rsx: Tweaks

- Optimize get_surface_subresource
- Add check_program_status time to draw call setup statistics. It can slow down games significantly

rpcs3/Emu/RSX/Common/surface_store.h

@@ -557,52 +557,63 @@ namespace rsx
 			if (surface_address > texaddr)
 				return false;
 
-			u32 offset = texaddr - surface_address;
+			const u32 offset = texaddr - surface_address;
-			if (texaddr >= surface_address)
+			if (offset == 0)
 			{
-				if (offset == 0)
+				*x = 0;
-				{
+				*y = 0;
-					is_subslice = true;
+				return true;
-				}
+			}
-				else
+			else
-				{
+			{
-					surface_format_info info;
+				surface_format_info info;
-					Traits::get_surface_info(surface, &info);
+				Traits::get_surface_info(surface, &info);
 
-					u32 range = info.rsx_pitch * info.surface_height;
+				u32 range = info.rsx_pitch * info.surface_height;
-					if (double_height) range *= 2;
+				if (double_height) range <<= 1;
 
-					if (offset < range)
+				if (offset < range)
+				{
+					const u32 y = (offset / info.rsx_pitch);
+					u32 x = (offset % info.rsx_pitch) / info.bpp;
+
+					if (scale_to_fit)
 					{
-						const u32 y = (offset / info.rsx_pitch);
+						const f32 x_scale = (f32)info.rsx_pitch / info.native_pitch;
-						u32 x = (offset % info.rsx_pitch) / info.bpp;
+						x = (u32)((f32)x / x_scale);
-
-						if (scale_to_fit)
-						{
-							const f32 x_scale = (f32)info.rsx_pitch / info.native_pitch;
-							x = (u32)((f32)x / x_scale);
-						}
-
-						x_offset = x;
-						y_offset = y;
-
-						if (double_height) y_offset /= 2;
-						is_subslice = true;
 					}
-				}
 
-				if (is_subslice)
+					x_offset = x;
-				{
+					y_offset = y;
-					*x = x_offset;
-					*y = y_offset;
 
-					return true;
+					if (double_height) y_offset /= 2;
+					is_subslice = true;
 				}
 			}
 
+			if (is_subslice)
+			{
+				*x = x_offset;
+				*y = y_offset;
+
+				return true;
+			}
+
 			return false;
 		}
 
+		//Fast hit test
+		inline bool surface_overlaps_address_fast(surface_type surface, u32 surface_address, u32 texaddr)
+		{
+			if (surface_address > texaddr)
+				return false;
+
+			const u32 offset = texaddr - surface_address;
+			const u32 range = surface->get_rsx_pitch() * surface->get_surface_height();
+
+			return (offset < range);
+		}
+
 		bool address_is_bound(u32 address, bool is_depth) const
 		{
 			if (is_depth)
@@ -629,7 +640,8 @@ namespace rsx
 			return true;
 		}
 
-		surface_subresource get_surface_subresource_if_applicable(u32 texaddr, u16 requested_width, u16 requested_height, u16 requested_pitch, bool scale_to_fit = false, bool crop = false, bool ignore_depth_formats = false, bool double_height = false)
+		surface_subresource get_surface_subresource_if_applicable(u32 texaddr, u16 requested_width, u16 requested_height, u16 requested_pitch,
+				bool scale_to_fit = false, bool crop = false, bool ignore_depth_formats = false, bool ignore_color_formats = false, bool double_height = false)
 		{
 			auto test_surface = [&](surface_type surface, u32 this_address, u16 &x_offset, u16 &y_offset, u16 &w, u16 &h, bool &clipped)
 			{
@@ -638,12 +650,6 @@ namespace rsx
 					surface_format_info info;
 					Traits::get_surface_info(surface, &info);
 
-					if (info.rsx_pitch != requested_pitch)
-						return false;
-
-					if (requested_width == 0 || requested_height == 0)
-						return true;
-
 					u16 real_width = requested_width;
 
 					if (scale_to_fit)
@@ -696,26 +702,55 @@ namespace rsx
 			u16  w;
 			u16  h;
 
-			for (auto &tex_info : m_render_targets_storage)
+			if (!ignore_color_formats)
 			{
-				u32 this_address = std::get<0>(tex_info);
+				for (auto &tex_info : m_render_targets_storage)
-				surface = std::get<1>(tex_info).get();
+				{
+					const u32 this_address = std::get<0>(tex_info);
+					if (texaddr < this_address)
+						continue;
 
-				if (test_surface(surface, this_address, x_offset, y_offset, w, h, clipped))
+					surface = std::get<1>(tex_info).get();
-					return { surface, x_offset, y_offset, w, h, address_is_bound(this_address, false), false, clipped };
+					if (surface->get_rsx_pitch() != requested_pitch)
+						continue;
+
+					if (requested_width == 0 || requested_height == 0)
+					{
+						if (!surface_overlaps_address_fast(surface, this_address, texaddr))
+							continue;
+						else
+							return{ surface, 0, 0, 0, 0, false, false, false };
+					}
+
+					if (test_surface(surface, this_address, x_offset, y_offset, w, h, clipped))
+						return{ surface, x_offset, y_offset, w, h, address_is_bound(this_address, false), false, clipped };
+				}
 			}
 
-			if (ignore_depth_formats)
+			if (!ignore_depth_formats)
-				return{};
-
-			//Check depth surfaces for overlap
-			for (auto &tex_info : m_depth_stencil_storage)
 			{
-				u32 this_address = std::get<0>(tex_info);
+				//Check depth surfaces for overlap
-				surface = std::get<1>(tex_info).get();
+				for (auto &tex_info : m_depth_stencil_storage)
+				{
+					const u32 this_address = std::get<0>(tex_info);
+					if (texaddr < this_address)
+						continue;
 
