[x265] [PATCH] copy_cnt_16: avx2 asm code, improved 514.32 cycles -> 313.66 cycles

[chen]

right
 

At 2014-09-18 19:08:23,praveen at multicorewareinc.com wrote:
># HG changeset patch
># User Praveen Tiwari
># Date 1411038475 -19800
># Node ID 532f798f98d7c7f5c493a819046a45e29b2da16a
># Parent  e723ecc1e5c99c451cbc8034514b9dc590a2d4ef
>copy_cnt_16: avx2 asm code, improved 514.32 cycles -> 313.66 cycles
>
>diff -r e723ecc1e5c9 -r 532f798f98d7 source/common/x86/asm-primitives.cpp
>--- a/source/common/x86/asm-primitives.cpp	Thu Sep 18 15:30:18 2014 +0530
>+++ b/source/common/x86/asm-primitives.cpp	Thu Sep 18 16:37:55 2014 +0530
>@@ -1730,7 +1730,7 @@
>          * code is updated, avx2 version will be enabled */
> 
>         p.copy_cnt[BLOCK_8x8] = x265_copy_cnt_8_avx2;
>-        // p.copy_cnt[BLOCK_16x16] = x265_copy_cnt_16_avx2;
>+        p.copy_cnt[BLOCK_16x16] = x265_copy_cnt_16_avx2;
>         // p.copy_cnt[BLOCK_32x32] = x265_copy_cnt_32_avx2;
> 
> 
>diff -r e723ecc1e5c9 -r 532f798f98d7 source/common/x86/blockcopy8.asm
>--- a/source/common/x86/blockcopy8.asm	Thu Sep 18 15:30:18 2014 +0530
>+++ b/source/common/x86/blockcopy8.asm	Thu Sep 18 16:37:55 2014 +0530
>@@ -4159,69 +4159,48 @@
> 
> 
> INIT_YMM avx2
>-cglobal copy_cnt_16, 3,5,5
>+cglobal copy_cnt_16, 3, 5, 5
>     add         r2d, r2d
>-    lea         r4, [r2 * 3]
>-    mov         r3d, 16/4
>-    ; NOTE: xorpd is faster than pxor
>+    lea         r3,  [r2 * 3]
>+    mov         r4d, 16/4
>+
>+    mova        m3, [pb_1]
>     xorpd       m4, m4
>-    xorpd       m3, m3
>-
>-.loop
>-    ; row 0
>+
>+.loop:
>+    ; row 0 - 1
>     movu        m0, [r1]
>-    movu        xm1, [r1 + mmsize/2]
>-    pmovsxwd    m2, xm0
>-    pmovsxwd    m1, xm1
>-    movu        [r0 + 0 * mmsize], m2
>-    movu        [r0 + 1 * mmsize], m1
>-
>-    ; row 1
>+    movu        [r0], m0
>     movu        m1, [r1 + r2]
>-    movu        xm2, [r1 + r2 + mmsize/2]
>+    movu        [r0 + 32], m1
>+
>     packsswb    m0, m1
>-    pcmpeqb     m0, m3
>+    pminub      m0, m3
>+
>+    ; row 2 - 3
>+    movu        m1, [r1 + r2 * 2]
>+    movu        [r0 + 64], m1
>+    movu        m2, [r1 + r3]
>+    movu        [r0 + 96], m2
>+
>+    packsswb    m1, m2
>+    pminub      m1, m3
>+    paddb       m0, m1
>     paddb       m4, m0
>-    pmovsxwd    m1, xm1
>-    pmovsxwd    m2, xm2
>-    movu        [r0 + 2 * mmsize], m1
>-    movu        [r0 + 3 * mmsize], m2
>-
>-    ; move output pointer here to avoid 128 bytes offset limit
>-    add         r0, 4 * mmsize
>-
>-    ; row 2
>-    movu        m0, [r1 + r2 * 2]
>-    movu        xm1, [r1 + r2 * 2 + mmsize/2]
>-    pmovsxwd    m2, xm0
>-    pmovsxwd    m1, xm1
>-    movu        [r0 + 0 * mmsize], m2
>-    movu        [r0 + 1 * mmsize], m1
>-
>-    ; row 3
>-    movu        m1, [r1 + r4]
>-    movu        xm2, [r1 + r4 + mmsize/2]
>-    packsswb    m0, m1
>-    pcmpeqb     m0, m3
>-    paddb       m4, m0
>-    pmovsxwd    m1, xm1
>-    pmovsxwd    m2, xm2
>-    movu        [r0 + 2 * mmsize], m1
>-    movu        [r0 + 3 * mmsize], m2
>-
>-    add         r0, 4 * mmsize
>-    lea         r1, [r1 + r2 * 4]
>-    dec         r3d
>-    jnz        .loop
>+
>+    add         r0, 128
>+    lea         r1, [r1 + 4 * r2]
>+    dec         r4d
>+    jnz         .loop
> 
>     ; get count
>-    vextracti128 xm0, m4, 1
>-    paddb        xm0, xm4
>-    movhlps     xm1, xm0
>-    paddb       xm0, xm1
>-    paddb       xm0, [pb_32]
>-    psadbw      xm0, xm3
>-    movd        eax, xm0
>+    xorpd        m0,  m0
>+    vextracti128 xm1, m4, 1
>+    paddb        xm4, xm1
>+    psadbw       xm4, xm0
>+    movhlps      xm1, xm4
>+    paddd        xm4, xm1
>+    movd         eax, xm4
>     RET
> 
> ;--------------------------------------------------------------------------------------
>diff -r e723ecc1e5c9 -r 532f798f98d7 source/common/x86/const-a.asm
>--- a/source/common/x86/const-a.asm	Thu Sep 18 15:30:18 2014 +0530
>+++ b/source/common/x86/const-a.asm	Thu Sep 18 16:37:55 2014 +0530
>@@ -29,6 +29,8 @@
> 
> SECTION_RODATA 32
> 
>+const pb_1,        times 32 db 1
>+
> const hsub_mul,    times 16 db 1, -1
> const pw_1,        times 16 dw 1
> const pw_16,       times 16 dw 16
>@@ -53,7 +55,6 @@
> const pb_64,       times 16 db 64
> const pb_01,       times  8 db 0,1
> const pb_0,        times 16 db 0
>-const pb_1,        times 32 db 1
> const pb_a1,       times 16 db 0xa1
> const pb_3,        times 16 db 3
> const pb_8,        times 16 db 8
>_______________________________________________
>x265-devel mailing list
>x265-devel at videolan.org
>https://mailman.videolan.org/listinfo/x265-devel
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.videolan.org/pipermail/x265-devel/attachments/20140919/c23fea89/attachment-0001.html>