rsx: Improve object reuse and lifetime management

rpcs3/Emu/RSX/Common/TextureUtils.cpp

@@ -796,3 +796,17 @@ size_t get_texture_size(const rsx::vertex_texture &texture)
 	return get_texture_size(texture.format(), texture.width(), texture.height(), texture.depth(),
 		texture.pitch(), texture.get_exact_mipmap_count(), texture.cubemap() ? 6 : 1);
 }
+
+u32 get_remap_encoding(const std::pair<std::array<u8, 4>, std::array<u8, 4>>& remap)
+{
+	u32 encode = 0;
+	encode |= (remap.first[0] << 0);
+	encode |= (remap.first[1] << 2);
+	encode |= (remap.first[2] << 4);
+	encode |= (remap.first[3] << 6);
+	encode |= (remap.second[0] << 8);
+	encode |= (remap.second[1] << 10);
+	encode |= (remap.second[2] << 12);
+	encode |= (remap.second[3] << 14);
+	return encode;
+}

rpcs3/Emu/RSX/Common/TextureUtils.h

@@ -139,3 +139,8 @@ size_t get_texture_size(const rsx::vertex_texture &texture);
 * Get packed pitch
 */
 u32 get_format_packed_pitch(u32 format, u16 width);
+
+/**
+* Reverse encoding
+*/
+u32 get_remap_encoding(const std::pair<std::array<u8, 4>, std::array<u8, 4>>& remap);

rpcs3/Emu/RSX/Common/texture_cache_utils.h

@@ -1,4 +1,4 @@
-#pragma once
+#pragma once
 
 #include "../rsx_cache.h"
 #include "texture_cache_predictor.h"
@@ -1114,6 +1114,9 @@ namespace rsx
 			invalidate_range();
 		}
 
+		virtual void dma_abort()
+		{}
+
 	public:
 		/**
 		 * Dirty/Unreleased Flag
@@ -1276,6 +1279,12 @@ namespace rsx
 
 		void reprotect(const utils::protection prot)
 		{
+			if (synchronized && !flushed)
+			{
+				// Abort enqueued transfer
+				dma_abort();
+			}
+
 			//Reset properties and protect again
 			flushed = false;
 			synchronized = false;
@@ -1286,6 +1295,12 @@ namespace rsx
 
 		void reprotect(const utils::protection prot, const std::pair<u32, u32>& range)
 		{
+			if (synchronized && !flushed)
+			{
+				// Abort enqueued transfer
+				dma_abort();
+			}
+
 			//Reset properties and protect again
 			flushed = false;
 			synchronized = false;

rpcs3/Emu/RSX/VK/VKGSRender.cpp

@@ -8,6 +8,7 @@
 #include "VKFormats.h"
 #include "VKCommonDecompiler.h"
 #include "VKRenderPass.h"
+#include "VKResourceManager.h"
 
 namespace
 {
@@ -640,7 +641,6 @@ VKGSRender::~VKGSRender()
 	m_rtts.destroy();
 	m_texture_cache.destroy();
 
-	m_resource_manager.destroy();
 	m_stencil_mirror_sampler.reset();
 
 	//Overlay text handler
@@ -1372,7 +1372,7 @@ void VKGSRender::end()
 
 					if (replace)
 					{
-						fs_sampler_handles[i] = m_resource_manager.find_sampler(*m_device, wrap_s, wrap_t, wrap_r, false, lod_bias, af_level, min_lod, max_lod,
+						fs_sampler_handles[i] = vk::get_resource_manager()->find_sampler(*m_device, wrap_s, wrap_t, wrap_r, false, lod_bias, af_level, min_lod, max_lod,
 							min_filter, mag_filter, mip_mode, border_color, compare_enabled, depth_compare_mode);
 					}
 				}
@@ -1422,7 +1422,7 @@ void VKGSRender::end()
 
 					if (replace)
 					{
-						vs_sampler_handles[i] = m_resource_manager.find_sampler(
+						vs_sampler_handles[i] = vk::get_resource_manager()->find_sampler(
 							*m_device,
 							VK_SAMPLER_ADDRESS_MODE_REPEAT, VK_SAMPLER_ADDRESS_MODE_REPEAT, VK_SAMPLER_ADDRESS_MODE_REPEAT,
 							unnormalized_coords,
@@ -2159,6 +2159,11 @@ void VKGSRender::flush_command_queue(bool hard_sync)
 		m_current_cb_index = (m_current_cb_index + 1) % VK_MAX_ASYNC_CB_COUNT;
 		m_current_command_buffer = &m_primary_cb_list[m_current_cb_index];
 
