Fix read from screen and depth texture

6 files changed, 158 insertions, 2 deletions

diff --git a/servers/rendering/renderer_rd/effects_rd.cpp b/servers/rendering/renderer_rd/effects_rd.cpp
index 0af8f8bacf..78bbdb046d 100644
--- a/servers/rendering/renderer_rd/effects_rd.cpp
+++ b/servers/rendering/renderer_rd/effects_rd.cpp
@@ -739,6 +739,26 @@ void EffectsRD::make_mipmap(RID p_source_rd_texture, RID p_dest_texture, const S
 	RD::get_singleton()->compute_list_end();
 }
 
+void EffectsRD::make_mipmap_raster(RID p_source_rd_texture, RID p_dest_framebuffer, const Size2i &p_size) {
+	ERR_FAIL_COND_MSG(!prefer_raster_effects, "Can't use the raster version of mipmap with the clustered renderer.");
+
+	memset(&blur_raster.push_constant, 0, sizeof(BlurRasterPushConstant));
+
+	BlurRasterMode mode = BLUR_MIPMAP;
+
+	blur_raster.push_constant.pixel_size[0] = 1.0 / float(p_size.x);
+	blur_raster.push_constant.pixel_size[1] = 1.0 / float(p_size.y);
+
+	RD::DrawListID draw_list = RD::get_singleton()->draw_list_begin(p_dest_framebuffer, RD::INITIAL_ACTION_KEEP, RD::FINAL_ACTION_READ, RD::INITIAL_ACTION_KEEP, RD::FINAL_ACTION_DISCARD);
+	RD::get_singleton()->draw_list_bind_render_pipeline(draw_list, blur_raster.pipelines[mode].get_render_pipeline(RD::INVALID_ID, RD::get_singleton()->framebuffer_get_format(p_dest_framebuffer)));
+	RD::get_singleton()->draw_list_bind_uniform_set(draw_list, _get_uniform_set_from_texture(p_source_rd_texture), 0);
+	RD::get_singleton()->draw_list_bind_index_array(draw_list, index_array);
+	RD::get_singleton()->draw_list_set_push_constant(draw_list, &blur_raster.push_constant, sizeof(BlurRasterPushConstant));
+
+	RD::get_singleton()->draw_list_draw(draw_list, true);
+	RD::get_singleton()->draw_list_end();
+}
+
 void EffectsRD::copy_cubemap_to_dp(RID p_source_rd_texture, RID p_dst_framebuffer, const Rect2 &p_rect, float p_z_near, float p_z_far, bool p_dp_flip) {
 	CopyToDPPushConstant push_constant;
 	push_constant.screen_rect[0] = p_rect.position.x;
@@ -1873,6 +1893,7 @@ EffectsRD::EffectsRD(bool p_prefer_raster_effects) {
 		// init blur shader (on compute use copy shader)
 
 		Vector<String> blur_modes;
+		blur_modes.push_back("\n#define MODE_MIPMAP\n"); // BLUR_MIPMAP
 		blur_modes.push_back("\n#define MODE_GAUSSIAN_BLUR\n"); // BLUR_MODE_GAUSSIAN_BLUR
 		blur_modes.push_back("\n#define MODE_GAUSSIAN_GLOW\n"); // BLUR_MODE_GAUSSIAN_GLOW
 		blur_modes.push_back("\n#define MODE_GAUSSIAN_GLOW\n#define GLOW_USE_AUTO_EXPOSURE\n"); // BLUR_MODE_GAUSSIAN_GLOW_AUTO_EXPOSURE
diff --git a/servers/rendering/renderer_rd/effects_rd.h b/servers/rendering/renderer_rd/effects_rd.h
index fc14050ccb..0c9b2efb7f 100644
--- a/servers/rendering/renderer_rd/effects_rd.h
+++ b/servers/rendering/renderer_rd/effects_rd.h
@@ -70,6 +70,8 @@ private:
 	bool prefer_raster_effects;
 
