On Mon, 18 Nov 2013, Sergey Ostanevich wrote:
> I would agree that the example is just for the case cost model makes
> correct estimation But how can we assure ourself that it won't have any
> mistakes in the future?
We call it bugs and not mistakes and we have bugzilla for it.
Richard.
> I believe it'll be Ok to introduce an extra flag as Jakub proposed for the
> dedicated simd-forced vectorization to use unlimited cost model. This
> can be default for -fopenmp or there should be a warning issued that
> compiler overrides user's request of vectorization. In such a case user
> can enforce vectorization (even with mentioned results :) with this
> unlimited cost model for simd.
>
>
>
> On Fri, Nov 15, 2013 at 6:24 PM, Richard Biener <rguent...@suse.de> wrote:
> > On Fri, 15 Nov 2013, Sergey Ostanevich wrote:
> >
> >> Richard,
> >>
> >> here's an example that causes trigger for the cost model.
> >
> > I hardly believe that (AVX2)
> >
> > .L9:
> > vmovups (%rsi), %xmm3
> > addl $1, %r8d
> > addq $256, %rsi
> > vinsertf128 $0x1, -240(%rsi), %ymm3, %ymm1
> > vmovups -224(%rsi), %xmm3
> > vinsertf128 $0x1, -208(%rsi), %ymm3, %ymm3
> > vshufps $136, %ymm3, %ymm1, %ymm3
> > vperm2f128 $3, %ymm3, %ymm3, %ymm2
> > vshufps $68, %ymm2, %ymm3, %ymm1
> > vshufps $238, %ymm2, %ymm3, %ymm2
> > vmovups -192(%rsi), %xmm3
> > vinsertf128 $1, %xmm2, %ymm1, %ymm2
> > vinsertf128 $0x1, -176(%rsi), %ymm3, %ymm1
> > vmovups -160(%rsi), %xmm3
> > vinsertf128 $0x1, -144(%rsi), %ymm3, %ymm3
> > vshufps $136, %ymm3, %ymm1, %ymm3
> > vperm2f128 $3, %ymm3, %ymm3, %ymm1
> > vshufps $68, %ymm1, %ymm3, %ymm4
> > vshufps $238, %ymm1, %ymm3, %ymm1
> > vmovups -128(%rsi), %xmm3
> > vinsertf128 $1, %xmm1, %ymm4, %ymm1
> > vshufps $136, %ymm1, %ymm2, %ymm1
> > vperm2f128 $3, %ymm1, %ymm1, %ymm2
> > vshufps $68, %ymm2, %ymm1, %ymm4
> > vshufps $238, %ymm2, %ymm1, %ymm2
> > vinsertf128 $0x1, -112(%rsi), %ymm3, %ymm1
> > vmovups -96(%rsi), %xmm3
> > vinsertf128 $1, %xmm2, %ymm4, %ymm4
> > vinsertf128 $0x1, -80(%rsi), %ymm3, %ymm3
> > vshufps $136, %ymm3, %ymm1, %ymm3
> > vperm2f128 $3, %ymm3, %ymm3, %ymm2
> > vshufps $68, %ymm2, %ymm3, %ymm1
> > vshufps $238, %ymm2, %ymm3, %ymm2
> > vmovups -64(%rsi), %xmm3
> > vinsertf128 $1, %xmm2, %ymm1, %ymm2
> > vinsertf128 $0x1, -48(%rsi), %ymm3, %ymm1
> > vmovups -32(%rsi), %xmm3
> > vinsertf128 $0x1, -16(%rsi), %ymm3, %ymm3
> > cmpl %r8d, %edi
> > vshufps $136, %ymm3, %ymm1, %ymm3
> > vperm2f128 $3, %ymm3, %ymm3, %ymm1
> > vshufps $68, %ymm1, %ymm3, %ymm5
> > vshufps $238, %ymm1, %ymm3, %ymm1
> > vinsertf128 $1, %xmm1, %ymm5, %ymm1
> > vshufps $136, %ymm1, %ymm2, %ymm1
> > vperm2f128 $3, %ymm1, %ymm1, %ymm2
> > vshufps $68, %ymm2, %ymm1, %ymm3
> > vshufps $238, %ymm2, %ymm1, %ymm2
> > vinsertf128 $1, %xmm2, %ymm3, %ymm1
> > vshufps $136, %ymm1, %ymm4, %ymm1
> > vperm2f128 $3, %ymm1, %ymm1, %ymm2
> > vshufps $68, %ymm2, %ymm1, %ymm3
> > vshufps $238, %ymm2, %ymm1, %ymm2
> > vinsertf128 $1, %xmm2, %ymm3, %ymm2
> > vaddps %ymm2, %ymm0, %ymm0
> > ja .L9
> >
> > is more efficient than
> >
> > .L3:
> > vaddss (%rcx,%rax), %xmm0, %xmm0
> > addq $32, %rax
> > cmpq %rdx, %rax
> > jne .L3
> >
> > ;)
> >
> >> As soon as
> >> elemental functions will appear and we update the vectorizer so it can
> >> accept
> >> an elemental function inside the loop - we will have the same
> >> situation as we have
> >> it now: cost model will bail out with profitability estimation.
