Use SendTexture to lower the overhead of the spout filter (#46)

* Use SendTexture() to avoid some extra memory copies

* Double-buffer to give good perf without a flush

* Use intermediate texture to handle sRGB

* Tidy

Authored-by: @millenium-cyborg 

Closes #42

Author: y2kcyborg

source/win-spout-filter.cpp

@@ -23,10 +23,11 @@ struct win_spout_filter
 	spoutDX* filter_sender;
 	obs_source_t* source_context;
 	const char* sender_name;
-	pthread_mutex_t mutex;
 	uint32_t width;
 	uint32_t height;
-	gs_texrender_t* texrender;
+	gs_texrender_t* texrender_curr;
+	gs_texrender_t* texrender_prev;
+	gs_texrender_t* texrender_intermediate;
 	gs_stagesurf_t* stagesurface;
 	video_t* video_output;
 	uint8_t* video_data;
@@ -80,17 +81,6 @@ void win_spout_filter_getdefaults(obs_data_t* defaults)
 		obs_module_text("defaultfiltername"));
 }
 
-void win_spout_filter_raw_video(void* data, video_data* frame)
-{
-	struct win_spout_filter* context = (win_spout_filter*)data;
-
-	if (!frame|| !frame->data[0]) return;
-
-	pthread_mutex_lock(&context->mutex);
-	context->filter_sender->SendImage(frame->data[0], context->width, context->height);
-	pthread_mutex_unlock(&context->mutex);
-}
-
 void win_spout_offscreen_render(void* data, uint32_t cx, uint32_t cy)
 {
 
@@ -104,76 +94,76 @@ void win_spout_offscreen_render(void* data, uint32_t cx, uint32_t cy)
 	uint32_t width = obs_source_get_base_width(target);
 	uint32_t height = obs_source_get_base_height(target);
 
-	gs_texrender_reset(context->texrender);
-
-	if (gs_texrender_begin(context->texrender, width, height))
-	{
+	// Render the target to an intemediate format in sRGB-aware format
+	gs_texrender_reset(context->texrender_intermediate);
+	if (gs_texrender_begin(context->texrender_intermediate, width, height)) {
 		struct vec4 background;
 		vec4_zero(&background);
 
 		gs_clear(GS_CLEAR_COLOR, &background, 0.0f, 0);
-		gs_ortho(0.0f, (float)width, 0.0f, (float)height, -100.0f, 100.0f);
+		gs_ortho(0.0f, (float)width, 0.0f, (float)height, -100.0f,
+			 100.0f);
 
 		gs_blend_state_push();
 		gs_blend_function(GS_BLEND_ONE, GS_BLEND_ZERO);
 
 		obs_source_video_render(target);
 
 		gs_blend_state_pop();
-		gs_texrender_end(context->texrender);
+		gs_texrender_end(context->texrender_intermediate);
+	}
+
+	// Use the default effect to render it back into a format Spout accepts
+	gs_texrender_reset(context->texrender_curr);
+	if (gs_texrender_begin(context->texrender_curr, width, height))
+	{
+		struct vec4 background;
+		vec4_zero(&background);
 
-		if (context->width != width || context->height != height)
-		{
+		gs_clear(GS_CLEAR_COLOR, &background, 0.0f, 0);
+		gs_ortho(0.0f, (float)width, 0.0f, (float)height, -100.0f, 100.0f);
 
-			gs_stagesurface_destroy(context->stagesurface);
-			context->stagesurface = gs_stagesurface_create(width, height, GS_BGRA);
+		gs_blend_state_push();
+		gs_blend_function(GS_BLEND_ONE, GS_BLEND_ZERO);
 
-			video_output_info video_out = { 0 };
-			video_out.format = VIDEO_FORMAT_BGRA;
-			video_out.width = width;
-			video_out.height = height;
-			video_out.fps_den = context->video_info.fps_den;
-			video_out.fps_num = context->video_info.fps_num;
-			video_out.cache_size = 16;
-			video_out.colorspace = VIDEO_CS_DEFAULT;
-			video_out.range = VIDEO_RANGE_DEFAULT;
-			video_out.name = obs_source_get_name(context->source_context);
+		// To get sRGB handling, render with the default effect
+		gs_effect_t *effect = obs_get_base_effect(OBS_EFFECT_DEFAULT);
+		gs_texture_t *tex = gs_texrender_get_texture(context->texrender_intermediate);
+		if (tex) {
+			const bool linear_srgb = gs_get_linear_srgb();
 
-			video_output_close(context->video_output);
+			const bool previous = gs_framebuffer_srgb_enabled();
+			gs_enable_framebuffer_srgb(linear_srgb);
 
