With the exception of the backwards branch for loops, all the control flow we
insert during instruction selection just predicates instructions rather than
actually jumping around. That means, for example, we execute both sides of the
if even for a uniform condition! That's inefficient. The solution is insert
jmp_exec_none instructions after control flow in order to skip unexecuted
regions, which is much faster than predicating them out. However, jmp_exec_none
is costly in itself, so we need to use a heuristic to determine when it's
actually beneficial.
This uses a very simple heuristic for this purpose. However, it is a massive
performance speed-up for Dolphin uber shaders: 39fps -> 67fps at 2x resolution.
Nearly a doubling of performance!
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
jmp_exec_none variant that jumps to the last instruction of the target block,
rather than the beginning. This is convenient for skipping over elses, while
still executing the block-final pop_exec instruction. Similarly for skipping
over loop bodies while still executing the block-final pop_exec, after break
instructions.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Add an optional pointer to a target block for these instructions. This does NOT
act like a logical branch, and does NOT get added to the logical control flow.
It is ignored wholesale until after RA, when physical edges may be inserted by a
pass we add later in this series.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Although this is well-motivated, perf effect seems to be neglible for Dolphin.
It does prevent the scheduler from making things worse by sinking these
instructions though, so as a way to prevent future problems this seems sensible.
The kind of problem this affects (late discard) isn't modelled in shader-db.
Nevertheless, nothing concerning there:
total instructions in shared programs: 1756699 -> 1756722 (<.01%)
instructions in affected programs: 10106 -> 10129 (0.23%)
helped: 21
HURT: 41
Inconclusive result (value mean confidence interval includes 0).
total bytes in shared programs: 11525404 -> 11525452 (<.01%)
bytes in affected programs: 72900 -> 72948 (0.07%)
helped: 27
HURT: 41
Inconclusive result (value mean confidence interval includes 0).
total halfregs in shared programs: 483394 -> 483286 (-0.02%)
halfregs in affected programs: 4945 -> 4837 (-2.18%)
helped: 88
HURT: 78
Inconclusive result (value mean confidence interval includes 0).
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
This fixes live range splitting with 3D textureGrad(), which involves vectors
larger than the natural 128-bit maximum and hence requires special handling.
Fixes this assert with a combination of debug flags and new patches:
unsigned int find_best_region_to_evict(struct ra_ctx *, unsigned int,
unsigned int *, unsigned int *):
Assertion `(rctx->bound % size) == 0 && "register file size must be aligned
to the maximum vector size"' failed
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
This lets us save a LOT of instructions at the cost of increased register
pressure. However, on my shader-db, this is still coming out ahead since no
shaders are hurt for thread count/spills, and only 1/10 of the shaders helped
for instruction count are hurt for register pressure. The shaders most hurt
for pressure have very low pressure (7 -> 15 is the worst case) and you need a
certain number of registers to use a 4 source instruction at all. Analyzing the
hurt shaders, nothing concerns me too much ... this isn't as bad as I feared.
So I think at this point it's worth ripping off the bandage, given the massive
potential for instruction count win. This is a big improvement for some of the
shaders I'm working on for my $SECRET_PROJECT.
total instructions in shared programs: 1784943 -> 1775169 (-0.55%)
instructions in affected programs: 644211 -> 634437 (-1.52%)
helped: 3498
HURT: 38
Instructions are helped.
total bytes in shared programs: 11720734 -> 11643224 (-0.66%)
bytes in affected programs: 4370986 -> 4293476 (-1.77%)
helped: 3572
HURT: 36
Bytes are helped.
total halfregs in shared programs: 474094 -> 475165 (0.23%)
halfregs in affected programs: 12821 -> 13892 (8.35%)
helped: 65
HURT: 247
Halfregs are HURT.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/25052>
Simple greedy thing that has the potential to inflate register pressure but
reduces instructions. Thanks to the recent loop work that turns if { break }
into while_icmp, this also implicitly handles fusing conditions into loops,
which is what actually prompted this.
Surprisingly, this helps register pressure on my shader-db (no change to thread
count), I guess by eliminating the boolean temps in case where the sources are
used multiple times.
total instructions in shared programs: 1786561 -> 1784943 (-0.09%)
instructions in affected programs: 128557 -> 126939 (-1.26%)
helped: 474
HURT: 13
Instructions are helped.
total bytes in shared programs: 11733236 -> 11720734 (-0.11%)
bytes in affected programs: 976034 -> 963532 (-1.28%)
helped: 521
HURT: 13
Bytes are helped.
total halfregs in shared programs: 474245 -> 474094 (-0.03%)
halfregs in affected programs: 1869 -> 1718 (-8.08%)
helped: 28
HURT: 7
Halfregs are helped.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/25052>
Apple doesn't do this, but it should be equivalent and it makes it easier to see
that we can use while_icmp for break_if_icmp in loops that don't use continue
(which Apple does do). So, the effect of this commit is to use while_icmp for
most breaks, which saves an instruction.
