zcull synchronization tweaks

- Implement forced reading when calling update method to sync partial lists
- Defer conditional render evaluation and use a read barrier to avoid extra work
- Fix HLE gcm library when binding tiles & zcull RAM

rpcs3/Emu/Cell/Modules/cellGcmSys.cpp

@@ -698,6 +698,7 @@ void cellGcmSetZcull(u8 index, u32 offset, u32 width, u32 height, u32 cullStart,
 	zcull.sFunc = sFunc;
 	zcull.sRef = sRef;
 	zcull.sMask = sMask;
+	zcull.binded = (zCullFormat > 0);
 
 	vm::_ptr<CellGcmZcullInfo>(m_config->zculls_addr)[index] = zcull.pack();
 }
@@ -1261,6 +1262,7 @@ s32 cellGcmSetTile(u8 index, u8 location, u32 offset, u32 size, u32 pitch, u8 co
 	tile.comp = comp;
 	tile.base = base;
 	tile.bank = bank;
+	tile.binded = (pitch > 0);
 
 	vm::_ptr<CellGcmTileInfo>(m_config->tiles_addr)[index] = tile.pack();
 	return CELL_OK;

rpcs3/Emu/RSX/RSXThread.cpp

@@ -275,6 +275,15 @@ namespace rsx
 
 	void thread::begin()
 	{
+		if (conditional_render_enabled && conditional_render_test_address)
+		{
+			// Evaluate conditional rendering test
+			zcull_ctrl->read_barrier(this, conditional_render_test_address, 4);
+			vm::ptr<CellGcmReportData> result = vm::cast(conditional_render_test_address);
+			conditional_render_test_failed = (result->value == 0);
+			conditional_render_test_address = 0;
+		}
+
 		rsx::method_registers.current_draw_clause.inline_vertex_array.resize(0);
 		in_begin_end = true;
 
@@ -2719,7 +2728,8 @@ namespace rsx
 					{
 						verify(HERE), query->pending;
 
-						if (!result && query->num_draws)
+						const bool implemented = (writer.type == CELL_GCM_ZPASS_PIXEL_CNT || writer.type == CELL_GCM_ZCULL_STATS3);
+						if (implemented && !result && query->num_draws)
 						{
 							get_occlusion_query_result(query);
 
@@ -2784,13 +2794,23 @@ namespace rsx
 			m_cycles_delay = min_zcull_cycles_delay;
 		}
 
-		void ZCULL_control::update(::rsx::thread* ptimer)
+		void ZCULL_control::update(::rsx::thread* ptimer, u32 sync_address)
 		{
 			m_tsc++;
 
 			if (m_pending_writes.empty())
 				return;
 
+			if (!sync_address)
+			{
+				const auto& front = m_pending_writes.front();
+				if (!front.sink || m_tsc < front.due_tsc)
+				{
+					// Avoid spamming backend with report status updates
+					return;
+				}
+			}
+
 			u32 stat_tag_to_remove = m_statistics_tag_id;
 			u32 processed = 0;
 			for (auto &writer : m_pending_writes)
@@ -2810,13 +2830,21 @@ namespace rsx
 				auto query = writer.query;
 				u32 result = m_statistics_map[writer.counter_tag];
 
+				const bool force_read = (sync_address != 0);
+				if (force_read && writer.sink == sync_address)
+				{
+					// Forced reads end here
+					sync_address = 0;
+				}
+
 				if (query)
 				{
 					verify(HERE), query->pending;
 
-					if (UNLIKELY(writer.due_tsc < m_tsc))
+					const bool implemented = (writer.type == CELL_GCM_ZPASS_PIXEL_CNT || writer.type == CELL_GCM_ZCULL_STATS3);
+					if (force_read || writer.due_tsc < m_tsc)
 					{
-						if (!result && query->num_draws)
+						if (implemented && !result && query->num_draws)
 						{
 							get_occlusion_query_result(query);
 
@@ -2834,12 +2862,7 @@ namespace rsx
 					}
 					else
 					{
-						if (result || !query->num_draws)
+						if (implemented && !result && query->num_draws)
-						{
-							//Not necessary to read the result anymore
-							discard_occlusion_query(query);
-						}
-						else
 						{
 							//Maybe we get lucky and results are ready
 							if (check_occlusion_query_status(query))
@@ -2857,6 +2880,11 @@ namespace rsx
 								break;
 							}
 						}
+						else
+						{
+							//Not necessary to read the result anymore
+							discard_occlusion_query(query);
+						}
 					}
 
 					query->pending = false;
@@ -2903,14 +2931,20 @@ namespace rsx
 				return;
 
 			const auto memory_end = memory_address + memory_range;
+			u32 sync_address = 0;
+
 			for (const auto &writer : m_pending_writes)
 			{
 				if (writer.sink >= memory_address && writer.sink < memory_end)
 				{
-					sync(ptimer);
+					sync_address = writer.sink;
-					return;
 				}
 			}
+
+			if (sync_address)
+			{
+				update(ptimer, sync_address);
+			}
 		}
 	}
 }

rpcs3/Emu/RSX/RSXThread.h

@@ -118,6 +118,11 @@ namespace rsx
 		lock_wait = 4 // Puller is processing a lock acquire
 	};
 
