When spilling a register, the number of temps can be increased when
introducing a temporal variable.
Those nodes are not elegible to be spilled, but we need to take care of
no accessing out-of-bounds of the arrays defined with a size equal to
the original number of temps.
Fixes address sanitizer error on
KHR-GLES3.shaders.uniform_block.random.all_shared_buffer.14 (and many
others).
v2 (Iago):
- Add clarification in assertion.
- Use `vir_get_temp` to increase num_temps.
v3 (Iago):
- Update clarification
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
Signed-off-by: Juan A. Suarez Romero <jasuarez@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/10643>
The current policy is to always favor accumulators if possible, however,
this is not always optimal.
Particularly, accumulators play a crucial role in enabling QPU instruction
merges, since these are limited to both the ADD and the ALU instructions
addressing at most 2 physical registers. For 2-src instructions, this means
that to be able to merge we need them to address at least 2 accumulators.
While favoring accumulators does help the case for instruction merges in
general, it is risky to assign accumulators to variables that have
long life spans. Doing so will make the accumulator unavailable for
any other instructions during that life span, and since we only have a few
accumulators, we can quickly run out and losing our capacity to merge
instructions for large parts of the qpu program.
On the other hand, we also want to avoid the extreme case were we keep
allocating physical registers to the point we run out, even if we have
accumulators available, since accumulators have additional restrictions
and may not be suitable for everything.
This change continues the policy of favoring accumulators, but it only
does so if the life span of the temps is short, to ensure that we can
recycle accumulators often across instructions and avoid running out
for sections of the QPU code, unless we are already running out of
physical registers.
total instructions in shared programs: 13654647 -> 13336921 (-2.33%)
instructions in affected programs: 11015919 -> 10698193 (-2.88%)
helped: 39758
HURT: 17325
Instructions are helped.
total threads in shared programs: 412046 -> 412038 (<.01%)
threads in affected programs: 16 -> 8 (-50.00%)
helped: 0
HURT: 4
Threads are HURT.
total uniforms in shared programs: 3745726 -> 3746003 (<.01%)
uniforms in affected programs: 17296 -> 17573 (1.60%)
helped: 76
HURT: 99
Uniforms are HURT.
total max-temps in shared programs: 2364430 -> 2359942 (-0.19%)
max-temps in affected programs: 109117 -> 104629 (-4.11%)
helped: 2893
HURT: 772
Max-temps are helped.
total spills in shared programs: 5727 -> 5746 (0.33%)
spills in affected programs: 221 -> 240 (8.60%)
helped: 1
HURT: 2
total fills in shared programs: 13121 -> 13139 (0.14%)
fills in affected programs: 466 -> 484 (3.86%)
helped: 1
HURT: 2
total sfu-stalls in shared programs: 33432 -> 34491 (3.17%)
sfu-stalls in affected programs: 18219 -> 19278 (5.81%)
helped: 4459
HURT: 5087
Inconclusive result
total inst-and-stalls in shared programs: 13688079 -> 13371412 (-2.31%)
inst-and-stalls in affected programs: 11030017 -> 10713350 (-2.87%)
helped: 39630
HURT: 17429
Inst-and-stalls are helped.
total nops in shared programs: 335753 -> 333708 (-0.61%)
nops in affected programs: 112659 -> 110614 (-1.82%)
helped: 8726
HURT: 7383
Inconclusive result
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/10686>
Once we have exhausted compile strategies at 4 threads and we start
enabling lower thread counts, there is no point in starting compiles
with 4 threads for them, we know these will fail, so let's start at
2 in these cases.
This also has another nice implication: if the driver compiles at 4
threads and fails to register allocate, we were allowing it to try
with 2 threads, but this would only retry the register allocation
process and would not really recompile the shader with 2 threads. This
is not optimal, because at 2 threads we have more TMU fifo space for
each thread and we can do more TMU pipelining, so we were missing that
opportunity.
This improves performance in Sponza by ~1.5% and also seems to help
UE4 slightly.
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/10647>
Until now, if we can't compile at 4 threads we would lower thread count
with optimizations disabled, however, lowering thread count doubles the
amount of registers available per thread, so that alone is already a big
relief for register pressure so it makes sense to enable optimizations
when we do that, and progressively disable them until we enable spilling
as a last resort.
