, some loader plugins implemented their own buffering (FLAC&MP3), some didn't require any (WAV), and some didn't buffer at all (QOA). This meant that in practice, while you could load arbitrary amounts of samples from some loader plugins, you couldn't do that with some others. Also, it was ill-defined how many samples you would actually get back from a get_more_samples call. This commit fixes that by introducing a layer of abstraction between the loader and its plugins (because that's the whole point of having the extra class!). The plugins now only implement a load_chunks() function, which is much simpler to implement and allows plugins to play fast and loose with what they actually return. Basically, they can return many chunks of samples, where one chunk is simply a convenient block of samples to load. In fact, some loaders such as FLAC and QOA have separate internal functions for loading exactly one chunk. The loaders *should* load as many chunks as necessary for the sample count to be reached or surpassed (the latter simplifies loading loops in the implementations, since you don't need to know how large your next chunk is going to be; a problem for e.g. FLAC). If a plugin has no problems returning data of arbitrary size (currently WAV), it can return a single chunk that exactly (or roughly) matches the requested sample count. If a plugin is at the stream end, it can also return less samples than was requested! The loader can handle all of these cases and may call into load_chunk multiple times. If the plugin returns an empty chunk list (or only empty chunks; again, they can play fast and loose), the loader takes that as a stream end signal. Otherwise, the loader will always return exactly as many samples as the user requested. Buffering is handled by the loader, allowing any underlying plugin to deal with any weird sample count requirement the user throws at it (looking at you, SoundPlayer!). This (not accidentally!) makes QOA work in SoundPlayer.
2023-02-27 00:05:14 +01:00
for (size_t i = 0; i < m_current_frame->sample_count; ++i) {
// left = right + side
samples[i] = { static_cast<float>(m_subframe_buffers[1][i] + m_subframe_buffers[0][i]) * sample_rescale,
static_cast<float>(m_subframe_buffers[1][i]) * sample_rescale };
}
break;
}
case FlacFrameChannelType::MidSideStereo: {
auto new_samples = TRY(FixedArray<Sample>::create(m_current_frame->sample_count));
samples.swap(new_samples);
// channels are mid (0) and side (1)
for (size_t i = 0; i < m_subframe_buffers[0].size(); ++i) {
i64 mid = m_subframe_buffers[0][i];
i64 side = m_subframe_buffers[1][i];
mid *= 2;
// prevent integer division errors
samples[i] = { static_cast<float>(mid + side) * .5f * sample_rescale,
static_cast<float>(mid - side) * .5f * sample_rescale };
}
break;
}
}
LibAudio: Move audio stream buffering into the loader Before, some loader plugins implemented their own buffering (FLAC&MP3), some didn't require any (WAV), and some didn't buffer at all (QOA). This meant that in practice, while you could load arbitrary amounts of samples from some loader plugins, you couldn't do that with some others. Also, it was ill-defined how many samples you would actually get back from a get_more_samples call. This commit fixes that by introducing a layer of abstraction between the loader and its plugins (because that's the whole point of having the extra class!). The plugins now only implement a load_chunks() function, which is much simpler to implement and allows plugins to play fast and loose with what they actually return. Basically, they can return many chunks of samples, where one chunk is simply a convenient block of samples to load. In fact, some loaders such as FLAC and QOA have separate internal functions for loading exactly one chunk. The loaders *should* load as many chunks as necessary for the sample count to be reached or surpassed (the latter simplifies loading loops in the implementations, since you don't need to know how large your next chunk is going to be; a problem for e.g. FLAC). If a plugin has no problems returning data of arbitrary size (currently WAV), it can return a single chunk that exactly (or roughly) matches the requested sample count. If a plugin is at the stream end, it can also return less samples than was requested! The loader can handle all of these cases and may call into load_chunk multiple times. If the plugin returns an empty chunk list (or only empty chunks; again, they can play fast and loose), the loader takes that as a stream end signal. Otherwise, the loader will always return exactly as many samples as the user requested. Buffering is handled by the loader, allowing any underlying plugin to deal with any weird sample count requirement the user throws at it (looking at you, SoundPlayer!). This (not accidentally!) makes QOA work in SoundPlayer.
