Vulkan API 杨瑜, 资深工程师

Transcription

1 Vulkan API 杨瑜, 资深工程师

2 Vulkan Overview (1/3) Some History ~2011 became apparent that the API is getting in the way - Console Developers programmed GPUs To-the-Metal 2012 Khronos started work on GLCommon 2014 Khronos accelerated Vulkan development when AMD donated Mantle Khronos switched focus to a new, clean modern API 2016 Vulkan 1.0 released Qualcomm s Role Qualcomm involved from the start Qualcomm is a promoter member (voting rights, etc) Activate participant in Vulkan working groups Propose extensions to the Vulkan API Qualcomm Snapdragon is a product of Qualcomm Technologies, Inc.

3 Vulkan Overview (2/3) Vulkan is a new graphics and compute API OpenGL introduced back in 1992 Vulkan launched in 2016 Designed for Modern GPUs Compute + Graphics capability Unified Memory Any OpenGL ES 3.1 or 4.x GPU Cross Platform Mobile/Desktop/Console/Embedded Mobile is a first class citizen (no ES type subset of API calls) Cross Operating System Android Windows 7,8,1 Linux

4 Vulkan Overview (3/3) Streamlined Greatly reduced CPU overhead Faster performance Lower overhead Less latency Multi-core Efficient Multithreading friendly Encouraged Explicit API Developer has direct control over the GPU Features aimed at increasing application (graphics & compute) control and performance.

5 Vulkan Overview (3/3) Consistency Improved reliability over OpenGL Consistency between implementations Validation Layers Conformance Architecture-neutral Full support for tile-based as well as direct renders Same API for desktop, mobile,console embedded Same API supports both graphics and compute (Compute support is mandatory in drivers) Features aimed at application portability and support for various architectures (mobile, desktop, automotive etc.)

6 Vulkan Advantages over OpenGL (1/5) Explicit GPU Control Driver is more predictable less hitching Error and validation checking controllable Easy to disable in production code Pre-compiled Shaders in SPIR-V format No front-end compiling done no differences between vendor compilers Simpler driver Driver is easier to developer and maintain Improved driver performance Lower driver overhead and latency Less cross driver vendor unpredictability Application GPU Driver API (Heavy with extra checks and memory mgmt. etc.) GPU hardware Applicatio n GPU Driver API (Thin) GPU hardware Language Front-end Compilers Initially GLSL Loadable debug and validation layers Vulkan 1.0 provides access to OpenGL ES 3.1/OpenGL 4.x-class GPU functionality, but with increased performance and flexibility

7 Vulkan Advantages over OpenGL (2/5) Layered Architecture Debug and Validation Layers loaded only when needed New utility layers can and will be developed by 3 rd parties Rich area for innovation SPIR-V Shader Language No shader compilation needed SPIR-V is byte code, not shader source Minimizes cross vendor differences Allows for other languages (e.g. HLSL) to be supported Application/Gam e Vulkan Loader Driver Adreno GPU Layer 1 Layer 2 Layer 3 Layer N

8 Vulkan Advantages over OpenGL (3/5) Reduced CPU Overhead to enable higher GPU usage OpenGL ES CPU utilization is high, causing higher power usage GPU utilization is more likely to be restricted waiting for CPU to complete Vulkan CPU overhead is drastically reduced. Lower power consumption on CPU. GPU spends less time waiting and more time rendering higher draw calls and primitive content Higher frame rates can be achieved for the same power envelope

9 Vulkan Advantages over OpenGL (4/5) Efficient, Predictable multi-threaded behavior Multithreading friendly and encouraged Easy to take advantage of multiple cores on device Portions of API designed for per-thread usage (e.g. CommandBuffers) Impossible to do in OpenGL because command buffer generation and submission are done at the same time (i.e. gldraw)

10 Vulkan OpenGL ES Vulkan Advantages over OpenGL (5/5) Pipeline State Reusability Vulkan Pipeline State Object (PSO)contains full rendering state information Frame 1 Frame 2 Frame 3 Frame.. Frame 1 - Set blend state - Set depth state - Set.. - Set.. - Set UBO - Draw - Set blend state - Set depth state - Set.. - Set.. - Set UBO - Draw - Set blend state - Set depth state - Set.. - Set.. - Set UBO - Draw Frame 2 - Create PSO - Set PSO - Set UBO - Draw - Set PSO - Set UBO - Draw - Set PSO - Set UBO - Draw Vulkan provides major advantages with State reusability which can be a big overhead in OpenGL ES when most of the render state is static.