This can slightly improve performance for some applications. Sponza,
for example, gets a ~1.5% boost. I see several UE4 shaders that also get
compiled to better code at 2 threads with this, but it is more difficult
to assess how much this improves performance in practice due to the large
variance in frame times that we observe with UE4 demos.
Also, if a compiler strategy disables an optimization that did not make
any progress in the previous compile attempt, we would end up re-compiling
the exact same shader code and failing again. This, patch keeps track of
which strategies won't make progress and skips them in that case to save
some CPU time during shader compiles.
Care should be taken to ensure that we try to compile with the default
NIR scheduler at minimum thread count at least once though, so a specific
strategy for this is added, to prevent the scenario where no optimizations
are used and we skip directly to the fallback scheduler if the default
strategy fails at 4 threads.
Similarly, we now also explicitly specify which strategies are allowed to do
TMU spills and make sure we take this into account when deciding to skip
strategies. This prevents the case where no optimizations are used in a
shader and we skip directly to the fallback scheduler after failing
compilation at 2 threads with the default NIR scheduler but without trying
to spill first.
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/10647>
The Vulkan driver was already creating and using its own set of options, so
the ones defined in the compiler are only used with GL, which is confusing.
Move them to the GL driver.
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/10647>
Right now this is useful for Vulkan onnly, because GL gets loop
unrolling from the GLSL compiler and/or mesa state tracker
NIR front-ends.
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/10647>
Vertex Shader has a store_out lowering pass that converts gallium driver
locations in offsets inside the VPM.
One of the consequences is that these offsets are consecutives; that is,
if the VS stores VARYING_SLOT_VAR0.xyz and VARYING_SLOT_VAR1.xyzw, there
isn't a hole in the VPM offsets for the un-stored VARYING_SLOT_VAR0.w.
Thus we need to change how the VPM offset is computed in the Geometry
Shader when loading the inputs.
This bug is exposed by !9050.
v2 (Iago):
- Include explanatory comment.
- Use assert.
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
Signed-off-by: Juan A. Suarez Romero <jasuarez@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/10129>
`qpu.raddr_b` is an unsigned int, so it is always positive, even after
casting to signed int.
Fixes CID#1438117 "Operands don't affect result
(CONSTANT_EXPRESSION_RESULT)":
"result_independent_of_operands: (int)inst->qpu.raddr_b >= -16 is
always true regardless of the values of its operands. This occurs as
the logical first operand of "&&".
v2:
- Use signed pointers (Iago)
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
Signed-off-by: Juan A. Suarez Romero <jasuarez@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/10131>
The term 'last' may be misleading because the offset represents
the current unifa offset, which is the offset used by the last
load plus 4 bytes, so rename these to use the term 'current'
instead.
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/10100>
This implements a NIR pass that groups together constant UBO loads
for the same UBO index in order of increasing offset when the distance
between them is small enough that it enables the "skip unifa write"
optimization.
This may increase register pressure because it can move UBO loads
earlier, so we also add a compiler strategy fallback to disable the
optimization if we need to drop thread count to compile the shader
with this optimization enabled.
total instructions in shared programs: 13557555 -> 13550300 (-0.05%)
instructions in affected programs: 814684 -> 807429 (-0.89%)
helped: 4485
HURT: 2377
Instructions are helped.
total uniforms in shared programs: 3777243 -> 3760990 (-0.43%)
uniforms in affected programs: 112554 -> 96301 (-14.44%)
helped: 7226
HURT: 36
Uniforms are helped.
total max-temps in shared programs: 2318133 -> 2333761 (0.67%)
max-temps in affected programs: 63230 -> 78858 (24.72%)
helped: 23
HURT: 3044
Max-temps are HURT.
total sfu-stalls in shared programs: 32245 -> 32567 (1.00%)
sfu-stalls in affected programs: 389 -> 711 (82.78%)
helped: 139
HURT: 451
Inconclusive result.
total inst-and-stalls in shared programs: 13589800 -> 13582867 (-0.05%)
inst-and-stalls in affected programs: 817738 -> 810805 (-0.85%)
helped: 4478
HURT: 2395
Inst-and-stalls are helped.
total nops in shared programs: 354365 -> 342202 (-3.43%)
nops in affected programs: 31000 -> 18837 (-39.24%)
helped: 4405
HURT: 265
Nops are helped.
