gfx-blur-dual-filtering: Implement asynchronous rendering

Asynchronous rendering allows the GPU to perform work while the CPU performs other work, and is significantly faster than lockstep immediate rendering. By reusing existing render targets we can see a performance improvement of up to 500%, while still doing the same things.

source/gfx/blur/gfx-blur-dual-filtering.cpp

@@ -184,9 +184,15 @@ gfx::blur::dual_filtering::dual_filtering()
 	: _data(::gfx::blur::dual_filtering_factory::get().data()), _size(0), _size_iterations(0)
 {
 	auto gctx = gs::context();
-	_rendertargets.resize(MAX_LEVELS + 1);
+	_rts.resize(MAX_LEVELS + 1);
 	for (std::size_t n = 0; n <= MAX_LEVELS; n++) {
-		_rendertargets[n] = std::make_shared<gs::rendertarget>(GS_RGBA32F, GS_ZS_NONE);
+		gs_color_format cf = GS_RGBA;
+#if 0
+		cf = GS_RGBA16F;
+#elif 0
+		cf = GS_RGBA32F;
+#endif
+		_rts[n] = std::make_shared<gs::rendertarget>(cf, GS_ZS_NONE);
 	}
 }
 
@@ -245,34 +251,38 @@ std::shared_ptr<::gs::texture> gfx::blur::dual_filtering::render()
 	gs_stencil_function(GS_STENCIL_BOTH, GS_ALWAYS);
 	gs_stencil_op(GS_STENCIL_BOTH, GS_ZERO, GS_ZERO, GS_ZERO);
 
+	uint32_t width  = _input_texture->get_width();
+	uint32_t height = _input_texture->get_height();
+
 	// Downsample
 	for (std::size_t n = 1; n <= actual_iterations; n++) {
 		// Idx 0 is a simply considered as a straight copy of the original and not rendered to.
+		auto gdm = gs::debug_marker(gs::debug_color_azure_radiance, "Down %lld", n);
 
 		// Select Texture
 		std::shared_ptr<gs::texture> tex_cur;
 		if (n > 1) {
-			tex_cur = _rendertargets[n - 1]->get_texture();
+			tex_cur = _rts[n - 1]->get_texture();
 		} else {
 			tex_cur = _input_texture;
 		}
 
 		// Reduce Size
-		std::uint32_t width  = tex_cur->get_width() / 2;
+		std::uint32_t owidth  = width >> n;
-		std::uint32_t height = tex_cur->get_height() / 2;
+		std::uint32_t oheight = height >> n;
-		if ((width <= 0) || (height <= 0)) {
+		if ((owidth <= 0) || (oheight <= 0)) {
 			actual_iterations = n - 1;
 			break;
 		}
 
 		// Apply
 		effect.get_parameter("pImage").set_texture(tex_cur);
-		effect.get_parameter("pImageSize").set_float2(float_t(width), float_t(height));
+		effect.get_parameter("pImageSize").set_float2(float_t(owidth), float_t(oheight));
-		effect.get_parameter("pImageTexel").set_float2(1.0f / width, 1.0f / height);
+		effect.get_parameter("pImageTexel").set_float2(1.0f / owidth, 1.0f / oheight);
-		effect.get_parameter("pImageHalfTexel").set_float2(0.5f / width, 0.5f / height);
+		effect.get_parameter("pImageHalfTexel").set_float2(0.5f / owidth, 0.5f / oheight);
 
 		{
-			auto op = _rendertargets[n]->render(width, height);
+			auto op = _rts[n]->render(owidth, oheight);
 			gs_ortho(0., 1., 0., 1., 0., 1.);
 			while (gs_effect_loop(effect.get_object(), "Down")) {
 				gs_draw_sprite(tex_cur->get_object(), 0, 1, 1);
@@ -282,38 +292,39 @@ std::shared_ptr<::gs::texture> gfx::blur::dual_filtering::render()
 
 	// Upsample
 	for (std::size_t n = actual_iterations; n > 0; n--) {
+		// Idx max is a simply considered as a straight copy of the downscale and not rendered to.
+		auto gdm = gs::debug_marker(gs::debug_color_azure_radiance, "Up %lld", n);
+
 		// Select Texture
-		std::shared_ptr<gs::texture> tex_cur = _rendertargets[n]->get_texture();
+		std::shared_ptr<gs::texture> tex_in = _rts[n]->get_texture();
 
 		// Get Size
-		std::uint32_t width  = tex_cur->get_width();
+		std::uint32_t iwidth  = width >> n;
-		std::uint32_t height = tex_cur->get_height();
+		std::uint32_t iheight = height >> n;
+		std::uint32_t owidth  = width >> (n - 1);
+		std::uint32_t oheight = height >> (n - 1);
 
 		// Apply
-		effect.get_parameter("pImage").set_texture(tex_cur);
+		effect.get_parameter("pImage").set_texture(tex_in);
-		effect.get_parameter("pImageSize").set_float2(float_t(width), float_t(height));
+		effect.get_parameter("pImageSize").set_float2(float_t(iwidth), float_t(iheight));
-		effect.get_parameter("pImageTexel").set_float2(1.0f / width, 1.0f / height);
+		effect.get_parameter("pImageTexel").set_float2(1.0f / iwidth, 1.0f / iheight);
-		effect.get_parameter("pImageHalfTexel").set_float2(0.5f / width, 0.5f / height);
+		effect.get_parameter("pImageHalfTexel").set_float2(0.5f / iwidth, 0.5f / iheight);
-
-		// Increase Size
-		width *= 2;
-		height *= 2;
 
 		{
-			auto op = _rendertargets[n - 1]->render(width, height);
+			auto op = _rts[n - 1]->render(owidth, oheight);
 			gs_ortho(0., 1., 0., 1., 0., 1.);
 			while (gs_effect_loop(effect.get_object(), "Up")) {
-				gs_draw_sprite(tex_cur->get_object(), 0, 1, 1);
+				gs_draw_sprite(tex_in->get_object(), 0, 1, 1);
 			}
 		}
 	}
 
 	gs_blend_state_pop();
 
-	return _rendertargets[0]->get_texture();
+	return _rts[0]->get_texture();
 }
 
 std::shared_ptr<::gs::texture> gfx::blur::dual_filtering::get()
 {
-	return _rendertargets[0]->get_texture();
+	return _rts[0]->get_texture();
 }

source/gfx/blur/gfx-blur-dual-filtering.hpp

@@ -88,7 +88,7 @@ namespace gfx {
 
 			std::shared_ptr<gs::texture> _input_texture;
 
-			std::vector<std::shared_ptr<gs::rendertarget>> _rendertargets;
+			std::vector<std::shared_ptr<gs::rendertarget>> _rts;
 
 			public:
 			dual_filtering();