+		if (!m_current_command_buffer->poke())
+		{
+			LOG_ERROR(RSX, "CB chain has run out of free entries!");
+		}
+
 		m_current_command_buffer->reset();
 
 		// Just in case a queued frame holds a ref to this cb, drain the present queue
@@ -2823,6 +2828,7 @@ void VKGSRender::close_and_submit_command_buffer(VkFence fence, VkSemaphore wait
 	{
 		auto open_query = m_occlusion_map[m_active_query_info->driver_handle].indices.back();
 		m_occlusion_query_pool.end_query(*m_current_command_buffer, open_query);
+		m_current_command_buffer->flags &= ~vk::command_buffer::cb_has_open_query;
 	}
 
 	m_current_command_buffer->end();
@@ -3601,6 +3607,18 @@ void VKGSRender::discard_occlusion_query(rsx::reports::occlusion_query_info* que
 	data.indices.clear();
 }
 
+void VKGSRender::emergency_query_cleanup(vk::command_buffer* commands)
+{
+	verify("Command list mismatch" HERE), commands == static_cast<vk::command_buffer*>(m_current_command_buffer);
+
+	if (m_current_command_buffer->flags & vk::command_buffer::cb_has_open_query)
+	{
+		auto open_query = m_occlusion_map[m_active_query_info->driver_handle].indices.back();
+		m_occlusion_query_pool.end_query(*m_current_command_buffer, open_query);
+		m_current_command_buffer->flags &= ~vk::command_buffer::cb_has_open_query;
+	}
+}
+
 bool VKGSRender::on_decompiler_task()
 {
 	return m_prog_buffer->async_update(8, *m_device, pipeline_layout).first;

rpcs3/Emu/RSX/VK/VKGSRender.h

@@ -72,7 +72,7 @@ struct command_buffer_chunk: public vk::command_buffer
 	VkDevice m_device = VK_NULL_HANDLE;
 
 	std::atomic_bool pending = { false };
-	std::atomic<u64> last_sync = { 0 };
+	u64 eid_tag = 0;
 	shared_mutex guard_mutex;
 
 	command_buffer_chunk() = default;
@@ -96,7 +96,7 @@ struct command_buffer_chunk: public vk::command_buffer
 
 	void tag()
 	{
-		last_sync = get_system_time();
+		eid_tag = vk::get_event_id();
 	}
 
 	void reset()
@@ -123,8 +123,11 @@ struct command_buffer_chunk: public vk::command_buffer
 
 			if (pending)
 			{
-				pending = false;
 				vk::reset_fence(&submit_fence);
+				vk::on_event_completed(eid_tag);
+
+				pending = false;
+				eid_tag = 0;
 			}
 		}
 
@@ -145,7 +148,10 @@ struct command_buffer_chunk: public vk::command_buffer
 		if (pending)
 		{
 			vk::reset_fence(&submit_fence);
+			vk::on_event_completed(eid_tag);
+
 			pending = false;
+			eid_tag = 0;
 		}
 
 		return ret;
@@ -288,66 +294,6 @@ struct flush_request_task
 	}
 };
 
