I noticed some unnecessary 64-bit ints in shaders that were using doubles.
Perhaps there's a different missing optimization that should run on the
actual pack/unpack instructions before they're lowered, or maybe I'm just
lowering them too early, but these seem simple enough that we might want
them even for hand-rolled pack/unpack pairs.
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27314>
The constant-folding definition and comments say that it takes the high
16 bits of the first source and low 16 bits of the second source, but
actually it's the opposite. The algebraic optimization, which actually
happens and needs to be correct, was correct but the comment above it
was wrong.
Note that in the way we use it when lowering multiplications, the
ordering doesn't matter.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22075>
The ffloor@64 case, which lowers to use ffract, is already ignored if
nir_lower_dfract is set. Do the same thing for ftrunc@64 and ffract@64
and let nir_lower_doubles take care of them directly instead.
Reviewed-by: Caio Oliveira <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/28702>
These don't seem useful, since they're already done in the early optimizations.
Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Georg Lehmann <dadschoorse@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27335>
No driver supports urol/uror on all bit sizes. Intel gen11+ only for 16
and 32 bit, Nvidia GV100+ only for 32 bit. Etnaviv can support it on 8,
16 and 32 bit.
Also turn the `lower` into a `has` option as only two drivers actually
support `uror` and `urol` at this momemt.
Fixes crashes with CL integer_rotate on iris and nouveau since we emit
urol for `rotate`.
v2: always lower 64 bit
Fixes: fe0965afa6 ("spirv: Don't use libclc for rotate")
Signed-off-by: Karol Herbst <kherbst@redhat.com>
Reviewed-by (Intel and nir): Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: David Heidelberg <david.heidelberg@collabora.com>
Acked-by: Yonggang Luo <luoyonggang@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27090>
Since nir_opt_algebraic runs on its own results, if the driver doesn't
have [su]dot_4x8_[ui]add then the [su]dot_4x8_[ui]add lowering rules
will kick in and lower that to what we had originally.
Reviewed-by: Jesse Natalie <jenatali@microsoft.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26533>
Right now some drivers are doing dsign lowering in GLSL and haven't had to
have a NIR path due to there not being a corresponding vulkan driver. We
want this in NIR now so that we can retire that batch of GLSL lowering
code.
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Acked-by: Faith Ekstrand <faith.ekstrand@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/25777>
These patterns are broken in the following scenario:
%1 = f2fmp %0
%2 = fddx %1
%3 = ... // non quad uniform
if %3 {
%4 = f2f32 %2
...
}
Which would turn into
%3 = ...
if %3 {
%4 = fddx %0
...
}
Yet another example that shows why derivative instructions should be
be intrinsics, not alu.
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Reviewed-by: Emma Anholt <emma@anholt.net>
Cc: mesa-stable
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/25014>
Undef will be replaced by NaN whenever it leads to elimination of FP
instructions. This implements the elimination part.
Reviewed-by: Faith Ekstrand <faith.ekstrand@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24792>
Passes dEQP-VK.spirv_assembly.*opquantize*.
Unlike the DXIL lowering, this should correctly handle NaNs. (I belive Dozen has
a bug here that is masked by running constant folding early and poor CTS
coverage.) It is also faster than the DXIL lowering for hardware that supports
f2f16 conversions natively. It is not as good as a backend implementation that
could flush-to-zero in hardware... but for a debug instruction it should be more
than good enough.
It might be slightly better to multiply with 0.0 to get the appropriate zero,
but NIR really likes optimizing that out ...
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Reviewed-by: Georg Lehmann <dadschoorse@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24616>
These are mostly just obvious patterns that somebody will eventually
want to add.
DG2, Tiger Lake, Ice Lake, Skylake, Broadwell, and Haswell had similar
results (Ice Lake shown)
total instructions in shared programs: 20570033 -> 20570026 (<.01%)
instructions in affected programs: 7363 -> 7356 (-0.10%)
helped: 6 / HURT: 0
total cycles in shared programs: 902118781 -> 902118854 (<.01%)
cycles in affected programs: 419132 -> 419205 (0.02%)
helped: 4 / HURT: 2
DG2, Tiger Lake, Ice Lake, and Skylake had similar results (Ice Lake shown)
Totals:
Instrs: 152819500 -> 152819380 (-0.00%)
Cycles: 15014627187 -> 15014624437 (-0.00%)
Totals from 115 (0.02% of 662497) affected shaders:
Instrs: 28963 -> 28843 (-0.41%)
Cycles: 404582 -> 401832 (-0.68%)
Reviewed-by: Matt Turner <mattst88@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19968>
v2: Fix a copy-and-paste bug s/('find_lsb', a)/a/ in the patterns. See
piglit!819.
DG2, Tiger Lake, Ice Lake, Skylake, and Broadwell had similar results (Ice Lake shown)
total instructions in shared programs: 20570063 -> 20570033 (<.01%)
instructions in affected programs: 452 -> 422 (-6.64%)
helped: 30 / HURT: 0
total cycles in shared programs: 902118723 -> 902118781 (<.01%)
cycles in affected programs: 1762 -> 1820 (3.29%)
helped: 0 / HURT: 29
DG2, Tiger Lake, Ice Lake, and Skylake had similar results (Ice Lake shown)
Totals:
Instrs: 152819969 -> 152819500 (-0.00%)
Cycles: 15014628652 -> 15014627187 (-0.00%); split: -0.00%, +0.00%
Totals from 469 (0.07% of 662497) affected shaders:
Instrs: 7644 -> 7175 (-6.14%)
Cycles: 31787 -> 30322 (-4.61%); split: -4.90%, +0.29%
Reviewed-by: Matt Turner <mattst88@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19968>
I was seeing u2u64 still in my final shader after pack/unpack were
lowered, which sounds to me like some other optimizations are missing
for detecting the post-lowering pack/unpack patterns, but let's at
least add some patterns for the simple cases.
