Update the vendored kernel to the bounded objects16 render-ahead

kernel/include/joc_stream.h

@@ -39,12 +39,12 @@ typedef struct joc_stream joc_stream;
 
 typedef enum joc_stream_input {
     JOC_STREAM_IN_EAC3 = 0,          /* bare E-AC-3 syncframes (the metadata stream) */
-    JOC_STREAM_IN_PCM_OBJECTS16 = 1, /* 16-channel objects16, decoded by the host   */
+    JOC_STREAM_IN_PCM_OBJECTS16 = 1, /* objects16 frames, planar [16][1536] each    */
     JOC_STREAM_IN_CORE_PCM = 3       /* the 5.1 core PCM of the pushed E-AC-3 frames */
 } joc_stream_input;
 
 typedef enum joc_stream_output {
-    JOC_STREAM_OUT_PCM_OBJECTS16 = 0, /* planar [16][samples] float32        */
+    JOC_STREAM_OUT_PCM_OBJECTS16 = 0, /* objects16 frames, planar [16][1536] each    */
     JOC_STREAM_OUT_SPEAKER = 1,       /* interleaved [samples][channels] f32         */
     JOC_STREAM_OUT_BINAURAL = 2       /* interleaved [samples][2] f32                */
 } joc_stream_output;

kernel/joc_kernel.vcxproj

@@ -121,11 +121,9 @@
 
       * x64 gets the AVX2 and AVX-512 units, JOC_SIMD_HAVE_SSE2 / _AVX2 /
         _AVX512, and a per-file /arch for those two files only;
-      * Win32 (x86) gets the AVX2 unit and JOC_SIMD_HAVE_AVX2.  AVX2 is not an
+      * Win32 (x86) gets the AVX2 unit and JOC_SIMD_HAVE_AVX2: AVX2 is not an
-        x86-64-only ISA and MSVC accepts /arch:AVX2 for x86, so gating it on the
+        x86-64-only ISA.  AVX-512 stays x64-only, because 32-bit mode addresses
-        pointer size left the 32-bit component on the scalar reference, which is
+        ZMM0-7 only.
-        too slow to hold a 4096-frame read inside its own 85.3 ms of audio.
-        AVX-512 stays x64-only: 32-bit mode addresses ZMM0-7 only.
       * src\simd\kernels_intrin_neon.cpp is excluded everywhere here: the CMake
         build lists it only for aarch64 (CMakeLists.txt lines 177-180), which no
         configuration of this project targets.

kernel/src/simd/cpu_probe.cpp

@@ -22,9 +22,7 @@ namespace joc::simd {
 namespace {
 
 // ---------------------------------------------------------------------- x86 --
-// Both pointer sizes are probed: AVX2 is not an x86-64-only ISA, and gating this
+// AVX2 is not an x86-64-only ISA, so both pointer sizes are probed.
-// on _M_X64 / __x86_64__ left every 32-bit x86 build reporting "no features",
-// which pinned the dispatcher to the scalar kernels.
 #if defined(_M_X64) || defined(_M_IX86) || defined(__x86_64__) || defined(__i386__)
 
 #if defined(_M_X64) || defined(_M_IX86)