+	enum FIFO_hint : u8
+	{
+		hint_conditional_render_eval = 1
+	};
+
 	u32 get_vertex_type_size_on_host(vertex_base_type type, u32 size);
 
 	u32 get_address(u32 offset, u32 location);
@@ -230,7 +235,7 @@ namespace rsx
 			queued_report_write* forwarder;
 			vm::addr_t sink;
 
-			u32 due_tsc;
+			u64 due_tsc;
 		};
 
 		struct ZCULL_control
@@ -249,7 +254,7 @@ namespace rsx
 
 			occlusion_query_info* m_current_task = nullptr;
 			u32 m_statistics_tag_id = 0;
-			u32 m_tsc = 0;
+			u64 m_tsc = 0;
 			u32 m_cycles_delay = max_zcull_cycles_delay;
 
 			std::vector<queued_report_write> m_pending_writes;
@@ -278,8 +283,8 @@ namespace rsx
 			// Conditionally sync any pending writes if range overlaps
 			void read_barrier(class ::rsx::thread* ptimer, u32 memory_address, u32 memory_range);
 
-			// Call once every 'tick' to update
+			// Call once every 'tick' to update, optional address provided to partially sync until address is processed
-			void update(class ::rsx::thread* ptimer);
+			void update(class ::rsx::thread* ptimer, u32 sync_address = 0);
 
 			// Draw call notification
 			void on_draw();
@@ -433,6 +438,7 @@ namespace rsx
 
 		atomic_t<s32> async_tasks_pending{ 0 };
 
+		u32  conditional_render_test_address = 0;
 		bool conditional_render_test_failed = false;
 		bool conditional_render_enabled = false;
 		bool zcull_stats_enabled = false;
@@ -482,6 +488,7 @@ namespace rsx
 		// sync
 		void sync();
 		void read_barrier(u32 memory_address, u32 memory_range);
+		virtual void sync_hint(FIFO_hint hint) {}
 		
 		gsl::span<const gsl::byte> get_raw_index_array(const std::vector<std::pair<u32, u32> >& draw_indexed_clause) const;
 		gsl::span<const gsl::byte> get_raw_vertex_buffer(const rsx::data_array_format_info&, u32 base_offset, const std::vector<std::pair<u32, u32>>& vertex_ranges) const;

rpcs3/Emu/RSX/VK/VKGSRender.cpp

@@ -816,20 +816,30 @@ bool VKGSRender::on_access_violation(u32 address, bool is_writing)
 
 		if (sync_timestamp > 0)
 		{
-			//Wait for earliest cb submitted after the sync timestamp to finish
+			// Wait for earliest cb submitted after the sync timestamp to finish
 			command_buffer_chunk *target_cb = nullptr;
 			for (auto &cb : m_primary_cb_list)
 			{
-				if (cb.pending && cb.last_sync >= sync_timestamp)
+				if (cb.last_sync >= sync_timestamp)
 				{
+					if (!cb.pending)
+					{
+						target_cb = nullptr;
+						break;
+					}
+
 					if (target_cb == nullptr || target_cb->last_sync > cb.last_sync)
+					{
 						target_cb = &cb;
 					}
 				}
+			}
 
 			if (target_cb)
+			{
 				target_cb->wait();
 			}
+		}
 
 		if (has_queue_ref)
 		{
@@ -1435,6 +1445,8 @@ void VKGSRender::end()
 		m_occlusion_query_pool.begin_query(*m_current_command_buffer, occlusion_id);
 		m_occlusion_map[m_active_query_info->driver_handle].indices.push_back(occlusion_id);
 		m_occlusion_map[m_active_query_info->driver_handle].command_buffer_to_wait = m_current_command_buffer;
+
+		m_current_command_buffer->flags |= cb_has_occlusion_task;
 	}
 
 	if (!upload_info.index_info)
@@ -1486,6 +1498,7 @@ void VKGSRender::end()
 	close_render_pass();
 	vk::leave_uninterruptible();
 
+	m_current_command_buffer->num_draws++;
 	m_rtts.on_write();
 
 	std::chrono::time_point<steady_clock> draw_end = steady_clock::now();
@@ -1834,6 +1847,22 @@ void VKGSRender::flush_command_queue(bool hard_sync)
 	open_command_buffer();
 }
 
