On Mon, 13 May 2024, Tamar Christina wrote:
> > -----Original Message-----
> > From: Richard Biener <richard.guent...@gmail.com>
> > Sent: Friday, May 10, 2024 2:07 PM
> > To: Richard Biener <rguent...@suse.de>
> > Cc: gcc-patches@gcc.gnu.org
> > Subject: Re: [PATCH] Allow patterns in SLP reductions
> >
> > On Fri, Mar 1, 2024 at 10:21 AM Richard Biener <rguent...@suse.de> wrote:
> > >
> > > The following removes the over-broad rejection of patterns for SLP
> > > reductions which is done by removing them from LOOP_VINFO_REDUCTIONS
> > > during pattern detection. That's also insufficient in case the
> > > pattern only appears on the reduction path. Instead this implements
> > > the proper correctness check in vectorizable_reduction and guides
> > > SLP discovery to heuristically avoid forming later invalid groups.
> > >
> > > I also couldn't find any testcase that FAILs when allowing the SLP
> > > reductions to form so I've added one.
> > >
> > > I came across this for single-lane SLP reductions with the all-SLP
> > > work where we rely on patterns to properly vectorize COND_EXPR
> > > reductions.
> > >
> > > Bootstrapped and tested on x86_64-unknown-linux-gnu, queued for stage1.
> >
> > Re-bootstrapped/tested, r15-361-g52d4691294c847
>
> Awesome!
>
> Does this now allow us to write new reductions using patterns? i.e.
> widening reductions?
Yes (SLP reductions, that is). This is really only for SLP reductions
(not SLP reduction chains, not non-SLP reductions). So it's just
a corner-case but since with SLP-only non-SLP reductions become
SLP reductions with a single lane that was important to fix ;)
Richard.
> Cheers,
> Tamar
> >
> > Richard.
> >
> > > Richard.
> > >
> > > * tree-vect-patterns.cc (vect_pattern_recog_1): Do not
> > > remove reductions involving patterns.
> > > * tree-vect-loop.cc (vectorizable_reduction): Reject SLP
> > > reduction groups with multiple lane-reducing reductions.
> > > * tree-vect-slp.cc (vect_analyze_slp_instance): When discovering
> > > SLP reduction groups avoid including lane-reducing ones.
> > >
> > > * gcc.dg/vect/vect-reduc-sad-9.c: New testcase.
> > > ---
> > > gcc/testsuite/gcc.dg/vect/vect-reduc-sad-9.c | 68 ++++++++++++++++++++
> > > gcc/tree-vect-loop.cc | 15 +++++
> > > gcc/tree-vect-patterns.cc | 13 ----
> > > gcc/tree-vect-slp.cc | 26 +++++---
> > > 4 files changed, 101 insertions(+), 21 deletions(-)
> > > create mode 100644 gcc/testsuite/gcc.dg/vect/vect-reduc-sad-9.c
> > >
> > > diff --git a/gcc/testsuite/gcc.dg/vect/vect-reduc-sad-9.c
> > b/gcc/testsuite/gcc.dg/vect/vect-reduc-sad-9.c
> > > new file mode 100644
> > > index 00000000000..3c6af4510f4
> > > --- /dev/null
> > > +++ b/gcc/testsuite/gcc.dg/vect/vect-reduc-sad-9.c
> > > @@ -0,0 +1,68 @@
> > > +/* Disabling epilogues until we find a better way to deal with scans. */
> > > +/* { dg-additional-options "--param vect-epilogues-nomask=0" } */
> > > +/* { dg-additional-options "-msse4.2" { target { x86_64-*-* i?86-*-* } }
> > > } */
> > > +/* { dg-require-effective-target vect_usad_char } */
> > > +
> > > +#include <stdarg.h>
> > > +#include "tree-vect.h"
> > > +
> > > +#define N 64
> > > +
> > > +unsigned char X[N] __attribute__ ((__aligned__(__BIGGEST_ALIGNMENT__)));
> > > +unsigned char Y[N] __attribute__ ((__aligned__(__BIGGEST_ALIGNMENT__)));
> > > +int abs (int);
> > > +
> > > +/* Sum of absolute differences between arrays of unsigned char types.