-				if (test_surface(surface, this_address, x_offset, y_offset, w, h, clipped))
+					surface = std::get<1>(tex_info).get();
-					return { surface, x_offset, y_offset, w, h, address_is_bound(this_address, true), true, clipped };
+					if (surface->get_rsx_pitch() != requested_pitch)
+						continue;
+
+					if (requested_width == 0 || requested_height == 0)
+					{
+						if (!surface_overlaps_address_fast(surface, this_address, texaddr))
+							continue;
+						else
+							return{ surface, 0, 0, 0, 0, false, true, false };
+					}
+
+					if (test_surface(surface, this_address, x_offset, y_offset, w, h, clipped))
+						return{ surface, x_offset, y_offset, w, h, address_is_bound(this_address, true), true, clipped };
+				}
 			}
 
 			return{};

rpcs3/Emu/RSX/Common/texture_cache.h

@@ -1087,7 +1087,7 @@ namespace rsx
 			}
 
 			//Check if src/dst are parts of render targets
-			auto dst_subres = m_rtts.get_surface_subresource_if_applicable(dst_address, dst.width, dst.clip_height, dst.pitch, true, true, false, dst.compressed_y);
+			auto dst_subres = m_rtts.get_surface_subresource_if_applicable(dst_address, dst.width, dst.clip_height, dst.pitch, true, true, false, false, dst.compressed_y);
 			dst_is_render_target = dst_subres.surface != nullptr;
 
 			if (dst_is_render_target && dst_subres.surface->get_native_pitch() != dst.pitch)
@@ -1099,7 +1099,7 @@ namespace rsx
 			}
 
 			//TODO: Handle cases where src or dst can be a depth texture while the other is a color texture - requires a render pass to emulate
-			auto src_subres = m_rtts.get_surface_subresource_if_applicable(framebuffer_src_address, src_w, src_h, src.pitch, true, true, false, src.compressed_y);
+			auto src_subres = m_rtts.get_surface_subresource_if_applicable(framebuffer_src_address, src_w, src_h, src.pitch, true, true, false, false, src.compressed_y);
 			src_is_render_target = src_subres.surface != nullptr;
 
 			if (src_is_render_target && src_subres.surface->get_native_pitch() != src.pitch)

rpcs3/Emu/RSX/GL/GLGSRender.cpp

@@ -318,12 +318,17 @@ namespace
 
 void GLGSRender::end()
 {
+	std::chrono::time_point<steady_clock> state_check_start = steady_clock::now();
+
 	if (skip_frame || !framebuffer_status_valid || (conditional_render_enabled && conditional_render_test_failed) || !check_program_state())
 	{
 		rsx::thread::end();
 		return;
 	}
 
+	std::chrono::time_point<steady_clock> state_check_end = steady_clock::now();
+	m_begin_time += (u32)std::chrono::duration_cast<std::chrono::microseconds>(state_check_end - state_check_start).count();
+
 	if (manually_flush_ring_buffers)
 	{
 		//Use approximations to reseve space. This path is mostly for debug purposes anyway
@@ -964,7 +969,7 @@ bool GLGSRender::check_program_state()
 			if (dirty_framebuffer)
 				return std::make_tuple(false, 0);
 
-			auto rsc = m_rtts.get_surface_subresource_if_applicable(texaddr, 0, 0, tex.pitch());
+			auto rsc = m_rtts.get_surface_subresource_if_applicable(texaddr, 0, 0, tex.pitch(), false, false, !is_depth, is_depth);
 			if (!rsc.surface || rsc.is_depth_surface != is_depth)
 				return std::make_tuple(false, 0);
 

rpcs3/Emu/RSX/VK/VKGSRender.cpp

@@ -995,6 +995,8 @@ void VKGSRender::end()
 		return;
 	}
 
+	std::chrono::time_point<steady_clock> state_check_start = steady_clock::now();
+
 	//Load program here since it is dependent on vertex state
 	if (!check_program_status())
 	{
@@ -1003,14 +1005,17 @@ void VKGSRender::end()
 		return;
 	}
 
+	std::chrono::time_point<steady_clock> state_check_end = steady_clock::now();
+	m_setup_time += (u32)std::chrono::duration_cast<std::chrono::microseconds>(state_check_end - state_check_start).count();
+
 	//Programs data is dependent on vertex state
-	std::chrono::time_point<steady_clock> vertex_start = steady_clock::now();
+	std::chrono::time_point<steady_clock> vertex_start = state_check_end;
 	auto upload_info = upload_vertex_data();
 	std::chrono::time_point<steady_clock> vertex_end = steady_clock::now();
 	m_vertex_upload_time += std::chrono::duration_cast<std::chrono::microseconds>(vertex_end - vertex_start).count();
 
 	//Load program
-	std::chrono::time_point<steady_clock> program_start = steady_clock::now();
+	std::chrono::time_point<steady_clock> program_start = vertex_end;
 	load_program(std::get<2>(upload_info), std::get<3>(upload_info));
 	std::chrono::time_point<steady_clock> program_stop = steady_clock::now();
 	m_setup_time += std::chrono::duration_cast<std::chrono::microseconds>(program_stop - program_start).count();
@@ -1841,7 +1846,7 @@ bool VKGSRender::check_program_status()
 			if (dirty_framebuffer)
 				return std::make_tuple(false, 0);
 
-			auto rsc = m_rtts.get_surface_subresource_if_applicable(texaddr, 0, 0, tex.pitch());
+			auto rsc = m_rtts.get_surface_subresource_if_applicable(texaddr, 0, 0, tex.pitch(), false, false, !is_depth, is_depth);
 			if (!rsc.surface || rsc.is_depth_surface != is_depth)
 				return std::make_tuple(false, 0);