2023-02-27 00:05:14 +01:00
return samples;
2021-11-27 17:00:19 +01:00
#undef FLAC_VERIFY
}
// 11.22.3. INTERCHANNEL SAMPLE BLOCK SIZE
2021-11-27 17:00:19 +01:00
ErrorOr<u32, LoaderError> FlacLoaderPlugin::convert_sample_count_code(u8 sample_count_code)
{
// single codes
switch (sample_count_code) {
case 0:
2021-11-27 17:00:19 +01:00
return LoaderError { LoaderError::Category::Format, static_cast<size_t>(m_current_sample_or_frame), "Reserved block size" };
case 1:
return 192;
case 6:
return FLAC_BLOCKSIZE_AT_END_OF_HEADER_8;
case 7:
return FLAC_BLOCKSIZE_AT_END_OF_HEADER_16;
}
if (sample_count_code >= 2 && sample_count_code <= 5) {
return 576 * AK::exp2(sample_count_code - 2);
}
return 256 * AK::exp2(sample_count_code - 8);
}
// 11.22.4. SAMPLE RATE
2021-11-27 17:00:19 +01:00
ErrorOr<u32, LoaderError> FlacLoaderPlugin::convert_sample_rate_code(u8 sample_rate_code)
{
switch (sample_rate_code) {
case 0:
return m_sample_rate;
case 1:
return 88200;
case 2:
return 176400;
case 3:
return 192000;
case 4:
return 8000;
case 5:
return 16000;
case 6:
return 22050;
case 7:
return 24000;
case 8:
return 32000;
case 9:
return 44100;
case 10:
return 48000;
case 11:
return 96000;
case 12:
return FLAC_SAMPLERATE_AT_END_OF_HEADER_8;
case 13:
return FLAC_SAMPLERATE_AT_END_OF_HEADER_16;
case 14:
return FLAC_SAMPLERATE_AT_END_OF_HEADER_16X10;
default:
2021-11-27 17:00:19 +01:00
return LoaderError { LoaderError::Category::Format, static_cast<size_t>(m_current_sample_or_frame), "Invalid sample rate code" };
}
}
// 11.22.6. SAMPLE SIZE
ErrorOr<u8, LoaderError> FlacLoaderPlugin::convert_bit_depth_code(u8 bit_depth_code)
{
switch (bit_depth_code) {
case 0:
return m_bits_per_sample;
case 1:
return 8;
case 2:
return 12;
case 3:
return LoaderError { LoaderError::Category::Format, static_cast<size_t>(m_current_sample_or_frame), "Reserved sample size" };
case 4:
return 16;
case 5:
return 20;
case 6:
return 24;
case 7:
return 32;
default:
return LoaderError { LoaderError::Category::Format, static_cast<size_t>(m_current_sample_or_frame), ByteString::formatted("Unsupported sample size {}", bit_depth_code) };
}
}
// 11.22.5. CHANNEL ASSIGNMENT
u8 frame_channel_type_to_channel_count(FlacFrameChannelType channel_type)
{
if (channel_type <= FlacFrameChannelType::Surround7p1)
return to_underlying(channel_type) + 1;
return 2;
}
// 11.25. SUBFRAME_HEADER
ErrorOr<FlacSubframeHeader, LoaderError> FlacLoaderPlugin::next_subframe_header(BigEndianInputBitStream& bit_stream, u8 channel_index)
{
u8 bits_per_sample = m_current_frame->bit_depth;
// For inter-channel correlation, the side channel needs an extra bit for its samples
switch (m_current_frame->channels) {
case FlacFrameChannelType::LeftSideStereo:
case FlacFrameChannelType::MidSideStereo:
if (channel_index == 1) {
++bits_per_sample;
}
break;
case FlacFrameChannelType::RightSideStereo:
if (channel_index == 0) {
++bits_per_sample;
}
break;
// "normal" channel types
default:
break;
}
// zero-bit padding
if (TRY(bit_stream.read_bit()) != 0)
