skia2/src/sksl

Overview
========

SkSL ("Skia Shading Language") is a variant of GLSL which is used as Skia's
internal shading language. SkSL is, at its heart, a single standardized version
of GLSL which avoids all of the various version and dialect differences found
in GLSL "in the wild", but it does bring a few of its own changes to the table.

Skia uses the SkSL compiler to convert SkSL code to GLSL, GLSL ES, or SPIR-V
before handing it over to the graphics driver.


Differences from GLSL
=====================

* Precision modifiers are not used. 'float', 'int', and 'uint' are always high
  precision. New types 'half', 'short', and 'ushort' are medium precision (we
  do not use low precision).
* Vector types are named <base type><columns>, so float2 instead of vec2 and
  bool4 instead of bvec4
* Matrix types are named <base type><columns>x<rows>, so float2x3 instead of
  mat2x3 and double4x4 instead of dmat4
* "@if" and "@switch" are static versions of if and switch. They behave exactly
  the same as if and switch in all respects other than it being a compile-time
  error to use a non-constant expression as a test.
* GLSL caps can be referenced via the syntax 'sk_Caps.<name>', e.g.
  sk_Caps.sampleVariablesSupport. The value will be a constant boolean or int,
  as appropriate. As SkSL supports constant folding and branch elimination, this
  means that an 'if' statement which statically queries a cap will collapse down
  to the chosen branch, meaning that:

    if (sk_Caps.externalTextureSupport)
        do_something();
    else
        do_something_else();

  will compile as if you had written either 'do_something();' or
  'do_something_else();', depending on whether that cap is enabled or not.
* no #version statement is required, and it will be ignored if present
* the output color is sk_FragColor (do not declare it)
* use sk_Position instead of gl_Position. sk_Position is in device coordinates
  rather than normalized coordinates.
* use sk_PointSize instead of gl_PointSize
* use sk_VertexID instead of gl_VertexID
* use sk_InstanceID instead of gl_InstanceID
* the fragment coordinate is sk_FragCoord, and is always relative to the upper
  left.
* use sk_Clockwise instead of gl_FrontFacing. This is always relative to an
  upper left origin.
* you do not need to include ".0" to make a number a float (meaning that
  "float2(x, y) * 4" is perfectly legal in SkSL, unlike GLSL where it would
  often have to be expressed "float2(x, y) * 4.0". There is no performance
  penalty for this, as the number is converted to a float at compile time)
* type suffixes on numbers (1.0f, 0xFFu) are both unnecessary and unsupported
* creating a smaller vector from a larger vector (e.g. float2(float3(1))) is
  intentionally disallowed, as it is just a wordier way of performing a swizzle.
  Use swizzles instead.
* Swizzle components, in addition to the normal rgba / xyzw components, can also
  be LTRB (meaning "left/top/right/bottom", for when we store rectangles in
  vectors), and may also be the constants '0' or '1' to produce a constant 0 or
  1 in that channel instead of selecting anything from the source vector.
  foo.rgb1 is equivalent to float4(foo.rgb, 1).
* All texture functions are named "sample", e.g. sample(sampler2D, float3) is
  equivalent to GLSL's textureProj(sampler2D, float3).
* Render target width and height are available via sk_Width and sk_Height
* some built-in functions and one or two rarely-used language features are not
  yet supported (sorry!)

SkSL is still under development, and is expected to diverge further from GLSL
over time.


SkSL Fragment Processors
========================

********************************************************************************
*** IMPORTANT: You must set gn arg "skia_compile_processors = true" to cause ***
*** .fp files to be recompiled! In order for compilation to succeed, you     ***
*** must run bin/fetch-clang-format (once) to install our blessed version.   ***
********************************************************************************

An extension of SkSL allows for the creation of fragment processors in pure
SkSL. The program defines its inputs similarly to a normal SkSL program (with
'in' and 'uniform' variables), but the 'main()' function represents only this
fragment processor's portion of the overall fragment shader.

