Bound the render-ahead so a demuxer cannot pace rendering
ci / ubuntu-latest (push) Failing after 14s
ci / macos-latest (push) Has been cancelled
ci / windows-latest (push) Has been cancelled

This commit is contained in:
TheM14 committed 2026-10-06 02:45:25 +08:00
1 parent 80543e6bd3
commit 30182ac670
4 files changed
+56 -10

No files matched your search

+9 -6
View File
@@ -7,10 +7,10 @@
#include "simd/cpu_probe.h"
#if defined(_M_X64)
#if defined(_M_X64) || defined(_M_IX86)
#include <immintrin.h>
#include <intrin.h>
#elif defined(__x86_64__)
#elif defined(__x86_64__) || defined(__i386__)
#include <cpuid.h>
#endif
@@ -21,10 +21,13 @@
namespace joc::simd {
namespace {
// ------------------------------------------------------------------- x86-64 --
#if defined(_M_X64) || defined(__x86_64__)
// ---------------------------------------------------------------------- x86 --
// Both pointer sizes are probed: AVX2 is not an x86-64-only ISA, and gating this
// 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)
#if defined(_M_X64) || defined(_M_IX86)
// CPUID tells us what the silicon can do; XCR0 tells us whether the OS saves the
// state the wider registers need. Both have to agree, or the first AVX
@@ -53,7 +56,7 @@ CpuFeatures probe_x86() noexcept {
return features;
}
#else // GCC/Clang on x86-64
#else // GCC/Clang on x86
// The compiler runtime performs the same CPUID + XGETBV probe (libgcc's cpuinfo
// checks XCR0 before it reports AVX), which keeps this file free of inline
+15 -3
View File
@@ -221,7 +221,7 @@ Status Stream::push_eac3(const std::uint8_t* data, std::size_t size, std::size_t
}
metadata_.push_back(entry);
}
return process_ready_frames();
return process_ready_frames(false);
}
Status Stream::push_bed(const float* interleaved6, std::size_t samples, std::size_t* consumed) {
@@ -235,7 +235,7 @@ Status Stream::push_bed(const float* interleaved6, std::size_t samples, std::siz
bed_pending_.insert(bed_pending_.end(), interleaved6,
interleaved6 + samples * kBedChannels);
}
return process_ready_frames();
return process_ready_frames(false);
}
Status Stream::push_objects16(const float* planar16, std::size_t samples, std::size_t* consumed) {
@@ -267,8 +267,13 @@ Status Stream::push_objects16(const float* planar16, std::size_t samples, std::s
return Status::success();
}
Status Stream::process_ready_frames() {
Status Stream::process_ready_frames(bool drain_all) {
while (bed_pending_.size() / kBedChannels >= kFrameSamples && !metadata_.empty()) {
// Stop before rendering what the caller is not about to take: the frames
// stay queued, in order, and are rendered by a later push or by flush().
if (!drain_all && buffered_samples() >= kMaxRenderAheadSamples) {
break;
}
const FrameMetadata entry = metadata_.front();
metadata_.pop_front();
@@ -429,6 +434,13 @@ 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:
// every frame still queued has to reach the renderer before its tail is
// drained, or the end of the file would be dropped.
const Status remaining = process_ready_frames(true);
if (!remaining.ok()) {
return remaining;
}
if (binaural_ready_) {
std::vector<double> tail;
const Status drained =
+11 -1
View File
@@ -85,8 +85,18 @@ public:
: 0u;
}
// A push renders every frame it makes ready, and the caller decides how far
// its demuxer runs ahead of playback. Without a bound, a demuxer that runs
// 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:
Status process_ready_frames();
// `drain_all` ignores kMaxRenderAheadSamples and renders every ready frame.
Status process_ready_frames(bool drain_all);
Status render_objects16(const std::vector<float>& objects16);
Status render_rosella_objects16(const std::vector<float>& objects16);
// Moves at most `limit` rendered stereo samples per channel out of the FIFO