From: Steinar H. Gunderson <sgunderson@bigfoot.com>
Date: Wed, 6 Mar 2019 19:07:39 +0000 (+0100)
Subject: Fix some unneeded copying of encoded JPEGs around (would take ~15% of uninterpolated... 
X-Git-Tag: 1.8.3~17
X-Git-Url: https://git.sesse.net/?p=nageru;a=commitdiff_plain;h=b57d57b88a1d7388c6eacf8cc0867f680123120a

Fix some unneeded copying of encoded JPEGs around (would take ~15% of uninterpolated export).
---

diff --git a/futatabi/video_stream.cpp b/futatabi/video_stream.cpp
index a820411..da57ebb 100644
--- a/futatabi/video_stream.cpp
+++ b/futatabi/video_stream.cpp
@@ -28,7 +28,7 @@ extern HTTPD *global_httpd;
 
 struct VectorDestinationManager {
 	jpeg_destination_mgr pub;
-	std::vector<uint8_t> dest;
+	string dest;
 
 	VectorDestinationManager()
 	{
@@ -62,7 +62,7 @@ struct VectorDestinationManager {
 	{
 		dest.resize(bytes_used + 4096);
 		dest.resize(dest.capacity());
-		pub.next_output_byte = dest.data() + bytes_used;
+		pub.next_output_byte = (uint8_t *)dest.data() + bytes_used;
 		pub.free_in_buffer = dest.size() - bytes_used;
 	}
 
@@ -78,7 +78,7 @@ struct VectorDestinationManager {
 };
 static_assert(std::is_standard_layout<VectorDestinationManager>::value, "");
 
-vector<uint8_t> encode_jpeg(const uint8_t *y_data, const uint8_t *cb_data, const uint8_t *cr_data, unsigned width, unsigned height)
+string encode_jpeg(const uint8_t *y_data, const uint8_t *cb_data, const uint8_t *cr_data, unsigned width, unsigned height)
 {
 	VectorDestinationManager dest;
 
@@ -659,15 +659,14 @@ void VideoStream::encode_thread_func()
 			pkt.size = jpeg.size();
 			pkt.flags = AV_PKT_FLAG_KEY;
 			mux->add_packet(pkt, qf.output_pts, qf.output_pts);
-
-			last_frame.assign(&jpeg[0], &jpeg[0] + jpeg.size());
+			last_frame = move(jpeg);
 		} else if (qf.type == QueuedFrame::FADED) {
 			glClientWaitSync(qf.fence.get(), /*flags=*/0, GL_TIMEOUT_IGNORED);
 
 			shared_ptr<Frame> frame = frame_from_pbo(qf.resources->pbo_contents, global_flags.width, global_flags.height);
 
 			// Now JPEG encode it, and send it on to the stream.
-			vector<uint8_t> jpeg = encode_jpeg(frame->y.get(), frame->cb.get(), frame->cr.get(), global_flags.width, global_flags.height);
+			string jpeg = encode_jpeg(frame->y.get(), frame->cb.get(), frame->cr.get(), global_flags.width, global_flags.height);
 
 			AVPacket pkt;
 			av_init_packet(&pkt);
@@ -687,7 +686,7 @@ void VideoStream::encode_thread_func()
 			}
 
 			// Now JPEG encode it, and send it on to the stream.
-			vector<uint8_t> jpeg = encode_jpeg(frame->y.get(), frame->cb.get(), frame->cr.get(), global_flags.width, global_flags.height);
+			string jpeg = encode_jpeg(frame->y.get(), frame->cb.get(), frame->cr.get(), global_flags.width, global_flags.height);
 			if (qf.flow_tex != 0) {
 				compute_flow->release_texture(qf.flow_tex);
 			}
diff --git a/futatabi/video_stream.h b/futatabi/video_stream.h
index 1b33587..26cb7c8 100644
--- a/futatabi/video_stream.h
+++ b/futatabi/video_stream.h
@@ -114,8 +114,6 @@ private:
 
 		// For original frames only. Made move-only so we know explicitly
 		// we don't copy these ~200 kB files around inadvertedly.
-		//
-		// TODO: Consider using vector<uint8_t> instead, so we save one copy.
 		std::unique_ptr<std::string> encoded_jpeg;
 
 		// For everything except original frames.
@@ -160,7 +158,7 @@ private:
 	GLuint last_flow_tex = 0;
 	FrameOnDisk last_frame1, last_frame2;
 
-	std::vector<uint8_t> last_frame;
+	std::string last_frame;
 };
 
 #endif  // !defined(_VIDEO_STREAM_H)