[Mesa-dev] [PATCH 05/12] nir/algebraic: Flag inexact optimizations
Jason Ekstrand
jason at jlekstrand.net
Wed Mar 23 18:40:52 UTC 2016
On Mar 23, 2016 2:07 AM, "Francisco Jerez" <currojerez at riseup.net> wrote:
>
> Jason Ekstrand <jason at jlekstrand.net> writes:
>
> > Many of our optimizations, while great for cutting shaders down to size,
> > aren't really precision-safe. This commit tries to flag all of the
> > inexact floating-point optimizations so they don't get run on values
that
> > are flagged "exact". It's a bit conservative and maybe flags some safe
> > optimizations as unsafe but that's better than missing one.
> > ---
> > src/compiler/nir/nir_opt_algebraic.py | 156
+++++++++++++++++-----------------
> > 1 file changed, 78 insertions(+), 78 deletions(-)
> >
> > diff --git a/src/compiler/nir/nir_opt_algebraic.py
b/src/compiler/nir/nir_opt_algebraic.py
> > index 3e7ea06..b229f03 100644
> > --- a/src/compiler/nir/nir_opt_algebraic.py
> > +++ b/src/compiler/nir/nir_opt_algebraic.py
> > @@ -61,19 +61,19 @@ optimizations = [
> > (('fabs', ('fneg', a)), ('fabs', a)),
> > (('iabs', ('iabs', a)), ('iabs', a)),
> > (('iabs', ('ineg', a)), ('iabs', a)),
> > - (('fadd', a, 0.0), a),
> > + (('~fadd', a, 0.0), a),
> > (('iadd', a, 0), a),
> > (('usadd_4x8', a, 0), a),
> > (('usadd_4x8', a, ~0), ~0),
> > - (('fadd', ('fmul', a, b), ('fmul', a, c)), ('fmul', a, ('fadd', b,
c))),
> > + (('~fadd', ('fmul', a, b), ('fmul', a, c)), ('fmul', a, ('fadd', b,
c))),
> > (('iadd', ('imul', a, b), ('imul', a, c)), ('imul', a, ('iadd', b,
c))),
> > - (('fadd', ('fneg', a), a), 0.0),
> > + (('~fadd', ('fneg', a), a), 0.0),
> > (('iadd', ('ineg', a), a), 0),
> > (('iadd', ('ineg', a), ('iadd', a, b)), b),
> > (('iadd', a, ('iadd', ('ineg', a), b)), b),
> > - (('fadd', ('fneg', a), ('fadd', a, b)), b),
> > - (('fadd', a, ('fadd', ('fneg', a), b)), b),
> > - (('fmul', a, 0.0), 0.0),
> > + (('~fadd', ('fneg', a), ('fadd', a, b)), b),
> > + (('~fadd', a, ('fadd', ('fneg', a), b)), b),
> > + (('~fmul', a, 0.0), 0.0),
> > (('imul', a, 0), 0),
> > (('umul_unorm_4x8', a, 0), 0),
> > (('umul_unorm_4x8', a, ~0), a),
> > @@ -81,33 +81,33 @@ optimizations = [
> > (('imul', a, 1), a),
> > (('fmul', a, -1.0), ('fneg', a)),
> > (('imul', a, -1), ('ineg', a)),
> > - (('ffma', 0.0, a, b), b),
> > - (('ffma', a, 0.0, b), b),
> > - (('ffma', a, b, 0.0), ('fmul', a, b)),
> > - (('ffma', a, 1.0, b), ('fadd', a, b)),
> > - (('ffma', 1.0, a, b), ('fadd', a, b)),
>
> I think 'ffma a 1 b'/'ffma 1 a b' -> 'fadd a b' could be made safe.
I think that should be OK.
