vk/gl: Synchronization improvements

- Properly wait for the buffer transfer operation to finish before map/readback!
- Change vkFence to vkEvent which works more like a GL fence which is what is needed.
- Implement supporting methods and functions
- Do not destroy fence by immediately waiting after copying to dma buffer

rpcs3/Emu/RSX/Common/texture_cache.h

@@ -2670,6 +2670,7 @@ namespace rsx
 			else
 			{
 				verify(HERE), dst_is_render_target;
+				dst_subres.surface->on_write();
 			}
 
 			if (rsx::get_resolution_scale_percent() != 100)

rpcs3/Emu/RSX/GL/GLTextureCache.h

@@ -61,7 +61,6 @@ namespace gl
 
 		texture::format format = texture::format::rgba;
 		texture::type type = texture::type::ubyte;
-		rsx::surface_antialiasing aa_mode = rsx::surface_antialiasing::center_1_sample;
 
 		u8 get_pixel_size(texture::format fmt_, texture::type type_)
 		{
@@ -157,7 +156,7 @@ namespace gl
 		using baseclass::cached_texture_section;
 
 		void create(u16 w, u16 h, u16 depth, u16 mipmaps, gl::texture* image, u32 rsx_pitch, bool read_only,
-				gl::texture::format gl_format, gl::texture::type gl_type, bool swap_bytes)
+				gl::texture::format gl_format = gl::texture::format::rgba, gl::texture::type gl_type = gl::texture::type::ubyte, bool swap_bytes = false)
 		{
 			auto new_texture = static_cast<gl::viewable_image*>(image);
 			ASSERT(!exists() || !is_managed() || vram_texture == new_texture);
@@ -166,11 +165,9 @@ namespace gl
 			if (read_only)
 			{
 				managed_texture.reset(vram_texture);
-				aa_mode = rsx::surface_antialiasing::center_1_sample;
 			}
 			else
 			{
-				aa_mode = static_cast<gl::render_target*>(image)->read_aa_mode;
 				ASSERT(managed_texture.get() == nullptr);
 			}
 
@@ -193,28 +190,6 @@ namespace gl
 			baseclass::on_section_resources_created();
 		}
 
-		void create_read_only(gl::viewable_image* image, u32 width, u32 height, u32 depth, u32 mipmaps, u16 pitch)
-		{
-			ASSERT(!exists() || !is_managed() || vram_texture == image);
-
-			verify(HERE), pitch;
-
-			//Only to be used for ro memory, we dont care about most members, just dimensions and the vram texture handle
-			this->width = width;
-			this->height = height;
-			this->depth = depth;
-			this->mipmaps = mipmaps;
-
-			managed_texture.reset(image);
-			vram_texture = image;
-
-			rsx_pitch = pitch;
-			real_pitch = 0;
-
-			// Notify baseclass
-			baseclass::on_section_resources_created();
-		}
-
 		void set_dimensions(u32 width, u32 height, u32 /*depth*/, u32 pitch)
 		{
 			this->width = width;
@@ -264,7 +239,9 @@ namespace gl
 				u32 real_width = width;
 				u32 real_height = height;
 
-				switch (aa_mode)
+				if (context == rsx::texture_upload_context::framebuffer_storage)
+				{
+					switch (static_cast<gl::render_target*>(vram_texture)->read_aa_mode)
 					{
 					case rsx::surface_antialiasing::center_1_sample:
 						break;
@@ -276,6 +253,7 @@ namespace gl
 						real_height *= 2;
 						break;
 					}
+				}
 
 				areai src_area = { 0, 0, 0, 0 };
 				const areai dst_area = { 0, 0, (s32)real_width, (s32)real_height };
@@ -376,16 +354,13 @@ namespace gl
 
 			verify(HERE), cmd.drv;
 			copy_texture(cmd, blocking);
-
-			if (blocking)
-			{
-				m_fence.wait_for_signal();
-			}
 		}
 
 		void* map_synchronized(u32 offset, u32 size)
 		{
-			AUDIT(synchronized);
+			AUDIT(synchronized && !m_fence.is_empty());
+
+			m_fence.wait_for_signal();
 
 			verify(HERE), (offset + size) <= pbo_size;
 			glBindBuffer(GL_PIXEL_PACK_BUFFER, pbo_id);
@@ -894,7 +869,7 @@ namespace gl
 			cached.set_image_type(type);
 			cached.set_gcm_format(gcm_format);
 
