Re: [PATCH 2/2] vect: Enhance cost evaluation in vect_transform_slp_perm_load_1

["Kewen.Lin" <linkw@linux.ibm.com>]

on 2023/5/23 14:19, Richard Biener wrote:
> On Tue, May 23, 2023 at 5:01 AM Kewen.Lin <linkw@linux.ibm.com> wrote:
>>
>> Hi Richi,
>>
>> Thanks for the review!
>>
>> on 2023/5/22 21:44, Richard Biener wrote:
>>> On Wed, May 17, 2023 at 8:15 AM Kewen.Lin <linkw@linux.ibm.com> wrote:
>>>>
>>>> Hi,
>>>>
>>>> Following Richi's suggestion in [1], I'm working on deferring
>>>> cost evaluation next to the transformation, this patch is
>>>> to enhance function vect_transform_slp_perm_load_1 which
>>>> could under-cost for vector permutation, since the costing
>>>> doesn't try to consider nvectors_per_build, it's inconsistent
>>>> with the transformation part.
>>>>
>>>> Bootstrapped and regtested on x86_64-redhat-linux,
>>>> aarch64-linux-gnu and powerpc64{,le}-linux-gnu.
>>>>
>>>> Is it ok for trunk?
>>>>
>>>> [1] https://gcc.gnu.org/pipermail/gcc-patches/2021-January/563624.html
>>>>
>>>> BR,
>>>> Kewen
>>>> -----
>>>> gcc/ChangeLog:
>>>>
>>>>         * tree-vect-slp.cc (vect_transform_slp_perm_load_1): Adjust the
>>>>         calculation on n_perms by considering nvectors_per_build.
>>>>
>>>> gcc/testsuite/ChangeLog:
>>>>
>>>>         * gcc.dg/vect/costmodel/ppc/costmodel-slp-perm.c: New test.
>>>> ---
>>>>  .../vect/costmodel/ppc/costmodel-slp-perm.c   | 23 +++++++
>>>>  gcc/tree-vect-slp.cc                          | 66 ++++++++++---------
>>>>  2 files changed, 57 insertions(+), 32 deletions(-)
>>>>  create mode 100644 gcc/testsuite/gcc.dg/vect/costmodel/ppc/costmodel-slp-perm.c
>>>>
>>>> diff --git a/gcc/testsuite/gcc.dg/vect/costmodel/ppc/costmodel-slp-perm.c b/gcc/testsuite/gcc.dg/vect/costmodel/ppc/costmodel-slp-perm.c
>>>> new file mode 100644
>>>> index 00000000000..e5c4dceddfb
>>>> --- /dev/null
>>>> +++ b/gcc/testsuite/gcc.dg/vect/costmodel/ppc/costmodel-slp-perm.c
>>>> @@ -0,0 +1,23 @@
>>>> +/* { dg-do compile } */
>>>> +/* { dg-require-effective-target vect_int } */
>>>> +/* { dg-require-effective-target powerpc_p9vector_ok } */
>>>> +/* Specify power9 to ensure the vectorization is profitable
>>>> +   and test point stands, otherwise it could be not profitable
>>>> +   to vectorize.  */
>>>> +/* { dg-additional-options "-mdejagnu-cpu=power9 -mpower9-vector" } */
>>>> +
>>>> +/* Verify we cost the exact count for required vec_perm.  */
>>>> +
>>>> +int x[1024], y[1024];
>>>> +
>>>> +void
>>>> +foo ()
>>>> +{
>>>> +  for (int i = 0; i < 512; ++i)
>>>> +    {
>>>> +      x[2 * i] = y[1023 - (2 * i)];
>>>> +      x[2 * i + 1] = y[1023 - (2 * i + 1)];
>>>> +    }
>>>> +}
>>>> +
>>>> +/* { dg-final { scan-tree-dump-times "2 times vec_perm" 1 "vect" } } */
>>>> diff --git a/gcc/tree-vect-slp.cc b/gcc/tree-vect-slp.cc
>>>> index e5c9d7e766e..af9a6dd4fa9 100644
>>>> --- a/gcc/tree-vect-slp.cc
>>>> +++ b/gcc/tree-vect-slp.cc
>>>> @@ -8115,12 +8115,12 @@ vect_transform_slp_perm_load_1 (vec_info *vinfo, slp_tree node,
>>>>
>>>>    mode = TYPE_MODE (vectype);
>>>>    poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype);
>>>> +  unsigned int nstmts = SLP_TREE_NUMBER_OF_VEC_STMTS (node);
>>>>
>>>>    /* Initialize the vect stmts of NODE to properly insert the generated
>>>>       stmts later.  */
>>>>    if (! analyze_only)
>>>> -    for (unsigned i = SLP_TREE_VEC_STMTS (node).length ();
>>>> -        i < SLP_TREE_NUMBER_OF_VEC_STMTS (node); i++)
>>>> +    for (unsigned i = SLP_TREE_VEC_STMTS (node).length (); i < nstmts; i++)
>>>>        SLP_TREE_VEC_STMTS (node).quick_push (NULL);
>>>>
>>>>    /* Generate permutation masks for every NODE. Number of masks for each NODE
>>>> @@ -8161,7 +8161,10 @@ vect_transform_slp_perm_load_1 (vec_info *vinfo, slp_tree node,
>>>>          (b) the permutes only need a single vector input.  */
>>>>        mask.new_vector (nunits, group_size, 3);