 	enum BlurRasterMode {
+		BLUR_MIPMAP,
+
 		BLUR_MODE_GAUSSIAN_BLUR,
 		BLUR_MODE_GAUSSIAN_GLOW,
 		BLUR_MODE_GAUSSIAN_GLOW_AUTO_EXPOSURE,
@@ -767,6 +769,7 @@ public:
 	void cubemap_roughness(RID p_source_rd_texture, RID p_dest_texture, uint32_t p_face_id, uint32_t p_sample_count, float p_roughness, float p_size);
 	void cubemap_roughness_raster(RID p_source_rd_texture, RID p_dest_framebuffer, uint32_t p_face_id, uint32_t p_sample_count, float p_roughness, float p_size);
 	void make_mipmap(RID p_source_rd_texture, RID p_dest_texture, const Size2i &p_size);
+	void make_mipmap_raster(RID p_source_rd_texture, RID p_dest_framebuffer, const Size2i &p_size);
 	void copy_cubemap_to_dp(RID p_source_rd_texture, RID p_dest_texture, const Rect2 &p_rect, float p_z_near, float p_z_far, bool p_dp_flip);
 	void luminance_reduction(RID p_source_texture, const Size2i p_source_size, const Vector<RID> p_reduce, RID p_prev_luminance, float p_min_luminance, float p_max_luminance, float p_adjust, bool p_set = false);
 	void luminance_reduction_raster(RID p_source_texture, const Size2i p_source_size, const Vector<RID> p_reduce, Vector<RID> p_fb, RID p_prev_luminance, float p_min_luminance, float p_max_luminance, float p_adjust, bool p_set = false);
diff --git a/servers/rendering/renderer_rd/forward_clustered/render_forward_clustered.cpp b/servers/rendering/renderer_rd/forward_clustered/render_forward_clustered.cpp
index 0d304556c2..5a0ad97c7c 100644
--- a/servers/rendering/renderer_rd/forward_clustered/render_forward_clustered.cpp
+++ b/servers/rendering/renderer_rd/forward_clustered/render_forward_clustered.cpp
@@ -1533,6 +1533,16 @@ void RenderForwardClustered::_render_scene(RenderDataRD *p_render_data, const Co
 		}
 	}
 
+	if (scene_state.used_screen_texture) {
+		// Copy screen texture to backbuffer so we can read from it
+		_render_buffers_copy_screen_texture(p_render_data);
+	}
+
+	if (scene_state.used_depth_texture) {
+		// Copy depth texture to backbuffer so we can read from it
+		_render_buffers_copy_depth_texture(p_render_data);
+	}
+
 	RENDER_TIMESTAMP("Render Transparent Pass");
 
 	RD::get_singleton()->draw_command_begin_label("Render Transparent Pass");
@@ -2219,7 +2229,8 @@ RID RenderForwardClustered::_setup_render_pass_uniform_set(RenderListType p_rend
 		RD::Uniform u;
 		u.binding = 9;
 		u.uniform_type = RD::UNIFORM_TYPE_TEXTURE;
-		RID texture = (false && rb && rb->depth.is_valid()) ? rb->depth : storage->texture_rd_get_default(RendererStorageRD::DEFAULT_RD_TEXTURE_WHITE);
+		RID dbt = rb ? render_buffers_get_back_depth_texture(p_render_data->render_buffers) : RID();
+		RID texture = (dbt.is_valid()) ? dbt : storage->texture_rd_get_default(RendererStorageRD::DEFAULT_RD_TEXTURE_WHITE);
 		u.ids.push_back(texture);
 		uniforms.push_back(u);
 	}
diff --git a/servers/rendering/renderer_rd/forward_mobile/render_forward_mobile.cpp b/servers/rendering/renderer_rd/forward_mobile/render_forward_mobile.cpp
index f810ae2454..ddd2a26d71 100644
--- a/servers/rendering/renderer_rd/forward_mobile/render_forward_mobile.cpp
+++ b/servers/rendering/renderer_rd/forward_mobile/render_forward_mobile.cpp
@@ -414,7 +414,8 @@ RID RenderForwardMobile::_setup_render_pass_uniform_set(RenderListType p_render_
 		RD::Uniform u;
 		u.binding = 9;
 		u.uniform_type = RD::UNIFORM_TYPE_TEXTURE;
-		RID texture = (false && rb && rb->depth.is_valid()) ? rb->depth : storage->texture_rd_get_default(RendererStorageRD::DEFAULT_RD_TEXTURE_WHITE);
+		RID dbt = rb ? render_buffers_get_back_depth_texture(p_render_data->render_buffers) : RID();
+		RID texture = (dbt.is_valid()) ? dbt : storage->texture_rd_get_default(RendererStorageRD::DEFAULT_RD_TEXTURE_WHITE);
 		u.ids.push_back(texture);
 		uniforms.push_back(u);
 	}
@@ -752,6 +753,16 @@ void RenderForwardMobile::_render_scene(RenderDataRD *p_render_data, const Color
 			RD::get_singleton()->draw_command_end_label(); // Render 3D Pass / Render Reflection Probe Pass
 		}
 