kernel/src/stream/stream.cpp

@@ -38,7 +38,10 @@ void Stream::reset_state() {
     reader_ = eac3::FrameReader();
     metadata_.clear();
     bed_pending_.clear();
+    bed_read_offset_ = 0;
     objects16_.clear();
+    objects_pending_.clear();
+    objects_read_offset_ = 0;
     output_.clear();
     read_offset_ = 0;
     info_ = Info();
@@ -245,54 +248,82 @@ Status Stream::push_objects16(const float* planar16, std::size_t samples, std::s
     if (planar16 == nullptr || samples == 0u) {
         return Status::success();
     }
-    // Rendered immediately: the host has already done the JOC rebuild.
+    // Queued as whole frames, each planar [16][kFrameSamples]: the shape the
+    // objects16 output writes, so a host can feed a batch back unchanged.
+    constexpr std::size_t kFrameValues =
+        static_cast<std::size_t>(JOC_OUTPUT_CHANNELS) * kFrameSamples;
     for (std::size_t offset = 0; offset < samples; offset += kFrameSamples) {
         const std::size_t count = std::min(kFrameSamples, samples - offset);
-        std::vector<float> frame(static_cast<std::size_t>(JOC_OUTPUT_CHANNELS) * kFrameSamples,
+        const std::size_t base = objects_pending_.size();
-                                 0.0f);
+        const float* frame = planar16 + (offset / kFrameSamples) * kFrameValues;
+        objects_pending_.resize(base + kFrameValues, 0.0f);
         for (std::size_t channel = 0; channel < JOC_OUTPUT_CHANNELS; ++channel) {
-            std::memcpy(frame.data() + channel * kFrameSamples,
+            std::memcpy(objects_pending_.data() + base + channel * kFrameSamples,
-                        planar16 + channel * samples + offset, count * sizeof(float));
+                        frame + channel * kFrameSamples, count * sizeof(float));
         }
-        ++info_.frames_in;
+    }
-        info_.samples_in += count;
+    return process_objects16_frames(false);
-        const Status rendered = render_objects16(frame);
+}
+
+Status Stream::process_objects16_frames(bool drain_all) {
+    constexpr std::size_t kFrameValues =
+        static_cast<std::size_t>(JOC_OUTPUT_CHANNELS) * kFrameSamples;
+    while (objects_pending_.size() - objects_read_offset_ >= kFrameValues) {
+        // Leave the rest queued, in order, for a later push or for flush().
+        if (!drain_all && buffered_samples() >= kMaxRenderAheadSamples) {
+            break;
+        }
+        const auto first = objects_pending_.begin() +
+                           static_cast<std::ptrdiff_t>(objects_read_offset_);
+        objects_frame_.assign(first, first + static_cast<std::ptrdiff_t>(kFrameValues));
+        objects_read_offset_ += kFrameValues;
+        const Status rendered = render_objects16(objects_frame_);
         if (!rendered.ok()) {
             return rendered;
         }
-        if (count != kFrameSamples) {
+        ++info_.frames_in;
-            break;  // a partial frame is dropped; the host should push whole frames
+        info_.samples_in += kFrameSamples;
     }
+    if (objects_read_offset_ == objects_pending_.size()) {
+        objects_pending_.clear();
+        objects_read_offset_ = 0;
+    } else if (objects_read_offset_ >= (1u << 20)) {
+        // Erasing from the front moves the remainder, so it is only worth doing
+        // once the consumed prefix is large enough to pay for the move.
+        objects_pending_.erase(
+            objects_pending_.begin(),
+            objects_pending_.begin() + static_cast<std::ptrdiff_t>(objects_read_offset_));
+        objects_read_offset_ = 0;
     }
     return Status::success();
 }
 
 Status Stream::process_ready_frames(bool drain_all) {
-    while (bed_pending_.size() / kBedChannels >= kFrameSamples && !metadata_.empty()) {
+    while (bed_pending_.size() - bed_read_offset_ >= kFrameSamples * kBedChannels &&
-        // Stop before rendering what the caller is not about to take: the frames
+           !metadata_.empty()) {
-        // stay queued, in order, and are rendered by a later push or by flush().
+        // Leave the rest queued, in order, for a later push or for flush().
         if (!drain_all && buffered_samples() >= kMaxRenderAheadSamples) {
             break;
         }
         const FrameMetadata entry = metadata_.front();
         metadata_.pop_front();
 