> > - (('flrp', a, b, 0.0), a),
> > - (('flrp', a, b, 1.0), b),
> > - (('flrp', a, a, b), a),
> > - (('flrp', 0.0, a, b), ('fmul', a, b)),
> > + (('~ffma', 0.0, a, b), b),
> > + (('~ffma', a, 0.0, b), b),
> > + (('~ffma', a, b, 0.0), ('fmul', a, b)),
> > + (('~ffma', a, 1.0, b), ('fadd', a, b)),
> > + (('~ffma', 1.0, a, b), ('fadd', a, b)),
> > + (('~flrp', a, b, 0.0), a),
> > + (('~flrp', a, b, 1.0), b),
> > + (('~flrp', a, a, b), a),
> > + (('~flrp', 0.0, a, b), ('fmul', a, b)),
> > (('flrp', a, b, c), ('fadd', ('fmul', c, ('fsub', b, a)), a),
'options->lower_flrp'),
> > (('ffract', a), ('fsub', a, ('ffloor', a)),
'options->lower_ffract'),
> > - (('fadd', ('fmul', a, ('fadd', 1.0, ('fneg', c))), ('fmul', b, c)),
('flrp', a, b, c), '!options->lower_flrp'),
> > - (('fadd', a, ('fmul', c, ('fadd', b, ('fneg', a)))), ('flrp', a, b,
c), '!options->lower_flrp'),
> > + (('~fadd', ('fmul', a, ('fadd', 1.0, ('fneg', c))), ('fmul', b,
c)), ('flrp', a, b, c), '!options->lower_flrp'),
> > + (('~fadd', a, ('fmul', c, ('fadd', b, ('fneg', a)))), ('flrp', a,
b, c), '!options->lower_flrp'),
> > (('ffma', a, b, c), ('fadd', ('fmul', a, b), c),
'options->lower_ffma'),
> > - (('fadd', ('fmul', a, b), c), ('ffma', a, b, c),
'!options->lower_ffma'),
> > + (('~fadd', ('fmul', a, b), c), ('ffma', a, b, c),
'!options->lower_ffma'),
> > # Comparison simplifications
> > - (('inot', ('flt', a, b)), ('fge', a, b)),
> > - (('inot', ('fge', a, b)), ('flt', a, b)),
> > - (('inot', ('feq', a, b)), ('fne', a, b)),
> > - (('inot', ('fne', a, b)), ('feq', a, b)),
> > - (('inot', ('ilt', a, b)), ('ige', a, b)),
> > - (('inot', ('ige', a, b)), ('ilt', a, b)),
> > - (('inot', ('ieq', a, b)), ('ine', a, b)),
> > - (('inot', ('ine', a, b)), ('ieq', a, b)),
>
> What's unsafe about the four integer comparison simplifications above?
Right. I got too excited. I'll leave those as safe.
> > - (('fge', ('fneg', ('fabs', a)), 0.0), ('feq', a, 0.0)),
>
> This one seems mostly safe too.
>
> > - (('bcsel', ('flt', a, b), a, b), ('fmin', a, b)),
> > - (('bcsel', ('flt', a, b), b, a), ('fmax', a, b)),
>
> What are the semantics of fmin/fmax? Are they defined as in GLSL with
> min(NaN, x) = NaN, min(x, NaN) = x? Or the other way around? Or is it
> defined to be commutative? In the first case the code above would
> indeed be unsafe *but* the related simplification:
I dug up the GLSL spec and it says the following for min:
Returns y if y < x, otherwise it returns x
so min(NaN, y) = y but min(x, NaN) = NaN
If we just change the optimizations to match the spec better, they would be
safe. Then we could add the non-safe ones as non-safe.
Sadly, I think our hardware implements the DX behaviour of "give me the
non-NaN value"
> | (('bcsel', ('flt', a, b), a, b), ('fmin', b, a)),
> | (('bcsel', ('flt', a, b), b, a), ('fmax', a, b)),
>
> would be completely safe.
