| Summary: | Fortran Complex reduction / multiplication not vectorized | ||
|---|---|---|---|
| Product: | gcc | Reporter: | Richard Biener <rguenth> |
| Component: | tree-optimization | Assignee: | Richard Biener <rguenth> |
| Status: | RESOLVED FIXED | ||
| Severity: | enhancement | CC: | bill.schmidt, dje, gcc-bugs, irar, pthaugen, sebastian.hegler, vincenzo.innocente |
| Priority: | P3 | Keywords: | alias, missed-optimization |
| Version: | 4.4.0 | ||
| Target Milestone: | 6.0 | ||
| Host: | Target: | ||
| Build: | Known to work: | ||
| Known to fail: | Last reconfirmed: | 2009-01-21 15:43:08 | |
| Bug Depends on: | 43434, 54939, 56766, 56902 | ||
| Bug Blocks: | 53947 | ||
| Attachments: | tree-slp-details dump | ||
Patch for 1) http://gcc.gnu.org/ml/gcc-patches/2008-08/msg00221.html Patch for 2) http://gcc.gnu.org/ml/gcc-patches/2008-08/msg00226.html I didn't yet start on 3), so 4) is unknown yet (as is 5, 6, ... ;)) 3) is because data-ref requires a constant step
else if (!simple_iv (loop, stmt, poffset, &offset_iv, false))
{
if (dump_file && (dump_flags & TDF_DETAILS))
fprintf (dump_file, "failed: evolution of offset is not affine.\n");
return;
but the step is (<unnamed-signed:64>) ((<unnamed-unsigned:64>) stride.3_36 * 16)
as we are dealing with general incoming arrays which are arbitrary striped.
Fixing this requires for example versioning for a constant stride.
Mine. I am working on adding versioning for non-constant strides. The testcase should be
subroutine to_product_of(self,a,b,a1,a2)
complex(kind=8) :: self (:)
complex(kind=8), intent(in) :: a(:,:)
complex(kind=8), intent(in) :: b(:)
integer a1,a2
do i = 1,a1
do j = 1,a2
self(i) = self(i) + a(j,i)*b(j)
end do
end do
end subroutine
to be meaningful - otherwise we are accessing a in non-continuous ways in the
inner loop which would prevent vectorization.
With the versioning for stride == 1 I get then
.L13:
movupd 16(%rax), %xmm1
movupd (%rax), %xmm3
incl %ecx
movupd (%rdx), %xmm4
addq $32, %rax
movapd %xmm3, %xmm0
unpckhpd %xmm1, %xmm3
unpcklpd %xmm1, %xmm0
movupd 16(%rdx), %xmm1
movapd %xmm4, %xmm2
addq $32, %rdx
movapd %xmm3, %xmm9
cmpl %ecx, %r8d
unpcklpd %xmm1, %xmm2
unpckhpd %xmm1, %xmm4
movapd %xmm4, %xmm1
movapd %xmm2, %xmm4
mulpd %xmm1, %xmm9
mulpd %xmm0, %xmm4
mulpd %xmm3, %xmm2
mulpd %xmm1, %xmm0
subpd %xmm9, %xmm4
addpd %xmm2, %xmm0
addpd %xmm4, %xmm6
addpd %xmm0, %xmm5
ja .L13
haddpd %xmm5, %xmm5
cmpl %r15d, %edi
movl -4(%rsp), %ecx
haddpd %xmm6, %xmm6
addsd %xmm5, %xmm8
addsd %xmm6, %xmm7
jne .L12
jmp .L14
for the innermost loop, followed by a tail loop (peel for niters). This is
about 15% faster on AMD K10 than the non-vectorized loop (if you disable
the cost-model and make sure to have enough iterations in the inner loop
to pay back for the extra guarding conditions).
So,
4) The vectorized version sucks because we have to use peeling for niters
because we need to unroll the loop once and cannot apply SLP here.
Q1: does SLP work with reductions at all?
Q2: does SLP do pattern recognition?
First of all we would need to recognize a complex reduction as a single
vectorized reduction. Second we need to vectorize the complex multiplication
with SLP, feeding the reduction with one resulting complex vector.
