1 files changed, 40 insertions, 27 deletions

diff --git a/servers/rendering/renderer_rd/environment/gi.cpp b/servers/rendering/renderer_rd/environment/gi.cpp
index d968736037..6da828df45 100644
--- a/servers/rendering/renderer_rd/environment/gi.cpp
+++ b/servers/rendering/renderer_rd/environment/gi.cpp
@@ -583,7 +583,8 @@ void GI::SDFGI::create(RID p_env, const Vector3 &p_world_position, uint32_t p_re
 		/* Buffers */
 
 		cascade.solid_cell_buffer = RD::get_singleton()->storage_buffer_create(sizeof(SDFGI::Cascade::SolidCell) * solid_cell_count);
-		cascade.solid_cell_dispatch_buffer = RD::get_singleton()->storage_buffer_create(sizeof(uint32_t) * 4, Vector<uint8_t>(), RD::STORAGE_BUFFER_USAGE_DISPATCH_INDIRECT);
+		cascade.solid_cell_dispatch_buffer_storage = RD::get_singleton()->storage_buffer_create(sizeof(uint32_t) * 4, Vector<uint8_t>());
+		cascade.solid_cell_dispatch_buffer_call = RD::get_singleton()->storage_buffer_create(sizeof(uint32_t) * 4, Vector<uint8_t>(), RD::STORAGE_BUFFER_USAGE_DISPATCH_INDIRECT);
 		cascade.lights_buffer = RD::get_singleton()->storage_buffer_create(sizeof(SDFGIShader::Light) * MAX(SDFGI::MAX_STATIC_LIGHTS, SDFGI::MAX_DYNAMIC_LIGHTS));
 		{
 			Vector<RD::Uniform> uniforms;
@@ -650,7 +651,7 @@ void GI::SDFGI::create(RID p_env, const Vector3 &p_world_position, uint32_t p_re
 				RD::Uniform u;
 				u.uniform_type = RD::UNIFORM_TYPE_STORAGE_BUFFER;
 				u.binding = 10;
-				u.append_id(cascade.solid_cell_dispatch_buffer);
+				u.append_id(cascade.solid_cell_dispatch_buffer_storage);
 				uniforms.push_back(u);
 			}
 			{
@@ -698,7 +699,7 @@ void GI::SDFGI::create(RID p_env, const Vector3 &p_world_position, uint32_t p_re
 				RD::Uniform u;
 				u.uniform_type = RD::UNIFORM_TYPE_STORAGE_BUFFER;
 				u.binding = 5;
-				u.append_id(cascade.solid_cell_dispatch_buffer);
+				u.append_id(cascade.solid_cell_dispatch_buffer_storage);
 				uniforms.push_back(u);
 			}
 			{
@@ -761,7 +762,7 @@ void GI::SDFGI::create(RID p_env, const Vector3 &p_world_position, uint32_t p_re
 			RD::Uniform u;
 			u.binding = 3;
 			u.uniform_type = RD::UNIFORM_TYPE_STORAGE_BUFFER;
-			u.append_id(cascade.solid_cell_dispatch_buffer);
+			u.append_id(cascade.solid_cell_dispatch_buffer_storage);
 			uniforms.push_back(u);
 		}
 		{
@@ -1129,7 +1130,8 @@ GI::SDFGI::~SDFGI() {
 		RD::get_singleton()->free(c.light_aniso_0_tex);
 		RD::get_singleton()->free(c.light_aniso_1_tex);
 		RD::get_singleton()->free(c.sdf_tex);
-		RD::get_singleton()->free(c.solid_cell_dispatch_buffer);
+		RD::get_singleton()->free(c.solid_cell_dispatch_buffer_storage);
+		RD::get_singleton()->free(c.solid_cell_dispatch_buffer_call);
 		RD::get_singleton()->free(c.solid_cell_buffer);
 		RD::get_singleton()->free(c.lightprobe_history_tex);
 		RD::get_singleton()->free(c.lightprobe_average_tex);
@@ -1238,6 +1240,10 @@ void GI::SDFGI::update(RID p_env, const Vector3 &p_world_position) {
 void GI::SDFGI::update_light() {
 	RD::get_singleton()->draw_command_begin_label("SDFGI Update dynamic Light");
 
+	for (uint32_t i = 0; i < cascades.size(); i++) {
+		RD::get_singleton()->buffer_copy(cascades[i].solid_cell_dispatch_buffer_storage, cascades[i].solid_cell_dispatch_buffer_call, 0, 0, sizeof(uint32_t) * 4);
+	}
+
 	/* Update dynamic light */
 
 	RD::ComputeListID compute_list = RD::get_singleton()->compute_list_begin();
@@ -1276,9 +1282,9 @@ void GI::SDFGI::update_light() {
 
