From: Steinar H. Gunderson <sgunderson@bigfoot.com>
Date: Wed, 12 Sep 2018 22:48:01 +0000 (+0200)
Subject: Release flow textures when we are done with them.
X-Git-Tag: 1.8.0~76^2~109
X-Git-Url: https://git.sesse.net/?p=nageru;a=commitdiff_plain;h=00a917a3feec0fdd99db1a58a7d1edf973a0eb57

Release flow textures when we are done with them.
---

diff --git a/flow.cpp b/flow.cpp
index 873bdf2..ece0d51 100644
--- a/flow.cpp
+++ b/flow.cpp
@@ -1019,11 +1019,14 @@ GLuint Interpolate::exec(GLuint image_tex, GLuint bidirectional_flow_tex, GLuint
 
 GLuint TexturePool::get_texture(GLenum format, GLuint width, GLuint height, GLuint num_layers)
 {
-	for (Texture &tex : textures) {
-		if (!tex.in_use && !tex.is_renderbuffer && tex.format == format &&
-		    tex.width == width && tex.height == height && tex.num_layers == num_layers) {
-			tex.in_use = true;
-			return tex.tex_num;
+	{
+		lock_guard<mutex> lock(mu);
+		for (Texture &tex : textures) {
+			if (!tex.in_use && !tex.is_renderbuffer && tex.format == format &&
+			    tex.width == width && tex.height == height && tex.num_layers == num_layers) {
+				tex.in_use = true;
+				return tex.tex_num;
+			}
 		}
 	}
 
@@ -1041,17 +1044,23 @@ GLuint TexturePool::get_texture(GLenum format, GLuint width, GLuint height, GLui
 	tex.num_layers = num_layers;
 	tex.in_use = true;
 	tex.is_renderbuffer = false;
-	textures.push_back(tex);
+	{
+		lock_guard<mutex> lock(mu);
+		textures.push_back(tex);
+	}
 	return tex.tex_num;
 }
 
 GLuint TexturePool::get_renderbuffer(GLenum format, GLuint width, GLuint height)
 {
-	for (Texture &tex : textures) {
-		if (!tex.in_use && tex.is_renderbuffer && tex.format == format &&
-		    tex.width == width && tex.height == height) {
-			tex.in_use = true;
-			return tex.tex_num;
+	{
+		lock_guard<mutex> lock(mu);
+		for (Texture &tex : textures) {
+			if (!tex.in_use && tex.is_renderbuffer && tex.format == format &&
+			    tex.width == width && tex.height == height) {
+				tex.in_use = true;
+				return tex.tex_num;
+			}
 		}
 	}
 
@@ -1064,12 +1073,16 @@ GLuint TexturePool::get_renderbuffer(GLenum format, GLuint width, GLuint height)
 	tex.height = height;
 	tex.in_use = true;
 	tex.is_renderbuffer = true;
-	textures.push_back(tex);
+	{
+		lock_guard<mutex> lock(mu);
+		textures.push_back(tex);
+	}
 	return tex.tex_num;
 }
 
 void TexturePool::release_texture(GLuint tex_num)
 {
+	lock_guard<mutex> lock(mu);
 	for (Texture &tex : textures) {
 		if (!tex.is_renderbuffer && tex.tex_num == tex_num) {
 			assert(tex.in_use);
@@ -1082,6 +1095,7 @@ void TexturePool::release_texture(GLuint tex_num)
 
 void TexturePool::release_renderbuffer(GLuint tex_num)
 {
+	lock_guard<mutex> lock(mu);
 	for (Texture &tex : textures) {
 		if (tex.is_renderbuffer && tex.tex_num == tex_num) {
 			assert(tex.in_use);
diff --git a/flow.h b/flow.h
index 6475d9b..f268316 100644
--- a/flow.h
+++ b/flow.h
@@ -9,6 +9,7 @@
 #include <epoxy/gl.h>
 #include <array>
 #include <map>
+#include <mutex>
 #include <vector>
 #include <utility>
 
@@ -383,6 +384,7 @@ private:
 	GLuint uniform_scale_factor;
 };
 
+// All operations, except construction and destruction, are thread-safe.
 class TexturePool {
 public:
 	GLuint get_texture(GLenum format, GLuint width, GLuint height, GLuint num_layers = 0);
@@ -398,7 +400,8 @@ private:
 		bool in_use = false;
 		bool is_renderbuffer = false;
 	};
-	std::vector<Texture> textures;
+	std::mutex mu;
+	std::vector<Texture> textures;  // Under mu.
 };
 
 class DISComputeFlow {
diff --git a/video_stream.cpp b/video_stream.cpp
index 441e36a..820a933 100644
--- a/video_stream.cpp
+++ b/video_stream.cpp
@@ -318,12 +318,16 @@ void VideoStream::schedule_interpolated_frame(int64_t output_pts, unsigned strea
 	check_error();
 	glGenerateTextureMipmap(resources.gray_tex);
 	check_error();
-	GLuint flow_tex = compute_flow->exec(resources.gray_tex, DISComputeFlow::FORWARD_AND_BACKWARD, DISComputeFlow::DO_NOT_RESIZE_FLOW);
+	qf.flow_tex = compute_flow->exec(resources.gray_tex, DISComputeFlow::FORWARD_AND_BACKWARD, DISComputeFlow::DO_NOT_RESIZE_FLOW);
 	check_error();
 
-	qf.output_tex = interpolate->exec(resources.input_tex, flow_tex, 1280, 720, alpha);
+	qf.output_tex = interpolate->exec(resources.input_tex, qf.flow_tex, 1280, 720, alpha);
 	check_error();
 
+	// We could have released qf.flow_tex here, but to make sure we don't cause a stall
+	// when trying to reuse it for the next frame, we can just as well hold on to it
+	// and release it only when the readback is done.
+
 	// Read it down (asynchronously) to the CPU.
 	glPixelStorei(GL_PACK_ROW_LENGTH, 0);
 	glBindBuffer(GL_PIXEL_PACK_BUFFER, resources.pbo);
@@ -377,6 +381,8 @@ void VideoStream::encode_thread_func()
 			glClientWaitSync(qf.fence.get(), /*flags=*/0, GL_TIMEOUT_IGNORED);
 
 			vector<uint8_t> jpeg = encode_jpeg((const uint8_t *)qf.resources.pbo_contents, 1280, 720);
+			compute_flow->release_texture(qf.flow_tex);
+			interpolate->release_texture(qf.output_tex);
 
 			AVPacket pkt;
 			av_init_packet(&pkt);
diff --git a/video_stream.h b/video_stream.h
index 953e28e..3645123 100644
--- a/video_stream.h
+++ b/video_stream.h
@@ -66,8 +66,8 @@ private:
 		int64_t input_second_pts;
 		float alpha;
 		InterpolatedFrameResources resources;
-		GLuint output_tex;
 		RefCountedGLsync fence;  // Set when the interpolated image is read back to the CPU.
+		GLuint flow_tex, output_tex;  // Released in the receiving thread; not really used for anything else.
 	};
 	std::deque<QueuedFrame> frame_queue;  // Under <queue_lock>.
 	std::mutex queue_lock;