(In reply to comment #5) > So, > 4) The vectorized version sucks because we have to use peeling for niters > because we need to unroll the loop once and cannot apply SLP here. What do you mean by "unroll the loop once"? > Q1: does SLP work with reductions at all? No. SLP currently originates from groups of strided stores. > Q2: does SLP do pattern recognition? Pattern recoginition is done before SLP, and SLP handles stmts that were marked as a part of a pattern. There is no SLP specific pattern recoginition. > First of all we would need to recognize a complex reduction as a single > vectorized reduction. Second we need to vectorize the complex multiplication > with SLP, feeding the reduction with one resulting complex vector. Subject: Re: Fortran Complex reduction / multiplication not vectorized On Sun, 25 Jan 2009, irar at il dot ibm dot com wrote: > > > ------- Comment #6 from irar at il dot ibm dot com 2009-01-25 09:12 ------- > (In reply to comment #5) > > So, > > 4) The vectorized version sucks because we have to use peeling for niters > > because we need to unroll the loop once and cannot apply SLP here. > > What do you mean by "unroll the loop once"? The vectorization factor is two, so we need two copies of the loop body (which means unrolling it once and creating an epilogue loop because niter may be odd) > > Q1: does SLP work with reductions at all? > > No. SLP currently originates from groups of strided stores. Ah, I see. In this loop we have two reductions, so to apply SLP we would need to see that we can use a group of reductions for SLP? > > Q2: does SLP do pattern recognition? > > Pattern recoginition is done before SLP, and SLP handles stmts that were marked > as a part of a pattern. There is no SLP specific pattern recoginition. Ok, but with a reduction it won't help me here. Can a loop be vectorized with just pattern recognition? Hm, if I remember correctly we detect scalar patterns and then vectorize them. We don't support detecting "vector patterns" from scalar code, correct? Thanks, Richard. (In reply to comment #7) > > > Q1: does SLP work with reductions at all? > > > > No. SLP currently originates from groups of strided stores. > Ah, I see. In this loop we have two reductions, so to apply SLP > we would need to see that we can use a group of reductions for SLP? Yes, I think this will work. > > > Q2: does SLP do pattern recognition? > > > > Pattern recoginition is done before SLP, and SLP handles stmts that were marked > > as a part of a pattern. There is no SLP specific pattern recoginition. > Ok, but with a reduction it won't help me here. > Can a loop be vectorized with just pattern recognition? Hm, if I > remember correctly we detect scalar patterns and then vectorize them. > We don't support detecting "vector patterns" from scalar code, correct? Yes, if I understand you correctly, we detect scalar patterns, but adding vector pattern detection does not seem to be complicated. (In reply to comment #4) > The testcase should be > subroutine to_product_of(self,a,b,a1,a2) > complex(kind=8) :: self (:) > complex(kind=8), intent(in) :: a(:,:) > complex(kind=8), intent(in) :: b(:) > integer a1,a2 > do i = 1,a1 > do j = 1,a2 > self(i) = self(i) + a(j,i)*b(j) > end do > end do > end subroutine > to be meaningful - otherwise we are accessing a in non-continuous ways in the > inner loop which would prevent vectorization. this change from a(i,j) to a(j,i) is not required if we try to vectorize the outer-loop, where the stride is 1. It's also a better way to vectorize the reduction. A few limitations on the way though are: 1) somehow don't let gcc create guard code around the innermost loop to check that it executes more than zero iterations. This creates a complicated control flow structure within the outer-loop. For now you have to have constant number of iterations for