 		RD::get_singleton()->compute_list_bind_uniform_set(compute_list, cascade.sdf_direct_light_dynamic_uniform_set, 0);
 		RD::get_singleton()->compute_list_set_push_constant(compute_list, &push_constant, sizeof(SDFGIShader::DirectLightPushConstant));
-		RD::get_singleton()->compute_list_dispatch_indirect(compute_list, cascade.solid_cell_dispatch_buffer, 0);
+		RD::get_singleton()->compute_list_dispatch_indirect(compute_list, cascade.solid_cell_dispatch_buffer_call, 0);
 	}
-	RD::get_singleton()->compute_list_end(RD::BARRIER_MASK_COMPUTE);
+	RD::get_singleton()->compute_list_end();
 	RD::get_singleton()->draw_command_end_label();
 }
 
@@ -1351,7 +1357,7 @@ void GI::SDFGI::update_probes(RID p_env, SkyRD::Sky *p_sky) {
 
 	render_pass++;
 
-	RD::ComputeListID compute_list = RD::get_singleton()->compute_list_begin(true);
+	RD::ComputeListID compute_list = RD::get_singleton()->compute_list_begin();
 	RD::get_singleton()->compute_list_bind_compute_pipeline(compute_list, gi->sdfgi_shader.integrate_pipeline[SDFGIShader::INTEGRATE_MODE_PROCESS]);
 
 	int32_t probe_divisor = cascade_size / SDFGI::PROBE_DIVISOR;
@@ -1368,14 +1374,11 @@ void GI::SDFGI::update_probes(RID p_env, SkyRD::Sky *p_sky) {
 		RD::get_singleton()->compute_list_dispatch_threads(compute_list, probe_axis_count * probe_axis_count, probe_axis_count, 1);
 	}
 
-	//end later after raster to avoid barriering on layout changes
-	//RD::get_singleton()->compute_list_end(RD::BARRIER_MASK_NO_BARRIER);
-
+	RD::get_singleton()->compute_list_end();
 	RD::get_singleton()->draw_command_end_label();
 }
 
 void GI::SDFGI::store_probes() {
-	RD::get_singleton()->barrier(RD::BARRIER_MASK_COMPUTE, RD::BARRIER_MASK_COMPUTE);
 	RD::get_singleton()->draw_command_begin_label("SDFGI Store Probes");
 
 	SDFGIShader::IntegratePushConstant push_constant;
@@ -1414,7 +1417,7 @@ void GI::SDFGI::store_probes() {
 		RD::get_singleton()->compute_list_dispatch_threads(compute_list, probe_axis_count * probe_axis_count * SDFGI::LIGHTPROBE_OCT_SIZE, probe_axis_count * SDFGI::LIGHTPROBE_OCT_SIZE, 1);
 	}
 
-	RD::get_singleton()->compute_list_end(RD::BARRIER_MASK_COMPUTE);
+	RD::get_singleton()->compute_list_end();
 
 	RD::get_singleton()->draw_command_end_label();
 }
@@ -1493,7 +1496,7 @@ void GI::SDFGI::update_cascades() {
 		cascade_data[i].pad = 0;
 	}
 
-	RD::get_singleton()->buffer_update(cascades_ubo, 0, sizeof(SDFGI::Cascade::UBO) * SDFGI::MAX_CASCADES, cascade_data, RD::BARRIER_MASK_COMPUTE);
+	RD::get_singleton()->buffer_update(cascades_ubo, 0, sizeof(SDFGI::Cascade::UBO) * SDFGI::MAX_CASCADES, cascade_data);
 }
 
 void GI::SDFGI::debug_draw(uint32_t p_view_count, const Projection *p_projections, const Transform3D &p_transform, int p_width, int p_height, RID p_render_target, RID p_texture, const Vector<RID> &p_texture_views) {
@@ -1636,7 +1639,7 @@ void GI::SDFGI::debug_draw(uint32_t p_view_count, const Projection *p_projection
 	copy_effects->copy_to_fb_rect(p_texture, texture_storage->render_target_get_rd_framebuffer(p_render_target), Rect2i(Point2i(), rtsize), true, false, false, false, RID(), p_view_count > 1);
 }
 
-void GI::SDFGI::debug_probes(RID p_framebuffer, const uint32_t p_view_count, const Projection *p_camera_with_transforms, bool p_will_continue_color, bool p_will_continue_depth) {
+void GI::SDFGI::debug_probes(RID p_framebuffer, const uint32_t p_view_count, const Projection *p_camera_with_transforms) {
 	RendererRD::MaterialStorage *material_storage = RendererRD::MaterialStorage::get_singleton();
 
 	// setup scene data
@@ -1651,7 +1654,7 @@ void GI::SDFGI::debug_probes(RID p_framebuffer, const uint32_t p_view_count, con
 			RendererRD::MaterialStorage::store_camera(p_camera_with_transforms[v], scene_data.projection[v]);
 		}
 
