#include "jpeg_frame_view.h"
-#include <jpeglib.h>
-#include <stdint.h>
-#include <unistd.h>
+#include "defs.h"
+#include "jpeg_destroyer.h"
+#include "post_to_main_thread.h"
+#include "video_stream.h"
+#include "ycbcr_converter.h"
+#include <QMouseEvent>
+#include <QScreen>
#include <atomic>
#include <condition_variable>
#include <deque>
+#include <jpeglib.h>
+#include <movit/init.h>
+#include <movit/resource_pool.h>
+#include <movit/util.h>
#include <mutex>
+#include <stdint.h>
#include <thread>
+#include <unistd.h>
#include <utility>
-#include <QMouseEvent>
-#include <QScreen>
-
-#include <movit/resource_pool.h>
-#include <movit/init.h>
-#include <movit/util.h>
-
-#include "defs.h"
-#include "post_to_main_thread.h"
+// Must come after the Qt stuff.
#include "vaapi_jpeg_decoder.h"
-#include "video_stream.h"
using namespace movit;
using namespace std;
+namespace {
+
// Just an arbitrary order for std::map.
struct JPEGIDLexicalOrder
{
{
if (a.stream_idx != b.stream_idx)
return a.stream_idx < b.stream_idx;
- if (a.pts != b.pts)
- return a.pts < b.pts;
- return a.interpolated < b.interpolated;
+ return a.pts < b.pts;
}
};
+inline size_t frame_size(const Frame &frame)
+{
+ size_t y_size = frame.width * frame.height;
+ size_t cbcr_size = y_size / frame.chroma_subsampling_x / frame.chroma_subsampling_y;
+ return y_size + cbcr_size * 2;
+}
+
struct LRUFrame {
shared_ptr<Frame> frame;
size_t last_used;
};
+struct PendingDecode {
+ JPEGFrameView *destination;
+
+ // For actual decodes (only if frame below is nullptr).
+ JPEGID primary, secondary;
+ float fade_alpha; // Irrelevant if secondary.stream_idx == -1.
+
+ // Already-decoded frames are also sent through PendingDecode,
+ // so that they get drawn in the right order. If frame is nullptr,
+ // it's a real decode.
+ shared_ptr<Frame> frame;
+};
+
+} // namespace
+
+thread JPEGFrameView::jpeg_decoder_thread;
mutex cache_mu;
map<JPEGID, LRUFrame, JPEGIDLexicalOrder> cache; // Under cache_mu.
-condition_variable any_pending_decodes, cache_updated;
-deque<pair<JPEGID, JPEGFrameView *>> pending_decodes; // Under cache_mu.
+size_t cache_bytes_used = 0; // Under cache_mu.
+condition_variable any_pending_decodes;
+deque<PendingDecode> pending_decodes; // Under cache_mu.
atomic<size_t> event_counter{0};
extern QGLWidget *global_share_widget;
+extern atomic<bool> should_quit;
shared_ptr<Frame> decode_jpeg(const string &filename)
{
jpeg_error_mgr jerr;
dinfo.err = jpeg_std_error(&jerr);
jpeg_create_decompress(&dinfo);
+ JPEGDestroyer destroy_dinfo(&dinfo);
FILE *fp = fopen(filename.c_str(), "rb");
if (fp == nullptr) {
jpeg_read_raw_data(&dinfo, data, v_mcu_size);
}
- (void) jpeg_finish_decompress(&dinfo);
- jpeg_destroy_decompress(&dinfo);
+ (void)jpeg_finish_decompress(&dinfo);
fclose(fp);
return frame;
void prune_cache()
{
// Assumes cache_mu is held.
- vector<size_t> lru_timestamps;
+ int64_t bytes_still_to_remove = cache_bytes_used - (size_t(CACHE_SIZE_MB) * 1024 * 1024) * 9 / 10;
+ if (bytes_still_to_remove <= 0) return;
+
+ vector<pair<size_t, size_t>> lru_timestamps_and_size;
for (const auto &key_and_value : cache) {
- lru_timestamps.push_back(key_and_value.second.last_used);
+ lru_timestamps_and_size.emplace_back(
+ key_and_value.second.last_used,
+ frame_size(*key_and_value.second.frame));
+ }
+ sort(lru_timestamps_and_size.begin(), lru_timestamps_and_size.end());
+
+ // Remove the oldest ones until we are below 90% of the cache used.