the inner-loop because of that, or insert a statement like "if (a2<=0) return;" before the loop... 2) use -fno-tree-sink cause otherwise it moves the loop iv increment to the latch block and the vectorizer likes to have the latch block empty... (see also PR33113 for related reference). > With the versioning for stride == 1 I get then > .L13: > movupd 16(%rax), %xmm1 > movupd (%rax), %xmm3 > incl %ecx > movupd (%rdx), %xmm4 > addq $32, %rax > movapd %xmm3, %xmm0 > unpckhpd %xmm1, %xmm3 > unpcklpd %xmm1, %xmm0 > movupd 16(%rdx), %xmm1 > movapd %xmm4, %xmm2 > addq $32, %rdx > movapd %xmm3, %xmm9 > cmpl %ecx, %r8d > unpcklpd %xmm1, %xmm2 > unpckhpd %xmm1, %xmm4 > movapd %xmm4, %xmm1 > movapd %xmm2, %xmm4 > mulpd %xmm1, %xmm9 > mulpd %xmm0, %xmm4 > mulpd %xmm3, %xmm2 > mulpd %xmm1, %xmm0 > subpd %xmm9, %xmm4 > addpd %xmm2, %xmm0 > addpd %xmm4, %xmm6 > addpd %xmm0, %xmm5 > ja .L13 > haddpd %xmm5, %xmm5 > cmpl %r15d, %edi > movl -4(%rsp), %ecx > haddpd %xmm6, %xmm6 > addsd %xmm5, %xmm8 > addsd %xmm6, %xmm7 > jne .L12 > jmp .L14 > for the innermost loop, followed by a tail loop (peel for niters). This is > about 15% faster on AMD K10 than the non-vectorized loop (if you disable > the cost-model and make sure to have enough iterations in the inner loop > to pay back for the extra guarding conditions). This one, as well as PR 33133, should be handled by "-floop-interchange". Fortran is row-major, so interchanging inner and outer loop would allow the loops to be coalesced into one, which in turn should be easily vectorized (if complex numbers can be vectorized, see PR 40770). Can you please give me some hints on how to find out if "-floop-interchange" actually does that? Thanks! Forget that about folding stuff into one loop, I didn't have my morning coffee yet. However, the rest still applies. I'm looking forward to some help in that regard. Thanks. On Fri, 25 Mar 2011, sebastian.hegler@tu-dresden.de wrote: > http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37021 > > --- Comment #11 from sebastian.hegler@tu-dresden.de 2011-03-25 11:38:37 UTC --- > Forget that about folding stuff into one loop, I didn't have my morning coffee > yet. However, the rest still applies. > > I'm looking forward to some help in that regard. Look at dump files (-fdump-tree-all-details). Richard. Link to vectorizer missed-optimization meta-bug. The following testcase shows the issue well:
_Complex double self[1024];
_Complex double a[1024][1024];
_Complex double b[1024];
void foo (void)
{
int i, j;
for (i = 0; i < 1024; i+=3)
for (j = 0; j < 1024; j+=3)
self[i] = self[i] + a[i][j]*b[j];
}
we have to get the complex multiplication pattern recognized by SLP
which looks like (without PRE):
<bb 3>:
<bb 4>:
# j_21 = PHI <j_13(3), 0(7)>
# self_I_RE_lsm.2_12 = PHI <_26(3), self_I_RE_lsm.2_7(7)>
# self_I_IM_lsm.3_28 = PHI <_27(3), self_I_IM_lsm.3_8(7)>
# ivtmp_16 = PHI <ivtmp_1(3), 342(7)>
_2 = REALPART_EXPR <a[i_20][j_21]>;
_18 = IMAGPART_EXPR <a[i_20][j_21]>;
_19 = REALPART_EXPR <b[j_21]>;
_17 = IMAGPART_EXPR <b[j_21]>;
_4 = _19 * _2;
_3 = _18 * _17;
_6 = _17 * _2;
_23 = _19 * _18;
_24 = _4 - _3;
_25 = _23 + _6;
_26 = _24 + self_I_RE_lsm.2_12;
_27 = _25 + self_I_IM_lsm.3_28;
j_13 = j_21 + 3;
ivtmp_1 = ivtmp_16 - 1;
if (ivtmp_1 != 0)
goto <bb 3>;
we fail to build the SLP tree for _25 = _23 + _6 because the matching
stmt is _24 = _4 - _3 which has a different operation (SSE4 addsub
would support vectorizing this). I don't see how we can easily
make this supported with the current pattern support ... the
support doesn't allow tieing together two SLP group members.
Simply allowing analysis to proceeed here reveals the fact that
the interleaving has a gap of 6 which makes the analysis fail.
Allowing it to proceed for ncopies == 1 (thus no actual interleaving
required) reveals the next check is slightly bogus in that case.