total instructions in shared programs: 1764199 -> 1764076 (<.01%)
instructions in affected programs: 24149 -> 24026 (-0.51%)
helped: 78
HURT: 0
Instructions are helped.
total bytes in shared programs: 11609306 -> 11608322 (<.01%)
bytes in affected programs: 164604 -> 163620 (-0.60%)
helped: 78
HURT: 0
Bytes are helped.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/25052>
The only role of the while_icmp at the end of a NIR loop is to make continue
jumps work. If, after emitting the loop, we learn that there are no continues,
there is no need to insert a while_icmp since it would be a no-op anyway.
total instructions in shared programs: 1764311 -> 1764199 (<.01%)
instructions in affected programs: 26321 -> 26209 (-0.43%)
helped: 82
HURT: 0
Instructions are helped.
total bytes in shared programs: 11609978 -> 11609306 (<.01%)
bytes in affected programs: 178842 -> 178170 (-0.38%)
helped: 82
HURT: 0
Bytes are helped.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/25052>
In general, loops need a push_exec at the start for correctness. However, a
push_exec at the top level (non-nested) is a no-op, so we can omit and save a
few cycles.
total instructions in shared programs: 1764350 -> 1764311 (<.01%)
instructions in affected programs: 7339 -> 7300 (-0.53%)
helped: 36
HURT: 0
Instructions are helped.
total bytes in shared programs: 11610212 -> 11609978 (<.01%)
bytes in affected programs: 48638 -> 48404 (-0.48%)
helped: 36
HURT: 0
Bytes are helped.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/25052>
Search for code like
if ... {
break
}
and replace with a break_if pseudo-instruction for optimized handling, since the
break_if lowering is better than the original code.
total instructions in shared programs: 1764596 -> 1764350 (-0.01%)
instructions in affected programs: 24540 -> 24294 (-1.00%)
helped: 78
HURT: 0
Instructions are helped.
total bytes in shared programs: 11611196 -> 11610212 (<.01%)
bytes in affected programs: 166458 -> 165474 (-0.59%)
helped: 78
HURT: 0
Bytes are helped.
shader-db probably understates the benefit here, since this optimizes the body
of loops.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/25052>
We use it for two different things. Pseudo-instructions are cheap, split it up
for easier optimization passes. This also fixes the schedule classes.. we can
move the cf_begin around if we want, it's inert.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/25052>
They're more trouble than they're worth for us. They were originally lifted
unthinkingly from ACO, where I assume they're necessary for software CF
lowering, but they're just an inconvenient convenience for us. Remove em.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/25052>
Since we register allocate in SSA, the number of registers required (register
demand) equals to the maximum number of simultaneous live values (register
pressure). So if we can reduce register pressure, we are guaranteed to reduce
register demand. Even an ineffective heuristic like randomly swapping
instructions can only reduce pressure as long as it's conservative.
This implements one such heuristic: in each block, schedule backwards, selecting
the free instruction that looks like it will reduce liveness the most. In other
words, the greedy algorithm to reduce register pressure. At the end of the
block, if we haven't actually reduced pressure, we bail. This isn't optimal, but
it's well-motivated and optimally handles special cases (like 0-source
instructions).
This is based on the scheduler I originally wrote for Mali.
In my Dolphin ubershader branch, this improved performance at native 4K by 10fps
(105fps->115fps) when I measured together with some other optimizations. On top
of my current next (which notably includes nir_opt_sink improvements), this
commit alone goes (53fps->54fps) which is considerably less impressive :-p
shader-db results are a win, but not as large as we might hope. Instruction
count win seems to be from the smaller live ranges being easier on RA (fewer
swaps / moves). The two shaders affected for thread count are from fifa mobile,
which go from 640 threads ->
1024 (full occupancy). In other words... this heuristic does an excellent job in
a small subset of shaders. The Dolphin ubershader win was real, though :~)
Note these shader-db wins are on top of a branch with the nir_opt_sink
improvements. Without that, the stats are much better... The schedulers have
some overlap, but they're better together.
total instructions in shared programs: 1766635 -> 1763496 (-0.18%)
instructions in affected programs: 445855 -> 442716 (-0.70%)
helped: 1963
HURT: 350
Instructions are helped.
total bytes in shared programs: 11597648 -> 11586924 (-0.09%)
bytes in affected programs: 3106230 -> 3095506 (-0.35%)
helped: 2003
HURT: 374
Bytes are helped.
total halfregs in shared programs: 504609 -> 481980 (-4.48%)
halfregs in affected programs: 138322 -> 115693 (-16.36%)
helped: 3405
HURT: 311
Halfregs are helped.
total threads in shared programs: 18839936 -> 18840704 (<.01%)
threads in affected programs: 1280 -> 2048 (60.00%)
helped: 2
HURT: 0
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/25052>
Some special registers imply scheduling constraints. We want to have a single
scheduling class per instruction in the IR, so fork off various get_sr variants
depending on what kind of SR we're reading, and validate that we use the right
kind.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/25052>
