vk: Keep the total number of allocated samplers under control

rpcs3/Emu/RSX/VK/VKDraw.cpp

@@ -285,8 +285,14 @@ void VKGSRender::load_texture_env()
 
 				if (replace)
 				{
-					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,
+					fs_sampler_handles[i] = vk::get_resource_manager()->get_sampler(
-						min_filter.filter, mag_filter, min_filter.mipmap_mode, border_color, compare_enabled, depth_compare_mode);
+						*m_device,
+						fs_sampler_handles[i],
+						wrap_s, wrap_t, wrap_r,
+						false,
+						lod_bias, af_level, min_lod, max_lod,
+						min_filter.filter, mag_filter, min_filter.mipmap_mode,
+						border_color, compare_enabled, depth_compare_mode);
 				}
 			}
 			else
@@ -338,8 +344,9 @@ void VKGSRender::load_texture_env()
 
 				if (replace)
 				{
-					vs_sampler_handles[i] = vk::get_resource_manager()->find_sampler(
+					vs_sampler_handles[i] = vk::get_resource_manager()->get_sampler(
 						*m_device,
+						vs_sampler_handles[i],
 						VK_SAMPLER_ADDRESS_MODE_REPEAT, VK_SAMPLER_ADDRESS_MODE_REPEAT, VK_SAMPLER_ADDRESS_MODE_REPEAT,
 						unnormalized_coords,
 						0.f, 1.f, min_lod, max_lod,

rpcs3/Emu/RSX/VK/VKPresent.cpp

@@ -223,6 +223,8 @@ void VKGSRender::frame_context_cleanup(vk::frame_context_t *ctx, bool free_resou
 			m_overlay_manager->dispose(uids_to_dispose);
 		}
 
+		vk::get_resource_manager()->trim();
+
 		vk::reset_global_resources();
 
 		ctx->buffer_views_to_clean.clear();

rpcs3/Emu/RSX/VK/VKResourceManager.cpp

@@ -31,6 +31,34 @@ namespace vk
 		return &g_resource_manager;
 	}
 
+	void resource_manager::trim()
+	{
+		// For any managed resources, try to keep the number of unused/idle resources as low as possible.
+		// Improves search times as well as keeping us below the hardware limit.
+		const auto limits = get_current_renderer()->gpu().get_limits();
+		const auto allocated_sampler_count = vmm_get_application_pool_usage(VMM_ALLOCATION_POOL_SAMPLER);
+		const auto max_allowed_samplers = std::min((limits.maxSamplerAllocationCount * 3u) / 4u, 2048u);
+
+		if (allocated_sampler_count > max_allowed_samplers)
+		{
+			ensure(max_allowed_samplers);
+			rsx_log.warning("Trimming allocated samplers. Allocated = %u, Max = %u", allocated_sampler_count, limits.maxSamplerAllocationCount);
+
+			auto& disposed_samplers_pool = get_current_eid_scope().m_disposed_samplers;
+			for (auto It = m_sampler_pool.begin(); It != m_sampler_pool.end();)
+			{
+				if (!It->second->has_refs())
+				{
+					disposed_samplers_pool.emplace_back(std::move(It->second));
+					It = m_sampler_pool.erase(It);
+					continue;
+				}
+
+				++It;
+			}
+		}
+	}
+
 	u64 get_event_id()
 	{
 		return g_event_ctr++;
@@ -212,4 +240,16 @@ namespace vk
 			vmm_handle_memory_pressure(load_severity);
 		}
 	}
+
+	void vmm_notify_object_allocated(vmm_allocation_pool pool)
+	{
+		ensure(pool >= VMM_ALLOCATION_POOL_SAMPLER);
+		g_vmm_stats.pool_usage[pool]++;
+	}
+
+	void vmm_notify_object_freed(vmm_allocation_pool pool)
+	{
+		ensure(pool >= VMM_ALLOCATION_POOL_SAMPLER);
+		g_vmm_stats.pool_usage[pool]--;
+	}
 }

rpcs3/Emu/RSX/VK/VKResourceManager.h

@@ -23,6 +23,7 @@ namespace vk
 		std::vector<std::unique_ptr<vk::image>> m_disposed_images;
 		std::vector<std::unique_ptr<vk::event>> m_disposed_events;
 		std::vector<std::unique_ptr<vk::query_pool>> m_disposed_query_pools;
+		std::vector<std::unique_ptr<vk::sampler>> m_disposed_samplers;
 
 		eid_scope_t(u64 _eid):
 			eid(_eid), m_device(g_render_device)
@@ -40,13 +41,19 @@ namespace vk
 			m_disposed_image_views.clear();
 			m_disposed_images.clear();
 			m_disposed_query_pools.clear();
+			m_disposed_samplers.clear();
 		}
 	};
 
 	class resource_manager
 	{
 	private:
-		std::unordered_map<u64, std::unique_ptr<vk::sampler>> m_sampler_pool;
+		struct cached_sampler_object_t : public vk::sampler, public rsx::ref_counted
+		{
+			using vk::sampler::sampler;
+		};
+
+		std::unordered_map<u64, std::unique_ptr<cached_sampler_object_t>> m_sampler_pool;
 		std::deque<eid_scope_t> m_eid_map;
 
 		eid_scope_t& get_current_eid_scope()
@@ -98,7 +105,8 @@ namespace vk
 			m_sampler_pool.clear();
 		}
 