Fixing that ends us with
t.c:9: note: Load permutation 0 0 1 0 1 1 0 1
t.c:9: note: Build SLP failed: unsupported load permutation _27 = _25 + self_I_IM_lsm.3_28;
... (to be continued)
Author: rguenth Date: Wed Mar 27 10:38:29 2013 New Revision: 197158 URL: http://gcc.gnu.org/viewcvs?rev=197158&root=gcc&view=rev Log: 2013-03-27 Richard Biener <rguenther@suse.de> PR tree-optimization/37021 * tree-vect-data-refs.c (vect_check_strided_load): Allow REALPART/IMAGPART_EXPRs around the supported refs. * tree-ssa-structalias.c (find_func_aliases): Assume that floating-point values are not used to transfer pointers. * gfortran.dg/vect/fast-math-pr37021.f90: New testcase. Added: trunk/gcc/testsuite/gfortran.dg/vect/fast-math-pr37021.f90 Modified: trunk/gcc/ChangeLog trunk/gcc/testsuite/ChangeLog trunk/gcc/tree-ssa-structalias.c trunk/gcc/tree-vect-data-refs.c We now vectorize this testcase by means of using strided loads, relying on store motion turning the store to self(i) in the innermost look into a reduction (no support for vectorized strided stores). The test gfortran.dg/vect/fast-math-pr37021.f90 fails on powerpc*-* (see http://gcc.gnu.org/ml/gcc-testresults/2013-04/msg00677.html ). Isn't it expected? AFAICT doubles are not vectorized, at least on a G5. Author: rguenth Date: Tue May 12 11:55:40 2015 New Revision: 223059 URL: https://gcc.gnu.org/viewcvs?rev=223059&root=gcc&view=rev Log: 2015-05-12 Richard Biener <rguenther@suse.de> PR tree-optimization/37021 * tree-vectorizer.h (struct _slp_tree): Add two_operators flag. (SLP_TREE_TWO_OPERATORS): New define. * tree-vect-slp.c (vect_create_new_slp_node): Initialize SLP_TREE_TWO_OPERATORS. (vect_build_slp_tree_1): Allow two mixing plus/minus in an SLP node. (vect_build_slp_tree): Adjust. (vect_analyze_slp_cost_1): Likewise. (vect_schedule_slp_instance): Vectorize mixing plus/minus by emitting two vector stmts and mixing the results. * gcc.target/i386/vect-addsub.c: New testcase. Added: trunk/gcc/testsuite/gcc.target/i386/vect-addsub.c Modified: trunk/gcc/ChangeLog trunk/gcc/testsuite/ChangeLog trunk/gcc/tree-vect-slp.c trunk/gcc/tree-vectorizer.h We now vectorize the original testcase with SLP. There is still PR56766 which causes us to fail to use addsubpd on x86_64 with SSE3. We still don't vectorize the original code example on Power. It appears that this is being disabled because of an alignment issue. The data references are being rejected by:
product.f:9:0: note: can't force alignment of ref: REALPART_EXPR <*a.0_24[_50]>
and similar for the other three DRs. This happens due to this code in vect_compute_data_ref_alignment:
if (base_alignment < TYPE_ALIGN (vectype))
{
/* Strip an inner MEM_REF to a bare decl if possible. */
if (TREE_CODE (base) == MEM_REF
&& integer_zerop (TREE_OPERAND (base, 1))
&& TREE_CODE (TREE_OPERAND (base, 0)) == ADDR_EXPR)
base = TREE_OPERAND (TREE_OPERAND (base, 0), 0);
if (!vect_can_force_dr_alignment_p (base, TYPE_ALIGN (vectype)))
{
if (dump_enabled_p ())
{
dump_printf_loc (MSG_NOTE, vect_location,
"can't force alignment of ref: ");
dump_generic_expr (MSG_NOTE, TDF_SLIM, ref);
dump_printf (MSG_NOTE, "\n");
}
return true;
}
Here TYPE_ALIGN (vectype) is 128 (Power vectors are normally aligned on a 128-bit value), and base_alignment is 64. a.0 is defined as:
complex(kind=8) [0:D.1831] * restrict a.0;
In both ELFv1 and ELFv2 ABIs for Power, a complex type is defined to have the same alignment as the underlying type. So "complex double" has 8-byte alignment.
On earlier versions of Power, the decision is fine, because unaligned accesses are expensive prior to POWER8. With POWER8, though, an unaligned access will (most of the time) perform as well as an aligned access. So ideally we would like to teach the vectorizer to allow vectorization here.