> > > + Detected as a sad pattern.
> > > + Vectorized on targets that support sad for unsigned chars. */
> > > +
> > > +__attribute__ ((noinline)) int
> > > +foo (int len, int *res2)
> > > +{
> > > + int i;
> > > + int result = 0;
> > > + int result2 = 0;
> > > +
> > > + for (i = 0; i < len; i++)
> > > + {
> > > + /* Make sure we are not using an SLP reduction for this. */
> > > + result += abs (X[2*i] - Y[2*i]);
> > > + result2 += abs (X[2*i + 1] - Y[2*i + 1]);
> > > + }
> > > +
> > > + *res2 = result2;
> > > + return result;
> > > +}
> > > +
> > > +
> > > +int
> > > +main (void)
> > > +{
> > > + int i;
> > > + int sad;
> > > +
> > > + check_vect ();
> > > +
> > > + for (i = 0; i < N/2; i++)
> > > + {
> > > + X[2*i] = i;
> > > + Y[2*i] = N/2 - i;
> > > + X[2*i+1] = i;
> > > + Y[2*i+1] = 0;
> > > + __asm__ volatile ("");
> > > + }
> > > +
> > > +
> > > + int sad2;
> > > + sad = foo (N/2, &sad2);
> > > + if (sad != (N/2)*(N/4))
> > > + abort ();
> > > + if (sad2 != (N/2-1)*(N/2)/2)
> > > + abort ();
> > > +
> > > + return 0;
> > > +}
> > > +
> > > +/* { dg-final { scan-tree-dump "vect_recog_sad_pattern: detected" "vect"
> > > } } */
> > > +/* { dg-final { scan-tree-dump-times "vectorized 1 loops" 1 "vect" } } */
> > > +
> > > diff --git a/gcc/tree-vect-loop.cc b/gcc/tree-vect-loop.cc
> > > index 35f1f8c7d42..13dcdba403a 100644
> > > --- a/gcc/tree-vect-loop.cc
> > > +++ b/gcc/tree-vect-loop.cc
> > > @@ -7703,6 +7703,21 @@ vectorizable_reduction (loop_vec_info loop_vinfo,
> > > return false;
> > > }
> > >
> > > + /* Lane-reducing ops also never can be used in a SLP reduction group
> > > + since we'll mix lanes belonging to different reductions. But it's
> > > + OK to use them in a reduction chain or when the reduction group
> > > + has just one element. */
> > > + if (lane_reduc_code_p
> > > + && slp_node
> > > + && !REDUC_GROUP_FIRST_ELEMENT (stmt_info)
> > > + && SLP_TREE_LANES (slp_node) > 1)
> > > + {
> > > + if (dump_enabled_p ())
> > > + dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
> > > + "lane-reducing reduction in reduction group.\n");
> > > + return false;
> > > + }
> > > +
> > > /* All uses but the last are expected to be defined in the loop.
