[Bug target/51179] New: poor vectorization on interlagos.

[Bug target/51179] New: poor vectorization on interlagos.

[Joost.VandeVondele at mat dot ethz.ch]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51179

             Bug #: 51179
           Summary: poor vectorization on interlagos.
    Classification: Unclassified
           Product: gcc
           Version: 4.6.1
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: target
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: Joost.VandeVondele@mat.ethz.ch


The following code executes significantly faster when compiled with the cray
compiler (gcc: 43.4s cray:7.7s for 100000000 calls)

SUBROUTINE smm_dnn_4_10_10_4_1_2_1(A,B,C)
   REAL(KIND=KIND(0.0D0))   :: C(4,10), B(10,10), A(4,10)
   INTEGER ::i,j,l
   DO j=           1 ,          10 ,           2
   DO l=           1 ,          10 ,           1
   DO i=           1 ,           4 ,           1
     C(i+0,j+0)=C(i+0,j+0)+A(i+0,l+0)*B(l+0,j+0)
     C(i+0,j+1)=C(i+0,j+1)+A(i+0,l+0)*B(l+0,j+1)
   ENDDO
   ENDDO
   ENDDO
 END SUBROUTINE

cray options: -h noomp -e m -F -ra -O2 -Oipa1 -v tst.f90
gfortran: -O3 -march=native -ffast-math 

which yields for gfortran:

-march=bdver1 -mcx16 -msahf -mno-movbe -maes -mpclmul -mpopcnt -mabm -mlwp
-mno-fma -mfma4 -mxop -mno-bmi -mno-tbm -mavx -msse4.2 -msse4.1