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/10100>
This adds a minimum thread count parameter to each compilation strategy with
the intention to limit the minimum allowed thread count that can be used to
register allocate with that strategy.
For now all strategies allow the minimum thread count supported by the
hardware, but we will be using this infrastructure to impose a more
strict limit in an upcoming optimization.
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/10100>
We will be using this distance to setup another optimization in a
follow-up patch.
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
x# Please enter the commit message for your changes. Lines starting
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/10100>
first_component is an uint, and thus if it takes value 0 we can't know
if it is because writemask has its first bit to 1, or all bits to 0.
As we want to ensure that at least one bit is set, apply the assertion
in writemask.
Fixes CID#1472829 "Macro compares unsigned to 0 (NO_EFFECT)".
v2:
- Restore "first_component <= last_component" assertion (Iago)
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
Signed-off-by: Juan A. Suarez Romero <jasuarez@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/10103>
A break/continue in a loop is typically emitted like this:
if (cond) {
break/continue;
} else {
}
If cond is uniform, we'll emit code for a uniform if statement and
that will emit a branch right before the if to jump directly to the
else (or the block after the else in this case, since the else is
empty) in case cond evaluates to false. This means we end up emitting
two consecutive branch instructions, one before the if and one for the
THEN block right after:
branch(!cond) -> jump to else (or after else) if cond is false
nop
nop
nop
branch -> unconditional jump to break/continue
nop
nop
nop
Instead, if we are in this scenario, we can do better by emitting the
conditional jump directly and avoiding the "jump to else" case:
branch(cond) -> jump to break/continue if cond is true
nop
nop
nop
We need to be careful when emitting the break/continue for the case
where all lanes are disabled to avoid infinite loops: if we have a
break we always want to take the jump, but we don't want to take it
if it is a continue.
total instructions in shared programs: 13563672 -> 13557348 (-0.05%)
instructions in affected programs: 348034 -> 341710 (-1.82%)
helped: 1158
HURT: 10
Instructions are helped.
total uniforms in shared programs: 3779137 -> 3777535 (-0.04%)
uniforms in affected programs: 90583 -> 88981 (-1.77%)
helped: 1169
HURT: 0
Uniforms are helped.
total max-temps in shared programs: 2317670 -> 2317575 (<.01%)
max-temps in affected programs: 1943 -> 1848 (-4.89%)
helped: 85
HURT: 4
Max-temps are helped.
total sfu-stalls in shared programs: 32247 -> 32247 (0.00%)
sfu-stalls in affected programs: 69 -> 69 (0.00%)
helped: 7
HURT: 9
Inconclusive result (value mean confidence interval includes 0).
total inst-and-stalls in shared programs: 13595919 -> 13589595 (-0.05%)
inst-and-stalls in affected programs: 350674 -> 344350 (-1.80%)
helped: 1154
HURT: 11
Inst-and-stalls are helped.
total nops in shared programs: 358202 -> 354325 (-1.08%)
nops in affected programs: 17367 -> 13490 (-22.32%)
helped: 1168
HURT: 1
Nops are helped.
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9948>
If we have an unconditional branch then we can try to fill up its
delay slots with the initial instructions of its successor block by
copying them into the delay slots and adjusting the branch offset to
skip the copied instructions.
total nops in shared programs: 365640 -> 364471 (-0.32%)
nops in affected programs: 15416 -> 14247 (-7.58%)
helped: 462
HURT: 0
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9918>
For this we do something similar to what we do with thrsw where we try to
move the branch instruction earlier so the previous instructions execute
in the delay slots of the branch.
Generally, we can do this with any instruction except:
- If the instruction reads a uniform: since our branches do as well and
uniforms come from an ordered FIFO stream.
- If the instruction writes flags, since our branch instruction will
probably read them.
