(originally part of previous patch, split out to separate patch by Rob)
v2: squash in some fixes from Eric
v3: Another fix from Eric for point coords.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
This avoids generation of undefined packing in qir and qpu instructions,
fixing a lot of rendering errors.
Fixes 8b36d107fd (vc4: Pack the unorm-packing bits into a src MUL
instruction when possible.)
Cc: mesa-stable@lists.freedesktop.org
Signed-off-by: Boyan Ding <boyan.j.ding@gmail.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Emil Velikov <emil.l.velikov@gmail.com>
Now that we do non-SSA QIR instructions, we can take a NIR SSA src that's
only used by the unorm packing and just stuff the pack bits into it.
total instructions in shared programs: 98136 -> 97974 (-0.17%)
instructions in affected programs: 4149 -> 3987 (-3.90%)
This helps ensure that the register allocator doesn't force the later pack
operations to insert extra MOVs.
total instructions in shared programs: 98170 -> 98159 (-0.01%)
instructions in affected programs: 2134 -> 2123 (-0.52%)
Now that we have NIR, most of the optimization we still need to do is
peepholes on instruction selection rather than general dataflow
operations. This means we want to be able to have QIR be a lot closer to
the actual QPU instructions, just with virtual registers. Allowing
multiple instructions writing the same register opens up a lot of
possibilities.
Previously, SFU values always moved to a temporary, and TLB color reads
and texture reads always lived in r4. Instead, we can have these results
just be normal temporaries, and the register allocator can leave the
values in r4 when they don't interfere with anything else using r4.
shader-db results:
total instructions in shared programs: 100809 -> 100040 (-0.76%)
instructions in affected programs: 42383 -> 41614 (-1.81%)
NIR brings us better optimization than I would have bothered to write
within the driver, developers sharing future optimization work, and the
ability to share device-specific lowering code that we and other
GLES2-level drivers need.
total uniforms in shared programs: 13421 -> 13422 (0.01%)
uniforms in affected programs: 62 -> 63 (1.61%)
total instructions in shared programs: 39961 -> 39707 (-0.64%)
instructions in affected programs: 15494 -> 15240 (-1.64%)
v2: Add missing imov support, and assert that there are no dest saturates.
v3: Rebase on the target-specific algebraic series.
v4: Rebase on gallium-includes-from-NIR changes in mater.
v5: Rebase on variables being in lists instead of hash tables.
v6: Squash in intermediate changes that used the NIR-to-TGSI pass (which
I'm not committing)
This cleans up some pointless operations generated by the in-driver mul24
lowering (commonly generated by making a vec4 index for a matrix in a
uniform array).
I could fill in other operations, but pretty much anything else ought to
be getting handled at the NIR level, I think.
total uniforms in shared programs: 13423 -> 13421 (-0.01%)
uniforms in affected programs: 346 -> 344 (-0.58%)
This lets us more intelligently decide which uniform values should be put
into temporaries, by choosing the most reused values to push to temps
first.
total uniforms in shared programs: 13457 -> 13433 (-0.18%)
uniforms in affected programs: 1524 -> 1500 (-1.57%)
total instructions in shared programs: 40198 -> 40019 (-0.45%)
instructions in affected programs: 6027 -> 5848 (-2.97%)
I noticed this opportunity because with the NIR work, some programs were
happening to make different uniform copy propagation choices that
significantly increased instruction counts.
Right now the places that used to emit a mov.sf just put the SF on the
previous instruction when it generated the source of the SF value. Even
without optimization to push the sf up further (and kill thus potentially
kill more MOVs), this gets us:
total uniforms in shared programs: 13455 -> 13457 (0.01%)
uniforms in affected programs: 3 -> 5 (66.67%)
total instructions in shared programs: 40296 -> 40198 (-0.24%)
instructions in affected programs: 12595 -> 12497 (-0.78%)
I'm using this in some WIP commits for doing blending in 8888 instead of
vec4. But it also gives us these results immediately, thanks to allowing
more uniforms/immediates in the arguments:
total instructions in shared programs: 41027 -> 40960 (-0.16%)
instructions in affected programs: 4381 -> 4314 (-1.53%)
Small immediates have the downside of taking over the raddr B field, so
you might have less chance to pack instructions together thanks to raddr B
conflicts. However, it also reduces some register pressure since it lets
you load 2 "uniform" values in one instruction (avoiding a previous load
of the constant value to a register), and increases some pairing for the
same reason.
total uniforms in shared programs: 16231 -> 13374 (-17.60%)
uniforms in affected programs: 10280 -> 7423 (-27.79%)
total instructions in shared programs: 40795 -> 41168 (0.91%)
instructions in affected programs: 25551 -> 25924 (1.46%)
In a previous version of this patch I had a reduction in instruction count
by forcing the other args alongside a SMALL_IMM to be in the A file or
accumulators, but that increases register pressure and had a bug in
handling FRAG_Z. In this patch is I just use raddr conflict resolution,
which is more expensive. I think I'd rather tweak allocation to have some
way to slightly prefer good choices for files in general, rather than risk
failing to register allocate by forcing things into register classes.
While depth test state is passed through the fragment shader as sideband,
data, the stencil test state has to be set by the fragment shader itself.
Many tests are still failing, but this gets most of hiz/ passing.
The goal here is to have an argument for the depth write opcode so that I
can do computed depth. In the process, this makes the calculations that
will be emitted more obvious in the QIR.
We potentially need to be careful that use of a value stored in r4 isn't
copy-propagated (or something) across another r4 write. That doesn't
appear to happen currently, and this makes the dataflow more obvious. It
also opens up not unpacking the r4 value, which will be useful for depth
textures.
The split between these two didn't make much sense. I'm going to want the
chance to look at uniform contents in optimization passes, and the QPU
emit I think is going to end up rewriting the uniforms stream.