11 Vulkan Disadvantages over OpenGL Explicit GPU Control All GPU operations needs to be specified in detail New API to learn Not OpenGL++ New concepts: synchronization, multipass, image layouts, Terminology, work-flow Not obvious how to port existing OpenGL applications No OpenGL state machine No 1 to 1 API call substitution Much more setup before rendering occurs? Device support growing, but not at 100% like OpenGL ES

12 Good Reasons to use Vulkan Your App is CPU bound and can be parallelized You want to minimize glitches and have predicated driver behavior You re starting work on a new graphics game No/little preexisting graphics code (reusing assets ok) Doesn t require the use of an existing game engine Or your game engine has support for Vulkan You want to create a new graphics engine Designed for modern GPUs Backwards hardware support not an issue

13 Developer Tips Threading (1/2) Threading Goals Key GPU Busy Take advantage of multiple cores on device Vulkan and Threading Threading is effective and encouraged with Vulkan API is designed from the ground up to be multithreading friendly Much of the spec details thread safety Obvious places to do threading Generating Command Buffers Non Obvious places to do threading Building Pipelines Loading and Updating Resources (e.g. Vertex, Uniform buffers) Other CPU intensive/time consuming operations (e.g. physics, animation)

14 Developer Tips Threading (2/2) Application must manage threads Divide work up Spawn threads (submitting to queues) Synchronize and Combine results Example: Command Pools Allocate command buffers from command pool Use one command pool per thread Have one master thread Have multiple slave threads that build multiple command buffers Master thread submits finished command buffers and handles presenting and swapping Command pool can be easily reset (once all command buffers from it have been executed)

15 Developer Tips Synchronization Application must handle Synchronization in Vulkan New responsibility driver handled this function in OpenGL Vulkan Synchronization Objects Semaphores Synchronize work across queues Synchronize work in a single queue Events and Barriers Synchronize work in a command buffer Synchronize work in a sequence of command buffers submitted to a single queue Fences Synchronize work between device and host

16 Developer Tips Synchronization (1/2) Probably one of the confusing topics in Vulkan Best advice is to study examples for patterns Barriers are useful for changing image layouts Submit multiple barriers into a single vkcmdpipelinebarrier call Events are useful for GPU/CPU synchronization Semaphore are useful for synchronizing queues Fences are useful for frame completion Particularly if using a multibuffer rendering sheme WaitIdle Avoid using vkqueuewaitidle Avoid using VkDeviceWaitIdle Not useful towards goal of always keeping the GPU busy Should only be used towards the app s teardown phase

17 Developer Tips - Pipelines Pipelines are monolithic objects that contain: State for the entire rendering pipeline Shader Stages/Vertex Input/Assembly Viewport/Rasterization/DepthStencil Multisample/ColorBlend Time consuming to build Build in Parallel Use pipeline caches Use derivative pipelines

18 Developer Tips Renderpass (1/2) Alternative technique for rendering in multiple passes Example: deferred lighting: Gbuffer Pass + Lighting Pass Gbuffer Pass Normal s Albedo Depth RenderPass: Full connected graph of rendering process Dependencies between passes vkcmdbeginrenderpass Lighting Pass Command Buffer Recording Gbuffer Commands vkcmdnextsubpass Lighting Commands vkcmdnextsubpass Postprocessing Commands vkcmdendrenderpass

19 Developer Tips Renderpass (2/2) Allows graphics driver and GPU to be much more efficient Renderpasses are particularly efficient on tilebased GPUs (like Qualcomm Adreno GPU) Minimizes transfers to/from Graphics Memory and System Memory Entire tile can be rendered at same time Only one resolve needed, regardless of the number of subpasses. Efficient use of attachments (render targets) Recommendation to use Vulkan renderpasses with multiple passes whenever possible Qualcomm Adreno GPU is a product of Qualcomm Technologies, Inc.

