This just eliminates tautological / contradictory compares that are used
for bcsel and other non-if-statement cases. If-statements are not
affected because removing flow control can cause the i965 instrution
scheduler to create some very long live ranges resulting in unncessary
spilling. This causes some shaders to fall of a performance cliff.
Since many small if-statements are already flattened to bcsel, this
optimization covers more than 68% of the possible cases (2417 shaders
helped for instructions on Skylake vs. 3554).
v2: Reorder and add whitespace to make the relationship between the
patterns more obvious. Suggested by Caio.
All Gen7+ platforms had similar results. (Ice Lake shown)
total instructions in shared programs: 16333474 -> 16322028 (-0.07%)
instructions in affected programs: 438559 -> 427113 (-2.61%)
helped: 1765
HURT: 0
helped stats (abs) min: 1 max: 275 x̄: 6.48 x̃: 4
helped stats (rel) min: 0.20% max: 36.36% x̄: 4.07% x̃: 1.82%
95% mean confidence interval for instructions value: -6.87 -6.10
95% mean confidence interval for instructions %-change: -4.30% -3.84%
Instructions are helped.
total cycles in shared programs: 367608554 -> 367511103 (-0.03%)
cycles in affected programs: 8368829 -> 8271378 (-1.16%)
helped: 1541
HURT: 129
helped stats (abs) min: 1 max: 4468 x̄: 66.78 x̃: 39
helped stats (rel) min: 0.01% max: 45.69% x̄: 4.10% x̃: 2.17%
HURT stats (abs) min: 1 max: 973 x̄: 42.25 x̃: 10
HURT stats (rel) min: 0.02% max: 64.39% x̄: 2.15% x̃: 0.60%
95% mean confidence interval for cycles value: -64.90 -51.81
95% mean confidence interval for cycles %-change: -3.89% -3.36%
Cycles are helped.
total spills in shared programs: 8867 -> 8868 (0.01%)
spills in affected programs: 18 -> 19 (5.56%)
helped: 0
HURT: 1
total fills in shared programs: 21900 -> 21903 (0.01%)
fills in affected programs: 78 -> 81 (3.85%)
helped: 0
HURT: 1
All Gen6 and earlier platforms had similar results. (Sandy Bridge shown)
total instructions in shared programs: 10829877 -> 10829247 (<.01%)
instructions in affected programs: 30240 -> 29610 (-2.08%)
helped: 177
HURT: 0
helped stats (abs) min: 1 max: 15 x̄: 3.56 x̃: 3
helped stats (rel) min: 0.37% max: 17.39% x̄: 2.68% x̃: 1.94%
95% mean confidence interval for instructions value: -3.93 -3.18
95% mean confidence interval for instructions %-change: -3.04% -2.32%
Instructions are helped.
total cycles in shared programs: 154036580 -> 154035437 (<.01%)
cycles in affected programs: 352402 -> 351259 (-0.32%)
helped: 96
HURT: 28
helped stats (abs) min: 1 max: 128 x̄: 14.73 x̃: 6
helped stats (rel) min: 0.03% max: 24.00% x̄: 1.51% x̃: 0.46%
HURT stats (abs) min: 1 max: 117 x̄: 9.68 x̃: 4
HURT stats (rel) min: 0.03% max: 2.24% x̄: 0.43% x̃: 0.23%
95% mean confidence interval for cycles value: -13.40 -5.03
95% mean confidence interval for cycles %-change: -1.62% -0.53%
Cycles are helped.
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
Optimizations that insert bitshift or bitwise operations should not be
applied on GPUs that don't support integer operations.
The .lower_bitshift could be used to control the bitshift related ones,
but there was also one bitwise optimization uncovered.
Since only lima and freedreno use this option and the use case is that
no bit operations are wanted, let's rename it to .lower_bitops and use
it to control all bitops related optimizations.
Signed-off-by: Erico Nunes <nunes.erico@gmail.com>
Reviewed-by: Jonathan Marek <jonathan@marek.ca>
The Mali400 pp doesn't implement fdot but has fsum3 and fsum4, which can
be used to optimize fdot lowering. fsum2 is not implemented and can be
further lowered to an add with the vector components.
Currently lima ppir handles this lowering internally, however this
happens in a very late stage and requires a big chunk of code compared
to a nir_opt_algebraic lowering.
By having fsum in nir, we can reduce ppir complexity and enable the
lowered ops to be part of other nir optimizations in the optimization
loop.
Signed-off-by: Erico Nunes <nunes.erico@gmail.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
When 'x' is the result of a scmp op:
x != 0.0 or x == 1.0: passthrough
x == 0.0 or x != 1.0: invert
Signed-off-by: Jonathan Marek <jonathan@marek.ca>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Add generic lowerings for fall_equalN/fany_nequalN. These should be optimal
for vec4 backends that doesn't have any special instructions for it, as
long as they support saturate.
Signed-off-by: Jonathan Marek <jonathan@marek.ca>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Add simple fdot2 optimizations that are missing.
