We turn these from bcsel into inot/b2f combos in order for other
optimisation passes to get further. Once we have finished turn
the ones that remain and are used in more than a single expression
back into a bcsel.
On BDW:
total instructions in shared programs: 13060965 -> 13060297 (-0.01%)
instructions in affected programs: 835701 -> 835033 (-0.08%)
helped: 670
HURT: 2
total cycles in shared programs: 256599536 -> 256598006 (-0.00%)
cycles in affected programs: 114655488 -> 114653958 (-0.00%)
helped: 419
HURT: 240
LOST: 0
GAINED: 1
The 2 HURT is because inserting bcsel creates the only use of
const 1.0 in two shaders from tri-of-friendship-and-madness.
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
This sequence shows up The Talos Principal, at least under Vulkan,
and prevents loop analysis from properly computing trip counts in a
few loops.
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
The previous power-of-two rules didn't catch idiv (because i965 doesn't
set lower_idiv) and imod cases. The udiv and umod cases should have
been caught, but I included them for orthogonality.
This fixes silly code observed from compute shaders with local_size_[xy]
= 1.
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=98299
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
According to Connor, it's safe to assume that the first operand of
bcsel, as well as the operand of b2f and b2i, must be well formed
booleans.
https://lists.freedesktop.org/archives/mesa-dev/2016-August/125658.html
With the previous improvements to a@bool handling, this now has no
change in shader-db instruction counts on Broadwell.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Some shaders include code that looks like:
uniform int i;
uniform vec4 bones[...];
foo(bones[i * 3], bones[i * 3 + 1], bones[i * 3 + 2]);
CSE would do some work on this:
x = i * 3
foo(bones[x], bones[x + 1], bones[x + 2]);
The compiler may then add '<< 4 + base' to the index calculations.
This results in expressions like
x = i * 3
foo(bones[x << 4], bones[(x + 1) << 4], bones[(x + 2) << 4]);
Just rearranging the math to produce (i * 48) + 16 saves an
instruction, and it allows CSE to do more work.
x = i * 48;
foo(bones[x], bones[x + 16], bones[x + 32]);
So, ~6 instructions becomes ~3.
Some individual shader-db results look pretty bad. However, I have a
really, really hard time believing the change in estimated cycles in,
for example, 3dmmes-taiji/51.shader_test after looking that change in
the generated code.
G45
total instructions in shared programs: 4020840 -> 4010070 (-0.27%)
instructions in affected programs: 177460 -> 166690 (-6.07%)
helped: 894
HURT: 0
total cycles in shared programs: 98829000 -> 98784990 (-0.04%)
cycles in affected programs: 3936648 -> 3892638 (-1.12%)
helped: 894
HURT: 0
Ironlake
total instructions in shared programs: 6418887 -> 6408117 (-0.17%)
instructions in affected programs: 177460 -> 166690 (-6.07%)
helped: 894
HURT: 0
total cycles in shared programs: 143504542 -> 143460532 (-0.03%)
cycles in affected programs: 3936648 -> 3892638 (-1.12%)
helped: 894
HURT: 0
Sandy Bridge
total instructions in shared programs: 8357887 -> 8339251 (-0.22%)
instructions in affected programs: 432715 -> 414079 (-4.31%)
helped: 2795
HURT: 0
total cycles in shared programs: 118284184 -> 118207412 (-0.06%)
cycles in affected programs: 6114626 -> 6037854 (-1.26%)
helped: 2478
HURT: 317
Ivy Bridge
total instructions in shared programs: 7669390 -> 7653822 (-0.20%)
instructions in affected programs: 388234 -> 372666 (-4.01%)
helped: 2795
HURT: 0
total cycles in shared programs: 68381982 -> 68263684 (-0.17%)
cycles in affected programs: 1972658 -> 1854360 (-6.00%)
helped: 2458
HURT: 307
Haswell
total instructions in shared programs: 7082636 -> 7067068 (-0.22%)
instructions in affected programs: 388234 -> 372666 (-4.01%)
helped: 2795
HURT: 0
total cycles in shared programs: 68282020 -> 68164158 (-0.17%)
cycles in affected programs: 1891820 -> 1773958 (-6.23%)
helped: 2459
HURT: 261
Broadwell
total instructions in shared programs: 9002466 -> 8985875 (-0.18%)
instructions in affected programs: 658784 -> 642193 (-2.52%)
helped: 2795
HURT: 5
total cycles in shared programs: 78503092 -> 78450404 (-0.07%)
cycles in affected programs: 2873304 -> 2820616 (-1.83%)
helped: 2275
HURT: 415
Skylake
total instructions in shared programs: 9156978 -> 9140387 (-0.18%)
instructions in affected programs: 682625 -> 666034 (-2.43%)
helped: 2795
HURT: 5
total cycles in shared programs: 75591392 -> 75550574 (-0.05%)
cycles in affected programs: 3192120 -> 3151302 (-1.28%)
helped: 2271
HURT: 425
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Thomas Helland <thomashelland90@gmail.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
I noticed this when I tried to do frexp(float(some_unsigned)) in the
ir_unop_find_lsb lowering pass. The code generated for frexp() uses
fabs, and this resulted in an extra instruction. Ultimately I ended up
not using frexp.