+ size_t lru_cutoff_point = 0;
+ for (const pair<size_t, size_t> &it : lru_timestamps_and_size) {
+ lru_cutoff_point = it.first;
+ bytes_still_to_remove -= it.second;
+ if (bytes_still_to_remove <= 0) break;
}
- size_t cutoff_point = CACHE_SIZE / 10; // Prune away the 10% oldest ones.
- nth_element(lru_timestamps.begin(), lru_timestamps.begin() + cutoff_point, lru_timestamps.end());
- size_t must_be_used_after = lru_timestamps[cutoff_point];
for (auto it = cache.begin(); it != cache.end(); ) {
- if (it->second.last_used < must_be_used_after) {
+ if (it->second.last_used <= lru_cutoff_point) {
+ cache_bytes_used -= frame_size(*it->second.frame);
it = cache.erase(it);
} else {
++it;
return nullptr;
}
- assert(!id.interpolated);
*did_decode = true;
shared_ptr<Frame> frame = decode_jpeg(filename_for_frame(id.stream_idx, id.pts));
unique_lock<mutex> lock(cache_mu);
+ cache_bytes_used += frame_size(*frame);
cache[id] = LRUFrame{ frame, event_counter++ };
- if (cache.size() > CACHE_SIZE) {
+ if (cache_bytes_used > size_t(CACHE_SIZE_MB) * 1024 * 1024) {
prune_cache();
}
return frame;
}
-void jpeg_decoder_thread()
+void jpeg_decoder_thread_func()
{
size_t num_decoded = 0, num_dropped = 0;
pthread_setname_np(pthread_self(), "JPEGDecoder");
- for ( ;; ) {
- JPEGID id;
- JPEGFrameView *dest;
+ while (!should_quit.load()) {
+ PendingDecode decode;
CacheMissBehavior cache_miss_behavior = DECODE_IF_NOT_IN_CACHE;
{
unique_lock<mutex> lock(cache_mu); // TODO: Perhaps under another lock?
any_pending_decodes.wait(lock, [] {
- return !pending_decodes.empty();
+ return !pending_decodes.empty() || should_quit.load();
});
- id = pending_decodes.front().first;
- dest = pending_decodes.front().second;
+ if (should_quit.load())
+ break;
+ decode = pending_decodes.front();
pending_decodes.pop_front();
size_t num_pending = 0;
- for (const pair<JPEGID, JPEGFrameView *> &decode : pending_decodes) {
- if (decode.second == dest) {
+ for (const PendingDecode &other_decode : pending_decodes) {
+ if (other_decode.destination == decode.destination) {
++num_pending;
}
}
}
}
- bool found_in_cache;
- shared_ptr<Frame> frame;
- if (id.interpolated) {
- // Interpolated frames are never decoded by us,
- // put directly into the cache from VideoStream.
- unique_lock<mutex> lock(cache_mu);
- cache_updated.wait(lock, [id] {
- return cache.count(id) != 0;
- });
- found_in_cache = true; // Don't count it as a decode.
+ if (decode.frame != nullptr) {
+ // Already decoded, so just show it.
+ decode.destination->setDecodedFrame(decode.frame, nullptr, 1.0f);
+ continue;
+ }
+
+ shared_ptr<Frame> primary_frame, secondary_frame;
+ bool drop = false;
+ for (int subframe_idx = 0; subframe_idx < 2; ++subframe_idx) {
+ const JPEGID &id = (subframe_idx == 0 ? decode.primary : decode.secondary);
+ if (id.stream_idx == (unsigned)-1) {
+ // No secondary frame.
+ continue;
+ }
- auto it = cache.find(id);
- assert(it != cache.end());
+ bool found_in_cache;
+ shared_ptr<Frame> frame = decode_jpeg_with_cache(id, cache_miss_behavior, &found_in_cache);
- it->second.last_used = event_counter++;
- frame = it->second.frame;
if (frame == nullptr) {
- // We inserted a nullptr as signal that the frame was never
- // interpolated and that we should stop waiting.
- // But don't let it linger in the cache anymore.