- If the instruction is in the delay slots of another thread switch,
branch, or unifa write, which is disallowed.
total instructions in shared programs: 13648140 -> 13613972 (-0.25%)
instructions in affected programs: 2209552 -> 2175384 (-1.55%)
helped: 6765
HURT: 0
Instructions are helped.
total max-temps in shared programs: 2318687 -> 2318436 (-0.01%)
max-temps in affected programs: 5046 -> 4795 (-4.97%)
helped: 152
HURT: 0
Max-temps are helped.
total inst-and-stalls in shared programs: 13680494 -> 13646326 (-0.25%)
inst-and-stalls in affected programs: 2220394 -> 2186226 (-1.54%)
helped: 6765
HURT: 0
Inst-and-stalls are helped.
total nops in shared programs: 399818 -> 365640 (-8.55%)
nops in affected programs: 127311 -> 93133 (-26.85%)
helped: 6765
HURT: 0
Nops are helped.
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9918>
Do not assign to a variable that won't be used.
Fixes CID#1451708 and CID#1451710 "Unused value (UNUSED_VALUE)".
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
Signed-off-by: Juan A. Suarez Romero <jasuarez@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9910>
We were using a write dependency to ensure ordering since LDTMUs sequences
are ordered, but by using a write dependency with TMU config we were also
preserving ordering with TMU config writes that are not a sequence
terminator, which is not required and reduces scheduling flexibility.
Instead, use a write dependency to ensure strict ordering of TMU reads,
but only a read depdency with TMU config.
With this change we also need to update CS barriers to also have a write
dependency with TMU reads to ensure that we don't move TMU reads around
CS barriers.
total instructions in shared programs: 13602500 -> 13597851 (-0.03%)
instructions in affected programs: 2681428 -> 2676779 (-0.17%)
helped: 6567
HURT: 4960
Instructions are helped.
total max-temps in shared programs: 2317927 -> 2317914 (<.01%)
max-temps in affected programs: 13861 -> 13848 (-0.09%)
helped: 355
HURT: 300
Inconclusive result (value mean confidence interval includes 0).
total sfu-stalls in shared programs: 32074 -> 32247 (0.54%)
sfu-stalls in affected programs: 848 -> 1021 (20.40%)
helped: 160
HURT: 327
Inconclusive result (%-change mean confidence interval includes 0).
total inst-and-stalls in shared programs: 13634574 -> 13630098 (-0.03%)
inst-and-stalls in affected programs: 2703041 -> 2698565 (-0.17%)
helped: 6558
HURT: 5020
Inst-and-stalls are helped.
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9856>
Instead of last TMU write. According to the documentation, the entries
in the output FIFO are pushed with the *final* input write for the
lookup, which is the one terminating the sequence. We flag these
with last_tmu_config.
This will allow us to move all TMU register writes for a lookup except
the last one ahead of the LDTMUs for the previous lookup, possibly
allowing us to pair up these writes the wrtmuc instructions for the
same lookup, turning code like this:
nop ; nop ; wrtmuc (tex[0].p0 | 0x3)
nop ; nop ; wrtmuc (tex[2].p1 | 0x1)
nop ; nop ; ldunif (ubo[2]+0xe0)
fadd r4, rf33, rf51 ; mov unifa, r5 ; ldunif (ubo[2]+0x110)
fmax rf34, 0, r4 ; nop
nop ; mov tmut, rf11
nop ; mov tmus, rf0
into:
nop ; mov tmut, rf11 ; wrtmuc (tex[0].p0 | 0x3)
nop ; nop ; wrtmuc (tex[2].p1 | 0x1)
nop ; nop ; ldunif (ubo[2]+0xe0)
fadd r4, rf33, rf51 ; mov unifa, r5 ; ldunif (ubo[2]+0x110)
fmax rf34, 0, r4 ; nop
nop ; mov tmus, rf0
total instructions in shared programs: 13648140 -> 13602500 (-0.33%)
instructions in affected programs: 3497402 -> 3451762 (-1.30%)
helped: 12044
HURT: 3484
Instructions are helped.
total max-temps in shared programs: 2318687 -> 2317927 (-0.03%)
max-temps in affected programs: 17234 -> 16474 (-4.41%)
helped: 615
HURT: 198
Max-temps are helped.
total sfu-stalls in shared programs: 32354 -> 32074 (-0.87%)
sfu-stalls in affected programs: 1462 -> 1182 (-19.15%)
helped: 461
HURT: 188
Sfu-stalls are helped.
total inst-and-stalls in shared programs: 13680494 -> 13634574 (-0.34%)
inst-and-stalls in affected programs: 3514405 -> 3468485 (-1.31%)
helped: 12062
HURT: 3486
Inst-and-stalls are helped.
