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>
