From owner-chemistry&$at$&ccl.net Thu Sep 29 18:49:05 2005
From: "CCL" <owner-chemistry---server.ccl.net>
To: CCL
Subject: CCL: Cleaning up dusty deck fortran and converting to C/C++
Message-Id: <-29405-050929160616-28406-34rr3HPj6mMmi67mbdt+PA---server.ccl.net>
X-Original-From: "Alex. A. Granovsky" <gran---classic.chem.msu.su>
Content-Transfer-Encoding: 7bit
Content-Type: text/plain;
	charset="iso-8859-1"
Date: Fri, 30 Sep 2005 00:04:31 +0400
MIME-Version: 1.0


Sent to CCL by: "Alex. A. Granovsky" [gran---classic.chem.msu.su]

--Replace strange characters with the "at" sign to recover email address--.

Dear Perry,

> You also need the exact GCC. Which version are you using?
gcc version 3.3.3

> Anyway, when I use -msse2, I seem to get references to %xmmN and such
> in my assembly language output. What are you trying?
> 
> Anyway, assuming gcc 3 or better:
> 
> cc -S -std=c99 -march=pentium4 -msse2 -mfpmath=sse -c bar.c
> 
> For me, that turns the core of the loop into a set of sse vector ops.

Using command line:
gcc -S -std=c99 -march=pentium4 -msse2 -mfpmath=sse -c test.c,
I have gotten extremely unoptimized inner loop using scalar sse2 instructions.
After adding -O3 option to the command line above, the inner loop becomes:

.L10:
        movapd  %xmm1, %xmm2
        mulsd   (%ebx,%eax), %xmm2
        movsd   %xmm2, (%esi,%eax)
        addl    $8, %eax
        subl    $1, %edx
        jne     .L10

As you can see, the multiplication and memory references are still using scalar SSE2 
and loop is not unrolled. 

Using icc, loop becomes unrolled, but still scalar and preserves strict memory access ordering. 

Using ifort 9.0 (Windows version, thus MS style listing) & original Fortran code:

$B1$11:                         ; Preds $B1$9 $B1$11
        movapd    xmm2, XMMWORD PTR [esi+ebx*8]                 ;11.17
        mulpd     xmm2, xmm0                                    ;11.8
        movapd    xmm3, XMMWORD PTR [esi+ebx*8+16]              ;11.17
        movapd    xmm4, XMMWORD PTR [esi+ebx*8+32]              ;11.17
        movapd    xmm5, XMMWORD PTR [esi+ebx*8+48]              ;11.17
        movapd    XMMWORD PTR [edi+ebx*8], xmm2                 ;11.8
        mulpd     xmm3, xmm0                                    ;11.8
        movapd    XMMWORD PTR [edi+ebx*8+16], xmm3              ;11.8
        mulpd     xmm4, xmm0                                    ;11.8
        movapd    XMMWORD PTR [edi+ebx*8+32], xmm4              ;11.8
        mulpd     xmm5, xmm0                                    ;11.8
        movapd    XMMWORD PTR [edi+ebx*8+48], xmm5              ;11.8
        add       ebx, 8                                        ;11.8
        cmp       ebx, ecx                                      ;11.8
        jb        $B1$11        ; Prob 97%                      ;11.8
        jmp       $B1$16        ; Prob 100%                     ;11.8
        ALIGN     4
        ALIGN     4
                                ; LOE eax ecx ebx esi edi xmm0 xmm1
$B1$14:                         ; Preds $B1$9 $B1$14
        movsd     xmm2, QWORD PTR [esi+ebx*8]                   ;11.17
        movhpd    xmm2, QWORD PTR [esi+ebx*8+8]                 ;11.17
        mulpd     xmm2, xmm0                                    ;11.8
        movsd     xmm3, QWORD PTR [esi+ebx*8+16]                ;11.17
        movhpd    xmm3, QWORD PTR [esi+ebx*8+24]                ;11.17
        movsd     xmm4, QWORD PTR [esi+ebx*8+32]                ;11.17
        movhpd    xmm4, QWORD PTR [esi+ebx*8+40]                ;11.17
        movsd     xmm5, QWORD PTR [esi+ebx*8+48]                ;11.17
        movhpd    xmm5, QWORD PTR [esi+ebx*8+56]                ;11.17
        movapd    XMMWORD PTR [edi+ebx*8], xmm2                 ;11.8
        mulpd     xmm3, xmm0                                    ;11.8
        movapd    XMMWORD PTR [edi+ebx*8+16], xmm3              ;11.8
        mulpd     xmm4, xmm0                                    ;11.8
        movapd    XMMWORD PTR [edi+ebx*8+32], xmm4              ;11.8
        mulpd     xmm5, xmm0                                    ;11.8
        movapd    XMMWORD PTR [edi+ebx*8+48], xmm5              ;11.8
        add       ebx, 8                                        ;11.8
        cmp       ebx, ecx                                      ;11.8
        jb        $B1$14        ; Prob 97%                      ;11.8

As you can see, there are two versions of the inner loop for two possible cases 
of data alignment, as it should be. Both are unrolled & vectorized, i.e, use vectorized 
memory access and multiplication commands. 

Thus, I would suppose that in C99 passing arrays does not imply that there are no aliasing 
across them, on the other hand it is now impossible to use restrict qualifier to explicitly 
declare this. This effectively disables advanced unrolling & vectorization

Best regards,
Alex Granovsky