vk: Use bulk allocator for FP and VP constant blocks to reduce wastage

rpcs3/Emu/RSX/Common/ring_buffer_helper.h

@@ -179,16 +179,27 @@ namespace rsx
 				, m_batch_size(batch_size)
 			{}
 
-			usz alloc()
+			usz alloc(u32 element_count = 1)
 			{
-				if (!m_capacity)
+				if (m_capacity < element_count)
 				{
+					ensure(element_count <= m_batch_size);
 					m_address = m_container.alloc<Alignment>(ElementSize * m_batch_size);
 					m_capacity = m_batch_size;
 				}
 
-				m_capacity--;
+				m_capacity -= element_count;
-				return std::exchange(m_address, m_address + ElementSize);
+				return std::exchange(m_address, m_address + (ElementSize * element_count));
+			}
+
+			usz alloc_bytes(usz size = ElementSize)
+			{
+				return alloc(static_cast<u32>(size / ElementSize));
+			}
+
+			u32 capacity() const
+			{
+				return m_capacity;
 			}
 
 		private:

rpcs3/Emu/RSX/VK/VKGSRender.cpp

@@ -568,6 +568,16 @@ VKGSRender::VKGSRender(utils::serial* ar) noexcept : GSRender(ar)
 		m_vertex_env_ring_info,
 		std::min<u32>(limits.maxUniformBufferRange / 96u, 1024u));
 
+	m_transform_constants_allocator = std::make_unique<rsx::data_heap::bulk_allocator<256, 16>>(
+		m_transform_constants_ring_info,
+		std::min<u32>(limits.maxUniformBufferRange / 16u, 8192u)
+	);
+
+	m_fragment_constants_allocator = std::make_unique<rsx::data_heap::bulk_allocator<256, 16>>(
+		m_fragment_constants_ring_info,
+		std::min<u32>(limits.maxUniformBufferRange / 16u, 8192u)
+	);
+
 	if (m_texbuffer_view_size < 0x800000)
 	{
 		// Warn, only possibly expected on macOS
@@ -2006,7 +2016,7 @@ void VKGSRender::load_program_env()
 		usz mem_offset = 0;
 		auto alloc_storage = [&](usz size) -> std::pair<void*, usz>
 		{
-			mem_offset = m_transform_constants_ring_info.alloc<256>(size);
+			mem_offset = m_transform_constants_allocator->alloc_bytes(size);
 			return std::make_pair(m_transform_constants_ring_info.map(mem_offset, size), size);
 		};
 
@@ -2028,7 +2038,7 @@ void VKGSRender::load_program_env()
 		// Fragment constants
 		if (fragment_constants_size)
 		{
-			m_fragment_constants_dynamic_offset = m_fragment_constants_ring_info.alloc<256>(fragment_constants_size);
+			m_fragment_constants_dynamic_offset = m_fragment_constants_allocator->alloc_bytes(fragment_constants_size);
 			auto buf = m_fragment_constants_ring_info.map(m_fragment_constants_dynamic_offset, fragment_constants_size);
 
 			m_prog_buffer->fill_fragment_constants_buffer({ reinterpret_cast<float*>(buf), fragment_constants_size },

rpcs3/Emu/RSX/VK/VKGSRender.h

@@ -162,6 +162,8 @@ private:
 	u64 m_stipple_array_dynamic_offset = 0;
 
 	std::unique_ptr<rsx::data_heap::bulk_allocator<256, 96>> m_vertex_env_allocator;
+	std::unique_ptr<rsx::data_heap::bulk_allocator<256, 16>> m_transform_constants_allocator;
+	std::unique_ptr<rsx::data_heap::bulk_allocator<256, 16>> m_fragment_constants_allocator;
 
 	std::vector<vk::frame_context_t> m_frame_context_storage;
 	u32 m_max_async_frames = 0u;