>
> > + (('~inot', ('flt', a, b)), ('fge', a, b)),
> > + (('~inot', ('fge', a, b)), ('flt', a, b)),
> > + (('~inot', ('feq', a, b)), ('fne', a, b)),
> > + (('~inot', ('fne', a, b)), ('feq', a, b)),
> > + (('~inot', ('ilt', a, b)), ('ige', a, b)),
> > + (('~inot', ('ige', a, b)), ('ilt', a, b)),
> > + (('~inot', ('ieq', a, b)), ('ine', a, b)),
> > + (('~inot', ('ine', a, b)), ('ieq', a, b)),
> > + (('~fge', ('fneg', ('fabs', a)), 0.0), ('feq', a, 0.0)),
> > + (('~bcsel', ('flt', a, b), a, b), ('fmin', a, b)),
> > + (('~bcsel', ('flt', a, b), b, a), ('fmax', a, b)),
> > (('bcsel', ('inot', 'a at bool'), b, c), ('bcsel', a, c, b)),
> > (('bcsel', a, ('bcsel', a, b, c), d), ('bcsel', a, b, d)),
> > (('fmin', a, a), a),
> > @@ -116,17 +116,17 @@ optimizations = [
> > (('imax', a, a), a),
> > (('umin', a, a), a),
> > (('umax', a, a), a),
> > - (('fmin', ('fmax', a, 0.0), 1.0), ('fsat', a),
'!options->lower_fsat'),
> > - (('fmax', ('fmin', a, 1.0), 0.0), ('fsat', a),
'!options->lower_fsat'),
> > + (('~fmin', ('fmax', a, 0.0), 1.0), ('fsat', a),
'!options->lower_fsat'),
> > + (('~fmax', ('fmin', a, 1.0), 0.0), ('fsat', a),
'!options->lower_fsat'),
> > (('fsat', a), ('fmin', ('fmax', a, 0.0), 1.0),
'options->lower_fsat'),
> > (('fsat', ('fsat', a)), ('fsat', a)),
> > - (('fsat', ('fmin', a, 1.0)), ('fsat', a)),
> > - (('fsat', ('fmax', a, 0.0)), ('fsat', a)),
> > + (('~fsat', ('fmin', a, 1.0)), ('fsat', a)),
> > + (('~fsat', ('fmax', a, 0.0)), ('fsat', a)),
> > (('fmin', ('fmax', ('fmin', ('fmax', a, 0.0), 1.0), 0.0), 1.0),
('fmin', ('fmax', a, 0.0), 1.0)),
> > - (('ior', ('flt', a, b), ('flt', a, c)), ('flt', a, ('fmax', b, c))),
> > - (('ior', ('flt', a, c), ('flt', b, c)), ('flt', ('fmin', a, b), c)),
> > - (('ior', ('fge', a, b), ('fge', a, c)), ('fge', a, ('fmin', b, c))),
> > - (('ior', ('fge', a, c), ('fge', b, c)), ('fge', ('fmax', a, b), c)),
>
> These would be safe if fmin/fmax were commutative, which I don't think
> is the case?
They are not commutative :-(
> > + (('~ior', ('flt', a, b), ('flt', a, c)), ('flt', a, ('fmax', b,
c))),
> > + (('~ior', ('flt', a, c), ('flt', b, c)), ('flt', ('fmin', a, b),
c)),
> > + (('~ior', ('fge', a, b), ('fge', a, c)), ('fge', a, ('fmin', b,
c))),
> > + (('~ior', ('fge', a, c), ('fge', b, c)), ('fge', ('fmax', a, b),
c)),
> > (('slt', a, b), ('b2f', ('flt', a, b)), 'options->lower_scmp'),
> > (('sge', a, b), ('b2f', ('fge', a, b)), 'options->lower_scmp'),
> > (('seq', a, b), ('b2f', ('feq', a, b)), 'options->lower_scmp'),
> > @@ -150,15 +150,15 @@ optimizations = [
> > (('ult', a, a), False),
> > (('uge', a, a), True),
> > # Logical and bit operations
> > - (('fand', a, 0.0), 0.0),
> > + (('~fand', a, 0.0), 0.0),
>
> This is safe if fand(x, y) is defined as (x != 0 && y != 0 ? 1 : 0).
> Could you remind me how it is defined?
