[go: up one dir, main page]

Translator: Implement an IR

The ANGLE compiler parses GLSL ES shaders and produces an Abstract
Syntax Tree (AST).  The use of AST is deceptively simple, as the
structure of this representation always matches the equivalent text
representation.  This simplicity can make the compiler initially more
approachable.

However, as compiler research, modern compiler best practices, and our
own experiences with the ANGLE compiler have shown, ASTs are harder to
analyze and transform. For example:

* Struct definitions may be in a number of surprising places in GLSL,
  such as a function’s return type. Reflecting this fact in the AST
  makes it tricky to apply transformations to the struct.

* The inability to include arbitrary blocks of code in many locations
  such as inside an `if` condition turns simple transformations (such as
  caching an expression with a side effect inside a temporary variable)
  into complex ones with many corner cases.

Instead, modern compilers include an Internal Representation (IR) that
is potentially far removed from the textual syntax of the language.
This comes with numerous benefits, such as the freedom to simplify and
rearrange type definitions, the flexibility of instruction flow
structure, etc, which in turn make more complex operations (such as
dead-code elimination) less burdensome.  In some projects, IRs
also allow the compiler to handle multiple textual formats (otherwise
known as front-ends).

While generating an IR from text is typically not complicated, one
possibly challenging aspect of IRs is going back to a textual
representation if needed (such as generating glsl, msl or hlsl in the
case of ANGLE), as the textual representation includes restrictions that
do not exist in the IR.

Given the number of bugs (security or otherwise) that constantly affect
ANGLE, there is a need for an IR to finally modernize this codebase.
The new WGSL backend for the translator is plagued with the same
complications that the SPIR-V generator has faced (such as lack of
ability to write to a swizzle expression), with extremely complicated
or in some cases no clear way of implementing them in the AST, so an IR
would simplify/unblock that work.
Finally, some needed optimizations (such as cross-stage dead-code
elimination) are needed for the ANGLE-for-Android project, but are
extremely difficult to do with the AST.

It’s time the translator was written to be based on an IR.  ANGLE
implements its own IR instead of reusing existing ones, such as SPIR-V,
mesa's NIR, or Tint's IR, because:

* It can be simpler e.g. compared to SPIR-V's highly flexible branching
* It can be tailored to ESSL e.g. compared to Tint's IR
* It can be in complete control e.g. compared to mesa's NIR

This change implements the following:

* GLSL->IR generation during parse
* IR transformation utilities, debug dump, documentation (see
  src/compiler/translator/ir/doc/README.md)
* An "astify" transformation that makes the IR suitable for AST and
  text-based output generation
* IR->AST generation after parse

The current AST generation during parse is still maintained because:

* It's used by validation, but that will ultimately be made unnecessary
  after the following point is no longer needed.
* It's simpler to keep the AST generation while the no-IR path is still
  supported.

With the rountrip through IR, a dozen transformations that try to make
the AST more malleable are made unnecessary.  However, at this point
when the IR is enabled, the overhead of creating it is present, while
the transformations are still complex and based on the AST.  The IR is
disabled until almost all transformations are ported and at least SPIR-V
generation is done directly from the IR.

To enable the IR, set `angle_ir = true` in `args.gn`.  Nearly all tests
pass with the IR enabled, though a few failures remain to be diagnosed.
One particular follow up fix is to port our compiler unittests to
end2end tests:

* They assume too much about the internals of the compiler, which is
  completely overhauled now
* The validation is too simplistic, where often transformation bugs are
  discovered only after trying to use the shader in a draw call (where
  multiple backends actually give the translator output to the native
  driver/compiler)

Bug: angleproject:349994211
Change-Id: I2ce917bf956d40af4545473b4a9068cf295878be
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/5898761
Reviewed-by: Geoff Lang <geofflang@chromium.org>
Commit-Queue: Shahbaz Youssefi <syoussefi@chromium.org>
Reviewed-by: Matthew Denton <mpdenton@chromium.org>
53 files changed
tree: 68fee60c08945a43cc8bf974915390f46b5afb31
  1. android/
  2. build_overrides/
  3. doc/
  4. extensions/
  5. gni/
  6. include/
  7. infra/
  8. samples/
  9. scripts/
  10. src/
  11. third_party/
  12. tools/
  13. util/
  14. .clang-format
  15. .clang-format-ignore
  16. .git-blame-ignore-revs
  17. .gitattributes
  18. .gitignore
  19. .gitmodules
  20. .gn
  21. .rustfmt.toml
  22. .style.yapf
  23. .vpython3
  24. .yapfignore
  25. additional_readme_paths.json
  26. Android.mk
  27. AUTHORS
  28. BUILD.gn
  29. codereview.settings
  30. CONTRIBUTORS
  31. DEPS
  32. DIR_METADATA
  33. dotfile_settings.gni
  34. LICENSE
  35. OWNERS
  36. PRESUBMIT.py
  37. README.chromium
  38. README.md
  39. unsafe_buffers_paths.txt
  40. WATCHLISTS
README.md