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9856>
The way we handle thrsw instructions is that we try to merge them
back into previously scheduled instructions to fill up its delay
slots. This is generally safe, because the thrsw won't happen until
after the delay slots, so we are not really changing the execution
order of the instructions and we just need to make sure we don't
violate a few specific restrictions.
If we have not managed to fill up all delay slots after doing this,
then we emit as many NOPs as needed to fill them. This is to ensure
that we don't schedule an instruction that needs to execute after the
thread switch before the thread switch happens. However, doing this
can lead to inefficient code, since some times the instructions we
schedule after a thrsw are indepdent of the thrsw and could be safely
executed in its delay slots.
This change removes the fixed NOP emission after a thrsw to fill
delay slots and instead adds code to ensure that our instruction
scheduling is aware of when it is scheduling instructions in the
delay slots of a previous thrsw to avoid selecting conflicting
instructions.
The only case were we still emit fixed NOPs is for the thread end that
we emit to terminate the program after scheduling all instructions
because we can't end the instruction stream before the thread end
is properly executed.
total instructions in shared programs: 13691004 -> 13648140 (-0.31%)
instructions in affected programs: 4345951 -> 4303087 (-0.99%)
helped: 19645
HURT: 652
Instructions are helped.
total max-temps in shared programs: 2319317 -> 2318687 (-0.03%)
max-temps in affected programs: 10510 -> 9880 (-5.99%)
helped: 532
HURT: 9
Max-temps are helped.
total sfu-stalls in shared programs: 31752 -> 32354 (1.90%)
sfu-stalls in affected programs: 840 -> 1442 (71.67%)
helped: 7
HURT: 467
Sfu-stalls are HURT.
total inst-and-stalls in shared programs: 13722756 -> 13680494 (-0.31%)
inst-and-stalls in affected programs: 4335590 -> 4293328 (-0.97%)
helped: 19453
HURT: 758
Inst-and-stalls are helped.
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9825>
Integer add/sub can be implemented as either an add or a mul instruction
but we always emit them as add instructions at VIR level. We can use this
flexibility to improve our QPU scheduling so we can be more effective
at instruction merging by converting these to mul instructions when we
are attempting to merge them with another add instruction.
total instructions in shared programs: 13721549 -> 13691004 (-0.22%)
instructions in affected programs: 3340493 -> 3309948 (-0.91%)
helped: 12805
HURT: 1656
Instructions are helped.
total max-temps in shared programs: 2319528 -> 2319317 (<.01%)
max-temps in affected programs: 5285 -> 5074 (-3.99%)
helped: 195
HURT: 3
Max-temps are helped.
total sfu-stalls in shared programs: 31616 -> 31752 (0.43%)
sfu-stalls in affected programs: 469 -> 605 (29.00%)
helped: 52
HURT: 161
Sfu-stalls are HURT.
total inst-and-stalls in shared programs: 13753165 -> 13722756 (-0.22%)
inst-and-stalls in affected programs: 3340383 -> 3309974 (-0.91%)
helped: 12782
HURT: 1666
Inst-and-stalls are helped.
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9769>
This maps the nir shader data.location to its final
data.driver_location. In general we are using the driver location as
index (like vattr_sizes on the same struct), so having this map is
useful if what we have is the data.location, and we don't have
available the original nir shader.
v2: use memset instead of for loop, and nir_foreach_shader_in_variable
instead of nir_foreach_variable_with_modes (Iago)
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9403>
As we plan to try to get directly the compiled variant from the cache,
it would be possible to not have available the nir shaders, so we add
this info on prog data.
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9403>
This pass splits writemaks with non-consecutive bits into multiple
store operations ensuring that each store only has consecutive
writemask bits set.
We can use this to simplify writemask handling in our backend removing
a loop solely intended to handle this case.
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9619>
This will make it so we pack consecutive scalar operations into a vector
operation, reducing the amount of load/store operations in the NIR program.
Our backend can handle vector load/stores, and doing so may be more efficient
since we don't need to setup individual load/stores all the time.
