rsx: Hack around using data regions as transfer targets

rpcs3/Emu/RSX/Common/TextureUtils.h

@@ -21,6 +21,13 @@ namespace rsx
 		srgb_nonlinear = 1
 	};
 
+	enum surface_usage_flags : u32
+	{
+		unknown = 0,
+		attachment = 1,
+		storage = 2
+	};
+
 	//Sampled image descriptor
 	struct sampled_image_descriptor_base
 	{

rpcs3/Emu/RSX/Common/surface_store.h

@@ -134,6 +134,8 @@ namespace rsx
 		GcmTileInfo *tile = nullptr;
 		rsx::surface_antialiasing write_aa_mode = rsx::surface_antialiasing::center_1_sample;
 
+		flags32_t usage = surface_usage_flags::unknown;
+
 		union
 		{
 			rsx::surface_color_format gcm_color_format;

rpcs3/Emu/RSX/Common/texture_cache.h

@@ -2459,40 +2459,50 @@ namespace rsx
 					return {};
 				}
 
-				if (list.back().is_clipped && !allow_clipped)
+				for (auto It = list.rbegin(); It != list.rend(); ++It)
 				{
-					for (auto It = list.rbegin(); It != list.rend(); ++It)
+					if (!(It->surface->usage & rsx::surface_usage_flags::attachment))
 					{
-						if (!It->is_clipped)
+						// HACK
-						{
+						// TODO: Properly analyse the input here to determine if it can properly fit what we need
-							return *It;
+						// This is a problem due to chunked transfer
-						}
+						// First 2 512x720 blocks go into a cpu-side buffer but suddenly when its time to render the final 256x720
-
+						// it falls onto some storage buffer in surface cache that has bad dimensions
-						auto _w = u32(It->width * It->surface->get_bpp()) / bpp;
+						// Proper solution is to always merge when a cpu resource is created (it should absorb the render targets in range)
-						auto _h = u32(It->height);
+						// We then should not have any 'dst-is-rendertarget' surfaces in use
-						get_rsx_dimensions(_w, _h, It->surface);
+						// Option 2: Make surfaces here part of surface cache and do not pad them for optimization
-
+						// Surface cache is good at merging for resolve operations. This keeps integrity even when drawing to the rendertgargets
-						if (_w < width)
+						// This option needs a lot more work
-						{
+						continue;
-							if ((_w * scale_x) <= 1.f)
-								continue;
-						}
-
-						if (_h < height)
-						{
-							if ((_h * scale_y) <= 1.f)
-								continue;
-						}
-
-						// Some surface exists, but its size is questionable
-						// Opt to re-upload (needs WCB/WDB to work properly)
-						break;
 					}
 
-					return {};
+					if (!It->is_clipped || allow_clipped)
+					{
+						return *It;
+					}
+
+					auto _w = u32(It->width * It->surface->get_bpp()) / bpp;
+					auto _h = u32(It->height);
+					get_rsx_dimensions(_w, _h, It->surface);
+
+					if (_w < width)
+					{
+						if ((_w * scale_x) <= 1.f)
+							continue;
+					}
+
+					if (_h < height)
+					{
+						if ((_h * scale_y) <= 1.f)
+							continue;
+					}
+
+					// Some surface exists, but its size is questionable
+					// Opt to re-upload (needs WCB/WDB to work properly)
+					break;
 				}
 
-				return list.back();
+				return {};
 			};
 
 			// Check if src/dst are parts of render targets

rpcs3/Emu/RSX/GL/GLRenderTargets.h

@@ -188,6 +188,7 @@ struct gl_render_target_traits
 		std::array<GLenum, 4> native_layout = { (GLenum)format.swizzle.a, (GLenum)format.swizzle.r, (GLenum)format.swizzle.g, (GLenum)format.swizzle.b };
 		result->set_native_component_layout(native_layout);
 
+		result->usage = rsx::surface_usage_flags::attachment;
 		result->set_cleared(false);
 		result->queue_tag(address);
 		result->add_ref();
@@ -215,6 +216,7 @@ struct gl_render_target_traits
 		result->set_native_component_layout(native_layout);
 		result->set_format(surface_depth_format);
 
+		result->usage = rsx::surface_usage_flags::attachment;
 		result->set_cleared(false);
 		result->queue_tag(address);
 		result->add_ref();
@@ -234,6 +236,7 @@ struct gl_render_target_traits
 			const auto new_h = rsx::apply_resolution_scale(prev.height, true, ref->get_surface_height());
 
 			sink.reset(new gl::render_target(new_w, new_h, internal_format));
+			sink->usage = rsx::surface_usage_flags::storage;
 			sink->add_ref();
 			sink->format_info = ref->format_info;
 			sink->set_native_pitch(prev.width * ref->get_bpp());
@@ -272,10 +275,17 @@ struct gl_render_target_traits
 		info->bpp = surface->get_bpp();
 	}
 