It seems like vect_supportable_dr_alignment ought to be considered as part of the SLP vectorization decision here, rather than just comparing the base alignment with the vector type alignment. Adding a check for that allows things to get a little further, but we still don't vectorize the block. (I haven't yet looked into why, but I assume more needs to be done downstream to handle this case.)
My understanding of the vectorizer is not yet very deep, so before going too far down the wrong path, I'd like your opinion on the best approach to fixing the problem. Thanks!
Bill
(In reply to Bill Schmidt from comment #20) > We still don't vectorize the original code example on Power. It appears > that this is being disabled because of an alignment issue. The data > references are being rejected by: > > product.f:9:0: note: can't force alignment of ref: REALPART_EXPR > <*a.0_24[_50]> > > and similar for the other three DRs. This happens due to this code in > vect_compute_data_ref_alignment: > > if (base_alignment < TYPE_ALIGN (vectype)) > { > /* Strip an inner MEM_REF to a bare decl if possible. */ > if (TREE_CODE (base) == MEM_REF > && integer_zerop (TREE_OPERAND (base, 1)) > && TREE_CODE (TREE_OPERAND (base, 0)) == ADDR_EXPR) > base = TREE_OPERAND (TREE_OPERAND (base, 0), 0); > > if (!vect_can_force_dr_alignment_p (base, TYPE_ALIGN (vectype))) > { > if (dump_enabled_p ()) > { > dump_printf_loc (MSG_NOTE, vect_location, > "can't force alignment of ref: "); > dump_generic_expr (MSG_NOTE, TDF_SLIM, ref); > dump_printf (MSG_NOTE, "\n"); > } > return true; > } > > Here TYPE_ALIGN (vectype) is 128 (Power vectors are normally aligned on a > 128-bit value), and base_alignment is 64. a.0 is defined as: > > complex(kind=8) [0:D.1831] * restrict a.0; > > In both ELFv1 and ELFv2 ABIs for Power, a complex type is defined to have > the same alignment as the underlying type. So "complex double" has 8-byte > alignment. > > On earlier versions of Power, the decision is fine, because unaligned > accesses are expensive prior to POWER8. With POWER8, though, an unaligned > access will (most of the time) perform as well as an aligned access. So > ideally we would like to teach the vectorizer to allow vectorization here. > > It seems like vect_supportable_dr_alignment ought to be considered as part > of the SLP vectorization decision here, rather than just comparing the base > alignment with the vector type alignment. Adding a check for that allows > things to get a little further, but we still don't vectorize the block. (I > haven't yet looked into why, but I assume more needs to be done downstream > to handle this case.) > > My understanding of the vectorizer is not yet very deep, so before going too > far down the wrong path, I'd like your opinion on the best approach to > fixing the problem. Thanks! I see it only failing due to cost issues (tried ppc64le and -mcpu=power8). The unaligned loads cost 3 and we end up with t.f90:8:0: note: Cost model analysis: Vector inside of loop cost: 40 Vector prologue cost: 8 Vector epilogue cost: 4 Scalar iteration cost: 12 Scalar outside cost: 6 Vector outside cost: 12 prologue iterations: 0 epilogue iterations: 0 t.f90:8:0: note: cost model: the vector iteration cost = 40 divided by the scalar iteration cost = 12 is greater or equal to the vectorization factor = 1. Note that we are (still) not very good in estimating the SLP cost as we account 4 vector loads here (because we essentially will end up with 4 different permutations used), so the "unaligned" part is accounted for too much and likely the permutation cost as well. Both are a limitation of the SLP data structures and not easily fixable. With -fvect-cost-model=unlimited I see both loops vectorized. > Bill (In reply to Richard Biener from comment #21) > (In reply to Bill Schmidt from comment #20) ...<snip>... > > I see it only failing due to cost issues (tried ppc64le and -mcpu=power8). > The unaligned loads cost 3 and we end up with > > t.f90:8:0: note: Cost model analysis: > Vector inside of loop cost: 40 > Vector prologue cost: 8 > Vector epilogue cost: 4 > Scalar iteration cost: 12 > Scalar outside cost: 6 > Vector outside cost: 12 > prologue iterations: 0 > epilogue iterations: 0 > t.f90:8:0: note: cost model: the vector iteration cost = 40 divided by the > scalar iteration cost = 12 is greater or equal to the vectorization factor = > 1. > > Note that we are (still) not very good in estimating the SLP cost as we > account 4 vector loads here (because we essentially will end up with > 4 different permutations used), so the "unaligned" part is accounted for > too much and likely the permutation cost as well. Both are a limitation > of the SLP data structures and not easily fixable. With > -fvect-cost-model=unlimited I see both loops vectorized. Yes, I get these same results for the loop vectorizer (using -O2 -ftree-vectorize -mcpu=power8 -ffast-math). But I was looking at the failure to do SLP vectorization. In comment 19 you indicated this was now working, presumably on x86, but for Power we fail to SLP-vectorize fast-math-pr37021.f90:9:0. However, with today's trunk my SLP dump looks slightly different so I need to have another look at whether this is still failing due to alignment or something else. I'll comment again when I've dug into it further. Created attachment 36261 [details]
tree-slp-details dump
Ah, I was looking at the code in the test suite this time, rather than the raw posted code, so the line numbers changed for the dejagnu commands. The statement number is now 12.