2021-11-27 17:00:19 +01:00
return LoaderError { LoaderError::Category::Format, static_cast<size_t>(m_current_sample_or_frame), "Zero bit padding" };
// 11.25.1. SUBFRAME TYPE
u8 subframe_code = TRY(bit_stream.read_bits<u8>(6));
2021-11-27 17:00:19 +01:00
if ((subframe_code >= 0b000010 && subframe_code <= 0b000111) || (subframe_code > 0b001100 && subframe_code < 0b100000))
return LoaderError { LoaderError::Category::Format, static_cast<size_t>(m_current_sample_or_frame), "Subframe type" };
FlacSubframeType subframe_type;
u8 order = 0;
// LPC has the highest bit set
if ((subframe_code & 0b100000) > 0) {
subframe_type = FlacSubframeType::LPC;
order = (subframe_code & 0b011111) + 1;
} else if ((subframe_code & 0b001000) > 0) {
// Fixed has the third-highest bit set
subframe_type = FlacSubframeType::Fixed;
order = (subframe_code & 0b000111);
} else {
subframe_type = (FlacSubframeType)subframe_code;
}
// 11.25.2. WASTED BITS PER SAMPLE FLAG
bool has_wasted_bits = TRY(bit_stream.read_bit());
u8 k = 0;
if (has_wasted_bits) {
bool current_k_bit = 0;
do {
current_k_bit = TRY(bit_stream.read_bit());
++k;
} while (current_k_bit != 1);
}
return FlacSubframeHeader {
subframe_type,
order,
k,
bits_per_sample
};
}
ErrorOr<void, LoaderError> FlacLoaderPlugin::parse_subframe(Vector<i64>& samples, FlacSubframeHeader& subframe_header, BigEndianInputBitStream& bit_input)
{
TRY(samples.try_ensure_capacity(m_current_frame->sample_count));
switch (subframe_header.type) {
case FlacSubframeType::Constant: {
// 11.26. SUBFRAME_CONSTANT
u64 constant_value = TRY(bit_input.read_bits<u64>(subframe_header.bits_per_sample - subframe_header.wasted_bits_per_sample));
dbgln_if(AFLACLOADER_DEBUG, " Constant subframe: {}", constant_value);
VERIFY(subframe_header.bits_per_sample - subframe_header.wasted_bits_per_sample != 0);
i64 constant = sign_extend(static_cast<u64>(constant_value), subframe_header.bits_per_sample - subframe_header.wasted_bits_per_sample);
for (u64 i = 0; i < m_current_frame->sample_count; ++i) {
samples.unchecked_append(constant);
}
break;
}
case FlacSubframeType::Fixed: {
dbgln_if(AFLACLOADER_DEBUG, " Fixed LPC subframe order {}", subframe_header.order);
2021-11-27 17:00:19 +01:00
samples = TRY(decode_fixed_lpc(subframe_header, bit_input));
break;
}
case FlacSubframeType::Verbatim: {
dbgln_if(AFLACLOADER_DEBUG, " Verbatim subframe");
2021-11-27 17:00:19 +01:00
samples = TRY(decode_verbatim(subframe_header, bit_input));
break;
}
case FlacSubframeType::LPC: {
dbgln_if(AFLACLOADER_DEBUG, " Custom LPC subframe order {}", subframe_header.order);
TRY(decode_custom_lpc(samples, subframe_header, bit_input));
break;
}
default:
2021-11-27 17:00:19 +01:00
return LoaderError { LoaderError::Category::Unimplemented, static_cast<size_t>(m_current_sample_or_frame), "Unhandled FLAC subframe type" };
}
for (size_t i = 0; i < samples.size(); ++i) {
samples[i] <<= subframe_header.wasted_bits_per_sample;
}
// Resamplers VERIFY that the sample rate is non-zero.