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
This was appearing in vc4 VS/CS in mupen64, due to vertex attrib lowering
producing some constants that were getting compared.
total instructions in shared programs: 112276 -> 112198 (-0.07%)
instructions in affected programs: 2239 -> 2161 (-3.48%)
total estimated cycles in shared programs: 283102 -> 283038 (-0.02%)
estimated cycles in affected programs: 2365 -> 2301 (-2.71%)
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
While mathematically correct, these two optimizations result in an
expression with substantially lower precision than the original. For any
positive finite floating-point value, log2(x) is well-defined and finite.
More precisely, it is in the range [-150, 150] so any sum of logarithms
log2(a) + log2(b) is also well-defined and finite as long as a and b are
both positive and finite. However, if a and b are either very small or
very large, their product may get flushed to infinity or zero causing
log2(a * b) to be nowhere close to log2(a) + log2(b).
This imprecision was causing incorrect rendering in Talos Principal because
part of its HDR rendering process involves doing 8 texture operations,
clamping the result to [0, 65000], taking a dot-product with a constant,
and then taking the log2. This is done 6 or 8 times and summed to produce
the final result which is written to a red texture. In cases where you
have a region of the screen that is very dark, it can end up getting a
result value of -inf which is not what is intended.
Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=96425
Cc: "11.1 11.2 12.0" <mesa-stable@lists.freedesktop.org>
Some optimizations, like converting integer multiply/divide into left/
right shifts, have additional constraints on the search expression.
Like requiring that a variable is a constant power of two. Support
these cases by allowing a fxn name to be appended to the search var
expression (ie. "a#32(is_power_of_two)").
Signed-off-by: Rob Clark <robclark@freedesktop.org>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
The i965 driver has its own pass for fusing mul+add combinations that's
much smarter than what nir_opt_algebraic can do so we don't want to get the
nir_opt_algebraic one just because we didn't set lower_ffma.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Split 32-bit and 64-bit fmod lowering as the drivers might need to
lower them separately inside NIR depending on the HW support.
Signed-off-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
There is no sense in having the double version of ldexp take a 64-bit
integer. Instead, let's just take a 32-bit int all the time. This also
matches what GLSL does where both variants of ldexp take a regular integer
for the exponent argument.
Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
Some hardware (i965 on Broadwell generation, for example) does not support
natively the execution of lrp instruction with double arguments.
Add 'lower_flrp64' flag to lower this instruction in that case.
v2:
- Rename lower_flrp_double to lower_flrp64 (Jason)
- Fix typo (Jason)
- Adapt the code to define bit_size information in the opcodes.
Signed-off-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
A later patch will add lower_flrp64 option to NIR.
Signed-off-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
The algorithm used is different from both the naive suggestion from the
GLSL spec and the one used in GLSL IR today. Unfortunately, the GLSL IR
implementation that we have today doesn't handle denormals (for those that
care) or the case where the float source is +-inf.
Reviewed-by: Matt Turner <mattst88@gmail.com>
Dolphin uses them a lot. Range tracking would be better in the long term,
but this two lines works fine for now.
Signed-off-by: Markus Wick <markus@selfnet.de>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Many shaders contain expression trees of the form:
const_1 * (value * const_2)
Reorganizing these to
(const_1 * const_2) * value
will allow constant folding to combine the constants. Sometimes, these
constants are 2 and 0.5, so we can remove a multiply altogether. Other
times, it can create more immediate constants, which can actually hurt.
Finding a good balance here is tricky. While much more could be done,
this simple patch seems to have a lot of positive benefit while having
a low downside.
shader-db results on Broadwell:
total instructions in shared programs: 8963768 -> 8961369 (-0.03%)
instructions in affected programs: 438318 -> 435919 (-0.55%)
helped: 1502
HURT: 245
total cycles in shared programs: 71527354 -> 71421516 (-0.15%)
cycles in affected programs: 11541788 -> 11435950 (-0.92%)
helped: 3445
HURT: 1224
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eduardo Lima Mitev <elima@igalia.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Oddly, this did not affect the shader where I first noticed the pattern.
That particular shader doesn't get its if-statement converted to a bcsel
because there are two assignments in the else-statement. This led to me
submitting https://bugs.freedesktop.org/show_bug.cgi?id=94747.
shader-db results:
Sandy Bridge
total instructions in shared programs: 8467384 -> 8467069 (-0.00%)
instructions in affected programs: 36594 -> 36279 (-0.86%)
helped: 46
HURT: 0
total cycles in shared programs: 117573448 -> 117568518 (-0.00%)
cycles in affected programs: 339114 -> 334184 (-1.45%)
helped: 46
HURT: 0
Ivy Bridge / Haswell / Broadwell / Skylake:
total instructions in shared programs: 7774258 -> 7773999 (-0.00%)
instructions in affected programs: 30874 -> 30615 (-0.84%)
helped: 46
HURT: 0
total cycles in shared programs: 65739190 -> 65734530 (-0.01%)
cycles in affected programs: 180380 -> 175720 (-2.58%)
helped: 45
HURT: 1
No change on G45 or Ironlake.
