v2: Go to extra effort to avoid flow control inserted to implement
short-circuit evaluation rules.
Results on the 308 shaders extracted from the fp64 portion of the OpenGL
CTS:
Tiger Lake and Ice Lake had similar results. (Tiger Lake shown)
total instructions in shared programs: 797779 -> 796849 (-0.12%)
instructions in affected programs: 3499 -> 2569 (-26.58%)
helped: 21
HURT: 0
helped stats (abs) min: 8 max: 112 x̄: 44.29 x̃: 44
helped stats (rel) min: 16.09% max: 33.15% x̄: 25.72% x̃: 24.62%
95% mean confidence interval for instructions value: -55.94 -32.63
95% mean confidence interval for instructions %-change: -28.14% -23.30%
Instructions are helped.
total cycles in shared programs: 6601355 -> 6588351 (-0.20%)
cycles in affected programs: 25376 -> 12372 (-51.25%)
helped: 21
HURT: 0
helped stats (abs) min: 156 max: 1410 x̄: 619.24 x̃: 526
helped stats (rel) min: 42.39% max: 53.98% x̄: 50.12% x̃: 50.75%
95% mean confidence interval for cycles value: -776.58 -461.89
95% mean confidence interval for cycles %-change: -51.57% -48.67%
Cycles are helped.
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net> [v1]
Reviewed-by: Matt Turner <mattst88@gmail.com> [v1]
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4142>
fsat is defined as min(max(a, 0.0), 1.0), and IEEE defines both min and
max to return the non-NaN value when one value is NaN. Based on this,
fsat should definitely return 0.0 for NaN.
Results on the 308 shaders extracted from the fp64 portion of the OpenGL
CTS:
Tiger Lake and Ice Lake had similar results. (Tiger Lake shown)
total instructions in shared programs: 841666 -> 841647 (<.01%)
instructions in affected programs: 122033 -> 122014 (-0.02%)
helped: 7
HURT: 0
helped stats (abs) min: 1 max: 4 x̄: 2.71 x̃: 3
helped stats (rel) min: 0.01% max: 0.02% x̄: 0.02% x̃: 0.01%
95% mean confidence interval for instructions value: -3.74 -1.69
95% mean confidence interval for instructions %-change: -0.02% -0.01%
Instructions are helped.
total cycles in shared programs: 6927246 -> 6926904 (<.01%)
cycles in affected programs: 1038987 -> 1038645 (-0.03%)
helped: 7
HURT: 0
helped stats (abs) min: 18 max: 72 x̄: 48.86 x̃: 54
helped stats (rel) min: 0.03% max: 0.05% x̄: 0.03% x̃: 0.03%
95% mean confidence interval for cycles value: -67.38 -30.33
95% mean confidence interval for cycles %-change: -0.05% -0.02%
Cycles are helped.
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Fixes: a42163cbbc ("compiler: Add lowering support for 64-bit saturate operations to software")
Closes: https://gitlab.freedesktop.org/mesa/mesa/issues/2585
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4142>
__fabs64 doesn't do anything special, and the value is still NaN
regardless of the value of the MSB. In a strict sense, it's possible
that both functions should set the "signal" bit.
lts on the 308 shaders extracted from the fp64 portion of the OpenGL
CTS:
Tiger Lake and Ice Lake had similar results. (Tiger Lake shown)
total instructions in shared programs: 844558 -> 843005 (-0.18%)
instructions in affected programs: 725975 -> 724422 (-0.21%)
helped: 53
HURT: 4
helped stats (abs) min: 1 max: 313 x̄: 29.87 x̃: 21
helped stats (rel) min: 0.01% max: 0.94% x̄: 0.30% x̃: 0.22%
HURT stats (abs) min: 4 max: 11 x̄: 7.50 x̃: 7
HURT stats (rel) min: 0.03% max: 0.09% x̄: 0.05% x̃: 0.04%
95% mean confidence interval for instructions value: -39.02 -15.47
95% mean confidence interval for instructions %-change: -0.34% -0.21%
Instructions are helped.
total cycles in shared programs: 6962024 -> 6944998 (-0.24%)
cycles in affected programs: 6185470 -> 6168444 (-0.28%)
helped: 59
HURT: 0
helped stats (abs) min: 64 max: 2863 x̄: 288.58 x̃: 208
helped stats (rel) min: 0.11% max: 0.87% x̄: 0.33% x̃: 0.27%
95% mean confidence interval for cycles value: -387.15 -190.00
95% mean confidence interval for cycles %-change: -0.38% -0.28%
Cycles are helped.
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4142>
Replace all of the bool(qSign) with qSign != 0u. Remove unnecessary
parenthesis from around most of the existing qSign != 0u.
This dramatically simplifies the next commit.
