rsx: Avoid calling slow functions every draw call

- Use TSC for timing where interval duration matters.
- Use atomic counter for ordering timestamps otherwise.

rpcs3/Emu/RSX/Common/surface_utils.h

@@ -147,6 +147,8 @@ namespace rsx
 		u8  samples_x = 1;
 		u8  samples_y = 1;
 
+		rsx::address_range memory_range;
+
 		std::unique_ptr<typename std::remove_pointer<image_storage_type>::type> resolve_surface;
 		surface_sample_layout sample_layout = surface_sample_layout::null;
 		surface_raster_type raster_type = surface_raster_type::linear;
@@ -348,7 +350,14 @@ namespace rsx
 
 		void queue_tag(u32 address)
 		{
+			ensure(native_pitch);
+			ensure(rsx_pitch);
+
 			base_addr = address;
+
+			const u32 internal_height = get_surface_height<rsx::surface_metrics::samples>();
+			const u32 excess = (rsx_pitch - native_pitch);
+			memory_range = rsx::address_range::start_length(base_addr, internal_height * rsx_pitch - excess);
 		}
 
 		void sync_tag()
@@ -394,6 +403,10 @@ namespace rsx
 				const auto sample_offset = (samples[n].y * rsx_pitch) + samples[n].x;
 				memory_tag_samples[n].first = (sample_offset + base_addr);
 			}
+
+			const u32 internal_height = get_surface_height<rsx::surface_metrics::samples>();
+			const u32 excess = (rsx_pitch - native_pitch);
+			memory_range = rsx::address_range::start_length(base_addr, internal_height * rsx_pitch - excess);
 		}
 
 		void sync_tag()
@@ -629,9 +642,7 @@ namespace rsx
 
 		inline rsx::address_range get_memory_range() const
 		{
-			const u32 internal_height = get_surface_height<rsx::surface_metrics::samples>();
-			const u32 excess = (rsx_pitch - native_pitch);
-			return rsx::address_range::start_length(base_addr, internal_height * rsx_pitch - excess);
+			return memory_range;
 		}
 
 		template <typename T>

rpcs3/Emu/RSX/Common/time.hpp

@@ -0,0 +1,21 @@
+#pragma once
+
+#include <util/asm.hpp>
+#include <util/sysinfo.hpp>
+
+extern u64 get_system_time();
+
+namespace rsx
+{
+	static inline u64 uclock()
+	{
+		if (const u64 freq = (utils::get_tsc_freq() / 1000000))
+		{
+			return utils::get_tsc() / freq;
+		}
+		else
+		{
+			return get_system_time();
+		}
+	}
+}

rpcs3/Emu/RSX/RSXFIFO.cpp

@@ -3,6 +3,7 @@
 #include "RSXFIFO.h"
 #include "RSXThread.h"
 #include "Capture/rsx_capture.h"
+#include "Common/time.hpp"
 #include "Emu/Cell/lv2/sys_rsx.h"
 
 namespace rsx
@@ -395,7 +396,7 @@ namespace rsx
 			{
 				if (performance_counters.state == FIFO_state::running)
 				{
-					performance_counters.FIFO_idle_timestamp = get_system_time();
+					performance_counters.FIFO_idle_timestamp = rsx::uclock();
 					performance_counters.state = FIFO_state::nop;
 				}
 
@@ -405,7 +406,7 @@ namespace rsx
 			{
 				if (performance_counters.state == FIFO_state::running)
 				{
-					performance_counters.FIFO_idle_timestamp = get_system_time();
+					performance_counters.FIFO_idle_timestamp = rsx::uclock();
 					performance_counters.state = FIFO_state::empty;
 				}
 				else
@@ -437,7 +438,7 @@ namespace rsx
 					//Jump to self. Often preceded by NOP
 					if (performance_counters.state == FIFO_state::running)
 					{
-						performance_counters.FIFO_idle_timestamp = get_system_time();
+						performance_counters.FIFO_idle_timestamp = rsx::uclock();
 						sync_point_request.release(true);
 					}
 
