| Commit message (Collapse) | Author | Age | Files | Lines |
|
|
|
|
|
|
|
| |
I had this exactly backwards, but apparently the piglit tests were all
landing in r0-r3 anyway.
Cc: "13.0" <mesa-stable@lists.freedesktop.org>
(cherry picked from commit 977d8b526b983c8d19df00af224033389f8ab7c8)
|
|
|
|
|
|
|
|
|
|
| |
It's much better to just skip the draw call entirely. Getting this
information out of register allocation will also be useful for
implementing threaded fragment shaders, which will need to retry
non-threaded if RA fails.
Cc: <mesa-stable@lists.freedesktop.org>
(cherry picked from commit 4d019bd703e7c20d56d5b858577607115b4926a3)
|
|
|
|
|
|
|
| |
We need the source to be in r0-r3, so make a new register class for it.
It will be up to the surrounding passes to make sure that the r0-r3
allocation of its source won't conflict with anything other class
requirements on that temp.
|
|
|
|
|
| |
Right now our CFG is always a trivial single basic block, but that will
change when enable loops.
|
|
|
|
|
|
|
|
| |
We have the prior list_foreach() all over the code, but I need to move
where instructions live as part of adding support for control flow. Start
by just converting to a helper iterator macro. (The simpler
"qir_for_each_inst()" will be used for the for-each-inst-in-a-block
iterator macro later)
|
|
|
|
|
|
|
|
|
|
|
|
| |
This paves the way for copy propagating our unpacks. We end up with a
small change on shader-db:
total instructions in shared programs: 89390 -> 89251 (-0.16%)
instructions in affected programs: 19041 -> 18902 (-0.73%)
which appears to be because we no longer convert MOVs for an FMAX dst,
r4.unpack, r4.unpack (instead of the previous MOV dst, r4.unpack), and
this ends up with a slightly better schedule.
|
|
|
|
| |
We can do 16a/16b from float as well. No difference on shader-db.
|
|
|
|
|
|
|
|
|
|
| |
I switched us to tracking whether the results *could* go to r4, but then
didn't make a separate register class for the class bits that included r4.
Switch the "any" class to actually be "any", and name the "any but r4"
class more appropriately.
total instructions in shared programs: 96798 -> 94680 (-2.19%)
instructions in affected programs: 62736 -> 60618 (-3.38%)
|
|
|
|
|
| |
total instructions in shared programs: 97580 -> 96798 (-0.80%)
instructions in affected programs: 52826 -> 52044 (-1.48%)
|
|
|
|
|
|
|
| |
I had QPU emit code to do it, but forgot to flag the register class.
total instructions in shared programs: 97974 -> 97590 (-0.39%)
instructions in affected programs: 25291 -> 24907 (-1.52%)
|
|
|
|
|
| |
total instructions in shared programs: 98159 -> 98136 (-0.02%)
instructions in affected programs: 12279 -> 12256 (-0.19%)
|
|
|
|
|
|
|
|
|
| |
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.
|
|
|
|
|
|
|
|
|
| |
Since i965 is now using make_reg_conflicts_transitive and doesn't need
q-value computations, they are disabled on i965. They are enabled
everywhere else so that they get the old behavior. This reduces the time
spent in eglInitialize() on BDW by around 10-15%.
Reviewed-by: Eric Anholt <eric@anholt.net>
|
|
|
|
|
|
|
|
|
|
|
| |
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%)
|
| |
|
|
|
|
| |
list.h is a nicer and more familiar set of list functions/macros.
|
|
|
|
| |
I want it from another location.
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Turns out this was harmful in code quality:
total instructions in shared programs: 39487 -> 38845 (-1.63%)
instructions in affected programs: 22522 -> 21880 (-2.85%)
This costs us yet another register, which is painful since it means more
programs might fail to compile). However, the alternative was causing us
trouble where we'd save/restore r3 while it contained a MIN-ed direct
texture offset, causing the kernel to fail to validate our shaders (such
as in GLB2.7).
|
| |
|
| |
|
|
|
|
|
| |
There is an equivalent unpack function without conversion to float if you
use an integer operation instead.
|
|
|
|
|
|
|
|
|
|
| |
This increases the cost of a raddr b conflict spill (save r3 to rb31, move
src1 to r3, move rb31 back to r3 when done, instead of just move src1 to
r3), but on average thanks to instruction pairing it's more worthwhile to
have another accumulator.
total instructions in shared programs: 46428 -> 46171 (-0.55%)
instructions in affected programs: 38030 -> 37773 (-0.68%)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The register allocator walks from the end of the nodes array looking for
trivially-allocatable things to put on the stack, meaning (assuming
everything is trivially colorable and gets put on the stack in a single
pass) the low node numbers get allocated first. The things allocated
first happen to get the lower-numbered registers, which is to say the fast
accumulators that can be paired more easily.
When we previously made the nodes match the temporary register numbers,
we'd end up putting the shader inputs (VS or FS) in the accumulators,
which are often long-lived values. By prioritizing the shortest-lived
values for allocation, we can get a lot more instructions that involve
accumulators, and thus fewer conflicts for raddr and WS.
total instructions in shared programs: 52870 -> 46428 (-12.18%)
instructions in affected programs: 52260 -> 45818 (-12.33%)
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The register allocator prefers low-index registers from vc4_regs[] in the
configuration we're using, which is good because it means we prioritize
allocating the accumulators (which are faster). On the other hand, it was
causing raddr conflicts because everything beyond r0-r2 ended up in
regfile A until you got massive register pressure. By interleaving, we
end up getting more instruction pairing from getting non-conflicting
raddrs and QPU_WSes.
total instructions in shared programs: 55957 -> 52719 (-5.79%)
instructions in affected programs: 46855 -> 43617 (-6.91%)
|
|
|
|
|
|
|
|
| |
They still provide register pressure since I haven't made a special class
for them, but since they're only live for one instruction it probably
doesn't matter.
This improves the readability of QPU assembly.
|
| |
|
|
|
|
|
|
|
|
|
| |
This will let me more reliably allocate a-file registers, which are going
to be even more in demand when I start using a-file unpacks.
Also fixes a bug where the reservation of payload registers (FRAG_Z/W) was
off by one but just caused failure to register allocate at all if the
off-by-one was fixed.
|
| |
|
|
|
|
| |
I need to get the non-reciprocal version of W for interpolation, anyway.
|
| |
|
|
I'm going to be rewriting it all, and having it mixed up with the
QIR-to-QPU opcode translation was messy.
|