-// TODO: This class will be expanded into a global allocator/collector eventually
-class resource_manager
-{
-private:
-	std::unordered_multimap<u64, std::unique_ptr<vk::sampler>> m_sampler_pool;
-
-	bool value_compare(const f32& a, const f32& b)
-	{
-		return fabsf(a - b) < 0.0000001f;
-	}
-
-public:
-
-	resource_manager() = default;
-	~resource_manager() = default;
-
-	void destroy()
-	{
-		m_sampler_pool.clear();
-	}
-
-	vk::sampler* find_sampler(VkDevice dev, VkSamplerAddressMode clamp_u, VkSamplerAddressMode clamp_v, VkSamplerAddressMode clamp_w,
-		VkBool32 unnormalized_coordinates, float mipLodBias, float max_anisotropy, float min_lod, float max_lod,
-		VkFilter min_filter, VkFilter mag_filter, VkSamplerMipmapMode mipmap_mode, VkBorderColor border_color,
-		VkBool32 depth_compare = VK_FALSE, VkCompareOp depth_compare_mode = VK_COMPARE_OP_NEVER)
-	{
-		u64 key = u16(clamp_u) | u64(clamp_v) << 3 | u64(clamp_w) << 6;
-		key |= u64(unnormalized_coordinates) << 9; // 1 bit
-		key |= u64(min_filter) << 10 | u64(mag_filter) << 11; // 1 bit each
-		key |= u64(mipmap_mode) << 12; // 1 bit
-		key |= u64(border_color) << 13; // 3 bits
-		key |= u64(depth_compare) << 16; // 1 bit
-		key |= u64(depth_compare_mode) << 17; // 3 bits
-
-		const auto found = m_sampler_pool.equal_range(key);
-		for (auto It = found.first; It != found.second; ++It)
-		{
-			const auto& info = It->second->info;
-			if (!value_compare(info.mipLodBias, mipLodBias) ||
-				!value_compare(info.maxAnisotropy, max_anisotropy) ||
-				!value_compare(info.minLod, min_lod) ||
-				!value_compare(info.maxLod, max_lod))
-			{
-				continue;
-			}
-
-			return It->second.get();
-		}
-
-		auto result = std::make_unique<vk::sampler>(
-			dev, clamp_u, clamp_v, clamp_w, unnormalized_coordinates,
-			mipLodBias, max_anisotropy, min_lod, max_lod,
-			min_filter, mag_filter, mipmap_mode, border_color,
-			depth_compare, depth_compare_mode);
-
-		auto It = m_sampler_pool.emplace(key, std::move(result));
-		return It->second.get();
-	}
-};
-
 class VKGSRender : public GSRender, public ::rsx::reports::ZCULL_control
 {
 private:
@@ -379,8 +325,6 @@ class VKGSRender : public GSRender, public ::rsx::reports::ZCULL_control
 	std::unique_ptr<vk::buffer_view> m_volatile_attribute_storage;
 	std::unique_ptr<vk::buffer_view> m_vertex_layout_storage;
 
-	resource_manager m_resource_manager;
-
 public:
 	//vk::fbo draw_fbo;
 	std::unique_ptr<vk::vertex_cache> m_vertex_cache;
@@ -459,8 +403,6 @@ class VKGSRender : public GSRender, public ::rsx::reports::ZCULL_control
 	shared_mutex m_flush_queue_mutex;
 	flush_request_task m_flush_requests;
 
-	std::atomic<u64> m_last_sync_event = { 0 };
-
 	bool m_render_pass_open = false;
 	u64  m_current_renderpass_key = 0;
 	VkRenderPass m_cached_renderpass = VK_NULL_HANDLE;
@@ -529,6 +471,9 @@ class VKGSRender : public GSRender, public ::rsx::reports::ZCULL_control
 	void get_occlusion_query_result(rsx::reports::occlusion_query_info* query) override;
 	void discard_occlusion_query(rsx::reports::occlusion_query_info* query) override;
 
