This is the mail archive of the gcc-patches@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [PATCH][tree-ssa-math-opts] Expand pow (x, CONST) using square roots when possible



On 08/05/15 11:18, Richard Biener wrote:
On Fri, May 1, 2015 at 6:02 PM, Kyrill Tkachov
<kyrylo.tkachov@foss.arm.com> wrote:
Hi all,

GCC has some logic to expand calls to pow (x, 0.75), pow (0.25) and pow (x,
(int)k + 0.5)
using square roots. So, for the above examples it would generate sqrt (x) *
sqrt (sqrt (x)),
sqrt (sqrt (x)) and powi (x, k) * sqrt (x) (assuming k > 0. For k < 0 it
will calculate the
reciprocal of that).

However, the implementation of these optimisations is done on a bit of an
ad-hoc basis with
the 0.25, 0.5, 0.75 cases hardcoded.
Judging by
https://gcc.gnu.org/wiki/summit2010?action=AttachFile&do=get&target=meissner2.pdf
these are the most commonly used exponents (at least in SPEC ;))

This patch generalises this optimisation into a (hopefully) more robust
algorithm.
In particular, it expands calls to pow (x, CST) by expanding the integer
part of CST
using a powi, like it does already, and then expanding the fractional part
as a product
of repeated applications of a square root if the fractional part can be
expressed
as a multiple of a power of 0.5.

I try to explain the algorithm in more detail in the comments in the patch
but, for example:

pow (x, 5.625) is not currently handled, but with this patch will be
expanded
to powi (x, 5) * sqrt (x) * sqrt (sqrt (sqrt (x))) because 5.625 == 5.0 +
0.5 + 0.5**3

Negative exponents are handled in either of two ways, depending on the
exponent value:
* Using a simple reciprocal.
   For example:
   pow (x, -5.625) == 1.0 / pow (x, 5.625)
     --> 1.0 / (powi (x, 5) * sqrt (x) * sqrt (sqrt (sqrt (x))))

* For pow (x, EXP) with negative exponent EXP with integer part INT and
fractional part FRAC:
pow (1.0 - FRAC) / powi (ceil (abs (EXP))).
   For example:
   pow (x, -5.875) == pow (x, 0.125) / powi (X, 6)
     --> sqrt (sqrt (sqrt (x))) / (powi (x, 6))


Since hardware square root instructions tend to be expensive, we may want to
reduce the number
of square roots we are willing to calculate. Since we reuse intermediate
square root results,
this boils down to restricting the depth of the square root chains. In all
the examples above
that depth is 3. I've made this maximum depth parametrisable in params.def.
By adjusting that
parameter we can adjust the resolution of this optimisation. So, if it's set
to '4' then we
will synthesize every exponent that is a multiple of 0.5**4 == 0.0625,
including negative
multiples. Currently, GCC will not try to expand negative multiples of
anything else than 0.5

I have tried to keep the existing functionality intact and activate this
only for
-funsafe-math-optimizations and only when the target has a sqrt instruction.
  An exception to that is pow (x, 0.5) which we prefer to transform to sqrt
even
when a hardware sqrt is not available, presumably because the library
function for
sqrt is usually faster than pow (?).
Yes.  It's also a safe transform - which you seem to put under
flag_unsafe_math_optimizations only with your patch.

It would be clearer to just leave the special-case

-  /* Optimize pow(x,0.5) = sqrt(x).  This replacement is always safe
-     unless signed zeros must be maintained.  pow(-0,0.5) = +0, while
-     sqrt(-0) = -0.  */
-  if (sqrtfn
-      && REAL_VALUES_EQUAL (c, dconsthalf)
-      && !HONOR_SIGNED_ZEROS (mode))
-    return build_and_insert_call (gsi, loc, sqrtfn, arg0);

in as-is.

Ok, I'll leave that case explicit.


You also removed the Os constraint which you should put back in.
Basically if !optimize_function_for_speed_p then generate at most
two calls to sqrt (iff the HW has a sqrt instruction).

I tried to move that logic into expand_with_sqrts but
I'll move it outside it. It seems that this boils down to
only 0.25, as any other 2xsqrt chain will also involve a
multiply or a divide which we currently avoid.


You fail to add a testcase that checks that the optimization applies.

I'll add one to scan the sincos dump.
I notice that we don't have a testuite check that the target has
a hw sqrt instructions. Would you like me to add one? Or can I make
the testcase aarch64-specific?


Otherwise the idea looks good though there must be a better way
to compute the series than by using real-arithmetic and forcefully
trying out all possibilities...

I get that feeling too. What I need is not only a way
of figuring out if the fractional part of the exponent can be
represented in this way, but also compute the depth of the
sqrt chain and the number of multiplies...
That being said, the current approach is O(maximum depth) and
I don't expect the depth to go much beyond 3 or 4 in practice.

Thanks for looking at it!
I'll respin the patch.

Kyrill


Richard.