ANGLE - Almost Native Graphics Layer Engine

The goal of ANGLE is to allow users of multiple operating systems to seamlessly run WebGL and other OpenGL ES content by translating OpenGL ES API calls to one of the hardware-supported APIs available for that platform. ANGLE currently provides translation from OpenGL ES 2.0, 3.0 and 3.1 to Vulkan, desktop OpenGL, OpenGL ES, Direct3D 9, and Direct3D 11. Future plans include ES 3.2, translation to Metal and MacOS, Chrome OS, and Fuchsia support.

Level of OpenGL ES support via backing renderers

Direct3D 9Direct3D 11Desktop GLGL ESVulkanMetal
OpenGL ES 2.0completecompletecompletecompletecompletecomplete
OpenGL ES 3.0completecompletecompletecompletecomplete
OpenGL ES 3.1incompletecompletecompletecomplete
OpenGL ES 3.2in progressin progresscomplete

Additionally, OpenGL ES 1.1 is implemented in the front-end using OpenGL ES 3.0 features. This version of the specification is thus supported on all platforms specified above that support OpenGL ES 3.0 with known issues.

Platform support via backing renderers

Direct3D 9Direct3D 11Desktop GLGL ESVulkanMetalWebGPU
Windowscompletecompletecompletecompletecomplete
Linuxcompletecomplete
Mac OS Xcompletecomplete [1]
iOScomplete [2]
Chrome OScompleteplanned
Androidcompletecomplete
Fuchsiacomplete

[1] Metal is supported on macOS 10.14+

[2] Metal is supported on iOS 12+

ANGLE v1.0.772 was certified compliant by passing the OpenGL ES 2.0.3 conformance tests in October 2011.

ANGLE has received the following certifications with the Vulkan backend:

  • OpenGL ES 2.0: ANGLE 2.1.0.d46e2fb1e341 (Nov, 2019)
  • OpenGL ES 3.0: ANGLE 2.1.0.f18ff947360d (Feb, 2020)
  • OpenGL ES 3.1: ANGLE 2.1.0.f5dace0f1e57 (Jul, 2020)
  • OpenGL ES 3.2: ANGLE 2.1.2.21688.59f158c1695f (Sept, 2023)

ANGLE also provides an implementation of the EGL 1.5 specification.

ANGLE is used as the default WebGL backend for both Google Chrome and Mozilla Firefox on Windows platforms. Chrome uses ANGLE for all graphics rendering on Windows, including the accelerated Canvas2D implementation and the Native Client sandbox environment.

Portions of the ANGLE shader compiler are used as a shader validator and translator by WebGL implementations across multiple platforms. It is used on Mac OS X, Linux, and in mobile variants of the browsers. Having one shader validator helps to ensure that a consistent set of GLSL ES shaders are accepted across browsers and platforms. The shader translator can be used to translate shaders to other shading languages, and to optionally apply shader modifications to work around bugs or quirks in the native graphics drivers. The translator targets Desktop GLSL, Vulkan GLSL, Direct3D HLSL, and even ESSL for native GLES2 platforms.

OpenCL Implementation

In addition to OpenGL ES, ANGLE also provides an optional OpenCL runtime built into the same output GLES lib.

This work/effort is currently work-in-progress/experimental.

This work provides the same benefits as the OpenGL implementation, having OpenCL APIs be translated to other HW-supported APIs available on that platform.

Level of OpenCL support via backing renderers

VulkanOpenCL
OpenCL 1.0in progressin progress
OpenCL 1.1in progressin progress
OpenCL 1.2in progressin progress
OpenCL 3.0in progressin progress

Each supported backing renderer above ends up being an OpenCL Platform for the user to choose from.

The OpenCL backend is a “passthrough” implementation which does not perform any API translation at all, instead forwarding API calls to other OpenCL driver(s)/implementation(s).

OpenCL also has an online compiler component to it that is used to compile OpenCL C source code at runtime (similarly to GLES and GLSL). Depending on the chosen backend(s), compiler implementations may vary. Below is a list of renderers and what OpenCL C compiler implementation is used for each:

  • Vulkan : clspv
  • OpenCL : Compiler is part of the native driver

Sources

ANGLE repository is hosted by Chromium project and can be browsed online or cloned with

git clone https://chromium.googlesource.com/angle/angle

Building

View the Dev setup instructions.

Contributing