+	// External callback in case we need to suddenly submit a commandlist unexpectedly, e.g in a violation handler
+	void emergency_query_cleanup(vk::command_buffer* commands);
+
 protected:
 	void begin() override;
 	void end() override;

rpcs3/Emu/RSX/VK/VKHelpers.cpp

@@ -1,9 +1,11 @@
 #include "stdafx.h"
 #include "VKHelpers.h"
+#include "VKGSRender.h"
 #include "VKCompute.h"
 #include "VKRenderPass.h"
 #include "VKFramebuffer.h"
 #include "VKResolveHelper.h"
+#include "VKResourceManager.h"
 #include "Utilities/mutex.h"
 
 namespace vk
@@ -247,6 +249,7 @@ namespace vk
 		vk::clear_renderpass_cache(dev);
 		vk::clear_framebuffer_cache();
 		vk::clear_resolve_helpers();
+		vk::get_resource_manager()->destroy();
 
 		g_null_texture.reset();
 		g_null_image_view.reset();
@@ -769,6 +772,14 @@ namespace vk
 		}
 	}
 
+	void do_query_cleanup(vk::command_buffer& cmd)
+	{
+		auto renderer = dynamic_cast<VKGSRender*>(rsx::get_current_renderer());
+		verify(HERE), renderer;
+
+		renderer->emergency_query_cleanup(&cmd);
+	}
+
 	void die_with_error(const char* faulting_addr, VkResult error_code)
 	{
 		std::string error_message;

rpcs3/Emu/RSX/VK/VKHelpers.h

@@ -190,6 +190,10 @@ namespace vk
 	VkResult wait_for_fence(VkFence pFence, u64 timeout = 0ull);
 	VkResult wait_for_event(VkEvent pEvent, u64 timeout = 0ull);
 
+	// Handle unexpected submit with dangling occlusion query
+	// TODO: Move queries out of the renderer!
+	void do_query_cleanup(vk::command_buffer& cmd);
+
 	void die_with_error(const char* faulting_addr, VkResult error_code);
 
 	struct memory_type_mapping
@@ -1008,6 +1012,9 @@ namespace vk
 				return;
 			}
 
+			// Check for hanging queries to avoid driver hang
+			verify("close and submit of commandbuffer with a hanging query!" HERE), (flags & cb_has_open_query) == 0;
+
 			if (!fence)
 			{
 				fence = m_submit_fence;
@@ -1291,11 +1298,19 @@ namespace vk
 				}
 			}
 
-			VkComponentMapping real_mapping = vk::apply_swizzle_remap
-			(
-				{native_component_map.a, native_component_map.r, native_component_map.g, native_component_map.b },
-				remap
-			);
+			VkComponentMapping real_mapping;
+			if (remap_encoding == 0xAAE4)
+			{
+				real_mapping = native_component_map;
+			}
+			else
+			{
+				real_mapping = vk::apply_swizzle_remap
+				(
+					{ native_component_map.a, native_component_map.r, native_component_map.g, native_component_map.b },
+					remap
+				);
+			}
 
 			const auto range = vk::get_image_subresource_range(0, 0, info.arrayLayers, info.mipLevels, aspect() & mask);
 

rpcs3/Emu/RSX/VK/VKOverlays.h

@@ -138,7 +138,7 @@ namespace vk
 			return fs_inputs;
 		}
 
-		virtual void get_dynamic_state_entries(VkDynamicState* state_descriptors, VkPipelineDynamicStateCreateInfo& info)
+		virtual void get_dynamic_state_entries(VkDynamicState* /*state_descriptors*/, VkPipelineDynamicStateCreateInfo& /*info*/)
 		{}
 
 		virtual std::vector<VkPushConstantRange> get_push_constants()

rpcs3/Emu/RSX/VK/VKResourceManager.cpp

@@ -0,0 +1,29 @@
+#include "stdafx.h"
+#include "VKResourceManager.h"
+
+namespace vk
+{
+	resource_manager g_resource_manager;
+	atomic_t<u64> g_event_ctr;
+
+	resource_manager* get_resource_manager()
+	{
+		return &g_resource_manager;
+	}
+
+	u64 get_event_id()
+	{
+		return g_event_ctr++;
+	}
+
+	u64 current_event_id()
+	{
+		return g_event_ctr.load();
+	}
+
+	void on_event_completed(u64 event_id)
+	{
+		// TODO: Offload this to a secondary thread
+		g_resource_manager.eid_completed(event_id);
+	}
+}