Yes, that is the definition.
> > (('iand', a, a), a),
> > (('iand', a, ~0), a),
> > (('iand', a, 0), 0),
> > (('ior', a, a), a),
> > (('ior', a, 0), a),
> > - (('fxor', a, a), 0.0),
> > + (('~fxor', a, a), 0.0),
>
> Same goes here.
Right.
> > (('ixor', a, a), 0),
> > - (('fxor', a, 0.0), a),
> > + (('~fxor', a, 0.0), a),
> > (('ixor', a, 0), a),
> > (('inot', ('inot', a)), a),
> > # DeMorgan's Laws
> > @@ -174,35 +174,35 @@ optimizations = [
> > (('iand', 0xff, ('ushr', a, 24)), ('ushr', a, 24)),
> > (('iand', 0xffff, ('ushr', a, 16)), ('ushr', a, 16)),
> > # Exponential/logarithmic identities
> > - (('fexp2', ('flog2', a)), a), # 2^lg2(a) = a
> > - (('flog2', ('fexp2', a)), a), # lg2(2^a) = a
> > + (('~fexp2', ('flog2', a)), a), # 2^lg2(a) = a
> > + (('~flog2', ('fexp2', a)), a), # lg2(2^a) = a
> > (('fpow', a, b), ('fexp2', ('fmul', ('flog2', a), b)),
'options->lower_fpow'), # a^b = 2^(lg2(a)*b)
> > - (('fexp2', ('fmul', ('flog2', a), b)), ('fpow', a, b),
'!options->lower_fpow'), # 2^(lg2(a)*b) = a^b
> > - (('fexp2', ('fadd', ('fmul', ('flog2', a), b), ('fmul', ('flog2',
c), d))),
> > - ('fmul', ('fpow', a, b), ('fpow', c, d)), '!options->lower_fpow'),
# 2^(lg2(a) * b + lg2(c) + d) = a^b * c^d
> > - (('fpow', a, 1.0), a),
> > - (('fpow', a, 2.0), ('fmul', a, a)),
> > - (('fpow', a, 4.0), ('fmul', ('fmul', a, a), ('fmul', a, a))),
> > - (('fpow', 2.0, a), ('fexp2', a)),
> > - (('fpow', ('fpow', a, 2.2), 0.454545), a),
> > - (('fpow', ('fabs', ('fpow', a, 2.2)), 0.454545), ('fabs', a)),
> > - (('fsqrt', ('fexp2', a)), ('fexp2', ('fmul', 0.5, a))),
> > - (('frcp', ('fexp2', a)), ('fexp2', ('fneg', a))),
> > - (('frsq', ('fexp2', a)), ('fexp2', ('fmul', -0.5, a))),
> > - (('flog2', ('fsqrt', a)), ('fmul', 0.5, ('flog2', a))),
> > - (('flog2', ('frcp', a)), ('fneg', ('flog2', a))),
> > - (('flog2', ('frsq', a)), ('fmul', -0.5, ('flog2', a))),
> > - (('flog2', ('fpow', a, b)), ('fmul', b, ('flog2', a))),
> > - (('fadd', ('flog2', a), ('flog2', b)), ('flog2', ('fmul', a, b))),
> > - (('fadd', ('flog2', a), ('fneg', ('flog2', b))), ('flog2', ('fdiv',
a, b))),
> > - (('fmul', ('fexp2', a), ('fexp2', b)), ('fexp2', ('fadd', a, b))),
> > + (('~fexp2', ('fmul', ('flog2', a), b)), ('fpow', a, b),
'!options->lower_fpow'), # 2^(lg2(a)*b) = a^b
> > + (('~fexp2', ('fadd', ('fmul', ('flog2', a), b), ('fmul', ('flog2',
c), d))),
> > + ('~fmul', ('fpow', a, b), ('fpow', c, d)),
'!options->lower_fpow'), # 2^(lg2(a) * b + lg2(c) + d) = a^b * c^d
