movups is not good. It'll be a lot faster (and portable) if you use movaps.

Process looks something like:
  * do the first few from a[0] until a's alignment interval as scalar
  * load the left of b's aligned pair
  * loop for each aligned vector in a
    - load a[n..n+4] aligned
    - load the right of b's pair
    - combine left~right and shift left to match elements against a
    - left = right
  * perform stragglers as scalar

Your benchmark is probably misleading too, because I suspect you are passing directly alloc-ed arrays into the function (which are 16 byte aligned).
movups will be significantly slower if the pointers supplied are not 16 byte aligned.
Also, results vary significantly between chip manufacturers and revisions.


On 18 August 2013 14:55, Ilya Yaroshenko <ilyayaroshenko@gmail.com> wrote:
On Saturday, 17 August 2013 at 19:38:52 UTC, John Colvin wrote:
On Saturday, 17 August 2013 at 19:24:52 UTC, Ilya Yaroshenko wrote:
BTW: -march=native automatically implies -mtune=native

Thanks, I`ll remove mtune)

It would be really interesting if you could try writing the same code in c, both a scalar version and a version using gcc's vector instrinsics, to allow us to compare performance and identify areas for D to improve.

I am lazy )

I have looked at assembler code:

float, scalar (main loop):
.L191:
        vmovss  xmm1, DWORD PTR [rsi+rax*4]
        vfmadd231ss     xmm0, xmm1, DWORD PTR [rcx+rax*4]
        add     rax, 1
        cmp     rax, rdi
        jne     .L191


float, vector (main loop):
.L2448:
        vmovups ymm5, YMMWORD PTR [rax]
        sub     rax, -128
        sub     r11, -128
        vmovups ymm4, YMMWORD PTR [r11-128]
        vmovups ymm6, YMMWORD PTR [rax-96]
        vmovups ymm7, YMMWORD PTR [r11-96]
        vfmadd231ps     ymm3, ymm5, ymm4
        vmovups ymm8, YMMWORD PTR [rax-64]
        vmovups ymm9, YMMWORD PTR [r11-64]
        vfmadd231ps     ymm0, ymm6, ymm7
        vmovups ymm10, YMMWORD PTR [rax-32]
        vmovups ymm11, YMMWORD PTR [r11-32]
        cmp     rdi, rax
        vfmadd231ps     ymm2, ymm8, ymm9
        vfmadd231ps     ymm1, ymm10, ymm11
        ja      .L2448

float, vector (full):
        https://gist.github.com/9il/6258443


It is pretty optimized)


____
Best regards

Ilya