The cray code looks nice:
0000000000000000 <smm_dnn_4_10_10_4_1_2_1_>:
   0:   48 89 7c 24 f8          mov    %rdi,-0x8(%rsp)
   5:   48 89 74 24 f0          mov    %rsi,-0x10(%rsp)
   a:   48 89 54 24 e8          mov    %rdx,-0x18(%rsp)
   f:   c5 fc 10 02             vmovups (%rdx),%ymm0
  13:   c5 fc 10 4a 20          vmovups 0x20(%rdx),%ymm1
  18:   c5 fc 10 52 40          vmovups 0x40(%rdx),%ymm2
  1d:   c5 fc 10 5a 60          vmovups 0x60(%rdx),%ymm3
  22:   c5 fc 10 a2 80 00 00    vmovups 0x80(%rdx),%ymm4
  29:   00 
  2a:   c5 fc 10 aa a0 00 00    vmovups 0xa0(%rdx),%ymm5
  31:   00 
  32:   c5 fc 10 b2 c0 00 00    vmovups 0xc0(%rdx),%ymm6
  39:   00 
  3a:   c5 fc 10 ba e0 00 00    vmovups 0xe0(%rdx),%ymm7
  41:   00 
  42:   c5 7c 10 82 00 01 00    vmovups 0x100(%rdx),%ymm8
  49:   00 
  4a:   c5 7c 10 8a 20 01 00    vmovups 0x120(%rdx),%ymm9
  51:   00 
  52:   31 c0                   xor    %eax,%eax
  54:   48 89 c1                mov    %rax,%rcx
  57:   66 0f 1f 84 00 00 00    nopw   0x0(%rax,%rax,1)
  5e:   00 00 
  60:   c4 62 7d 19 94 c6 d0    vbroadcastsd 0x2d0(%rsi,%rax,8),%ymm10
  67:   02 00 00 
  6a:   c5 7c 10 1c 0f          vmovups (%rdi,%rcx,1),%ymm11
  6f:   c4 43 a5 69 c9 a0       vfmaddpd %ymm9,%ymm10,%ymm11,%ymm9
  75:   c4 62 7d 19 94 c6 80    vbroadcastsd 0x280(%rsi,%rax,8),%ymm10
  7c:   02 00 00 
  7f:   c4 43 a5 69 c0 a0       vfmaddpd %ymm8,%ymm10,%ymm11,%ymm8
  85:   c4 62 7d 19 94 c6 30    vbroadcastsd 0x230(%rsi,%rax,8),%ymm10
  8c:   02 00 00 
  8f:   c4 e3 a5 69 ff a0       vfmaddpd %ymm7,%ymm10,%ymm11,%ymm7
  95:   c4 62 7d 19 94 c6 e0    vbroadcastsd 0x1e0(%rsi,%rax,8),%ymm10
  9c:   01 00 00 
  9f:   c4 e3 a5 69 f6 a0       vfmaddpd %ymm6,%ymm10,%ymm11,%ymm6
  a5:   c4 62 7d 19 94 c6 90    vbroadcastsd 0x190(%rsi,%rax,8),%ymm10
  ac:   01 00 00 
  af:   c4 e3 a5 69 ed a0       vfmaddpd %ymm5,%ymm10,%ymm11,%ymm5
  b5:   c4 62 7d 19 94 c6 40    vbroadcastsd 0x140(%rsi,%rax,8),%ymm10
  bc:   01 00 00 
  bf:   c4 e3 a5 69 e4 a0       vfmaddpd %ymm4,%ymm10,%ymm11,%ymm4
  c5:   c4 62 7d 19 94 c6 f0    vbroadcastsd 0xf0(%rsi,%rax,8),%ymm10
  cc:   00 00 00 
  cf:   c4 e3 a5 69 db a0       vfmaddpd %ymm3,%ymm10,%ymm11,%ymm3
  d5:   c4 62 7d 19 94 c6 a0    vbroadcastsd 0xa0(%rsi,%rax,8),%ymm10
  dc:   00 00 00 
  df:   c4 e3 a5 69 d2 a0       vfmaddpd %ymm2,%ymm10,%ymm11,%ymm2
  e5:   c4 62 7d 19 54 c6 50    vbroadcastsd 0x50(%rsi,%rax,8),%ymm10
  ec:   c4 e3 a5 69 c9 a0       vfmaddpd %ymm1,%ymm10,%ymm11,%ymm1
  f2:   c4 62 7d 19 14 c6       vbroadcastsd (%rsi,%rax,8),%ymm10
  f8:   c4 e3 a5 69 c0 a0       vfmaddpd %ymm0,%ymm10,%ymm11,%ymm0
  fe:   48 83 c1 20             add    $0x20,%rcx
 102:   48 ff c0                inc    %rax
 105:   48 83 f8 0a             cmp    $0xa,%rax
 109:   0f 8c 51 ff ff ff       jl     60 <smm_dnn_4_10_10_4_1_2_1_+0x60>
 10f:   c5 78 11 8a 20 01 00    vmovups %xmm9,0x120(%rdx)
 116:   00 
 117:   c4 63 7d 19 8a 30 01    vextractf128 $0x1,%ymm9,0x130(%rdx)
 11e:   00 00 01 
 121:   c5 78 11 82 00 01 00    vmovups %xmm8,0x100(%rdx)
 128:   00 
 129:   c4 63 7d 19 82 10 01    vextractf128 $0x1,%ymm8,0x110(%rdx)
 130:   00 00 01 
 133:   c5 f8 11 ba e0 00 00    vmovups %xmm7,0xe0(%rdx)
 13a:   00 
 13b:   c4 e3 7d 19 ba f0 00    vextractf128 $0x1,%ymm7,0xf0(%rdx)
 142:   00 00 01 
 145:   c5 f8 11 b2 c0 00 00    vmovups %xmm6,0xc0(%rdx)
 14c:   00 
 14d:   c4 e3 7d 19 b2 d0 00    vextractf128 $0x1,%ymm6,0xd0(%rdx)
 154:   00 00 01 
 157:   c5 f8 11 aa a0 00 00    vmovups %xmm5,0xa0(%rdx)
 15e:   00 
 15f:   c4 e3 7d 19 aa b0 00    vextractf128 $0x1,%ymm5,0xb0(%rdx)
 166:   00 00 01 
 169:   c5 f8 11 a2 80 00 00    vmovups %xmm4,0x80(%rdx)
 170:   00 
 171:   c4 e3 7d 19 a2 90 00    vextractf128 $0x1,%ymm4,0x90(%rdx)
 178:   00 00 01 
 17b:   c5 f8 11 5a 60          vmovups %xmm3,0x60(%rdx)
 180:   c4 e3 7d 19 5a 70 01    vextractf128 $0x1,%ymm3,0x70(%rdx)
 187:   c5 f8 11 52 40          vmovups %xmm2,0x40(%rdx)
 18c:   c4 e3 7d 19 52 50 01    vextractf128 $0x1,%ymm2,0x50(%rdx)
 193:   c5 f8 11 4a 20          vmovups %xmm1,0x20(%rdx)
 198:   c4 e3 7d 19 4a 30 01    vextractf128 $0x1,%ymm1,0x30(%rdx)
 19f:   c5 f8 11 02             vmovups %xmm0,(%rdx)
 1a3:   c4 e3 7d 19 42 10 01    vextractf128 $0x1,%ymm0,0x10(%rdx)
 1aa:   c5 f8 77                vzeroupper 
 1ad:   c3                      retq   
 1ae:   66 90                   xchg   %ax,%ax