Signed-off-by: Jonathan Marek <jonathan@marek.ca>
Reviewed-by: Thomas Helland <thomashelland90@gmail.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
For backends that don't have a 'fdph' instructions
Signed-off-by: Jonathan Marek <jonathan@marek.ca>
Reviewed-by: Thomas Helland <thomashelland90@gmail.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
This version has less ops for the same precision.
Signed-off-by: Jonathan Marek <jonathan@marek.ca>
Reviewed-by: Vasily Khoruzhick <anarsoul@gmail.com>
Acked-by: Matt Turner <mattst88@gmail.com>
This is to allow optimizations in nir_opt_algebraic not otherwise possible
Signed-off-by: Jonathan Marek <jonathan@marek.ca>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Acked-by: Matt Turner <mattst88@gmail.com>
These seem like obvious enough optimizations in the world of multiple
integer bit sizes. The only known thing which hits these at the moment
is some Vulkan CTS tests for 16-bit SSBO values which like to up-cast
and check for equality. However, it's something that's bound to come up
as we start seeing more integers in shaders.
The optimizations of comparisons of casted values with constants are
something which we would ideally do with range analysis. However,
lacking that, we can do it in opt_algebraic as long as one side is a
constant.
In dEQP-VK.ssbo.phys.layout.random.16bit.scalar.13, this commit, along
with the previous commit, reduce the number of instructions emitted on
Skylake from 55328 to 44546, a reduction of 20%.
Acked-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
We could, in theory, add the same optimization for 64-bit unpack
operations but that's likely to fight with 64-bit integer lowering on
platforms which require it so it will require more infrastructure before
that will be a good idea.
Reviewed-by: Matt Turner <mattst88@gmail.com>
Search patterns that are expected to have too many (e.g., the giant
bitfield_reverse pattern) can be added to a white list.
This would have saved me a few hours debugging. :(
v2: Implement the expected-failure annotation as a property of the
search-replace pattern instead of as a property of the whole list of
patterns. Suggested by Connor.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Reviewed-by: Dylan Baker <dylan@pnwbakers.com>
The bfi/bfm behavior change replaced the bfi/bfm usage in
lower_bitfield_insert_to_shifts with actual shifts like the name says,
but it failed to handle the offset=0, bits==32 case in the new
lowering.
v2: Use 31 < bits instead of bits == 32, to get the 31 < (iand bits,
31) -> false optimization.
Fixes regressions in dEQP-GLES31.*bitfield_insert* on freedreno.
Fixes: 165b7f3a44 ("nir: define behavior of nir_op_bfm and nir_op_u/ibfe according to SM5 spec.")
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
The helpers are needed so we can use the syntax `instr(cond)` in the
algebraic rules. Add simple rule for dropping a pair of mul-div of
the same value when wrapping is guaranteed to not happen.
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
There doesn't seem to be any reason to keep these opcodes around:
* fnot/fxor are not used at all.
* fand/for are only used in lower_alu_to_scalar, but easily replaced
Signed-off-by: Jonathan Marek <jonathan@marek.ca>
Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
This lets us use the optimization pattern
(('ult', 31, ('iand', b, 31)), False) to remove the
bcsel instruction for code originating in D3D shaders.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
The [iu]bfe and bfm instructions are defined to only use the five
least significant bits.
This optimizes a common pattern from D3D -> SPIR-V translation.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
That is: the five least significant bits provide the values of
'bits' and 'offset' which is the case for all hardware currently
supported by NIR and using the bfm/bfe instructions.
This patch also changes the lowering of bitfield_insert/extract
using shifts to not use bfm and removes the flag 'lower_bfm'.
Tested-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
These optimizations are based on the fact that
'and(a,b) <= umin(a,b)'.
For AMD, this series moves the optimization from LLVM to NIR,
so currently no vkpipeline-db changes here.
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
The rules added in patch 3addd7c are inverted:
It should be:
(al * bh) << 16 + c
instead of:
(ah * bl) << 16 + c
Fixes a number of regressions under
dEQP-GLES31.functional.draw_indirect.compute_interop.large.*
on Freedreno.
Reviewed-by: Rob Clark <robdclark@gmail.com>
For umul_low (al * bl), zero is returned if the low 16-bits word of either
source is zero.
for imadsh_mix16 (ah * bl << 16 + c), c is returned if either 'ah' or 'bl'
is zero.
A couple of nir_search_helpers are added:
is_upper_half_zero() returns true if the highest word of all components of
an integer NIR alu src are zero.
is_lower_half_zero() returns true if the lowest word of all components of
an integer nir alu src are zero.
Reviewed-by: Eric Anholt <eric@anholt.net>
We originally had a single lower_fmod option. In commit 2ab2d2e5, Sam
split 32 and 64-bit lowering into separate flags, with the rationale
that some drivers might want different options there. This left 16-bit
unhandled, so Iago added a lower_fmod16 option in commit ca31df6f.
Now that lower_fmod64 is gone (in favor of nir_lower_doubles and
nir_lower_dmod), we re-combine lower_fmod16 and lower_fmod32 into a
single lower_fmod flag again. I'm not aware of any hardware which
need lowering for one bitsize and not the other.
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
nir_lower_doubles offers a wide variety of fp64 lowering, including
lowering fmod@64. The version there also better handles imprecisions
due to lowered frcp@64. Let's consolidate on one version.
