From 42aa4c5000f6d1023d8ef31be915983f540b989a Mon Sep 17 00:00:00 2001
From: kd-11 <karokidii@gmail.com>
Date: Sat, 12 Oct 2019 13:32:49 +0300
Subject: [PATCH] gl: Vendor-specific tuning

---
 rpcs3/Emu/RSX/GL/GLHelpers.h      | 13 ++++++-
 rpcs3/Emu/RSX/GL/GLTextureCache.h | 58 ++++++++-----------------------
 2 files changed, 27 insertions(+), 44 deletions(-)

diff --git a/rpcs3/Emu/RSX/GL/GLHelpers.h b/rpcs3/Emu/RSX/GL/GLHelpers.h
index 02e9b8bf83..d35fce0c7a 100644
--- a/rpcs3/Emu/RSX/GL/GLHelpers.h
+++ b/rpcs3/Emu/RSX/GL/GLHelpers.h
@@ -764,7 +764,8 @@ namespace gl
 
 		void allocate(GLsizeiptr size, const void* data_, memory_type type, GLenum usage)
 		{
-			if (get_driver_caps().ARB_buffer_storage_supported)
+			if (const auto& caps = get_driver_caps();
+				caps.ARB_buffer_storage_supported)
 			{
 				target target_ = current_target();
 				save_binding_state save(target_, *this);
@@ -789,6 +790,16 @@ namespace gl
 					}
 				}
 
+				if ((flags & GL_MAP_READ_BIT) && !caps.vendor_AMD)
+				{
+					// This flag stops NVIDIA from allocating read-only memory in VRAM.
+					// NOTE: On AMD, allocating client-side memory via CLIENT_STORAGE_BIT or
+					// making use of GL_AMD_pinned_memory brings everything down to a crawl.
+					// Afaict there is no reason for this; disabling pixel pack/unpack operations does not alleviate the problem.
+					// The driver seems to eventually figure out the optimal storage location by itself.
+					flags |= GL_CLIENT_STORAGE_BIT;
+				}
+
 				glBufferStorage((GLenum)target_, size, data_, flags);
 				m_size = size;
 			}
diff --git a/rpcs3/Emu/RSX/GL/GLTextureCache.h b/rpcs3/Emu/RSX/GL/GLTextureCache.h
index 98c8c655db..729beb4c5c 100644
--- a/rpcs3/Emu/RSX/GL/GLTextureCache.h
+++ b/rpcs3/Emu/RSX/GL/GLTextureCache.h
@@ -50,8 +50,7 @@ namespace gl
 		friend baseclass;
 
 		fence m_fence;
-		u32 pbo_id = 0;
-		u32 pbo_size = 0;
+		buffer pbo;
 
 		gl::viewable_image* vram_texture = nullptr;
 
@@ -66,23 +65,16 @@ namespace gl
 			const u32 vram_size = src->pitch() * src->height();
 			const u32 buffer_size = align(vram_size, 4096);
 
-			if (pbo_id)
+			if (pbo)
 			{
-				if (pbo_size >= buffer_size)
+				if (pbo.size() >= buffer_size)
 					return;
 
-				glDeleteBuffers(1, &pbo_id);
-				pbo_id = 0;
-				pbo_size = 0;
+				pbo.remove();
 			}
 
-			glGenBuffers(1, &pbo_id);
-
-			glBindBuffer(GL_PIXEL_PACK_BUFFER, pbo_id);
-			glBufferStorage(GL_PIXEL_PACK_BUFFER, buffer_size, nullptr, GL_MAP_READ_BIT);
+			pbo.create(buffer::target::pixel_pack, buffer_size, nullptr, buffer::memory_type::host_visible, GL_STREAM_READ);
 			glBindBuffer(GL_PIXEL_PACK_BUFFER, GL_NONE);
-
-			pbo_size = buffer_size;
 		}
 
 	public:
@@ -157,7 +149,7 @@ namespace gl
 			init_buffer(src);
 
 			glGetError();
-			glBindBuffer(GL_PIXEL_PACK_BUFFER, pbo_id);
+			pbo.bind(buffer::target::pixel_pack);
 
 			if (context == rsx::texture_upload_context::dma)
 			{
@@ -179,7 +171,7 @@ namespace gl
 
 			pixel_pack_settings pack_settings;
 			pack_settings.alignment(1);
-			pack_settings.swap_bytes(pack_unpack_swap_bytes);
+			//pack_settings.swap_bytes(pack_unpack_swap_bytes);
 
 			src->copy_to(nullptr, format, type, pack_settings);
 			real_pitch = src->pitch();
@@ -276,25 +268,6 @@ namespace gl
 			dma_transfer(cmd, target_texture, {}, {}, rsx_pitch);
 		}
 
-		void fill_texture(gl::texture* tex)
-		{
-			if (!synchronized)
-			{
-				//LOG_WARNING(RSX, "Request to fill texture rejected because contents were not read");
-				return;
-			}
-
-			u32 min_width = std::min((u16)tex->width(), width);
-			u32 min_height = std::min((u16)tex->height(), height);
-
-			glBindTexture(GL_TEXTURE_2D, tex->id());
-			glPixelStorei(GL_UNPACK_SWAP_BYTES, pack_unpack_swap_bytes);
-			glBindBuffer(GL_PIXEL_UNPACK_BUFFER, pbo_id);
-			glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, min_width, min_height, (GLenum)format, (GLenum)type, nullptr);
-			glBindBuffer(GL_PIXEL_UNPACK_BUFFER, GL_NONE);
-		}
-
-
 		/**
 		 * Flush
 		 */
@@ -304,8 +277,9 @@ namespace gl
 
 			m_fence.wait_for_signal();
 
-			verify(HERE), (offset + size) <= pbo_size;
-			glBindBuffer(GL_PIXEL_PACK_BUFFER, pbo_id);
+			verify(HERE), (offset + size) <= pbo.size();
+			pbo.bind(buffer::target::pixel_pack);
+
 			return glMapBufferRange(GL_PIXEL_PACK_BUFFER, offset, size, GL_MAP_READ_BIT);
 		}
 
@@ -373,21 +347,19 @@ namespace gl
 		 */
 		void destroy()
 		{
-			if (!is_locked() && pbo_id == 0 && vram_texture == nullptr && m_fence.is_empty() && !managed_texture)
+			if (!is_locked() && !pbo && vram_texture == nullptr && m_fence.is_empty() && !managed_texture)
 				//Already destroyed
 				return;
 
-			if (pbo_id != 0)
+			if (pbo)
 			{
-				//Destroy pbo cache since vram texture is managed elsewhere
-				glDeleteBuffers(1, &pbo_id);
+				// Destroy pbo cache since vram texture is managed elsewhere
+				pbo.remove();
 				scaled_texture.reset();
 			}
-			managed_texture.reset();
 
+			managed_texture.reset();
 			vram_texture = nullptr;
-			pbo_id = 0;
-			pbo_size = 0;
 
 			if (!m_fence.is_empty())
 			{