vk: Batch query copy requests to reduce number of vulkan commands used

rpcs3/Emu/RSX/VK/VKGSRender.cpp

@@ -2894,7 +2894,7 @@ void VKGSRender::begin_conditional_rendering(const std::vector<rsx::reports::occ
 			if (!query_info.indices.empty())
 			{
 				const auto& index = query_info.indices.front();
-				m_occlusion_query_manager->get_query_result_indirect(*m_current_command_buffer, index, m_cond_render_buffer->value, 0);
+				m_occlusion_query_manager->get_query_result_indirect(*m_current_command_buffer, index, 1, m_cond_render_buffer->value, 0);
 
 				vk::insert_buffer_memory_barrier(*m_current_command_buffer, m_cond_render_buffer->value, 0, 4,
 					VK_PIPELINE_STAGE_TRANSFER_BIT, dst_stage,
@@ -2912,16 +2912,58 @@ void VKGSRender::begin_conditional_rendering(const std::vector<rsx::reports::occ
 	{
 		// We'll need to do some result aggregation using a compute shader.
 		auto scratch = vk::get_scratch_buffer(*m_current_command_buffer, num_hw_queries * 4);
+
+		// Range latching. Because of how the query pool manages allocations using a stack, we get an inverse sequential set of handles/indices that we can easily group together.
+		// This drastically boosts performance on some drivers like the NVIDIA proprietary one that seems to have a rather high cost for every individual query transer command.
+		struct { u32 first, last; } query_range = { umax, 0 };
+
+		auto copy_query_range_impl = [&]()
+		{
+			const auto count = (query_range.last - query_range.first + 1);
+			m_occlusion_query_manager->get_query_result_indirect(*m_current_command_buffer, query_range.first, count, scratch->value, dst_offset);
+			dst_offset += count * 4;
+		};
+
 		for (usz i = first; i < last; ++i)
 		{
 			auto& query_info = m_occlusion_map[sources[i]->driver_handle];
 			for (const auto& index : query_info.indices)
 			{
-				m_occlusion_query_manager->get_query_result_indirect(*m_current_command_buffer, index, scratch->value, dst_offset);
-				dst_offset += 4;
+				// First iteration?
+				if (query_range.first == umax)
+				{
+					query_range = { index, index };
+					continue;
+				}
+
+				// Head?
+				if ((query_range.first - 1) == index)
+				{
+					query_range.first = index;
+					continue;
+				}
+
+				// Tail?
+				if ((query_range.last + 1) == index)
+				{
+					query_range.last = index;
+					continue;
+				}
+
+				// Flush pending queue. In practice, this is never reached and we fall out to the spill block outside the loops
+				copy_query_range_impl();
+
+				// Start a new range for the current index
+				query_range = { index, index };
 			}
 		}
 
+		if (query_range.first != umax)
+		{
+			// Dangling queries, flush
+			copy_query_range_impl();
+		}
+
 		// Sanity check
 		ensure(dst_offset <= scratch->size());
 

rpcs3/Emu/RSX/VK/VKQueryPool.cpp

@@ -168,11 +168,11 @@ namespace vk
 		return query_info.data;
 	}
 
-	void query_pool_manager::get_query_result_indirect(vk::command_buffer& cmd, u32 index, VkBuffer dst, VkDeviceSize dst_offset)
+	void query_pool_manager::get_query_result_indirect(vk::command_buffer& cmd, u32 index, u32 count, VkBuffer dst, VkDeviceSize dst_offset)
 	{
 		// We're technically supposed to stop any active renderpasses before streaming the results out, but that doesn't matter on IMR hw
 		// On TBDR setups like the apple M series, the stop is required (results are all 0 if you don't flush the RP), but this introduces a very heavy performance loss.
-		vkCmdCopyQueryPoolResults(cmd, *query_slot_status[index].pool, index, 1, dst, dst_offset, 4, VK_QUERY_RESULT_WAIT_BIT);
+		vkCmdCopyQueryPoolResults(cmd, *query_slot_status[index].pool, index, count, dst, dst_offset, 4, VK_QUERY_RESULT_WAIT_BIT);
 	}
 
 	void query_pool_manager::free_query(vk::command_buffer&/*cmd*/, u32 index)

rpcs3/Emu/RSX/VK/VKQueryPool.h

@@ -47,7 +47,7 @@ namespace vk
 
 		bool check_query_status(u32 index);
 		u32  get_query_result(u32 index);
-		void get_query_result_indirect(vk::command_buffer& cmd, u32 index, VkBuffer dst, VkDeviceSize dst_offset);
+		void get_query_result_indirect(vk::command_buffer& cmd, u32 index, u32 count, VkBuffer dst, VkDeviceSize dst_offset);
 
 		u32 allocate_query(vk::command_buffer& cmd);
 		void free_query(vk::command_buffer&/*cmd*/, u32 index);