rsx: Bug fixes

rsx: Give more info when ring buffer allocations fail

rpcs3/Emu/RSX/Common/ring_buffer_helper.h

@@ -69,7 +69,12 @@ public:
 	template<int Alignement>
 	size_t alloc(size_t size)
 	{
-		if (!can_alloc<Alignement>(size)) fmt::throw_exception("Working buffer not big enough" HERE);
+		if (!can_alloc<Alignement>(size))
+		{
+			fmt::throw_exception("Working buffer not big enough, buffer_length=%d allocated=%d requested=%d guard=%d largest_pool=%d" HERE,
+					m_size, m_current_allocated_size, size, m_min_guard_size, m_largest_allocated_pool);
+		}
+
 		size_t alloc_size = align(size, Alignement);
 		size_t aligned_put_pos = align(m_put_pos, Alignement);
 

rpcs3/Emu/RSX/GL/GLGSRender.cpp

@@ -487,7 +487,7 @@ void GLGSRender::on_init_thread()
 	LOG_NOTICE(RSX, "%s", (const char*)glGetString(GL_SHADING_LANGUAGE_VERSION));
 	LOG_NOTICE(RSX, "%s", (const char*)glGetString(GL_VENDOR));
 
-	auto gl_caps = gl::get_driver_caps();
+	auto& gl_caps = gl::get_driver_caps();
 
 	if (!gl_caps.ARB_texture_buffer_supported)
 	{
@@ -499,6 +499,11 @@ void GLGSRender::on_init_thread()
 		fmt::throw_exception("Failed to initialize OpenGL renderer. ARB_direct_state_access or EXT_direct_state_access is required but not supported by your GPU");
 	}
 
+	if (!gl_caps.ARB_depth_buffer_float_supported && g_cfg.video.force_high_precision_z_buffer)
+	{
+		LOG_WARNING(RSX, "High precision Z buffer requested but your GPU does not support GL_ARB_depth_buffer_float. Option ignored.");
+	}
+
 	//Use industry standard resource alignment values as defaults
 	m_uniform_buffer_offset_align = 256;
 	m_min_texbuffer_alignment = 256;

rpcs3/Emu/RSX/GL/GLHelpers.h

@@ -78,11 +78,12 @@ namespace gl
 		bool ARB_buffer_storage_supported = false;
 		bool ARB_texture_buffer_supported = false;
 		bool ARB_shader_draw_parameters_supported = false;
+		bool ARB_depth_buffer_float_supported = false;
 		bool initialized = false;
 
 		void initialize()
 		{
-			int find_count = 5;
+			int find_count = 6;
 			int ext_count = 0;
 			glGetIntegerv(GL_NUM_EXTENSIONS, &ext_count);
 
@@ -127,6 +128,13 @@ namespace gl
 					find_count --;
 					continue;
 				}
+
+				if (ext_name == "GL_ARB_depth_buffer_float")
+				{
+					ARB_depth_buffer_float_supported = true;
+					find_count--;
+					continue;
+				}
 			}
 
 			//Workaround for intel drivers which have terrible capability reporting
@@ -1137,7 +1145,7 @@ namespace gl
 			uint_10_10_10_2 = GL_UNSIGNED_INT_10_10_10_2,
 			uint_2_10_10_10_rev = GL_UNSIGNED_INT_2_10_10_10_REV,
 			uint_24_8 = GL_UNSIGNED_INT_24_8,
-			float32z_s8int = GL_DEPTH32F_STENCIL8,
+			float32_uint8 = GL_FLOAT_32_UNSIGNED_INT_24_8_REV,
 
 			sbyte = GL_BYTE,
 			sshort = GL_SHORT,

rpcs3/Emu/RSX/GL/GLRenderTargets.cpp

@@ -57,8 +57,8 @@ depth_format rsx::internals::surface_depth_format_to_gl(rsx::surface_depth_forma
 	default:
 		LOG_ERROR(RSX, "Surface depth buffer: Unsupported surface depth format (0x%x)", (u32)depth_format);
 	case rsx::surface_depth_format::z24s8:
-		if (g_cfg.video.force_high_precision_z_buffer)
+		if (g_cfg.video.force_high_precision_z_buffer && ::gl::get_driver_caps().ARB_depth_buffer_float_supported)
-			return{ ::gl::texture::type::float32z_s8int, ::gl::texture::format::depth_stencil, ::gl::texture::internal_format::depth32f_stencil8 };
+			return{ ::gl::texture::type::float32_uint8, ::gl::texture::format::depth_stencil, ::gl::texture::internal_format::depth32f_stencil8 };
 		else
 			return{ ::gl::texture::type::uint_24_8, ::gl::texture::format::depth_stencil, ::gl::texture::internal_format::depth24_stencil8 };
 	}