> > + (('~fpow', a, 1.0), a),
> > + (('~fpow', a, 2.0), ('fmul', a, a)),
> > + (('~fpow', a, 4.0), ('fmul', ('fmul', a, a), ('fmul', a, a))),
> > + (('~fpow', 2.0, a), ('fexp2', a)),
> > + (('~fpow', ('fpow', a, 2.2), 0.454545), a),
> > + (('~fpow', ('fabs', ('fpow', a, 2.2)), 0.454545), ('fabs', a)),
> > + (('~fsqrt', ('fexp2', a)), ('fexp2', ('fmul', 0.5, a))),
> > + (('~frcp', ('fexp2', a)), ('fexp2', ('fneg', a))),
> > + (('~frsq', ('fexp2', a)), ('fexp2', ('fmul', -0.5, a))),
> > + (('~flog2', ('fsqrt', a)), ('fmul', 0.5, ('flog2', a))),
> > + (('~flog2', ('frcp', a)), ('fneg', ('flog2', a))),
> > + (('~flog2', ('frsq', a)), ('fmul', -0.5, ('flog2', a))),
> > + (('~flog2', ('fpow', a, b)), ('fmul', b, ('flog2', a))),
> > + (('~fadd', ('flog2', a), ('flog2', b)), ('flog2', ('fmul', a, b))),
> > + (('~fadd', ('flog2', a), ('fneg', ('flog2', b))), ('flog2',
('fdiv', a, b))),
> > + (('~fmul', ('fexp2', a), ('fexp2', b)), ('fexp2', ('fadd', a, b))),
> > # Division and reciprocal
> > - (('fdiv', 1.0, a), ('frcp', a)),
> > - (('fdiv', a, b), ('fmul', a, ('frcp', b)), 'options->lower_fdiv'),
>
> This one is safe and I guess necessary if the back-end requires fdiv to
> be lowered?
Thanks. I must have gotten too eager to flag things.
> > - (('frcp', ('frcp', a)), a),
> > - (('frcp', ('fsqrt', a)), ('frsq', a)),
> > - (('fsqrt', a), ('frcp', ('frsq', a)), 'options->lower_fsqrt'),
>
> Same here.
Yup
> > - (('frcp', ('frsq', a)), ('fsqrt', a), '!options->lower_fsqrt'),
> > + (('~fdiv', 1.0, a), ('frcp', a)),
> > + (('~fdiv', a, b), ('fmul', a, ('frcp', b)), 'options->lower_fdiv'),
> > + (('~frcp', ('frcp', a)), a),
> > + (('~frcp', ('fsqrt', a)), ('frsq', a)),
> > + (('~fsqrt', a), ('frcp', ('frsq', a)), 'options->lower_fsqrt'),
> > + (('~frcp', ('frsq', a)), ('fsqrt', a), '!options->lower_fsqrt'),
> > # Boolean simplifications
> > (('ieq', 'a at bool', True), a),
> > (('ine', 'a at bool', True), ('inot', a)),
> > @@ -221,8 +221,8 @@ optimizations = [
> >
> > # Conversions
> > (('i2b', ('b2i', a)), a),
> > - (('f2i', ('ftrunc', a)), ('f2i', a)),
> > - (('f2u', ('ftrunc', a)), ('f2u', a)),
> > + (('~f2i', ('ftrunc', a)), ('f2i', a)),
> > + (('~f2u', ('ftrunc', a)), ('f2u', a)),
> >
> These seem safe to me.
I think they probably are. I think there may be a problem when the number
of mantissa bits is bigger than the number of integer bits. In that case,
the out-of-bounds behaviour may get interesting. However, we're not
handling that case yet so I think it's safe.
