rsx/vk: Implement bulk aligned allocator

- Avoids wasting space and allows use of natural arrays in shaders

rpcs3/Emu/RSX/Common/ring_buffer_helper.h

@@ -3,7 +3,9 @@
 #include "Utilities/StrFmt.h"
 #include "util/asm.hpp"
 
-/**
+namespace rsx
+{
+	/**
 	 * Ring buffer memory helper :
 	 * There are 2 "pointers" (offset inside a memory buffer to be provided by class derivative)
 	 * PUT pointer "points" to the start of allocatable space.
@@ -11,9 +13,9 @@
 	 * Space between GET and PUT is used by the GPU ; this structure check that this memory is not overwritten.
 	 * User has to update the GET pointer when synchronisation happens.
 	 */
-class data_heap
-{
-protected:
+	class data_heap
+	{
+	protected:
 		/**
 		* Internal implementation of allocation test
 		* Does alloc cross get position?
@@ -70,13 +72,13 @@ protected:
 		usz m_min_guard_size;          // If an allocation touches the guard region, reset the heap to avoid going over budget
 
 		char* m_name;
-public:
+	public:
 		data_heap() = default;
 		~data_heap() = default;
 		data_heap(const data_heap&) = delete;
 		data_heap(data_heap&&) = delete;
 
-	void init(usz heap_size, const char* buffer_name = nullptr, usz min_guard_size=0x10000)
+		void init(usz heap_size, const char* buffer_name = nullptr, usz min_guard_size = 0x10000)
 		{
 			m_name = const_cast<char*>(buffer_name ? buffer_name : "<unnamed>");
 
@@ -166,4 +168,35 @@ public:
 		{
 			return m_size;
 		}
-};
+
+		// Bulk static allocator. Allows to allocate one large block and subdivide
+		// [ 0, 1, 2, 3 ] <pad> [ 4, 5, 6, 7 ] ...
+		template <usz Alignment, usz ElementSize = Alignment>
+		struct bulk_allocator
+		{
+			bulk_allocator(data_heap& container, u32 batch_size = 1)
+				: m_container(container)
+				, m_batch_size(batch_size)
+			{}
+
+			usz alloc()
+			{
+				if (!m_capacity)
+				{
+					m_address = m_container.alloc<Alignment>(ElementSize * m_batch_size);
+					m_capacity = m_batch_size;
+				}
+
+				m_capacity--;
+				return std::exchange(m_address, m_address + ElementSize);
+			}
+
+		private:
+			data_heap& m_container;
+			usz m_address = 0;
+
+			u32 m_capacity = 0;
+			u32 m_batch_size = 1;
+		};
+	};
+}

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

@@ -15,7 +15,7 @@ namespace vk
 
 	void data_heap::create(VkBufferUsageFlags usage, usz size, rsx::flags32_t flags, const char* name, usz guard, VkBool32 notify)
 	{
-		::data_heap::init(size, name, guard);
+		rsx::data_heap::init(size, name, guard);
 
 		const auto& memory_map = g_render_device->get_memory_mapping();
 
@@ -135,7 +135,7 @@ namespace vk
 		auto memory_index = m_prefer_writethrough ? memory_map.device_bar : memory_map.host_visible_coherent;
 
 		// Update heap information and reset the allocator
-		::data_heap::init(aligned_new_size, m_name, m_min_guard_size);
+		rsx::data_heap::init(aligned_new_size, m_name, m_min_guard_size);
 
 		// Discard old heap and create a new one. Old heap will be garbage collected when no longer needed
 		auto gc = get_resource_manager();
@@ -188,7 +188,7 @@ namespace vk
 		return after_usage < limit;
 	}
 
-	void* data_heap::map(usz offset, usz size)
+	void* data_heap::map_impl(usz offset, usz size)
 	{
 		if (!_ptr)
 		{

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

@@ -20,7 +20,7 @@ namespace vk
 		heap_pool_force_vram_shadow  = (1 << 2),
 	};
 
-	class data_heap : public ::data_heap
+	class data_heap : public rsx::data_heap
 	{
 	private:
 		usz initial_size = 0;
@@ -41,6 +41,8 @@ namespace vk
 		bool grow(usz size) override;
 		bool can_allocate_heap(const vk::memory_type_info& target_heap, usz size, int max_usage_percent);
 
+		void* map_impl(usz offset, usz size);
+
 	public:
 		std::unique_ptr<buffer> heap;
 
@@ -51,9 +53,16 @@ namespace vk
 		void create(VkBufferUsageFlags usage, usz size, rsx::flags32_t flags, const char* name, usz guard = 0x10000, VkBool32 notify = VK_FALSE);
 		void destroy();
 
-		void* map(usz offset, usz size);
+		template <typename T = void>
+		T* map(usz offset, usz size)
+		{
+			return reinterpret_cast<T*>(map_impl(offset, size));
+		}
+
 		void unmap(bool force = false);
 
+		void sync(const vk::command_buffer& cmd);
+
 		template<int Alignment, typename T = char>
 			requires std::is_trivially_destructible_v<T>
 		std::pair<usz, T*> alloc_and_map(usz count)
@@ -63,8 +72,6 @@ namespace vk
 			return { addr, reinterpret_cast<T*>(map(addr, size_bytes)) };
 		}
 
-		void sync(const vk::command_buffer& cmd);
-
 		template <usz Alignment>
 		VkDescriptorBufferInfoEx window(usz offset, usz range, u64 window_size) const
 		{