diff options
| author | Ian Romanick <[email protected]> | 2020-03-04 14:21:02 -0800 |
|---|---|---|
| committer | Marge Bot <[email protected]> | 2020-03-18 20:36:29 +0000 |
| commit | 617a69107ee58e23ace06093bc49fa2c86b7dd4b (patch) | |
| tree | e598904931d40b1d3f761561d1af04b5b4cd4b83 | |
| parent | abf28d6a70c3219e41c904806f77ea92d31bdb0f (diff) | |
soft-fp64/ffloor: Simplify the >= 0 comparison
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: 797951 -> 797779 (-0.02%)
instructions in affected programs: 126482 -> 126310 (-0.14%)
helped: 15
HURT: 0
helped stats (abs) min: 1 max: 20 x̄: 11.47 x̃: 10
helped stats (rel) min: <.01% max: 0.60% x̄: 0.28% x̃: 0.29%
95% mean confidence interval for instructions value: -14.79 -8.14
95% mean confidence interval for instructions %-change: -0.40% -0.16%
Instructions are helped.
total cycles in shared programs: 6601437 -> 6601355 (<.01%)
cycles in affected programs: 1089336 -> 1089254 (<.01%)
helped: 15
HURT: 0
helped stats (abs) min: 2 max: 12 x̄: 5.47 x̃: 6
helped stats (rel) min: <.01% max: 0.04% x̄: 0.01% x̃: 0.01%
95% mean confidence interval for cycles value: -7.06 -3.87
95% mean confidence interval for cycles %-change: -0.02% <.01%
Cycles are helped.
Reviewed-by: Matt Turner <[email protected]>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4142>
| -rw-r--r-- | src/compiler/glsl/float64.glsl | 14 |
1 files changed, 13 insertions, 1 deletions
diff --git a/src/compiler/glsl/float64.glsl b/src/compiler/glsl/float64.glsl index 5b0a9dc0c28..67e0e657c55 100644 --- a/src/compiler/glsl/float64.glsl +++ b/src/compiler/glsl/float64.glsl @@ -1714,7 +1714,19 @@ __ftrunc64(uint64_t __a) uint64_t __ffloor64(uint64_t a) { - bool is_positive = __fge64(a, 0ul); + /* The big assumtion is that when 'a' is NaN, __ftrunc(a) returns a. Based + * on that assumption, NaN values that don't have the sign bit will safely + * return NaN (identity). This is guarded by RELAXED_NAN_PROPAGATION + * because otherwise the NaN should have the "signal" bit set. The + * __fadd64 will ensure that occurs. + */ + bool is_positive = +#if defined RELAXED_NAN_PROPAGATION + int(unpackUint2x32(a).y) >= 0 +#else + __fge64(a, 0ul) +#endif + ; uint64_t tr = __ftrunc64(a); if (is_positive || __feq64(tr, a)) { |