+		if (scene_state.used_screen_texture) {
+			// Copy screen texture to backbuffer so we can read from it
+			_render_buffers_copy_screen_texture(p_render_data);
+		}
+
+		if (scene_state.used_depth_texture) {
+			// Copy depth texture to backbuffer so we can read from it
+			_render_buffers_copy_depth_texture(p_render_data);
+		}
+
 		// transparent pass
 
 		RD::get_singleton()->draw_command_begin_label("Render Transparent Subpass");
diff --git a/servers/rendering/renderer_rd/renderer_scene_render_rd.cpp b/servers/rendering/renderer_rd/renderer_scene_render_rd.cpp
index 06e5ff4a0b..5016c99a1e 100644
--- a/servers/rendering/renderer_rd/renderer_scene_render_rd.cpp
+++ b/servers/rendering/renderer_rd/renderer_scene_render_rd.cpp
@@ -1513,6 +1513,38 @@ void RendererSceneRenderRD::_allocate_blur_textures(RenderBuffers *rb) {
 	}
 }
 
+void RendererSceneRenderRD::_allocate_depth_backbuffer_textures(RenderBuffers *rb) {
+	ERR_FAIL_COND(!rb->depth_back_texture.is_null());
+
+	{
+		RD::TextureFormat tf;
+		if (rb->view_count > 1) {
+			tf.texture_type = RD::TEXTURE_TYPE_2D_ARRAY;
+		}
+		// We're not using this as a depth stencil, just copying our data into this. May need to look into using a different format on mobile, maybe R16?
+		tf.format = RD::DATA_FORMAT_R32_SFLOAT;
+
+		tf.width = rb->width;
+		tf.height = rb->height;
+		tf.usage_bits = RD::TEXTURE_USAGE_SAMPLING_BIT;
+		tf.array_layers = rb->view_count; // create a layer for every view
+
+		tf.usage_bits |= RD::TEXTURE_USAGE_CAN_COPY_TO_BIT | RD::TEXTURE_USAGE_STORAGE_BIT;
+		tf.usage_bits |= RD::TEXTURE_USAGE_COLOR_ATTACHMENT_BIT; // set this as color attachment because we're copying data into it, it's not actually used as a depth buffer
+
+		rb->depth_back_texture = RD::get_singleton()->texture_create(tf, RD::TextureView());
+	}
+
+	if (!_render_buffers_can_be_storage()) {
+		// create framebuffer so we can write into this...
+
+		Vector<RID> fb;
+		fb.push_back(rb->depth_back_texture);
+
+		rb->depth_back_fb = RD::get_singleton()->framebuffer_create(fb, RD::INVALID_ID, rb->view_count);
+	}
+}
+
 void RendererSceneRenderRD::_allocate_luminance_textures(RenderBuffers *rb) {
 	ERR_FAIL_COND(!rb->luminance.current.is_null());
 