A pathological case is:
dEQP-VK.spirv_assembly.instruction.compute.opcopymemory.array
which goes from 862 instructions to only 573 by converting
all scalar SSBO load/store operations to vec4 operations.
total instructions in shared programs: 13752607 -> 13733627 (-0.14%)
instructions in affected programs: 367117 -> 348137 (-5.17%)
helped: 1168
HURT: 371
Instructions are helped.
total threads in shared programs: 412230 -> 412272 (0.01%)
threads in affected programs: 54 -> 96 (77.78%)
helped: 23
HURT: 2
Threads are helped.
total uniforms in shared programs: 3790248 -> 3784601 (-0.15%)
uniforms in affected programs: 57417 -> 51770 (-9.84%)
helped: 1420
HURT: 19
Uniforms are helped.
total max-temps in shared programs: 2322170 -> 2322714 (0.02%)
max-temps in affected programs: 14353 -> 14897 (3.79%)
helped: 185
HURT: 306
Max-temps are HURT.
total spills in shared programs: 5940 -> 6010 (1.18%)
spills in affected programs: 65 -> 135 (107.69%)
helped: 0
HURT: 11
total fills in shared programs: 13372 -> 13494 (0.91%)
fills in affected programs: 75 -> 197 (162.67%)
helped: 0
HURT: 11
total sfu-stalls in shared programs: 31505 -> 31521 (0.05%)
sfu-stalls in affected programs: 751 -> 767 (2.13%)
helped: 210
HURT: 246
Inconclusive result (value mean confidence interval includes 0).
total inst-and-stalls in shared programs: 13784112 -> 13765148 (-0.14%)
inst-and-stalls in affected programs: 360283 -> 341319 (-5.26%)
helped: 1125
HURT: 366
Inst-and-stalls are helped.
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9619>
V3D 4.x allows more flexibility, so take advantage of that. Generally,
we can reorder any writes in the same sequence, so long as they are
not the sequence terminator (which must always be last, since it is
the one triggering the operation), and TMUD writes, since these must
be ordered with respect to each other.
total instructions in shared programs: 13735183 -> 13731927 (-0.02%)
instructions in affected programs: 903057 -> 899801 (-0.36%)
helped: 2358
HURT: 746
Instructions are helped.
total max-temps in shared programs: 2322020 -> 2322009 (<.01%)
max-temps in affected programs: 619 -> 608 (-1.78%)
helped: 19
HURT: 11
Inconclusive result (value mean confidence interval includes 0).
total sfu-stalls in shared programs: 31494 -> 31489 (-0.02%)
sfu-stalls in affected programs: 182 -> 177 (-2.75%)
helped: 40
HURT: 40
Inconclusive result (value mean confidence interval includes 0).
total inst-and-stalls in shared programs: 13766677 -> 13763416 (-0.02%)
inst-and-stalls in affected programs: 901343 -> 898082 (-0.36%)
helped: 2349
HURT: 746
Inst-and-stalls are helped.
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9555>
Instead of using a write depdency. We use last_tmu_config to ensure ordering
of instructions participating in different TMU sequences. To this end,
all sequence terminators flag a write dependency on last_tmu_config, but
wrtmuc is not a sequence terminator, so we can be more flexible by flagging
it as a read depedency. This would prevent it to be moved into a previous
sequence (since it cannot be moved past the previous sequence terminator due
to the read depedency), but it allows it to be reordered with instructions in
the same sequence, which allows us to pair it up more effectively. Particularly,
it allows to pair up a wrtmuc with the sequence terminator of the same sequence,
turning code like this:
nop ; mov tmut, r0 ; thrsw; wrtmuc (tex[0].p0 | 0x3)
nop ; nop ; wrtmuc (tex[0].p1 | 0x0)
nop ; mov tmus, r1
Into this:
nop ; mov tmut, r0 ; thrsw; wrtmuc (tex[0].p0 | 0x3)
nop ; mov tmus, r1 ; wrtmuc (tex[0].p1 | 0x0)
total instructions in shared programs: 13755738 -> 13735183 (-0.15%)
instructions in affected programs: 2510921 -> 2490366 (-0.82%)
helped: 10963
HURT: 485
Instructions are helped.