-        std::vector<float> bed5(static_cast<std::size_t>(JOC_CORE_CHANNELS) * kFrameSamples, 0.0f);
+        const float* bed = bed_pending_.data() + bed_read_offset_;
-        std::vector<float> lfe(kFrameSamples, 0.0f);
+        bed5_.resize(static_cast<std::size_t>(JOC_CORE_CHANNELS) * kFrameSamples);
+        lfe_.resize(kFrameSamples);
         for (std::size_t sample = 0; sample < kFrameSamples; ++sample) {
             for (std::size_t channel = 0; channel < JOC_CORE_CHANNELS; ++channel) {
-                bed5[channel * kFrameSamples + sample] =
+                bed5_[channel * kFrameSamples + sample] =
-                    bed_pending_[sample * kBedChannels + kCoreChannels[channel]];
+                    bed[sample * kBedChannels + kCoreChannels[channel]];
             }
-            lfe[sample] = bed_pending_[sample * kBedChannels + kLfeChannel];
+            lfe_[sample] = bed[sample * kBedChannels + kLfeChannel];
         }
-        bed_pending_.erase(bed_pending_.begin(),
+        bed_read_offset_ += kFrameSamples * kBedChannels;
-                           bed_pending_.begin() + static_cast<std::ptrdiff_t>(kFrameSamples *
+        compact_bed_pending();
-                                                                             kBedChannels));
 
         std::string error;
-        const Status rebuilt = joc::rebuild_objects16(rebuilder_, entry.params, bed5.data(),
+        const Status rebuilt = joc::rebuild_objects16(rebuilder_, entry.params, bed5_.data(),
-                                                      lfe.data(), gain_, &objects16_, &error);
+                                                      lfe_.data(), gain_, &objects16_, &error);
         if (!rebuilt.ok()) {
             return Status::fail(rebuilt.code(), stage::kDsp, error);
         }
@@ -307,6 +338,22 @@ Status Stream::process_ready_frames(bool drain_all) {
     return Status::success();
 }
 
+void Stream::compact_bed_pending() {
+    if (bed_read_offset_ == 0) {
+        return;
+    }
+    if (bed_read_offset_ == bed_pending_.size()) {
+        bed_pending_.clear();
+        bed_read_offset_ = 0;
+    } else if (bed_read_offset_ >= (1u << 20)) {
+        // Erasing from the front moves the remainder, so it is only worth doing
+        // once the consumed prefix is large enough to pay for the move.
+        bed_pending_.erase(bed_pending_.begin(),
+                           bed_pending_.begin() + static_cast<std::ptrdiff_t>(bed_read_offset_));
+        bed_read_offset_ = 0;
+    }
+}
+
 Status Stream::render_objects16(const std::vector<float>& objects16) {
     if (config_.output == JOC_STREAM_OUT_PCM_OBJECTS16) {
         output_.insert(output_.end(), objects16.begin(), objects16.end());
@@ -324,6 +371,7 @@ Status Stream::render_objects16(const std::vector<float>& objects16) {
         if (!stepped.ok()) {
             return Status::fail(stepped.code(), stage::kRender, error);
         }
+        output_.reserve(output_.size() + speaker_.output.size());
         for (const double value : speaker_.output) {
             output_.push_back(static_cast<float>(value));
         }
@@ -344,13 +392,13 @@ Status Stream::render_objects16(const std::vector<float>& objects16) {
     if (!submitted.ok()) {
         return submitted;
     }
-    std::vector<double> produced;
+    binaural_.take_output(&produced_);
-    binaural_.take_output(&produced);
+    output_.reserve(output_.size() + produced_.size());
-    for (const double value : produced) {
+    for (const double value : produced_) {
         output_.push_back(static_cast<float>(value));
     }
     info_.frames_out++;
-    info_.samples_out += produced.size() / 2u;
+    info_.samples_out += produced_.size() / 2u;
     return Status::success();
 }
 