Attaching the details dump for SLP.
On Thu, 27 Aug 2015, wschmidt at gcc dot gnu.org wrote:
> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=37021
>
> --- Comment #22 from Bill Schmidt <wschmidt at gcc dot gnu.org> ---
> (In reply to Richard Biener from comment #21)
> > (In reply to Bill Schmidt from comment #20)
>
> ...<snip>...
> >
> > I see it only failing due to cost issues (tried ppc64le and -mcpu=power8).
> > The unaligned loads cost 3 and we end up with
> >
> > t.f90:8:0: note: Cost model analysis:
> > Vector inside of loop cost: 40
> > Vector prologue cost: 8
> > Vector epilogue cost: 4
> > Scalar iteration cost: 12
> > Scalar outside cost: 6
> > Vector outside cost: 12
> > prologue iterations: 0
> > epilogue iterations: 0
> > t.f90:8:0: note: cost model: the vector iteration cost = 40 divided by the
> > scalar iteration cost = 12 is greater or equal to the vectorization factor =
> > 1.
> >
> > Note that we are (still) not very good in estimating the SLP cost as we
> > account 4 vector loads here (because we essentially will end up with
> > 4 different permutations used), so the "unaligned" part is accounted for
> > too much and likely the permutation cost as well. Both are a limitation
> > of the SLP data structures and not easily fixable. With
> > -fvect-cost-model=unlimited I see both loops vectorized.
>
> Yes, I get these same results for the loop vectorizer (using -O2
> -ftree-vectorize -mcpu=power8 -ffast-math). But I was looking at the failure
> to do SLP vectorization. In comment 19 you indicated this was now working,
> presumably on x86, but for Power we fail to SLP-vectorize
> fast-math-pr37021.f90:9:0.
Err, I meant loop SLP vectorization as opposed to loop vectorization
with interleaving... Basic-block SLP doesn't work because (at least)
it does not handle reductions yet (I have done some early work here
but wasn't able to finish it)
Ah, thank you for the clarification. So does this require -fvect-cost-model=unlimited on all targets? If so, then I'll move on; otherwise I'll have a look at the Power-specific cost issues. (In reply to Bill Schmidt from comment #25) > Ah, thank you for the clarification. So does this require > -fvect-cost-model=unlimited on all targets? If so, then I'll move on; > otherwise I'll have a look at the Power-specific cost issues. Though, reading back, I see your comment on this not being easily fixable, so I guess I know the answer. Thanks again for your help. |
subroutine to_product_of(self,a,b,a1,a2) complex(kind=8) :: self (:) complex(kind=8), intent(in) :: a(:,:) complex(kind=8), intent(in) :: b(:) integer a1,a2 self = ZERO do i = 1,a1 do j = 1,a2 self(i) = self(i) + a(i,j)*b(j) end do end do end subroutine There are numerous reasons this is not vectorized at the moment: 1) aliasing doesn't tell apart the data of self, a and b 2) the reduction is obfuscated by extra SSA_NAME copies from store-motion (after you fix 1) 3) REALPART_EXPR < (*ptr)[i] > is not handled by data dependency analysis 4) ... ?