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
This pattern was noticed in glmark's jellyfish scene.
v2: Add inexact qualifier due to NaN behaviour.
Minimal shader-db changes (slightly helped).
Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Reviewed-by: Elie Tournier <tournier.elie@gmail.com>
Add a "lower_bitshift" option, which disables optimizations introducing
bitshifts and lowers ishl by constant to a multiply, so that we don't have
to deal with bitshifts in int_to_float lowering.
Signed-off-by: Jonathan Marek <jonathan@marek.ca>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
This line is no longer relevant now that booleans are 1-bit, and in fact
causes issues (infinite progress loop between algebraic optimizations
and copy prop) with constant vector masks.
No shader-db changes on Intel platforms (Jason).
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
The goal is to avoid having an extra MOV instruction to perform the
saturate. Doing the subtraction first allows the saturate to be applied
to the ADD instruction making the MOV unnecessary. Values generated in
different block and values from non-ALU instructions (e.g., texture
instructions) almost always need the extra MOV.
Multiply instructions are restricted because doing this rearrangement
can interfere with the generation of flrp and ffma instructions.
v2: Now that the final method has been selected, squash three commits
into one.
All Intel platforms has similar results. (Ice Lake shown)
total instructions in shared programs: 17223214 -> 17219386 (-0.02%)
instructions in affected programs: 1524376 -> 1520548 (-0.25%)
helped: 2686
HURT: 26
helped stats (abs) min: 1 max: 32 x̄: 1.44 x̃: 1
helped stats (rel) min: 0.03% max: 16.67% x̄: 0.54% x̃: 0.37%
HURT stats (abs) min: 1 max: 2 x̄: 1.69 x̃: 2
HURT stats (rel) min: 0.33% max: 1.67% x̄: 0.54% x̃: 0.35%
95% mean confidence interval for instructions value: -1.46 -1.36
95% mean confidence interval for instructions %-change: -0.56% -0.50%
Instructions are helped.
total cycles in shared programs: 360811571 -> 360791896 (<.01%)
cycles in affected programs: 103650214 -> 103630539 (-0.02%)
helped: 1557
HURT: 675
helped stats (abs) min: 1 max: 1773 x̄: 41.44 x̃: 16
helped stats (rel) min: <.01% max: 26.77% x̄: 1.37% x̃: 0.64%
HURT stats (abs) min: 1 max: 1513 x̄: 66.44 x̃: 14
HURT stats (rel) min: <.01% max: 46.16% x̄: 2.00% x̃: 0.49%
95% mean confidence interval for cycles value: -14.82 -2.81
95% mean confidence interval for cycles %-change: -0.50% -0.20%
Cycles are helped.
LOST: 2
GAINED: 0
Reviewed-by: Matt Turner <mattst88@gmail.com> [v1]
Reviewed-by: Thomas Helland <thomashelland90@gmail.com>
The value-range tracking pass that is coming is not clever enough to
know that the result of the ffma must be non-negative. Making it that
smart will require quite a bit of work. It might be possible to add a
special case that detects that a whole tree of fadd(fmul(fsat(a),
fneg(fsat(a))), 1.0) cannot be negative.
For cases when the comparison is used in the domain guard for a
square-root (see nir/algebraic: Simplify fsqrt domain guard), the
compare may be converted to a fmax. This patch also handles that case.
All of the affected cases are in DiRT: Showdown.
All Gen7+ platforms had similar results. (Ice Lake shown)
total instructions in shared programs: 17225365 -> 17225303 (<.01%)
instructions in affected programs: 40051 -> 39989 (-0.15%)
helped: 62
HURT: 0
helped stats (abs) min: 1 max: 1 x̄: 1.00 x̃: 1
helped stats (rel) min: 0.07% max: 0.66% x̄: 0.27% x̃: 0.26%
95% mean confidence interval for instructions value: -1.00 -1.00
95% mean confidence interval for instructions %-change: -0.31% -0.22%
Instructions are helped.
total cycles in shared programs: 360842788 -> 360842595 (<.01%)
cycles in affected programs: 1818081 -> 1817888 (-0.01%)
helped: 29
HURT: 22
helped stats (abs) min: 1 max: 206 x̄: 20.66 x̃: 14
helped stats (rel) min: <.01% max: 9.55% x̄: 0.87% x̃: 0.42%
HURT stats (abs) min: 1 max: 108 x̄: 18.45 x̃: 7
HURT stats (rel) min: <.01% max: 4.48% x̄: 0.56% x̃: 0.19%
95% mean confidence interval for cycles value: -14.48 6.91
95% mean confidence interval for cycles %-change: -0.71% 0.21%
Inconclusive result (value mean confidence interval includes 0).
No changes on any other Intel platform.
Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Thomas Helland <thomashelland90@gmail.com>
This change also enables a later change (nir/algebraic: Replace
1-fsat(a) with fsat(1-a)) to affect more shaders.
Almost all of the affected shaders are in Bioshock Infinite, and all of
those shaders all require GLSL 4.10.