+void VKGSRender::sync_hint(rsx::FIFO_hint hint)
+{
+	if (hint == rsx::FIFO_hint::hint_conditional_render_eval)
+	{
+		if (m_current_command_buffer->flags & cb_has_occlusion_task)
+		{
+			// Occlusion test result evaluation is coming up, avoid a hard sync
+			if (!m_flush_requests.pending())
+			{
+				m_flush_requests.post(false);
+				m_flush_requests.remove_one();
+			}
+		}
+	}
+}
+
 void VKGSRender::advance_queued_frames()
 {
 	//Check all other frames for completion and clear resources
@@ -3289,10 +3318,13 @@ void VKGSRender::end_occlusion_query(rsx::reports::occlusion_query_info* query)
 
 	//Avoid stalling later if this query is already tied to a report
 	if (query->num_draws && query->owned && !m_flush_requests.pending())
+	{
+		if (0)//m_current_command_buffer->flags & cb_has_occlusion_task)
 		{
 			m_flush_requests.post(false);
 			m_flush_requests.remove_one();
 		}
+	}
 }
 
 bool VKGSRender::check_occlusion_query_status(rsx::reports::occlusion_query_info* query)

rpcs3/Emu/RSX/VK/VKGSRender.h

@@ -47,11 +47,19 @@ namespace vk
 
 extern u64 get_system_time();
 
+enum command_buffer_data_flag
+{
+	cb_has_occlusion_task = 1
+};
+
 struct command_buffer_chunk: public vk::command_buffer
 {
 	VkFence submit_fence = VK_NULL_HANDLE;
 	VkDevice m_device = VK_NULL_HANDLE;
 
+	u32 num_draws = 0;
+	u32 flags = 0;
+
 	std::atomic_bool pending = { false };
 	std::atomic<u64> last_sync = { 0 };
 	shared_mutex guard_mutex;
@@ -90,12 +98,17 @@ struct command_buffer_chunk: public vk::command_buffer
 			wait();
 
 		CHECK_RESULT(vkResetCommandBuffer(commands, 0));
+		num_draws = 0;
+		flags = 0;
 	}
 
 	bool poke()
 	{
 		reader_lock lock(guard_mutex);
 
+		if (!pending)
+			return true;
+
 		if (vkGetFenceStatus(m_device, submit_fence) == VK_SUCCESS)
 		{
 			lock.upgrade();
@@ -117,14 +130,8 @@ struct command_buffer_chunk: public vk::command_buffer
 		if (!pending)
 			return;
 
-		switch(vkGetFenceStatus(m_device, submit_fence))
+		// NOTE: vkWaitForFences is slower than polling fence status at least on NV
-		{
+		while (vkGetFenceStatus(m_device, submit_fence) == VK_NOT_READY);
-		case VK_SUCCESS:
-			break;
-		case VK_NOT_READY:
-			CHECK_RESULT(vkWaitForFences(m_device, 1, &submit_fence, VK_TRUE, UINT64_MAX));
-			break;
-		}
 
 		lock.upgrade();
 
@@ -406,6 +413,8 @@ public:
 	void write_buffers();
 	void set_viewport();
 
+	void sync_hint(rsx::FIFO_hint hint) override;
+
 	void begin_occlusion_query(rsx::reports::occlusion_query_info* query) override;
 	void end_occlusion_query(rsx::reports::occlusion_query_info* query) override;
 	bool check_occlusion_query_status(rsx::reports::occlusion_query_info* query) override;

rpcs3/Emu/RSX/rsx_methods.cpp

@@ -521,9 +521,10 @@ namespace rsx
 				return;
 			}
 
-			rsx->sync();
+			// Defer conditional render evaluation
-			vm::ptr<CellGcmReportData> result = address_ptr;
+			rsx->sync_hint(FIFO_hint::hint_conditional_render_eval);
-			rsx->conditional_render_test_failed = (result->value == 0);
+			rsx->conditional_render_test_address = address_ptr;
+			rsx->conditional_render_test_failed = false;
 		}
 
 		void set_zcull_render_enable(thread* rsx, u32, u32 arg)
@@ -1809,8 +1810,6 @@ namespace rsx
 		bind<NV4097_SET_DEPTH_MASK, nv4097::set_surface_options_dirty_bit>();
 		bind<NV4097_SET_COLOR_MASK, nv4097::set_surface_options_dirty_bit>();
 		bind<NV4097_WAIT_FOR_IDLE, nv4097::sync>();
-		bind<NV4097_ZCULL_SYNC, nv4097::sync>();
-		bind<NV4097_SET_CONTEXT_DMA_REPORT, nv4097::sync>();
 		bind<NV4097_INVALIDATE_L2, nv4097::set_shader_program_dirty>();
 		bind<NV4097_SET_SHADER_PROGRAM, nv4097::set_shader_program_dirty>();
 		bind<NV4097_SET_TRANSFORM_PROGRAM_START, nv4097::set_transform_program_start>();