From 1b305bf789f95d18f382249ac972cbc73f4f9717 Mon Sep 17 00:00:00 2001
From: kd-11 <karokidii@gmail.com>
Date: Sat, 28 May 2022 22:28:40 +0300
Subject: [PATCH] gl: Workaround for poor AMD OpenGL performance

- Turns out the AMD driver really hates it if you render with a mapped index buffer.
  The driver internally seems to make a copy of the consumed indices and uses that. Very slow.
  I was able to isolate this after observing that glDrawArrays is not entirely shit, but glDrawElements duration scaled linearly with the number of vertices.
---
 rpcs3/Emu/RSX/GL/GLGSRender.cpp |  4 +-
 rpcs3/Emu/RSX/GL/GLHelpers.h    | 69 +++++++++++++++++++++++++++++++++
 2 files changed, 71 insertions(+), 2 deletions(-)

diff --git a/rpcs3/Emu/RSX/GL/GLGSRender.cpp b/rpcs3/Emu/RSX/GL/GLGSRender.cpp
index c9904a9e16..a30b477c45 100644
--- a/rpcs3/Emu/RSX/GL/GLGSRender.cpp
+++ b/rpcs3/Emu/RSX/GL/GLGSRender.cpp
@@ -236,14 +236,14 @@ void GLGSRender::on_init_thread()
 		m_vertex_env_buffer = std::make_unique<gl::ring_buffer>();
 		m_texture_parameters_buffer = std::make_unique<gl::ring_buffer>();
 		m_vertex_layout_buffer = std::make_unique<gl::ring_buffer>();
-		m_index_ring_buffer = std::make_unique<gl::ring_buffer>();
+		m_index_ring_buffer = gl_caps.vendor_AMD ? std::make_unique<gl::transient_ring_buffer>() : std::make_unique<gl::ring_buffer>();
 		m_vertex_instructions_buffer = std::make_unique<gl::ring_buffer>();
 		m_fragment_instructions_buffer = std::make_unique<gl::ring_buffer>();
 		m_raster_env_ring_buffer = std::make_unique<gl::ring_buffer>();
 	}
 
 	m_attrib_ring_buffer->create(gl::buffer::target::texture, 256 * 0x100000);
-	m_index_ring_buffer->create(gl::buffer::target::element_array, 64 * 0x100000);
+	m_index_ring_buffer->create(gl::buffer::target::element_array, 16 * 0x100000);
 	m_transform_constants_buffer->create(gl::buffer::target::uniform, 64 * 0x100000);
 	m_fragment_constants_buffer->create(gl::buffer::target::uniform, 16 * 0x100000);
 	m_fragment_env_buffer->create(gl::buffer::target::uniform, 16 * 0x100000);
diff --git a/rpcs3/Emu/RSX/GL/GLHelpers.h b/rpcs3/Emu/RSX/GL/GLHelpers.h
index c91da6a9fb..6d9e605c4b 100644
--- a/rpcs3/Emu/RSX/GL/GLHelpers.h
+++ b/rpcs3/Emu/RSX/GL/GLHelpers.h
@@ -812,6 +812,11 @@ namespace gl
 
 	public:
 
+		virtual void bind()
+		{
+			buffer::bind();
+		}
+
 		virtual void recreate(GLsizeiptr size, const void* data = nullptr)
 		{
 			if (m_id)
@@ -890,6 +895,8 @@ namespace gl
 
 		virtual void unmap() {}
 
+		virtual void flush() {}
+
 		//Notification of a draw command
 		virtual void notify()
 		{
@@ -1011,6 +1018,68 @@ namespace gl
 		void notify() override {}
 	};
 
+	// A non-persistent ring buffer
+	// Internally maps and unmaps data. Uses persistent storage just like the regular persistent variant
+	// Works around drivers that have issues using mapped data for specific sources (e.g AMD proprietary driver with index buffers)
+	class transient_ring_buffer : public ring_buffer
+	{
+		bool dirty = false;
+
+		void* map_internal(u32 offset, u32 length)
+		{
+			flush();
+
+			dirty = true;
+			return DSA_CALL2_RET(MapNamedBufferRange, m_id, offset, length, GL_MAP_WRITE_BIT | GL_MAP_UNSYNCHRONIZED_BIT);
+		}
+
+	public:
+
+		void bind() override
+		{
+			flush();
+			buffer::bind();
+		}
+
+		void recreate(GLsizeiptr size, const void* data = nullptr) override
+		{
+			if (m_id)
+			{
+				m_fence.wait_for_signal();
+				remove();
+			}
+
+			buffer::create();
+			save_binding_state save(current_target(), *this);
+			DSA_CALL2(NamedBufferStorage, m_id, size, data, GL_MAP_WRITE_BIT);
+
+			m_data_loc = 0;
+			m_size = ::narrow<u32>(size);
+			m_memory_type = memory_type::host_visible;
+		}
+
+		std::pair<void*, u32> alloc_from_heap(u32 alloc_size, u16 alignment) override
+		{
+			ensure(m_memory_mapping == nullptr);
+			const auto allocation = ring_buffer::alloc_from_heap(alloc_size, alignment);
+			return { map_internal(allocation.second, alloc_size), allocation.second };
+		}
+
+		void flush() override
+		{
+			if (dirty)
+			{
+				buffer::unmap();
+				dirty = false;
+			}
+		}
+
+		void unmap() override
+		{
+			flush();
+		}
+	};
+
 	class buffer_view
 	{
 		buffer* m_buffer = nullptr;