-		RD::get_singleton()->buffer_update(debug_probes_scene_data_ubo, 0, sizeof(SDFGIShader::DebugProbesSceneData), &scene_data, RD::BARRIER_MASK_RASTER);
+		RD::get_singleton()->buffer_update(debug_probes_scene_data_ubo, 0, sizeof(SDFGIShader::DebugProbesSceneData), &scene_data);
 	}
 
 	// setup push constant
@@ -1718,7 +1721,7 @@ void GI::SDFGI::debug_probes(RID p_framebuffer, const uint32_t p_view_count, con
 
 	SDFGIShader::ProbeDebugMode mode = p_view_count > 1 ? SDFGIShader::PROBE_DEBUG_PROBES_MULTIVIEW : SDFGIShader::PROBE_DEBUG_PROBES;
 
-	RD::DrawListID draw_list = RD::get_singleton()->draw_list_begin(p_framebuffer, RD::INITIAL_ACTION_CONTINUE, p_will_continue_color ? RD::FINAL_ACTION_CONTINUE : RD::FINAL_ACTION_READ, RD::INITIAL_ACTION_CONTINUE, p_will_continue_depth ? RD::FINAL_ACTION_CONTINUE : RD::FINAL_ACTION_READ);
+	RD::DrawListID draw_list = RD::get_singleton()->draw_list_begin(p_framebuffer, RD::INITIAL_ACTION_LOAD, RD::FINAL_ACTION_STORE, RD::INITIAL_ACTION_LOAD, RD::FINAL_ACTION_STORE);
 	RD::get_singleton()->draw_command_begin_label("Debug SDFGI");
 
 	RD::get_singleton()->draw_list_bind_render_pipeline(draw_list, gi->sdfgi_shader.debug_probes_pipeline[mode].get_render_pipeline(RD::INVALID_FORMAT_ID, RD::get_singleton()->framebuffer_get_format(p_framebuffer)));
@@ -1861,7 +1864,7 @@ void GI::SDFGI::pre_process_gi(const Transform3D &p_transform, RenderDataRD *p_r
 		}
 	}
 
-	RD::get_singleton()->buffer_update(gi->sdfgi_ubo, 0, sizeof(SDFGIData), &sdfgi_data, RD::BARRIER_MASK_COMPUTE);
+	RD::get_singleton()->buffer_update(gi->sdfgi_ubo, 0, sizeof(SDFGIData), &sdfgi_data);
 
 	/* Update dynamic lights in SDFGI cascades */
 
@@ -1983,7 +1986,7 @@ void GI::SDFGI::pre_process_gi(const Transform3D &p_transform, RenderDataRD *p_r
 		}
 
 		if (idx > 0) {
-			RD::get_singleton()->buffer_update(cascade.lights_buffer, 0, idx * sizeof(SDFGIShader::Light), lights, RD::BARRIER_MASK_COMPUTE);
+			RD::get_singleton()->buffer_update(cascade.lights_buffer, 0, idx * sizeof(SDFGIShader::Light), lights);
 		}
 
 		cascade_dynamic_light_count[i] = idx;
@@ -2046,6 +2049,8 @@ void GI::SDFGI::render_region(Ref<RenderSceneBuffersRD> p_render_buffers, int p_
 		push_constant.cascade = cascade;
 
 		if (cascades[cascade].dirty_regions != SDFGI::Cascade::DIRTY_ALL) {
+			RD::get_singleton()->buffer_copy(cascades[cascade].solid_cell_dispatch_buffer_storage, cascades[cascade].solid_cell_dispatch_buffer_call, 0, 0, sizeof(uint32_t) * 4);
+
 			RD::ComputeListID compute_list = RD::get_singleton()->compute_list_begin();
 
 			//must pre scroll existing data because not all is dirty
@@ -2053,7 +2058,7 @@ void GI::SDFGI::render_region(Ref<RenderSceneBuffersRD> p_render_buffers, int p_
 			RD::get_singleton()->compute_list_bind_uniform_set(compute_list, cascades[cascade].scroll_uniform_set, 0);
 
 			RD::get_singleton()->compute_list_set_push_constant(compute_list, &push_constant, sizeof(SDFGIShader::PreprocessPushConstant));
-			RD::get_singleton()->compute_list_dispatch_indirect(compute_list, cascades[cascade].solid_cell_dispatch_buffer, 0);
+			RD::get_singleton()->compute_list_dispatch_indirect(compute_list, cascades[cascade].solid_cell_dispatch_buffer_call, 0);
 			// no barrier do all together
 
