Skip to content

Latest commit

 

History

History

impeller

--------------------------------------------------------------------------------
--------------------------------------------------------------------------------
-----------------   ___                      _ _            --------------------
-----------------  |_ _|_ __ ___  _ __   ___| | | ___ _ __  --------------------
-----------------   | || '_ ` _ \| '_ \ / _ \ | |/ _ \ '__| --------------------
-----------------   | || | | | | | |_) |  __/ | |  __/ |    --------------------
-----------------  |___|_| |_| |_| .__/ \___|_|_|\___|_|    --------------------
-----------------                |_|                        --------------------
--------------------------------------------------------------------------------
--------------------------------------------------------------------------------

⚠️ Impeller is a Prototype and Work-In-Progress. Proceed with caution. ⚠️

Impeller is a rendering runtime for Flutter with the following objectives:

  • Predictable Performance: All shader compilation and reflection is performed offline at build time. All pipeline state objects are built upfront. Caching is explicit and under the control of the engine.
  • Instrumentable: All graphics resources (textures, buffers, pipeline state objects, etc..) are tagged and labeled. Animations can be captured and persisted to disk without affecting per-frame rendering performance.
  • Portable: Not tied to a specific client rendering API. Shaders are authored once and converted as necessary.
  • Uses Modern Graphics APIs Effectively: Makes heavy use of (but doesn’t depend on) features available in Modern APIs like Metal and Vulkan.
  • Makes Effective Use of Concurrency: Can distribute single-frame workloads across multiple threads if necessary.

Project Organization

Impeller is a meta-framework. While a user of Impeller may choose to include the whole enchilada (in //impeller/:impeller), the various sub-frameworks have clearly defined responsibilities and adhere to a strict hierarchy.

Impeller itself may not depend on anything in //flutter except //flutter/fml and flutter/display_list. FML is a base library for C++ projects and Impeller implements the display list dispatcher interface to make it easy for Flutter to swap the renderer with Impeller. Impeller is meant to be used by the Flow (//flutter/flow) subsystem. Hence the name. The tessellator and geometry libraries are exceptions - they unconditionally may not depend on anything from //flutter.

An overview of the major sub-frameworks, their responsibilities, and, relative states of completion:

  • //impeller/compiler: The offline shader compiler. Takes GLSL 4.60 shader source code and converts it into a backend specific shader representation (like Metal Shading Language). It also generates C++ bindings that callers can include as a GN source_sets so there is no runtime shader reflection either. The target is an executable called impellerc which is never shipped into the binary or as an artifact.
  • //impeller/renderer: The very lowest level of the renderer that is still backend agnostic. Allows users to build a renderer from scratch with few restrictions. Has utilities for creating allocators, generating pipeline state objects from bindings generated by //impeller/compiler, setting up render passes, managing jumbo uniform-buffers, tessellators, etc..
    • //impeller/renderer/backend: Contains all the implementation details for a specific client rendering API. The interfaces in these targets are meant to be private for non-WSI user targets. No Impeller sub-frameworks may depend on these targets.
  • //impeller/archivist: Allows persisting objects to disk as performantly as possible (usually on a background thread). The framework is meant to be used for storing frame meta-data and related profiling/instrumentation information. Collection of information should succeed despite process crashes and retrieval of traces must not use inordinate amounts of time or memory (which usually leads to crashes).
  • //impeller/geometry: All (or, most of) the math! This C++ mathematics library is used extensively by Impeller and its clients. The reasonably interesting bit about this library is that all types can be used interchangeably in device and host memory. Various Impeller subsystems understand these types and can take care of packing and alignment concerns w.r.t these types.
  • //impeller/playground: When working with graphics APIs, it is often necessary to visually verify rendering results as a specific feature is being worked upon. Moreover, it is useful to attach frame debuggers or profilers to specific test cases. The playground framework provides Google Test fixtures that open the current state of various rendering related objects in a window in which rendering results can be visualized, or, to which frame debuggers can be attached. Most Impeller sub-frameworks that have a test harness also have a custom playground subclass. This sub-framework is only meant to provide utilities for tests and will not be compiled into any shipping binary.
  • //impeller/entity: Sits one level above //impeller/renderer and provides a framework for building 2D renderers. Most of the pipeline state objects generated from shaders authored at build time reside in this framework. The render-pass optimization and pass-rewriting framework also resides there. This allows authoring composable 2D rendering optimizations (like collapsing passes, or, eliding them completely).
  • //impeller/aiks: Aiks wraps //impeller/entity into an API that resembles Skia. This makes it easy to mechanically replace Skia calls with their Impeller counterparts even though the //impeller/entity framework API is different from Skia. This presence of this sub-framework is probably short-lived as integration of Impeller into Flutter should likely happen via a custom Display List implementation in //impeller/display_list. The roadblocks to this today are graphics package agnosticism in the Display List interface.
  • //impeller/display_list: The replacement for //impeller/aiks to serve in the integration of Impeller in //flutter/flow. This is pending graphics package agnosticism in the Impeller interface. This sub-framework primarily provides a custom implementation of the flutter::DisplayListDispatcher that forwards Flutter rendering intent to Impeller.
  • //impeller/base: Contains C++ utilities that are used throughout the Impeller family of frameworks. Ideally, these should go in //flutter/fml but their use is probably not widespread enough to at this time.
  • //impeller/image: The Impeller renderer works with textures whose memory is resident in device memory. However, pending the migration of //flutter/display_list to graphics package agnosticism and the subsequent migration of the image decoders to work with the package agnostic types, there needs to be a way for tests and such to decode compressed image data. This sub-framework provides that functionality. This sub-framework is slated for removal and must not be used outside of tests.
  • //fixtures: Contains test fixtures used by the various test harnesses. This depends on //flutter/testing.
  • //tools: Contains all GN rules and python scripts for working with Impeller. These include GN rules processing GLSL shaders, including reflected shader information as source set targets, and, including compiled shader intermediate representations into the final executable as binary blobs for easier packaging.

The Offline Shader Compilation Pipeline

  • Shaders are authored once in GLSL 4.60. This choice of shading language is consistent across all backends. Shader code resides in the Impeller source tree like any other source file.
  • At build time, the Impeller Shader Compiler (impellerc) converts the GLSL into SPIRV. No optimizations are performed on the generated SPIRV at this stage. This is to preserve all debugging and instrumentation information.
  • Using the SPIRV, a backend specific transpiler converts the SPIRV to the appropriate high-level shading language. This is controlled using flags to the impellerc.
  • All the files generated in the high-level shading language are compiled, optimized, and linked into a single binary blob.
  • The binary blob containing the compiled and optimized high-level shading language is included as a hex dump (see xxd.py) into a C source file with a generated GN target. Executable targets that want to include the compiled code in their binaries just need to depend on the generated GN target. This eases any shader packaging concerns.
  • In parallel, the SPIRV is processed by a reflector. This produces C++ translation units that allow for the easy creation of pipeline state objects at runtime. The headers for these translation units include any structs (with appropriate padding and alignment) such that uniform data as well as vertex information can be specified to the shader without having to deal with bindings, vertex descriptors, etc.. This also makes iterating on shaders easier as changes to the shader interface lead to compile time errors.
  • The C++ translation units generated from reflected shader information are made available to callers as a generated GN target that callers may use if necessary. It is possible for callers to perform reflection at runtime but there are no Impeller components that do this currently.

Shader Compilation Pipeline