> >
> > Yes.
> >
> >> Still we have no chance to get info on how efficient the bar() function
> >> when it
> >> is in vector form.
> >
> > Well I assume you mean that the speedup when vectorizing the elemental
> > will offset whatever wreckage we cause with vectorizing the rest of the
> > statements. I'd say you can at least compare to unrolling by
> > the vectorization factor, building the vector inputs to the elemental
> > from scalars, distributing the vector result from the elemental to
> > scalars.
> >
> >> I believe I should repeat: #pragma omp simd is intended for introduction
> >> of an
> >> instruction-level parallel region on developer's request, hence should
> >> be treated
> >> in same manner as #pragma omp parallel. Vectorizer cost model is an
> >> obstacle
> >> here, not a help.
> >
> > Surely not if there isn't an elemental call in it. With it the
> > cost model of course will have not enough information to decide.
> >
> > But still, what's the difference to the case where we cannot vectorize
> > the function? What happens if we cannot vectorize the elemental?
> > Do we have to build scalar versions for all possible vector sizes?
> >
> > Richard.
> >
> >> Regards,
> >> Sergos
> >>
> >>
> >> On Fri, Nov 15, 2013 at 1:08 AM, Richard Biener <rguent...@suse.de> wrote:
> >> > Sergey Ostanevich <sergos....@gmail.com> wrote:
> >> >>this is only for the whole file? I mean to have a particular loop
> >> >>vectorized in a
> >> >>file while all others - up to compiler's cost model. is there such a
> >> >>machinery?
> >> >
> >> > No, there is not.
> >> >
> >> > Richard.
> >> >
> >> >>Sergos
> >> >>
> >> >>On Thu, Nov 14, 2013 at 12:39 PM, Richard Biener <rguent...@suse.de>
> >> >>wrote:
> >> >>> On Wed, 13 Nov 2013, Sergey Ostanevich wrote:
> >> >>>
> >> >>>> I will get some tests.
> >> >>>> As for cost analysis - simply consider the pragma as a request to
> >> >>>> vectorize. How can I - as a developer - enforce it beyond the
> >> >>pragma?
> >> >>>
> >> >>> You can disable the cost model via -fvect-cost-model=unlimited
> >> >>>
> >> >>> Richard.
> >> >>>
> >> >>>> On Wed, Nov 13, 2013 at 12:55 PM, Richard Biener <rguent...@suse.de>
> >> >>wrote:
> >> >>>> > On Tue, 12 Nov 2013, Sergey Ostanevich wrote:
> >> >>>> >
> >> >>>> >> The reason patch was in its original state is because we want
> >> >>>> >> to notify user that his assumption of profitability may be wrong.
> >> >>>> >> This is not a part of any spec and as far as I know ICC does not
> >> >>>> >> notify user about the case. Still it can be a good hint for those
> >> >>>> >> users who tries to get as much as possible performance.
> >> >>>> >>
> >> >>>> >> Richard's comment on the vectorization problems is about the same
> >> >>-
> >> >>>> >> to inform user that his attempt to force vectorization is failed.
> >> >>>> >>
> >> >>>> >> As for profitable or not - sometimes I believe it's impossible to
> >> >>be
> >> >>>> >> precise. For OMP we have case of a vector version of a function
> >> >>>> >> and we have no chance to figure out whether it is profitable to
> >> >>use
> >> >>>> >> it or to loose it. If we can't map the loop for any vector length
> >> >>>> >> other than 1 - I believe in this case we have to bail out and
> >> >>report.
> >> >>>> >> Is it about 'never profitable'?
> >> >>>> >
> >> >>>> > For example. I think we should report non-vectorized loops
> >> >>>> > that are marked with force_vect anyway, with
> >> >>-Wdisabled-optimization.
> >> >>>> > Another case is that a loop may be profitable to vectorize if
> >> >>>> > the ISA supports a gather instruction but otherwise not. Or if
> >> >>the
> >> >>>> > ISA supports efficient vector construction from N not loop
> >> >>>> > invariant scalars (for vectorization of strided loads).
> >> >>>> >
> >> >>>> > Simply disregarding all of the cost analysis sounds completely
> >> >>>> > bogus to me.
> >> >>>> >
> >> >>>> > I'd simply go for the diagnostic for now, not changing anything
> >> >>else.
> >> >>>> > We want to have a good understanding about why the cost model is
> >> >>>> > so bad that we have to force to ignore it for #pragma simd - thus
> >> >>we
> >> >>>> > want testcases.
> >> >>>> >
> >> >>>> > Richard.
> >> >>>> >
> >> >>>> >>
> >> >>>> >> On Tue, Nov 12, 2013 at 6:35 PM, Richard Biener
> >> >><rguent...@suse.de> wrote:
> >> >>>> >> > On 11/12/13 3:16 PM, Jakub Jelinek wrote:
> >> >>>> >> >> On Tue, Nov 12, 2013 at 05:46:14PM +0400, Sergey Ostanevich
> >> >>wrote:
> >> >>>> >> >>> ivdep just substitutes all cross-iteration data analysis,
> >> >>>> >> >>> nothing related to cost model. ICC does not cancel its
> >> >>>> >> >>> cost model in case of #pragma ivdep
> >> >>>> >> >>>
> >> >>>> >> >>> as for the safelen - OMP standart treats it as a limitation
> >> >>>> >> >>> for the vector length. this means if no safelen is present
> >> >>>> >> >>> an arbitrary vector length can be used.