-	static void prepare_rtt_for_drawing(gl::command_context&, gl::render_target*) {}
+	static void prepare_rtt_for_drawing(gl::command_context&, gl::render_target* rtt)
+	{
+		rtt->usage |= rsx::surface_usage_flags::attachment;
+	}
+
+	static void prepare_ds_for_drawing(gl::command_context&, gl::render_target* ds)
+	{
+		ds->usage |= rsx::surface_usage_flags::attachment;
+	}
+
 	static void prepare_rtt_for_sampling(gl::command_context&, gl::render_target*) {}
-	
-	static void prepare_ds_for_drawing(gl::command_context&, gl::render_target*) {}
 	static void prepare_ds_for_sampling(gl::command_context&, gl::render_target*) {}
 
 	static
@@ -292,6 +302,7 @@ struct gl_render_target_traits
 		surface->queue_tag(address);
 		surface->set_cleared(false);
 		surface->last_use_tag = 0;
+		surface->usage = rsx::surface_usage_flags::unknown;
 	}
 
 	static

rpcs3/Emu/RSX/VK/VKRenderTargets.h

@@ -197,6 +197,7 @@ namespace rsx
 			change_image_layout(cmd, rtt.get(), VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL, vk::get_image_subresource_range(0, 0, 1, 1, VK_IMAGE_ASPECT_COLOR_BIT));
 
 			rtt->set_format(format);
+			rtt->usage = rsx::surface_usage_flags::attachment;
 			rtt->native_component_map = fmt.second;
 			rtt->rsx_pitch = (u16)pitch;
 			rtt->native_pitch = (u16)width * get_format_block_size_in_bytes(format);
@@ -235,7 +236,9 @@ namespace rsx
 				VK_IMAGE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT| VK_IMAGE_USAGE_TRANSFER_SRC_BIT| VK_IMAGE_USAGE_TRANSFER_DST_BIT|VK_IMAGE_USAGE_SAMPLED_BIT,
 				0));
 
+
 			ds->set_format(format);
+			ds->usage = rsx::surface_usage_flags::attachment;
 			ds->native_component_map = { VK_COMPONENT_SWIZZLE_R, VK_COMPONENT_SWIZZLE_R, VK_COMPONENT_SWIZZLE_R, VK_COMPONENT_SWIZZLE_R };
 			change_image_layout(cmd, ds.get(), VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL, range);
 
@@ -277,6 +280,7 @@ namespace rsx
 
 				sink->add_ref();
 				sink->format_info = ref->format_info;
+				sink->usage = rsx::surface_usage_flags::storage;
 				sink->native_component_map = ref->native_component_map;
 				sink->native_pitch = u16(prev.width * ref->get_bpp());
 				sink->surface_width = prev.width;
@@ -315,6 +319,7 @@ namespace rsx
 		{
 			surface->change_layout(cmd, VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL);
 			surface->frame_tag = 0;
+			surface->usage |= rsx::surface_usage_flags::attachment;
 		}
 
 		static void prepare_rtt_for_sampling(vk::command_buffer& cmd, vk::render_target *surface)
@@ -326,6 +331,7 @@ namespace rsx
 		{
 			surface->change_layout(cmd, VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL);
 			surface->frame_tag = 0;
+			surface->usage |= rsx::surface_usage_flags::attachment;
 		}
 
 		static void prepare_ds_for_sampling(vk::command_buffer& cmd, vk::render_target *surface)
@@ -345,6 +351,7 @@ namespace rsx
 			surface->queue_tag(address);
 			surface->dirty = true;
 			surface->last_use_tag = 0;
+			surface->usage = rsx::surface_usage_flags::unknown;
 		}
 
 		static void notify_surface_invalidated(const std::unique_ptr<vk::render_target> &surface)

rpcs3/Emu/RSX/rsx_utils.h

@@ -27,6 +27,11 @@ namespace rsx
 	using utils::page_end;
 	using utils::next_page;
 
+	using flags64_t = uint64_t;
+	using flags32_t = uint32_t;
+	using flags16_t = uint16_t;
+	using flags8_t = uint8_t;
+
 	// Definitions
 	class thread;
 	extern thread* g_current_renderer;