> > # Byte extraction
> > (('ushr', a, 24), ('extract_u8', a, 3),
'!options->lower_extract_byte'),
> > @@ -235,17 +235,17 @@ optimizations = [
> > (('iand', 0xffff, a), ('extract_u16', a, 0),
'!options->lower_extract_word'),
> >
> > # Subtracts
> > - (('fsub', a, ('fsub', 0.0, b)), ('fadd', a, b)),
> > + (('~fsub', a, ('fsub', 0.0, b)), ('fadd', a, b)),
> > (('isub', a, ('isub', 0, b)), ('iadd', a, b)),
> > (('ussub_4x8', a, 0), a),
> > (('ussub_4x8', a, ~0), 0),
> > - (('fsub', a, b), ('fadd', a, ('fneg', b)), 'options->lower_sub'),
> > + (('~fsub', a, b), ('fadd', a, ('fneg', b)), 'options->lower_sub'),
>
> Required for lowering.
Yup
> > (('isub', a, b), ('iadd', a, ('ineg', b)), 'options->lower_sub'),
> > - (('fneg', a), ('fsub', 0.0, a), 'options->lower_negate'),
> > + (('~fneg', a), ('fsub', 0.0, a), 'options->lower_negate'),
>
> Same here.
Yup
> > (('ineg', a), ('isub', 0, a), 'options->lower_negate'),
> > - (('fadd', a, ('fsub', 0.0, b)), ('fsub', a, b)),
> > + (('~fadd', a, ('fsub', 0.0, b)), ('fsub', a, b)),
> > (('iadd', a, ('isub', 0, b)), ('isub', a, b)),
> > - (('fabs', ('fsub', 0.0, a)), ('fabs', a)),
> > + (('~fabs', ('fsub', 0.0, a)), ('fabs', a)),
>
> This one seems safe.
Depends on whether or not a + 0.0 == a. I think it should but glennk
seemed to think it wasn't. I'd rather be on the safe side for now.
> > (('iabs', ('isub', 0, a)), ('iabs', a)),
> >
> > # Misc. lowering
> > @@ -372,10 +372,10 @@ for op in ['flt', 'fge', 'feq', 'fne',
> > # they help code generation but do not necessarily produce code that is
> > # more easily optimizable.
> > late_optimizations = [
> > - (('flt', ('fadd', a, b), 0.0), ('flt', a, ('fneg', b))),
>
> This one also seems safe,
I've come to the conclusion that it isn't. If you have a -inf and inf,
then a+b is going to be one of NaN, inf, or 0. In any case the expression
on the left is commutative while the one on the right isn't.
> > - (('fge', ('fadd', a, b), 0.0), ('fge', a, ('fneg', b))),
> > - (('feq', ('fadd', a, b), 0.0), ('feq', a, ('fneg', b))),
> > - (('fne', ('fadd', a, b), 0.0), ('fne', a, ('fneg', b))),
>
> ...but these other ones are indeed unsafe for some sign combinations of
> a and b when they are not finite.
>
> > + (('~flt', ('fadd', a, b), 0.0), ('flt', a, ('fneg', b))),
> > + (('~fge', ('fadd', a, b), 0.0), ('fge', a, ('fneg', b))),
> > + (('~feq', ('fadd', a, b), 0.0), ('feq', a, ('fneg', b))),
> > + (('~fne', ('fadd', a, b), 0.0), ('fne', a, ('fneg', b))),
> > (('fdot2', a, b), ('fdot_replicated2', a, b),
'options->fdot_replicates'),
> > (('fdot3', a, b), ('fdot_replicated3', a, b),
'options->fdot_replicates'),
> > (('fdot4', a, b), ('fdot_replicated4', a, b),
'options->fdot_replicates'),
>
> I've also looked through the remaining algebraic optimizations in this
> pass, and things look safe except for the ones you have already marked
> unsafe in this patch -- I'll wait until you answer the questions I
> pointed out before I send you a R-b for it though.
>
> > --
> > 2.5.0.400.gff86faf
> >
> > _______________________________________________
> > mesa-dev mailing list
> > mesa-dev at lists.freedesktop.org
> > https://lists.freedesktop.org/mailman/listinfo/mesa-dev
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.freedesktop.org/archives/mesa-dev/attachments/20160323/4d8aabdd/attachment-0001.html>
More information about the mesa-dev
mailing list