gcc's code looks more involved:

smm_dnn_4_10_10_4_1_2_1_:
.LFB0:
        pushq   %rbp
.LCFI0:
        movl    $1, %eax
        movq    %rsp, %rbp
.LCFI1:
        andq    $-32, %rsp
        subq    $616, %rsp
.LCFI2:
        vmovupd 96(%rdi), %ymm0
        vmovupd (%rdi), %ymm3
        vmovupd 32(%rdi), %ymm1
        vmovsd  280(%rdi), %xmm13
        vmovupd 64(%rdi), %ymm2
        vmovsd  288(%rdi), %xmm15
        vmovsd  256(%rdi), %xmm4
        vmovsd  264(%rdi), %xmm6
        vmovsd  272(%rdi), %xmm7
        vmovupd 128(%rdi), %ymm12
        vmovsd  %xmm13, 296(%rsp)
        vmovupd 160(%rdi), %ymm11
        vperm2f128      $32, %ymm1, %ymm3, %ymm13
        vmovsd  %xmm15, 288(%rsp)
        vperm2f128      $49, %ymm1, %ymm3, %ymm1
        vmovsd  %xmm4, 320(%rsp)
        vperm2f128      $32, %ymm0, %ymm2, %ymm15
        vmovsd  296(%rdi), %xmm4
        vperm2f128      $49, %ymm0, %ymm2, %ymm2
        vmovsd  %xmm6, 312(%rsp)
        vmovaps %ymm1, 40(%rsp)
        vunpcklpd       %ymm1, %ymm13, %ymm1
        vmovsd  304(%rdi), %xmm6
        vunpcklpd       %ymm2, %ymm15, %ymm0
        vmovsd  %xmm7, 304(%rsp)
        vmovsd  312(%rdi), %xmm7
        vmovaps %ymm2, -24(%rsp)
        vperm2f128      $32, %ymm0, %ymm1, %ymm2
        vmovupd 192(%rdi), %ymm10
        vperm2f128      $49, %ymm0, %ymm1, %ymm0
        vmovsd  %xmm4, 280(%rsp)
        vmovsd  %xmm6, 336(%rsp)
        vmovaps %ymm13, %ymm4
        vmovsd  %xmm7, 328(%rsp)
        vmovaps %ymm15, %ymm6
        vmovaps %ymm2, %ymm7
        vunpcklpd       %ymm0, %ymm2, %ymm8
        vmovupd 224(%rdi), %ymm9
        vmovaps %ymm13, 72(%rsp)
        vmovaps %ymm15, 8(%rsp)
        vmovaps %ymm2, -56(%rsp)
        vmovaps %ymm0, -88(%rsp)
        vxorps  %xmm0, %xmm0, %xmm0
.L3:
        vunpckhpd       40(%rsp), %ymm4, %ymm3
        vmovupd (%rsi), %ymm4
        vunpckhpd       -24(%rsp), %ymm6, %ymm1
        vunpckhpd       -88(%rsp), %ymm7, %ymm5
        vperm2f128      $32, %ymm1, %ymm3, %ymm2
        vperm2f128      $49, %ymm1, %ymm3, %ymm1
        vfmaddpd        %ymm0, %ymm5, %ymm4, %ymm15
        vfmaddpd        %ymm0, %ymm8, %ymm4, %ymm3
        vunpcklpd       %ymm1, %ymm2, %ymm6
        vunpckhpd       %ymm1, %ymm2, %ymm2
        vmovupd 80(%rsi), %ymm1
        vfmaddpd        %ymm0, %ymm6, %ymm4, %ymm13
        vfmaddpd        %ymm0, %ymm2, %ymm4, %ymm4
        vmovaps %ymm15, 200(%rsp)
        vmovsd  320(%rsp), %xmm15
        vfmaddpd        %ymm0, %ymm8, %ymm1, %ymm14
        vfmaddpd        %ymm0, %ymm6, %ymm1, %ymm6
        vfmaddpd        %ymm0, %ymm5, %ymm1, %ymm5