if (m_current_frame->sample_rate == 0 || m_sample_rate == 0
|| m_current_frame->sample_rate == m_sample_rate)
return {};
ResampleHelper<i64> resampler(m_current_frame->sample_rate, m_sample_rate);
samples = resampler.resample(samples);
return {};
}
// 11.29. SUBFRAME_VERBATIM
// Decode a subframe that isn't actually encoded, usually seen in random data
ErrorOr<Vector<i64>, LoaderError> FlacLoaderPlugin::decode_verbatim(FlacSubframeHeader& subframe, BigEndianInputBitStream& bit_input)
{
Vector<i64> decoded;
decoded.ensure_capacity(m_current_frame->sample_count);
if (subframe.bits_per_sample <= subframe.wasted_bits_per_sample) {
return LoaderError {
LoaderError::Category::Format,
TRY(m_stream->tell()),
"Effective verbatim bits per sample are zero"sv,
};
}
for (size_t i = 0; i < m_current_frame->sample_count; ++i) {
decoded.unchecked_append(sign_extend(
TRY(bit_input.read_bits<u64>(subframe.bits_per_sample - subframe.wasted_bits_per_sample)),
subframe.bits_per_sample - subframe.wasted_bits_per_sample));
}
return decoded;
}
// 11.28. SUBFRAME_LPC
// Decode a subframe encoded with a custom linear predictor coding, i.e. the subframe provides the polynomial order and coefficients
ErrorOr<void, LoaderError> FlacLoaderPlugin::decode_custom_lpc(Vector<i64>& decoded, FlacSubframeHeader& subframe, BigEndianInputBitStream& bit_input)
{
// LPC must provide at least as many samples as its order.
if (subframe.order > m_current_frame->sample_count)
return LoaderError { LoaderError::Category::Format, static_cast<size_t>(m_current_sample_or_frame), "Too small frame for LPC order" };
decoded.ensure_capacity(m_current_frame->sample_count);
if (subframe.bits_per_sample <= subframe.wasted_bits_per_sample) {
return LoaderError {
LoaderError::Category::Format,
TRY(m_stream->tell()),
"Effective verbatim bits per sample are zero"sv,
};
}
// warm-up samples
for (auto i = 0; i < subframe.order; ++i) {
decoded.unchecked_append(sign_extend(
TRY(bit_input.read_bits<u64>(subframe.bits_per_sample - subframe.wasted_bits_per_sample)),
subframe.bits_per_sample - subframe.wasted_bits_per_sample));
}
// precision of the coefficients
u8 lpc_precision = TRY(bit_input.read_bits<u8>(4));
2021-11-27 17:00:19 +01:00
if (lpc_precision == 0b1111)
return LoaderError { LoaderError::Category::Format, static_cast<size_t>(m_current_sample_or_frame), "Invalid linear predictor coefficient precision" };
lpc_precision += 1;
// shift needed on the data (signed!)
i8 lpc_shift = static_cast<i8>(sign_extend(TRY(bit_input.read_bits<u8>(5)), 5));
Vector<i64, 32> coefficients;
coefficients.ensure_capacity(subframe.order);
// read coefficients
for (auto i = 0; i < subframe.order; ++i) {
u64 raw_coefficient = TRY(bit_input.read_bits<u64>(lpc_precision));
i64 coefficient = sign_extend(raw_coefficient, lpc_precision);
coefficients.unchecked_append(coefficient);
}
dbgln_if(AFLACLOADER_DEBUG, " {}-bit {} shift coefficients: {}", lpc_precision, lpc_shift, coefficients);
TRY(decode_residual(decoded, subframe, bit_input));
// approximate the waveform with the predictor
for (size_t i = subframe.order; i < m_current_frame->sample_count; ++i) {
// (see below)
Checked<i64> sample = 0;
for (size_t t = 0; t < subframe.order; ++t) {
// It's really important that we compute in 64-bit land here.
// Even though FLAC operates at a maximum bit depth of 32 bits, modern encoders use super-large coefficients for maximum compression.