> > > The last use is the reduction variable. In case of nested cycle
> > > this
> > > assumption is not true: we use reduc_index to record the index of
> > > the
> > > diff --git a/gcc/tree-vect-patterns.cc b/gcc/tree-vect-patterns.cc
> > > index d562f57920f..fe1ffba8688 100644
> > > --- a/gcc/tree-vect-patterns.cc
> > > +++ b/gcc/tree-vect-patterns.cc
> > > @@ -7172,7 +7172,6 @@ vect_pattern_recog_1 (vec_info *vinfo,
> > > vect_recog_func *recog_func, stmt_vec_info
> > > stmt_info)
> > > {
> > > gimple *pattern_stmt;
> > > - loop_vec_info loop_vinfo;
> > > tree pattern_vectype;
> > >
> > > /* If this statement has already been replaced with pattern statements,
> > > @@ -7198,8 +7197,6 @@ vect_pattern_recog_1 (vec_info *vinfo,
> > > return;
> > > }
> > >
> > > - loop_vinfo = dyn_cast <loop_vec_info> (vinfo);
> > > -
> > > /* Found a vectorizable pattern. */
> > > if (dump_enabled_p ())
> > > dump_printf_loc (MSG_NOTE, vect_location,
> > > @@ -7208,16 +7205,6 @@ vect_pattern_recog_1 (vec_info *vinfo,
> > >
> > > /* Mark the stmts that are involved in the pattern. */
> > > vect_mark_pattern_stmts (vinfo, stmt_info, pattern_stmt,
> > > pattern_vectype);
> > > -
> > > - /* Patterns cannot be vectorized using SLP, because they change the
> > > order of
> > > - computation. */
> > > - if (loop_vinfo)
> > > - {
> > > - unsigned ix, ix2;
> > > - stmt_vec_info *elem_ptr;
> > > - VEC_ORDERED_REMOVE_IF (LOOP_VINFO_REDUCTIONS (loop_vinfo), ix,
> > ix2,
> > > - elem_ptr, *elem_ptr == stmt_info);
> > > - }
> > > }
> > >
> > >
> > > diff --git a/gcc/tree-vect-slp.cc b/gcc/tree-vect-slp.cc
> > > index dabd8407aaf..d9961945c1c 100644
> > > --- a/gcc/tree-vect-slp.cc
> > > +++ b/gcc/tree-vect-slp.cc
> > > @@ -3597,14 +3597,24 @@ vect_analyze_slp_instance (vec_info *vinfo,
> > > = as_a <loop_vec_info> (vinfo)->reductions;
> > > scalar_stmts.create (reductions.length ());
> > > for (i = 0; reductions.iterate (i, &next_info); i++)
> > > - if ((STMT_VINFO_RELEVANT_P (next_info)
> > > - || STMT_VINFO_LIVE_P (next_info))
> > > - /* ??? Make sure we didn't skip a conversion around a
> > > reduction
> > > - path. In that case we'd have to reverse engineer that
> > > conversion
> > > - stmt following the chain using reduc_idx and from the PHI
> > > - using reduc_def. */
> > > - && STMT_VINFO_DEF_TYPE (next_info) == vect_reduction_def)
> > > - scalar_stmts.quick_push (next_info);
> > > + {
> > > + gassign *g;
> > > + next_info = vect_stmt_to_vectorize (next_info);
> > > + if ((STMT_VINFO_RELEVANT_P (next_info)
> > > + || STMT_VINFO_LIVE_P (next_info))
> > > + /* ??? Make sure we didn't skip a conversion around a
> > > reduction
> > > + path. In that case we'd have to reverse engineer that
> > > + conversion stmt following the chain using reduc_idx and
> > > from
> > > + the PHI using reduc_def. */
> > > + && STMT_VINFO_DEF_TYPE (next_info) == vect_reduction_def
> > > + /* Do not discover SLP reductions for lane-reducing ops,
> > > that
> > > + will fail later. */
> > > + && (!(g = dyn_cast <gassign *> (STMT_VINFO_STMT
> > > (next_info)))
> > > + || (gimple_assign_rhs_code (g) != DOT_PROD_EXPR
> > > + && gimple_assign_rhs_code (g) != WIDEN_SUM_EXPR
> > > + && gimple_assign_rhs_code (g) != SAD_EXPR)))
> > > + scalar_stmts.quick_push (next_info);
> > > + }
> > > /* If less than two were relevant/live there's nothing to SLP. */
> > > if (scalar_stmts.length () < 2)
> > > return false;
> > > --
> > > 2.35.3
>
--
Richard Biener <rguent...@suse.de>
SUSE Software Solutions Germany GmbH,
Frankenstrasse 146, 90461 Nuernberg, Germany;
GF: Ivo Totev, Andrew McDonald, Werner Knoblich; (HRB 36809, AG Nuernberg)
