[Mesa-dev] [PATCH 0/6] glsl: Add support for varying structs.

[Jordan Justen]

Reviewed-by: Jordan Justen <jordan.l.justen at intel.com>

On Mon, Jan 21, 2013 at 2:16 PM, Paul Berry <stereotype441 at gmail.com> wrote:
> This patch series adds support for varying structs, which are a
> required part of GLSL ES 3.00 and GLSL 1.50.
>
> I can see two principal ways to implement this feature: a "flattening"
> approach, and a "packing" approach.  In the "flattening" approach, the
> linker replaces each varying struct with a set of varyings, one for
> each field in the struct, and adjusts the shader code to refer to the
> new set of varyings (this is what a programmer would likely try to do
> when manually porting a shader that uses structs to a version of GLSL
> that didn't support them).  In the "packing" approach, we keep the
> varying structs as single variables, and lay out their contents
> contiguously in GPU memory (in much the same way that a CPU compiler
> would handle an ordinary C struct).
>
> In principle, the "flattening" approach carries some potential
> performance benefits, since it frees the linker to assign each struct
> element a location that is properly aligned to its type.  However, it
> is much more difficult to implement, since it requires a lowering pass
> to traverse the entire shader and adjust all references to the varying
> structs, including function in/out parameters and whole structure
> assignment.  To make matters worse, there would be subtle interactions
> between the linker and the lowering pass, since error checking would
> have to be performed based on the un-lowered shaders, but the location
> assignment would have to be applied to the lowered varyings.
>
> Because of these difficulties, I've used the "packing" approach in
> this initial implementation--this allowed me to take advantage of the
> existing lower_packed_varyings() function to pack and unpack varying
> structs, so the code delta is fairly small.  We can always switch to
> the "flattening" approach if, in the future, we discover some programs
> whose performance would be dramatically improved by it.
>
> Patch 1/6 eliminates the ambiguity in the ir_variable_mode enum
> between "in/out" variables that are inputs/outputs of functions and
> inputs/outputs of shaders.  This makes possible patch 2/6, which
> modifies opt_structure_splitting to leave varying structs alone (we
> can't have it splitting varying structs, otherwise the linker won't be
> able to find them).  Patches 3/6 and 4/6 update the linker and
> lower_packed_varyings, respectively, to handle varying structs, and
> patch 5/6 disables transform feedback for varying structs (we haven't
> yet heard from Khronos how this is supposed to work).  Finally, patch
> 6/6 enables the feature when the GLSL version is 1.50 or 3.00 ES.
>
>
> [PATCH 1/6] glsl: Eliminate ambiguity between function ins/outs and shader ins/outs
> [PATCH 2/6] glsl: Disable structure splitting for shader ins/outs.
> [PATCH 3/6] glsl: Generalize compute_packing_order for varying structs.
> [PATCH 4/6] glsl: Update lower_packed_varyings to handle varying structs.
> [PATCH 5/6] glsl: Disable transform feedback of varying structs.
> [PATCH 6/6] glsl: Allow varying structs in GLSL ES 3.00 and GLSL 1.50.