vulkan: Improve primitive restart workaround for vega/polaris

- For some reason the hardware forgets that primitive restart is enabled and tries to actually read vertex index 65535
- Works correctly if uint32 vertex indices are used instead of uint16 for cases where primitive restart is active

rpcs3/Emu/RSX/VK/VKGSRender.h

@@ -25,11 +25,12 @@ namespace vk
 	using shader_cache = rsx::shaders_cache<vk::pipeline_props, VKProgramBuffer>;
 }
 
-//Heap allocation sizes in MB - each 'frame' owns a private heap, one of each kind
+//Heap allocation sizes in MB
+//NOTE: Texture uploads can be huge, upto 16MB for a single texture (4096x4096px)
 #define VK_ATTRIB_RING_BUFFER_SIZE_M 256
+#define VK_TEXTURE_UPLOAD_RING_BUFFER_SIZE_M 256
 #define VK_UBO_RING_BUFFER_SIZE_M 64
 #define VK_INDEX_RING_BUFFER_SIZE_M 64
-#define VK_TEXTURE_UPLOAD_RING_BUFFER_SIZE_M 128
 
 #define VK_MAX_ASYNC_CB_COUNT 64
 #define VK_MAX_ASYNC_FRAMES 2

rpcs3/Emu/RSX/VK/VKHelpers.cpp

@@ -15,6 +15,7 @@ namespace vk
 
 	atomic_t<bool> g_cb_no_interrupt_flag { false };
 	atomic_t<bool> g_drv_no_primitive_restart_flag { false };
+	atomic_t<bool> g_drv_force_32bit_indices{ false };
 
 	u64 g_num_processed_frames = 0;
 	u64 g_num_total_frames = 0;
@@ -280,8 +281,9 @@ namespace vk
 	void set_current_renderer(const vk::render_device &device)
 	{
 		g_current_renderer = device;
+		const auto gpu_name = g_current_renderer.gpu().name();
 
-		const std::array<std::string, 8> black_listed =
+/*		const std::array<std::string, 8> black_listed =
 		{
 			// Black list all polaris unless its proven they dont have a problem with primitive restart
 			"RX 580",
@@ -294,7 +296,6 @@ namespace vk
 			"RX Vega",
 		};
 
-		const auto gpu_name = g_current_renderer.gpu().name();
 		for (const auto& test : black_listed)
 		{
 			if (gpu_name.find(test) != std::string::npos)
@@ -302,6 +303,11 @@ namespace vk
 				g_drv_no_primitive_restart_flag = !g_cfg.video.vk.force_primitive_restart;
 				break;
 			}
+		}*/
+
+		if (gpu_name.find("AMD") != std::string::npos)
+		{
+			g_drv_force_32bit_indices = true;
 		}
 	}
 
@@ -310,6 +316,11 @@ namespace vk
 		return g_drv_no_primitive_restart_flag;
 	}
 
+	bool force_32bit_index_buffer()
+	{
+		return g_drv_force_32bit_indices;
+	}
+
 	void change_image_layout(VkCommandBuffer cmd, VkImage image, VkImageLayout current_layout, VkImageLayout new_layout, VkImageSubresourceRange range)
 	{
 		//Prepare an image to match the new layout..

rpcs3/Emu/RSX/VK/VKHelpers.h

@@ -65,7 +65,9 @@ namespace vk
 	vk::render_device *get_current_renderer();
 	void set_current_renderer(const vk::render_device &device);
 
+	//Compatibility workarounds
 	bool emulate_primitive_restart();
+	bool force_32bit_index_buffer();
 
 	VkComponentMapping default_component_map();
 	VkImageSubresource default_image_subresource();

rpcs3/Emu/RSX/VK/VKVertexBuffers.cpp

@@ -123,6 +123,10 @@ namespace
 
 		vertex_input_state operator()(const rsx::draw_indexed_array_command& command)
 		{
+			const bool primitive_restart_enabled = rsx::method_registers.restart_index_enabled();
+			const bool emulate_primitive_restart = primitive_restart_enabled && vk::emulate_primitive_restart();
+			const bool expand_indices_to_32bit = primitive_restart_enabled && !emulate_primitive_restart && vk::force_32bit_index_buffer();
+
 			bool primitives_emulated = false;
 			VkPrimitiveTopology prims = vk::get_appropriate_topology(
 				rsx::method_registers.current_draw_clause.primitive, primitives_emulated);
@@ -131,7 +135,8 @@ namespace
 				rsx::index_array_type::u32 :
 				rsx::method_registers.index_type();
 
-			u32 type_size = gsl::narrow<u32>(get_index_type_size(index_type));
+			rsx::index_array_type upload_type = expand_indices_to_32bit ? rsx::index_array_type::u32 : index_type;
+			u32 type_size = gsl::narrow<u32>(get_index_type_size(upload_type));
 
 			u32 index_count = rsx::method_registers.current_draw_clause.get_elements_count();
 			if (primitives_emulated)
@@ -143,7 +148,7 @@ namespace
 
 			gsl::span<gsl::byte> dst;
 			std::vector<gsl::byte> tmp;
-			if (rsx::method_registers.restart_index_enabled() && vk::emulate_primitive_restart())
+			if (emulate_primitive_restart || (expand_indices_to_32bit && index_type == rsx::index_array_type::u16))
 			{
 				tmp.resize(upload_size);
 				dst = tmp;
@@ -154,7 +159,7 @@ namespace
 			}
 
 			std::optional<std::tuple<VkDeviceSize, VkIndexType>> index_info =
-				std::make_tuple(offset_in_index_buffer, vk::get_index_type(index_type));
+				std::make_tuple(offset_in_index_buffer, vk::get_index_type(upload_type));
 
 			/**
 			* Upload index (and expands it if primitive type is not natively supported).
@@ -175,10 +180,13 @@ namespace
 				return{ prims, 0, 0, 0, 0, index_info };
 			}
 
-			if (rsx::method_registers.restart_index_enabled() && vk::emulate_primitive_restart())
+			if (tmp.size() > 0)
+			{
+				if (emulate_primitive_restart)
 				{
 					//Emulate primitive restart by breaking up the draw calls
 					rsx::method_registers.current_draw_clause.alternate_first_count_commands.resize(0);
+					rsx::method_registers.current_draw_clause.alternate_first_count_commands.reserve(index_count / 3);
 
 					if (index_type == rsx::index_array_type::u16)
 						rsx::split_index_list(reinterpret_cast<u16*>(tmp.data()), index_count, (u16)UINT16_MAX, rsx::method_registers.current_draw_clause.alternate_first_count_commands);
@@ -187,6 +195,22 @@ namespace
 
 					memcpy(buf, tmp.data(), tmp.size());
 				}
+				else
+				{
+					//Force 32-bit indices
+					verify(HERE), index_type == rsx::index_array_type::u16;
+					u32* dst = reinterpret_cast<u32*>(buf);
+					u16* src = reinterpret_cast<u16*>(tmp.data());
+					for (u32 n = 0; n < index_count; ++n)
+					{
+						const auto index = src[n];
+						if (index == UINT16_MAX)
+							dst[n] = UINT32_MAX;
+						else
+							dst[n] = index;
+					}
+				}
+			}
 
 			m_index_buffer_ring_info.unmap();