        vfmaddpd        %ymm0, %ymm2, %ymm1, %ymm1
        vperm2f128      $32, %ymm11, %ymm12, %ymm2
        vmovaps %ymm13, -120(%rsp)
        vmovsd  64(%rsi), %xmm13
        vmovaps %ymm4, 136(%rsp)
        vmovaps %ymm6, 232(%rsp)
        vfmaddsd        (%rdx), %xmm15, %xmm13, %xmm15
        vmovaps %ymm1, 104(%rsp)
        vperm2f128      $49, %ymm11, %ymm12, %ymm1
        vmovaps %ymm5, 168(%rsp)
        vperm2f128      $32, %ymm9, %ymm10, %ymm5
        vunpcklpd       %ymm1, %ymm2, %ymm6
        vmovsd  %xmm13, 344(%rsp)
        vunpckhpd       %ymm1, %ymm2, %ymm2
        vperm2f128      $49, %ymm9, %ymm10, %ymm1
        vunpcklpd       %ymm1, %ymm5, %ymm4
        vmovsd  %xmm15, 352(%rsp)
        vunpckhpd       %ymm1, %ymm5, %ymm1
        vperm2f128      $32, %ymm4, %ymm6, %ymm5
        vperm2f128      $49, %ymm4, %ymm6, %ymm4
        vunpcklpd       %ymm4, %ymm5, %ymm7
        vunpckhpd       %ymm4, %ymm5, %ymm5
        vperm2f128      $32, %ymm1, %ymm2, %ymm4
        vperm2f128      $49, %ymm1, %ymm2, %ymm1
        vmovupd 32(%rsi), %ymm2
        vunpcklpd       %ymm1, %ymm4, %ymm6
        vunpckhpd       %ymm1, %ymm4, %ymm4
        vmovupd 112(%rsi), %ymm1
        vfmaddpd        %ymm3, %ymm7, %ymm2, %ymm3
        vfmaddpd        %ymm14, %ymm7, %ymm1, %ymm7
        vhaddpd %ymm3, %ymm3, %ymm3
        vhaddpd %ymm7, %ymm7, %ymm7
        vperm2f128      $1, %ymm3, %ymm3, %ymm15
        vaddpd  %ymm15, %ymm3, %ymm3
        vmovaps %ymm3, 584(%rsp)
        vmovsd  352(%rsp), %xmm3
        vaddsd  584(%rsp), %xmm3, %xmm3
        vmovsd  144(%rsi), %xmm15
        vmovsd  %xmm3, 264(%rsp)
        vmovsd  320(%rsp), %xmm3
        vfmaddsd        32(%rdx), %xmm3, %xmm15, %xmm13
        vperm2f128      $1, %ymm7, %ymm7, %ymm3
        vaddpd  %ymm3, %ymm7, %ymm3
        vmovaps %ymm3, 552(%rsp)
        vmovsd  312(%rsp), %xmm3
        vaddsd  552(%rsp), %xmm13, %xmm13
        vmovsd  %xmm13, 272(%rsp)
        vmovsd  344(%rsp), %xmm13
        vfmaddsd        8(%rdx), %xmm3, %xmm13, %xmm7
        vfmaddpd        -120(%rsp), %ymm6, %ymm2, %ymm13
        vhaddpd %ymm13, %ymm13, %ymm13
        vperm2f128      $1, %ymm13, %ymm13, %ymm3
        vaddpd  %ymm3, %ymm13, %ymm3
        vmovaps %ymm3, 520(%rsp)
        vaddsd  520(%rsp), %xmm7, %xmm7
        vmovsd  %xmm7, 352(%rsp)
        vfmaddpd        232(%rsp), %ymm6, %ymm1, %ymm6
        vmovsd  312(%rsp), %xmm13
        vfmaddsd        40(%rdx), %xmm13, %xmm15, %xmm7
        vhaddpd %ymm6, %ymm6, %ymm6
        vperm2f128      $1, %ymm6, %ymm6, %ymm3
        vaddpd  %ymm3, %ymm6, %ymm3
