[Mesa-dev] [PATCH 15/22] nir: Use De Morgan's Law on logic compounded comparisons

Wed Feb 28 09:07:29 UTC 2018

Reviewed-by: Samuel Iglesias Gonsálvez <siglesias at igalia.com>


On 24/02/18 00:56, Ian Romanick wrote:
> From: Ian Romanick <ian.d.romanick at intel.com>
>
> The replacement of the comparison operators must happen during this
> step.  If it does not, the next pass of nir_opt_algebraic will reapply
> De Morgan's Law in the "opposite direction" before performing dead code
> elimination.  The resulting infinite loop will eventually get OOM
> killed.
>
> Haswell, Broadwell, and Skylake had similar results. (Broadwell shown)
> total instructions in shared programs: 14808185 -> 14808036 (<.01%)
> instructions in affected programs: 13758 -> 13609 (-1.08%)
> helped: 39
> HURT: 0
> helped stats (abs) min: 1 max: 10 x̄: 3.82 x̃: 3
> helped stats (rel) min: 0.44% max: 1.55% x̄: 0.98% x̃: 1.01%
> 95% mean confidence interval for instructions value: -4.67 -2.97
> 95% mean confidence interval for instructions %-change: -1.09% -0.88%
> Instructions are helped.
>
> total cycles in shared programs: 559438333 -> 559435832 (<.01%)
> cycles in affected programs: 199160 -> 196659 (-1.26%)
> helped: 42
> HURT: 3
> helped stats (abs) min: 2 max: 184 x̄: 61.50 x̃: 51
> helped stats (rel) min: 0.02% max: 6.94% x̄: 1.41% x̃: 1.40%
> HURT stats (abs)   min: 2 max: 40 x̄: 27.33 x̃: 40
> HURT stats (rel)   min: 0.05% max: 0.74% x̄: 0.51% x̃: 0.74%
> 95% mean confidence interval for cycles value: -71.47 -39.69
> 95% mean confidence interval for cycles %-change: -1.64% -0.93%
> Cycles are helped.
>
> Sandy Bridge and Ivy Bridge had similar results. (Ivy Bridge shown)
> total instructions in shared programs: 11811776 -> 11811553 (<.01%)
> instructions in affected programs: 15201 -> 14978 (-1.47%)
> helped: 39
> HURT: 0
> helped stats (abs) min: 1 max: 20 x̄: 5.72 x̃: 6
> helped stats (rel) min: 0.44% max: 2.53% x̄: 1.30% x̃: 1.26%
> 95% mean confidence interval for instructions value: -7.21 -4.23
> 95% mean confidence interval for instructions %-change: -1.48% -1.12%
> Instructions are helped.
>
> total cycles in shared programs: 257617270 -> 257614589 (<.01%)
> cycles in affected programs: 212107 -> 209426 (-1.26%)
> helped: 45
> HURT: 0
> helped stats (abs) min: 2 max: 180 x̄: 59.58 x̃: 54
> helped stats (rel) min: 0.02% max: 6.02% x̄: 1.30% x̃: 1.32%
> 95% mean confidence interval for cycles value: -74.02 -45.14
> 95% mean confidence interval for cycles %-change: -1.59% -1.01%
> Cycles are helped.
>
> Iron Lake
> total instructions in shared programs: 7886648 -> 7886515 (<.01%)
> instructions in affected programs: 14106 -> 13973 (-0.94%)
> helped: 29
> HURT: 0
> helped stats (abs) min: 1 max: 10 x̄: 4.59 x̃: 4
> helped stats (rel) min: 0.35% max: 1.83% x̄: 0.90% x̃: 0.81%
> 95% mean confidence interval for instructions value: -5.65 -3.52
> 95% mean confidence interval for instructions %-change: -1.03% -0.76%
> Instructions are helped.
>
> total cycles in shared programs: 178100812 -> 178100396 (<.01%)
> cycles in affected programs: 67970 -> 67554 (-0.61%)
> helped: 29
> HURT: 0
> helped stats (abs) min: 2 max: 40 x̄: 14.34 x̃: 12
> helped stats (rel) min: 0.15% max: 1.69% x̄: 0.58% x̃: 0.54%
> 95% mean confidence interval for cycles value: -18.30 -10.39
> 95% mean confidence interval for cycles %-change: -0.71% -0.45%
> Cycles are helped.
>
> GM45
> total instructions in shared programs: 4857939 -> 4857872 (<.01%)
> instructions in affected programs: 7426 -> 7359 (-0.90%)
> helped: 15
> HURT: 0
> helped stats (abs) min: 1 max: 10 x̄: 4.47 x̃: 4
> helped stats (rel) min: 0.33% max: 1.80% x̄: 0.87% x̃: 0.77%
> 95% mean confidence interval for instructions value: -6.06 -2.87
> 95% mean confidence interval for instructions %-change: -1.06% -0.67%
> Instructions are helped.
>
> total cycles in shared programs: 122167930 -> 122167654 (<.01%)
> cycles in affected programs: 43118 -> 42842 (-0.64%)
> helped: 15
> HURT: 0
> helped stats (abs) min: 4 max: 40 x̄: 18.40 x̃: 16
> helped stats (rel) min: 0.15% max: 1.69% x̄: 0.62% x̃: 0.54%
> 95% mean confidence interval for cycles value: -25.03 -11.77
> 95% mean confidence interval for cycles %-change: -0.82% -0.41%
> Cycles are helped.
>
> Signed-off-by: Ian Romanick <ian.d.romanick at intel.com>
> ---
>  src/compiler/nir/nir_opt_algebraic.py | 9 +++++++++
>  1 file changed, 9 insertions(+)
>
> diff --git a/src/compiler/nir/nir_opt_algebraic.py b/src/compiler/nir/nir_opt_algebraic.py
> index 314b064..af3ad2c 100644
> --- a/src/compiler/nir/nir_opt_algebraic.py
> +++ b/src/compiler/nir/nir_opt_algebraic.py
> @@ -24,6 +24,7 @@
>  #    Jason Ekstrand (jason at jlekstrand.net)
>  
>  import nir_algebraic
> +import itertools
>  
>  # Convenience variables
>  a = 'a'
> @@ -545,6 +546,14 @@ optimizations = [
>       'options->lower_unpack_snorm_4x8'),
>  ]
>  
> +invert = {'feq': 'fne', 'fne': 'feq', 'fge': 'flt', 'flt': 'fge' }
> +
> +for left, right in list(itertools.combinations(invert.keys(), 2)) + zip(invert.keys(), invert.keys()):
> +   optimizations.append((('inot', ('ior(is_used_once)', (left, a, b), (right, c, d))),
> +                         ('iand', (invert[left], a, b), (invert[right], c, d))))
> +   optimizations.append((('inot', ('iand(is_used_once)', (left, a, b), (right, c, d))),
> +                         ('ior', (invert[left], a, b), (invert[right], c, d))))
> +
>  def fexp2i(exp, bits):
>     # We assume that exp is already in the right range.
>     if bits == 32:

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.freedesktop.org/archives/mesa-dev/attachments/20180228/5e60baf5/attachment-0001.html>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 801 bytes
Desc: OpenPGP digital signature
URL: <https://lists.freedesktop.org/archives/mesa-dev/attachments/20180228/5e60baf5/attachment-0001.sig>