After generating VIR, we leave c->cursor pointing at the end of the
shader. If the shader had dead code at the end (for example from preamble
instructions in a shader with no side effects), we would assertion fail
that we were leaving the cursor pointing at freed memory. Since anything
following DCE should be setting up a new cursor anyway, just clear the
cursor at the start.
The thrsw will invalidate rtop, just like accumulators and flags. Caught
by simulator assertions in CS imulextended/umulextended tests.
Fixes: 90269ba353 ("broadcom/vc5: Use THRSW to enable multi-threaded shaders.")
The HW apparently has some issues (or at least a much more complicated VCM
calculation) with non-combined segments, and the closed source driver also
uses combined I/O. Until I get the last CTS failure resolved (which does
look plausibly like some VPM stomping), let's use combined I/O too.
Instead of a single i2b and b2i, we now have i2b32 and b2iN where N is
one if 8, 16, 32, or 64. This leads to having a few more opcodes but
now everything is consistent and booleans aren't a weird special case
anymore.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
The way nir_lower_clip_vs() works with store_output intrinsics makes a
ton of assumptions about the driver_location field.
In i965 and iris, I'd rather do this lowering early and work with
variables. v3d may want to switch to that as well, and ir3 could too,
but I'm not sure exactly what would need updating. For now, handle
both methods.
Reviewed-by: Eric Anholt <eric@anholt.net>
We were doing this late after nir_lower_io, but we can just reuse the core
code. By doing it at this stage, we won't even set up the VS attributes
as inputs, reducing our VPM size.
We always emit 4 slots per slot because things like color output and
position processing in the epilogue will potentially look up more values
than the variable declaration had. However, when we get a .location_frac
!= 0, we don't want to overwrite components of the following
.driver_location.
For supporting scalar VPM i/o at the NIR level, we need to do a pass over
the vars to figure out how big each attribute is after DCE. Once we've
done that, we can just walk over c->vattr_sizes[] instead of bothering
with vars.
There were two bugs working together to make things mostly work: I wasn't
dividing the VPM output size available by the size of a batch (vertex),
but I also had the size of the VPM reduced by a factor of 8.
Fixes dEQP-GLES3.functional.vertex_array_objects.all_attributes and it
seems also my intermittent varying failures.
Fixes: 1561e4984e ("v3d: Emit the VCM_CACHE_SIZE packet.")
Found when debugging register spilling -- we would try to spill the dest
of a STVPMV, inserting spill code after entering the last segment. In
fact, we were likely to to choose to do this, given that the STVPMV "dest"
temp was never read from, making it cheap to spill.
Cc: "18.2" <mesa-stable@lists.freedesktop.org>
The simulator complained that we had write responses outstanding at shader
end. It seems that a TMU read does not guarantee that previous TMU writes
by the thread have completed, which surprised me.
Cc: "18.2" <mesa-stable@lists.freedesktop.org>
This is useful for periodically testing out register spilling to see how
it goes on simple shaders, rather than only failing on insanely
complicated ones.
This instruction is used to ensure that TMU stores have been processed
before moving on. In particular, you need any TMU ops to be done by the
time the shader ends.
This is controlled by a new nir_shader_compiler_options flag, and fixes
dEQP-GLES3.functional.shaders.builtin_variable.pointcoord on V3D.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
We can do one per instruction, and we have to be careful not to overwrite
raddr_b, but this greatly reduces the pressure on uniform loads
(particularly around ldvpm/stvpm instructions).
total instructions in shared programs: 90768 -> 88220 (-2.81%)
instructions in affected programs: 82711 -> 80163 (-3.08%)
Sometimes when iterating over sources, we might want to check if it's the
implicit one. We wouldn't want to match on a non-implicit src using this
function.
These instructions let us write directly to the phys regfile, instead of
just R4. That lets us avoid moving out of R4 to avoid conflicting with
other SFU results, and to avoid conflicting with thread switches.
There is still an extra instruction of latency, which is not represented
in the scheduler at the moment. If you use the result before it's ready,
the QPU will just stall, unlike the magic R4 mode where you'd read the
previous value. That means that the following shader-db results aren't
quite representative (since we now cause some stalls instead of emitting
nops), but they're impressive enough that I'm happy with the change.
total instructions in shared programs: 95669 -> 91275 (-4.59%)
instructions in affected programs: 82590 -> 78196 (-5.32%)
Similarly to VC4's implementation, by not picking r0 immediately upon
freeing it, we give the scheduler more of a chance to fit later writes in
earlier. I'm not clear on whether there's any real cost to picking phys
over accumulators, so keep that behavior for now.
shader-db:
total instructions in shared programs: 96831 -> 95669 (-1.20%)
instructions in affected programs: 77254 -> 76092 (-1.50%)
This restriction existed in V3D 2.x, but lifting it was a major change in
3.x.
shader-db results:
total instructions in shared programs: 98117 -> 96831 (-1.31%)
instructions in affected programs: 48520 -> 47234 (-2.65%)
total instructions in shared programs: 98578 -> 98119 (-0.47%)
instructions in affected programs: 27571 -> 27112 (-1.66%)
and it also eliminates most spills/fills on the CTS's randomized uniform
usage testcases.
There's a convenient "FTOC" instruction for generating the coverage now,
unlike vc4. This fixes
dEQP-GLES3.functional.multisample.fbo_4_samples.proportionality_alpha_to_coverage