2022-01-06 07:07:15 -07:00
// These will easily overflow 32 bits and cause strange white noise that abruptly stops intermittently (at the end of a frame).
// The simple fix of course is to do intermediate computations in 64 bits, but we additionally use saturating arithmetic.
// These considerations are not in the original FLAC spec, but have been added to the IETF standard: https://datatracker.ietf.org/doc/html/draft-ietf-cellar-flac-03#appendix-A.3
sample.saturating_add(Checked<i64>::saturating_mul(static_cast<i64>(coefficients[t]), static_cast<i64>(decoded[i - t - 1])));
}
decoded[i] += lpc_shift >= 0 ? (sample.value() >> lpc_shift) : (sample.value() << -lpc_shift);
}
return {};
}
// 11.27. SUBFRAME_FIXED
// Decode a subframe encoded with one of the fixed linear predictor codings
ErrorOr<Vector<i64>, LoaderError> FlacLoaderPlugin::decode_fixed_lpc(FlacSubframeHeader& subframe, BigEndianInputBitStream& bit_input)
{
// LPC must provide at least as many samples as its order.
if (subframe.order > m_current_frame->sample_count)
return LoaderError { LoaderError::Category::Format, static_cast<size_t>(m_current_sample_or_frame), "Too small frame for LPC order" };
Vector<i64> decoded;
decoded.ensure_capacity(m_current_frame->sample_count);
if (subframe.bits_per_sample <= subframe.wasted_bits_per_sample) {
return LoaderError {
LoaderError::Category::Format,
TRY(m_stream->tell()),
"Effective verbatim bits per sample are zero"sv,
};
}
// warm-up samples
for (auto i = 0; i < subframe.order; ++i) {
decoded.unchecked_append(sign_extend(
TRY(bit_input.read_bits<u64>(subframe.bits_per_sample - subframe.wasted_bits_per_sample)),
subframe.bits_per_sample - subframe.wasted_bits_per_sample));
}
2021-11-27 17:00:19 +01:00
TRY(decode_residual(decoded, subframe, bit_input));
dbgln_if(AFLACLOADER_DEBUG, " decoded length {}, {} order predictor, now at file offset {:x}", decoded.size(), subframe.order, TRY(m_stream->tell()));
// Skip these comments if you don't care about the neat math behind fixed LPC :^)
// These coefficients for the recursive prediction formula are the only ones that can be resolved to polynomial predictor functions.
// The order equals the degree of the polynomial - 1, so the second-order predictor has an underlying polynomial of degree 1, a straight line.
// More specifically, the closest approximation to a polynomial is used, and the degree depends on how many previous values are available.
// This makes use of a very neat property of polynomials, which is that they are entirely characterized by their finitely many derivatives.
// (Mathematically speaking, the infinite Taylor series of any polynomial equals the polynomial itself.)
// Now remember that derivation is just the slope of the function, which is the same as the difference of two close-by values.
// Therefore, with two samples we can calculate the first derivative at a sample via the difference, which gives us a polynomial of degree 1.
// With three samples, we can do the same but also calculate the second derivative via the difference in the first derivatives.
// This gives us a polynomial of degree 2, as it has two "proper" (non-constant) derivatives.
// This can be continued for higher-order derivatives when we have more coefficients, giving us higher-order polynomials.
// In essence, it's akin to a Lagrangian polynomial interpolation for every sample (but already pre-solved).
// The coefficients for orders 0-3 originate from the SHORTEN codec:
// http://mi.eng.cam.ac.uk/reports/svr-ftp/auto-pdf/robinson_tr156.pdf page 4
// The coefficients for order 4 are undocumented in the original FLAC specification(s), but can now be found in
// https://datatracker.ietf.org/doc/html/draft-ietf-cellar-flac-03#section-10.2.5
// FIXME: Share this code with predict_fixed_lpc().