Within an '.fp' fragment processor file:

* C++ code can be embedded in sections of the form:

  @section_name { <arbitrary C++ code> }

  Supported section are:
    @header            (in the .h file, outside the class declaration)
    @headerEnd         (at the end of the .h file)
    @class             (in the .h file, inside the class declaration)
    @cpp               (in the .cpp file)
    @cppEnd            (at the end of the .cpp file)
    @constructorParams (extra parameters to the constructor, comma-separated)
    @constructor       (replaces the default constructor)
    @initializers      (constructor initializer list, comma-separated)
    @emitCode          (extra code for the emitCode function)
    @fields            (extra private fields, each terminated with a semicolon)
    @make              (replaces the default Make function)
    @clone             (replaces the default clone() function)
    @setData(<pdman>)  (extra code for the setData function, where <pdman> is
                        the name of the GrGLSLProgramDataManager)
    @test(<testData>)  (the body of the TestCreate function, where <testData> is
                        the name of the GrProcessorTestData* parameter)
    @coordTransform(<sampler>)
                       (the matrix to attach to the named sampler2D's
                        GrCoordTransform)
    @samplerParams(<sampler>)
                       (the sampler params to attach to the named sampler2D)
* global 'in' variables represent data passed to the fragment processor at
  construction time. These variables become constructor parameters and are
  stored in fragment processor fields. By default float2/half2 maps to SkPoints,
  and float4/half4 maps to SkRects (in x, y, width, height) order. Similarly,
  int2/short2 maps to SkIPoint and int4/half4 maps to SkIRect. Use ctype
  (below) to override this default mapping.
* global variables support an additional 'ctype' layout key, providing the type
  they should be represented as from within the C++ code. For instance, you can
  use 'layout(ctype=SkPMColor4f) in half4 color;' to create a variable that looks
  like a half4 on the SkSL side of things, and a SkPMColor4f on the C++ side of
  things.
* 'uniform' variables become, as one would expect, top-level uniforms. By
  default they do not have any data provided to them; you will need to provide
  them with data via the @setData section.
* 'in uniform' variables are uniforms that are automatically wired up to
  fragment processor constructor parameters. The fragment processor will accept
  a parameter representing the uniform's value, and automatically plumb it
  through to the uniform's value in its generated setData() function.
* 'in uniform' variables support a 'tracked' flag in the layout that will
  have the generated code automatically implement state tracking on the uniform
  value to minimize GPU calls.
* the 'sk_TransformedCoords2D' array provides access to 2D transformed
  coordinates. sk_TransformedCoords2D[0] is equivalent to calling
  fragBuilder->ensureCoords2D(args.fTransformedCoords[0]) (and the result is
  cached, so you need not worry about using the value repeatedly).
* Uniform variables support an additional 'when' layout key.
  'layout(when=foo) uniform int x;' means that this uniform will only be
  emitted when the 'foo' expression is true.
* 'in' variables support an additional 'key' layout key.
  'layout(key) in uniform int x;' means that this uniform should be included in
  the program's key. Matrix variables additionally support 'key=identity',
  which causes the key to consider only whether or not the matrix is an
  identity matrix.
* child processors can be declared with 'in fragmentProcessor <name>;', and can
  be invoked by calling 'sample(<name>)' or 'sample(<name>, <inputColor>)'.
  The first variant emits the child with a solid white input color. The second
  variant emits the child with the result of the 2nd argument's expression,
  which must evaluate to a half4. The process function returns a half4.
* By default, fragment processors must be non-null. The type for a nullable
  fragment processor is 'fragmentProcessor?', as in
  'in fragmentProcessor? <name>'. You can check for the presence of such a
  fragment processor by comparing it to 'null'.


Creating a new .fp file
=======================

1. Ensure that you have set gn arg "skia_compile_processors = true"
2. Create your new .fp file, generally under src/gpu/effects.
3. Add the .fp file to sksl.gni.
4. Build Skia. This will cause the .fp file to be compiled, resulting in a new
   .cpp and .h file for the fragment processor.
5. Add the .cpp and .h files to gpu.gni.
6. Add the new processor's ClassID (k<ProcessorName>_ClassID) to
   GrProcessor::ClassID.
7. At this point you can reference the new fragment processor from within Skia.

Once you have done this initial setup, simply re-build Skia to pick up any
changes to the .fp file.