rsx: Texture cache fixes - Optionally attempt to merge framebuffers into an atlas if partial resources are missing - Support for data update requests to the temporary subresource handler This is useful for framebuffer feedback loops where a new copy is needed after every draw call (resource is always dirty)

rpcs3/Emu/RSX/Common/surface_store.h

@@ -35,6 +35,20 @@ namespace rsx
 		{}
 	};
 
+	template <typename surface_type>
+	struct surface_overlap_info_t
+	{
+		surface_type surface = nullptr;
+		bool is_depth = false;
+
+		u16 src_x = 0;
+		u16 src_y = 0;
+		u16 dst_x = 0;
+		u16 dst_y = 0;
+		u16 width = 0;
+		u16 height = 0;
+	};
+
 	struct surface_format_info
 	{
 		u32 surface_width;
@@ -120,6 +134,7 @@ namespace rsx
 		using command_list_type = typename Traits::command_list_type;
 		using download_buffer_object = typename Traits::download_buffer_object;
 		using surface_subresource = surface_subresource_storage<surface_type>;
+		using surface_overlap_info = surface_overlap_info_t<surface_type>;
 
 		std::unordered_map<u32, surface_storage_type> m_render_targets_storage = {};
 		std::unordered_map<u32, surface_storage_type> m_depth_stencil_storage = {};
@@ -846,5 +861,61 @@ namespace rsx
 
 			return{};
 		}
+
+		std::vector<surface_overlap_info> get_merged_texture_memory_region(u32 texaddr, u32 required_width, u32 required_height, u32 required_pitch, u32 bpp)
+		{
+			std::vector<surface_overlap_info> result;
+			const u32 limit = texaddr + (required_pitch * required_height);
+
+			auto process_list_function = [&](std::unordered_map<u32, surface_storage_type>& data, bool is_depth)
+			{
+				for (auto &tex_info : data)
+				{
+					auto this_address = std::get<0>(tex_info);
+					if (this_address > limit)
+						continue;
+
+					auto surface = std::get<1>(tex_info).get();
+					const auto pitch = surface->get_rsx_pitch();
+					if (pitch != required_pitch)
+						continue;
+
+					const auto texture_size = pitch * surface->get_surface_height();
+					if ((this_address + texture_size) <= texaddr)
+						continue;
+
+					surface_overlap_info info;
+					info.surface = surface;
+					info.is_depth = is_depth;
+
+					if (this_address < texaddr)
+					{
+						auto offset = texaddr - this_address;
+						info.src_y = (offset / required_pitch);
+						info.src_x = (offset % required_pitch) / bpp;
+						info.dst_x = 0;
+						info.dst_y = 0;
+						info.width = std::min<u32>(required_width, surface->get_surface_width() - info.src_x);
+						info.height = std::min<u32>(required_height, surface->get_surface_height() - info.src_y);
+					}
+					else
+					{
+						auto offset = this_address - texaddr;
+						info.src_x = 0;
+						info.src_y = 0;
+						info.dst_y = (offset / required_pitch);
+						info.dst_x = (offset % required_pitch) / bpp;
+						info.width = std::min<u32>(surface->get_surface_width(), required_width - info.dst_x);
+						info.height = std::min<u32>(surface->get_surface_height(), required_height - info.dst_y);
+					}
+
+					result.push_back(info);
+				}
+			};
+
+			process_list_function(m_render_targets_storage, false);
+			process_list_function(m_depth_stencil_storage, true);
+			return result;
+		}
 	};
 }

rpcs3/Emu/RSX/Common/texture_cache.h

@@ -251,10 +251,22 @@ namespace rsx
 			u32 address_range = 0;
 		};
 
+		struct copy_region_descriptor
+		{
+			image_resource_type src;
+			u16 src_x;
+			u16 src_y;
+			u16 dst_x;
+			u16 dst_y;
+			u16 w;
+			u16 h;
+		};
+
 		struct deferred_subresource
 		{
 			image_resource_type external_handle = 0;
 			std::array<image_resource_type, 6> external_cubemap_sources;
+			std::vector<copy_region_descriptor> sections_to_copy;
 			u32 base_address = 0;
 			u32 gcm_format = 0;
 			u16 x = 0;
@@ -262,6 +274,8 @@ namespace rsx
 			u16 width = 0;
 			u16 height = 0;
 			bool is_cubemap = false;
+			bool is_copy_cmd = false;
+			bool update_cached = false;
 
