vk: Fix slow bo upload + copy when using RADV

- This reverts commit dbcf1b5a03bfef18f404aa6827515af319abd744.
- Avoid unmapping buffers used for GPU upload/download. It is very slow
  on some platforms.

rpcs3/Emu/RSX/VK/VKDMA.cpp

@@ -34,10 +34,16 @@ namespace vk
 			return inheritance_info.parent->map_range(range);
 		}
 
+		if (memory_mapping == nullptr)
+		{
+			memory_mapping = static_cast<u8*>(allocated_memory->map(0, VK_WHOLE_SIZE));
+			ensure(memory_mapping);
+		}
+
 		ensure(range.start >= base_address);
 		u32 start = range.start;
 		start -= base_address;
-		return allocated_memory->map(start, range.length());
+		return memory_mapping + start;
 	}
 
 	void dma_block::unmap()
@@ -49,6 +55,7 @@ namespace vk
 		else
 		{
 			allocated_memory->unmap();
+			memory_mapping = nullptr;
 		}
 	}
 
@@ -73,6 +80,7 @@ namespace vk
 
 			auto gc = vk::get_resource_manager();
 			gc->dispose(allocated_memory);
+			memory_mapping = nullptr;
 		}
 	}
 
@@ -107,8 +115,7 @@ namespace vk
 		auto dst = vm::get_super_ptr(range.start);
 		std::memcpy(dst, src, range.length());
 
-		// TODO: Clear page bits
+		// NOTE: Do not unmap. This can be extremely slow on some platforms.
-		unmap();
 	}
 
 	void dma_block::load(const utils::address_range& range)
@@ -124,8 +131,7 @@ namespace vk
 		auto dst = map_range(range);
 		std::memcpy(dst, src, range.length());
 
-		// TODO: Clear page bits to sychronized
+		// NOTE: Do not unmap. This can be extremely slow on some platforms.
-		unmap();
 	}
 
 	std::pair<u32, buffer*> dma_block::get(const utils::address_range& range)

rpcs3/Emu/RSX/VK/VKDMA.h

@@ -22,6 +22,7 @@ namespace vk
 		inheritance_info;
 
 		u32 base_address = 0;
+		u8* memory_mapping = nullptr;
 		std::unique_ptr<buffer> allocated_memory;
 
 		virtual void allocate(const render_device& dev, usz size);

rpcs3/Emu/RSX/VK/VKTexture.cpp

@@ -880,7 +880,7 @@ namespace vk
 		VkImageAspectFlags flags, vk::data_heap &upload_heap, u32 heap_align, rsx::flags32_t image_setup_flags)
 	{
 		const bool requires_depth_processing = (dst_image->aspect() & VK_IMAGE_ASPECT_STENCIL_BIT) || (format == CELL_GCM_TEXTURE_DEPTH16_FLOAT);
-		rsx::texture_uploader_capabilities caps{ .supports_zero_copy = true, .alignment = heap_align };
+		rsx::texture_uploader_capabilities caps{ .alignment = heap_align };
 		rsx::texture_memory_info opt{};
 		bool check_caps = true;
 
@@ -896,15 +896,6 @@ namespace vk
 		std::vector<std::pair<VkBuffer, u32>> upload_commands;
 		copy_regions.reserve(subresource_layout.size());
 
-#ifndef _WIN32
-		// RADV workaround. Buffer-to-buffer transfers are extremely slow and invoke memmove in vkCmdCopyBuffer.
-		if (const auto vendor = vk::get_driver_vendor();
-			vendor == driver_vendor::RADV)
-		{
-			caps.supports_zero_copy = false;
-		}
-#endif
-
 		for (const rsx::subresource_layout &layout : subresource_layout)
 		{
 			const auto [row_pitch, upload_pitch_in_texel] = calculate_upload_pitch(format, heap_align, dst_image, layout);
@@ -922,7 +913,7 @@ namespace vk
 			{
 				caps.supports_byteswap = (image_linear_size >= 1024);
 				caps.supports_hw_deswizzle = caps.supports_byteswap;
-				caps.supports_zero_copy = caps.supports_zero_copy && caps.supports_byteswap;
+				caps.supports_zero_copy = caps.supports_byteswap;
 				caps.supports_vtc_decoding = false;
 				check_caps = false;
 			}