@@ -456,7 +457,7 @@ namespace rsx
 					//Jump to self. Often preceded by NOP
 					if (performance_counters.state == FIFO_state::running)
 					{
-						performance_counters.FIFO_idle_timestamp = get_system_time();
+						performance_counters.FIFO_idle_timestamp = rsx::uclock();
 						sync_point_request.release(true);
 					}
 
@@ -513,7 +514,7 @@ namespace rsx
 			}
 
 			// Update performance counters with time spent in idle mode
-			performance_counters.idle_time += (get_system_time() - performance_counters.FIFO_idle_timestamp);
+			performance_counters.idle_time += (rsx::uclock() - performance_counters.FIFO_idle_timestamp);
 		}
 
 		do

rpcs3/Emu/RSX/RSXThread.cpp

@@ -7,6 +7,7 @@
 #include "Common/BufferUtils.h"
 #include "Common/texture_cache.h"
 #include "Common/surface_store.h"
+#include "Common/time.hpp"
 #include "Capture/rsx_capture.h"
 #include "rsx_methods.h"
 #include "gcm_printing.h"
@@ -483,6 +484,7 @@ namespace rsx
 		// This whole thing becomes a mess if we don't have a provoking attribute.
 		const auto vertex_id = vertex_push_buffers[0].get_vertex_id();
 		vertex_push_buffers[attribute].set_vertex_data(attribute, vertex_id, subreg_index, type, size, value);
+		m_graphics_state |= rsx::pipeline_state::push_buffer_arrays_dirty;
 	}
 
 	u32 thread::get_push_buffer_vertex_count() const
@@ -507,7 +509,9 @@ namespace rsx
 	void thread::end()
 	{
 		if (capture_current_frame)
+		{
 			capture::capture_draw_memory(this);
+		}
 
 		in_begin_end = false;
 		m_frame_stats.draw_calls++;
@@ -517,12 +521,17 @@ namespace rsx
 		m_graphics_state |= rsx::pipeline_state::framebuffer_reads_dirty;
 		ROP_sync_timestamp = rsx::get_shared_tag();
 
-		for (auto & push_buf : vertex_push_buffers)
+		if (m_graphics_state & rsx::pipeline_state::push_buffer_arrays_dirty)
 		{
-			//Disabled, see https://github.com/RPCS3/rpcs3/issues/1932
-			//rsx::method_registers.register_vertex_info[index].size = 0;
+			for (auto& push_buf : vertex_push_buffers)
+			{
+				//Disabled, see https://github.com/RPCS3/rpcs3/issues/1932
+				//rsx::method_registers.register_vertex_info[index].size = 0;
 
-			push_buf.clear();
+				push_buf.clear();
+			}
+
+			m_graphics_state &= ~rsx::pipeline_state::push_buffer_arrays_dirty;
 		}
 
 		element_push_buffer.clear();
@@ -630,7 +639,7 @@ namespace rsx
 
 		fifo_ctrl = std::make_unique<::rsx::FIFO::FIFO_control>(this);
 
-		last_flip_time = get_system_time() - 1000000;
+		last_flip_time = rsx::uclock() - 1000000;
 
 		vblank_count = 0;
 
@@ -642,7 +651,7 @@ namespace rsx
 #else
 			constexpr u32 host_min_quantum = 500;
 #endif
-			u64 start_time = get_system_time();
+			u64 start_time = rsx::uclock();
 
 			u64 vblank_rate = g_cfg.video.vblank_rate;
 			u64 vblank_period = 1'000'000 + u64{g_cfg.video.vblank_ntsc.get()} * 1000;
@@ -653,7 +662,7 @@ namespace rsx
 			while (!is_stopped())
 			{
 				// Get current time
-				const u64 current = get_system_time();
+				const u64 current = rsx::uclock();
 