 			deferred_subresource()
 			{}
@@ -370,6 +384,8 @@ namespace rsx
 		virtual void set_up_remap_vector(section_storage_type& section, const std::pair<std::array<u8, 4>, std::array<u8, 4>>& remap_vector) = 0;
 		virtual void insert_texture_barrier(commandbuffer_type&, image_storage_type* tex) = 0;
 		virtual image_view_type generate_cubemap_from_images(commandbuffer_type&, u32 gcm_format, u16 size, const std::array<image_resource_type, 6>& sources) = 0;
+		virtual image_view_type generate_atlas_from_images(commandbuffer_type&, u32 gcm_format, u16 width, u16 height, const std::vector<copy_region_descriptor>& sections_to_copy) = 0;
+		virtual void update_image_contents(commandbuffer_type&, image_view_type dst, image_resource_type src, u16 width, u16 height) = 0;
 
 		constexpr u32 get_block_size() const { return 0x1000000; }
 		inline u32 get_block_address(u32 address) const { return (address & ~0xFFFFFF); }
@@ -1214,14 +1230,19 @@ namespace rsx
 					found_desc.width != desc.width || found_desc.height != desc.height)
 					continue;
 
+				if (desc.update_cached)
+					update_image_contents(cmd, It->second.second, desc.external_handle, desc.width, desc.height);
+
 				return It->second.second;
 			}
 
 			image_view_type result = 0;
-			if (!desc.is_cubemap)
+			if (desc.is_copy_cmd)
-				result = create_temporary_subresource_view(cmd, &desc.external_handle, desc.gcm_format, desc.x, desc.y, desc.width, desc.height);
+				result = generate_atlas_from_images(cmd, desc.gcm_format, desc.width, desc.height, desc.sections_to_copy);
-			else
+			else if (desc.is_cubemap)
 				result = generate_cubemap_from_images(cmd, desc.gcm_format, desc.width, desc.external_cubemap_sources);
+			else
+				result = create_temporary_subresource_view(cmd, &desc.external_handle, desc.gcm_format, desc.x, desc.y, desc.width, desc.height);
 
 			if (result)
 			{
@@ -1238,7 +1259,7 @@ namespace rsx
 
 		template <typename render_target_type, typename surface_store_type>
 		sampled_image_descriptor process_framebuffer_resource(commandbuffer_type& cmd, render_target_type texptr, u32 texaddr, u32 gcm_format, surface_store_type& m_rtts,
-				u16 tex_width, u16 tex_height, rsx::texture_dimension_extended extended_dimension, bool is_depth)
+				u16 tex_width, u16 tex_height, u16 tex_pitch, rsx::texture_dimension_extended extended_dimension, bool is_depth)
 		{
 			const u32 format = gcm_format & ~(CELL_GCM_TEXTURE_UN | CELL_GCM_TEXTURE_LN);
 			const auto surface_width = texptr->get_surface_width();
@@ -1321,7 +1342,43 @@ namespace rsx
 				scale_y = 0.f;
 			}
 