rpcs3/Emu/RSX/VK/VKGSRender.cpp

@@ -959,6 +959,22 @@ void VKGSRender::end()
 	vkCmdBindPipeline(*m_current_command_buffer, VK_PIPELINE_BIND_POINT_GRAPHICS, m_program->pipeline);
 	vkCmdBindDescriptorSets(*m_current_command_buffer, VK_PIPELINE_BIND_POINT_GRAPHICS, pipeline_layout, 0, 1, &descriptor_sets, 0, nullptr);
 
+	if (auto ds = std::get<1>(m_rtts.m_bound_depth_stencil))
+	{
+		if (ds->dirty)
+		{
+			//Clear this surface before drawing on it
+			VkClearValue depth_clear_value;
+			depth_clear_value.depthStencil.depth = 1.f;
+			depth_clear_value.depthStencil.stencil = 255;
+
+			VkClearRect clear_rect = { 0, 0, m_framebuffer_to_clean.back()->width(), m_framebuffer_to_clean.back()->height(), 0, 1 };
+			VkClearAttachment clear_desc = { ds->attachment_aspect_flag, 0, depth_clear_value };
+			vkCmdClearAttachments(*m_current_command_buffer, 1, &clear_desc, 1, &clear_rect);
+			ds->dirty = false;
+		}
+	}
+
 	std::optional<std::tuple<VkDeviceSize, VkIndexType> > index_info = std::get<2>(upload_info);
 
 	if (!index_info)
@@ -1141,6 +1157,12 @@ void VKGSRender::clear_surface(u32 mask)
 	vkCmdClearAttachments(*m_current_command_buffer, (u32)clear_descriptors.size(), clear_descriptors.data(), (u32)clear_regions.size(), clear_regions.data());
 
 	vkCmdEndRenderPass(*m_current_command_buffer);
+
+	if (mask & 0x3)
+	{
+		if (std::get<0>(m_rtts.m_bound_depth_stencil) != 0)
+			std::get<1>(m_rtts.m_bound_depth_stencil)->dirty = false;
+	}
 }
 
 void VKGSRender::sync_at_semaphore_release()

rpcs3/Emu/RSX/VK/VKRenderTargets.h

@@ -11,6 +11,7 @@ namespace vk
 {
 	struct render_target : public image
 	{
+		bool dirty = false;
 		u16 native_pitch = 0;
 		VkImageAspectFlags attachment_aspect_flag = VK_IMAGE_ASPECT_COLOR_BIT;
 
@@ -143,19 +144,9 @@ namespace rsx
 
 		static void invalidate_rtt_surface_contents(vk::command_buffer*, vk::render_target*) {}
 		
-		static void invalidate_depth_surface_contents(vk::command_buffer* pcmd, vk::render_target *ds)
+		static void invalidate_depth_surface_contents(vk::command_buffer* /*pcmd*/, vk::render_target *ds)
 		{
-			VkImageLayout old_layout = ds->current_layout;
+			ds->dirty = true;
-			VkImageSubresourceRange range = vk::get_image_subresource_range(0, 0, 1, 1, ds->attachment_aspect_flag);
-			change_image_layout(*pcmd, ds, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, range);
-
-			//Clear the surface before drawing on it
-			VkClearDepthStencilValue clear_depth = {};
-			clear_depth.depth = 1.f;
-			clear_depth.stencil = 255;
-
-			vkCmdClearDepthStencilImage(*pcmd, ds->value, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, &clear_depth, 1, &range);
-			change_image_layout(*pcmd, ds, old_layout, range);
 		}
 
 		static bool rtt_has_format_width_height(const std::unique_ptr<vk::render_target> &rtt, surface_color_format format, size_t width, size_t height)