>>>>        nelts_to_build = mask.encoded_nelts ();
>>>> -      nvectors_per_build = SLP_TREE_VEC_STMTS (node).length ();
>>>> +      /* It's possible to obtain zero nstmts during analyze_only, so make
>>>> +        it at least one to ensure the later computation for n_perms
>>>> +        proceed.  */
>>>> +      nvectors_per_build = nstmts > 0 ? nstmts : 1;
>>>>        in_nlanes = DR_GROUP_SIZE (stmt_info) * 3;
>>>>      }
>>>>    else
>>>> @@ -8252,40 +8255,39 @@ vect_transform_slp_perm_load_1 (vec_info *vinfo, slp_tree node,
>>>>                   return false;
>>>>                 }
>>>>
>>>> -             ++*n_perms;
>>>> -
>>>> +             tree mask_vec = NULL_TREE;
>>>>               if (!analyze_only)
>>>> -               {
>>>> -                 tree mask_vec = vect_gen_perm_mask_checked (vectype, indices);
>>>> +               mask_vec = vect_gen_perm_mask_checked (vectype, indices);
>>>>
>>>> -                 if (second_vec_index == -1)
>>>> -                   second_vec_index = first_vec_index;
>>>> +             if (second_vec_index == -1)
>>>> +               second_vec_index = first_vec_index;
>>>>
>>>> -                 for (unsigned int ri = 0; ri < nvectors_per_build; ++ri)
>>>> +             for (unsigned int ri = 0; ri < nvectors_per_build; ++ri)
>>>> +               {
>>>> +                 ++*n_perms;
>>>
>>> So the "real" change is doing
>>>
>>>       *n_perms += nvectors_per_build;
>>>
>>> and *n_perms was unused when !analyze_only?  And since at
>>
>> Yes, although both !analyze_only and analyze_only calls pass n_perms in, now
>> only the call sites with analyze_only will use the returned n_perms further.
>>
>>> analysis time we (sometimes?) have zero nvectors you have to
>>> fixup above?  Which cases are that?
>>
>> Yes, the fixup is to avoid to result in unexpected n_perms in function
>> vect_optimize_slp_pass::internal_node_cost。  One typical case is
>> gcc.dg/vect/bb-slp-50.c, without special casing zero, slp2 fails to optimize
>> out one more vec_perm unexpectedly.
>>
>> In vect_optimize_slp_pass::internal_node_cost, it checks if the returned n_perms
>> is zero or not (vec_perm not needed or needed).
>>
>>       if (!vect_transform_slp_perm_load_1 (m_vinfo, node, tmp_perm, vNULL,
>>                                            nullptr, vf, true, false, &n_perms))
>>         {
>>           auto rep = SLP_TREE_REPRESENTATIVE (node);
>>           if (out_layout_i == 0)
>>             {
>>               /* Use the fallback cost if the load is an N-to-N permutation.
>>                  Otherwise assume that the node will be rejected later
>>                  and rebuilt from scalars.  */
>>               if (STMT_VINFO_GROUPED_ACCESS (rep)
>>                   && (DR_GROUP_SIZE (DR_GROUP_FIRST_ELEMENT (rep))
>>                       == SLP_TREE_LANES (node)))
>>                 return fallback_cost;
>>               return 0;
>>             }
>>           return -1;
>>         }
>>
>>       /* See the comment above the corresponding VEC_PERM_EXPR handling.  */
>>       return n_perms == 0 ? 0 : 1;
>>
>> In vect_optimize_slp_pass::forward_pass (), it only considers the case that
>> factor > 0 (there is some vec_perm needed).
>>
>>               /* Accumulate the cost of using LAYOUT_I within NODE,
>>                  both for the inputs and the outputs.  */
>>               int factor = internal_node_cost (vertex.node, layout_i,
>>                                                layout_i);
>>               if (factor < 0)
>>                 {
>>                   is_possible = false;
>>                   break;
>>                 }
>>               else if (factor)
>>                 layout_costs.internal_cost.add_serial_cost
>>                   ({ vertex.weight * factor, m_optimize_size });
> 
> Ah, OK - thanks for clarifying.
> 
> The patch is OK.

Thanks!  Committed in r14-1151.

BR,
Kewen