Results on the 308 shaders extracted from the fp64 portion of the OpenGL
CTS:
Tiger Lake and Ice Lake had similar results. (Tiger Lake shown)
total instructions in shared programs: 848109 -> 848106 (<.01%)
instructions in affected programs: 53 -> 50 (-5.66%)
helped: 1
HURT: 0
total cycles in shared programs: 6969145 -> 6969125 (<.01%)
cycles in affected programs: 396 -> 376 (-5.05%)
helped: 1
HURT: 0
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4142>
findMSB returns -1 for an input of zero, so 31 - findMSB(a) is
sufficient on its own.
There's only one user of findMSB in shader-db, and it does not match
this pattern.
TODO: Add a pattern in the backend code generator that emits 31 -
nir_op_ufind_msb(a) as if it were nir_op_uclz. That should save a couple
instructions.
Results on the 308 shaders extracted from the fp64 portion of the OpenGL
CTS:
Tiger Lake and Ice Lake had similar results. (Tiger Lake shown)
total instructions in shared programs: 859509 -> 848109 (-1.33%)
instructions in affected programs: 841058 -> 829658 (-1.36%)
helped: 97
HURT: 0
helped stats (abs) min: 3 max: 1161 x̄: 117.53 x̃: 72
helped stats (rel) min: 0.98% max: 6.74% x̄: 1.70% x̃: 1.35%
95% mean confidence interval for instructions value: -147.21 -87.84
95% mean confidence interval for instructions %-change: -1.94% -1.46%
Instructions are helped.
total cycles in shared programs: 7072275 -> 6969145 (-1.46%)
cycles in affected programs: 6955767 -> 6852637 (-1.48%)
helped: 97
HURT: 0
helped stats (abs) min: 32 max: 10900 x̄: 1063.20 x̃: 560
helped stats (rel) min: 1.18% max: 7.58% x̄: 1.84% x̃: 1.45%
95% mean confidence interval for cycles value: -1339.43 -786.96
95% mean confidence interval for cycles %-change: -2.11% -1.57%
Cycles are helped.
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4142>
This doesn't help a lot of shaders, but it helps those few a LOT.
This could also be implemented using bcsel. That version is very
slightly worse because the generated SEL instruction wants to have two
immediate sources, so one of them usually needs an extra MOV instruction
to load.
Results on the 308 shaders extracted from the fp64 portion of the OpenGL
CTS:
Tiger Lake and Ice Lake had similar results. (Tiger Lake shown)
total instructions in shared programs: 929619 -> 928859 (-0.08%)
instructions in affected programs: 1651 -> 891 (-46.03%)
helped: 8
HURT: 0
helped stats (abs) min: 38 max: 152 x̄: 95.00 x̃: 95
helped stats (rel) min: 42.70% max: 86.36% x̄: 49.88% x̃: 44.66%
95% mean confidence interval for instructions value: -132.97 -57.03
95% mean confidence interval for instructions %-change: -62.28% -37.49%
Instructions are helped.
total cycles in shared programs: 7280180 -> 7272912 (-0.10%)
cycles in affected programs: 12960 -> 5692 (-56.08%)
helped: 8
HURT: 0
helped stats (abs) min: 352 max: 1456 x̄: 908.50 x̃: 910
helped stats (rel) min: 52.45% max: 91.19% x̄: 59.24% x̃: 55.15%
95% mean confidence interval for cycles value: -1274.03 -542.97
95% mean confidence interval for cycles %-change: -70.06% -48.41%
Cycles are helped.
Reviewed-by: Matt Turner <mattst88@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4142>
The current implementation was broken for any integers between 2^24
and 2^30 (it would return zero for me on ICL). The reason is that for
such integers we wouldn't take the 'if (0 <= shiftCount)' early return
path, however 'shiftCount + 7' would be positive, leading to a
negative 'count' argument passed to __shift64RightJamming(), which
would give undefined results.
This reworks the affected conversion functions to use either
__shortShift64Left() or __shift64RightJamming() based on the sign of
the final shift count, which should avoid the problem. In addition
this should qualify as a clean-up/optimization -- This implementation
of the conversion functions translates to 7 instructions less than the
original on Intel hardware.
This fixes the 'KHR-GL46.shader_ballot_tests.ShaderBallotFunctionBallot'
conformance tests on soft fp64 hardware with large enough subgroup
size (>16).
Fixes: d5cf6e92b4 "glsl: Add built-in functions to do uint64_to_fp32(uint64_t)"
Fixes: c9d333a6b7 "glsl: Add built-in functions to do int64_to_fp32(int64_t)"
Cc: Sagar Ghuge <sagar.ghuge@intel.com>
Reviewed-by: Sagar Ghuge <sagar.ghuge@intel.com>
v2: use mix and findMSB to optimise.
v3: [Sagar] Fix zFrac0 == 0u case in __normalizeRoundAndPackFloat64
Signed-off-by: Elie Tournier <elie.tournier@collabora.com>
The following patches will add implementations of various
double-precision operations to this file.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