All Intel platforms had similar results. (Ice Lake shown)
total instructions in shared programs: 17228584 -> 17228376 (<.01%)
instructions in affected programs: 31438 -> 31230 (-0.66%)
helped: 105
HURT: 0
helped stats (abs) min: 1 max: 5 x̄: 1.98 x̃: 1
helped stats (rel) min: 0.08% max: 1.53% x̄: 0.73% x̃: 0.70%
95% mean confidence interval for instructions value: -2.20 -1.76
95% mean confidence interval for instructions %-change: -0.80% -0.67%
Instructions are helped.
total cycles in shared programs: 360936431 -> 360935690 (<.01%)
cycles in affected programs: 420100 -> 419359 (-0.18%)
helped: 71
HURT: 21
helped stats (abs) min: 1 max: 160 x̄: 19.28 x̃: 10
helped stats (rel) min: <.01% max: 9.78% x̄: 0.95% x̃: 0.48%
HURT stats (abs) min: 1 max: 198 x̄: 29.90 x̃: 10
HURT stats (rel) min: 0.05% max: 8.36% x̄: 1.24% x̃: 0.90%
95% mean confidence interval for cycles value: -16.77 0.66
95% mean confidence interval for cycles %-change: -0.85% -0.06%
Inconclusive result (value mean confidence interval includes 0).
Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Thomas Helland <thomashelland90@gmail.com>
This doesn't make any real difference now, but future work (not in this
series) will add a LOT of ffma patterns. Having to duplicate all of
them for ffma(a, b, c) and ffma(b, a, c) is just terrible.
No shader-db changes on any Intel platform.
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
In a previous verion of this patch, Jason commented,
"Re-associating based on whether or not something has a constant
value of 1.0 seems a bit sneaky. I think it's well within the rules
but it seems like something that could bite you."
That is possibly true. The reassociation will generate different
results if fabs(b) >= 2**24 and fabs(c) < 0.5. The delta increases as
fabs(c) approaches 0.
However, i965 has done this same reassociation indirectly for years.
We would previously allow nir_op_flrp on all pre-Gen11 hardware even
though Gen4 and Gen5 do not have a LRP instruction. Optimizations in
nir_opt_algebraic would convert expressions like a+c(b-a) into flrp(a,
b, c). On Gen7+, the hardware performs the same arithmetic as
a(1-c)+bc. Gen6 seems to implement LRP as a+c(b-a). On Gen4 and
Gen5, we would lower LRP to a sequence of instructions that implement
a(1-c)+bc. The lowering happens after all constant folding, so we
would litterally generate a 1+(-1) instruction sequence in this
scenario: one instruction to load either 1 or -1 in a register, and
another instruction to add either -1 or 1 to it.
This patch just cuts out the middle man. Do the reassociation that
we've always done, but do it explicitly at a time when we can benefit
from other optimizations.
A few cases that were hurt by "nir: Lower flrp(±1, b, c) and flrp(a,
±1, c) differently" are restored by this patch. This includes a few
shaders in ET:QW.
I tried a similar thing for open-coded flrp(-1, b, c), and it hurt
instructions on 35 shaders for ILK without helping any. The helped /
hurt cycles was about even.
No changes on any other Intel platforms.
Iron Lake and GM45 had similar results. (Iron Lake shown)
total instructions in shared programs: 8172020 -> 8164367 (-0.09%)
instructions in affected programs: 1089851 -> 1082198 (-0.70%)
helped: 3285
HURT: 64
helped stats (abs) min: 1 max: 6 x̄: 2.35 x̃: 2
helped stats (rel) min: 0.13% max: 12.00% x̄: 1.15% x̃: 0.83%
HURT stats (abs) min: 1 max: 1 x̄: 1.00 x̃: 1
HURT stats (rel) min: 0.24% max: 0.64% x̄: 0.39% x̃: 0.38%
95% mean confidence interval for instructions value: -2.32 -2.25
95% mean confidence interval for instructions %-change: -1.16% -1.09%
Instructions are helped.
total cycles in shared programs: 188758338 -> 188719974 (-0.02%)
cycles in affected programs: 20004922 -> 19966558 (-0.19%)
helped: 3012
HURT: 477
helped stats (abs) min: 2 max: 142 x̄: 13.41 x̃: 12
helped stats (rel) min: 0.01% max: 6.37% x̄: 0.52% x̃: 0.24%
HURT stats (abs) min: 2 max: 328 x̄: 4.27 x̃: 4
HURT stats (rel) min: <.01% max: 1.55% x̄: 0.14% x̃: 0.11%
95% mean confidence interval for cycles value: -11.38 -10.62
95% mean confidence interval for cycles %-change: -0.46% -0.41%
Cycles are helped.
Reviewed-by: Matt Turner <mattst88@gmail.com>
I tried to be very careful while updating all the various drivers, but I
don't have any of that hardware for testing. :(
i965 is the only platform that sets always_precise = true, and it is
only set true for fragment shaders. Gen4 and Gen5 both set lower_flrp32
only for vertex shaders. For fragment shaders, nir_op_flrp is lowered
during code generation as a(1-c)+bc. On all other platforms 64-bit
nir_op_flrp and on Gen11 32-bit nir_op_flrp are lowered using the old
nir_opt_algebraic method.
