[x265] [PATCH] copy_cnt_16, AVX2 asm code as per new interface, performance improved from 14.22x to 23.57x on HASWELL-I5

[chen]

At 2014-09-10 16:05:40,praveen at multicorewareinc.com wrote:
># HG changeset patch
># User Praveen Tiwari
># Date 1410336330 -19800
># Node ID c5b3e04e4eba2fcc4298c225d11ab25e0da82558
># Parent  d29cb300975a491287abdfb6abd2a9d3141e99f0
>copy_cnt_16, AVX2 asm code as per new interface, performance improved from 14.22x to 23.57x on HASWELL-I5

for comment, use cycles is better, our testbench will show cycles, the speed up factor depends on CPU type and compiler.

> INIT_YMM avx2
>-cglobal copy_cnt_16, 3,5,5
>+cglobal copy_cnt_16, 3,5,7
>     add         r2d, r2d
>-    lea         r4, [r2 * 3]
>-    mov         r3d, 16/4
>-    ; NOTE: xorpd is faster than pxor
>-    xorpd       m4, m4
>-    xorpd       m3, m3
>-
>-.loop
>-    ; row 0
>+    lea         r3,  [r2 * 3]
>+    mov         r4d, 256/128
>+
>+    xorpd       m5, m5

m5 for psadbw only, why you spent a register before loop?
>+    xorpd       m6, m6
>+
>+.loop:
>+    ; row 0 - 1
>     movu        m0, [r1]
>+    movu        [r0], m0
>     movu        m1, [r1 + r2]
>+    movu        [r0 + 32], m1
>+
>+    vpacksswb    m0, m1

remove prefix 'v' is better, unless it is new AVX2 only instruction, x86inc.asm have rename macro
>+    pminub       m0, [pb_1]

in my demo, I use this style because I just use memory 2 times
In here, you use 4x2=8 times, so buffer memory into register is better
>+
>+    ; row 2 - 3
>+    movu        m1, [r1 + r2 * 2]
>+    movu        [r0 + 64], m1
>+    movu        m2, [r1 + r3]
>+    movu        [r0 + 96], m2
>+
>+    vpacksswb    m1, m2
>+    pminub       m1, [pb_1]
>+    paddb        m0, m1
>+
>+    ; row 4 - 5
>+    lea        r1, [r1 + r2 * 4]
>+    movu       m2, [r1]
>+    movu       [r0 + 128], m2
>+    movu       m3, [r1 + r2]
>+    movu       [r0 + 160], m3
>+
>+    vpacksswb    m2, m3
>+    pminub       m2, [pb_1]
>+
>+    ; row 6 - 7
>+    movu       m3, [r1 + r2 * 2]
>+    movu       [r0 + 192], m3
>+    movu       m4, [r1 + r3]
>+    movu       [r0 + 224], m4

offset more than 128 will encode as 4 bytes
>+
>+    vpacksswb    m3, m4
>+    pminub       m3, [pb_1]
>+    paddb        m2, m3
>+
>+    paddb        m0, m2
>+    paddb        m6, m0
>+
>+    add         r0, 256

offset more than 128 will encode as 4 bytes
>+    lea         r1, [r1 + 4 * r2]
>+    dec         r4d
>     jnz        .loop
> 
>     ; get count
>+    vextracti128 xm1, m6, 1
>+    paddb        xm6, xm1
>+    psadbw       xm6, xm5
>+    movhlps      xm1, xm6
>+    paddd        xm6, xm1
>+    movd         eax, xm6
>     RET
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.videolan.org/pipermail/x265-devel/attachments/20140911/85630d96/attachment-0001.html>