Having seen the glibc implementation of a fully IEEE-754-compliant pow
function, I think we
would prefer synthesising the pow call whenever we can for -ffast-math.

I have seen this optimisation trigger a few times in SPEC2k6, in particular
in 447.dealII
and 481.wrf where it replaced calls to powf (x, -0.25), pow (x, 0.125) and
pow (x, 0.875)
with square roots, multiplies and, in the case of -0.25, divides.
On 481.wrf I saw it remove a total of 22 out of 322 calls to pow

On 481.wrf on aarch64 I saw about a 1% improvement.
The cycle count on x86_64 was also smaller, but not by a significant amount
(the same calls to
pow were eliminated).

In general, I think this can shine if multiple expandable calls to pow
appear together.
So, for example for code:
double
baz (double a)
{
   return __builtin_pow (a, -1.25) + __builtin_pow (a, 5.75) - __builtin_pow
(a, 3.375);
}

we can generate:
baz:
         fsqrt   d3, d0
         fmul    d4, d0, d0
         fmov    d5, 1.0e+0
         fmul    d6, d0, d4
         fsqrt   d2, d3
         fmul    d1, d0, d2
         fsqrt   d0, d2
         fmul    d3, d3, d2
         fdiv    d1, d5, d1
         fmul    d3, d3, d6
         fmul    d2, d2, d0
         fmadd   d0, d4, d3, d1
         fmsub   d0, d6, d2, d0
         ret

reusing the sqrt results and doing more optimisations rather than the
current:
baz:
         stp     x29, x30, [sp, -48]!
         fmov    d1, -1.25e+0
         add     x29, sp, 0
         stp     d8, d9, [sp, 16]
         fmov    d9, d0
         str     d10, [sp, 32]
         bl      pow
         fmov    d8, d0
         fmov    d0, d9
         fmov    d1, 5.75e+0
         bl      pow
         fmov    d10, d0
         fmov    d0, d9
         fmov    d1, 3.375e+0
         bl      pow
         fadd    d8, d8, d10
         ldr     d10, [sp, 32]
         fsub    d0, d8, d0
         ldp     d8, d9, [sp, 16]
         ldp     x29, x30, [sp], 48
         ret


Of course gcc could already do that if the exponents were to fall in the set
{0.25, 0.75, k+0.5}
but with this patch that set can be greatly expanded.

I suppose if we're really lucky we might even open up new vectorisation
opportunities.
For example:
void
vecfoo (float *a, float *b)
{
   for (int i = 0; i < N; i++)
     a[i] = __builtin_powf (a[i], 1.25f) - __builtin_powf (b[i], 3.625);
}

will now be vectorisable if we have a hardware vector sqrt instruction.
Though I'm not sure
how often this would appear in real code.


I would like advice on some implementation details:
- The new function representable_as_half_series_p is used to check if the
fractional
part of an exponent can be represented as a multiple of a power of 0.5. It
does so
by using REAL_VALUE_TYPE arithmetic and a loop. I wonder whether there is a
smarter
way of doing this, considering that REAL_VALUE_TYPE holds the bit
representation of the
floating point number?

- Are there any correctness cases that I may have missed? This patch gates
the optimisation
on -funsafe-math-optimizations, but maybe there are some edge cases that I
missed?

- What should be the default value of the max-pow-sqrt-depth param? In this
patch it's 5 which
on second thought strikes me as a bit aggressive. To catch exponents that
are multiples of
0.25 we need it to be 2. For multiples of 0.125 it has to be 3 etc... I
suppose it depends on
how likely more fine-grained powers are to appear in real code, how
expensive the C library
implementation of pow is, and how expensive are the sqrt instructions in
hardware.


Bootstrapped and tested on x86_64, aarch64, arm (pending)
SPEC2k6 built and ran fine. Can anyone suggest anything other codebase that
might
be of interest to look at?

Thanks,
Kyrill

2015-05-01  Kyrylo Tkachov  <kyrylo.tkachov@arm.com>

     * params.def (PARAM_MAX_POW_SQRT_DEPTH): New param.
     * tree-ssa-math-opts.c: Include params.h
     (pow_synth_sqrt_info): New struct.
     (representable_as_half_series_p): New function.
     (get_fn_chain): Likewise.
     (print_nested_fn): Likewise.
     (dump_fractional_sqrt_sequence): Likewise.
     (dump_integer_part): Likewise.
     (expand_pow_as_sqrts): Likewise.
     (gimple_expand_builtin_pow): Use above to attempt to expand
     pow as series of square roots.  Removed now unused variables.

2015-05-01  Kyrylo Tkachov  <kyrylo.tkachov@arm.com>

     * gcc.dg/pow-sqrt.x: New file.
     * gcc.dg/pow-sqrt-1.c: New test.
     * gcc.dg/pow-sqrt-2.c: Likewise.
     * gcc.dg/pow-sqrt-3.c: Likewise.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]