-			cached.create_read_only(image, width, height, depth, mipmaps, pitch);
+			cached.create(width, height, depth, mipmaps, image, pitch, true);
 			cached.set_dirty(false);
 
 			if (context != rsx::texture_upload_context::blit_engine_dst)

rpcs3/Emu/RSX/VK/VKGSRender.cpp

@@ -871,10 +871,6 @@ bool VKGSRender::on_access_violation(u32 address, bool is_writing)
 		const bool is_rsxthr = std::this_thread::get_id() == m_rsx_thread;
 		bool has_queue_ref = false;
 
-		u64 sync_timestamp = 0ull;
-		for (const auto& tex : result.sections_to_flush)
-			sync_timestamp = std::max(sync_timestamp, tex->get_sync_timestamp());
-
 		if (!is_rsxthr)
 		{
 			//Always submit primary cb to ensure state consistency (flush pending changes such as image transitions)
@@ -882,7 +878,7 @@ bool VKGSRender::on_access_violation(u32 address, bool is_writing)
 
 			std::lock_guard lock(m_flush_queue_mutex);
 
-			m_flush_requests.post(sync_timestamp == 0ull);
+			m_flush_requests.post(false);
 			has_queue_ref = true;
 		}
 		else if (!vk::is_uninterruptible())
@@ -895,33 +891,6 @@ bool VKGSRender::on_access_violation(u32 address, bool is_writing)
 			//LOG_ERROR(RSX, "Fault in uninterruptible code!");
 		}
 
-		if (sync_timestamp > 0)
-		{
-			// 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.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(GENERAL_WAIT_TIMEOUT);
-			}
-		}
-
 		if (has_queue_ref)
 		{
 			//Wait for the RSX thread to process request if it hasn't already
@@ -3520,9 +3489,18 @@ bool VKGSRender::scaled_image_from_memory(rsx::blit_src_info& src, rsx::blit_dst
 	//Verify enough memory exists before attempting to handle data transfer
 	check_heap_status();
 
+	const auto old_speculations_count = m_texture_cache.get_num_cache_speculative_writes();
 	if (m_texture_cache.blit(src, dst, interpolate, m_rtts, *m_current_command_buffer))
 	{
 		m_samplers_dirty.store(true);
+		m_current_command_buffer->flags |= cb_has_blit_transfer;
+
+		if (m_texture_cache.get_num_cache_speculative_writes() > old_speculations_count)
+		{
+			// A speculative write happened, flush while the dma resource is valid
+			// TODO: Deeper investigation as to why this can trigger problems
+			flush_command_queue();
+		}
 		return true;
 	}
 

rpcs3/Emu/RSX/VK/VKGSRender.h

@@ -50,7 +50,8 @@ extern u64 get_system_time();
 
 enum command_buffer_data_flag
 {
-	cb_has_occlusion_task = 1
+	cb_has_occlusion_task = 1,
+	cb_has_blit_transfer = 2
 };
 
 struct command_buffer_chunk: public vk::command_buffer

rpcs3/Emu/RSX/VK/VKHelpers.cpp

@@ -654,6 +654,42 @@ namespace vk
 		}
 	}
 
+	VkResult wait_for_event(VkEvent event, u64 timeout)
+	{
+		u64 t = 0;
+		while (true)
+		{
+			switch (const auto status = vkGetEventStatus(*g_current_renderer, event))
+			{
+			case VK_EVENT_SET:
+				return VK_SUCCESS;
+			case VK_EVENT_RESET:
+				break;
+			default:
+				die_with_error(HERE, status);
+				return status;
+			}
+
+			if (timeout)
+			{
+				if (!t)
+				{
+					t = get_system_time();
+					continue;
+				}
+
+				if ((get_system_time() - t) > timeout)
+				{
+					LOG_ERROR(RSX, "[vulkan] vk::wait_for_event has timed out!");
+					return VK_TIMEOUT;
+				}
+			}
+
+			//std::this_thread::yield();
+			_mm_pause();
+		}
+	}
+
 	void die_with_error(const char* faulting_addr, VkResult error_code)
 	{
 		std::string error_message;