> >> >>>> >> >>
> >> >>>> >> >> I was talking about GCC loop->safelen, which is INT_MAX for
> >> >>#pragma omp simd
> >> >>>> >> >> without safelen clause or #pragma simd without vectorlength
> >> >>clause.
> >> >>>> >> >>
> >> >>>> >> >>> so I believe loop->force_vect is the only trigger to
> >> >>disregard
> >> >>>> >> >>> the cost model
> >> >>>> >> >>
> >> >>>> >> >> Anyway, in that case I think the originally posted patch is
> >> >>wrong,
> >> >>>> >> >> if we want to treat force_vect as disregard all the cost model
> >> >>and
> >> >>>> >> >> force vectorization (well, the name of the field already kind
> >> >>of suggest
> >> >>>> >> >> that), then IMHO we should treat it the same as
> >> >>-fvect-cost-model=unlimited
> >> >>>> >> >> for those loops.
> >> >>>> >> >
> >> >>>> >> > Err - the user may have a specific sub-architecture in mind
> >> >>when using
> >> >>>> >> > #pragma simd, if you say we should completely ignore the cost
> >> >>model
> >> >>>> >> > then should we also sorry () if we cannot vectorize the loop
> >> >>(either
> >> >>>> >> > because of GCC deficiencies or lack of sub-target support)?
> >> >>>> >> >
> >> >>>> >> > That said, at least in the cases that the cost model says the
> >> >>loop
> >> >>>> >> > is never profitable to vectorize we should follow its advice.
> >> >>>> >> >
> >> >>>> >> > Richard.
> >> >>>> >> >
> >> >>>> >> >> Thus (untested):
> >> >>>> >> >>
> >> >>>> >> >> 2013-11-12 Jakub Jelinek <ja...@redhat.com>
> >> >>>> >> >>
> >> >>>> >> >> * tree-vect-loop.c (vect_estimate_min_profitable_iters):
> >> >>Use
> >> >>>> >> >> unlimited cost model also for force_vect loops.
> >> >>>> >> >>
> >> >>>> >> >> --- gcc/tree-vect-loop.c.jj 2013-11-12 12:09:40.000000000
> >> >>+0100
> >> >>>> >> >> +++ gcc/tree-vect-loop.c 2013-11-12 15:11:43.821404330
> >> >>+0100
> >> >>>> >> >> @@ -2702,7 +2702,7 @@ vect_estimate_min_profitable_iters (loop
> >> >>>> >> >> void *target_cost_data = LOOP_VINFO_TARGET_COST_DATA
> >> >>(loop_vinfo);
> >> >>>> >> >>
> >> >>>> >> >> /* Cost model disabled. */
> >> >>>> >> >> - if (unlimited_cost_model ())
> >> >>>> >> >> + if (unlimited_cost_model () || LOOP_VINFO_LOOP
> >> >>(loop_vinfo)->force_vect)
> >> >>>> >> >> {
> >> >>>> >> >> dump_printf_loc (MSG_NOTE, vect_location, "cost model
> >> >>disabled.\n");
> >> >>>> >> >> *ret_min_profitable_niters = 0;
> >> >>>> >> >>
> >> >>>> >> >> Jakub
> >> >>>> >> >>
> >> >>>> >> >
> >> >>>> >>
> >> >>>> >>
> >> >>>> >
> >> >>>> > --
> >> >>>> > Richard Biener <rguent...@suse.de>
> >> >>>> > SUSE / SUSE Labs
> >> >>>> > SUSE LINUX Products GmbH - Nuernberg - AG Nuernberg - HRB 16746
> >> >>>> > GF: Jeff Hawn, Jennifer Guild, Felix Imend
> >> >>>>
> >> >>>>
> >> >>>
> >> >>> --
> >> >>> Richard Biener <rguent...@suse.de>
> >> >>> SUSE / SUSE Labs
> >> >>> SUSE LINUX Products GmbH - Nuernberg - AG Nuernberg - HRB 16746
> >> >>> GF: Jeff Hawn, Jennifer Guild, Felix Imend"orffer
> >> >
> >> >
> >>
> >
> > --
> > Richard Biener <rguent...@suse.de>
> > SUSE / SUSE Labs
> > SUSE LINUX Products GmbH - Nuernberg - AG Nuernberg - HRB 16746
> > GF: Jeff Hawn, Jennifer Guild, Felix Imend"orffer
>
>
--
Richard Biener <rguent...@suse.de>
SUSE / SUSE Labs
SUSE LINUX Products GmbH - Nuernberg - AG Nuernberg - HRB 16746
GF: Jeff Hawn, Jennifer Guild, Felix Imend"orffer