 				// Calculate the time at which we need to send a new VBLANK signal
 				const u64 post_event_time = start_time + (local_vblank_count + 1) * vblank_period / vblank_rate;
@@ -715,7 +724,7 @@ namespace rsx
 				if (Emu.IsPaused())
 				{
 					// Save the difference before pause
-					start_time = get_system_time() - start_time;
+					start_time = rsx::uclock() - start_time;
 
 					while (Emu.IsPaused() && !is_stopped())
 					{
@@ -723,7 +732,7 @@ namespace rsx
 					}
 
 					// Restore difference
-					start_time = get_system_time() - start_time;
+					start_time = rsx::uclock() - start_time;
 				}
 			}
 		});
@@ -2602,7 +2611,7 @@ namespace rsx
 
 	void thread::recover_fifo(u32 line, u32 col, const char* file, const char* func)
 	{
-		const u64 current_time = get_system_time();
+		const u64 current_time = rsx::uclock();
 
 		if (recovered_fifo_cmds_history.size() == 20u)
 		{
@@ -2659,7 +2668,7 @@ namespace rsx
 
 		// Some cases do not need full delay
 		remaining = utils::aligned_div(remaining, div);
-		const u64 until = get_system_time() + remaining;
+		const u64 until = rsx::uclock() + remaining;
 
 		while (true)
 		{
@@ -2691,7 +2700,7 @@ namespace rsx
 				busy_wait(100);
 			}
 
-			const u64 current = get_system_time();
+			const u64 current = rsx::uclock();
 
 			if (current >= until)
 			{
@@ -2922,7 +2931,7 @@ namespace rsx
 		//Average load over around 30 frames
 		if (!performance_counters.last_update_timestamp || performance_counters.sampled_frames > 30)
 		{
-			const auto timestamp = get_system_time();
+			const auto timestamp = rsx::uclock();
 			const auto idle = performance_counters.idle_time.load();
 			const auto elapsed = timestamp - performance_counters.last_update_timestamp;
 
@@ -3086,7 +3095,7 @@ namespace rsx
 
 		if (limit)
 		{
-			const u64 time = get_system_time() - Emu.GetPauseTime();
+			const u64 time = rsx::uclock() - Emu.GetPauseTime();
 			const u64 needed_us = static_cast<u64>(1000000 / limit);
 
 			if (int_flip_index == 0)
@@ -3124,7 +3133,7 @@ namespace rsx
 
 		flip(m_queued_flip);
 
-		last_flip_time = get_system_time() - 1000000;
+		last_flip_time = rsx::uclock() - 1000000;
 		flip_status = CELL_GCM_DISPLAY_FLIP_STATUS_DONE;
 		m_queued_flip.in_progress = false;
 
@@ -3630,7 +3639,7 @@ namespace rsx
 					}
 				}
 
-				if (m_tsc = get_system_time(); m_tsc < m_next_tsc)
+				if (m_tsc = rsx::uclock(); m_tsc < m_next_tsc)
 				{
 					return;
 				}

rpcs3/Emu/RSX/RSXThread.h

@@ -129,6 +129,8 @@ namespace rsx
 		polygon_stipple_pattern_dirty = 0x8000,  // Rasterizer stippling pattern changed
 		line_stipple_pattern_dirty = 0x10000,    // Line stippling pattern changed
 
