[Mesa-dev] [PATCH 08/21] etnaviv: nir: add virtual register classes

[Philipp Zabel]

On Tue, 2018-06-05 at 11:39 -0400, Rob Clark wrote:
> On Tue, Jun 5, 2018 at 10:38 AM, Philipp Zabel <p.zabel at pengutronix.de> wrote:
> > Since all threads share a global temporary vec4 register file, it is
> > important to reduce temporary register use of shaders.
> > Using source swizzles and destination write mask of ALU operations we
> > can layer smaller virtual registers on top of the physical base
> > registers that overlap with their base register and partially with each
> > other:
> > 
> >  +----+---------+-------------+---------+
> >  |VEC4|  VEC3   |    VEC2     | SCALAR  |
> >  +----+---------+-------------+---------+
> >  |  X | X X X   | X X X       | X       |
> >  |  Y | Y Y   Y | Y     Y Y   |   Y     |
> >  |  Z | Z   Z Z |   Z   Z   Z |     Z   |
> >  |  W |   W W W |     W   W W |       W |
> >  +----+---------+-------------+---------+
> > 
> > There are four possible virtual vec3 registers that leave the remaining
> > component usable as a scalar virtual register, six possible vec2
> > registers, and four possible scalar registers that only use a single
> > component.
> > 
> > This patch adds an interference graph for virtual registers to the
> > register allocator, using information about SSA interference and virtual
> > register overlap. If possible, SSAs with smaller num_components are
> > allocated from the unused components of already partially used temporary
> > registers.
> > 
> > Signed-off-by: Philipp Zabel <p.zabel at pengutronix.de>
> > Signed-off-by: Michael Tretter <m.tretter at pengutronix.de>
> > ---
> 
> so one quick note, constructing the register classes can be
> expensive.. you probably only want to do this once and then re-use for
> each shader

Thank you. Yes, that should be easily fixable.

The main reason we haven't already done this is that I hadn't added the
dummy RA node that grabs the forbidden register 0 in fragment shaders
until the last minute. Previously we had left out the register from the
graph altogether.

regards
Philipp