@@ -1579,6 +1611,16 @@ void RendererSceneRenderRD::_free_render_buffer_data(RenderBuffers *rb) {
 		rb->depth_texture = RID();
 	}
 
+	if (rb->depth_back_fb.is_valid()) {
+		RD::get_singleton()->free(rb->depth_back_fb);
+		rb->depth_back_fb = RID();
+	}
+
+	if (rb->depth_back_texture.is_valid()) {
+		RD::get_singleton()->free(rb->depth_back_texture);
+		rb->depth_back_texture = RID();
+	}
+
 	for (int i = 0; i < 2; i++) {
 		for (int m = 0; m < rb->blur[i].mipmaps.size(); m++) {
 			// do we free the texture slice here? or is it enough to free the main texture?
@@ -1879,6 +1921,58 @@ void RendererSceneRenderRD::_process_ssao(RID p_render_buffers, RID p_environmen
 	storage->get_effects()->generate_ssao(rb->depth_texture, p_normal_buffer, rb->ssao.depth, rb->ssao.depth_slices, rb->ssao.ao_deinterleaved, rb->ssao.ao_deinterleaved_slices, rb->ssao.ao_pong, rb->ssao.ao_pong_slices, rb->ssao.ao_final, rb->ssao.importance_map[0], rb->ssao.importance_map[1], p_projection, settings, uniform_sets_are_invalid, rb->ssao.downsample_uniform_set, rb->ssao.gather_uniform_set, rb->ssao.importance_map_uniform_set);
 }
 
+void RendererSceneRenderRD::_render_buffers_copy_screen_texture(const RenderDataRD *p_render_data) {
+	RenderBuffers *rb = render_buffers_owner.getornull(p_render_data->render_buffers);
+	ERR_FAIL_COND(!rb);
+
+	RD::get_singleton()->draw_command_begin_label("Copy screen texture");
+
+	if (rb->blur[0].texture.is_null()) {
+		_allocate_blur_textures(rb);
+	}
+
+	// @TODO IMPLEMENT MULTIVIEW, all effects need to support stereo buffers or effects are only applied to the left eye
+
+	bool can_use_storage = _render_buffers_can_be_storage();
+
+	if (can_use_storage) {
+		storage->get_effects()->copy_to_rect(rb->texture, rb->blur[0].mipmaps[0].texture, Rect2i(0, 0, rb->width, rb->height));
+		for (int i = 1; i < rb->blur[0].mipmaps.size(); i++) {
+			storage->get_effects()->make_mipmap(rb->blur[0].mipmaps[i - 1].texture, rb->blur[0].mipmaps[i].texture, Size2i(rb->blur[0].mipmaps[i].width, rb->blur[0].mipmaps[i].height));
+		}
+	} else {
+		storage->get_effects()->copy_to_fb_rect(rb->texture, rb->blur[0].mipmaps[0].fb, Rect2i(0, 0, rb->width, rb->height));
+		for (int i = 1; i < rb->blur[0].mipmaps.size(); i++) {
+			storage->get_effects()->make_mipmap_raster(rb->blur[0].mipmaps[i - 1].texture, rb->blur[0].mipmaps[i].fb, Size2i(rb->blur[0].mipmaps[i].width, rb->blur[0].mipmaps[i].height));
+		}
+	}
+
+	RD::get_singleton()->draw_command_end_label();
+}
+
+void RendererSceneRenderRD::_render_buffers_copy_depth_texture(const RenderDataRD *p_render_data) {
+	RenderBuffers *rb = render_buffers_owner.getornull(p_render_data->render_buffers);
+	ERR_FAIL_COND(!rb);
+
+	RD::get_singleton()->draw_command_begin_label("Copy depth texture");
+
+	if (rb->depth_back_texture.is_null()) {
+		_allocate_depth_backbuffer_textures(rb);
+	}
+
+	// @TODO IMPLEMENT MULTIVIEW, all effects need to support stereo buffers or effects are only applied to the left eye
+
+	bool can_use_storage = _render_buffers_can_be_storage();
+
+	if (can_use_storage) {
+		storage->get_effects()->copy_to_rect(rb->depth_texture, rb->depth_back_texture, Rect2i(0, 0, rb->width, rb->height));
+	} else {
+		storage->get_effects()->copy_to_fb_rect(rb->depth_texture, rb->depth_back_fb, Rect2i(0, 0, rb->width, rb->height));
+	}
+
+	RD::get_singleton()->draw_command_end_label();
+}
+
 void RendererSceneRenderRD::_render_buffers_post_process_and_tonemap(const RenderDataRD *p_render_data) {
 	RenderBuffers *rb = render_buffers_owner.getornull(p_render_data->render_buffers);
 	ERR_FAIL_COND(!rb);
@@ -2227,6 +2321,15 @@ RID RendererSceneRenderRD::render_buffers_get_back_buffer_texture(RID p_render_b
 	return rb->blur[0].texture;
 }
 