+		push_buffer_arrays_dirty = 0x20000,   // Push buffers have data written to them (immediate mode vertex buffers)
+
 		fragment_program_dirty = fragment_program_ucode_dirty | fragment_program_state_dirty,
 		vertex_program_dirty = vertex_program_ucode_dirty | vertex_program_state_dirty,
 		invalidate_pipeline_bits = fragment_program_dirty | vertex_program_dirty,

rpcs3/Emu/RSX/VK/VKGSRender.cpp

@@ -1085,7 +1085,7 @@ void VKGSRender::check_heap_status(u32 flags)
 			m_texture_upload_buffer_ring_info.reset_allocation_stats();
 			m_raster_env_ring_info.reset_allocation_stats();
 			m_current_frame->reset_heap_ptrs();
-			m_last_heap_sync_time = get_system_time();
+			m_last_heap_sync_time = rsx::get_shared_tag();
 		}
 		else
 		{

rpcs3/Emu/RSX/VK/VKGSRenderTypes.hpp

@@ -244,7 +244,7 @@ namespace vk
 			texture_upload_heap_ptr = texture_loc;
 			rasterizer_env_heap_ptr = rasterizer_loc;
 
-			last_frame_sync_time = get_system_time();
+			last_frame_sync_time = rsx::get_shared_tag();
 		}
 
 		void reset_heap_ptrs()

rpcs3/Emu/RSX/rsx_methods.cpp

@@ -3,6 +3,7 @@
 #include "RSXThread.h"
 #include "rsx_utils.h"
 #include "rsx_decode.h"
+#include "Common/time.hpp"
 #include "Emu/Cell/PPUCallback.h"
 #include "Emu/Cell/lv2/sys_rsx.h"
 #include "Emu/RSX/Common/BufferUtils.h"
@@ -106,7 +107,7 @@ namespace rsx
 				rsx->flush_fifo();
 			}
 
-			u64 start = get_system_time();
+			u64 start = rsx::uclock();
 			while (sema != arg)
 			{
 				if (rsx->is_stopped())
@@ -118,7 +119,7 @@ namespace rsx
 				{
 					if (rsx->is_paused())
 					{
-						const u64 start0 = get_system_time();
+						const u64 start0 = rsx::uclock();
 
 						while (rsx->is_paused())
 						{
@@ -126,11 +127,11 @@ namespace rsx
 						}
 
 						// Reset
-						start += get_system_time() - start0;
+						start += rsx::uclock() - start0;
 					}
 					else
 					{
-						if ((get_system_time() - start) > tdr)
+						if ((rsx::uclock() - start) > tdr)
 						{
 							// If longer than driver timeout force exit
 							rsx_log.error("nv406e::semaphore_acquire has timed out. semaphore_address=0x%X", addr);
@@ -143,7 +144,7 @@ namespace rsx
 			}
 
 			rsx->fifo_wake_delay();
-			rsx->performance_counters.idle_time += (get_system_time() - start);
+			rsx->performance_counters.idle_time += (rsx::uclock() - start);
 		}
 
 		void semaphore_release(thread* rsx, u32 /*reg*/, u32 arg)

rpcs3/emucore.vcxproj

@@ -479,6 +479,7 @@
     <ClInclude Include="Emu\RSX\Common\bitfield.hpp" />
     <ClInclude Include="Emu\RSX\Common\profiling_timer.hpp" />
     <ClInclude Include="Emu\RSX\Common\simple_array.hpp" />
+    <ClInclude Include="Emu\RSX\Common\time.hpp" />
     <ClInclude Include="Emu\RSX\Overlays\overlay_edit_text.hpp" />
     <ClInclude Include="Emu\RSX\Overlays\overlay_list_view.hpp" />
     <ClInclude Include="Emu\RSX\Overlays\overlay_media_list_dialog.h" />

rpcs3/emucore.vcxproj.filters

@@ -2053,6 +2053,9 @@
     <ClInclude Include="Emu\RSX\Overlays\overlay_media_list_dialog.h">
       <Filter>Emu\GPU\RSX\Overlays</Filter>
     </ClInclude>
+    <ClInclude Include="Emu\RSX\Common\time.hpp">
+      <Filter>Emu\GPU\RSX\Common</Filter>
+    </ClInclude>
   </ItemGroup>
   <ItemGroup>
     <None Include="Emu\RSX\Common\Interpreter\FragmentInterpreter.glsl">