@@ -371,16 +419,15 @@ Status Stream::render_rosella_objects16(const std::vector<float>& objects16) {
     if (!submitted.ok()) {
         return submitted;
     }
-    std::vector<double> produced;
+    rosella_.take_output(&produced_);
-    rosella_.take_output(&produced);
+    if (!produced_.empty()) {
-    if (!produced.empty()) {
+        rosella_pending_.insert(rosella_pending_.end(), produced_.begin(), produced_.end());
-        rosella_pending_.insert(rosella_pending_.end(), produced.begin(), produced.end());
     }
     release_rosella_output(kFrameSamples);
     info_.frames_out++;
     // Counted as the runtime produces it, which is also how the SOFA path counts:
     // the totals are identical, only the frame they appear on differs.
-    info_.samples_out += produced.size() / 2u;
+    info_.samples_out += produced_.size() / 2u;
     return Status::success();
 }
 
@@ -393,6 +440,7 @@ void Stream::release_rosella_output(std::size_t limit) {
         return;
     }
     const std::size_t values = count * 2u;
+    output_.reserve(output_.size() + values);
     for (std::size_t index = 0; index < values; ++index) {
         output_.push_back(static_cast<float>(rosella_pending_[rosella_read_offset_ + index]));
     }
@@ -434,13 +482,16 @@ Status Stream::pull(float* destination, std::size_t capacity_samples, std::size_
 }
 
 Status Stream::flush() {
-    // Input has ended, so the render-ahead bound has nothing left to wait for:
+    // Input has ended, so drain what the cap held back: nothing else will
-    // every frame still queued has to reach the renderer before its tail is
+    // trigger rendering.
-    // drained, or the end of the file would be dropped.
     const Status remaining = process_ready_frames(true);
     if (!remaining.ok()) {
         return remaining;
     }
+    const Status objects = process_objects16_frames(true);
+    if (!objects.ok()) {
+        return objects;
+    }
     if (binaural_ready_) {
         std::vector<double> tail;
         const Status drained =
@@ -448,6 +499,7 @@ Status Stream::flush() {
         if (!drained.ok()) {
             return drained;
         }
+        output_.reserve(output_.size() + tail.size());
         for (const double value : tail) {
             output_.push_back(static_cast<float>(value));
         }
@@ -464,6 +516,7 @@ Status Stream::flush() {
         // tail only sounds after it.  The program samples were already counted by
         // render_rosella_objects16, so only the tail is added here.
         release_rosella_output(rosella_pending_samples());
+        output_.reserve(output_.size() + tail.size());
         for (const double value : tail) {
             output_.push_back(static_cast<float>(value));
         }

kernel/src/stream/stream.h

@@ -85,13 +85,8 @@ public:
                    : 0u;
     }
 
-    // A push renders every frame it makes ready, and the caller decides how far
+    // Cap on rendered samples that have not been pulled.  A push renders what it
-    // its demuxer runs ahead of playback.  Without a bound, a demuxer that runs
+    // makes ready, so a caller that feeds faster than it pulls renders ahead.
-    // far ahead turns its whole read-ahead burst into latency on whichever pull()
-    // happens to follow it: the samples are not wasted, but they are rendered at
-    // the worst possible moment.  Rendering therefore stops once this many
-    // samples are rendered and unpulled; flush() lifts the bound so the frames
-    // still waiting when the input ends are drained rather than dropped.
     static constexpr std::size_t kMaxRenderAheadSamples = 16384;
 
 private:
@@ -106,6 +101,10 @@ private:
         return rosella_ready_ ? (rosella_pending_.size() - rosella_read_offset_) / 2u : 0u;
     }
     void reset_state();
+    // Drops the bed samples that have already been rendered, keeping the rest.
+    void compact_bed_pending();
+    // Renders the queued objects16 frames, bounded by kMaxRenderAheadSamples.
+    Status process_objects16_frames(bool drain_all);
 