switch (subframe.order) {
case 0:
// s_0(t) = 0
for (u32 i = subframe.order; i < m_current_frame->sample_count; ++i)
decoded[i] += 0;
break;
case 1:
// s_1(t) = s(t-1)
for (u32 i = subframe.order; i < m_current_frame->sample_count; ++i)
decoded[i] += decoded[i - 1];
break;
case 2:
// s_2(t) = 2s(t-1) - s(t-2)
for (u32 i = subframe.order; i < m_current_frame->sample_count; ++i)
decoded[i] += 2 * decoded[i - 1] - decoded[i - 2];
break;
case 3:
// s_3(t) = 3s(t-1) - 3s(t-2) + s(t-3)
for (u32 i = subframe.order; i < m_current_frame->sample_count; ++i)
decoded[i] += 3 * decoded[i - 1] - 3 * decoded[i - 2] + decoded[i - 3];
break;
case 4:
// s_4(t) = 4s(t-1) - 6s(t-2) + 4s(t-3) - s(t-4)
for (u32 i = subframe.order; i < m_current_frame->sample_count; ++i)
decoded[i] += 4 * decoded[i - 1] - 6 * decoded[i - 2] + 4 * decoded[i - 3] - decoded[i - 4];
break;
default:
return LoaderError { LoaderError::Category::Format, static_cast<size_t>(m_current_sample_or_frame), ByteString::formatted("Unrecognized predictor order {}", subframe.order) };
}
return decoded;
}
// 11.30. RESIDUAL
// Decode the residual, the "error" between the function approximation and the actual audio data
MaybeLoaderError FlacLoaderPlugin::decode_residual(Vector<i64>& decoded, FlacSubframeHeader& subframe, BigEndianInputBitStream& bit_input)
{
// 11.30.1. RESIDUAL_CODING_METHOD
auto residual_mode = static_cast<FlacResidualMode>(TRY(bit_input.read_bits<u8>(2)));
u8 partition_order = TRY(bit_input.read_bits<u8>(4));
size_t partitions = 1 << partition_order;
dbgln_if(AFLACLOADER_DEBUG, " {}-bit Rice partitions, {} total (order {})", residual_mode == FlacResidualMode::Rice4Bit ? "4"sv : "5"sv, partitions, partition_order);
if (partitions > m_current_frame->sample_count)
return LoaderError { LoaderError::Category::Format, static_cast<size_t>(m_current_sample_or_frame), "Too many Rice partitions, each partition must contain at least one sample" };
// “The partition order MUST be such that the block size is evenly divisible by the number of partitions.”
// FIXME: Check “The partition order also MUST be such that the (block size >> partition order) is larger than the predictor order.”
if (m_current_frame->sample_count % partitions != 0)
return LoaderError { LoaderError::Category::Format, TRY(m_stream->tell()), "Block size is not evenly divisible by number of partitions" };
if (residual_mode == FlacResidualMode::Rice4Bit) {
// 11.30.2. RESIDUAL_CODING_METHOD_PARTITIONED_EXP_GOLOMB
// decode a single Rice partition with four bits for the order k
for (size_t i = 0; i < partitions; ++i) {
// FIXME: Write into the decode buffer directly.
auto rice_partition = TRY(decode_rice_partition(4, partitions, i, subframe, bit_input));
decoded.extend(move(rice_partition));
}
} else if (residual_mode == FlacResidualMode::Rice5Bit) {
// 11.30.3. RESIDUAL_CODING_METHOD_PARTITIONED_EXP_GOLOMB2
// five bits equivalent
for (size_t i = 0; i < partitions; ++i) {
// FIXME: Write into the decode buffer directly.