- cache.erase(it);
+ assert(cache_miss_behavior == RETURN_NULLPTR_IF_NOT_IN_CACHE);
+ drop = true;
+ break;
}
- } else {
- frame = decode_jpeg_with_cache(id, cache_miss_behavior, &found_in_cache);
- }
- if (frame == nullptr) {
- assert(id.interpolated || cache_miss_behavior == RETURN_NULLPTR_IF_NOT_IN_CACHE);
+ if (!found_in_cache) {
+ ++num_decoded;
+ if (num_decoded % 1000 == 0) {
+ fprintf(stderr, "Decoded %zu images, dropped %zu (%.2f%% dropped)\n",
+ num_decoded, num_dropped, (100.0 * num_dropped) / (num_decoded + num_dropped));
+ }
+ }
+ if (subframe_idx == 0) {
+ primary_frame = move(frame);
+ } else {
+ secondary_frame = move(frame);
+ }
+ }
+ if (drop) {
++num_dropped;
continue;
}
- if (!found_in_cache) {
- ++num_decoded;
- if (num_decoded % 1000 == 0) {
- fprintf(stderr, "Decoded %zu images, dropped %zu (%.2f%% dropped)\n",
- num_decoded, num_dropped, (100.0 * num_dropped) / (num_decoded + num_dropped));
- }
- }
-
// TODO: Could we get jitter between non-interpolated and interpolated frames here?
- dest->setDecodedFrame(frame);
+ decode.destination->setDecodedFrame(primary_frame, secondary_frame, decode.fade_alpha);
}
}
+void JPEGFrameView::shutdown()
+{
+ any_pending_decodes.notify_all();
+ jpeg_decoder_thread.join();
+}
+
JPEGFrameView::JPEGFrameView(QWidget *parent)
- : QGLWidget(parent, global_share_widget) {
+ : QGLWidget(parent, global_share_widget)
+{
}
-void JPEGFrameView::setFrame(unsigned stream_idx, int64_t pts, bool interpolated)
+void JPEGFrameView::setFrame(unsigned stream_idx, int64_t pts, int secondary_stream_idx, int64_t secondary_pts, float fade_alpha)
{
- current_stream_idx = stream_idx;
+ if (secondary_stream_idx != -1) assert(secondary_pts != -1);
+ current_stream_idx = stream_idx; // TODO: Does this interact with fades?
unique_lock<mutex> lock(cache_mu);
- pending_decodes.emplace_back(JPEGID{ stream_idx, pts, interpolated }, this);
+ PendingDecode decode;
+ decode.primary = JPEGID{ stream_idx, pts };
+ decode.secondary = JPEGID{ (unsigned)secondary_stream_idx, secondary_pts };
+ decode.fade_alpha = fade_alpha;
+ decode.destination = this;
+ pending_decodes.push_back(decode);
any_pending_decodes.notify_all();
}
-void JPEGFrameView::insert_interpolated_frame(unsigned stream_idx, int64_t pts, shared_ptr<Frame> frame)
+void JPEGFrameView::setFrame(shared_ptr<Frame> frame)
{
- JPEGID id{ stream_idx, pts, true };
-
- // We rely on the frame not being evicted from the cache before
- // jpeg_decoder_thread() sees it and can display it (otherwise,
- // that thread would hang). With a default cache of 1000 elements,
- // that would sound like a reasonable assumption.
unique_lock<mutex> lock(cache_mu);
- cache[id] = LRUFrame{ std::move(frame), event_counter++ };
- cache_updated.notify_all();
+ PendingDecode decode;
+ decode.frame = std::move(frame);
+ decode.destination = this;
+ pending_decodes.push_back(decode);
+ any_pending_decodes.notify_all();
}
ResourcePool *resource_pool = nullptr;
static once_flag once;
call_once(once, [] {
resource_pool = new ResourcePool;
- std::thread(&jpeg_decoder_thread).detach();
+ jpeg_decoder_thread = std::thread(jpeg_decoder_thread_func);
});
+ ycbcr_converter.reset(new YCbCrConverter(YCbCrConverter::OUTPUT_TO_RGBA, resource_pool));
+
ImageFormat inout_format;
inout_format.color_space = COLORSPACE_sRGB;
inout_format.gamma_curve = GAMMA_sRGB;
- ycbcr_format.luma_coefficients = YCBCR_REC_709;
- ycbcr_format.full_range = false;
- ycbcr_format.num_levels = 256;
- ycbcr_format.chroma_subsampling_x = 2;
- ycbcr_format.chroma_subsampling_y = 1;
- ycbcr_format.cb_x_position = 0.0f; // H.264 -- _not_ JPEG, even though our input is MJPEG-encoded
- ycbcr_format.cb_y_position = 0.5f; // Irrelevant.