 			RD::get_singleton()->compute_list_bind_compute_pipeline(compute_list, gi->sdfgi_shader.preprocess_pipeline[SDFGIShader::PRE_PROCESS_SCROLL_OCCLUSION]);
@@ -2142,7 +2147,7 @@ void GI::SDFGI::render_region(Ref<RenderSceneBuffersRD> p_render_buffers, int p_
 
 		//clear dispatch indirect data
 		uint32_t dispatch_indirct_data[4] = { 0, 0, 0, 0 };
-		RD::get_singleton()->buffer_update(cascades[cascade].solid_cell_dispatch_buffer, 0, sizeof(uint32_t) * 4, dispatch_indirct_data);
+		RD::get_singleton()->buffer_update(cascades[cascade].solid_cell_dispatch_buffer_storage, 0, sizeof(uint32_t) * 4, dispatch_indirct_data);
 
 		RD::ComputeListID compute_list = RD::get_singleton()->compute_list_begin();
 
@@ -2451,6 +2456,15 @@ void GI::SDFGI::render_static_lights(RenderDataRD *p_render_data, Ref<RenderScen
 		}
 	}
 
+	for (uint32_t i = 0; i < p_cascade_count; i++) {
+		ERR_CONTINUE(p_cascade_indices[i] >= cascades.size());
+
+		SDFGI::Cascade &cc = cascades[p_cascade_indices[i]];
+		if (light_count[i] > 0) {
+			RD::get_singleton()->buffer_copy(cc.solid_cell_dispatch_buffer_storage, cc.solid_cell_dispatch_buffer_call, 0, 0, sizeof(uint32_t) * 4);
+		}
+	}
+
 	/* Static Lights */
 	RD::ComputeListID compute_list = RD::get_singleton()->compute_list_begin();
 
@@ -2482,7 +2496,7 @@ void GI::SDFGI::render_static_lights(RenderDataRD *p_render_data, Ref<RenderScen
 		if (dl_push_constant.light_count > 0) {
 			RD::get_singleton()->compute_list_bind_uniform_set(compute_list, cc.sdf_direct_light_static_uniform_set, 0);
 			RD::get_singleton()->compute_list_set_push_constant(compute_list, &dl_push_constant, sizeof(SDFGIShader::DirectLightPushConstant));
-			RD::get_singleton()->compute_list_dispatch_indirect(compute_list, cc.solid_cell_dispatch_buffer, 0);
+			RD::get_singleton()->compute_list_dispatch_indirect(compute_list, cc.solid_cell_dispatch_buffer_call, 0);
 		}
 	}
 
@@ -3716,7 +3730,7 @@ void GI::setup_voxel_gi_instances(RenderDataRD *p_render_data, Ref<RenderSceneBu
 	if (p_voxel_gi_instances.size() > 0) {
 		RD::get_singleton()->draw_command_begin_label("VoxelGIs Setup");
 
-		RD::get_singleton()->buffer_update(voxel_gi_buffer, 0, sizeof(VoxelGIData) * MIN((uint64_t)MAX_VOXEL_GI_INSTANCES, p_voxel_gi_instances.size()), voxel_gi_data, RD::BARRIER_MASK_COMPUTE);
+		RD::get_singleton()->buffer_update(voxel_gi_buffer, 0, sizeof(VoxelGIData) * MIN((uint64_t)MAX_VOXEL_GI_INSTANCES, p_voxel_gi_instances.size()), voxel_gi_data);
 
 		RD::get_singleton()->draw_command_end_label();
 	}
@@ -3804,11 +3818,11 @@ void GI::process_gi(Ref<RenderSceneBuffersRD> p_render_buffers, const RID *p_nor
 		scene_data.screen_size[0] = internal_size.x;
 		scene_data.screen_size[1] = internal_size.y;
 
-		RD::get_singleton()->buffer_update(rbgi->scene_data_ubo, 0, sizeof(SceneData), &scene_data, RD::BARRIER_MASK_COMPUTE);
+		RD::get_singleton()->buffer_update(rbgi->scene_data_ubo, 0, sizeof(SceneData), &scene_data);
 	}
 
 	// Now compute the contents of our buffers.
-	RD::ComputeListID compute_list = RD::get_singleton()->compute_list_begin(true);
+	RD::ComputeListID compute_list = RD::get_singleton()->compute_list_begin();
 
 	// Render each eye separately.
 	// We need to look into whether we can make our compute shader use Multiview but not sure that works or makes a difference..
@@ -4038,8 +4052,7 @@ void GI::process_gi(Ref<RenderSceneBuffersRD> p_render_buffers, const RID *p_nor
 		}
 	}
 
-	//do barrier later to allow oeverlap
-	//RD::get_singleton()->compute_list_end(RD::BARRIER_MASK_NO_BARRIER); //no barriers, let other compute, raster and transfer happen at the same time
+	RD::get_singleton()->compute_list_end();
 	RD::get_singleton()->draw_command_end_label();
 }