Reviewed-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/23173>
v2: Require that the constant value be representable as either uint16_t
or int16_t. Suggested by Matt.
v3: Remove redundant patterns. Noticed by Matt.
shader-db:
DG2
total instructions in shared programs: 23103767 -> 23103577 (<.01%)
instructions in affected programs: 51822 -> 51632 (-0.37%)
helped: 98 / HURT: 15
total cycles in shared programs: 842347714 -> 842380017 (<.01%)
cycles in affected programs: 1942595 -> 1974898 (1.66%)
helped: 97 / HURT: 32
Nearly all of the affected shaders (around 9,900) are shaders in
Cyberpunk 2077. It's about an even split between vertex and fragment
shaders. The majority of the remaining affected shaders (3,600) are
from Strange Brigade. This was also a nearly even split between
fragment and vertex.
All but two of the lost shaders are SIMD32 fragment shaders in
Cyberpunk 2077. The other two are SIMD32 fragment shaders in Dota2.
fossil-db:
DG2
Instructions in all programs: 196379107 -> 196248608 (-0.1%)
helped: 13467 / HURT: 1210
Cycles in all programs: 13931355281 -> 13929955971 (-0.0%)
helped: 11801 / HURT: 2922
Lost: 90
Reviewed-by: Matt Turner <mattst88@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/23262>
There should not be any isub at this point due to lowerings that happened
ages before getting to late algebraic.
shader-db:
DG2
total instructions in shared programs: 23103769 -> 23103767 (<.01%)
instructions in affected programs: 65 -> 63 (-3.08%)
helped: 1 / HURT: 0
total cycles in shared programs: 842348074 -> 842347714 (<.01%)
cycles in affected programs: 28572 -> 28212 (-1.26%)
helped: 3 / HURT: 0
One compute shader in Assassin's Creed Odyssey was affected.
fossil-db:
DG2
Instructions in all programs: 196400668 -> 196400676 (+0.0%)
helped: 8 / HURT: 5
Cycles in all programs: 13931740724 -> 13931758003 (+0.0%)
helped: 8 / HURT: 7
Reviewed-by: Matt Turner <mattst88@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/23262>
The DXIL backend would like to distinguish between casts to 16-bit
that must cast, vs those that may. If a shader only ever produces
16-bit types from mediump casts and ALU ops on those values, then
the resulting shader can be annotated with DXIL's min-precision
qualifier, basically telling the driver to use 16-bit precision if
it's faster for them. If it uses concrete 16-bit casts, or loads/
stores to externally-visible memory, then it must use the "native"
16-bit flag, which is not supported on all hardware.
Acked-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/23344>
Otherwise this can cause optimizations to fight resulting in infinite
optimization loops with opt_algebraic, constant_folding, and copy_prop.
Fixes: 368be872 ("nir/algebraic: shrink 64-bit bitwise operations with 0/-1 constant half")
Reviewed-by: Emma Anholt <emma@anholt.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/23192>
To be helpful, the thing inside the fsat has to be used with and without
the fsat. Otherwise it just moves a saturate destination modifier
around. To not be harmful, the fsat has to only be used by the bcsel.
All Broadwell and newer Intel platforms had similar results. (Ice Lake shown)
total instructions in shared programs: 20174475 -> 20174449 (<.01%)
instructions in affected programs: 3913 -> 3887 (-0.66%)
helped: 13 / HURT: 0
total cycles in shared programs: 866844832 -> 866844719 (<.01%)
cycles in affected programs: 46037 -> 45924 (-0.25%)
helped: 10 / HURT: 1
All Intel platforms had similar results. (Ice Lake shown)
Instructions in all programs: 161491468 -> 161491372 (-0.0%)
helped: 31 / HURT: 8
Cycles in all programs: 10933090736 -> 10933024716 (-0.0%)
helped: 32 / HURT: 18
Reviewed-by: Matt Turner <mattst88@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22169>
As noted on [1], lowering patterns of the form
floatS -> floatB -> floatS ==> floatS
cannot require precision since this may cause flush denorming.
[1] 3f779013 ("nir: Add an algebraic optimization for float->double->float")
Fixes: b86305bb ("nir/algebraic: collapse conversion opcodes (many patterns)")
Acked-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Reviewed-by: Georg Lehmann <dadschoorse@gmail.com>
Signed-off-by: Isabella Basso <isabellabdoamaral@usp.br>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20965>
Conversions on smaller bit sizes should also be collapsed when composed.
This also adds more patterns on the
intS -> intB -> floatB ==> intS -> floatB
lowering so as to deal with any int size C > B instead of a fixed intB.
Closes: #7776
Acked-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Reviewed-by: Georg Lehmann <dadschoorse@gmail.com>
Signed-off-by: Isabella Basso <isabellabdoamaral@usp.br>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20965>
Some patterns were outside the list of optimizations.
Fixes: b86305bb ("nir/algebraic: collapse conversion opcodes (many patterns)")
Acked-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Reviewed-by: Georg Lehmann <dadschoorse@gmail.com>
Signed-off-by: Isabella Basso <isabellabdoamaral@usp.br>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20965>