        vmovsd  304(%rsp), %xmm6
        vmovaps %ymm3, 488(%rsp)
        vmovsd  344(%rsp), %xmm3
        vaddsd  488(%rsp), %xmm7, %xmm13
        vfmaddsd        16(%rdx), %xmm6, %xmm3, %xmm7
        vfmaddpd        200(%rsp), %ymm5, %ymm2, %ymm3
        vfmaddpd        168(%rsp), %ymm5, %ymm1, %ymm5
        vfmaddpd        136(%rsp), %ymm4, %ymm2, %ymm2
        vfmaddpd        104(%rsp), %ymm4, %ymm1, %ymm1
        vmovsd  288(%rsp), %xmm4
        vhaddpd %ymm3, %ymm3, %ymm3
        vhaddpd %ymm5, %ymm5, %ymm5
        vhaddpd %ymm2, %ymm2, %ymm2
        vhaddpd %ymm1, %ymm1, %ymm1
        vperm2f128      $1, %ymm3, %ymm3, %ymm6
        vaddpd  %ymm6, %ymm3, %ymm3
        vmovaps %ymm3, 456(%rsp)
        vperm2f128      $1, %ymm5, %ymm5, %ymm3
        vaddpd  %ymm3, %ymm5, %ymm3
        vaddsd  456(%rsp), %xmm7, %xmm14
        vmovsd  304(%rsp), %xmm7
        vfmaddsd        48(%rdx), %xmm7, %xmm15, %xmm6
        vmovsd  296(%rsp), %xmm7
        vmovaps %ymm3, 424(%rsp)
        vmovsd  344(%rsp), %xmm3
        vfmaddsd        24(%rdx), %xmm7, %xmm3, %xmm5
        vperm2f128      $1, %ymm2, %ymm2, %ymm3
        vaddpd  %ymm3, %ymm2, %ymm2
        vaddsd  424(%rsp), %xmm6, %xmm6
        vmovaps %ymm2, 392(%rsp)
        vperm2f128      $1, %ymm1, %ymm1, %ymm2
        vaddpd  %ymm2, %ymm1, %ymm1
        vfmaddsd        56(%rdx), %xmm7, %xmm15, %xmm15
        vaddsd  392(%rsp), %xmm5, %xmm5
        vmovaps %ymm1, 360(%rsp)
        vmovsd  72(%rsi), %xmm2
        vmovsd  152(%rsi), %xmm1
        addq    $160, %rsi
        vaddsd  360(%rsp), %xmm15, %xmm15
        vfmaddsd        264(%rsp), %xmm4, %xmm2, %xmm3
        vmovsd  %xmm3, (%rdx)
        vfmaddsd        272(%rsp), %xmm4, %xmm1, %xmm3
        vmovsd  %xmm3, 32(%rdx)
        vmovsd  280(%rsp), %xmm3
        vfmaddsd        352(%rsp), %xmm3, %xmm2, %xmm7
        vmovsd  %xmm7, 8(%rdx)
        vfmaddsd        %xmm13, %xmm3, %xmm1, %xmm7
        vfmaddsd        %xmm6, 336(%rsp), %xmm1, %xmm6
        vfmaddsd        %xmm5, 328(%rsp), %xmm2, %xmm5
        vfmaddsd        %xmm15, 328(%rsp), %xmm1, %xmm1
        vmovsd  %xmm7, 40(%rdx)
        vfmaddsd        %xmm14, 336(%rsp), %xmm2, %xmm7
        vmovsd  %xmm6, 48(%rdx)
        vmovsd  %xmm5, 24(%rdx)
        vmovsd  %xmm1, 56(%rdx)
        vmovsd  %xmm7, 16(%rdx)
        addq    $64, %rdx
        cmpl    $9, %eax
        je      .L1
        addl    $2, %eax
        vmovaps 72(%rsp), %ymm4
        vmovaps 8(%rsp), %ymm6
        vmovaps -56(%rsp), %ymm7
        jmp     .L3
        .p2align 5,,7
        .p2align 3
.L1:
        leave
.LCFI3:

[Bug target/51179] poor vectorization on interlagos.

[ubizjak at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51179

--- Comment #1 from Uros Bizjak <ubizjak at gmail dot com> 2011-11-22 10:19:40 UTC ---
What about current 4.7 SVN?

[Bug tree-optimization/51179] poor vectorization on interlagos.

[ubizjak at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51179

Uros Bizjak <ubizjak at gmail dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2011-11-22
                 CC|                            |irar at il dot ibm.com
          Component|target                      |tree-optimization
     Ever Confirmed|0                           |1

--- Comment #2 from Uros Bizjak <ubizjak at gmail dot com> 2011-11-22 11:33:24 UTC ---
We can start here with something that hopefully resembles your original fortran
code:

--cut here--
double C[10][4], B[10][10], A[10][4];

void test (void)
{
  int i = 0, j = 0, l = 0;

//for (; j < 10; j += 2)
//  for (; l < 10; l++)
      for (; i < 4; i++)
        {
          C[j+0][i] = C[j+0][i] + A[l][i] * B[j+0][l];
          C[j+1][i] = C[j+1][i] + A[l][i] * B[j+1][l];
        }
}
--cut here--

gcc -O3 -ffast-math -mfma4 -mavx:

test:
        vmovapd A(%rip), %ymm0
        vbroadcastsd    B(%rip), %ymm1
        vfmaddpd        C(%rip), %ymm1, %ymm0, %ymm1
        vmovapd %ymm1, C(%rip)
        vbroadcastsd    B+80(%rip), %ymm1
        vfmaddpd        C+32(%rip), %ymm1, %ymm0, %ymm0
        vmovapd %ymm0, C+32(%rip)
        vzeroupper
        ret

Nice.