+RID RendererSceneRenderRD::render_buffers_get_back_depth_texture(RID p_render_buffers) {
+	RenderBuffers *rb = render_buffers_owner.getornull(p_render_buffers);
+	ERR_FAIL_COND_V(!rb, RID());
+	if (!rb->depth_back_texture.is_valid()) {
+		return RID(); //not valid at the moment
+	}
+	return rb->depth_back_texture;
+}
+
 RID RendererSceneRenderRD::render_buffers_get_ao_texture(RID p_render_buffers) {
 	RenderBuffers *rb = render_buffers_owner.getornull(p_render_buffers);
 	ERR_FAIL_COND_V(!rb, RID());
diff --git a/servers/rendering/renderer_rd/renderer_scene_render_rd.h b/servers/rendering/renderer_rd/renderer_scene_render_rd.h
index 236e4b4320..5306df36f0 100644
--- a/servers/rendering/renderer_rd/renderer_scene_render_rd.h
+++ b/servers/rendering/renderer_rd/renderer_scene_render_rd.h
@@ -134,6 +134,8 @@ protected:
 
 	void _pre_opaque_render(RenderDataRD *p_render_data, bool p_use_ssao, bool p_use_gi, RID p_normal_roughness_buffer, RID p_voxel_gi_buffer);
 
+	void _render_buffers_copy_screen_texture(const RenderDataRD *p_render_data);
+	void _render_buffers_copy_depth_texture(const RenderDataRD *p_render_data);
 	void _render_buffers_post_process_and_tonemap(const RenderDataRD *p_render_data);
 	void _post_process_subpass(RID p_source_texture, RID p_framebuffer, const RenderDataRD *p_render_data);
 	void _disable_clear_request(const RenderDataRD *p_render_data);
@@ -492,6 +494,9 @@ private:
 		WeightBuffers weight_buffers[4]; // Only used in raster
 		RID base_weight_fb; // base buffer for weight
 
+		RID depth_back_texture;
+		RID depth_back_fb; // only used on mobile
+
 		struct Luminance {
 			Vector<RID> reduce;
 			RID current;
@@ -535,6 +540,7 @@ private:
 
 	void _free_render_buffer_data(RenderBuffers *rb);
 	void _allocate_blur_textures(RenderBuffers *rb);
+	void _allocate_depth_backbuffer_textures(RenderBuffers *rb);
 	void _allocate_luminance_textures(RenderBuffers *rb);
 
 	void _render_buffers_debug_draw(RID p_render_buffers, RID p_shadow_atlas, RID p_occlusion_buffer);
@@ -1183,6 +1189,7 @@ public:
 
 	RID render_buffers_get_ao_texture(RID p_render_buffers);
 	RID render_buffers_get_back_buffer_texture(RID p_render_buffers);
+	RID render_buffers_get_back_depth_texture(RID p_render_buffers);
 	RID render_buffers_get_voxel_gi_buffer(RID p_render_buffers);
 	RID render_buffers_get_default_voxel_gi_buffer();
 	RID render_buffers_get_gi_ambient_texture(RID p_render_buffers);