-		vk::sampler* find_sampler(const vk::render_device& dev, VkSamplerAddressMode clamp_u, VkSamplerAddressMode clamp_v, VkSamplerAddressMode clamp_w,
+		vk::sampler* get_sampler(const vk::render_device& dev, vk::sampler* previous,
+			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)
@@ -115,20 +123,30 @@ namespace vk
 			key |= u64(encode_fxp<true>(mipLodBias)) << 44; // 13 bits
 			key |= u64(max_anisotropy) << 57;               // 4 bits
 
+			if (previous)
+			{
+				auto as_cached_object = static_cast<cached_sampler_object_t*>(previous);
+				ensure(as_cached_object->has_refs());
+				as_cached_object->release();
+			}
+
 			if (const auto found = m_sampler_pool.find(key);
 				found != m_sampler_pool.end())
 			{
+				found->second->add_ref();
 				return found->second.get();
 			}
 
-			auto result = std::make_unique<vk::sampler>(
+			auto result = std::make_unique<cached_sampler_object_t>(
 				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.first->second.get();
+			auto ret = It.first->second.get();
+			ret->add_ref();
+			return ret;
 		}
 
 		void dispose(std::unique_ptr<vk::buffer>& buf)
@@ -162,6 +180,11 @@ namespace vk
 			get_current_eid_scope().m_disposed_query_pools.emplace_back(std::move(pool));
 		}
 
+		void dispose(std::unique_ptr<vk::sampler>& sampler)
+		{
+			get_current_eid_scope().m_disposed_samplers.emplace_back(std::move(sampler));
+		}
+
 		void eid_completed(u64 eid)
 		{
 			while (!m_eid_map.empty())
@@ -177,6 +200,8 @@ namespace vk
 				}
 			}
 		}
+
+		void trim();
 	};
 
 	struct vmm_allocation_t

rpcs3/Emu/RSX/VK/vkutils/memory.h

@@ -18,6 +18,7 @@ namespace vk
 			VMM_ALLOCATION_POOL_TEXTURE_CACHE,
 			VMM_ALLOCATION_POOL_SWAPCHAIN,
 			VMM_ALLOCATION_POOL_SCRATCH,
+			VMM_ALLOCATION_POOL_SAMPLER,
 		};
 	}
 
@@ -173,6 +174,7 @@ namespace vk
 		void* m_host_pointer;
 	};
 
+	// Tracking for memory usage. Constrained largely by amount of VRAM + shared video memory.
 	void vmm_notify_memory_allocated(void* handle, u32 memory_type, u64 memory_size, vmm_allocation_pool pool);
 	void vmm_notify_memory_freed(void* handle);
 	void vmm_reset();
@@ -182,5 +184,9 @@ namespace vk
 	bool vmm_handle_memory_pressure(rsx::problem_severity severity);
 	rsx::problem_severity vmm_determine_memory_load_severity();
 
+	// Tracking for host memory objects. Allocated count is more important than actual memory amount.
+	void vmm_notify_object_allocated(vmm_allocation_pool pool);
+	void vmm_notify_object_freed(vmm_allocation_pool pool);
+
 	mem_allocator_base* get_current_mem_allocator();
 }

rpcs3/Emu/RSX/VK/vkutils/sampler.cpp

@@ -1,3 +1,4 @@
+#include "memory.h"
 #include "sampler.h"
 #include "../../rsx_utils.h"
 
@@ -27,11 +28,13 @@ namespace vk
 		info.borderColor = border_color;
 
 		CHECK_RESULT(vkCreateSampler(m_device, &info, nullptr, &value));
+		vmm_notify_object_allocated(VMM_ALLOCATION_POOL_SAMPLER);
 	}
 
 	sampler::~sampler()
 	{
 		vkDestroySampler(m_device, value, nullptr);
+		vmm_notify_object_freed(VMM_ALLOCATION_POOL_SAMPLER);
 	}
 
 	bool sampler::matches(VkSamplerAddressMode clamp_u, VkSamplerAddressMode clamp_v, VkSamplerAddressMode clamp_w,