No changes on any other Intel platforms.
v2: Add panfrost changes.
Iron Lake and GM45 had similar results. (Iron Lake shown)
total cycles in shared programs: 188647754 -> 188647748 (<.01%)
cycles in affected programs: 5096 -> 5090 (-0.12%)
helped: 3
HURT: 0
helped stats (abs) min: 2 max: 2 x̄: 2.00 x̃: 2
helped stats (rel) min: 0.12% max: 0.12% x̄: 0.12% x̃: 0.12%
Reviewed-by: Matt Turner <mattst88@gmail.com>
These have been popping up more and more with the OpenCL work and other
bits causing extra conversions to/from 64-bit.
Reviewed-by: Karol Herbst <kherbst@redhat.com>
The mali utgard pp doesn't support a sign instruction.
In the ARM offline shader compiler, the sign function is implemented
using sub(gt(0.0, a), lt(0.0, a)).
This is a generic optimization, so implement it in the nir level when
lower_fsign is set, alongside the lowering for isign.
Signed-off-by: Erico Nunes <nunes.erico@gmail.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
All of the affected shaders are in Mad Max. I noticed this while
looking at some other things. I tried a couple similar patterns, but
the affect on cycles was general negative. It may be worth revisiting
this later.
v2: Rebase on 1-bit Boolean changes.
All Gen7+ platforms had similar results. (Skylake shown)
total instructions in shared programs: 15282073 -> 15282053 (<.01%)
instructions in affected programs: 1192 -> 1172 (-1.68%)
helped: 14
HURT: 0
helped stats (abs) min: 1 max: 2 x̄: 1.43 x̃: 1
helped stats (rel) min: 1.16% max: 2.17% x̄: 1.65% x̃: 1.39%
95% mean confidence interval for instructions value: -1.73 -1.13
95% mean confidence interval for instructions %-change: -1.91% -1.38%
Instructions are helped.
total cycles in shared programs: 372595954 -> 372594532 (<.01%)
cycles in affected programs: 11477 -> 10055 (-12.39%)
helped: 14
HURT: 0
helped stats (abs) min: 76 max: 122 x̄: 101.57 x̃: 104
helped stats (rel) min: 7.76% max: 15.62% x̄: 12.94% x̃: 14.78%
95% mean confidence interval for cycles value: -111.05 -92.09
95% mean confidence interval for cycles %-change: -14.90% -10.98%
Cycles are helped.
No changes on any Gen6 or earlier platforms.
Reviewed-by: Matt Turner <mattst88@gmail.com>
All of the affected shaders are in Mad Max. The inner part of the
pattern is itself an open-coded sign(a). I tried using that as a
pattern, but the results were not good. A bunch of shaders were helped
for instructions, but overall cycles, spill, and fills were hurt.
v2: Rebase on 1-bit Boolean changes.
v3: Fix order of copysign() parameters in comments and commit message.
Noticed by Matt.
All Gen7+ platforms had similar results. (Skylake shown)
total instructions in shared programs: 15282141 -> 15282073 (<.01%)
instructions in affected programs: 6106 -> 6038 (-1.11%)
helped: 17
HURT: 0
helped stats (abs) min: 4 max: 4 x̄: 4.00 x̃: 4
helped stats (rel) min: 1.02% max: 2.20% x̄: 1.15% x̃: 1.06%
95% mean confidence interval for instructions value: -4.00 -4.00
95% mean confidence interval for instructions %-change: -1.30% -1.00%
Instructions are helped.
total cycles in shared programs: 372597886 -> 372595954 (<.01%)
cycles in affected programs: 32701 -> 30769 (-5.91%)
helped: 17
HURT: 0
helped stats (abs) min: 6 max: 216 x̄: 113.65 x̃: 118
helped stats (rel) min: 0.40% max: 21.86% x̄: 6.20% x̃: 5.83%
95% mean confidence interval for cycles value: -152.84 -74.45
95% mean confidence interval for cycles %-change: -8.89% -3.51%
Cycles are helped.
No changes on any Gen6 or earlier platforms.
Reviewed-by: Matt Turner <mattst88@gmail.com>
In compute shaders if no derivative group is defined, the derivatives
will always be zero. Specified in NV_compute_shader_derivatives.
To make the check more convenient, add a "info" local variable to the
generated code so we can refer to it in the Python rules. (Jason)
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
The new OR pattern has been seen in the wild and can end up being
generated by GLSLang. Not sure about the other two new patterns but we
may as well throw them in for completeness. While we're here, we can
drop the '@bool' specifier from the one pattern because specifying True
already implies 1-bit which basically implies boolean.
Shader-db results on Kaby Lake:
total instructions in shared programs: 15321227 -> 15321129 (<.01%)
instructions in affected programs: 3594 -> 3496 (-2.73%)
helped: 6
HURT: 0
total cycles in shared programs: 357481321 -> 357479725 (<.01%)
cycles in affected programs: 44109 -> 42513 (-3.62%)
helped: 6
HURT: 0
VkPipeline-DB results on Kaby Lake:
total instructions in shared programs: 3770504 -> 3769734 (-0.02%)
instructions in affected programs: 19058 -> 18288 (-4.04%)
helped: 163
HURT: 0
total cycles in shared programs: 1417583701 -> 1417569727 (<.01%)
cycles in affected programs: 750958 -> 736984 (-1.86%)
helped: 158
HURT: 1
Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>
Now that we have one-bit booleans, we don't need to rely on looking at
parent instructions in order to figure out if a value is a Boolean most
of the time. We can drop these specifiers and now the optimizations
will apply more generally.