20 Developer Tips Hardware Scaling (1) Possibly the easiest (and overlooked) way to improve graphics performance Nonlinear relationship between resolution and number of pixels Fewer pixels rendered = improved graphics performance

21 Developer Tips Hardware Scaling (2) Hardware Scaling: Render at a resolution smaller than screen size Use hardware display system to do the scaling Strike a balance between lowering the resolution and acceptable quality Render scene at lower resolution, but HUD/Text at screen resolution Vulkan hardware scaling is trivial to do Allocate swapchain images at resolution smaller than surface size Surface size returned by vkgetphysicaldevicesurfacecapabilitieskhr Also returned by Android framework code. 1280x x x x2160 Display Surface Swapchain Images Base Resolution 2.25x more pixels 4x more pixels 9x more pixels

22 Developer Tips Memory Allocation (1) In Vulkan, your application needs to allocate and bind memory for various objects Different than in OpenGL where the driver did this behind the scenes Vulkan Objects requiring memory allocations - Textures and Buffers Vulkan memory has different properties Device Local memory that is physically connected to the device Host Visible memory that is visible to the host Host Coherent - memory that is automatically flushed to the host or device Host Cached memory that is cached on the host (memory accesses are slower than uncached) Lazily Allocated - memory that only allows device access to the memory Adreno Supports a single heap of memory with these types 1 - Device Local 2 - Device Local + Host Visible + Host Cached 3 - Device Local + Host Visible + Host Cached + Host Coherent

23 Developer Tips Memory Allocation (2) Vulkan requires you to explicit allocate memory blocks On Adreno, page alignment is large Allocate memory by calling vkallocatememory Recommend to minimize the # of these allocations Two strategies to doing this 1. Suballocation Allocate single memory allocation with vkallocatememory Suballocate from this block when memory is needed 2. Offsets Allocate single memory allocation with vkallocatememory Use offsets from this block when memory is needed Want to recycle memory allocations Instead of destroying and reallocating Using a custom memory allocator not recommended Doubtful of any performance improvement VkAllocateMemory VkAllocateMemory VkAllocateMemory VkAllocateMemory VkAllocateMemory Suballocate VkAllocateMemory Offsets Bad Better Best X

24 Developer Tips Validation Layers Robust Error Checking not included in API You can check return codes VkError not really recommended. Validation Layers should be used for error checking Optional can enable for debugging builds, disable for production builds Comprehensive checking: threading parameter checking object tracking image swapchain, Not perfect, but coverage and accuracy improving over time Implemented in Vulkan as extension layers Loaded by application at runtime Hook into Vulkan API processing Callback into your application when errors detected. Your app can log reports of the errors Application/Gam e Vulkan Loader Driver Adreno GPU Threading Parameter Validation Object Tracking Core Validation Image Swapchain Unique Objects

25 Adreno SDK for Vulkan Freely available from QDN website 16 samples + simple Vulkan framework Android Studio IDE (Windows only) Cover various topics: Initializing Vulkan Setting up swapchains Creating devices and queues Synchronizing the rendering steps Drawing a simple triangle Creating image processing effects Rendering in multiple passes with a renderpass Compiling SPIR-V shaders Using compute shaders Using push constants and more

26 Vulkan links Learning Vulkan is challenging, but growing number of resources on web Khronos website Up to date specification Reference pages Overview presentations Videos from recorded events Lunar G SDK (no Android support, but good documentation) Android NDK Guide on Vulkan Sample Libraries (beyond our Adreno SDK for Vulkan) Sascha Williams Norbert Nopper

27 THANK YOU Follow us on: For more information, visit us at: & Nothing in these materials is an offer to sell any of the components or devices referenced herein Qualcomm Technologies, Inc. and/or its affiliated companies. All Rights Reserved. Qualcomm Snapdragon and Qualcomm Adreno GPU are products of Qualcomm Technologies, Inc. Qualcomm, Snapdragon and Adreno are trademarks of Qualcomm Incorporated, registered in the United States and other countries, used with permission. Other produc ts and brand names may be trademarks or registered trademarks of their respective owners. References in this presentation to Qualcomm may mean Qualcomm Incorporated, Qualcomm Technologies, Inc., and/or other subsi diaries or business units within the Qualcomm corporate structure, as applicable. Qualcomm Incorporated includes Qualcomm s licensing business, QTL, and the vast majority of its patent portfolio. Qualcomm Technologies, Inc., a wholly-owned subsidiary of Qualcomm Incorporated, operates, along with its subsidiaries, substantially all of Qualcomm s engineering, research and development functions, and substantially all of its product and services businesses, inc luding its semiconductor business, QCT.