I also tried these expressions, but none of them affected any shaders in
shader-db:
(('bcsel', a, 'a@bool', 'b@bool'), ('ior', a, b)),
(('bcsel', a, 'b@bool', False), ('iand', a, b)),
(('bcsel', a, 'b@bool', 'a@bool'), ('iand', a, b)),
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Many of our optimizations, while great for cutting shaders down to size,
aren't really precision-safe. This commit tries to flag all of the
inexact floating-point optimizations so they don't get run on values that
are flagged "exact". It's a bit conservative and maybe flags some safe
optimizations as unsafe but that's better than missing one.
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
The previous transformation got the arguments to fmin backwards. When NaNs
are involved, the GLSL min/max aren't commutative so it matters.
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
The fxor opcode is required to return 1.0f or 0.0f but the input variable
may not be 1.0f or 0.0f.
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
No shader-db changes, but this is symmetric with the previous commit.
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
This also prevented some regressions with other patches in my local
tree.
Broadwell / Skylake
total instructions in shared programs: 8980835 -> 8980833 (-0.00%)
instructions in affected programs: 45 -> 43 (-4.44%)
helped: 1
HURT: 0
total cycles in shared programs: 70077904 -> 70077900 (-0.00%)
cycles in affected programs: 122 -> 118 (-3.28%)
helped: 1
HURT: 0
No changes on earlier platforms.
v2: Modify the comments to look more like a proof.
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
On Intel platforms that don't set lower_flrp, using bcsel instead of
flrp seems to be a small amount worse. On those platforms, the use of
flrp, bcsel, and multiply of b2f is still an active area of research.
In review, Matt suggested this is because bcsel turns into CMP+SEL, and
because of the flag register we can't schedule instructions well.
shader-db results:
G4X / Ironlake
total instructions in shared programs: 4016538 -> 4012279 (-0.11%)
instructions in affected programs: 161556 -> 157297 (-2.64%)
helped: 1077
HURT: 1
total cycles in shared programs: 84328296 -> 84315862 (-0.01%)
cycles in affected programs: 4174570 -> 4162136 (-0.30%)
helped: 926
HURT: 53
Unsurprisingly, no changes on later platforms.
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
No shader-db changes, but does recognize some extract_u16 which enables
the next patch to optimize some code.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
Two shaders that appear in Unigine benchmarks (Heaven and Valley) unpack
three bytes from an integer and convert each into a float:
float((val >> 16u) & 0xffu)
float((val >> 8u) & 0xffu)
float((val >> 0u) & 0xffu)
Instead of shifting, masking, and type converting like this:
shr(8) g15<1>UD g25<8,8,1>UD 0x00000010UD
and(8) g16<1>UD g15<8,8,1>UD 0x000000ffUD
mov(8) g17<1>F g16<8,8,1>UD
shr(8) g18<1>UD g25<8,8,1>UD 0x00000008UD
and(8) g19<1>UD g18<8,8,1>UD 0x000000ffUD
mov(8) g20<1>F g19<8,8,1>UD
and(8) g21<1>UD g25<8,8,1>UD 0x000000ffUD
mov(8) g22<1>F g21<8,8,1>UD
i965 can simply extract a byte and convert to float in a single
instruction:
mov(8) g17<1>F g25.2<32,8,4>UB
mov(8) g20<1>F g25.1<32,8,4>UB
mov(8) g22<1>F g25.0<32,8,4>UB
This patch implements the first step: recognizing byte extraction. A
later patch will optimize out the conversion to float.
instructions in affected programs: 28568 -> 27450 (-3.91%)
helped: 7
cycles in affected programs: 210076 -> 203144 (-3.30%)
helped: 7
This patch decreases the number of instructions in the two Unigine
programs by:
#1721: 4520 -> 4374 instructions (-3.23%)
#1706: 3752 -> 3582 instructions (-4.53%)
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
Helps 11 shaders in UnrealEngine4 demos.
I seriously hope they would have given us bitfieldReverse() if we
exposed GL 4.0 (but we do expose ARB_gpu_shader5, so why not use that
anyway?).
instructions in affected programs: 4875 -> 4633 (-4.96%)
cycles in affected programs: 270516 -> 244516 (-9.61%)
I suspect there's a *lot* of room to improve nir_search/opt_algebraic's
handling of this. We'd actually like to match, e.g., step2 by matching
step1 once and then doing a pointer comparison for the second instance
of step1, but unfortunately we generate an enormous tuple for instead.
The .text size increases by 6.5% and the .data by 17.5%.
text data bss dec hex filename
22957 45224 0 68181 10a55 nir_libnir_la-nir_opt_algebraic.o
24461 53160 0 77621 12f35 nir_libnir_la-nir_opt_algebraic.o
I'd be happy to remove this if Unreal4 uses bitfieldReverse() if it is
in a GL 4.0 context once we expose GL 4.0.
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>