Shader-DB results on Kaby Lake:
total instructions in shared programs: 15321168 -> 15321227 (<.01%)
instructions in affected programs: 8836 -> 8895 (0.67%)
helped: 1
HURT: 31
total cycles in shared programs: 357481781 -> 357481321 (<.01%)
cycles in affected programs: 146524 -> 146064 (-0.31%)
helped: 22
HURT: 10
total spills in shared programs: 23675 -> 23673 (<.01%)
spills in affected programs: 11 -> 9 (-18.18%)
helped: 1
HURT: 0
total fills in shared programs: 32040 -> 32036 (-0.01%)
fills in affected programs: 27 -> 23 (-14.81%)
helped: 1
HURT: 0
No change in VkPipeline-DB
Looking at the instructions hurt, a bunch of them seem to be a case
where doing exactly the right thing in NIR ends up doing the wrong-ish
thing in the back-end because flags are dumb. In particular, there's a
case where we have a MUL followed by a CMP followed by a SEL and when we
turn that SEL into an OR, it uses the GRF result of the CMP rather than
the flag result so the CMP can't be merged with the MUL. Those shaders
appear to schedule better according to the cycle estimates so I guess
it's a win? Also it helps spilling in one Car Chase compute shader.
Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>
No shader-db changes on any Intel platform.
v2: Use a loop to generate patterns. Suggested by Jason.
v3: Fix a copy-and-paste bug in the extract_[ui] of ishl loop that would
replace an extract_i8 with and extract_u8. This broke ~180 tests. This
bug was introduced in v2.
Reviewed-by: Matt Turner <mattst88@gmail.com> [v1]
Reviewed-by: Dylan Baker <dylan@pnwbakers.com> [v2]
Acked-by: Jason Ekstrand <jason@jlekstrand.net> [v2]
No shader-db changes on any Intel platform.
v2: Use a loop to generate patterns. Suggested by Jason.
v3: Fix a copy-and-paste bug in the extract_[ui] of ishl loop that would
replace an extract_i8 with and extract_u8. This broke ~180 tests. This
bug was introduced in v2.
Reviewed-by: Matt Turner <mattst88@gmail.com> [v1]
Reviewed-by: Dylan Baker <dylan@pnwbakers.com> [v2]
Acked-by: Jason Ekstrand <jason@jlekstrand.net> [v2]
No shader-db changes on any Intel platform.
Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Dylan Baker <dylan@pnwbakers.com>
Acked-by: Jason Ekstrand <jason@jlekstrand.net>
v2 (Topi):
- Make bit-size handling order be 16-bit, 32-bit, 64-bit
- Clamp lower exponent range at -28 instead of -30.
Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Shader-db results on Kaby Lake:
total instructions in shared programs: 15225213 -> 15222365 (-0.02%)
instructions in affected programs: 43524 -> 40676 (-6.54%)
helped: 203
HURT: 0
Lots of shaders in Shadow Warrior had this pattern along with Deus Ex,
Civ, Shadow of Mordor, and several others.
Reviewed-by: Kristian H. Kristensen <hoegsberg@chromium.org>
This broke piles of image load store tests (179 failures on CI,
mesa_master build #15546, previous build right before this landed
was green). I'd rather not leave the tree on fire over the weekend,
so let's revert for now, and we can figure out what happened next week.
No shader-db changes on any Intel platform.
v2: Use a loop to generate patterns. Suggested by Jason.
Reviewed-by: Matt Turner <mattst88@gmail.com> [v1]
Reviewed-by: Dylan Baker <dylan@pnwbakers.com>
Acked-by: Jason Ekstrand <jason@jlekstrand.net>
No shader-db changes on any Intel platform.
v2: Use a loop to generate patterns. Suggested by Jason.
Reviewed-by: Matt Turner <mattst88@gmail.com> [v1]
Reviewed-by: Dylan Baker <dylan@pnwbakers.com>
Acked-by: Jason Ekstrand <jason@jlekstrand.net>
No shader-db changes on any Intel platform.
Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Dylan Baker <dylan@pnwbakers.com>
Acked-by: Jason Ekstrand <jason@jlekstrand.net>
Not complete, mostly just adding things as I encounter them in CTS. But
not getting far enough yet to hit most of the OpenCL.std instructions.
Anyway, this is better than nothing and covers the most common builtins.
v2: add hadd proof from Jason
move some of the lowering into opt_algebraic and create new nir opcodes
simplify nextafter lowering
fix normalize lowering for inf
rework upsample to use nir_pack_bits
add missing files to build systems
v3: split lines of iadd/sub_sat expressions
Signed-off-by: Karol Herbst <kherbst@redhat.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Optimize a situation where we only need lower 32 bits from 64 bit
result.