+			if (internal_width > surface_width || internal_height > surface_height)
+			{
+				auto bpp = get_format_block_size_in_bytes(format);
+				auto overlapping = m_rtts.get_merged_texture_memory_region(texaddr, tex_width, tex_height, tex_pitch, bpp);
+
+				if (overlapping.size() > 1)
+				{
+					const auto w = rsx::apply_resolution_scale(internal_width, true);
+					const auto h = rsx::apply_resolution_scale(internal_height, true);
+
+					sampled_image_descriptor result = { texptr->get_surface(), texaddr, format, 0, 0, w, h,
+						texture_upload_context::framebuffer_storage, is_depth, scale_x, scale_y,
+						rsx::texture_dimension_extended::texture_dimension_2d };
+
+					result.external_subresource_desc.is_copy_cmd = true;
+					result.external_subresource_desc.sections_to_copy.reserve(overlapping.size());
+
+					for (auto &section : overlapping)
+					{
+						result.external_subresource_desc.sections_to_copy.push_back
+						({
+							section.surface->get_surface(),
+							rsx::apply_resolution_scale(section.src_x, true),
+							rsx::apply_resolution_scale(section.src_y, true),
+							rsx::apply_resolution_scale(section.dst_x, true),
+							rsx::apply_resolution_scale(section.dst_y, true),
+							rsx::apply_resolution_scale(section.width, true),
+							rsx::apply_resolution_scale(section.height, true)
+						});
+					}
+
+					return result;
+				}
+			}
+
 			bool requires_processing = surface_width > internal_width || surface_height > internal_height;
+			bool update_subresource_cache = false;
 			if (!requires_processing)
 			{
 				//NOTE: The scale also accounts for sampling outside the RTT region, e.g render to one quadrant but send whole texture for sampling
@@ -1343,6 +1400,7 @@ namespace rsx
 							{
 								LOG_WARNING(RSX, "Attempting to sample a currently bound render target @ 0x%x", texaddr);
 								requires_processing = true;
+								update_subresource_cache = true;
 								break;
 							}
 							else
@@ -1362,6 +1420,7 @@ namespace rsx
 						{
 							LOG_WARNING(RSX, "Attempting to sample a currently bound depth surface @ 0x%x", texaddr);
 							requires_processing = true;
+							update_subresource_cache = true;
 						}
 						else
 						{
@@ -1376,8 +1435,12 @@ namespace rsx
 			{
 				const auto w = rsx::apply_resolution_scale(internal_width, true);
 				const auto h = rsx::apply_resolution_scale(internal_height, true);
-				return{ texptr->get_surface(), texaddr, format, 0, 0, w, h, texture_upload_context::framebuffer_storage,
+
+				sampled_image_descriptor result = { texptr->get_surface(), texaddr, format, 0, 0, w, h, texture_upload_context::framebuffer_storage,
 					is_depth, scale_x, scale_y, rsx::texture_dimension_extended::texture_dimension_2d };
+
+				result.external_subresource_desc.update_cached = update_subresource_cache;
+				return result;
 			}
 
 			return{ texptr->get_view(), texture_upload_context::framebuffer_storage, is_depth, scale_x, scale_y, rsx::texture_dimension_extended::texture_dimension_2d };
@@ -1400,8 +1463,9 @@ namespace rsx
 			const auto extended_dimension = tex.get_extended_texture_dimension();
 			u16 depth = 0;
 			u16 tex_height = (u16)tex.height();
-			u16 tex_pitch = tex.pitch();
 			const u16 tex_width = tex.width();
+			u16 tex_pitch = is_compressed_format? (u16)(get_texture_size(tex) / tex_height) : tex.pitch(); //NOTE: Compressed textures dont have a real pitch (tex_size = (w*h)/6)
+			if (tex_pitch == 0) tex_pitch = tex_width * get_format_block_size_in_bytes(format);
 
 			switch (extended_dimension)
 			{
@@ -1428,7 +1492,7 @@ namespace rsx
 				{
 					if (test_framebuffer(texaddr + texptr->raster_address_offset))
 					{
-						return process_framebuffer_resource(cmd, texptr, texaddr, tex.format(), m_rtts, tex_width, tex_height, extended_dimension, false);
+						return process_framebuffer_resource(cmd, texptr, texaddr, tex.format(), m_rtts, tex_width, tex_height, tex_pitch, extended_dimension, false);
 					}
 					else
 					{
@@ -1441,7 +1505,7 @@ namespace rsx
 				{
 					if (test_framebuffer(texaddr + texptr->raster_address_offset))
 					{
-						return process_framebuffer_resource(cmd, texptr, texaddr, tex.format(), m_rtts, tex_width, tex_height, extended_dimension, true);
+						return process_framebuffer_resource(cmd, texptr, texaddr, tex.format(), m_rtts, tex_width, tex_height, tex_pitch, extended_dimension, true);
 					}
 					else
 					{
@@ -1451,9 +1515,6 @@ namespace rsx
 				}
 			}
 
-			tex_pitch = is_compressed_format? (u16)(get_texture_size(tex) / tex_height) : tex_pitch; //NOTE: Compressed textures dont have a real pitch (tex_size = (w*h)/6)
-			if (tex_pitch == 0) tex_pitch = tex_width * get_format_block_size_in_bytes(format);
-
 			const bool unnormalized = (tex.format() & CELL_GCM_TEXTURE_UN) != 0;
 			f32 scale_x = (unnormalized) ? (1.f / tex_width) : 1.f;
 			f32 scale_y = (unnormalized) ? (1.f / tex_height) : 1.f;

rpcs3/Emu/RSX/GL/GLTextureCache.h

@@ -654,7 +654,7 @@ namespace gl
 			m_temporary_surfaces.resize(0);
 		}
 
-		u32 create_temporary_subresource_impl(u32 src_id, GLenum sized_internal_fmt, GLenum dst_type, u16 x, u16 y, u16 width, u16 height)
+		u32 create_temporary_subresource_impl(u32 src_id, GLenum sized_internal_fmt, GLenum dst_type, u16 x, u16 y, u16 width, u16 height, bool copy = true)
 		{
 			u32 dst_id = 0;
 