total max-temps in shared programs: 2322828 -> 2322020 (-0.03%)
max-temps in affected programs: 11303 -> 10495 (-7.15%)
helped: 608
HURT: 19
Max-temps are helped.
total sfu-stalls in shared programs: 31545 -> 31494 (-0.16%)
sfu-stalls in affected programs: 235 -> 184 (-21.70%)
helped: 62
HURT: 11
Sfu-stalls are helped.
total inst-and-stalls in shared programs: 13787283 -> 13766677 (-0.15%)
inst-and-stalls in affected programs: 2525187 -> 2504581 (-0.82%)
helped: 10989
HURT: 477
Inst-and-stalls are helped.
v2: add a comment explaining the read depdency (Piñeiro).
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9555>
This restores many of the hurt shaders from the previous patch at the
expense of re-adding ldvary tracking in the scheduler.
total instructions in shared programs: 13760415 -> 13755738 (-0.03%)
instructions in affected programs: 1207560 -> 1202883 (-0.39%)
helped: 5080
HURT: 1731
Instructions are helped.
total max-temps in shared programs: 2322991 -> 2322828 (<.01%)
max-temps in affected programs: 5063 -> 4900 (-3.22%)
helped: 229
HURT: 108
Max-temps are helped.
total sfu-stalls in shared programs: 31827 -> 31545 (-0.89%)
sfu-stalls in affected programs: 478 -> 196 (-59.00%)
helped: 304
HURT: 21
Sfu-stalls are helped.
total inst-and-stalls in shared programs: 13792242 -> 13787283 (-0.04%)
inst-and-stalls in affected programs: 1220856 -> 1215897 (-0.41%)
helped: 5162
HURT: 1697
Inst-and-stalls are helped.
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9471>
We get optimal ldvary pipelining by doing the following:
1) Carefully merge a paired ldvary into the previous instruction when
possible.
2) When the above succeeds, flag the ldvary as scheduled immediately so
we can merge one of its children into the current instruction.
3) When scheduling ldvary sequences, only pick up instructions that are
part of the sequence to avoid picking up something that prevents
successful pipelining.
This patch skips 3) assuming some hurt shaders in exchange for better
scheduling flexibility during ldvary sequences. Besides eliminating most
of the code dedicated to special handling ldvary sequences, this also
usually allows us to produce better code by merging instructions that are
unrelated to ldvary sequences into the ldvary sequences, which is
particularly effective to fill up the gaps produced when scheduling the
first and last ldvary sequences as well as the gaps produced by flat
and noperspective varyings sequences that don't have both mul and add
instructions.
Notice that there are some hurt shaders, because some times the extra
scheduler flexibility can lead to picking up instructions that will
break a sequence without compensating for that, typically an ldunif
that prevents us from doing the fixup for a follow-up ldvary. We will
try to correct some of these cases with the next patch.
total instructions in shared programs: 13786037 -> 13760415 (-0.19%)
instructions in affected programs: 3201387 -> 3175765 (-0.80%)
helped: 16155
HURT: 4146
Instructions are helped.
total max-temps in shared programs: 2324834 -> 2322991 (-0.08%)
max-temps in affected programs: 22160 -> 20317 (-8.32%)
helped: 1340
HURT: 103
Max-temps are helped.
total sfu-stalls in shared programs: 30685 -> 31827 (3.72%)
sfu-stalls in affected programs: 782 -> 1924 (146.04%)
helped: 253
HURT: 1416
Inconclusive result.
total inst-and-stalls in shared programs: 13816722 -> 13792242 (-0.18%)
inst-and-stalls in affected programs: 3171642 -> 3147162 (-0.77%)
helped: 15331
HURT: 4179
Inst-and-stalls are helped.
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9471>
These checks depend on prev_inst being set, so move them down below
with all the other checks with the same requirement.
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9471>
This replaces the new_src parameter of nir_ssa_def_rewrite_uses_after()
with an SSA def, and rewrites all the users as needed.
Acked-by: Alyssa Rosenzweig <alyssa@collabora.com>
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9383>
This commit replaces the new_src parameter of nir_ssa_def_rewrite_uses()
with an SSA def, removes nir_ssa_def_rewrite_uses_ssa(), and rewrites
all the users as needed.