Now uncomment the second loop ("l" index) and this kernel will break:

        < ... lots of code deleted ... >
.L3:
        vmovupd (%r8,%rax), %xmm1
        addl    $1, %esi
        vinsertf128     $0x1, 16(%r8,%rax), %ymm1, %ymm1
        vfmaddpd        %ymm0, %ymm5, %ymm1, %ymm0
        vmovapd %ymm0, (%rbx,%rax)
        vmovupd (%rcx,%rax), %xmm0
        vinsertf128     $0x1, 16(%rcx,%rax), %ymm0, %ymm0
        vfmaddpd        %ymm0, %ymm4, %ymm1, %ymm0
        vmovupd %xmm0, (%rcx,%rax)
        vextractf128    $0x1, %ymm0, 16(%rcx,%rax)
        addq    $32, %rax
        cmpl    %r10d, %esi
        jb      .L3
        < ... lots of code deleted ... >

This already happens in the tree optimizers (vectorizer), RTL is just following
this trail.

Confirmed as a vectorizer problem.

[Bug tree-optimization/51179] poor vectorization on interlagos.

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51179

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |jakub at gcc dot gnu.org

--- Comment #3 from Jakub Jelinek <jakub at gcc dot gnu.org> 2011-11-22 17:13:26 UTC ---
Your testcase doesn't ressemble the original, the inner for cycles need
clearing of the iteration variable.
double C[10][4], B[10][10], A[10][4];

void
test (void)
{
  int i = 0, j = 0, l = 0;

  for (l = 0; l < 10; l++)
    for (j = 0; j < 10; j += 2)
      for (i = 0; i < 4; i++)
        {
          C[j + 0][i] = C[j + 0][i] + A[l][i] * B[j + 0][l];
          C[j + 1][i] = C[j + 1][i] + A[l][i] * B[j + 1][l];
        }
}

is IMHO just a matter whether graphite can -floop-interchange this or not.
If you swap manually the l and j for lines, the generated code looks better,
though for some reason we unroll even the l loop which increases register
pressure too much.

[Bug tree-optimization/51179] poor vectorization on interlagos.

[Joost.VandeVondele at mat dot ethz.ch]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51179

--- Comment #4 from Joost VandeVondele <Joost.VandeVondele at mat dot ethz.ch> 2011-11-22 18:34:03 UTC ---
Created attachment 25887
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=25887
general code

the more general code used to find the most efficient matrix multiply for sizes
4,10,10

[Bug tree-optimization/51179] poor vectorization on interlagos.

[Joost.VandeVondele at mat dot ethz.ch]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51179

--- Comment #5 from Joost VandeVondele <Joost.VandeVondele at mat dot ethz.ch> 2011-11-22 18:34:48 UTC ---
(In reply to comment #3)
> is IMHO just a matter whether graphite can -floop-interchange this or not.
> If you swap manually the l and j for lines, the generated code looks better,
> though for some reason we unroll even the l loop which increases register
> pressure too much.

Unfortunately, the issue is not just loop ordering or loop unrolling. I have a
code generator which tries systematically all possible loop orderings, and all
possible unroll factors. For this testcase (matrix sizes 4,10,10) the best cray
output (this one) runs at 10.8 Gflops. The best gcc compiled version runs at
4.7 Gflops (smm_dnn_4_10_10_1_1_10_2). I attach the test code, which I use for
testing.

[Bug tree-optimization/51179] poor vectorization on interlagos.

[dominiq at lps dot ens.fr]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51179

--- Comment #6 from Dominique d'Humieres <dominiq at lps dot ens.fr> 2011-11-22 21:08:47 UTC ---
> ... I attach the test code, which I use for testing.

Compiling the code with -O3 gives the following ICE

pr51179_1.f90: In function 'tiny_find':
pr51179_1.f90:3594:0: internal compiler error: in check_loop_closed_ssa_use, at
tree-ssa-loop-manip.c:422

(if nobody beats me, I'll try to reduce the code and open a new pr). On a
2.5Ghz Core2Duo, I get 4.1Gflops with -O2 -ftree-vectorize.

[Bug tree-optimization/51179] poor vectorization on interlagos.

[ubizjak at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51179

--- Comment #7 from Uros Bizjak <ubizjak at gmail dot com> 2011-11-22 22:00:36 UTC ---
(In reply to comment #3)
> Your testcase doesn't ressemble the original, the inner for cycles need
> clearing of the iteration variable.