@@ -686,6 +686,8 @@ namespace gl
 
 			m_temporary_surfaces.push_back(dst_id);
 
+			if (copy)
+			{
 				//Empty GL_ERROR
 				glGetError();
 
@@ -698,6 +700,7 @@ namespace gl
 					LOG_WARNING(RSX, "Failed to copy image subresource with GL error 0x%X", err);
 					return 0;
 				}
+			}
 
 			return dst_id;
 		}
@@ -815,6 +818,26 @@ namespace gl
 			return dst_id;
 		}
 
+		u32 generate_atlas_from_images(void*&, u32 gcm_format, u16 width, u16 height, const std::vector<copy_region_descriptor>& sections_to_copy) override
+		{
+			const GLenum ifmt = gl::get_sized_internal_format(gcm_format);
+			auto result = create_temporary_subresource_impl(sections_to_copy.front().src, ifmt, GL_TEXTURE_2D, 0, 0, width, height, false);
+
+			for (const auto &region : sections_to_copy)
+			{
+				glCopyImageSubData(region.src, GL_TEXTURE_2D, 0, region.src_x, region.src_y, 0,
+					result, GL_TEXTURE_2D, 0, region.dst_x, region.dst_y, 0, region.w, region.h, 1);
+			}
+
+			return result;
+		}
+
+		void update_image_contents(void*&, u32 dst, u32 src, u16 width, u16 height) override
+		{
+			glCopyImageSubData(src, GL_TEXTURE_2D, 0, 0, 0, 0,
+					dst, GL_TEXTURE_2D, 0, 0, 0, 0, width, height, 1);
+		}
+
 		cached_texture_section* create_new_texture(void*&, u32 rsx_address, u32 rsx_size, u16 width, u16 height, u16 depth, u16 mipmaps, u32 gcm_format,
 				rsx::texture_upload_context context, rsx::texture_dimension_extended type, rsx::texture_create_flags flags,
 				const std::pair<std::array<u8, 4>, std::array<u8, 4>>& /*remap_vector*/) override

rpcs3/Emu/RSX/VK/VKTextureCache.h

@@ -564,7 +564,7 @@ namespace vk
 			tex.destroy();
 		}
 
-		vk::image_view* create_temporary_subresource_view_impl(vk::command_buffer& cmd, vk::image* source, VkImageType image_type, VkImageViewType view_type, u32 gcm_format, u16 x, u16 y, u16 w, u16 h)
+		vk::image_view* create_temporary_subresource_view_impl(vk::command_buffer& cmd, vk::image* source, VkImageType image_type, VkImageViewType view_type, u32 gcm_format, u16 x, u16 y, u16 w, u16 h, bool copy = true)
 		{
 			VkImageAspectFlags aspect = VK_IMAGE_ASPECT_COLOR_BIT;
 
@@ -600,6 +600,8 @@ namespace vk
 			VkImageSubresourceRange view_range = { aspect & ~(VK_IMAGE_ASPECT_STENCIL_BIT), 0, 1, 0, 1 };
 			view.reset(new vk::image_view(*vk::get_current_renderer(), image->value, view_type, dst_format, source->native_component_map, view_range));
 
+			if (copy)
+			{
 				VkImageLayout old_src_layout = source->current_layout;
 
 				vk::change_image_layout(cmd, image.get(), VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, subresource_range);
@@ -615,6 +617,7 @@ namespace vk
 				vkCmdCopyImage(cmd, source->value, source->current_layout, image->value, image->current_layout, 1, &copy_rgn);
 				vk::change_image_layout(cmd, image.get(), VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, subresource_range);
 				vk::change_image_layout(cmd, source, old_src_layout, subresource_range);
+			}
 
 			const u32 resource_memory = w * h * 4; //Rough approximate
 			m_discardable_storage.push_back({ image, view });
@@ -706,6 +709,85 @@ namespace vk
 			return m_discardable_storage.back().view.get();
 		}
 