-			context->width = width;
-			context->height = height;
-			video_output_open(&context->video_output, &video_out);
-			video_output_connect(context->video_output, nullptr, win_spout_filter_raw_video, context);
+			gs_eparam_t *image =
+				gs_effect_get_param_by_name(effect, "image");
+			if (linear_srgb)
+				gs_effect_set_texture_srgb(image, tex);
+			else
+				gs_effect_set_texture(image, tex);
 
+			while (gs_effect_loop(effect, "Draw"))
+				gs_draw_sprite(tex, 0, width, height);
 
+			gs_enable_framebuffer_srgb(previous);
 		}
 
-		struct video_frame output_frame;
-		if (video_output_lock_frame(context->video_output,
-			&output_frame, 1, obs_get_video_frame_time()))
-		{
-			if (context->video_data) {
-				gs_stagesurface_unmap(context->stagesurface);
-				context->video_data = nullptr;
-			}
-
-			gs_stage_texture(context->stagesurface,
-				gs_texrender_get_texture(context->texrender));
-			gs_stagesurface_map(context->stagesurface,
-				&context->video_data, &context->video_linesize);
-
-			uint32_t linesize = output_frame.linesize[0];
-			for (uint32_t i = 0; i < context->height; ++i) {
-				uint32_t dst_offset = linesize * i;
-				uint32_t src_offset = context->video_linesize * i;
-				memcpy(output_frame.data[0] + dst_offset,
-					context->video_data + src_offset,
-					linesize);
-			}
-
-			video_output_unlock_frame(context->video_output);
+		gs_blend_state_pop();
+		gs_texrender_end(context->texrender_curr);
+
+		gs_texture_t *prev_tex =
+			gs_texrender_get_texture(context->texrender_prev);
+		if (prev_tex) {
+			context->filter_sender->SendTexture((
+				ID3D11Texture2D *)gs_texture_get_obj(prev_tex));
 		}
+
+		// Swap the buffers
+		// Double-buffering avoids the need for a flush, and also fixes
+		// some issues related to G-Sync.
+		gs_texrender_t *tmp = context->texrender_curr;
+		context->texrender_curr = context->texrender_prev;
+		context->texrender_prev = tmp;
 	}
 }
 
@@ -194,7 +184,11 @@ void* win_spout_filter_create(obs_data_t* settings, obs_source_t* source)
 	struct win_spout_filter* context = (win_spout_filter*)bzalloc(sizeof(win_spout_filter));
 
 	context->source_context = source;
-	context->texrender = gs_texrender_create(GS_BGRA, GS_ZS_NONE);
+	// Use a Spout-compatible texture format
+	context->texrender_curr = gs_texrender_create(GS_BGRA_UNORM, GS_ZS_NONE);
+	context->texrender_prev = gs_texrender_create(GS_BGRA_UNORM, GS_ZS_NONE);
+	context->texrender_intermediate =
+		gs_texrender_create(GS_BGRA, GS_ZS_NONE);
 	context->sender_name = obs_data_get_string(settings, FILTER_PROP_NAME);
 	context->video_data = nullptr;
 
@@ -203,15 +197,9 @@ void* win_spout_filter_create(obs_data_t* settings, obs_source_t* source)
 	obs_get_video_info(&context->video_info);
 	win_spout_filter_update(context, settings);
 
-	
-
 	if (openDX11(context))
 	{
-		pthread_mutex_init_value(&context->mutex);
-		if (pthread_mutex_init(&context->mutex, NULL) == 0)
-		{
-			return context;
-		}
+		return context;
 	}
 
 	blog(LOG_ERROR, "Failed to create spout output!");
@@ -234,8 +222,9 @@ void win_spout_filter_destroy(void* data)
 		video_output_close(context->video_output);
 		gs_stagesurface_unmap(context->stagesurface);
 		gs_stagesurface_destroy(context->stagesurface);
-		gs_texrender_destroy(context->texrender);
-		pthread_mutex_destroy(&context->mutex);
+		gs_texrender_destroy(context->texrender_intermediate);
+		gs_texrender_destroy(context->texrender_prev);
+		gs_texrender_destroy(context->texrender_curr);
 		bfree(context);
 	}
 }
@@ -259,7 +248,7 @@ struct obs_source_info create_spout_filter_info()
 	struct obs_source_info win_spout_filter_info = {};
 	win_spout_filter_info.id = "win_spout_filter";
 	win_spout_filter_info.type = OBS_SOURCE_TYPE_FILTER;
-	win_spout_filter_info.output_flags = OBS_SOURCE_VIDEO;
+	win_spout_filter_info.output_flags = OBS_SOURCE_VIDEO | OBS_SOURCE_SRGB;
 	win_spout_filter_info.get_name = win_spout_filter_getname;
 	win_spout_filter_info.get_properties = win_spout_filter_getproperties;
 	win_spout_filter_info.get_defaults = win_spout_filter_getdefaults;