Signed-off-by: Sagar Ghuge <sagar.ghuge@intel.com>
Suggested-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
On Gen 8 and 9, "mul" instruction supports 64 bit destination type. We
can reduce our 64x64 int multiplication from 4 instructions to 3.
Also instead of emitting two mul instructions, we can emit single mul
instuction and extract low/high 32 bits from 64 bit result for
[i/u]mulExtended
v2: 1) Allow lower_mul_high64 to use new opcode (Jason Ekstrand)
2) Add lower_mul_2x32_64 flag (Matt Turner)
3) Remove associative property as bit size is different (Connor
Abbott)
v3: Fix indentation and variable naming convention (Jason Ekstrand)
Signed-off-by: Sagar Ghuge <sagar.ghuge@intel.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
I noticed this while looking at a shader that was affected by Tim's
"more loop unrolling" series.
In review, Tim Arceri asked:
> Why the hurt on Gen6+ is this something that should be in the late
> optimisations pass?
As far as I can tell, it's just because our scheduler is terrible. In
all the fragment shaders that I looked at (some hurt shaders were from
other stages), only one of the SIMD8 or SIMD16 version would be hurt.
In many of those case, the other SIMD width is improved (e.g.,
shaders/closed/steam/brutal-legend/3990.shader_test).
Often it looks like the scheduler decides to differently schedule a SEND
the occurs somewhere early in the shader. Once that happens, everything
is different.
I looked at one vertex shader that was hurt (from Goat Simulator). In
that case, both the floor and fract are used. The optimization
eliminates the add, and it should allow better scheduling. In the area
of the FRC and RNDD instructions, the scheduler does the right thing.
However, later in the shader a MAD and and ADD get scheduled
differently, and that makes it slightly worse.
In light of this, I tried adding some "is_used_once" mark-up, and that
did not fix all the cycles regressions. It also did a lot more harm
than good on SKL (helped 82 vs. hurt 241).
All Gen6+ platforms had similar results. (Skylake shown)
total instructions in shared programs: 15437001 -> 15435259 (-0.01%)
instructions in affected programs: 213651 -> 211909 (-0.82%)
helped: 988
HURT: 0
helped stats (abs) min: 1 max: 27 x̄: 1.76 x̃: 1
helped stats (rel) min: 0.15% max: 11.54% x̄: 1.14% x̃: 0.59%
95% mean confidence interval for instructions value: -1.89 -1.63
95% mean confidence interval for instructions %-change: -1.23% -1.05%
Instructions are helped.
total cycles in shared programs: 383007378 -> 382997063 (<.01%)
cycles in affected programs: 1650825 -> 1640510 (-0.62%)
helped: 679
HURT: 302
helped stats (abs) min: 1 max: 348 x̄: 23.39 x̃: 14
helped stats (rel) min: 0.04% max: 28.77% x̄: 1.61% x̃: 0.98%
HURT stats (abs) min: 1 max: 250 x̄: 18.43 x̃: 7
HURT stats (rel) min: 0.04% max: 25.86% x̄: 1.41% x̃: 0.53%
95% mean confidence interval for cycles value: -13.05 -7.98
95% mean confidence interval for cycles %-change: -0.86% -0.50%
Cycles are helped.
Iron Lake and GM45 had similar results. (GM45 shown)
total instructions in shared programs: 5043616 -> 5043010 (-0.01%)
instructions in affected programs: 119691 -> 119085 (-0.51%)
helped: 432
HURT: 0
helped stats (abs) min: 1 max: 27 x̄: 1.40 x̃: 1
helped stats (rel) min: 0.10% max: 8.11% x̄: 0.66% x̃: 0.39%
95% mean confidence interval for instructions value: -1.58 -1.23
95% mean confidence interval for instructions %-change: -0.72% -0.59%
Instructions are helped.
total cycles in shared programs: 128139812 -> 128135762 (<.01%)
cycles in affected programs: 3829724 -> 3825674 (-0.11%)
helped: 602
HURT: 0
helped stats (abs) min: 2 max: 486 x̄: 6.73 x̃: 6
helped stats (rel) min: 0.02% max: 4.85% x̄: 0.19% x̃: 0.10%
95% mean confidence interval for cycles value: -8.40 -5.05
95% mean confidence interval for cycles %-change: -0.22% -0.16%
Cycles are helped.
Reviewed-by: Elie Tournier <tournier.elie@gmail.com>
I have not investigated the result of doing this during code
generation. That should be possible, but it would be a bit more
effort.