auto rice_partition = TRY(decode_rice_partition(5, partitions, i, subframe, bit_input));
decoded.extend(move(rice_partition));
}
2021-11-27 17:00:19 +01:00
} else
return LoaderError { LoaderError::Category::Format, static_cast<size_t>(m_current_sample_or_frame), "Reserved residual coding method" };
return {};
}
// 11.30.2.1. EXP_GOLOMB_PARTITION and 11.30.3.1. EXP_GOLOMB2_PARTITION
// Decode a single Rice partition as part of the residual, every partition can have its own Rice parameter k
ALWAYS_INLINE ErrorOr<Vector<i64>, LoaderError> FlacLoaderPlugin::decode_rice_partition(u8 partition_type, u32 partitions, u32 partition_index, FlacSubframeHeader& subframe, BigEndianInputBitStream& bit_input)
{
// 11.30.2.2. EXP GOLOMB PARTITION ENCODING PARAMETER and 11.30.3.2. EXP-GOLOMB2 PARTITION ENCODING PARAMETER
u8 k = TRY(bit_input.read_bits<u8>(partition_type));
u32 residual_sample_count;
if (partitions == 0)
residual_sample_count = m_current_frame->sample_count - subframe.order;
else
residual_sample_count = m_current_frame->sample_count / partitions;
if (partition_index == 0) {
if (subframe.order > residual_sample_count)
return LoaderError { LoaderError::Category::Format, static_cast<size_t>(m_current_sample_or_frame), "First Rice partition must advertise more residuals than LPC order" };
residual_sample_count -= subframe.order;
}
Vector<i64> rice_partition;
rice_partition.resize(residual_sample_count);
// escape code for unencoded binary partition
if (k == (1 << partition_type) - 1) {
u8 unencoded_bps = TRY(bit_input.read_bits<u8>(5));
if (unencoded_bps != 0) {
for (size_t r = 0; r < residual_sample_count; ++r) {
rice_partition[r] = sign_extend(TRY(bit_input.read_bits<u32>(unencoded_bps)), unencoded_bps);
}
}
} else {
for (size_t r = 0; r < residual_sample_count; ++r) {
rice_partition[r] = TRY(decode_unsigned_exp_golomb(k, bit_input));
}
}
return rice_partition;
}
// Decode a single number encoded with Rice/Exponential-Golomb encoding (the unsigned variant)
ALWAYS_INLINE ErrorOr<i32> decode_unsigned_exp_golomb(u8 k, BigEndianInputBitStream& bit_input)
{
u8 q = 0;
while (TRY(bit_input.read_bit()) == 0)
++q;
// least significant bits (remainder)
u32 rem = TRY(bit_input.read_bits<u32>(k));
u32 value = q << k | rem;
return rice_to_signed(value);
}
ErrorOr<u64> read_utf8_char(BigEndianInputBitStream& input)
{
u64 character;
u8 start_byte = TRY(input.read_value<u8>());
// Signal byte is zero: ASCII character
if ((start_byte & 0b10000000) == 0) {
return start_byte;
} else if ((start_byte & 0b11000000) == 0b10000000) {
return Error::from_string_literal("Illegal continuation byte");
}
// This algorithm supports the theoretical max 0xFF start byte, which is not part of the regular UTF-8 spec.
u8 length = 1;
while (((start_byte << length) & 0b10000000) == 0b10000000)
++length;
// This is technically not spec-compliant, but if we take UTF-8 to its logical extreme,
// we can say 0xFF means there's 7 following continuation bytes and no data at all in the leading character.
if (length == 8) [[unlikely]] {
character = 0;
} else {
u8 bits_from_start_byte = 8 - (length + 1);
u8 start_byte_bitmask = AK::exp2(bits_from_start_byte) - 1;
character = start_byte_bitmask & start_byte;
}
for (u8 i = length - 1; i > 0; --i) {
u8 current_byte = TRY(input.read_value<u8>());
character = (character << 6) | (current_byte & 0b00111111);
}
return character;
}
i64 sign_extend(u32 n, u8 size)
{
// negative
if ((n & (1 << (size - 1))) > 0) {
return static_cast<i64>(n | (0xffffffffffffffffLL << size));
}
// positive
return n;
}
i32 rice_to_signed(u32 x)
{
// positive numbers are even, negative numbers are odd
// bitmask for conditionally inverting the entire number, thereby "negating" it
i32 sign = -static_cast<i32>(x & 1);
// copies the sign's sign onto the actual magnitude of x
return static_cast<i32>(sign ^ (x >> 1));
}
}