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Acked-by: Alyssa Rosenzweig <alyssa@collabora.com>
Acked-By: Mike Blumenkrantz <michael.blumenkrantz@gmail.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9383>
We were checking that the previous instruction doesn't write flags,
but we also need to check it doesn't read them.
Fixes: 1784dd22a3 ('broadcom/compiler: pipeline smooth ldvary sequences')
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9431>
When we were only able to pipeline smooth varyings, if we had to disable
ldvary pipelining in the middle of a sequence it would stay disabled for
the rest of the program, to prevent us from prioritizing scheduling of
ldvary instructions that we would not be able to pipeline effectively.
Now that we can pipeline all ldvary sequences we can change this.
This change re-enables ldvary pipelining upon finding the next
ldvary in the program in the hopes that we can continue pipelining
succesfully. To do this, we track the number of ldvary instructions we
emitted so far and compare that to the number of inputs in the fragment
shader we are scheduling. This also allows us to simplify our ldvary
tracking at nir to vir time, since that is all now handled in the QPU
scheduler.
total instructions in shared programs: 13817048 -> 13810783 (-0.05%)
instructions in affected programs: 810114 -> 803849 (-0.77%)
helped: 4843
HURT: 591
Instructions are helped.
total max-temps in shared programs: 2326612 -> 2326300 (-0.01%)
max-temps in affected programs: 4689 -> 4377 (-6.65%)
helped: 285
HURT: 7
Max-temps are helped.
total sfu-stalls in shared programs: 30942 -> 30865 (-0.25%)
sfu-stalls in affected programs: 207 -> 130 (-37.20%)
helped: 120
HURT: 42
Sfu-stalls are helped.
total inst-and-stalls in shared programs: 13847990 -> 13841648 (-0.05%)
inst-and-stalls in affected programs: 825378 -> 819036 (-0.77%)
helped: 4899
HURT: 590
Inst-and-stalls are helped.
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9404>
We had an optimization in place to skip a unifa write if the address
happens to be right after the last ldunifa read address, but we can
take this further and update the unifa address by emitting ldunifa
instructions if needed to skip a unifa write that is close enough.
This is because a unifa write involves 4 cycles: 1 for the write
and 3 delay slots before we can emit the first ldunifa.
So if we have code like this:
unifa addr + 0
ldunifa.r0
unifa addr + 12
ldunifa.r1
In practice we end up with QPU like this:
unifa addr + 0
nop
nop
nop
ldunifa.r0
unifa addr + 12
nop
nop
nop
ldunifa.r1
And with this patch we get:
unifa addr + 0
nop
nop
nop
ldunifa.r0 <--- reads offset 0
ldunifa.- <--- reads offset 4
ldunifa.- <--- reads offset 8
ldunifa.r1 <--- reads offset 12
Of course, QPU scheduling might find ways to fill the NOPs to some
extent and remove some of the gains, but generally speaking, this is
still usually a win.
Going by shader-db results, allowing the next unifa address to be up
to 12 bytes after the address resulting from the last ldunifa read
shows the best results:
total instructions in shared programs: 13817048 -> 13812202 (-0.04%)
instructions in affected programs: 602701 -> 597855 (-0.80%)
helped: 1750
HURT: 760
Instructions are helped.
total uniforms in shared programs: 3795485 -> 3793200 (-0.06%)
uniforms in affected programs: 43930 -> 41645 (-5.20%)
helped: 898
HURT: 0
Uniforms are helped.
total max-temps in shared programs: 2326612 -> 2326621 (<.01%)
max-temps in affected programs: 651 -> 660 (1.38%)
helped: 10
HURT: 21
Inconclusive result (value mean confidence interval includes 0).
total sfu-stalls in shared programs: 30942 -> 30906 (-0.12%)
sfu-stalls in affected programs: 627 -> 591 (-5.74%)
helped: 186
HURT: 158
Inconclusive result (value mean confidence interval includes 0).
total inst-and-stalls in shared programs: 13847990 -> 13843108 (-0.04%)
inst-and-stalls in affected programs: 601404 -> 596522 (-0.81%)
helped: 1747
HURT: 757
Inst-and-stalls are helped.
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9384>
We can't remove unused ldunifa that are not the first or last
in a sequence, but we can still ignore their destination
to reduce register pressure.
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9384>