All Gen6+ platforms had nearly identical results. (Skylake shown)
total cycles in shared programs: 370961508 -> 370961367 (<.01%)
cycles in affected programs: 5174 -> 5033 (-2.73%)
helped: 2
HURT: 0
Iron Lake and GM45 had similar results. (Iron Lake shown)
total instructions in shared programs: 8206587 -> 8206589 (<.01%)
instructions in affected programs: 1325 -> 1327 (0.15%)
helped: 0
HURT: 2
total cycles in shared programs: 187657422 -> 187657428 (<.01%)
cycles in affected programs: 11566 -> 11572 (0.05%)
helped: 0
HURT: 2
This change has almost no effect right now. However, removing this
patch (but leaving the patch "intel/fs: Generate if instructions with
inverted conditions") after adding a patch that removes !(a < b) -> (a
>= b) optimizations (like
https://patchwork.freedesktop.org/patch/264787/) has the following
results on Skylake:
Skylake
total instructions in shared programs: 15071804 -> 15071806 (<.01%)
instructions in affected programs: 640 -> 642 (0.31%)
helped: 0
HURT: 2
total cycles in shared programs: 369914348 -> 369916569 (<.01%)
cycles in affected programs: 27900 -> 30121 (7.96%)
helped: 4
HURT: 15
helped stats (abs) min: 2 max: 112 x̄: 30.00 x̃: 3
helped stats (rel) min: 0.28% max: 12.28% x̄: 3.34% x̃: 0.40%
HURT stats (abs) min: 2 max: 758 x̄: 156.07 x̃: 81
HURT stats (rel) min: 0.20% max: 74.30% x̄: 16.29% x̃: 16.91%
95% mean confidence interval for cycles value: 12.68 221.11
95% mean confidence interval for cycles %-change: 3.09% 21.23%
Cycles are HURT.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
This is a common pattern from HLSL->SPIRV translation
and supported in HW by all current NIR backends.
vkpipeline-db results anv (SKL):
total instructions in shared programs: 6403130 -> 6402380 (-0.01%)
instructions in affected programs: 204084 -> 203334 (-0.37%)
helped: 208
HURT: 0
total cycles in shared programs: 1915629582 -> 1918198408 (0.13%)
cycles in affected programs: 1158892682 -> 1161461508 (0.22%)
helped: 107
HURT: 86
shader-db results on i965 (KBL):
total instructions in shared programs: 15284592 -> 15284568 (<.01%)
instructions in affected programs: 81683 -> 81659 (-0.03%)
helped: 24
HURT: 0
total cycles in shared programs: 375013622 -> 375013932 (<.01%)
cycles in affected programs: 40169618 -> 40169928 (<.01%)
helped: 13
HURT: 9
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
The idea here is to reassociate a * (b * c) into (a * c) * b, when
b is a non-constant value, but a and c are constants, allowing them
to be combined.
But nothing was enforcing that 'b' must be non-constant, which meant
that running opt_algebraic in a loop would never terminate if the IR
contained non-folded constant expressions like 256 * 0.5 * 2. Normally,
we call constant folding in such a loop too, but IMO it's better for
nir_opt_algebraic to be robust and not rely on that.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=109581
Fixes: 32e266a9a5 i965: Compile fp64 funcs only if we do not have 64-bit hardware support
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
The quotation marks around 1.0 cause it to be treated as a string
instead of a floating point value. The generator then treats it as an
arbitrary variable replacement, so any iand involving a ('ineg', ('b2i',
a)) matches.
v2: Remove misleading comment about sized literals (suggested by
Timothy). Add assertion that the name of a varible is entierly
alphabetic (suggested by Jason).
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Tested-by: Timothy Arceri <tarceri@itsqueeze.com> [v1]
Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com> [v1]
Fixes: 6bcd2af086 ("nir/algebraic: Add some optimizations for D3D-style Booleans")
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=109075
D3D Booleans use a 32-bit 0/-1 representation. Because this previously
matched NIR exactly, we didn't have to really optimize for it. Now that
we have 1-bit Booleans, we need some specific optimizations to chew
through the D3D12-style Booleans.
Shader-db results on Kaby Lake:
total instructions in shared programs: 15136811 -> 14967944 (-1.12%)
instructions in affected programs: 2457021 -> 2288154 (-6.87%)
helped: 8318
HURT: 10
total cycles in shared programs: 373544524 -> 359701825 (-3.71%)
cycles in affected programs: 151029683 -> 137186984 (-9.17%)
helped: 7749
HURT: 682
total loops in shared programs: 4431 -> 4399 (-0.72%)
loops in affected programs: 32 -> 0
helped: 21
HURT: 0
total spills in shared programs: 10290 -> 10051 (-2.32%)
spills in affected programs: 2532 -> 2293 (-9.44%)
helped: 18
HURT: 18
total fills in shared programs: 22203 -> 21732 (-2.12%)
fills in affected programs: 3319 -> 2848 (-14.19%)
helped: 18
HURT: 18
Note that a large chunk of the improvement fixing regressions caused by
switching to 1-bit Booleans. Previously, our ability to optimize D3D
booleans was improved by using the D3D representation directly in NIR.
Now that NIR does 1-bit bools, we need a few more optimizations.
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Reviewed-by: Eric Anholt <eric@anholt.net>
Tested-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Tested-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
This just makes it nicely scale across bit sizes.
Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Tested-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Later in this series, bool is not going to imply 32-bit.
Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Tested-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
This was originally added for the out-of-tree Mali driver but I think
we've all agreed it's easy enough for them to just do in their back-end.
Reviewed-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Tested-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Shader-db results on Kaby Lake:
total instructions in shared programs: 15072525 -> 15072525 (0.00%)
instructions in affected programs: 0 -> 0
helped: 0
HURT: 0
This helps prevent regressions in later commits.
Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Tested-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
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>