     Config config_;
     Info info_;
@@ -113,7 +112,12 @@ private:
     std::deque<FrameMetadata> metadata_;
     FrameMetadata pending_metadata_;
     std::vector<float> bed_pending_;
-    std::vector<std::uint8_t> frame_copy_;
+    std::size_t bed_read_offset_ = 0;
+    std::vector<float> bed5_;
+    std::vector<float> lfe_;
+    std::vector<float> objects_pending_;
+    std::size_t objects_read_offset_ = 0;
+    std::vector<float> objects_frame_;
     std::vector<float> objects16_;
     std::vector<float> output_;
     std::size_t read_offset_ = 0;
@@ -125,6 +129,7 @@ private:
     hrtf::RosellaRuntime rosella_;
     std::vector<double> rosella_pending_;
     std::size_t rosella_read_offset_ = 0;
+    std::vector<double> produced_;
     bool speaker_enabled_ = false;
     bool binaural_enabled_ = false;
     bool binaural_ready_ = false;

src/joc_decode.cpp

@@ -20,15 +20,10 @@ namespace joc_decode {
 namespace {
 
 constexpr std::size_t kEac3Chunk = 96u * 1024u;      // bytes read per push
-// The core renders every frame it is handed, and it renders it during the push,
+// The core renders during the push, so a read queues about as much as it is
-// so a read must not queue more frames than the caller is about to take: a 96 KB
+// about to consume rather than everything it just read.
-// read is around thirty syncframes, i.e. a second of audio rendered to satisfy
-// one 4096-frame read.  This is that read (2.67 syncframes) rounded up, so each
-// read hands the core about as much as it is about to consume.
 constexpr std::uint64_t kEac3FramesPerRead = 3u;
-// Rendered audio the caller has not taken yet.  Once this much is waiting there
+// Rendered audio the caller has not taken yet.
-// is nothing to gain from queueing more input: the core would render it now and
-// the caller would not ask for it for several more reads.
 constexpr std::size_t kMaxRenderedAheadSamples = 4096u;
 constexpr std::size_t kBedFramesChunk = 8192u;       // staging capacity, in frames
 constexpr std::size_t kBedChannels = 6;              // ffmpeg -ac 6
@@ -984,11 +979,7 @@ std::size_t Engine::read(float* destination, std::size_t frames, std::string* er
             impl.eac3_eof = true;
         }
 
-        // The core renders during the push, so input is queued only while the
+        // Queue more input only while less than one read's worth is waiting.
-        // caller still has less than one read's worth of rendered audio waiting.
-        // Pushing past that is what turns a single read into a second of work:
-        // the samples are rendered early rather than wrongly, and the read that
-        // pays for them overruns its own audio.
         std::size_t rendered_ahead = 0;
         {
             joc_stream_status_info pending{};
@@ -1049,13 +1040,9 @@ std::size_t Engine::read(float* destination, std::size_t frames, std::string* er
                     offset += bytes;
                     ++complete;
                 }
-                // The chunk is cut at a frame boundary for two independent
+                // Both cuts land on a frame boundary because the renderer's output
-                // reasons: a configured input limit has to stop exactly where it
+                // depends on how many frames it was given: the input limit has to
-                // says, and a read must not queue more frames than it is about to
+                // stop where it says, and a read queues only its own budget.
-                // consume.  Both cuts land on a frame boundary because the
-                // renderer's output depends on how many frames it was given, so a
-                // cut that overshoots would not reproduce a run that stopped
-                // earlier.
                 std::uint64_t allowed = complete;
                 if (impl.settings.input_frame_limit != 0) {
                     const std::uint64_t room =
@@ -1096,10 +1083,8 @@ std::size_t Engine::read(float* destination, std::size_t frames, std::string* er
                     std::memmove(impl.eac3_buffer.data(), impl.eac3_buffer.data() + push_bytes,
                                  impl.eac3_carry);
                 }
-                // The pipe can end while complete syncframes are still waiting in
+                // The pipe can end with complete syncframes still buffered; they
-                // the buffer: they are queued by the next pass, so the input is
+                // are queued next, so end of input is only real once none remain.
-                // only over once nothing but a partial frame is left.  Ending here
-                // instead would drop them, and with them the end of the file.
                 if (got == 0 && pushed_frames >= complete) impl.eac3_eof = true;
             }
         }