Ah, indeed... fingers were too fast.

One additional data point with -O2 -ftree-vectorize -mfma4 -mavx with all
loops:

        movslq  %r8d, %rax
        movl    $C+32, %edx
        xorl    %esi, %esi
        leaq    B(,%rax,8), %rcx
        movl    $C, %eax
.L3:
>>      vmovsd  80(%rcx), %xmm1
        addl    $2, %esi
        vmovapd A(%rdi), %ymm0
>>      vmovddup        %xmm1, %xmm1
        vbroadcastsd    (%rcx), %ymm2
        addq    $160, %rcx
>>      vinsertf128     $1, %xmm1, %ymm1, %ymm1
        vfmaddpd        (%rax), %ymm2, %ymm0, %ymm2
        vmovapd %ymm2, (%rax)
        addq    $64, %rax
        vfmaddpd        (%rdx), %ymm1, %ymm0, %ymm0
        vmovapd %ymm0, (%rdx)
        addq    $64, %rdx
        cmpl    $10, %esi
        jne     .L3

This could be just "vbroadcastsd 80(%rcx), %ymm1". For some reason combine pass
does not form it.

[Bug tree-optimization/51179] poor vectorization on interlagos.

[Joost.VandeVondele at mat dot ethz.ch]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51179

--- Comment #8 from Joost VandeVondele <Joost.VandeVondele at mat dot ethz.ch> 2011-11-23 08:34:59 UTC ---
(In reply to comment #6)
> (if nobody beats me, I'll try to reduce the code and open a new pr). 
If reproduced the ICE with 4.7, and started a delta reduce. It goes very
slowly, so not sure if this is going to beat you ;-).

[Bug tree-optimization/51179] poor vectorization on interlagos.

[Joost.VandeVondele at mat dot ethz.ch]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51179

--- Comment #9 from Joost VandeVondele <Joost.VandeVondele at mat dot ethz.ch> 2011-11-23 17:19:28 UTC ---
(In reply to comment #8)
> (In reply to comment #6)
> > (if nobody beats me, I'll try to reduce the code and open a new pr). 
> If reproduced the ICE with 4.7, and started a delta reduce. It goes very
> slowly, so not sure if this is going to beat you ;-).

FYI, this ICE on tiny_find is now PR51285

[Bug tree-optimization/51179] poor vectorization on interlagos.

[Joost.VandeVondele at mat dot ethz.ch]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51179

--- Comment #10 from Joost VandeVondele <Joost.VandeVondele at mat dot ethz.ch> 2011-11-23 20:11:17 UTC ---
(In reply to comment #1)
> What about current 4.7 SVN?

The fastest 4x10 . 10x10 multiply as found with tiny_find.f90 yields somewhat
better results with 4.7, but not quite as effective as Cray's compiler.

Cray: 11.363 Gflops
4.7.0: 6.250 Gflops
4.6.1: 5.102 Gflops

[Bug tree-optimization/51179] poor vectorization on interlagos.

[Joost.VandeVondele at mat dot ethz.ch]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51179

--- Comment #11 from Joost VandeVondele <Joost.VandeVondele at mat dot ethz.ch> 2012-06-30 11:26:59 UTC ---
It looks like this problem is solved in the current 4.7 and 4.8 branches. At
least on an avx machine, the best performance found by the code in comment #4
jumps from 5.3Gflops in 4.6 to 13.9Glfops in 4.7/4.8. Great work.

I can't test this right now on interlagos, but I guess this could be OK as
well.

[Bug tree-optimization/51179] poor vectorization on interlagos.

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51179

Richard Guenther <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |RESOLVED
         Resolution|                            |FIXED
   Target Milestone|---                         |4.7.2

--- Comment #12 from Richard Guenther <rguenth at gcc dot gnu.org> 2012-07-19 10:36:13 UTC ---
Thus, fixed.