- ycbcr_format.cr_x_position = 0.0f;
- ycbcr_format.cr_y_position = 0.5f;
-
- // Planar Y'CbCr decoding chain.
- planar_chain.reset(new EffectChain(1280, 720, resource_pool));
- ycbcr_planar_input = (movit::YCbCrInput *)planar_chain->add_input(new YCbCrInput(inout_format, ycbcr_format, 1280, 720, YCBCR_INPUT_PLANAR));
- planar_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
- planar_chain->set_dither_bits(8);
- planar_chain->finalize();
-
- // Semiplanar Y'CbCr decoding chain (for images coming from VA-API).
- semiplanar_chain.reset(new EffectChain(1280, 720, resource_pool));
- ycbcr_semiplanar_input = (movit::YCbCrInput *)semiplanar_chain->add_input(new YCbCrInput(inout_format, ycbcr_format, 1280, 720, YCBCR_INPUT_SPLIT_Y_AND_CBCR));
- semiplanar_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
- semiplanar_chain->set_dither_bits(8);
- semiplanar_chain->finalize();
-
overlay_chain.reset(new EffectChain(overlay_base_width, overlay_base_height, resource_pool));
overlay_input = (movit::FlatInput *)overlay_chain->add_input(new FlatInput(inout_format, FORMAT_GRAYSCALE, GL_UNSIGNED_BYTE, overlay_base_width, overlay_base_height));
}
check_error();
- if (current_frame->is_semiplanar) {
- semiplanar_chain->render_to_screen();
- } else {
- planar_chain->render_to_screen();
- }
+ current_chain->render_to_screen();
if (overlay_image != nullptr) {
if (overlay_input_needs_refresh) {
}
}
-void JPEGFrameView::setDecodedFrame(std::shared_ptr<Frame> frame)
+namespace {
+
+} // namespace
+
+void JPEGFrameView::setDecodedFrame(shared_ptr<Frame> frame, shared_ptr<Frame> secondary_frame, float fade_alpha)
{
- post_to_main_thread([this, frame] {
+ post_to_main_thread([this, frame, secondary_frame, fade_alpha] {
current_frame = frame;
- ycbcr_format.chroma_subsampling_x = frame->chroma_subsampling_x;
- ycbcr_format.chroma_subsampling_y = frame->chroma_subsampling_y;
-
- if (frame->is_semiplanar) {
- ycbcr_semiplanar_input->change_ycbcr_format(ycbcr_format);
- ycbcr_semiplanar_input->set_width(frame->width);
- ycbcr_semiplanar_input->set_height(frame->height);
- ycbcr_semiplanar_input->set_pixel_data(0, frame->y.get());
- ycbcr_semiplanar_input->set_pixel_data(1, frame->cbcr.get());
- ycbcr_semiplanar_input->set_pitch(0, frame->pitch_y);
- ycbcr_semiplanar_input->set_pitch(1, frame->pitch_chroma);
+ current_secondary_frame = secondary_frame;
+
+ if (secondary_frame != nullptr) {
+ current_chain = ycbcr_converter->prepare_chain_for_fade(frame, secondary_frame, fade_alpha);
} else {
- ycbcr_planar_input->change_ycbcr_format(ycbcr_format);
- ycbcr_planar_input->set_width(frame->width);
- ycbcr_planar_input->set_height(frame->height);
- ycbcr_planar_input->set_pixel_data(0, frame->y.get());
- ycbcr_planar_input->set_pixel_data(1, frame->cb.get());
- ycbcr_planar_input->set_pixel_data(2, frame->cr.get());
- ycbcr_planar_input->set_pitch(0, frame->pitch_y);
- ycbcr_planar_input->set_pitch(1, frame->pitch_chroma);
- ycbcr_planar_input->set_pitch(2, frame->pitch_chroma);
+ current_chain = ycbcr_converter->prepare_chain_for_conversion(frame);
}
update();
});