+		vk::image_view* generate_atlas_from_images(vk::command_buffer& cmd, u32 gcm_format, u16 width, u16 height, const std::vector<copy_region_descriptor>& sections_to_copy) override
+		{
+			auto result = create_temporary_subresource_view_impl(cmd, sections_to_copy.front().src, VK_IMAGE_TYPE_2D,
+					VK_IMAGE_VIEW_TYPE_2D, gcm_format, 0, 0, width, height, false);
+
+			VkImage dst = result->info.image;
+			VkImageAspectFlags aspect = VK_IMAGE_ASPECT_COLOR_BIT;
+
+			switch (sections_to_copy.front().src->info.format)
+			{
+			case VK_FORMAT_D16_UNORM:
+				aspect = VK_IMAGE_ASPECT_DEPTH_BIT;
+				break;
+			case VK_FORMAT_D24_UNORM_S8_UINT:
+			case VK_FORMAT_D32_SFLOAT_S8_UINT:
+				aspect = VK_IMAGE_ASPECT_DEPTH_BIT | VK_IMAGE_ASPECT_STENCIL_BIT;
+				break;
+			}
+
+			VkImageSubresourceRange subresource_range = { aspect & ~(VK_IMAGE_ASPECT_STENCIL_BIT), 0, 1, 0, 1 };
+			vk::change_image_layout(cmd, dst, VK_IMAGE_LAYOUT_UNDEFINED, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, subresource_range);
+
+			for (const auto &region : sections_to_copy)
+			{
+				VkImageLayout old_src_layout = region.src->current_layout;
+				vk::change_image_layout(cmd, region.src, VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL, subresource_range);
+
+				VkImageCopy copy_rgn;
+				copy_rgn.srcOffset = { region.src_x, region.src_y, 0 };
+				copy_rgn.dstOffset = { region.dst_x, region.dst_y, 0 };
+				copy_rgn.dstSubresource = { aspect, 0, 0, 1 };
+				copy_rgn.srcSubresource = { aspect, 0, 0, 1 };
+				copy_rgn.extent = { region.w, region.h, 1 };
+
+				vkCmdCopyImage(cmd, region.src->value, VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL, dst, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL,
+					1, &copy_rgn);
+
+				vk::change_image_layout(cmd, region.src, old_src_layout, subresource_range);
+			}
+
+			vk::change_image_layout(cmd, dst, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, subresource_range);
+			return result;
+		}
+
+		void update_image_contents(vk::command_buffer& cmd, vk::image_view* dst_view, vk::image* src, u16 width, u16 height) override
+		{
+			VkImage dst = dst_view->info.image;
+			VkImageAspectFlags aspect = VK_IMAGE_ASPECT_COLOR_BIT;
+
+			switch (src->info.format)
+			{
+			case VK_FORMAT_D16_UNORM:
+				aspect = VK_IMAGE_ASPECT_DEPTH_BIT;
+				break;
+			case VK_FORMAT_D24_UNORM_S8_UINT:
+			case VK_FORMAT_D32_SFLOAT_S8_UINT:
+				aspect = VK_IMAGE_ASPECT_DEPTH_BIT | VK_IMAGE_ASPECT_STENCIL_BIT;
+				break;
+			}
+
+			VkImageSubresourceRange subresource_range = { aspect, 0, 1, 0, 1 };
+			vk::change_image_layout(cmd, dst, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, subresource_range);
+
+			VkImageLayout old_src_layout = src->current_layout;
+			vk::change_image_layout(cmd, src, VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL, subresource_range);
+
+			VkImageCopy copy_rgn;
+			copy_rgn.srcOffset = { 0, 0, 0 };
+			copy_rgn.dstOffset = { 0, 0, 0 };
+			copy_rgn.dstSubresource = { aspect & ~(VK_IMAGE_ASPECT_DEPTH_BIT), 0, 0, 1 };
+			copy_rgn.srcSubresource = { aspect & ~(VK_IMAGE_ASPECT_DEPTH_BIT), 0, 0, 1 };
+			copy_rgn.extent = { width, height, 1 };
+
+			vkCmdCopyImage(cmd, src->value, VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL, dst, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, 1, &copy_rgn);
+
+			vk::change_image_layout(cmd, src, old_src_layout, subresource_range);
+			vk::change_image_layout(cmd, dst, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, subresource_range);
+		}
+
 		cached_texture_section* create_new_texture(vk::command_buffer& cmd, u32 rsx_address, u32 rsx_size, u16 width, u16 height, u16 depth, u16 mipmaps, u32 gcm_format,
 				rsx::texture_upload_context context, rsx::texture_dimension_extended type, rsx::texture_create_flags flags,
 				const std::pair<std::array<u8, 4>, std::array<u8, 4>>& remap_vector) override