rpcs3/Emu/RSX/VK/VKHelpers.h

@@ -181,6 +181,7 @@ namespace vk
 	// Fence reset with driver workarounds in place
 	void reset_fence(VkFence *pFence);
 	VkResult wait_for_fence(VkFence pFence, u64 timeout = 0ull);
+	VkResult wait_for_event(VkEvent pEvent, u64 timeout = 0ull);
 
 	void die_with_error(const char* faulting_addr, VkResult error_code);
 

rpcs3/Emu/RSX/VK/VKTextureCache.h

@@ -36,7 +36,7 @@ namespace vk
 		std::unique_ptr<vk::viewable_image> managed_texture = nullptr;
 
 		//DMA relevant data
-		VkFence dma_fence = VK_NULL_HANDLE;
+		VkEvent dma_fence = VK_NULL_HANDLE;
 		vk::render_device* m_device = nullptr;
 		vk::viewable_image *vram_texture = nullptr;
 		std::unique_ptr<vk::buffer> dma_buffer;
@@ -82,9 +82,9 @@ namespace vk
 			{
 				dma_buffer.reset();
 
-				if (dma_fence != nullptr)
+				if (dma_fence != VK_NULL_HANDLE)
 				{
-					vkDestroyFence(*m_device, dma_fence, nullptr);
+					vkDestroyEvent(*m_device, dma_fence, nullptr);
 					dma_fence = VK_NULL_HANDLE;
 				}
 			}
@@ -164,9 +164,9 @@ namespace vk
 
 			if (dma_fence == VK_NULL_HANDLE)
 			{
-				VkFenceCreateInfo createInfo = {};
-				createInfo.sType = VK_STRUCTURE_TYPE_FENCE_CREATE_INFO;
-				vkCreateFence(*m_device, &createInfo, nullptr, &dma_fence);
+				VkEventCreateInfo createInfo = {};
+				createInfo.sType = VK_STRUCTURE_TYPE_EVENT_CREATE_INFO;
+				vkCreateEvent(*m_device, &createInfo, nullptr, &dma_fence);
 			}
 
 			if (dma_buffer.get() == nullptr)
@@ -297,17 +297,33 @@ namespace vk
 
 			if (manage_cb_lifetime)
 			{
+				VkFence submit_fence;
+				VkFenceCreateInfo create_info{};
+				create_info.sType = VK_STRUCTURE_TYPE_FENCE_CREATE_INFO;
+				vkCreateFence(*m_device, &create_info, nullptr, &submit_fence);
+
 				cmd.end();
-				cmd.submit(submit_queue, {}, dma_fence, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT);
+				cmd.submit(submit_queue, {}, submit_fence, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT);
 
 				// Now we need to restart the command-buffer to restore it to the way it was before...
-				vk::wait_for_fence(dma_fence);
-				vk::reset_fence(&dma_fence);
+				vk::wait_for_fence(submit_fence);
 				CHECK_RESULT(vkResetCommandBuffer(cmd, 0));
 
+				// Cleanup
+				vkDestroyFence(*m_device, submit_fence, nullptr);
+				vkSetEvent(*m_device, dma_fence);
 				if (cmd.access_hint != vk::command_buffer::access_type_hint::all)
+				{
+					// If this is a primary CB, restart it
 					cmd.begin();
 				}
+			}
+			else
+			{
+				// Only used when doing speculation
+				verify(HERE), vkGetEventStatus(*m_device, dma_fence) == VK_EVENT_RESET;
+				vkCmdSetEvent(cmd, dma_fence, VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT | VK_PIPELINE_STAGE_TRANSFER_BIT);
+			}
 
 			synchronized = true;
 			sync_timestamp = get_system_time();
@@ -333,6 +349,10 @@ namespace vk
 		{
 			AUDIT(synchronized);
 
+			// Synchronize, reset dma_fence after waiting
+			vk::wait_for_event(dma_fence, GENERAL_WAIT_TIMEOUT);
+			vkResetEvent(*m_device, dma_fence);
+
 			return dma_buffer->map(offset, size);
 		}